Datasets:

kjappelbaum

/

chemnlp-arxiv

like 1

# Introduction There are six essential elements of every triangle-three angles and three sides. The method of constructing a triangle varies according to the facts which are known about its sides and angles. It is important to know what is the minimum knowledge about the sides and angles which is necessary to construct a particular triangle. Clearly all triangles constructed in the same way with the same data must be identically equal, i. e. they must be of exactly the same size and shape and their areas must be the same. Triangles which are equal in all respects are called *congruent triangles*. The four sets of minimal conditions for two triangles to be congruent are set out in the following geometric criteria. - ***Criterion A**. Two triangles are congruent if two sides and the included angle of one triangle are respectively equal to two sides and the included angle of the other.* - ***Criterion B**. Two triangles are congruent if two angles and a side of one triangle are respectively equal to two angles and a side of the other.* - ***Criterion C**. Two triangles are congruent if the three sides of one triangle are respectively equal to the three sides of the other.* - ***Criterion D**. Two triangles are congruent if two sides and the angle opposite the greater side of one triangle are respectively equal to two sides and the angle opposite the greater side of the other.* It should be noted that in *criteria* *A* and *D* the sets of corresponding equal elements are two sides and an angle. In fact the angle given may be any one of the three angles of the triangle. The problem to "*Construct a triangle with two of its sides \(\,a\,\) and \(\,b,\) \(\,a<b,\) and angle \(\,\alpha\,\) opposite the smaller side*\" has not a unique solution. There can be two triangles each of them satisfying the given conditions. In the present paper we compare not congruent triangles with respect to given sets of corresponding elements and answer the question what are the geometric properties characterizing such couples of triangles. # Theoretical basis of the proposed method for comparing triangles In \(\,\triangle\, ABC\,\) and \(\,\triangle\, A_1B_1C_1\,\) it is convenient to use the notations \(\,AB=c,\,\) \(BC=a,\,\) \(CA=b;\) \(\;A_1B_1=c_1,\,\) \(B_1C_1=a_1,\,\) \(C_1A_1=b_1.\) Let \(\,\theta\,\) and \(\,\theta_1\,\) be two corresponding angles of these triangles. If for \(\triangle\, ABC\) and \(\triangle\, A_1B_1C_1\) the relations \(a=a_1,\; b=b_1\) and \(\theta=\theta_1\) between the corresponding elements hold, we consider the following cases. - The angle \(\theta\) is included between the sides \(a\) and \(b\), i. e. \(\theta= \angle ACB\), and respectively \(\theta_1= \angle A_1C_1B_1\). In this case the triangles are congruent in view of *Criterion A*. - Let \(a=b\) and correspondingly \(a_1=b_1\), i. e. \(\triangle\, ABC\) and \(\triangle\, A_1B_1C_1\) are isosceles. Since \(\theta=\theta_1\), the triangles are congruent as a consequence of *Criterion A*. - Let \(a > b\), correspondingly \(a_1 > b_1\), and the angle \(\theta\) is opposite the greater side \(a\). In this case the triangles are congruent in view of *Criterion D*. - Let \(a > b\), correspondingly \(a_1 > b_1\), and the angle \(\theta\) is opposite the smaller side \(b\). In this case the triangles are either congruent or not. - If the triangles are congruent, then the angles opposite the greater sides are necessarily equal. It could happen that the sum of the equal angles opposite the greater sides is two right angles. If so, the triangles are right-angled. - If the triangles are not congruent, then we show that the sum of the angles opposite the greater sides is always two right angles. We prove the following *Proof*. Since \(\triangle\, ABC\) and \(\triangle\, ABD\) are not congruent, then \(AC< AB\) (and hence \(AD< AB\)). Let us denote \(\angle ACB=\alpha\) and \(\angle ADB=\beta\). There are two possibilities for the location of \(\triangle\, ABC\) and \(\triangle\, ABD\) with respect to the straight line \(AB\). \((i)\) *The points \(C\) and \(D\) lie on opposite sides of \(AB\)*. The symmetry with respect to the straight line \(AB\) transforms \(\triangle\, ABD\) into its congruent \(\triangle\, ABG\) which lies on one and the same side of the axis of symmetry \(AB\) like \(\triangle\, ABC\) (fig. 1). Since \(\triangle\, ABC\ncong \triangle\, ABD\), then \(\triangle\, ABC\ncong \triangle\, ABG\). The condition \(\,\angle ABC=\angle ABD\) states that the straight line \(AB\) is the bisector of \(\angle DBC\). From the symmetry with respect to \(AB\) it follows that \(G\in BC\) and \(BG\neq BC\). Let, for instance, \(G/BC\). (The case \(C/BG\) is analogical.) It is clear that if the conditions of Lemma 2.1 are fulfilled for \(\triangle\, ABC\) and \(\triangle\, ABD\), then they are also valid for \(\triangle\, ABC\) and \(\triangle\, ABG\) and vice versa. Let us consider \(\triangle\, ABC\) and \(\triangle\, ABG\). The side \(AB\) and \(\angle ABC\) are common for both triangles. In view of the symmetry with respect to \(AB\) and \(AC=AD\), we get \(AD=AG=AC\). Hence, \(\triangle\, ACG\) is isosceles and \(\angle ACG=\alpha=\angle AGC\). The angles \(\angle AGC\) and \(\angle AGB=\angle ADB=\beta\) are adjacent and hence \(\angle AGC+\angle AGB=\angle ACB+\angle ADB=\alpha+\beta=180^0\). \((ii)\) *The points \(C\) and \(D\) lie on one and the same side of \(AB\)*. We consider \(\triangle\, ABC\) and \(\triangle\, ABG\) (in this case \(D\equiv G\)). The proof of the statement is as in \((i)\). Based on the above arguments we can formulate a theorem, which is a generalization of *criteria* *A* and *D* for congruence of triangles (see also, p. 12). The denotations \(AB=c,\,\) \(BC=a,\,\) \(CA=b;\) \(A_1B_1=c_1,\,\) \(B_1C_1=a_1,\,\) \(C_1A_1=b_1\) are usually used in \(\triangle \,ABC\) and \(\triangle\, A_1B_1C_1\) respectively. Lemma 2.1 and Theorem 2.4 can be used as alternative methods of comparing different triangles. # Application of Theorem 2.4 to two geometric problems The solutions of next selected problems are based on Theorem 2.4. *Proof*. It is given that the center \(G\) of the circumscribing circle \(k\) of \(\triangle\, FDE\) lies on the bisector of \(\angle ACB\) (fig. 2). Since \(\triangle\, CGD\) and \(\triangle \,CGF\) have a common side \(CG\), equal corresponding angles \(\angle DCG=\angle FCG\) and equal corresponding sides \(DG=FG\) (as radii of \(k\)), the conditions of Theorem 2.4 are satisfied. Then \(\triangle\, CGD\) and \(\triangle\, CGF\) are either congruent, or not congruent. \((i)\) If \(\triangle\, CGD\) and \(\triangle\, CGF\) are congruent, then \(CD=CF\) and hence \(CA=CB\), i. e. \(\triangle \,ABC\) is isosceles. \((ii)\) If \(\triangle\, CGD\) and \(\triangle\, CGF\) are not congruent, then in view of Lemma 2.1 \(\angle CDG+\angle CFG=180^0\) and the quadrilateral \(CDGF\) can be inscribed in a circle \(k'\) (fig. 2). It is easy to be seen that \(\triangle\, EFD\cong \triangle\, CDF\) and the circumscribing circles \(k\) and \(k'\) have equal radii. The circles \(k\) and \(k'\) are symmetrically located with respect to their common chord \(FD\). Since the center \(G\) of \(k\) lies on \(k'\), then the center \(G'\) of \(k'\) lies on \(k\). Hence, \(\triangle\, DGG'\cong\triangle\, FGG'\), both triangles are equilateral, \(\angle DGF=120^0\) and \(\angle ACB=60^0\). *Proof*. We use the denotations \(\angle BAC=2\alpha\), \(\angle ABC=2 \beta\), \(\angle ACB=2\gamma\). Since \(J\) is the cut point of the angle bisectors \(\,AA_1\,\) and \(\,BB_1\,\) of \(\triangle\, ABC\), then the straight line \(CJ\) is the bisector of \(\angle ACB\) and \(\alpha+\beta+\gamma=90^0\) (fig. 3). Since \(\angle CB_1J\) is an exterior angle of \(\triangle\, ABB_1\), then \(\angle CB_1J=2\alpha+\beta\). Since \(\angle CA_1J\) is an exterior angle of \(\triangle\, ABA_1\), then \(\angle CA_1J=2\beta+\alpha\). Let us compare \(\triangle\, CA_1J\) and \(\triangle\, CB_1J\). They have a common side \(CJ\), corresponding equal sides \(JA_1=JB_1\) and angles \(\angle A_1CJ=\angle B_1CJ\). The conditions of Theorem 2.4 are satisfied. Then \(\triangle\, CA_1J\) and \(\triangle\, CB_1J\) are either congruent, or not. \((i)\) If \(\triangle\, CA_1J\) and \(\triangle\, CB_1J\) are congruent, then their corresponding elements are equal, in particular \[\angle CB_1J=\angle CA_1J\;\Leftrightarrow\; 2\alpha+\beta=2\beta+\alpha\;\Leftrightarrow\; \alpha=\beta.\] Hence, \(\triangle\, ABC\) is isosceles with \(CA=CB\). \((ii)\) If \(\triangle\, CA_1J\) and \(\triangle\, CB_1J\) are not congruent, then with respect to Lemma 2.1 \[\angle CB_1J+\angle CA_1J=180^0 \;\Leftrightarrow\; (2\alpha+\beta)+(2\beta+\alpha)=180^0 \;\Leftrightarrow\; \alpha+\beta=60^0.\] Hence, \(\angle ACB=60^0\). # Groups of problems In this section we illustrate the composing technology of new problems as an interpretation of specific logical models. Our aim is the *basic problem* in each of the groups under consideration to be with (exclusive or not exclusive) disjunction as a logical structure in the conclusion and its proof to be based on Lemma 2.1 or on Theorem 2.4. ## Problems of group I Suitable logical models for formulation of *equivalent* problems and *generating* problems of a given problem are described in detail in [@NM1; @NM2]. The basic statements we need in this group of problems are: - \(t:=\{\) *A square with center \(O\) is inscribed in a \(\triangle\, ABC\) in the following way: the vertexes of the square lie on the sides of the triangle, in addition two of them lie on the side \(AB\).\(\}\)* - \(p:=\{\angle ACB=90^0\}\) - \(q:=\{CA=CB\}\) - \(r:=\{\angle ACO=\angle BCO\}\) We describe the logical scheme for the composition of the Basic problem 4.4, which has not exclusive disjunction as a logical structure in the conclusion: - First we formulate and prove the *generating* problems-Problem 4.1 with logical structure \(t\wedge p\rightarrow r\) and Problem 4.3 with logical structure \(t\wedge q\rightarrow r\). - To generate problems with logical structure \(\; (*)\quad t\wedge(p \vee q)\rightarrow r\) we use the logical equivalence \[(t\wedge p\rightarrow r)\wedge (t\wedge q\rightarrow r)\;\Leftrightarrow\; t\wedge (p\vee q) \rightarrow r.\] - Finally, the formulated *inverse* problem-Basic problem 4.4-to the problem with structure \((*)\) has the logical structure \(t\wedge r\rightarrow p \vee q\). *Proof*. Let the quadrilateral \(MNPQ\), \(M\in AB\), \(\,N\in AB\), \(\, P\in BC\), \(\,Q\in AC\), be the inscribed in \(\triangle\, ABC\) square (fig. 4). Since the diagonals of a square are equal, intersect at right angles, bisect each other and bisect the opposite angles, then \(OP=OQ\) and \(\angle POQ=90^0\). The quadrilateral \(OPCQ\) can be inscribed in a circle \(k\) with diameter \(PQ\). To the equal chords \(OQ\) and \(OP\) of \(k\) correspond equal angles, i. e. \(\angle ACO=\angle BCO\). *Proof*. Let the quadrilateral \(MNPQ\), \(M\in AB\), \(\,N\in AB\), \(\, P\in BC\), \(\,Q\in AC\), be the inscribed in \(\triangle\, ABC\) rectangle (fig. 5). Since the diagonals of a rectangle are equal and bisect each other, then \(OM=ON=OP=OQ\). Let \(CH\perp AB,\, H\in AB\). Provided that \(\triangle\, ABC\) is isosceles with \(CA=CB\), the point \(H\) is the middle point of \(AB\) and the straight line \(CH\) is the bisector of \(\angle ACB\). Because \(MQ \parallel NP,\; NP \parallel CH\) and \(MQ=NP\), it follows that \(\triangle\, AMQ \cong \triangle\, BNP\) (by *Criterion B*) and \(AM=BN\). Hence, \(H\) is also the middle point of \(MN\). Since \(\triangle\, MON\) is isosceles, then its median \(OH\) is also an altitude, i. e. \(OH\perp MN\). This means that \(O\in CH\) and \(\angle ACO=\angle BCO\). A special case of Problem 4.2 is Problem 4.3 with a logical structure \(t\wedge q\,\rightarrow \,r\). Now we formulate and prove the *Basic problem* in this group. *Proof*. Let the quadrilateral \(MNPQ\), \(M\in AB\), \(\,N\in AB\), \(\, P\in BC\), \(\,Q\in AC\), be the inscribed in \(\triangle\, ABC\) square (fig. 6). Since the diagonals of any square are equal, intersect at right angles, bisect each other and bisect the opposite angles, then \(OP=OQ\) and \(\angle OPQ=\angle OQP=45^0\). We compare \(\triangle\, CQO\) and \(\triangle\, CPO\). They have a common side \(CO\), respectively equal sides \(OQ=OP\) and angles \(\angle QCO=\angle PCO\). We compute that \(\angle CQO=\angle CAB+45^0\) and \(\angle CPO=\angle CBA+45^0\) as exterior angles of \(\triangle\, QAN\) and \(\triangle\, PBM\) respectively. In view of Theorem 2.4 \(\;\triangle\, CQO\) and \(\triangle\, CPO\) are either congruent or not. - If \(\triangle\, CQO\) and \(\triangle\, CPO\) are congruent, then \(\angle CQO=\angle CPO\) and hence \(\angle CAB=\angle CBA\), i. e. \(CA=CB\) and \(\triangle\, ABC\) is isosceles. In this case \(\angle ACB\) is either a right angle and \(\triangle\, ABC\) is isosceles right-angled, or not a right angle and \(\triangle\, ABC\) is only isosceles. - If \(\triangle\, CQO\) and \(\triangle\, CPO\) are not congruent then, in view of Lemma 2.1, \(\angle CQO+\angle CPO=180^0\) and hence \(\angle CAB+\angle CBA=90^0\), i. e. \(\triangle\, ABC\) is right-angled with \(\angle ACB=90^0\). We reformulate Problem 4.4 by keeping the condition of homogeneity of the conclusion. ## Problems of group II By formulating appropriate statements and giving suitable logical models we get two *generating* problems that are necessary for the construction of the Basic problem 4.9. The basic statements we need are: - \(t:=\{\)*In \(\triangle \,ABC\) the straight lines \(AA_1,\,A_1\in BC\), and \(BB_1,\, B_1\in AC\), are the bisectors of \(\angle CAB\) and \(\angle CBA\) respectively.\(\}\)* - \(p:=\{\angle ACB=60^0\}\) - \(q:=\{\angle CAB=120^0\}\) - \(r:=\{\angle BB_1A_1=30^0\}\) Since the sum of the angles of any triangle is equal to two right angles, the statements \(\,p\,\) and \(\,q\,\) are mutually exclusive. Hence, if \(\,p\,\) is true, so is \(\,\neg q\,\) and vice versa. We describe the logical scheme for the composition of the Basic problem 4.9, which has exclusive disjunction as a logical structure in the conclusion: - First we formulate and prove the *generating* problems-Problem 4.7 with logical structure \(t\wedge p\rightarrow r\) and Problem 4.8 with logical structure \(t\wedge q\rightarrow r\). - Since the statements \(\,p\,\) and \(\,q\,\) are mutually exclusive then the equivalences \(\;p\wedge \neg q\;\Leftrightarrow\; p\) and \(\; \neg p\wedge q \;\Leftrightarrow\; q\;\) are true. As a consequence of these facts problems with logical structures \(\;t\wedge p\;\rightarrow\; r\;\) and \(\;t\wedge (p\wedge \neg q)\;\rightarrow\; r\;\) are equivalent. So the problems with logical structures \(\;t\wedge q\;\rightarrow\; r\;\) and \(\;t\wedge (q\wedge \neg p)\;\rightarrow\; r\). To generate problems with logical structure \(\; (**)\quad t\wedge(p \veebar q)\rightarrow r\) we use the logical equivalence \[(t\wedge(p\wedge \neg q)\rightarrow r)\wedge (t\wedge(\neg p\wedge q)\rightarrow r)\quad \Leftrightarrow\quad t\wedge(p \veebar q)\rightarrow r.\] - Finally, the formulated *inverse* problem-the Basic problem 4.9-to the problem with structure \((**)\) has the logical structure \(t\wedge r\rightarrow p \veebar q\). *Proof*. Let \(\angle BAA_1=\angle CAA_1=\alpha\), \(\,\angle ABB_1=\angle CBB_1=\beta\), \(\, J=AA_1\cap BB_1\). Since \(J\) is the cut point of the angle bisectors of \(\triangle\, ABC\), then \(\angle JCA=\angle JCB=\gamma=30^0\) (fig. 7). Because \(\alpha+\beta+\gamma=90^0\) it follows that \(\angle AJB=120^0\). Hence, the quadrilateral \(CA_1JB_1\) can be inscribed in a circle. Then \(\angle JA_1B_1=\angle JCB_1=30^0\) and \(\;\angle JB_1A_1=\angle JCA_1=30^0\) as angles in the same segment of this circle. *Proof*. Let \(J=AA_1\cap BB_1\), \(\, E=A_1B_1\cap CJ\), \(\, C_1=CJ\cap AB\). Since \(\angle BAC=120^0\), then its adjacent angles have a measure of \(60^0\). It is easy to be seen that the point \(B_1\) is equidistant from the straight lines \(BA\), \(\,BC\), \(\,AA_1\) and that the straight line \(A_1B_1\) is the bisector of \(\angle CA_1A\) (fig. 8). The proof that the straight line \(A_1C_1\) is the bisector of \(\angle BA_1A\) is analogical. It follows that \(\angle B_1A_1C_1\) is a right angle (the bisectors of any two adjacent angles are perpendicular to each other) (see also [@M], p. 194, Problem 156). As a consequence we get that \(E\) is the intersection point of the angle bisectors \(CJ\) and \(A_1B_1\) of \(\triangle\, AA_1C\) and hence \(\angle JAE=\angle EAB_1=30^0\). Let \(\varphi=\angle CA_1B_1=\angle B_1A_1A\) and \(\gamma=\angle C_1CA=\angle C_1CB\). Then \(\angle A_1B_1C=60^0+\varphi\) as an exterior angle of \(\triangle\, A_1B_1A\), the sum of the angles of \(\triangle\, AA_1C\) is \(60^0+2\varphi+2\gamma=180^0\), i. e. \(\varphi+\gamma=60^0\) and hence \(\angle JEB_1=120^0\). Thus, the quadrilateral \(AJEB_1\) can be inscribed in a circle. We conclude that \(\angle JAE=\angle JB_1E=30^0\) as angles in the same segment of this circle. Hence, \(\angle BB_1A_1=30^0\). Now we formulate and prove the *Basic problem* in this group. *Proof*. Let us denote \(\angle BAA_1=\angle CAA_1=\alpha\), \(\;\angle ABB_1=\angle CBB_1=\beta\), \(AA_1\cap BB_1=J\). Since \(J\) is the cut point of the angle bisectors of \(\triangle\, ABC\), then the straight line \(CJ\) is the bisector of \(\angle ACB\). Denoting \(\;\gamma=\angle JCA=\angle JCB\;\) we get \(\alpha+\beta+\gamma=90^0\) (fig. 9). Let the point \(A'\) be orthogonally symmetric to the point \(A_1\) with respect to the axis \(BB_1\). It follows that \(A'\neq A\). (If \(A'\equiv A\) then \(\triangle\, ABC\) does not exist.) The straight line \(BB_1\) is the bisector of \(\angle ABC\) and consequently \(A'\in AB\) and \(B_1A_1=B_1A'\). On the other hand, \(\angle BB_1A_1= 30^0\) and hence \(\triangle\, A_1B_1A'\) is equilateral. We compute \(\angle AA'B_1=30^0+\beta\) (as an exterior angle of \(\triangle\, A'BB_1\)), \(\,\angle AA'A_1=90^0+\beta\) (as an exterior angle of \(\triangle\, A'BE\)), \(\,\angle AB_1A' =60^0+\gamma-\alpha\,\) and \(\,\angle AB_1A_1=120^0+\gamma-\alpha\). Let us compare \(\triangle\, AA_1B_1\) and \(\triangle\, AA_1A'\). They have a common side \(AA_1\) and corresponding equal sides \(A_1B_1=A_1A'\) and angles \(\angle B_1AA_1=\angle A'AA_1=\alpha\). In view of Theorem 2.4 we have the possibilities: - \(\triangle\, AA_1B_1\) and \(\triangle\, AA_1A'\) are congruent. Then \(\angle AB_1A_1=\angle AA'A_1\), i. e. \(120^0+\gamma-\alpha=90^0+\beta\). Hence, \(2 \gamma=\angle ACB=60^0\). - \(\triangle\, AA_1B_1\) and \(\triangle\, AA_1A'\) are not congruent. By Lemma 2.1 it follows that \(\angle AB_1A_1+\angle AA'A_1=180^0\), i. e. \((120^0+\gamma-\alpha)+(90^0+\beta)=180^0\). Hence, \(2 \alpha=\angle BAC=120^0\). In order to formulate a special type equivalent problem to this Basic problem we prove *Proof.* \[(\neg(p\veebar q))\;\Leftrightarrow\; \neg((p\wedge\neg q)\vee(\neg p\wedge q))\] \[\Leftrightarrow\; (p\vee \neg q)\wedge (\neg p\vee q)\;\Leftrightarrow\; p\wedge (\neg p\vee q)\vee\neg q\wedge (\neg p\vee q)\] \[\Leftrightarrow\; (p\wedge \neg p)\vee (p\wedge q) \vee (\neg q\wedge\neg p) \vee (q\wedge \neg q)\;\Leftrightarrow\; \neg p\wedge\neg q.\] Because of this Proposition problems with logical structures \(\; t\wedge (\neg(p\veebar q))\;\rightarrow\; \neg r\;\) and \(\;t\wedge (\neg p\wedge\neg q)\;\rightarrow\; \neg r\;\) are equivalent. The following problem is equivalent to the Basic problem 4.9. *Proof*. Assuming the truth of the contrary statement, i. e. \(\angle BB_1A_1 = 30^0\), the solution of this problem leads to the solution of the Basic problem 4.9. Acknowledgements. The first author is partially supported by Sofia University Grant 99/2013. The second author is partially supported by Sofia University Grant 159/2013.

{'timestamp': '2014-10-28T01:15:51', 'yymm': '1401', 'arxiv_id': '1401.5905', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5905'}

# Introduction and related work {#sec:related} Simplicial complexes were first introduced by Poincaré in 1895  to study the topology of spaces of arbitrary dimension, and are basic tools for algebraic topology , image analysis  and discrete surfaces , among many other domains. In the form of meshes they are widely used in many contexts to express tridimensional data. Some graphs can be represented as a form of simplicial complexes, and we can build simplicial complexes based on regular, matricial, images. This versatility is the reason we chose to use simplicial complexes as the operating space. Considering operators on simplicial complex spaces, it is fairly common to change the complexity of the mesh structure . Even when additional data is associated with the elements of the complex, they are mostly used to guide the change in the structure, the values themselves are not changed. Here, we pursuit a different option, our objective is to filter values associated to the elements of the complex, without changing its structure, using the framework of mathematical morphology. Mathematical morphology was introduced by Matheron and Serra in 1964 and it is one of the most important frameworks for non-linear image processing, providing tools for many applications. It was later extended by Heijmans and Ronse  using complete lattices, allowing the use of more complex digital structures, such as graphs , hypergraphs  and simplicial complexes . The use of a digital structure as support to image processing is not new. In , Vincent uses the lattice approach to mathematical morphology to define morphological operators on neighborhood graphs, where the graph structure is used to define neighborhood relationships between unorganized data, expressed as vertices. By allowing the propagation of values from vertices to the edges, therefore using the graph structure to express more than just neighborhood relation, Cousty   obtained different morphological operators, including openings, closings and alternating sequential filters. Those operators are capable of dealing with smaller noise structures, acting in a smaller size than the classical operators. Similarly, Meyer and Stawiaski  and Meyer and Angulo  obtain a new approach to image segmentation and levellings, respectively. Recently, Bloch and Bretto  introduced mathematical morphology on hypergraphs, defining lattices and operators. Their lattices and operators are similar to the ones presented here, taking into account the differences between hypergraphs and complexes. This work is focused on mathematical morphology on simplicial complexes, specifically to process values associated to elements of the complex in an unified manner, without altering the structure itself. In , Loménie and Stamon explore mathematical morphology operators on mesh spaces from point spaces. However, the complex only provides structural information, while the information itself is associated only to triangles or edges of the mesh. Our approach for mathematical morphology on simplicial complexes has been studied before and this article is an extension of the conference article, where interesting new operators were introduced. These operators, called *dimensional operators* can be used as building blocks for new operators. In this work we explore these operators, introducing composition properties and defining new morphological operators. The proofs are omitted here, but they are available in . We also revisit the related work, showing that most of the operators from the literature can be expressed by the dimensional operators. # Basic theoretical concepts {#ch:basic} The objective of this work is to explore the dimensional operators for mathematical morphology on simplicial complex spaces. To this end, we start by reminding useful definitions about simplicial complexes and mathematical morphology. ## Simplicial complexes {#sec:sc} One of the most known forms of complex is the concept of *mesh*, often used to express tridimensional data on various domains, such as computer aided design, animation and computer graphics in general. However, in this work we prefer to approach complexes by the combinatorial definition of an abstract complex . The basic element of a complex is a *simplex*. In this work, a simplex is a finite, nonempty set. The *dimension* of a simplex \(x\), denoted by \(dim(x)\), is the number of its elements minus one. A simplex of dimension \(n\) is also called an *n-simplex*. We call *simplicial complex*, or simply *complex*, any set \(X\) of simplices such that, for any \(x\in X\), any non-empty subset of \(x\) also belongs to \(X\). The *dimension* of a complex is equal to the greatest dimension of its simplices and, by convention, we set the *dimension* of the empty set to \(-1\). In the following, a complex of dimension \(n\) is also called an *\(n\)-complex*. Figure [\[figSimA\]](#figSimA){reference-type="ref" reference="figSimA"} (resp. b, and c) graphically represents a simplex \(x = \{a\}\) (resp. \(y = \{a,b\}\) and \(z = \{a,b,c\}\)) of dimension \(0\) (resp. \(1\), \(2\)). Figure [\[figSimD\]](#figSimD){reference-type="ref" reference="figSimD"} shows a set of simplices composed of one \(2\)-simplex (\(\left\{a,b,c\right\}\)), three \(1\)-simplices (\(\{a,b\}, \{b,c\}\) and \(\{a,c\}\)) and three \(0\)-simplices (\(\{a\}, \{b\}\) and \(\{c\}\)). **Important notations.** In this work, the symbol denotes a non-empty \(n\)-complex, with \(n\in\mathbb{N}\). The set of all subsets of \(\C\) is denoted by . Any subset of \(\C\) that is also a complex is called a *subcomplex (of \(\C\))*. We denote by \(\subcomplexes\) the set of all subcomplexes of \(\C\). If \(X\) is a subset of \(\C\), we denote by \(\overline{X}\) the *complement* of \(X\) (in ): \(\overline{X}=\C \backslash X\). The complement of a subcomplex of \(\C\) is usually not a subcomplex. Any subset \(X\) of \(\C\) whose complement \(\overline{X}\) is a subcomplex is called a *star*. We denote by \(\stars\) the set of all stars in \(\C\). The intersection \(\subcomplexes\,\cap\,\stars\) is non-empty since it always contains at least \(\emptyset\) and \(\C\). In the domain of simplicial complexes, some operators are well known, such as the *closure* and *star* . We define the closure \(\hat{x}\) and the star \(\check{x}\) of a simplex \(x\) as: \[\begin{aligned} \forall x \in \C,\,\hat{x}=&\left\{ y \st y \subseteq x, y \neq \emptyset \right\}\label{eq:hatx}\\ \forall x \in \C,\,\check{x}=&\left\{ y\in \C \st x \subseteq y\right\}\label{eq:checkx} \end{aligned}\] In other words, the closure operator gives as result the set of all simplices that are subsets of the simplex \(x\), and the star gives as result the set of all simplices of \(\C\) that contain the simplex \(x\). These operators can be easily extended to sets of simplices. The operators \(Cl:\PC \rightarrow \PC\) and \(St: \PC \rightarrow \PC\) are defined by: \[\begin{aligned} \forall X \in \PC, \ Cl =&\bigcup \{\hat{x} \st x\in X\}\label{eq:ClSetPC}\\ \forall X \in \PC ,\ St =&\bigcup\{ \check{x} \st x\in X\} \label{eq:StSetPC} \end{aligned}\] ## Mathematical morphology {#sec:mm} We approach mathematical morphology through the framework of lattices . We start with the concept of partially ordered set. It is composed by a set and a binary relation. The binary relation is defined only between certain pairs of elements of the set, representing precedence, and must be reflexive, antisymmetric and transitive. A *lattice* is a partially ordered set with a least upper bound, called *supremum*, and a greatest lower bound, called *infimum*. For instance, the set \(\mathcal{P}(S)=\{\{a,b,c\}, \{a,b\}, \{a,c\}, \{b,c\}, \{a\}, \{b\}, \{c\},\emptyset\}\), that is the power set of the set \(S=\{a,b,c\}\), ordered by the inclusion relation, is a lattice. The supremum of two elements of this lattice is given by the union operator and the infimum by the intersection operator. This lattice can be denoted by \(\langle\mathcal{P}(S),\bigcup,\bigcap,\subseteq\rangle\). In mathematical morphology (see, , ), any operator that associates elements of a lattice \(\L[1]\) to elements of a lattice \(\L[2]\) is called a *dilation* if it commutes with the supremum. Similarly, an operator that commutes with the infimum is called an *erosion*. Let \(\L[1]\) and \(\L[2]\) be two lattices whose order relations and suprema are denoted by \(\leq_1\), \(\leq_2\), \(\vee_1\), and \(\vee_2\). Two operators \(\alpha:\DomImL{2}{1}\) and \(\beta:\DomImL{1}{2}\) form an *adjunction* \((\beta,\alpha)\) if \(\alpha(a)\leq_{1} b \ \iff\ a \leq_{2} \beta(b)\) for every element \(a\) in \(\mathcal{L}_{2}\) and \(b\) in \(\mathcal{L}_{1}\). It is well known (see, , ) that, given two operators \(\alpha\) and \(\beta\), if the pair \((\beta,\alpha)\) is an adjunction, then \(\beta\) is an erosion and \(\alpha\) is a dilation. Furthermore, if \(\alpha\) is a dilation, there is a unique erosion \(\beta\), called the *adjoint of \(\alpha\)*, such that \((\beta,\alpha)\) is an adjunction. This erosion is characterized by: \[\forall a \in \L[1],\, \beta(a)=\vee_2 \left\{ b \in \L[2] \st \alpha(b) \leq_1 a \right\} \label{eq:AdjDef}\] In certain cases, we will denote the adjoint operator of an operator \(\alpha\) by \(\alpha\adjoint\), to explicit the relationship between them. In mathematical morphology, an operator \(\alpha\), acting from a lattice \(\L_1\) to \(L_2\), that is increasing (\(\forall a,b\in\L_1,\ a\leq b \implies \alpha(a)\leq \alpha(b)\)) and idempotent (\(\forall a\in\L_1,\ \alpha(a)=\alpha(\alpha(a))\)) is a *filter*. If a filter is anti-extensive (\(\forall a\in\L_1,\ \alpha(a)\leq a\)) it is called an *opening*. Similarly, an extensive filter (\(\forall a\in\L_1,\ a\leq \alpha(a)\)) is called a *closing*. One way of obtaining openings and closings is by combining dilations and erosions . Let \(\alpha:\DomIm{\L}{\L}\) be a dilation. Then the operator \(\zeta=\alpha\adjoint\alpha\) is a closing and the operator \(\psi=\alpha\alpha\adjoint\) is an opening. Both operators act on \(\L\). A family of openings \(\Psi=\{\psi_{\lambda} \st \lambda \in \mathbb{N}\}\) acting on \(\L\), is a *granulometry* if, given two positive integers \(i\) and \(j\), we have \(i \geq j \implies \psi_{i}(a) \subseteq \psi_{j}(a)\), for any \(a\in\L\) . Similarly, a family of closings \(Z=\{\zeta_{\lambda} \st \lambda \geq 0\}\), is a *anti-granulometry* if, given two positive integers \(i\) and \(j\), we have \(i \leq j \implies \zeta_{i}(a) \subseteq \zeta_{j}(a)\), for any \(a\in\L\). A family of filters \(\{ \alpha_\lambda, \lambda \in \N\}\) is a family of *alternating sequential filters* if, given two positive integers \(i\) and \(j\), we have \(i > j \implies \alpha_{i} \alpha_{j} = \alpha_{i}\). Let \(\Psi=\{\psi_{\lambda},\,\lambda\in\N\}\) be a granulometry and \(Z=\{\zeta_{\lambda},\,\lambda\in\N\}\) be an anti-granulometry. We can construct two alternating sequential filters by composing operators from both families. Let \(i\in\N\) and \(a\in\L\): \[\begin{aligned} \nu_i(a)= &\paren{\psi_{i}\zeta_{i}}\paren{\psi_{i-1}\zeta_{i-1}}\dots\paren{\psi_{1}\zeta_{1}} (a) \label{eq:asfDef}\\ \nu_i'(a)= &\paren{\zeta_{i}\psi_{i}}\paren{\zeta_{i-1}\psi_{i-1}}\dots\paren{\zeta_{1}\psi_{1}} (a) \label{eq:asfAltDef} \end{aligned}\] # Dimensional operators {#sec:DimOps} In , we introduced four new basic operators that act on simplices of given dimensions. These operators can be composed into new operators which behavior can be finely controlled. We proceed with a brief reminder of their definition and explore some new properties. We start by introducing a new notation that allows only simplices of a given dimension to be retrieved. Let \(X \subseteq \C\) and let \(i \in [0,n]\), we denote by \(\dimension{X}{i}\) the set of all \(i\)-simplices of \(X\): \(\dimension{X}{i} = \{x \in X \st dim(x) = i\}\). In particular, \(\C[i]\) is the set of all \(i\)-simplices of \(\C\). We denote by \(\PC[i]\) the set of all subsets of \(\C[i]\). Let \(i\in\N\) such that \(i\in [0,n]\). The structure \(\left\langle \PC[i], \bigcup,\bigcap,\subseteq \right\rangle\) is a lattice. In other words, \(\dij(X)\) is the set of all \(j\)-simplices of \(\C\) that include an \(i\)-simplex of \(X\), \(\dji(X)\) is the set of all \(i\)-simplices of \(\C\) that are included in a \(j\)-simplex of \(X\), \(\eij(X)\) is the set of all \(j\)-simplices of \(\C\) whose subsets of dimension \(i\) all belong to \(X\), and \(\eji(X)\) is the set of all \(i\)-simplices of \(\C\) that are not contained in any \(j\)-simplex of \(\overline{X}\). The dimensional operators can also be recovered using the classical star and closure operators: The dimensional operators can be useful when the considered data is associated only with simplices of a given dimension of the complex, which is fairly common. In this situation, these operators can be used to propagate the values to the other dimensions of the complex, or even filter the values directly, depending on the application. However, since the objective of this work is to find interesting operators acting on subcomplexes, we mostly use these operators as building blocks to define new operators. The following adjunction property can be proved by constructing the adjoint erosion of the dilation operators and verifying that they correspond to the provided erosion, the properties regarding duality are trivial results from property [\[p:DimOpAlt\]](#p:DimOpAlt){reference-type="ref" reference="p:DimOpAlt"}. We can use the dimensional operators from definition [\[def:DimOp\]](#def:DimOp){reference-type="ref" reference="def:DimOp"} to define new operators, leading to new dilations, erosions, openings, closings and alternating sequential filters. Before we start composing these operators, let us consider the following results, that can guide the exploration of new compositions. Property [\[p:CompOp\]](#p:CompOp){reference-type="ref" reference="p:CompOp"} states that any composition of the same operator is equivalent to the operator acting from the initial to the final dimension. The proof of this property can be done by contradiction, where if \(\dilp{j}{k}\dilp{i}{j}(X)\not=\dilp{i}{k}(X)\) is true, our space is not a simplicial complex. To explore the possible combinations of the operators from definition [\[def:DimOp\]](#def:DimOp){reference-type="ref" reference="def:DimOp"}, we start by considering only operators acting on the same dimension. The following property can be deduced from property [\[p:DimOpAlt\]](#p:DimOpAlt){reference-type="ref" reference="p:DimOpAlt"}: Property [\[p:OpInd\]](#p:OpInd){reference-type="ref" reference="p:OpInd"} states that the result of the compositions of dilations and erosions that use a higher intermediary dimension is independent of the dimension chosen. Therefore, we can obtain only one dilation, one erosion, one opening and one closing using those compositions. However, this is not entirely true when we consider a lower dimension as intermediary dimension for the compositions. The following property can be deduced from property [\[p:DimOpAlt\]](#p:DimOpAlt){reference-type="ref" reference="p:DimOpAlt"}: Property [\[p:FamilyDownUp\]](#p:FamilyDownUp){reference-type="ref" reference="p:FamilyDownUp"} states that compositions from dilations and erosions using a lower intermediary dimension are equal, independent of the chosen dimension and that compositions of only dilations and erosions are related, but not always equivalent. ## Extension to weighted complexes In this section, we extend the operators we defined to weighted simplicial complexes. Let \(k_{min}\) and \(k_{max}\) be two distinct, positive integers. We define the set \(\K\) as the set of the integers between these two numbers, \(\K=\{k\in\N \st k_{min}\leq k\leq k_{max}\}\). Now, let \(M\) be a map from \(\C\) to \(\K\), that associates an element of \(\K\) to every element of the simplicial complex \(\C\). Let \(x\in\C\), in this work, \(M(x)\) is called the *value* of the simplex \(x\). We can extend the notion of subcomplexes and stars to the domain of weighted complexes. A map \(M\) from \(\C\) in \(\K\) is a *simplicial stack* (see ) if the value of each simplex is smaller than or equal to the value of the simplices it includes, \(\forall x\in X,\,\forall y\subseteq x, M(x)\geq M(y)\). On the other hand, when the comparison is reversed, when \(\forall x\in \C,\,\forall y\subseteq x, M(x)\leq M(y)\), we say that \(M\) is a *starred stack*. The *dual \(\overline{M}\)* of a map \(M\) is defined using the value \(k_{max}\): \(\forall x\in \C, \overline{M}(x)=k_{max}-M(x)\). Let \(M\) be a map from an arbitrary set \(E\) in \(\K\) and let \(k\in\K\). We denote by \(M[k]\) the set of elements in \(E\) with value greater than or equal to \(k\), \(M[k]=\{x\in E \st M(x)\geq k\}\). This set is called the *\(k\)-threshold* of \(M\). The following lemma, which can be easily proved from the definitions, clarifies the links between stars, complexes, and the \(k\)-thresholds of simplicial stacks and starred stacks. We approach the problem of extending the dimensional operators to weighted complexes using threshold decomposition and stack reconstruction (see,  ). The main idea of this method is that, if the considered operator is increasing, we can apply it to each \(k\)-threshold and then combine the results to obtain the final values. More precisely, let \(E_1\) and \(E_2\) be two sets and \(\alpha\) an increasing operator from \(\mathcal{P}(E_1)\) to \(\mathcal{P}(E_2)\), the *extended stack operator of \(\alpha\)*, also denoted by \(\alpha\), is: \[\begin{aligned} \forall &M:E_1 \rightarrow \K, \forall x \in E_2,\nonumber\\ &[\alpha(M)](x) =\max\{ k \in \K \st x \in \alpha(M[k]) \} \label{eq:stack} \end{aligned}\] As erosions and dilations, the dimensional operators are increasing. Thus, they can be extended to maps. Their extended stack operators are characterized by the following property. ## Revisiting the related work {#sec:rev} In section [3](#sec:DimOps){reference-type="ref" reference="sec:DimOps"} we defined operators acting between specific dimensions of the complex. Here, we use these operators, considering in particular property [\[p:DimOpGray\]](#p:DimOpGray){reference-type="ref" reference="p:DimOpGray"}, to express operators from the literature. We start by the classical star and closure operators. Let \(X\subseteq\C\). \[\begin{aligned} St(X)= &\bigcup\left\{\dij(\dimension{X}{i}) \st i,j\in\N ,i\leq j\right\}\\ Cl(X)= &\bigcup\left\{\dji(\dimension{X}{j}) \st i,j\in\N ,i\leq j\right\} \end{aligned}\] defined operators acting on a vertex weighted graph \((V,E, f)\), where \(V\) is a finite set (of vertices), \(E\) is a set of unordered pairs of \(V\), called edges, and \(f\) is a map from \(V\) in \(\K\). By abuse of terminology, \(f\) is called a weighted graph. Let \(v \in V\), the set of neighbors of a vertex \(v\) is given by \(N_E(v)=\left\{v'\in V \st \{v,v'\} \in E\right\}\). The dilated graph \(\Gamma(f)\) and the eroded graph \(\Gamma^0(f)\) of the graph \(f\) are given, for any vertex \(v\), by: 1. \([\Gamma(f)](v) = \max\left\{f(v') \st v'\in N_E(v)\cup\left\{v\right\}\right\}\); 2. \([\Gamma^0(f)](v) = \min\left\{f(v') \st v'\in N_E(v) \cup \left\{v\right\}\right\}\). In other words, these operators replace the value of each vertex with the maximum (or minimum) value of its neighbors, as morphological operators often do. To be able to draw a parallel between these operators and the dimensional operators presented in this work, let us consider the 1-complex \(\C\) defined as the union of the vertex and edge sets of the graph \(G\): \(\C = V \cup E\). Observe then that \(C_0 = V\) and \(\C_1 = E\) and that \(f\) is a map weighting the 0-simplices of \(\C\). Using the dimensional operators, we can recover the operators from \[\begin{aligned} \Gamma(f) = & \dilm{1}{0}\dilp{0}{1}(f)\\ \Gamma^0(f) = & \erom{1}{0}\erop{0}{1}(f) \end{aligned}\] From these basic dilations and erosions derives several interesting operators, which, thanks to the previous relations, can be recovered using the operators of this article. So far the graphs were used only to provide structural information about the considered space. By considering the edges and vertices in an uniform way, allowing the propagation of the values also to the edges of the graph, both Cousty   and Meyer and Stawiaski  obtained new operators. Cousty   considered a graph \(\G=(\G^{\bullet},\G^{\times})\). For any \(X^{\times}\subseteq\G^{\times}\) and \(Y^{\bullet}\subseteq\G^{\bullet}\), the operators \(\varepsilon^{\times}\), \(\delta^{\times}\), \(\varepsilon^{\bullet}\) and \(\delta^{\bullet}\) are defined by: \[\begin{aligned} \varepsilon^{\times}(Y^{\bullet})=&\left\{e_{x,y}\in\G^{\times}\st x\in Y^{\bullet} \text{ and } y\in Y^{\bullet}\right\}\\ \delta^{\times}(Y^{\bullet})= &\left\{e_{x,y}\in\G^{\times} \st x\in Y^{\bullet} \text{ or } y \in Y^{\bullet}\right\}\\ \varepsilon^{\bullet}(X^{\times})=&\left\{x\in\G^{\bullet}\st \forall e_{x,y} \in\G^{\times}, e_{x,y}\in X^{\times}\right\}\\ \delta^{\bullet}(X^{\times})=&\left\{x\in\G^{\bullet}\st \exists e_{x,y}\in X^{\times}\right\} \end{aligned}\] If the considered space \(\C\) is the 1-complex \(\C= \G^{\bullet} \cup \G^{\times}\), using the dimensional operators, we have: \[\begin{aligned} \varepsilon^{\times}(Y^{\bullet}) = &\erop{0}{1}(Y^{\bullet})\\ \delta^{\times}(Y^{\bullet}) = &\dilp{0}{1}(Y^{\bullet})\\ \varepsilon^{\bullet}(X^{\times}) = &\erom{1}{0}(X^{\times})\\ \delta^{\bullet}(X^{\times}) = &\dilm{1}{0}(X^{\times}) \end{aligned}\] Later, in , Cousty extended these operators to weighted graphs, but the relations presented here are still true. Meyer, Angulo and Stawiaski  also defined operators capable of dealing with weighted graphs. They consider the space as a graph \(G=(N,E)\), where \(N=\left\{n_1, n_2, \dots, n_{|N|}\right\}\) is the set of vertices and \(E=\left\{e_{ij} \st i,j \in \mathbb{N}^{+}, 0<i<j\leq|N|\right\}\) is the set of edges. For two functions \(n\) and \(e\) weighting the vertices and edges of \(G\), they consider the following operators: \[\begin{aligned} \left[\varepsilon_{en}n\right]_{ij} =& n_i \wedge n_j\\ \left[\delta_{ne}e \right]_{i\ } =& \bigvee\nolimits_{k \text{ neighbors of } i} \left\{e_{ik}\right\}\\ \left[\varepsilon_{ne}e\right]_{i\ } =& \bigwedge\nolimits_{k \text{ neighbors of } i} \left\{e_{ik}\right\}\\ \left[\delta_{en}n \right]_{ij} =& n_i \vee n_j \end{aligned}\] If the considered space \(\C\) is the 1-complex \(\C= \{1, \ldots, |N|\} \cup \{\{i,j\} \st e_{ij} \in E\}\), using the dimensional operators from definition [\[def:DimOp\]](#def:DimOp){reference-type="ref" reference="def:DimOp"}, we have: \[\begin{aligned} \left[\varepsilon_{en}n\right]_{ij} =&{[\erop{0}{1}(n)](\{i,j\})}\\ \left[\delta_{en}n\right]_{ij} =&{[\dilp{0}{1}(n)](\{i,j\})}\\ \left[\varepsilon_{ne}e\right]_{i} =&{[\erom{1}{0}(e)](\{i\})}\\ \left[\delta_{ne}e\right]_{i} = &{[\dilm{1}{0}(e)](\{i\})}. \end{aligned}\] Meyer, Angulo, and Stawiaski  defined several operators based on the four presented above, all of them are recoverable by the dimensional operators. They also defined operators that rely on the particular structure of the hexagonal grid and cannot be easily expressed using our operators. ## Morphological operators on \(\subcomplexes\) using a higher intermediary dimension In this section, we define new operators, acting on subcomplexes, whose result is a complex of the same dimension of its argument, using an higher intermediary dimension, exploring the effects of property [\[p:OpInd\]](#p:OpInd){reference-type="ref" reference="p:OpInd"}. For instance, if we consider a complex \(X\) of dimension \(i\), with \(i\in\N,\,0<i\leq n\), we would like the dilation of \(X\) to be also an \(i\)-complex. To that end, the operators proposed next act independently on each dimension of the complex: As expected, the set \(\Dim{(\dilG(X))}{i}\), made of the \(i\)-simplices of \(\dilG(X)\), depends only on the set \(\Dim{X}{i}\), made of the \(i\)-simplices of \(X\). Intuitively, for \(i < n\), the set \(\Dim{(\dilG(X))}{i}\) contains all \(i\)-simplices of \(\C\) that either belong to \(\Dim{X}{i}\) or are contained in a \((i+1)\)-simplex that includes an \(i\)-simplex of \(\Dim{X}{i}\). For \(i=n\), the operator will return all \(n\)-simplices that contains an \((n-1)\)-simplex of X. Some results of the operators \(\dilG\) and \(\eroG\), along with the results of the operators \(\dilation\) and \(\erosion\) introduced by , are depicted as gray simplices in the figure [\[fig:EroDilG\]](#fig:EroDilG){reference-type="ref" reference="fig:EroDilG"}. As expected, these operators result in a subcomplex more similar to the argument. The dilation included less simplices into the set, while the erosion removed less simplices of the set. It can easily be proven that the operators \(\eroG\) and \(\dilG\) act on \(\subcomplexes\) and form an adjunction. Therefore, we can compose them to define new operators. Similarly to the operators defined in  and , the parameter \(i\) controls how much of the complex will be affected by the operator. Figure [\[fig:OpClG\]](#fig:OpClG){reference-type="ref" reference="fig:OpClG"} illustrates the operators \(\opG[i]\) and \(\clG[i]\) on two subcomplexes, depicted in gray. Since the dilation and erosion used to compose these operators affect less elements than the classical operators, we can expect the same behavior from them as well. It can be easily proven that the operators from definition [\[def:OpClG\]](#def:OpClG){reference-type="ref" reference="def:OpClG"} act on complexes, that \(\opG\) is an opening and \(\clG\) is a closing. The families of operators formed by these two operators, indexed by the integer \(i\), are a granulometry and an anti-granulometry, respectively. Therefore, they can be used to define new alternating sequential filters acting on \(\C\). In this section we explored operators acting on subcomplexes composed by dimensional operators using a higher intermediary dimension. We defined an adjunction, families of openings and closings. We composed these granulometry and anti-granulometry into two alternating sequential filters. ## Morphological operators on \(\subcomplexes\) using a lower intermediary dimension We just explored compositions of dimensional operators using a higher intermediary dimension. We will explore compositions that use a lower intermediary dimension. As theorem [\[p:FamilyDownUp\]](#p:FamilyDownUp){reference-type="ref" reference="p:FamilyDownUp"} suggests, we can define a family of different operators, using the variation of the temporary dimension as parameter. However, we chose to explore only the operators that affects the smallest possible number of simplices, because such operators usually lead to more controlled filters. Additionally, one would need a space of higher dimensionality in order to properly exploit these families. However, the following property states that the operators from definition [\[def:DilEroS\]](#def:DilEroS){reference-type="ref" reference="def:DilEroS"} are the same operators from definition [\[def:dilG\]](#def:dilG){reference-type="ref" reference="def:dilG"}. This property can be proved by analysing the elements of the space that are included or removed by each operator. Following this property, the operators obtained using a lower intermediary dimension are identical to the ones obtained using a higher intermediary dimension. For this reason, we will only illustrate the results of one of them in the next section. # Illustrations of some operators {#c:exp} We defined various operators and filters acting on subcomplexes. In this section we illustrate these operators, acting on values associated with elements of a mesh and on subcomplexes created from regular images. More results and quantitative comparisons are available in . ## Illustration on a tridimensional mesh {#sec:mesh} As illustration, we processed the curvature values associated with a \(3\)D mesh, courtesy of the French Museum Center for Research. We computed the curvature for the vertices and propagated these values to the edges and triangles, following the procedure described in , resulting in values between \(0\) and \(1\). These values were then processed using our filters. For visualization purposes only, we thresholded the values at \(0.51\), as shown in black on figure [\[fig:StThres\]](#fig:StThres){reference-type="ref" reference="fig:StThres"} that depicts the thresholded set for the original curvature data. The renderings presented in this section consider only the values associated with the vertices of the mesh, and no interpolation was used. ## Illustration on binary regular images {#sec:IlusIm} In this section we consider the application of our alternating sequential filters on regular images. For this end, we need to create a simplicial complex based on the image. Several methods can be used and the choice is application dependent. Here, we create a vertex for each pixel, with edges between the vertices, six for each vertex, corresponding to an hexagonal grid. Triangles are placed between three vertices, so each vertex belongs to six triangles. We then consider the greatest complex that can be made using the value of the vertices. For visualization purposes, the images presented depict only the values associated with the vertices. We compare our results with the literature considering the same image used by Cousty  , shown on figure [\[fig:ZebOrig\]](#fig:ZebOrig){reference-type="ref" reference="fig:ZebOrig"}. The noisy image shown on figure [\[fig:ZebNoisy\]](#fig:ZebNoisy){reference-type="ref" reference="fig:ZebNoisy"} was processed. Figure [\[fig:ZebIllG\]](#fig:ZebIllG){reference-type="ref" reference="fig:ZebIllG"} shows some results of the operators \(\asfG\) and \(\asfG[']\), along with some results from  and , for visual comparison. The operator \(\asfG\) removed most of the features of the zebra and left some noise on the background. The operator \(\asfG[']\) removed most of the background noise, while preserving some of the gaps between the stripes. However it also removed the smaller features of the object and left small holes. From these results, we may conclude that our operators are, for this type of image, on a competitive level with the operators presented in the literature. Additionally, Mennillo   used the dimensional operators for document processing as a pre-processing stage to boost OCR performance with encouraging results. While the results for regular image processing are good, they do not fully exploit the structure of the simplicial complex, nor the flexibility of the operators. We expect the filtering results to be even better when considering more complex scenarios. ## Illustration on a grayscale image We now consider a grayscale image, with a small part shown on figure [\[fig:ExGray\]](#fig:ExGray){reference-type="ref" reference="fig:ExGray"}. This image is from Jon Salisbury at the english language wikipedia, released under Creative Commons license and is a photomicrograph of bone marrow showing abnormal mononuclear megakaryocytes, typical of \(5q-\) syndrome. The image was converted from RGB to grayscale. Medical meaning aside, this image was chosen because it has many features of various sizes. Figure [\[fig:MicroCl\]](#fig:MicroCl){reference-type="ref" reference="fig:MicroCl"} shows the results of the considered closing operators, with size \(4\). As expected, all operators removed the small noise of the image. The result of the operator \(\clG[4]\) was identical to the classical operator, because we only depict the values of the vertices, while the operator \(\phi^{cm}_{12/3}\) was generally less abrasive, maintaining the level of deeper holes that were raised by the classical operator. # Conclusion {#ch:conclusion} In this work we explored the dimensional operators, presenting composition properties and defining new operators. We created new dilations, erosions, openings, closings and alternating sequential filters. We also used these operators to express operators from the related literature, acting on digital objects, such as graphs and simplicial complexes.

{'timestamp': '2014-01-23T02:06:44', 'yymm': '1401', 'arxiv_id': '1401.5602', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5602'}

# Abstract {#abstract .unnumbered} Understanding how people interact and socialize is important in many contexts from disease control to urban planning. Datasets that capture this specific aspect of human life have increased in size and availability over the last few years. We have yet to understand, however, to what extent such electronic datasets may serve as a valid proxy for real life social interactions. For an observational dataset, gathered using mobile phones, we analyze the problem of identifying transient and non-important links, as well as how to highlight important social interactions. Applying the Bluetooth signal strength parameter to distinguish between observations, we demonstrate that weak links, compared to strong links, have a lower probability of being observed at later times, while such links---on average---also have lower link-weights and probability of sharing an online friendship. Further, the role of link-strength is investigated in relation to social network properties. # Introduction {#introduction .unnumbered} Recognizing genuine social connections is a central issue within multiple disciplines. When do connections happen? Where do they take place? And with whom is an individual connected? These questions are important when working to understand and design urban areas, studying close-contact spreading of infectious diseases, or organizing teams of knowledge workers. In spite of their importance, measuring social ties in the real world can be difficult. In classical social science the standard approach is to use self-reported data. This method, however, is only practical for relatively small groups and suffers from cognitive biases, errors of perception, and ambiguities. Further, it has been shown that the ability to capture behavioral patterns via self-report data is limited in many contexts. A different approach for uncovering social behavior is to use digital records from emails and cell phone communication. Although such analyses have improved our understanding of social ties, they have left many important questions unanswered---are electronic traces a valid proxy for real social connections? Eagle et al.  began to answer this question by including a spatial component as part of their data, using the short range (\(\sim 10\,m\)) Bluetooth sensor embedded in study participants' smartphones to measure physical proximity. Their results show that proximity data closely reflects social interactions in many cases. But since it is easy to think of examples where reciprocal Bluetooth detection does not correspond to social interaction (e.g. transient co-location in dining hall) the question remains, which observations correspond to actual social interactions and which are just noise? Multiple alternatives have been proposed to Bluetooth for sensor-driven measurement of social interactions, each with particular strengths and weaknesses. For example, Radio Frequency Identification (RFID) badges have short interaction ranges (\(1-4\,m\)) and measure only face-to-face interactions, thus solving many of the resolution problems posed by Bluetooth. This approach, however, confines interactions to occur within specific areas covered by special radio receivers and requires participants to wear custom radio tags on their chests at all times---unlike Bluetooth which is ubiquitous across many types of modern electronic devices. Our investigation digs into the role of Bluetooth signal strength, using a dataset obtained from applications running on the cell phones of 134 students at a large academic institution. Each phone records and sends data to researchers about call and text logs, Bluetooth devices in nearby proximity, WiFi hotspots in proximity, cell towers, GPS location, and battery usage. In addition, we combine the data collected via the phones with online data, such as social graphs from Facebook for a majority of the participants. The study continuously gathers data, but in this paper we focus on Bluetooth proximity data gathered for 119 days during the academic year of 2012-2013. Specifically, we focus on the received signal strength parameter and propose a methodology that applies signal strength to distinguish between social and non-social interactions. We concentrate on the signal parameter because it is present in a majority of digitally recorded proximity datasets and in addition, it also suggests a rough estimate for the distance between two devices. Applying the method on our data, we compare the findings to a null model and demonstrate how removing links with low signal strength influences network structure. Moreover, we use estimated link-weights and an online dataset to validate the friendship-quality of removed links. # Materials and Methods {#materials-and-methods .unnumbered} ## Dataset {#dataset .unnumbered} We distributed phones among students from four study lines (majors), where each major was chosen based on the fraction of students interested in participating in the project. This selection method yielded a coverage of \(>93 \%\) of students per study line, enabling us to capture a dense sample of the social interactions between subjects. Such high coverage of internal connections within a social group, with respect to the density of social interactions combined with the duration of observation, has not been achieved in earlier studies. The data collector application installed on each phone follows a predefined scanning time table, which specifies the activation and duration of each probe. Proximity data is obtained by using the Bluetooth probe. Every 300 seconds each phone performs a Bluetooth scan that lasts 30 seconds. During the scan it registers all discoverable devices within its vicinity (\(5-10 m\)) along with the associated received signal strength indicator (RSSI). Recorded proximity data is of the form (\(i\), \(j\), \(t\), \(s\)), denoting that person \(i\) has observed \(j\) at time \(t\) with signal strength \(s\). Only links between experiment participants are considered, comprising a dataset of \(2\,183\,434\) time ordered edges between \(134\) nodes, see Table [1](#tab:data){reference-type="ref" reference="tab:data"} for more information. Data collection, anonymization, and storage was approved by the Danish Data Protection Agency, and complies with both local and EU regulations. Written informed consent was obtained via electronic means, where all invited participants digitally signed the form with their university credentials. Along with the mobile phone study we also collected Facebook graphs of the participants. Not all users donated their data since this was voluntary, however we obtained a user participation of \(\sim88\%\) (119 users and 1018 Facebook friendships). For the missing \(12\%\) of users, we assume they do not share any online friendships with the bulk of participants. ::: ## Identifying links {#identifying-links .unnumbered} Independent of starting conditions, the scanning framework on one phone will drift out of sync with the framework on other phones after a certain amount of time, thus the phones will inevitably scan in a desynchronized manner. This desynchronization can mainly be attributed to: internal drift in the time-protocol of each phone, depletion of the battery, and users manually turning phones off. To account for irregular scans, we divide time into windows (bins) of fixed width and aggregate the Bluetooth observations within each time-window into a weighted adjacency matrix. The complete adjacency matrix is thus given by: \(W=\left(W^{(\Delta t_1)},W^{(\Delta t_2)},\dots,W^{(\Delta t_n)}\right)\), where each link is weighted by its signal strength and where \(\Delta t_i\) indicates window number \(i\). These matrices generally assume a non-symmetric form, i.e. person \(A\) might observe \(B\) with signal strength \(s\) while person \(B\) observes \(A\) with strength \(s'\), or not at all. The scanning frequency of the application sets a natural lower limit of the network resolution to \(5\) minutes. If we are interested in the social dynamics at a different temporal resolution we can aggregate the adjacency matrices and retain entries according to some heuristic (e.g. with the strongest signal). Depending on the level of description (monthly, weekly, daily, hourly, or every 5 minutes) the researcher must think carefully about the definition of a network connection. Frameworks for finding the best temporal resolution, so called *natural timescales* have for specific problems been investigated by Clauset and Eagle, and Sulo et al.. In this paper, however, we are interested in the identification and removal of non-important proximity links, so aggregating multiple time-windows is not a concern here. Henceforth we solely work with 5 minutes time-bins. The Bluetooth probe logs all discoverable devices within a sphere with a radius of 5-10 meters---walls and floor divisions reduce the radius, but the reduction in signal depends on the construction materials. Blindly taking proximity observations as a ground truth for social interactions will introduce both false negative and false positive links in the social network. False negative links are typically induced by hardware errors beyond our control, thus we focus on identifying false positive links. We therefore propose to identify non-social or noisy proximity links via the signal strength parameter. The parameter can be thought of as a proxy for the relative distance between devices, since most people carry their phones on them, it will in principle also suggests the separation distance between individuals. Previous work has applied Bluetooth signals to estimate the position of individuals but studies by Hay, and Hossein et al.  have revealed signal strength as an unsuitable candidate for accurately estimating location. However, the complexity of the problem can greatly be reduced by focusing on the relative distance between individuals rather than position. In theory, the transmitted power between two antennae is inversely proportional to the distance squared between them. Reality is more complicated, due to noise and reflection caused by obstacles. We use the ideal result as a reference while we perform empirical measurements to determine how signal strength depends on distance. Two devices are placed on the ground in a simulated classroom setting, where we are able to control the relative distance between them. The resulting measurements are plotted in Fig. [\[fig:distance\]](#fig:distance){reference-type="ref" reference="fig:distance"}A. As is evident from the figure, there is a large variance in the measured signal strength values for each fixed distance. However, as both phones exhibit the same variance we can exclude faulty hardware; further, environmental noise such as interference from other devices, or solar radiation can also be dismissed since there appear no daily patterns in the data. But we observe multiple bands or so-called modes onto which measurements collapse, Ladd et al.  noted a similar behavior for the received signal strength of WiFi connections, both are phenomena caused by non-Gaussian distributed noise. The empirical measurements form a foundation for understanding signal variance as a function of distance, but they were performed in a controlled environment. In reality, there are a multitude of ways to carry a smartphone: some carry it around in a pocket, others in a bag. Liu and Striegel investigated how these various scenarios influence the received signal strength---their results indicate only minor variations, hence we conclude that the general behavior is similar to the measurements shown in the figure. Further, social interactions are not only limited to office environments, so we have re-produced the experiment in outdoors and in basement-like settings; the results are similar. Bi-directional observations yield at most two observations per dyad per 5-minute time-bin, we can average over the measurements (Fig [\[fig:distance\]](#fig:distance){reference-type="ref" reference="fig:distance"}B), or take the maximal value (Fig [\[fig:distance\]](#fig:distance){reference-type="ref" reference="fig:distance"}C). Fig. [\[fig:distance_dist\]](#fig:distance_dist){reference-type="ref" reference="fig:distance_dist"} shows the distributions of signal strength for each respective distance. For raw data, Fig. [\[fig:distance_dist\]](#fig:distance_dist){reference-type="ref" reference="fig:distance_dist"}A, we observe a localized zero-distance distribution while the 1, 2, and 3-m distributions overlap considerably. Averaging over values per time-bin smoothes out and compresses the distributions, but the bulk of the distributions still overlap (Fig. [\[fig:distance_dist\]](#fig:distance_dist){reference-type="ref" reference="fig:distance_dist"}B). Taking only the maximal signal value into account separates the distributions more effectively (Fig. [\[fig:distance_dist\]](#fig:distance_dist){reference-type="ref" reference="fig:distance_dist"}C). The reasoning behind choosing the maximal signal value is that phones are physically at different locations and we expect the distance to be maximally reflected in the distributions. Thus, by thresholding observations on signal strength, we can filter out proximity links that are likely to be further away than a certain distance. By doing so we are able to emphasize links that are more probable of being genuine social interactions, while minimizing noise and filtering away non-social proximity links. From the behavioral data we count the number of appearances per dyad and assign the values as weights for each link. Link weights follow a heavy-tailed distribution, with a majority of pairs only observed a few times (low weights), a social behavior that has previously been observed by Onnela et al.. Based on their weight we divide links into two categories: weak and strong. A link is defined as 'weak' if it has been observed (on average) less than once per day during the data collection period, remaining links are characterized as 'strong'. An effective threshold should maximize the number of removed weak links, while minimizing the loss of strong links. Fig. [\[fig:threshold_curve\]](#fig:threshold_curve){reference-type="ref" reference="fig:threshold_curve"} depicts the number of weak and strong links as a function of threshold value. We observe that, as we increase the threshold, the number of weak links decreases linearly, while the number of strong links remains roughly constant and then drops off suddenly. Taking into account both the maximum-value distance distributions (Fig. [\[fig:distance_dist\]](#fig:distance_dist){reference-type="ref" reference="fig:distance_dist"}C) and link weights (Fig. [\[fig:threshold_curve\]](#fig:threshold_curve){reference-type="ref" reference="fig:threshold_curve"}), we choose the value (\(-80 \,dBm\)) that optimizes the ratio between strong and weak links. In a large majority of cases, this corresponds to interactions that occur within a radius of \(0-2\) meters---a distance which Hall notes as a typical social distance for interactions among close acquaintances. ## Removing links {#removing-links .unnumbered} This section outlines various strategies for removing non-social links from the network. Fig. [\[fig:models\]](#fig:models){reference-type="ref" reference="fig:models"}A shows an illustration of the raw proximity data for a single time-bin, a link is drawn if either \(i\rightarrow j\) or \(j\rightarrow i\). Thickness of a link represents the strength of the received signal. For the thresholded network (Fig. [\[fig:models\]](#fig:models){reference-type="ref" reference="fig:models"}B) we remove links according to the strength of the signal (where we assume the weaker the signal the greater the relative distance between two persons). To estimate the effect of the threshold we compare it to a null model, where we remove the same number of links, but where the links are chosen at random, illustrated in Fig. [\[fig:models\]](#fig:models){reference-type="ref" reference="fig:models"}C. To minimize any noise the random removal might cause, we repeat the procedure \(n=100\) times, each time choosing a new set of random links, with statistics averaged over the 100 repetitions. As a reference, to check whether thresholding actually emphasizes social proximity links, we additionally compare it to a control network, where we remove the same amount of links, but where the links have signal strengths *above* or *equal* to the threshold, Fig. [\[fig:models\]](#fig:models){reference-type="ref" reference="fig:models"}D. This procedure is also repeated \(n\) times. In a situation where there are more links below the threshold than above, we will remove fewer links for the latter compared to the other networks. # Results {#results .unnumbered} ## Network properties {#network-properties .unnumbered} Now that we have determined a threshold for filtering out non-social proximity links, let us study the effects on the network properties. Thresholding weak links does not significantly influence the number of nodes present (\(N\)) in the network (Fig. [\[fig:threshold\]](#fig:threshold){reference-type="ref" reference="fig:threshold"}A), while the number of links (\(M\)) is substantially reduced (Fig. [\[fig:threshold\]](#fig:threshold){reference-type="ref" reference="fig:threshold"}B). On average we remove \(2.38\) nodes and \(32.18\) links per time-bin. Social networks differ topologically from other kinds of networks by having a larger than expected number of triangles, thus clustering is a key component in determining the effects of thresholding. Fig. [\[fig:clust\]](#fig:clust){reference-type="ref" reference="fig:clust"} suggests that we are, in fact, keeping real social interactions: random removal disentangles the network and dramatically decreases the clustering coefficient, while thresholding conserves most of the average clustering. Calculating the average ratio (\(\langle \langle c_T \rangle / \langle c_N \rangle \rangle\)) between clustering in the thresholded (\(\langle c_T \rangle\)) and the null networks (\(\langle c_N \rangle\)) reveals that \(c_T\) on average is \(2.38\) larger. These findings emphasize that a selection process based on signal strength greatly differs from a random one. ## Link evaluation {#link-evaluation .unnumbered} Sorting links by signal strength and disregarding weak ones greatly reduces the number of links, but do we remove the correct links, i.e. do we get rid of noisy, non-social links? The fact that clustering remains high in spite of removing a large fraction of links is a good sign, but we want to investigate this question more directly. To do so, we divide the problem into two timescales; a short one where we consider the probability that a removed link might reappear a few time-steps later, and a long where we evaluate the quality of a removed link according to certain network properties. Let's first consider the short time-scale. We assume that human interactions take place on a time-scale that is mostly longer than the 5-minute time-bins we analyze here. Thus, if a noisy link is removed, the probability that it will re-appear in one of the immediately following time-steps should be low, since no interaction is assumed to take place. Howbeit we expect the probability to be significantly greater than zero, since even weak (non-social) links imply physical proximity. Similarly, if we (accidentally) remove a social link, the probability that it will appear again should be high, since the social activity is expected to continue to take place. Let us formalize this notion. Consider a link \(e\) that is removed at time \(t\), the probability that the link will appear in the next time-step is \(p(t+1|e,t)\). Generalizing this we can write the probability that any removed link will appear in all the following \(n\) time-steps as: \[p(t+1,\ldots,t+n|t)=\frac{\text{no. links removed at } t \text{ present at } t+1 \cap \ldots \cap t+n}{\text{no. links removed at } t} \label{eq:reappear}\] Fig. [\[fig:link_evaluation\]](#fig:link_evaluation){reference-type="ref" reference="fig:link_evaluation"}A illustrates that thresholded links in subsequent time-steps are observed less frequently then both null and control links. To compare with the worst possible condition, we compare data from each thresholded time-bin with the *raw data* from the next bin (where the raw data contains many weak links). In spite of this, we observe a clear advantage of distinguishing between links with weak and strong signal strengths. If we look at values for \(t+1\), the first subsequent time-step, the probability of re-occurrence in the thresholded network is about \(12\%\) lower than for the null model, and as we look to later time-steps, the gap widens. A different set of social dynamics unfolds on longer timescales where the class schedule imposes certain links to appear periodically, e.g every week. Here we determine impact of removing links in two ways. First, we use total link weights and second, we use online friendship status. Friends meet frequently; we capture this behavior by using the total number of observations of a certain dyad to estimate the weight of a friendship (again, counted in the raw network). Thus, we evaluate the quality of a removed links by considering its total weight compared to the weight of other links present in the same time-bin. However, since multiple links are removed per time-bin we are more interested in the average, \[\label{eq:proximity} q_t=\frac{\text{Avg. weight of removed links at } t}{\text{Avg. weight of all links present at } t}\] This estimates, per time-bin, whether removed links on average have weights below, close to, or above the mean. Note that the measure is intended to estimate the quality of removed links and is therefore not defined for bins where zero links are removed. Fig. [\[fig:link_evaluation\]](#fig:link_evaluation){reference-type="ref" reference="fig:link_evaluation"}B indicates difference in link selection processes. Choosing links at random (null network) removes both strong and weak links with equal probability, thus on average this corresponds to the mean weight of links present. Compared to null, the thresholded network removes links with weights below average, indicating that removed links are less frequently observed and therefore also less likely to be real friendships. The control case displays an diametrical behavior, on average, it removes links with higher weights. The second method to evaluate the link-selection processes compares the set of removed links with the structure of an online social network, i.e. if a removed proximity link has an equivalent online counterpart. We estimate the quality by measuring the fraction of removed links with respect to those present at time \(t\). \[\label{eq:online} q_t^{\text{FB}}=\frac{\text{no. of FB links removed at } t }{\text{no. of FB links present at } t}\] The quality measure is essentially a ratio, i.e. it can assume values \(0 \leq q_t^{\text{FB}} \leq 1\) depending on the fraction of links that are removed. Bins with zero Facebook friendships are disregarded since they contain no information regarding the online social network. Fig. [\[fig:link_evaluation\]](#fig:link_evaluation){reference-type="ref" reference="fig:link_evaluation"}C shows that random removal (null network), on average, removes \(\sim 43 \%\) of online friendships, while the thresholded network removes \(\sim 33 \%\), a \(10\) percent point difference. For comparison, the control network removes \(\sim 44\%\) of the online links. Further, redoing the analysis for a dataset comprised only of users for which we have both proximity and online data for, does not significantly alter the results. Facebook links are not necessary good indicators for strong friendships, but are more likely to correspond to real social interactions. In spite of this, both Fig. [\[fig:link_evaluation\]](#fig:link_evaluation){reference-type="ref" reference="fig:link_evaluation"}B and C support that distinguishing between strong and weak proximity links tends to emphasize real social interactions: on average thresholded links have lower edge weights and remove fewer Facebook friendships compared to both the null-model and the control. # Discussion {#discussion .unnumbered} The availability of electronic datasets is increasing, so the question of how well can we use these electronic *clicks* to infer actual social interactions is important for effectively understanding processes such as relational dynamics, and contagion. Sorting links based on their signal strength allows us to distinguish between strong and weak ties, and we have argued that thresholding the network emphasizes social proximity links while eliminating some noise. Simply thresholding links based on signal strength is not a perfect solution. In certain settings we remove real social connections while noisy links are retained. Our results indicate that the proposed framework is better at identifying strong links than removing them. A trend which the link-reappearance probability, link-weights, and online friendship analysis support. Compared to the baseline we achieve better results than just assuming all proximity observations as real social interactions. But determining whether a close proximity link corresponds to an actual friendship interaction is much more difficult. Multiple scenarios exist where people are in close contact but are not friends, one obvious example is queuing. Each human interaction has a specific social context, so an understanding of the underlying social fabric is required to fully discern when a close proximity link is an actual social meeting. This brings us back to the question of how to determine a real friendship from digital observations (cf. ). Close proximity may not be the best indicator of friendship; call logs, text logs, and geographical positions are all factors which coupled with information from the Bluetooth probe could give us a better insight into social dynamics and interactions.

{'timestamp': '2014-05-29T02:02:56', 'yymm': '1401', 'arxiv_id': '1401.5836', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5836'}

## Semi-classical Dynamics The dynamics of the system is evolved via a semi-classical path integral approach (Path Integral Langevin dynamics) introduced in Ref.. This approach is similar to Path Integral Monte Carlo and takes advantage of the quantum to classical mapping of the partition function as Feynman path integrals. While in Monte Carlo, the paths are sampled via local update moves, the goal of Path Integral Langevin dynamics is to update the paths via molecular dynamics updates. The challenge of this approach is an efficient thermalization as low temperatures lead to large frequencies in the path integral an therefore small time-steps are required. This is solved by transforming into a normal mode picture and treat the center of mass mode differently as the high modes. In the following we repeat the derivation from. The partition function of the system is \[\langle A \rangle = \frac{1}{Z}tr[e^{-\beta H} A],\] with \[Z = \frac{1}{(2 \pi \hbar)^f} \int d^f \mathbf{p} \int d^f \mathbf{q} e^{-\beta_n H_n(\mathbf{q},\mathbf{p})}.\] Here, \(f=Nn\) and \(\beta_n = \beta/n\). The full Hamiltonian consists of system Hamiltonian and interaction Hamiltonian \[H_n(\mathbf{q},\mathbf{p}) = H_n^0(\mathbf{q},\mathbf{p}) + V_n(\mathbf{q}).\] The system part is \[H_n^0(\mathbf{q},\mathbf{p}) = \sum_{i=1}^{N}\sum_{j=1}^n \left(\frac{(p_i^{(j)})^2}{2 m_i} + \frac{1}{2}m_i \frac{1}{\beta_n^2 \hbar^2} [q_i^{(j)}-q_i^{j-1}]^2 \right)\] While sophisticated methods have been developed to sample this Hamiltonian with dynamics, an additional coupling to a thermal bath are more challenging as the frequencies in the path space separate from the frequency of the thermal bath. In Path Integral Langevin Dynamics, this is solved by propagating the Hamiltonian in the usual path integral space but the dissipative part of the thermal bath acts on the slowest mode of the normal modes of the path integral. This requires the normal mode represent of the Hamiltonian which is derived from the transformation \[\begin{aligned} \tilde{p}_k = \sum_{j=1}^n p_i^{(j)} C_{jk} \; {\text and} \;\tilde{q}_k = \sum_{j=1}^n q_i^{(j)} C_{jk} \end{aligned}\] where the matrix elements of \(c_{jk}\) are \[C_{jk}=\begin{cases} \sqrt{1/n}, & \text{if \(k=0\)}\\ \sqrt{2/n} \cos(2 \pi jk/n), & \text{if \(1 \leq k \leq n/2-1\)}\\ \sqrt{1/n}(-1)^j, & \text{if \(k = n/2\)}\\ \sqrt{2/n} \sin(2 \pi jk/n), & \text{if \(n/2+1\leq k\leq n-1\)}. \end{cases}\] In the normal mode representation the Hamiltonian reads as: \[H_n^0(\mathbf{q},\mathbf{p}) = \sum_{i=1}^{N}\sum_{k=0}^{n-1} \left(\frac{(\tilde{p}_i^{(j)})^2}{2 m_i} + \frac{1}{2}m_i \frac{\sin(k \pi/n)}{\beta_n^2 \hbar^2} (\tilde{q}_i^{k})^2 \right)\] The Liouvillian is written as system part and interaction part. The coupling to the bath acts in the space of normal modes. \[L = L_0 + L_V.\] Here, \(L_0 =-[H_n^0(\mathbf{q},\mathbf{p}).]\) and \(L_V =-[V_n^0(\mathbf{q},\mathbf{p}).]\). \[e^{\Delta t L} = e^{(\Delta t/2)L_V}e^{\Delta t L_0}e^{(\Delta t/2)L_V}.\] While \(L_0\) and \(L_V\) are associated with momentum and position dynamics in the path integral the dissipative part is added in the equations of motion in the space of normal modes. This leads to the following equations of motion \[e^{\Delta t L} = e^{(\Delta t/2)L_\gamma}e^{(\Delta t/2)L_V}e^{\Delta t L_0}e^{(\Delta t/2)L_V}e^{(\Delta t/2)L_\gamma}\] While \(L_0\) and \(L_V\) are associated with momentum and position dynamics in the path integral, \(L_\gamma\) is defined the space of normal mode variable. This requires four transformation between normal mode and path integral space per timestep. The step associated with \(L_V\) is \[p_i^{(j)}(t + \Delta t) = p_i^{(j)}(t)-\Delta t \frac{\partial V}{\partial q_i^{(j)}}\] The evolution of the path integral with \(L_0\) defined in normal mode space corresponds to \[\begin{aligned} \tilde{p}_i^{(k)}(t + \Delta t) = \tilde{p}_i^{(j)}(t) \cos(\omega_k \Delta t)-\tilde{q}_i^{(j)}(t) m_i \omega_k \sin(\omega_k \Delta t). \\ \nonumber \tilde{q}_i^{(k)}(t + \Delta t) = \tilde{p}_i^{(j)}(t)\frac{1}{m_i \omega_k} \sin(\omega_k \Delta t)-\tilde{q}_i^{(j)}(t) \cos(\omega_k \Delta t). \end{aligned}\] The last step, \(L_\gamma\) describes thermalization in the normal mode space and leads to the step \[\tilde{p}_i^{(k)}(t + \Delta t) = \tilde{p}_i^{(k)}(t) e^{-\Delta t/2 \gamma^{(k)}} + \sqrt{\frac{m_i}{\beta_n} (1-e^{-\Delta t \gamma^{(k)}})}\] ## Spectroscopy The transition from the quasi-long range crystal phase to the hexatic phase requires the knowledge of the tails of the structure factor or, as described in the main text, the interaction between dislocations. The transition to the liquid phase, however, can be resolved with standard spectroscopy tools like Bragg scattering \[see e.g. \] from the static structure factor defined as \[S(\mathbf{k}) = \frac{1}{N} \left \langle \sum_{jk} \mathrm{e}^{-i \mathbf{k} (\mathbf{R}_j-\mathbf{R}_k)} \right \rangle.\] The quasi-long-range nature of the correlation is clearly resolved from \(S(\bf{k})\) with a sharp peak at the first Bravis vector \[Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"}(a)\]. In the liquid phase no significant structures are left and \(S(\bf{k})\) is structure-less \[Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"}(b)\].

{'timestamp': '2014-01-23T02:09:18', 'yymm': '1401', 'arxiv_id': '1401.5682', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5682'}

# Introduction Consider two directed graphical models (or directed acyclic graphs, DAGs) on random variables \((A,B,C)\): \[\label{eq:thr} A \to B \to C \qquad B \leftarrow A \to C\] (See  for background on graphical models.) In this paper, we will say that these two models are *isomorphic* (as graphical models). Roughly, this means that after relabeling (\(A\leftrightarrow B\)), the two resulting models describe the same collection of joint distributions \(P_{A,B,C}\). Note that the so defined isomorphism notion is weaker than the (directed) graph isomorphism: the graphs in [\[eq:thr\]](#eq:thr){reference-type="eqref" reference="eq:thr"} are not isomorphic. On the other hand, there does not exist any relabeling making [\[eq:thr\]](#eq:thr){reference-type="eqref" reference="eq:thr"} equivalent to \[\label{eq:thr2} B \to C \leftarrow A\] In fact, a simple exercise in \(d\)-separation criterion shows that [\[eq:thr\]](#eq:thr){reference-type="eqref" reference="eq:thr"} and [\[eq:thr2\]](#eq:thr2){reference-type="eqref" reference="eq:thr2"} list all possible isomorphism classes of directed *tree* models on three variables. However, note that the above DAGs are all isomorphic as undirected graphs. The goal of this paper is to provide (computational) answer to: *What are the isomorphism classes of directed graphical models on \(n\) nodes?* Note that when variables \((A,B,C)\) are jointly Gaussian and zero-mean, then conditions such as [\[eq:thr\]](#eq:thr){reference-type="eqref" reference="eq:thr"} can be stated as algebraic constraints on the covariance matrix: \[\label{eq:thr3} \mathbb{E}[AB] \mathbb{E}[BC] = \mathbb{E}[AC] \mathbb{E}[B^2].\] This suggests that checking isomorphism of models can be carried out via algebraic methods. Indeed, one needs to recall (see ) that graphical models equality can be tested by restricting to Gaussian random variables. In this paper, we associate with every DAG two subsets of covariance matrices: - all non-singular covariance matrices satisfying DAG constraints (denoted \(\mathrm{loc}(G)\cap\Sigma^{++}\) below) - all covariance matrices satisfying DAG constraints (denoted \(\mathrm{loc}(G)\) below) We give an *analytic* result: while \(\mathrm{loc}(G)\) is not necessarily (Euclidean) closed, closures of both sets coincide. Next, we switch to the *algebraic* part. Due to the analytic fact above, much simpler equations for non-singular matrices can be used to completely characterize the Zariski closure of \(\mathrm{loc}(G)\) (denoted \(X_G\) below). Aesthetically pleasing is the fact that \(X_G\) is always an irreducible complex variety (affine and rational). Furthermore, two graphical models \(G\) and \(G'\) define the same set of conditional independence constraints if and only if \(X_G = X_{G'}\). For large graphs it is important to reduce the number of equations needed to describe \(X_G\). The natural set of equations (denoted \(I_G\) below) turns out to be too small: its solution set \(V(I_G)\) contains \(X_G\) and a number of spurious components. We show how to get rid of these spurious components, proving that \[X_G = V((I_G:\theta_0^m))\,,\] where \(\theta_0\) is an explicit polynomial (and establishing Conjecture 3.3 of Sullivant ). This provides a convenient method for computing \(X_G\). After these preparations, we give our main result: isomorphism question \(G\stackrel{?}{\sim} G'\) is equivalent to comparing intersections of \(X_G\) and \(X_{G'}\) with a certain invariant variety. We give a randomized algorithm for this and apply it to provide a list of isomorphism classes on 4,5, and 6 nodes. The question of DAG isomorphism does not seem to have appeared elsewhere, though the closely related question of DAG equivalence (or Markov equivalence ) is well-studied. As mentioned in, the natural space to work with when doing model selection or averaging over DAGs is that of their equivalence classes. In practice, the number of DAGs in an equivalence class encountered during model selection can be large. Some examples are presented in Table 4 of where, for instance, the learning algorithm discovers equivalence classes on 402 nodes with more than \(7\times 10^{21}\) members. However, one should keep in mind that not all equivalence classes are as large. In fact, Steinsky showed, by a recursive method, that roughly 7 percent of equivalence classes of graphs with 500 nodes or less consist only of one element. Recent results indicate that this ratio is valid asymptotically as well. Such classes appear to be the most common type of equivalences classes of DAGs. In general, it is expected, based on observations on small graphs, that the ratio of equivalence classes to DAGs be around 0.27. Nevertheless, some equivalence classes (such as those that appear in ) are quite large. This has motivated the need to represent DAGs, and among the representatives that are relevant in this regard are the essential graphs[^1] and the characteristic imsets. Both these methods have a combinatorial flavor and this work provides an algebraic alternative. The word algebraic here means commutative-algebraic, unlike in . We remark that the two mentioned methods can also be applied to solve the isomorphism problem. For instance, using the results in one can reduce DAG isomorphism to the isomorphism of certain directed multi-graphs. In fact, this gives a sense of the inherent computational difficulty involved in working with the isomorphism class of a DAG. While the notion of Markov equivalence makes sense in the setting of, there are situations where it is natural to want to work with the isomorphism class of a DAG--the Markov equivalence class modulo permutations of variables. For instance, the recent results in imply that there is a precise sense in which the isomorphism classes of all large graphs that admit efficient inference are related to graphs that look like the (unlabeled) trees listed in Figure [\[fig_classes\]](#fig_classes){reference-type="ref" reference="fig_classes"}, but are far from the complete DAGs. An exact description of such models, however, is problematic by the subsequent results in. It thus appears reasonable to find good ways to approximate them, and for that we resort to the family of projective varieties. It is also important to mention that the idea of associating an algebraic variety to a conditional independence (CI) model has been previously explored in a number of publications, among which we will discuss . Some of our preparatory propositions can be found in the literature in slightly weaker forms and we attempt to give references. The main novelties are: - We essentially leverage the directed-graph structure of the model (as opposed to general CI models) to infer stronger algebraic claims. In particular, our treatment is base independent--although for readability we present results for the varieties over \(\mathbb{C}\). - We present a computational procedure for answering the isomorphism question. ## Preliminaries Directed acyclic graphical models are constraints imposed on a set of probability distributions: Given a collection of such models, it is often of interest to find representatives for their isomorphism classes (see also )--these are models that have the same compatible distributions modulo labelings of variables: We shall mainly focus on characterizing isomorphism classes of DAGs. A related question is that of understanding the structure of conditional independence constraints--see for the case of discrete random variables, positive discrete random variables, non-singular Gaussians, and general Gaussians. Let \(H=H_1\times\cdots\times H_n\) be a product measure space endowed with the \(\sigma\)-algebra \(\mathcal{H}=\mathcal{H}_1\otimes\cdots\otimes\mathcal{H}_n\). We assume that \(H_i\) is measurably isomorphic to \(\mathbb{R}\) and that \(\mathcal{H}_i\) is a Borel \(\sigma\)-algebra for all \(i\). The next property, factorization, relies on a digraph structure and pertains only to DAGs: Given a DAG \(G\), we denote by \(\mathrm{Fac}(G)\) the set of distributions that factorize w.r.t \(G\). It is known (c.f. Section II.A) that \[\mathrm{Fac}(G)=\mathrm{Loc}(G).\] This means that two DAGs are equal (isomorphic) in the above sense if and only if they factorize the same set of distributions (modulo the labeling of the variables). ## Notation - \(\mathcal{N}\) is the set of real valued Gaussians - \(\mathcal{N}^{+}\) is the non-singular subset of \(\mathcal{N}\). - \(\Sigma=[\sigma_{ij}]\) is the affine space \(\mathbb{C}^{{n+1 \choose 2}}\) of Hermitian \(n\times n\) matrices. - \(\Sigma^+\) is the positive semi-definite (PSD) subset of \(\Sigma\). - \(\Sigma^{++}\) is the positive definite (PD) subset of \(\Sigma\). - \(\Sigma^{\mdot{}\,}\) is the subset of matrices in \(\Sigma\) with non-zero principal minors[^3]. - \(\hat{\Sigma}\) is the subset of \(\Sigma\) consisting of matrices with ones along the diagonal. We also set \(\hat{\Sigma}^+\stackrel{\vartriangle}{=}\hat{\Sigma}\cap \Sigma^+\), \(\hat{\Sigma}^{++}\stackrel{\vartriangle}{=}\hat{\Sigma}\cap \Sigma^{++}\), and \(\hat{\Sigma}^{\mdot{}\,}\stackrel{\vartriangle}{=}\hat{\Sigma}\cap \Sigma^{\mdot{}\,}\). - \(\mathrm{loc}(G)\) is the set of covariance matrices in \(\mathrm{Loc}(G)\cap \mathcal{N}\). - \(f_G\) is the rational parametrization defined in II.B. - Given \(S\subset \Sigma\), \([S]_{}\) and \([S]_Z\) are its standard and Zariski closures[^4], respectively. - Given \(S\subset \Sigma\), \(I(S)\) is the ideal of polynomials that vanish on \([S]_Z\). - Given an ideal \(I\), the associated algebraic set is given by \[{V}(I)=\{ x\in \mathbb{C}^n| f(x)=0 \quad \forall f\in I\}.\] - \(X_G\stackrel{\vartriangle}{=} [\mathrm{loc}(G)]_Z\), \(\mpf_G\stackrel{\vartriangle}{=} I(X_G)\), \(\hat{X}_G\stackrel{\vartriangle}{=} [\mathrm{loc}(G)\cap \hat{\Sigma}]_Z\). - \(Y_G\) is the closure of \(\hat{X}_G\) inside \(\mathbb{P}^{n \choose 2}\). ## Overview of main results Our main purpose is to show that the computational tools in algebra are relevant for addressing the following problem: Our starting point is to show that isomorphism and \(\mathcal{N}^{+}\)-isomorphism are equivalent for DAGs (see Section II.C). It is well known that checking \(\mathcal{N}^{+}\)-equivalence reduces to checking equality of algebraic subsets inside the positive definite cone (see for instance ). This follows from the next proposition: Proposition [\[pro_minors\]](#pro_minors){reference-type="ref" reference="pro_minors"} enables us to think algebraically and/or geometrically when deciding Gaussian equivalence. Indeed, it states that \(\mathrm{loc}(G)\cap \Sigma^{++}\) can be identified with the positive definite subset of the real solutions to the polynomial equations generated by the implied relations in \(G\). Working with such subsets, however, is not convenient from a computational point of view. This motivates the next problem: Let \({J}_{G}\) be the ideal generated by the minors \(|\sigma_{iK,jK}|\) of the implied relations \(i\perp\!\!\!\perp j|K\in \mathcal{C}_G\) inside \(\mathbb{C}[\Sigma]\). Similarly, the minors of imposed relations of \(G\) generate an ideal \(I_G\subset J_G\) in \(\mathbb{C}[\Sigma]\). Note that this ideal coincides with that generated by the toposorted imposed relations. The corresponding ideals generated inside \(\mathbb{C}[\hat{\Sigma}]\) are denoted by \(\hat{I}_G,\hat{J}_G\). With the established notation, for example, Proposition [\[pro_minors\]](#pro_minors){reference-type="ref" reference="pro_minors"} implies \[\label{eq:vcmloc} {V}(I_G) \cap \Sigma^{++}\cap \mathbb{R}^{n+1\choose 2} = \mathrm{loc}(G) \cap \Sigma^{++}.\] We address the above problem by identifying \(X_G\) with an irreducible component of \({V}(I_G)\). It is a curious fact that the points in \({V}(I_G)\cap \hat{\Sigma}^{++}\) correspond to covariances of circularly symmetric Gaussians that satisfy the CI constraints of \(G/\mathbb{C}\). Thus if we work with complex Gaussians, we may avoid intersecting with the reals in ([\[eq:vcmloc\]](#eq:vcmloc){reference-type="ref" reference="eq:vcmloc"}). In Section II, we first prove some geometric results, which can be summarized in the following diagram The same inclusions hold if we replace \((I_G,\Sigma)\) with \((\hat{I}_G,\hat{\Sigma})\), \(\Sigma^{\mdot}\,\) with \(\Sigma^{++}\), or \(I_G\) with \(J_G\). It is known that \([\mathrm{loc}(G)\cap \Sigma^{++}]_Z\) is a complex irreducible rational algebraic variety, cf. . Here we further show that it coincides with \(X_G\) and characterize \(\mathfrak{p}_G=I(X_G)\) in two different ways: as the saturated ideal of \(I_G\) at \(\theta_0=\prod_{A\subset [n]}(|{\Sigma}_{AA}|)\) (Conjecture 3.3 in ), and as the unique minimal prime of \(I_G\) contained in the maximal ideal \(\mathfrak{m}_I\) at the identity. We thus have the following relations inside \(\mathbb{C}[{\Sigma}]\): \[\begin{aligned} &I_G\subset J_G\subset S^{-1}J_G\cap {\mathbb{C}}[\Sigma]= S^{-1}I_G\cap \mathbb{C}[\Sigma]\\ &=I(\mathrm{loc}(G)\cap\Sigma^{++})=\mathfrak{p}_G\subset \mathfrak{m}_I, \end{aligned}\] where \(S=\{\theta_0^n|n>0\}\). One can replace \((\mathrm{loc}(G),{\Sigma})\) with \((\mathrm{loc}(G)\cap \hat{\Sigma},\hat{\Sigma})\) in the above. We note that the above relations hold verbatim over \(\mathbb{Z}[\Sigma]\) and other base rings. Our main statement, shown in [2.5](#sec:algrep){reference-type="ref" reference="sec:algrep"}, is that two DAGs \(G,G'\) are isomorphic if and only if \[S^{-1}I_G\cap \mathbb{C}[\Sigma]^\Pi=S^{-1}I_{G'}\cap \mathbb{C}[\Sigma]^\Pi.\] We use the above results to provide a randomized algorithm for testing DAG isomorphism in [2.7](#sec:rand){reference-type="ref" reference="sec:rand"}. Section III concerns the special case of directed tree models. In particular, we show that \(\hat{I}_T\) is a prime ideal for a tree model \(T\) and hence \(\hat{I}_T=I(\mathrm{loc}(T)\cap \hat{\Sigma})\). This is analogous to primality of \(J_T\), the ideal of implied relations, shown in (see Corollary 2.4 and Theorem 5.8). We thus have that two directed tree models \(T\) and \(T'\) are equal if and only if \(\hat{I}_T=\hat{I}_{T'}\). Moreover, we use our randomized algorithm to list the isomorphism classes of directed tree models for \(n=4,5,\) and \(6\) nodes. We then use the algorithm to list the isomorphism classes of directed tree models for \(n=4\) and \(5\) nodes. We also include the list for \(n=6\) in the extended version of the paper. There, we further discuss some special properties of directed tree models. In particular, we show that \(\hat{I}_T\) is a prime ideal for a tree model \(T\) and hence \(\hat{I}_T=I(\mathrm{loc}(T)\cap \hat{\Sigma})\). This is analogous to primality of \(J_T\), the ideal of implied relations, shown in (see Corollary 2.4 and Theorem 5.8). The number of isomorphism classes of directed tree models found by our procedure is \(1,1,2,5,14,42,142...\) for \(n\ge 1\). Curiously, the first 6 numbers are Catalan but the 7th is not. # Main results ## Factorization and local Markov properties In this section we show that isomorphic DAGs factorize the same set of probability distributions modulo the labeling of the variables. A theorem of Lauritzon (see ) says that a non-singular measure satisfies the local Markov property if and only if its density factorizes. Let \((H,\mathcal{H})\) be as in Definition [\[def_fac\]](#def_fac){reference-type="ref" reference="def_fac"}. One can further state: ## Weak limits of factorable Gaussians This section provides a characterization of the singular distributions in \(\mathrm{loc}(G)\) as the weak limit of sequences in \(\mathrm{loc}(G)\cap \Sigma^{++}\). Note that since \((H,\mathcal{H})\) is a topological space, weak convergence \(P_{X_n}\stackrel{w}\rightarrow P_X\) is well-defined. We note that in the case of Gaussians \(X_n\sim N(0,\sigma_n), X\sim N(0,\sigma)\), \(P_{X_n}\stackrel{w}{\rightarrow} P_X\) is equivalent to \(\sigma_n\rightarrow \sigma\) in the standard metric.To characterize \(\mathrm{loc}(G)\cap \partial \mathcal{N}^+\), we shall find it useful to work with the parametrization \[X_i=\sum_{j<i} \alpha_{ij} X_j+\omega_iZ_i, \label{eq_rational}\] where \(Z_i\)'s are independent standard Gaussians. Suppose that \(\alpha_{ij}= 0\) for all \((i,j)\notin E\), where \(E\) denotes the set of (directed) edges of \(G\). Then this parametrization gives a polynomial map \(f_G:\mathbb{R}^{|E|+n}\mapsto \mathbb{R}^{n+1 \choose 2}\), sending \(\{\alpha_{ij},\omega_i\}\) to \(\mathrm{cov}(X)\). Indeed, starting from ([\[eq_rational\]](#eq_rational){reference-type="ref" reference="eq_rational"}), one can write \[\sigma_{ik}=\sum_{j<i}\alpha_{ij} \sigma_{jk}+\omega_i\gamma_{ik}\] where \(\gamma_{ik}=\mathrm{Cov}(Z_i,X_k)\), \(\sigma_{ik}=\mathrm{Cov}(X_i,X_k)\). Note that \(\gamma_{ik}=0\) for \(k<i\). With this notation, we can write \[\begin{aligned} \gamma_{ki}&=\sum_{j>i}\alpha_{ij}\gamma_{kj}+\omega_i\delta_{ik}=\sum_{j>i}\gamma_{kj}\alpha^*_{ji}+\omega_i\delta_{ik}. \end{aligned}\] Set \(\Gamma:=[\gamma_{ij}],A:=[\alpha_{ij}],\Omega:=[\omega_{ii}],\Sigma:=[\sigma_{ij}]\). We can write the above equations in matrix form: \[\Sigma=A\Sigma+\Omega\Gamma, \quad \Gamma=\Gamma A^*+\Omega.\] Hence, \[\Sigma=(I-A)^{-1}\Omega^2(I-A^*)^{-1}.\] The image of \(f_G\) is Zariski dense in \([\mathrm{loc}(G)\cap \Sigma^{++}]_Z\): The next Proposition shows that \[X_G=[\mathrm{loc}(G)\cap \Sigma^{++}]_Z.\] This proposition shows that \(X_G\) contains all \(G\)-factorable Gaussians. There are, however, (singular) covariances on \(X_G\) that are not \(G\)-compatible. In other words, unlike independence, conditional independence is not preserved under weak limits as shown in the following example. ## DAG isomorphism The next result states that isomorphism of DAGs can be decided inside \(\mathcal{N}^+\)[^7]: This property is also known as the faithfulness of Gaussians in the statistics literature (cf.  ). Let us point out that, in general, \(\mathcal{N}^{+}\)-isomorphic models are not \(\mathcal{N}\)-isomorphic as shown in the next example. ## DAG varieties and ideals Here we provide some algebraic and geometric descriptions for \(\mathrm{loc}(G)\): Chasing \(\theta_0\) from \(\mathbb{C}[\Sigma]\) to \(\mathbb{C}\) in two different ways gives \(\theta_0\notin\ker \varphi^*\), and thus \(\varphi^*(\theta_0)=h/g^m\) for \(h\neq 0\). Localizing \(\mathbb{C}[\Sigma_{\mathrm{edge}}][1/g]\) at \(h\) gives a diagram Note that \(\varphi^*_{gh}\) sends \(\theta_0\) to a unit. Hence, by the universal property of localization (see ), it extends to a map \[\begin{aligned} \varphi_{gh}^{*,e}:S^{-1}\mathbb{C}[\Sigma] \to \mathbb{C}[\Sigma_{\mathrm{edge}}][1/gh]\label{eq:morp}\\ \sigma_{ij}\mapsto \varphi^*(\sigma_{ij}),1/\theta_0\rightarrow g^m/h \end{aligned}\] that is onto and has \(S^{-1}I_G\) as kernel. To verify the latter claim, take \(s\) in the kernel of ([\[eq:morp\]](#eq:morp){reference-type="ref" reference="eq:morp"}) and write it as \(s=p_\mathrm{edge}+\sum_{ij}q_{ij}.(\sigma_{ij}-\tilde h_{ij}/\tilde u_{ij})\) where \(q_{ij}\in S^{-1}\mathbb{C}[\Sigma]\) with \((i,j)\mbox{--non-edge}\), and \(p_\mathrm{edge}\) is a polynomial in \(\sigma_{\mathrm{edge}}\). This can be done by virtue of the binomial theorem: \[\sigma_{ij}^n=(\frac{\tilde{h}_{ij}}{\tilde{u}_{ij}}-\frac{\tilde{h}_{ij}}{\tilde{u}_{ij}}+\sigma_{ij})^n=(\frac{\tilde{h}_{ij}}{\tilde{u}_{ij}})^n+n(\frac{\tilde{h}_{ij}}{\tilde{u}_{ij}})^{n-1}(\sigma_{ij}-\frac{\tilde{h}_{ij}}{\tilde{u}_{ij}})+\cdots\] Then \(\varphi^*(s)=p_\mathrm{edge}\), and thus \(\varphi^*(s)=0\) gives \(p_\mathrm{edge}=0\), that is \(s\in S^{-1}I_G\). The reverse inclusion is obvious. This establishes isomorphism of rings \[S^{-1} \mathbb{C}[\Sigma]/S^{-1}I_G = \mathbb{C}[\Sigma_{\mathrm{edge}}][1/gh],\] which implies that \(S^{-1} I_G\) is prime, and that each local ring of \(X_G \cap \Sigma^{\mdot}\) is regular (since all local rings of \(\mathbb{C}[\Sigma_{\mathrm{edge}}][1/gh]\) are regular). Geometrically our proof corresponds to constructing a birational isomorphism: where \(\mathcal{U}=D(gh)\) is a distinguished open, \(\varphi|_\mathcal{U}\) is obtained by restriction of the map given in [\[eq:ltt\]](#eq:ltt){reference-type="eqref" reference="eq:ltt"}(\(\varphi\) is regular on \(\mathcal{U}\)), and \(\pi\) is the projection from \(\Sigma\) to \(\Sigma_{\mathrm{edge}}\). see  for details. ◻ The next example shows how Theorem [\[cor_conj\]](#cor_conj){reference-type="ref" reference="cor_conj"} can be used to construct \(\mathfrak{p}_G\) from \(I_G\): The ideal of imposed relations is generated by relations \(1\perp\!\!\!\perp 3|2\) and \(4\perp\!\!\!\perp 1|(2,3)\): \[I_G=\langle |\sigma_{12,23}|,|\sigma_{123,423}|\rangle.\] It has primary components \[\begin{aligned} I_{G,1}&=\langle \sigma_{12}\sigma_{23}-\sigma_{13}\sigma_{22}, \sigma_{12}\sigma_{24}-\sigma_{14}\sigma_{22}, \sigma_{13}\sigma_{24}-\sigma_{14}\sigma_{23}\rangle \end{aligned}\] and \[I_{G,2}=\langle \sigma_{12}\sigma_{33}-\sigma_{13}\sigma_{23}, \sigma_{12}\sigma_{33}-\sigma_{13}\sigma_{23}, \sigma_{22}\sigma_{33}-\sigma_{23}^2\rangle.\] It can be seen that only \(I_{G,1}\) intersects \(\Sigma^{\mdot}\)Ẇe thus have \(\mathfrak{p}_G=I_{G,1}\). Furthermore, \(I_{G,1}\) is the unique ideal contained in the maximal ideal at the identity of \(\Sigma\), and is also equal to the saturation of \(I_{G}\) at \(f=\sigma_{22}(\sigma_{22}\sigma_{33}-\sigma_{23}^2)\). The ideal of implied relations is generated by relations \(1\perp\!\!\!\perp 3|2,1\perp\!\!\!\perp 4|2,1\perp\!\!\!\perp 4|(2,3)\), and \(1\perp\!\!\!\perp 3|(2,4)\): \[J_G=\langle |\sigma_{12,23}|,|\sigma_{12,42}|,|\sigma_{123,423}|,|\sigma_{124,324}|\rangle.\] It has primary components \[J_{G,1}=\langle \sigma_{13}\sigma_{22}-\sigma_{12}\sigma_{23},\sigma_{14}\sigma_{22}-\sigma_{12}\sigma_{24},\sigma_{14}\sigma_{23}-\sigma_{13}\sigma_{24} \rangle\] and \[J_{G,2}=\langle \sigma_{12}, \sigma_{22},\sigma_{24},\sigma_{23}^2\rangle.\] Again one can check that \(J_{G,1}=I_{G,1}=S^{-1}J_G\cap \mathbb{C}[\Sigma]\) is the unique component that is contained in the maximal ideal at the identity. Finally, we can see that \[S^{-1}\hat{I}_G\cap \mathbb{C}[\hat{\Sigma}]=\langle\hat{\sigma}_{12}\hat{\sigma}_{23}-\hat{\sigma}_{13}, \hat{\sigma}_{12}\hat{\sigma}_{24}-\hat{\sigma}_{14},\hat{\sigma}_{13}\hat{\sigma}_{24}-\hat{\sigma}_{14}\hat{\sigma}_{23}\rangle\] is the unique irreducible component of \(V(\hat{I}_G)\) that contains the origin of \(\hat{\Sigma}\).\ [\[ex:thm1\]]{#ex:thm1 label="ex:thm1"} ## Algebraic representation {#sec:algrep} Here, we put together the results of the previous sections to give an algebraic criteria for testing isomorphism of graphical models. We start by a result on equivalence of DAGs: In what follows, \(\Pi=\{\pi_s\}_{s\in S_n}\) is the permutation group with induced action on \(\mathbb{C}[{\Sigma}]\): \(\pi_s(f(\sigma_{ij}))=f(\sigma_{s(i)s(j))})\) where \(s\in S_n\) is a permutation of indices. The invariant subring \(\{f\in {\mathbb{C}}[{\Sigma}]\,|\, f\circ \pi_s=f\,\, \forall s\}\) is denoted by \({\mathbb{C}}[{\Sigma}]^\Pi\). We can now state our main result: As before, we can work over \(\mathbb{C}[\hat{\Sigma}]\). We remark that \(\hat{S}^{-1}\hat{I}_G\) can be replaced with \(\hat{J}_G\) in the above. The presentation of the results in this form is a matter of convenience. Indeed, there is a simple way to generate \(\hat{I}_G\): - Traverse the graph in the order of the topological sort and set \(\hat{I}_i:= \sum_{{j<i}, j\notin K}\langle |\hat{\sigma}_{iK,jK}|\rangle\) with \(K:=\mathbf{pa}(i)\) for all \(i\). - Output \(\hat{I}_{G}=\sum_i \hat{I}_i\). However, extracting the implied relations of a DAG requires more work. We also point out that the extra components of \(V(\hat{I}_G)\) (or \(V(I_G)\)) are not invariant across the isomorphism class of \(G\): is isomorphic to the DAG \(G\) in Example [\[ex:thm1\]](#ex:thm1){reference-type="ref" reference="ex:thm1"}, but \(\hat{I}_{G'}\) is a prime ideal. Note however that \(J_{G'}=J_{\pi(G)}\) where \(\pi\) is the permutation \((14)(23)\). Thus it is necessary to compute the saturation ideal in ([\[eq:tpr\]](#eq:tpr){reference-type="ref" reference="eq:tpr"}). We shall see, however, in section [2.7](#sec:rand){reference-type="ref" reference="sec:rand"} that one can avoid computing the saturation ideal, and more importantly, the subring intersection by a probabilistic procedure. [\[ex:inv\]]{#ex:inv label="ex:inv"} If \(G\) and \(G'\) are two DAGs on \(n\) nodes and \(E\) edges, it is easy to verify that the algorithm can decide equivalence of \(G\) and \(G'\) in time \(O(n^3|E^c|)\). To test if a sampled point lies on \(\hat{X}_G\), one needs to solve a sequence of \(|E^c|\) linear equations. Each equation involves one missing edge and one minor of size at most \((n-1)\times (n-1)\), which requires \(O(n^3)\) operations to compute (note that we can do the arithmetics over rationals). This is comparable, in terms of complexity, with \(O(n^4|E|)\) operations needed in the essential graph method (see ). # Directed tree models We study the case of directed tree models in this section. The main property is the following: This is analogous to primality of \(J_T\) shown in (see Corollary 2.4 and Theorem 5.8). A direct consequence is that two tree models are isomorphic if and only if \(\hat{I}_T\cap \mathbb{C}[\hat{\Sigma}]^{\Pi}=\hat{I}_{T'}\cap \mathbb{C}[\hat{\Sigma}]^\Pi.\) It also follows from the proposition that \(\hat{I}_T=\hat{J}_T\). In II.E, we give a procedure to generate \(\hat{I}_T\). To prove primality of \(\hat{I}_T\), we introduce a second procedure that uses lower degree generators by modifying the first step: - Set \({L}_0=\emptyset\). - Traverse the graph in order of the sort and set \(K_i:=\mathbf{pa}(i)\) for all \(i\). - For all \(j<i\), let \(K'\in K_i\) be the smallest subset that \(d\)-separates \(j\) and \(i\). Set \(L_i=L_{i-1}\cup \{f| f= \langle |\sigma_{iK',jK'}|\rangle\}\). - Output \(\hat{I}'_G:=\sum_{f\in {L}_n} \langle f \rangle\). Let us verify \(\hat{I}_T=\hat{I}'_T\) for a tree model \(T\). Suppose the claim holds on \(n-1\) nodes and let the first procedure reach node \(n\). Set \(K=\mathrm{pa}(i)\) and let \(k\) be the separator of nodes \(i\) and \(n\) for some \(i<n\). The generated polynomial is \[|\sigma_{ iK,nK}|=\left|\begin{array}{cc} \sigma_{in}& \sigma_{iK} \\ \sigma_{Kn} & \sigma_{KK} \end{array} \right|.\] This can be expanded as \[\begin{aligned} |\sigma_{ iK,nK}|&=\left|\begin{array}{cc} \sigma_{in}& \sigma_{ik} \\ \sigma_{kn} & 1 \end{array} \right|+\sum_{j\in \{K-k\}} \sigma_{ij}f_j+\sum_{j,j'\in K} \sigma_{jj'}f_{jj'}, \end{aligned}\] which is in the ideal of the second procedure since \(i\) is not connected to \(K-k\) and the parents of \(n\) are not connected either. Conversely, take the minor \[\left|\begin{array}{cc} \sigma_{in}& \sigma_{ik} \\ \sigma_{kn} & 1 \end{array} \right|\] produced by the second procedure and note that \(i\perp\!\!\!\perp K-k\) and \(k'\perp\!\!\!\perp k''\) for distinct \(k',k''\in K\). By the inductive hypothesis, the corresponding linear forms are in \(I_T\). Using the above expansion for \(|\sigma_{ iK,nK}|\) one can see that this reduced minor is in the ideal of the first procedure as well. Thus the two ideals are equal when \(T\) is a tree. The first procedure is easy to implement and is what we use to generate the ideals. The second procedure is easy to analyze and has useful properties that we exploit later. For instance, it shows that the ideal of imposed relations of a tree model is generated by quadratic polynomials of type \(\sigma_{ij}-\sigma_{ik}\sigma_{kj}\) or linear forms \(\sigma_{ij}\). We note that the equality of \(\hat{I}_T\) and \(\hat{I}'_T\) does not generalize to all DAGs. For instance, in the case of Example [\[ex:thm1\]](#ex:thm1){reference-type="ref" reference="ex:thm1"}, \(\hat{I}_T\) is not prime whereas \(\hat{I}'_T\) is prime for all DAGs with \(n\le 4\) nodes. We are now ready to prove that the ideal of the imposed relations is prime for tree models:\ One can further show that for a certain lexicographic order the generators of \(\hat{I}_G\) form a Gr\(\ddot{\textup{o}}\)bner  basis. Set \(d_{ij}=\min_{\pi \in \mathcal{D}(i,j|\emptyset)} |\pi|\), that is \(d_{ij}\) is the length of the shortest d-path from \(i\) to \(j\). Note that if \(i\perp\!\!\!\perp j|k\) is an imposed relation, then \(d_{ij}>1\). Order the variables as follows: \(\sigma_{ij}\succ\sigma_{i'j'}\) if \(d_{ij}>d_{i'j'}\) or \(d_{ij}=d_{i'j'}\) but \((i,j)\succ_{\mathbb{Z}^2}(i'j')\), where \(\succ_{\mathbb{Z}^2}\) is any order on \(\mathbb{Z}^2\). We also set \(\sigma_{ij}\succ 1\) for all variables.\ Let \(\alpha=(\alpha_{ij})\) be a vector. Denote by \(\sigma^\alpha\) the monomial \(\prod \sigma_{ij}^{\alpha_{ij}}\). Now define the relation \(\sigma^\alpha\succ_\mathrm{dag} \sigma^\beta\) if the first non-zero coordinate of \(\alpha-\beta\) is positive. Note that \(\succ_\mathrm{dag}\) is a lexicographic order. The above order has a pleasant property: the leading monomial of quadratic relations \(\sigma_{ij}-\sigma_{ik}\sigma_{kj}\) generated by CI relations of \(T\) is always the linear form \(\sigma_{ij}\). We use this to prove that the quadratic and linear imposed relations used to generate \(\hat{I}'_T\) form a Gr\(\ddot{\textup{o}}\)bner  basis w.r.t to this order: It was asked in (see Conjecture 5.9) if there exists a Gr\(\ddot{\textup{o}}\)bner  basis consisting of square free terms of degree one and two for \(J_T\). The above proposition shows that this is the case for \(\hat{I}_T\) (or \(\hat{J}_T\)). Finally, using the procedure in section [2.7](#sec:rand){reference-type="ref" reference="sec:rand"}, we list the isomorphism classes of trees on \(4,5,\) and \(6\) nodes. See Figure [\[fig_classes\]](#fig_classes){reference-type="ref" reference="fig_classes"}. Our computations are done in Magma. We use this theorem to list the isomorphism classes of trees on \(4\) and \(5\) nodes. See Figure [\[fig_classes\]](#fig_classes){reference-type="ref" reference="fig_classes"}. # Marginalization Exact inference on a DAG \(G\) is the problem of extracting marginals of a \(G\)-compatible distribution. In this section, we establish a connection between the obstructions to embeddings of \(Y_G\) and obstructions to efficient inference on \(G\). Take for instance the DAG To extract the marginal on nodes 1 and 3, one needs to eliminate nodes 2 and 4. A well known problem is that the order of elimination matters. For instance, eliminating node 4 leads to a "sparse" graph whereas eliminating node 2 gives a "dense" graph A fact is that the computational effort in sequential eliminations is controlled by the sparsity of the subgraphs that appear in the process. The problem is that there is no good way of finding the right elimination order. What is worse is that, by the recent results in, it is even hard to say if such an order exists or not. The difficulty of inference in a DAG is controlled by the so called tree-widths of its underlying graph. This is hard to compute for large graphs, and it thus makes sense to settle for an easier question: *what are some necessary conditions for a good elimination order to exist?* One observation is that having small tree-width implies that, under some suitable ordering, the graphs that appear in the elimination process have small tree-width as well, and one can take this as a measure of efficiency for inference. In the above example, we see that the marginals of interest in \(G\) are supported on the (unlabeled) Markov chain: However, the same cannot be said about the complete graph on four nodes. It is thus useful to know how simple the space of models that support the marginals of a certain DAG can be. This gives us a notion of complexity of marginals in DAGs and it is clear that this notion depends on the class of the model, and not on how the model is represented as a graph. We formalize this notion as follows: In other words, \(M\) lies below \(G\) if the marginal of any \(G\)-compatible distribution on the subspace associated to \(M\) factorizes w.r.t it. We note that the elimination of a DAG need not be a DAG (see ), but in cases that it is, the notions of minimal model and elimination model coincide. The main question of interest is to decide when \([M]\) can lie below \([G]\). It is clear that an enumerative approach to answer this question is problematic, even for a small graph \(M\), as the number of possibilities grow exponentially with the size of the eliminated subset (which grows as \(M\) gets smaller). The next proposition gives a necessary condition. We use the following proposition later: This proposition shows that when \(m=1\) the Segre embedding \(\mathbb{P}^n\times\mathbb{P}^m\hookrightarrow \mathbb{P}^{nm+n+m+1}\) uses the minimal target dimension[^10]. This can be useful for our purposes. and \(M_n\) be a disconnected V-structure on \(n+1\) nodes One can check that \(Y_{G_n}\simeq\mathbb{P}^{n}\times\mathbb{P}^1\) and that \(Y_{M_n^e}\simeq\mathbb{P}^{2n}\). Then \([M_n]\) does not lie below \([G_n]\) by the above proposition. This example is tight, in the sense that \([M_{n-1}]\) does lie below \([G_n]\) as eliminating the nodes \(n+1\) and \(n+2\) makes \(G_n\) completely disconnected. In particular, \(G_2\) does not lie above \(\bullet\quad \bullet\to \bullet\). We want to show that for all \(n\), \([G]\) does not lie above \([M]\) where \([M]\) is the class of a V-structure \(\bullet \rightarrow \bullet \leftarrow \bullet\). The extension \([M^e]\) of a V-structure to four nodes is the DAG Suppose that \([M]\) is below \([G]\) to obtain an embedding \(\iota:Y_G\hookrightarrow Y_{M^e}\). We note that \(\Pi_G\) is smooth everywhere except at the vertex of the affine cone over \(\mathbb{P}^1\times\mathbb{P}^1\). Blowing up \(Y_{M^e}\simeq\mathbb{P}^5\) at \(\iota(p)\) gives a commutative diagram: Since the property of being a closed immersion is stable under base change, it follows that the map \(\mathrm{Bl}_p(Y_G)\to \mathrm{Bl}_{\iota(p)}(\mathbb{P}^5)\) is an embedding (see corollary II.7.15 in We note that \(\mathrm{Bl}_p(\Pi_G)\) is smooth, and has the structure of a \(\mathbb{P}^1\)-bundle over \(\mathbb{P}^2\times \mathbb{P}^1\). Likewise, the blow-up \(\mathrm{Bl}_{\iota(p)}(\mathbb{P}^5)\) is a \(\mathbb{P}^1\)-bundle over \(\mathbb{P}^4\). In other words, after blow ups, we get projective bundles over DAG varieties where \(G'\) looks like and \(M'\) is One can check that \(\mathrm{Pic}(\mathrm{Bl}_{\iota(p)}(Y_{G}))=\mathbb{Z}^3\) while \(\mathrm{Pic}(\mathrm{Bl}_p(Y_{M^e}))=\mathbb{Z}^2\). If an embedding exists, \(\mathrm{Bl}(Y_{G})\) must be isomorphic to a smooth codimension one subvariety in \(\mathrm{Bl}_p(Y_{M^e})\), which has Picard group isomorphic to \(\mathbb{Z}^2\) by the Lefschetz's hyperplane theorem. This example shows that if two Markov chain relations fit in a DAG, they cannot combine to produce a pure independence relation. The fact that the relations fit in a DAG is essential here. lies above \(M:\bullet\to\bullet\to\bullet\). There is an embedding \(\iota:Y_G\hookrightarrow \mathbb{P}^{15}\) defined by the ideal sheaf \(\mathcal{I}_G\). If \([G]\) lies above \([M]\), then there must exist an irreducible quadric hyper-surface in \(\mathbb{P}^{15}\) that contains the image of \(\iota\). This means that the map \(\Gamma(\mathbb{P}^{15},\mathcal{O}_{\mathbb{P}^{15}}(2))\to \Gamma(\mathbb{P}^{15},\iota_*\mathcal{O}_{Y_G}(2))\) has an irreducible quadratic polynomial in its kernel, which corresponds to a global section of \(\mathcal{I}_G(2)\). Using Magma, we compute the Hilbert polynomial of \(\mathcal{I}_G\) and conclude that \(h^0\mathcal{I}_G(1)=2\) and \(h^0\mathcal{I}_G(2)=31\). Thus there are 2 linearly independent hyperplanes[^11] that contain the image of \(\iota\). These in turn give 31 reducible (and thus no irreducible) quadrics containing the image. This shows that \([G]\) does not lie above \([M]\). This section justifies the following questions: 1) *what are some invariants of DAG varieties that are useful for ruling out embeddings of the above type?* 2) *how are such invariants related to the combinatorial data of the DAG?* 3) *when can the numbers \(a_i\) be read from the Hilbert polynomial of \(\mathcal{I}_G\)?* We plan to come back to these questions in a future paper.\ [^1]: The essential graphs were originally known as completed patterns and were introduced in as the maximal invariants associated to equivalence classes of DAGs. They were also studied in under the name maximally oriented graphs. [^2]: We mostly work with \(k=\mathbb{R}\) or \(\mathbb{C}\). Since \(\mathbb{R}\) and \(\mathbb{C}\) are measurably isomorphic, it does not matter which one we pick. We write \({G}/k\) if we need to emphasize the base field \(k\). [^3]: Note that \(\Sigma^{\mdot{}\,}\) is Zariski open, while \(\Sigma^{+},\Sigma^{++}\) are described by inequalities. [^4]: The closure is always taken inside the affine complex space. [^5]: A vector random variable is said to be non-singular if its distribution admits a density w.r.t. product Lebesgue measure. [^6]: This follows directly from the lower semi-continuity of divergence. [^7]: This statement generalizes to the class of chain graphs. See for details. [^8]: For instance, take \((i,j)<(i',j')\) if \(i<i'\) or \(i=i'\) and \(j<j'\) in the topological sort. [^9]: By an embedding we mean a closed immersion. [^10]: In the general case, there are embeddings \(\mathbb{P}^n\times\mathbb{P}^m\hookrightarrow \mathbb{P}^{2(m+n)-1}\) and these can be shown to have the smallest target dimension. [^11]: Two hyperplanes are said to be linearly independent if they are defined by linearly independent sections of \(O(1)\).

{'timestamp': '2014-12-24T02:07:20', 'yymm': '1401', 'arxiv_id': '1401.5551', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5551'}

# Introduction Quantum Chromodynamics is a truly remarkable theory, describing the spectroscopy and dynamics of hadron and nuclear physics in terms of fundamental, but confined, quark and gluonic non-Abelian gauge fields. The existence of heavy quarks gives a new dimension to QCD, allowing tests of the theory and the nature of color confinement in many diverse ways. The discovery of the charged \(Z^+\) states , which can be interpreted as a \(c \bar c u \bar d\) heavy tetraquark state , is just a hint of the vast array of exotic color-singlet resonances or bound states of light and heavy quarks possible in QCD. The very strong spin correlation  \(A_{NN}\) appearing at \(\sqrt s \simeq~ 3 ~{\rm and} ~5 ~{\rm GeV}\) in large-angle elastic proton-proton scattering \(pp \to pp\) can be understood as due to the excitation of charge \(Q=+2\) baryon number \(B=2\) \(|uud uud Q \bar Q \rangle\) resonances in the \(s\)-channel at the strangeness and charm thresholds . Similarly, one expects color-singlet \(Q=+4\) hexaquark resonances  \(|uuu uuu\rangle\), \(|uuu uuc\rangle,\) and \(uuu ucc\rangle\) in which the six \(3_C\) quarks are bound as a color-singlet S-wave configuration, analogous to the "hidden-color\" configurations  which dominate the dynamics of the deuteron at short distances. In addition, the attractive multi-gluon exchange van der Waals potential leads to the prediction  of "nuclear-bound quarkonium\" states such as \([Q \bar Q A].\) Hadrons in QCD are eigenstates of the light-front Hamiltonian \(H^{QCD}_{LF} |\Psi_H\rangle = M^2_H |\Psi_{LF}\rangle\), the evolution operator in light-front time \(\tau = t +z/c\). The hadronic eigenfunction can be projected on the \(n\)-particle eigensolutions \(| n \rangle\) of the free QCD Hamiltonian to generate the frame-independent LF Fock state wavefunctions \(\psi^H_n(x_i, \vec k_{\perp i}, S^z_i ) = \langle\Psi_H | n\rangle\), where \(x_i= {k^0+k^z \over P^0 +P^z}={k^+_i \over P^+}\) (with \(\sum_i^n x_i =1\)) are the quark and gluon light-front momentum fractions, the \(\vec k_{\perp i}\) (with \(\sum^n_i \vec k_{\perp i} = 0_\perp\)) are the transverse momenta, and the \(S^z_i\) are the constituent spin-projections in the \(\hat z\) direction which satisfy \(J^z\) angular momentum conservation. Given its LF wavefunction, one can compute the quark and gluon composition of a hadron and thus virtually all its hadronic observables. For example, the square of the LFWFs generate structure functions and their overlaps determine the elastic and inelastic form factors. For a review see ref.. The entire hadronic spectrum in QCD (1+1), including higher Fock states, can be systematically computed using the discretized light-cone quantization (DLCQ) method  In general, the hadron eigensolution has an infinite number of distinct Fock states; higher Fock states such as \(|uud Q \bar Q\rangle\) generate the heavy sea quark distributions, both the "extrinsic\" contributions corresponding to gluon splitting \(g \to Q \bar Q\) predicted by DGLAP evolution, plus the "intrinsic\" contributions  in which the heavy quarks are multi-connected to the valence quarks. One can show from the operator product expansion  that the probability of intrinsic heavy quarks in a light hadron decreases as \(1/ M^2_Q\), corresponding to the twist-6 operator \(G^3_{\mu \nu}\) in non-Abelian QCD. In contrast, the fall-off in Abelian QED is \(1/ M^4_Q\), corresponding to the twist-8 Euler-Heisenberg light-by-light scattering operator \(F^4_{\mu \nu}\). # The Unique Color-Confining Potential The QCD Lagrangian has no explicit mass scale if all quark masses are set to zero. The classical QCD Lagrangian and its action thus have a fundamental conformal symmetry in the limit of massless quarks. Remarkably, as first shown by de Alfaro, Fubini, and Furlan , the action will retain its conformal invariance even if the Hamiltonian is augmented by mass terms proportional to the dilatation operator \(D\) and the special conformal operator \(K\). If one applies the dAFF formalism to light-front Hamiltonian theory, the \(q\bar q\) interaction has a unique form of a confining harmonic oscillator potential  which depends on a single mass-scale parameter \(\kappa\). The result is a nonperturbative relativistic light-front quantum mechanical wave equation for mesons which incorporates color confinement, a mass gap, and other essential spectroscopic and dynamical features of hadron physics. The same light-front equations arise from the holographic mapping of the soft-wall model modification of AdS\(_5\) space--with a unique dilaton profile \(e^{\kappa^2 z^2}\)--to QCD (3+1) at fixed light-front time \(\tau\). "Light-front holography\"  thus provides a precise relation between amplitudes in the fifth dimension of AdS space and light-front wavefunctions. The effective \(q \bar q\) color-confining interaction between light quarks in the LF Hamiltonian derived from AdS/QCD and light-front holography has the unique form of a two-dimension harmonic oscillator \(U(\zeta^2) = \kappa^4 \zeta^2 + 2\kappa^2(J-1)\) where the invariant variable \(\zeta^2= b^2_\perp x(1-x)\) is conjugate to the invariant mass squared. A complimentary argument for the form of the LF-confining potential is given in ref. . The eigensolutions of this "Light-Front Schrödinger Equation\" correspond to a massless pion for zero quark mass and linear Regge trajectories with the same slope in the radial quantum number \(n\) and orbital angular momentum \(L\) . The resulting LF wavefunctions have remarkable phenomenological features. For example, the proton eigensolution of the corresponding LF Dirac equation has equal probability to have relative quark-diquark relative angular momentum \(L=0\) and \(L=1\). Forshaw and Sandapen  have shown that the resulting predictions for \(\rho\) electroproduction agree with experiment. The quark counting rules  for hard exclusive processes are first-principle features of AdS/QCD . The shape of the QCD running coupling in the nonperturbative domain is also predicted , in agreement with effective charge phenomenology for \(Q^2 < 1~{\rm GeV}^2\). The AdS/QCD light front approach has recently been extended to heavy quarks, successfully describing the spectroscopy, wavefunctions and decays of heavy hadrons . # Anomalous Heavy-Quark Measurements A number of recent experimental results involving heavy quarks appear to be in striking disagreement with conventional expectations. ## Top/Anti-Top Asymmetry and PMC Renormalization Scale Setting The \(t\) versus \(\bar t\) momentum asymmetry measured at the Tevatron by CDF   and by D0  in \(\bar p p \to t \bar t X\) disagrees with canonical PQCD predictions by more than \(3 \sigma.\) However, as Xing-Gang Wu and I have shown in refs.  , this disagreement can be attributed to an arbitrary, scheme-dependent choice of the renormalization scale of the QCD running coupling constant \(\alpha_s(\mu^2)\). In contrast, when one uses the scheme-independent Principle of Maximum Conformality (PMC)  to set the scale, the discrepancy between pQCD prediction and experiment is reduced to 1 \(\sigma.\) The running coupling in a gauge theory sums the terms involving the \(\beta\) function; thus when the renormalization scale is set properly, all non-conformal \(\beta \ne 0\) terms in a perturbative expansion arising from renormalization are summed into the running coupling. The remaining terms in the perturbative series will then be identical to those of a conformal theory; i.e., the corresponding theory with \(\beta=0\). As discussed by Di Giustino, Wu, Mojaza, and myself , the resulting scale-fixed predictions using the PMC are independent of the choice of renormalization scheme-- as required by the renormalization group. The PMC is the principle  which underlies the BLM scale-setting method.  The PMC/BLM scales are fixed order-by-order and the scales then automatically determine the number \(n_f\) of effective flavors in the \(\beta\)-function analytically . The results avoid the divergent renormalon resummation  and agree with QED scale-setting in the Abelian limit. In the case of QED, the PMC scale is proportional to the photon virtuality and thus sums all vacuum polarization corrections to all orders. Different schemes lead to different effective PMC/BLM scales, but the final results are scheme independent. The PMC procedure is also valid for multi-scale processes. One can introduce a generalization of conventional dimensional regularization, the \({\cal R}_\delta\) schemes. For example, if one generalizes the \(\bar{MS}\) scheme by subtracting \(\ln 4 \pi-\gamma_E-\delta\) instead of just \(\ln 4 \pi-\gamma_E\) the new terms generated in the pQCD series that are proportional to \(\delta\) expose the \(\beta\) terms and thus the renormalization scheme dependence. Thus the \({\cal R}_\delta\) schemes uncover the renormalization scheme and scale ambiguities of pQCD predictions, expose the general pattern of nonconformal terms, and allow one to systematically determine the argument of the running coupling order-by-order in pQCD in a form which can be readily automatized . The resulting PMC scales and the finite-order PMC predictions are to high accuracy independent of the choice of the initial renormalization scale. The PMC satisfies all of the principles of the renormalization group: reflectivity, symmetry, and transitivity, and it thus eliminates an unnecessary source of systematic error in pQCD predictions . The BLM/PMC also provides scale-fixed, scheme-independent high-precision connections between observables, such as the "Generalized Crewther Relation" , as well as other "Commensurate Scale Relations" . The renormalizations scales for multi-scale amplitudes are also determined by the PMC. For example, the PMC/BLM scale  for the running coupling appearing in the final state for heavy-quark production at threshold is proportional to the relative velocity within the \(Q \bar Q\) pair, very different than the usual assumption that the renormalization scale is of order of the heavy quark mass. The elimination of the renormalization scheme ambiguity thus improves the accuracy of pQCD tests and increases the sensitivity of LHC experiments and other measurements to new physics beyond the Standard Model. ## High-\(x\) Strangeness Distributions The strange quark \(s(x,Q^2) + \bar s(x,Q^2)\) distribution in the proton measured by HERMES in \(\gamma^* p \to K X\) reactions  has significant support at large \(x_{bj}\), in contradiction with usual expectations. The strange quark in the proton is seen to have two distinct components: a fast-falling contribution, consistent with gluon splitting to \(s \bar s\), and an approximately flat component up to \(x < 0.5\). See fig. [\[Hermes\]](#Hermes){reference-type="ref" reference="Hermes"}. As emphasized by Chang and Peng , the "intrinsic\" component  at high \(x\) is in agreement with expectations derived from the nonperturbative 5-quark light-front (LF) Fock state \(|uuds\bar s\rangle\) of the proton. The dominant configuration in \(x_i\) and \(k_{\perp i}\) will minimizes the total invariant mass: \({\cal M}^2= \sum_{i=1}^5 {k^2_{\perp i } +m_i^2\over x_i}\); i.e. minimal rapidity differences of the constituents. Equal rapidity implies that the light-front momentum fractions are proportional to the quark transverse mass: \(x_i \propto m_{\perp i}\) with \(m_{\perp i} = \sqrt {k^2_{\perp i} + m^2_i}.\) The heavy quarks in the\(|uudQ\bar Q\rangle\) Fock state thus have the largest LF momentum fraction \(x_i\). The high-\(x\) intrinsic strange quarks can reinteract with the valence quarks in the \(|uud s \bar s\rangle\) Fock state since all of the constituents in the LF Fock state tend to have the same rapidity. This leads to a \(s(x,Q^2)\) versus \(\bar s(x,Q^2)\) asymmetry in both momentum and spin, as also expected when one identifies  the \(|uud s \bar s\rangle\) Fock state with the analogous \(| K^+(\bar s u) \Lambda(sud)\rangle\) hadronic state. Similarly, the \(\bar u(x) \ne \bar d(x)\) asymmetry can be identified with the nonperturbative dynamics of the \(|uud q \bar q\rangle\) Fock state. ## High Transverse Momentum Heavy Quark Jet Production and the Physics Consequences of Intrinsic Heavy Quarks The cross sections for high transverse photon plus a charm jet cross section \(\bar p p \to \gamma c X\) and also \(Z^0\) plus a charm jet \(\bar p p \to Z^0 c X\) measured at the Tevatron  for \(p_T^\gamma > 60~ GeV/c\) appear to be substantially larger than predicted using conventional charm PDF distributions. In contrast, the corresponding rate for \(\bar p p \to b + \gamma X\) agrees well with NLO PQCD predictions. The dominant underlying 2 to 2 subprocesses  are \(g c \to \gamma c\) and \(g c \to Z^0 c\), which depend critically on the assumed parametrization of the charm quark PDFs at \(x>0.1\) and high \(Q^2 \sim 10^4~{\rm GeV}^2\). The charm distribution in the proton predicted by QCD includes an intrinsic component --the five-quark Fock state \(|uud c \bar c\rangle\) derived from multigluonic couplings of the \(c \bar c\) pair to the proton's valence quarks. The intrinsic charm quarks appear at large \(x\) since this minimizes the off-shellness of the LFWF. In fact, the EMC determination  of \(c(x,Q^2)\) at \(x=0.42\) and \(Q^2 = 75~{\rm GeV}^2\) is approximately 30 times larger than predicted by the soft distribution from gluon splitting \(g \to c \bar c\). CTEQ parametrizations  include the intrinsic charm as measured by EMC. The photon plus charm-jet anomaly could possibly be explained if one allows for a substantial intrinsic contribution to the charm structure function in \(g c \to c \gamma\) at \(Q^2 \sim 10^4~{\rm GeV}^2\), but one requires a factor of two increase in strength compared to the CTEQ PDF. The reduction of the charm distribution at large \(x\) due to DGLAP evolution is likely to have been overestimated because one conventionally takes \(m_c=0\): it is clearly important to evolve \(c(x,Q^2)\) to the high \(Q^2 \sim 10^4~{\rm GeV}^2\) domain using massive charm quark in the DGLAP evolution equations; The argument of the running coupling and the effective number of flavors in the QCD \(\beta\) function in the DGLAP evolution equations can be set using the scheme-independent PMC method. These questions are now being investigated by Gang Li and myself. The ratio of intrinsic charm to intrinsic bottom scales as \(m_c^2/m_b^2 \simeq 1/10,\) using the operator product expansion in non-Abelian QCD,  so that intrinsic bottom plays a minor role in the Tevatron measurements. In the case of a hadronic high energy proton collision, such as \(pp \to \Lambda_c X\) the high-\(x\) intrinsic charm quark in the proton's \(|uud c \bar c\rangle\) Fock state can coalesce with the co-moving \(ud\) valence quarks in a projectile proton to produce a forward \(\Lambda_c(cud)\) baryon at the combined high momentum fraction \(x_F = x_u + x_d + x_c\). Similarly, the coalescence of comoving \(b\) and \(\bar u\) quarks from the \(|uud \bar b b\rangle\) intrinsic bottom Fock state can explain the high \(x_F\) production of the \(\Lambda_b(udb)\), which was first observed at the ISR collider at CERN  in association with a positron from the decay of the associated high-\(x_F\) B meson. A similar mechanism predicts quarkonium hadroproduction at high \(x_F\) . The NA3 experiment  has observed the hadroproduction of **two** \(J/\psi\)s at high \(x_F\), a signal for seven-quark Fock states such as \(|uud c \bar c c\bar c\rangle\) in the proton . The intrinsic contributions can explain both the open-charm and open-bottom hadron production at high momentum fractions, and it can also account for single and double \(J/\psi\) hadroproduction measured by NA3 at high \(x_F\) . Measurements by the SELEX collaboration  have led to the discovery of a set of doubly-charmed spin-1/2 and spin-3/2 baryons with quantum numbers that can be identified as \(|ccu \rangle\) and \(|ccd \rangle\) bound states--again a signal for seven-quark Fock states such as \(|uud c \bar c c\bar c\rangle\) in the proton. Surprisingly, the mass splittings of the \(ccu\) and \(ccd\) states measured by SELEX are much larger than expected from known QCD isospin-breaking effects . One speculative proposal  is that these doubly charmed baryons have the configuration \(c~q~c\) where the light quark \(q\) is exchanged between the heavy quarks, as in a linear molecule. This configuration may enhance the Coulomb repulsion of the \(c~u~c\) relative to \(c~d~c\). It is clearly important to have experimental confirmation of the SELEX results. The presence of intrinsic heavy quarks in the proton leads to a novel mechanism for the inclusive and diffractive Higgs production \(pp \to p p H\) where the Higgs boson carries a large fraction of the projectile proton momentum.  This high \(x_F\) production mechanism is based on the subprocess \((Q \bar Q) g \to H\) where the Higgs couples to the sum of the momentum of the \(Q \bar Q\) pair in the \(\vert uud Q \bar Q \rangle\) intrinsic heavy quark Fock state of the colliding proton; it thus can be produced with approximately \(80\%\) of the projectile proton's momentum. High-\(x_F\) Higgs production could be measured at the LHC using far-forward detectors or arranging the proton beams to collide at a significant crossing angle or at different beam energies. The same mechanism can produce the Higgs at large \(x_F\) in \(\gamma p \to X H\) collisions at the LHeC. Intrinsic charm in light hadrons also provides a solution to the \(J/\psi \to \rho \pi\) puzzle . The conventional assumption is that quarkonium states decay into light hadrons via annihilation into virtual gluons, the OZI rule. However, this hypothesis leads to the identical decays of the \(J/\psi\) and the \(\Psi^\prime\), up to a factor of \(12\%\) from the wavefunction at the origin squared--in strong disagreement with measurement. Worse, the decay \(J/\psi \to \rho \pi\) is predicted to be suppressed by hadron-helicity conservation , when in fact it is the largest two-body hadronic decay. The \(J/\psi \to \rho \pi\) puzzle can be explained if the \(c \bar c\) does not annihilate but instead flows into the intrinsic charm Fock state \(|q \bar q c \bar c\rangle\) of one of the final state meson. In contrast, the \(\psi^\prime \to \rho \pi\) decay is suppressed by change of sign in the decay amplitude from the node in the \(\psi^\prime\) wavefunction. Intrinsic charm also affects B-decays in a novel way . The presence of intrinsic charm in the hadronic light-front wave function of the \(B\), even at a few percent level, provides new, competitive decay mechanisms for B decays which are nominally CKM suppressed. The impact of intrinsic heavy quark distributions in the proton on new physics searches at the high intensity frontier is discussed in ref. . Precision measurements of the charm and bottom distributions in hadrons are clearly of prime interest. This can be done in lepton-scattering facilities such as COMPASS, the LHeC, and the new lepton-ion colliders proposed at BNL and JLab. The proposed fixed target program "AFTER\"   at the LHC will allow remarkable accessibility to heavy-quark phenomena in the target and projectile fragmentation regions of rapidity. For example, heavy charm and bottom hadrons will be produced with relatively low rapidities in the easily instrumented target rapidity domain at AFTER. Just as important, the existence of charm quarks at high momentum fraction in the proton wavefunction implies enhanced production of open and hidden charm states in the threshold regime. Since the produced quarks and gluons are produced at threshold at small relative rapidity differences, this provides an important opportunity  to create exotic heavy quark states at upcoming facilities such as JLab at 12 GeV, PANDA, and NICA. # Nuclear Suppression of Quarkonium and Di-Gluon Saturation Since its radius is small, the cross section for the interaction of a charmonium state in nuclear matter is expected to be only a few millibarns, as expected from QCD color transparency . Thus the usual expectation is that hadroproduction cross sections are approximately linear in the number of nucleons \(A\). However, the production cross section \(p A \to J/\psi X\) measured by LHcB  and ALICE  at forward rapidity \(y \sim 4\) shows an unexpectedly strong nuclear suppression, close to \(A^{2/3}\). This effect cannot be accounted for by shadowing of the nuclear gluon distribution. Arleo and Peigne  suggest that the strong nuclear suppression of \(J/\psi\) production in \(pA\) collisions can be explained as a manifestation of the "color-octet\" model: the \(c \bar c\) propagates through the nucleus as a color-octet, and its nuclear energy loss will be proportional to its energy if the induced gluon radiation is coherent on the entire nucleus. The color-octet \(c \bar c\) is assumed to convert to the color-singlet \(J/\psi\) after exciting the nucleus. HuaXing Zhu and I have postulated an alternative mechanism: We assume that the basic QCD mechanism for \(J/\psi\) production at small transverse momentum is \([gg] g \to J/\psi\) where the \([gg]\) is a color-octet di-gluon from the proton. See fig. [\[Digluon\]](#Digluon){reference-type="ref" reference="Digluon"}. The propagating color-octet di-gluon has a large interaction cross section, and it thus interacts primarily at the nucleus front surface, giving a production cross section \(\sigma(p A \to J/\psi X) \propto A^{2/3}.\) It should be noted that since \(g(x,Q^2)\) falls rapidly, two gluons in the di-gluon, each with \(x \sim 0.01\), have a higher probability than a single gluon with \(x \sim 0.02\). The di-gluon mechanism is expected to diminish in strength at increasing \(p_T.\) The di-gluon subprocess is the color-octet analog of the color-singlet two-gluon exchange mechanism  underlying diffractive processes like \(\ell p \to \ell p X\). The di-gluon multiparton subprocess is analogous to the higher-twist subprocess \([q\bar] q q \to \gamma^* q\) which dominates the \(\pi N \to \ell \bar \ell X\) Drell-Yan reaction at high \(x_F\), accounting for the observed dramatic change from transverse to longitudinal virtual photon polarization . Similarly, multiparton "direct\" subprocesses can account  for the observed anomalous power-law fall-off of high \(p_T\) inclusive hadron production cross sections \({ d \sigma\over d^3p/E}(p p \to h X)\) at fixed \(x_T = 2{p_T\over \sqrt s}\) and fixed \(\theta_{CM}\). The \(pA \to J/\psi X\) cross sections measured in fixed-target experiments at CERN and FermiLab at high \(x_F\) also show strong nuclear suppression at high \(x_F\). The ratio of the nuclear and proton target cross sections has the form \(A^{\alpha(x_F)}\), where \(x_F\) is Feynman fractional longitudinal momentum of the \(J/\psi\). At small \(x_F\), \(\alpha(x_F)\) is slightly smaller than one, but at \(x_F \sim 1\), it decreases to \(\alpha=2/3\). These results are again surprising since (1) the \(\alpha= 2/3\) is characteristic of a strongly interacting hadron, not a small-size quarkonium state; and (2) the functional dependence \(A^{\alpha(x_F)}\) contradicts pQCD factorization . The observed nuclear suppression, in combination with the anomalously nearly flat cross section at high \(x_F\), points to a QCD mechanism based on the intrinsic charm Fock state . QCD predicts that the color-configuration of the heavy quark pair \(Q \bar Q\) in the intrinsic five-quark Fock state is primarily a color-octet. The intrinsic heavy quark Fock state of the proton: \(|(uud)_{8_C} (c \bar c)_{8_C} \rangle\) thus interacts primarily with the \(A^{2/3}\) nucleons at the front surface because of the large color-dipole moment of the color-octet \(c \bar c\). The \(c \bar c\) color octet thus interacts primarily on a front-surface nucleon, changes to a color singlet, and then propagates through the nucleus as a \(J/\psi\) at high \(x_F\). See fig. [\[AdepIC\]](#AdepIC){reference-type="ref" reference="AdepIC"}. # Conclusions The phenomenological disagreements with conventional pQCD predictions discussed in this contribution are not necessarily due to new physics beyond the standard model; instead they may point to features of QCD itself, such as intrinsic heavy quarks at high \(x\) and the need to determine the appropriate renormalization scales. I have outlined a number of novel physics consequences of intrinsic heavy quarks, such as the hadroproduction of the Higgs and exotic heavy quark states, both at high \(x_F\) and at threshold. Multi-parton subprocesses, such as di-gluon initiated reactions, can also play an important role. I have discussed how AdS/QCD and light-front holography provide a new analytic approach to the QCD confinement potential, hadron spectroscopy, hadron dynamics, and the origin of the QCD mass scale. I have also emphasized that the renormalization scale ambiguity can be consistently eliminated at finite orders in pQCD using the scheme-independent PMC procedure, thus eliminating an unnecessary source of theoretical systematic error.

{'timestamp': '2014-01-29T02:00:34', 'yymm': '1401', 'arxiv_id': '1401.5886', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5886'}

{'timestamp': '2014-03-04T02:08:17', 'yymm': '1401', 'arxiv_id': '1401.5361', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5361'}

# Introduction Self-propelled particles convert chemical, magnetic, or radiation energy into directed motion. In this way, kinetic energy is intrinsically introduced into the corresponding system on the single particle level and length scale, in contrast to a process of macroscopic stirring or shaking from outside. Far-reaching consequences on the single particle behavior as well as on their collective properties emerge. Single artificially generated Janus particles can start to self-propel mostly due to phoretic effects when one of the two sides is selectively heated by laser light or catalyzes a chemical reaction. In a liquid environment, such single colloidal particles obey rotational as well as translational diffusion. On time scales that are short when compared to the characteristic rotational diffusion time, self-propulsion shows up as directed motion. On significantly longer time scales, the overall diffusion coefficient is considerably enhanced. Also the consequences of a possible additional particle deformability were studied extensively as it may occur, for example, for self-propelled droplets on surfaces or in a bulk fluid. Another example of artificial self-propelled particles are granular hoppers that transform vertical vibration energy into directed horizontal motion. On the biological side, the conversion of chemical energy into directed motion allows bacteria to individually search for food or to swim towards or away from illuminating light. Likewise, chemical energy is used by bacteria, amoebae, or tissue cells to crawl on substrates. Propulsion mechanisms are manifold and reach from the beating of a single flagellum via twisting deformations of whole filamental bacterial bodies to the synchronization of the motion of thousands of cilia. When many self-propelled particles act together, collective modes of migration emerge, like swarms that move as single entities, traveling density bands, or lanes of counterpropagating particles. Even turbulent states arise that are based on the energy conversion on the single particle length scale. Already without adhesive interactions and in the absence of alignment mechanisms, clusters of self-propelled particles can form in high-density systems. These clusters appear when there is no interaction that aligns the migration directions of different particles so that they can mutually block their motion. For monodisperse two-dimensional systems, hexagonal packing can arise. Confining walls can support a local clustering or trapping of self-propelled particles at the boundaries. If the particle interaction is strong enough compared to their self-propulsion, then a formation of active crystals composed of self-propelled particles is to be expected. For this scenario, particle simulations demonstrated the emergence of crystal-like objects that collectively migrate as a single structure. In contrast, the approach that we recently introduced and analyze here in breadth is a complementary continuum one. It combines the description of crystalline materials by the phase field crystal model with ideas from the Toner-Tu model to characterize active crystalline systems of self-propelling objects. To our knowledge, there is no field-theoretical approach preceding Ref.  that includes the translational ordering in active crystals of self-propelled particles. The phase field crystal model is a microscopic field approach introduced by Elder et al. to efficiently model processes in crystalline materials with particle resolution, but on diffusive time scales. In this way it can reproduce results from molecular dynamics simulations, yet in a computationally much more efficient way. Motivated by dynamical density functional theory, the phase field crystal model can in principle be viewed as a microscopic theory describing a manifold of solidification and crystallization phenomena. Examples are defect motion in strained crystals and multidomain structures, epitaxial growth, crystal pinning, crystal growth of binary alloys, branching in crystal growth, or even crack propagation. Quite contrarily, the Toner-Tu model is a macroscopic hydrodynamic-like continuum description of systems of self-propelling objects. Local spatial--i.e. translational--order of the single particles is not resolved or explicitly taken into account. The focus in this model is on the orientational order of the migration directions and thus on the macroscopic emergence of collective motion. It was argued that, if locally the migration directions spontaneously order, long-ranged orientational order of the migration directions and therefore global collective motion into one single collective direction emerges even in two spatial dimensions. Our present work starts from an "active phase field crystal model" that we recently introduced in a previous study. It combines ideas from the two outlined theories, the equilibrium phase field crystal model and the Toner-Tu approach namely. We extend our previous investigation by providing for example more details about inverse active lattice structures and the migration speeds of the collectively migrating textures. Our main focus here is on the stability of traveling active crystals. We therefore perform a linear stability analysis of these structures and demonstrate linear stability even on large length scales that are not accessible by direct numerical simulations. Furthermore, for particles self-propelling on a substrate and surrounded by a thin fluid film, we include the influence of hydrodynamic interactions: they can disturb or destroy the crystalline and orientational order. # Active phase field crystal model As a first step we derive in this section our active phase field crystal model on the basis of dynamical density functional theory for Brownian systems. We consider a system of self-propelled particles in two spatial dimensions. Examples would be amoebae or tissue cells crawling on a substrate or granular hoppers that move horizontally on a vibrating plate and cannot jump over each other. Such systems can directly exchange momentum with the ground. The same should also be true for low Reynolds number swimmers in a thin film of fluid on a substrate. We assume that each particle migrates with an effective active drive of magnitude \(v\) into an individual direction that is characterized by a unit vector \(\mathbf{\hat{u}}\). The one-particle density distribution function is therefore given by \(\rho(\mathbf{x},\mathbf{\hat{u}},\tau)\), with \(\mathbf{x}\) the spatial coordinate and \(\tau\) the time. In two spatial dimensions, the orientation of \(\mathbf{\hat{u}}\) can be parameterized by a single angle \(\vartheta\). Thus the dynamical equation for the one-particle density follows as \[\begin{aligned} \frac{\partial\rho(\mathbf{x},\mathbf{\hat{u}},\tau)}{\partial \tau} & = & \nabla\cdot\mathbf{\tilde{D}_T}\cdot\Big( \beta\, \rho(\mathbf{x},\mathbf{\hat{u}},\tau) \,\nabla \frac{\delta\tilde{\mathcal{F}}}{\delta\rho(\mathbf{x},\mathbf{\hat{u}},\tau)}\Big) \nonumber\\ &&{}+\tilde{D}_r\,\partial_{\varphi}\,\Big( \beta\, \rho(\mathbf{x},\mathbf{\hat{u}},\tau)\, \partial_{\varphi} \frac{\delta\tilde{\mathcal{F}}}{\delta\rho(\mathbf{x},\mathbf{\hat{u}},\tau)}\Big) \nonumber\\ &&{}-\nabla\cdot\mathbf{\tilde{D}_T}\cdot\Big( \rho(\mathbf{x},\mathbf{\hat{u}},\tau)\, v\mathbf{\hat{u}}\,/\tilde{D}_{\|}\Big). \label{eqrho} \end{aligned}\] In this equation \(\beta=(k_BT)^{-1}\) sets the inverse thermal energy, \(\mathbf{\tilde{D}_T}=\tilde{D}_{\|}\mathbf{\hat{u}}\mathbf{\hat{u}}+\tilde{D}_{\bot}(\mathbf{I}-\mathbf{\hat{u}}\mathbf{\hat{u}})\) is the translational diffusion tensor, \(\tilde{D}_r\) denotes the rotational diffusion constant, and \(\tilde{\mathcal{F}}\) is an equilibrium energy functional that will be specified below. The effect of self-propulsion enters through the last term on the right-hand side via the magnitude of the active drive \(v\). Our next step is to derive dynamical equations for the particle density and for the polar orientational order of the self-propulsion directions. We denote the average value of the one-particle density distribution function \(\rho(\mathbf{x},\mathbf{\hat{u}},\tau)\) as \(\bar{\rho}\), i.e. \[\bar{\rho} = \frac{1}{V\Omega}\int\rho(\mathbf{x},\mathbf{\hat{u}},\tau)\,d\mathbf{\hat{u}}\,d\mathbf{x},\] which is a conserved quantity in a closed system. Here, \(V\) is the spatial volume of the system and \(\Omega\) is the surface area of the unit sphere. The modulation of the particle density \(\bar{\rho}\tilde{\phi}_1(\mathbf{x},\tau)\) around \(\bar{\rho}\) is obtained by integrating out the orientational degrees of freedom from \(\rho(\mathbf{x},\mathbf{\hat{u}},\tau)\) and subtracting \(\bar{\rho}\), \[\bar{\rho}\tilde{\phi}_1(\mathbf{x},\tau) = \frac{1}{\Omega}\int\rho(\mathbf{x},\mathbf{\hat{u}},\tau)\,d\mathbf{\hat{u}}-\bar{\rho}.\] Furthermore, we take into account the first orientational moment of \(\rho(\mathbf{x},\mathbf{\hat{u}},\tau)\), which we denote as \[\bar{\rho}\tilde{\mathbf{P}}(\mathbf{x},\tau) = \frac{d}{\Omega}\int\rho(\mathbf{x},\mathbf{\hat{u}},\tau)\,\mathbf{\hat{u}}\,d\mathbf{\hat{u}},\] with \(d\in\{2,3\}\) the dimensionality of the system. From this definition it becomes evident that the polarization field \(\tilde{\mathbf{P}}(\mathbf{x},\tau)\) describes the local orientational order of the active driving directions \(\mathbf{\hat{u}}\) of the self-propelled particles. Similarly to the procedure in Refs. , we expand the one-particle density with respect to its orientational dependence as \[\label{rhoexpand} \rho(\mathbf{x},\mathbf{\hat{u}},\tau)\approx\bar{\rho}+\bar{\rho}\,\tilde{\phi}_1(\mathbf{x},\tau)+\bar{\rho}\,\mathbf{\hat{u}}\cdot\tilde{\mathbf{P}}(\mathbf{x},\tau) +\dots\;.\] To proceed we insert this expansion into Eq. ([\[eqrho\]](#eqrho){reference-type="ref" reference="eqrho"}). From the resulting equation, the different Fourier modes in terms of the angle \(\varphi\) are calculated. In this way we obtain dynamical equations for \(\bar{\rho}\tilde{\phi}_1\) and \(\bar{\rho}\tilde{\mathbf{P}}\) similar to the ones in Refs. , now, however, including the active drive. Applying the approximation of constant mobility and assuming an isotropic translational diffusion \(\tilde{D}_{\|}\approx \tilde{D}_{\bot}\approx \tilde{D}\), the resulting expressions are considerably simplified. We finally find \[\begin{aligned} \partial_{\tau}(\bar{\rho}\tilde{\phi}_1) &=& \frac{\beta \tilde{D}\bar{\rho}}{2\pi}\nabla^2\frac{\delta\tilde{\mathcal{F}}}{\delta(\bar{\rho}\tilde{\phi}_1)}-\frac{v}{2}\,\nabla\cdot(\bar{\rho}\tilde{\mathbf{P}}), \label{eqpsi1dft}\\[.1cm] \partial_{\tau}(\bar{\rho}\tilde{\mathbf{P}}) &=& \frac{\beta \tilde{D}\bar{\rho}}{\pi}\nabla^2\frac{\delta\tilde{\mathcal{F}}}{\delta(\bar{\rho}\tilde{\mathbf{P}})}-\frac{\beta \tilde{D}_r \bar{\rho}}{\pi}\frac{\delta\tilde{\mathcal{F}}}{\delta(\bar{\rho}\tilde{\mathbf{P}})} \nonumber\\[.05cm] & & {}-v\nabla(\bar{\rho}\tilde{\phi}_1). \label{eqPdft} \end{aligned}\] In the following we specify the free energy functional \(\tilde{\mathcal{F}}\). More precisely, we introduce \[\label{Fsum} \tilde{\mathcal{F}}=\tilde{\mathcal{F}}_{pfc}+\tilde{\mathcal{F}}_{\tilde{\mathbf{P}}}\] as a sum of a translational part \(\tilde{\mathcal{F}}_{pfc}\) and an orientational part \(\tilde{\mathcal{F}}_{\tilde{\mathbf{P}}}\). On the one hand, \[\label{Fpfc_phi} \tilde{\mathcal{F}}_{pfc} = \int d^2\!x\,\left\{\frac{1}{2}\phi\left[ a\Delta T+\lambda(q_0^2+\nabla^2)^2\right]\phi + \frac{u}{4}\phi^4 \right\},\] is the free energy introduced as the phase field crystal model. Here, the order parameter \(\phi\) is related to the field \(\tilde{\phi}_1\) in Eq. ([\[rhoexpand\]](#rhoexpand){reference-type="ref" reference="rhoexpand"}) via \(\phi=\bar{\phi}+\phi_1\) and \(\phi_1=\bar{\rho}\tilde{\phi}_1\). The parameters \(a\), \(\lambda\), and \(u\) determine the magnitude of the energetic contributions, whereas \(\Delta T\) is a measure for the temperature. Minimizing this functional for an equilibrium system leads to fluid-like states at higher temperatures or high values of \(|\bar{\phi}|\), hexagonally crystalline states at lower temperatures and intermediate values of \(|\bar{\phi}|\), and lamellar states at lower temperatures and low values of \(|\bar{\phi}|\). The characteristic length scale of the hexagonally crystalline and lamellar structures is determined by the magnitude of \(q_0^{-1}\) in Eq. ([\[Fpfc_phi\]](#Fpfc_phi){reference-type="ref" reference="Fpfc_phi"}). On the other hand, the orientational part is given by \[\label{FtildeP} \tilde{\mathcal{F}}_{\tilde{\mathbf{P}}} = \int d^2\!x\,\left\{ \frac{1}{2}\tilde{C}_1(\bar{\rho}\tilde{\mathbf{P}})^2 + \frac{1}{4}\tilde{C}_4\left[(\bar{\rho}\tilde{\mathbf{P}})^2\right]^2 \right\}.\] The concept behind this free energy is similar to the original approach by Toner and Tu. We have \(\tilde{C}_4\geq0\) due to thermodynamic stability. If \(\tilde{C}_1<0\) and \(\tilde{C}_4>0\), the system can reduce its free energy by an orientational ordering of the self-propulsion directions, given by a nonzero solution of \(\bar{\rho}\tilde{\mathbf{P}}\). This corresponds to an intrinsic tendency towards ordered collective motion. If \(\tilde{C}_1>0\), orientational ordering of the self-propulsion directions always costs orientational free energy. Consequently there is a tendency to avoid collectively ordered motion. This case shows a richer behavior than the other one and will mainly be considered below. Predominantly using the Ramakrishnan-Yussouff expansion, the phase field crystal model could be connected to density functional theory; for a review see Ref. . Expressions for the phenomenological parameters in terms of microscopic correlation functions could be derived. Afterwards the procedure was extended to the case of an additional orientational order. Likewise microscopic expressions for the corresponding phenomenological parameters were listed in these references, and we do not repeat them here. To finally obtain our dynamical equations, we introduce Eqs. ([\[Fsum\]](#Fsum){reference-type="ref" reference="Fsum"})--([\[FtildeP\]](#FtildeP){reference-type="ref" reference="FtildeP"}) into Eqs. ([\[eqpsi1dft\]](#eqpsi1dft){reference-type="ref" reference="eqpsi1dft"}) and ([\[eqPdft\]](#eqPdft){reference-type="ref" reference="eqPdft"}). Moreover we apply the rescaling rules \[\phi=\left(\lambda q_0^4 u^{-1}\right)^{\frac{1}{2}}\psi, \quad a\Delta T=\lambda q_0^4\,\varepsilon, \quad \mathbf{x} = q_0^{-1}\mathbf{r}.\] This rescales \(\phi=\bar{\phi}+\phi_1\) to \(\psi=\bar{\psi}+\psi_1\). All lengths from now on are measured in units of the characteristic length scale \(q_0^{-1}\) of the phase field crystal structures. An additional simplification of the notation follows from the further rescaling \[\begin{aligned} \tau & = & \frac{2\pi}{\beta\tilde{D}\bar{\rho}}\frac{1}{\lambda q_0^6}\,t, \\ v & = & \frac{\beta\tilde{D}\bar{\rho}}{2^{\frac{1}{2}}\pi}\,\lambda q_0^5 \,v_0, \\ \tilde{\mathbf{P}} & = & \frac{1}{\bar{\rho}}\, q_0^2 \left(2\frac{\lambda}{u}\right)^{\!\frac{1}{2}}\,\mathbf{P}, \\ \tilde{C}_1 & = & \frac{\lambda q_0^4}{2}\,C_1, \\[0.05cm] \tilde{C}_4 & = & \frac{u}{4}\,C_4, \\[.1cm] \tilde{D}_r & = & \tilde{D}q_0^2\,D_r. \end{aligned}\] At the end of this procedure we obtain our dynamical order parameter equations in the form \[\begin{aligned} \partial_{t}\psi_1 &=& \nabla^2\Big\{\!\left[\varepsilon+(1+\nabla^2)^2+3\bar{\psi}^2\right]\psi_1 + 3\bar{\psi}\psi_1^2+\psi_1^3\,\Big\} \nonumber\\ & & {}-v_0\,\nabla\cdot\mathbf{P}, \label{eqpsi1}\\[.1cm] \partial_{t}\mathbf{P} &=& \nabla^2\left( C_1\mathbf{P}+C_4\mathbf{P}^2\mathbf{P} \right)-D_r\left( C_1\mathbf{P}+C_4\mathbf{P}^2\mathbf{P} \right)\nonumber\\[.05cm] & & {}-v_0\nabla\psi_1. \label{eqP} \end{aligned}\] The field \(\psi_1\) is a measure for the local particle density, whereas the field \(\mathbf{P}\) measures the local orientational order of the active driving directions of the self-propelled particles. \(\varepsilon\) controls the temperature, \(\bar{\psi}\) the mean density, and \(v_0\) the strength of the active drive of the individual particles. The ordering behavior of the active driving directions of the self-propelled particles is determined by \(C_1\) and \(C_4\), whereas \(D_r\) describes the orientational diffusion of the active drive directions. # Predictions of the model As noted above, the equilibrium phase field crystal model based on Eq. ([\[Fpfc_phi\]](#Fpfc_phi){reference-type="ref" reference="Fpfc_phi"}) shows hexagonally crystalline structures in a certain range of temperature \(\varepsilon\) and mean density \(\bar{\psi}\). Changing the temperature or mean density, hexagonal textures can be transformed into lamellar ones. In the non-equilibrium case considered here, such a transition can also be achieved by increasing the active drive \(v_0\) of the individual self-propelled particles. This scenario is depicted in the upper row of Fig. [\[snapshots\]](#snapshots){reference-type="ref" reference="snapshots"}. The snapshots shown in the figure were obtained by numerically solving Eqs. ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}), starting from random initial conditions. We start with a low value of the active drive \(v_0\) and a hexagonal structure in panel (a). For the parameter values listed in the figure caption, density peaks form that are symmetrically located within "+1"-defects of the polarization field. The overall structure remains at rest. This changes with increasing \(v_0\), when the symmetry of the density peak positions with respect to the "+1"-defects of the polarization field is spontaneously broken. This spontaneous shift of the density peaks out of the centers of the "+1"-defects of the polarization field represents the symmetry breaking necessary for a net active propulsion to emerge. As a whole, the structure starts to migrate (travel) collectively (b). Further increasing \(v_0\) deforms the traveling hexagonal crystal to a traveling rhombic and then to a traveling quadratic crystal (c). Finally it leads to a transition to traveling lamellae (d). Typically, when thinking of crystalline textures in this way, we would identify each density peak with one particle. However, also the situation of cluster crystals is included by our approach. In that case several particles occupy one lattice site. Such active cluster crystalline structures were observed before in a system of deformable self-propelled particles and in a modified Vicsek model confined between two parallel walls. Furthermore, traveling crystalline textures have been observed previously in three-component reaction-diffusion systems. When we thus identify each density peak in the top row of Fig. [\[snapshots\]](#snapshots){reference-type="ref" reference="snapshots"} with a cloud of self-propelled particles, we can similarly interpret the corresponding active inverted textures in the bottom row. In each column of Fig. [\[snapshots\]](#snapshots){reference-type="ref" reference="snapshots"} the two snapshots can be mapped onto each other by noting a special symmetry relation of Eqs. ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}). The equations remain invariant under the simultaneous transformations \(\bar{\psi}\rightarrow-\bar{\psi}\), \(\psi_1\rightarrow-\psi_1\), and \(\mathbf{P}\rightarrow-\mathbf{P}\). Consequently the series of patterns in the bottom row of Fig. [\[snapshots\]](#snapshots){reference-type="ref" reference="snapshots"} follows directly as an inversion of the top row patterns. We will concentrate on the characterization of hexagonal, rhombic, and quadratic crystalline textures in the following. Nevertheless the results apply equally for the inverted structures due to the symmetry property of the dynamical equations. Active honeycomb textures are formed for example by flagellated marine bacteria. To quantify the collective motion of the active crystals, we tracked the migration of each individual density peak in the sample. For this purpose, the center of each density peak was determined at fixed time intervals. On the one hand, this time interval was short enough so that identical density peaks could be identified before and after each migration step. On the other hand, it should be long enough so that the migration step is larger than the mesh size of the calculation grid. It was shown previously that--at least for the finite system sizes investigated--a collectively traveling single crystal develops from the random initial conditions via an intermediate multidomain structure. At the end of this coarsening process towards the active single crystal, all density peaks migrate collectively into the same direction. The migration speed \(v_m\) of the resulting active single crystal follows from a sample average of the peak velocity magnitudes of all density peaks, \(v_m=\sum_{i=1}^{N_p} \|\mathbf{v}_i\|/{N_p}\). Here \(N_p\) denotes the total number of all density peaks in the sample. The velocities \(\mathbf{v}_i\) of the individual density peaks, with \(i=1..,N_p\), follow from the tracking procedure described above. Corresponding results are depicted in Fig. [\[vmvsv0\]](#vmvsv0){reference-type="ref" reference="vmvsv0"} for characteristic values of the parameters \(C_1\) and \(C_4\). As mentioned above, these parameters control the spontaneous orientational ordering of the migration directions of the individual self-propelled particles. We plot in Fig. [\[vmvsv0\]](#vmvsv0){reference-type="ref" reference="vmvsv0"} the collective migration speed \(v_m\) of the single crystal as a function of the active drive \(v_0\) of the individual self-propelled particles. For \(C_1>0\) the active driving directions of the individual particles do not order spontaneously. At low values of the active drive \(v_0\), the overall hexagonally crystalline structure remains at rest, i.e. \(v_m=0\). This corresponds to the situation in Fig. [\[snapshots\]](#snapshots){reference-type="ref" reference="snapshots"}(a). Only at a nonzero threshold value \(v_{0,c}\) do the crystals start to travel collectively, see also Fig. [\[snapshots\]](#snapshots){reference-type="ref" reference="snapshots"}(b). Here, the nonzero value of \(C_4>0\) only plays a minor quantitative role, and a similar curve is obtained for \(C_4=0\). In contrast, there is no threshold when the active driving directions of the individual particles order spontaneously, i.e. \(C_1<0\) and \(C_4>0\). The whole crystalline structure migrates at each nonzero value of the active drive \(v_0\) of the individual particles. Even more, the whole crystal propels collectively at the speed of the individual particles, i.e. \(v_m=v_0\). In addition to this curve of negative \(C_1\), we show in Fig. [\[vmvsv0\]](#vmvsv0){reference-type="ref" reference="vmvsv0"} an intermediate case of \(C_1=0\) and \(C_4>0\). There, the nonlinear \(C_4\)-term leads to a small residual threshold behavior at low values of \(v_0\). During the remaining part of this paper we will be concerned with the case of \(C_1>0\) and \(C_4=0\). It shows the richer behavior due to the emergence of resting as well as traveling structures that are separated by the finite threshold value \(v_{0,c}\). # Linear stability analysis Introducing a hydrodynamic-like continuum theory, Toner and Tu studied the collective motion of flocks of self-propelling objects. Their focus was mainly on the orientational ordering of the self-propulsion directions, and not on a translational, possibly crystalline order of the individual objects. The theory contains nonlinear convective terms due to self-propulsion. Analyzing their model, Toner and Tu found that long-ranged orientational order of the self-propulsion directions can arise in two spatial dimensions, in contrast to what is known for orientational ordering in equilibrium systems. Such nonlinear convective terms that could lead to long-ranged orientational order of the self-propulsion directions are absent in our approach. Therefore the question arises whether a very large crystal will still migrate as a single structure into one direction, or whether it will break up into domains of different migration directions. Instead of the nonlinear convective terms, however, our approach contains a kind of elastic interaction. The phase field crystal free energy functional \(\mathcal{F}_{pfc}\) of Eq. ([\[Fpfc_phi\]](#Fpfc_phi){reference-type="ref" reference="Fpfc_phi"}) acts in the form of a chemical potential in the dynamical equation ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) for the density. It enforces a locally spatially periodic arrangement of the density peaks with a characteristic wave number. Deviations from this order are energetically penalized. It has been demonstrated in particle simulations that solely elastic interactions between self-propelled particles, without any explicit alignment interaction, can induce global collective motion into one common migration direction. There the elastic interactions guided the particles to find a global mode of migration, even under the influence of moderate stochastic noise. In this way, elasticity enforced global collective motion. Due to the finiteness of the system sizes it is difficult, however, to make a rigid statement about what happens on very large length scales on the basis of the particle simulations. Our continuum field approach offers an advantage in this respect. The explicit dynamical equations ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}) contain all length scales. In principle, their investigation is not limited to a certain system size. An arbitrary large system can be analyzed. In reality, however, to get an explicit solution, we have to solve the equations on a finite numerical lattice. Nevertheless, we can extract from our equations the information of whether an arbitrarily large single crystal with a single migration direction is linearly stable. Our procedure will be described in the following. As mentioned before, we confine ourselves for this purpose to the case in which the self-propulsion directions do not spontaneously order, i.e. \(C_1>0\) (\(C_4=0\)). As a first step we linearize the dynamical equations ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}) with respect to a small perturbation \(\delta \psi_1(\mathbf{r},t)\) and \(\delta \mathbf{P}(\mathbf{r},t)\) of the order parameter fields, \[\begin{aligned} \psi_1(\mathbf{r},t) & = & \psi_{10}(\mathbf{r},t) + \delta \psi_1(\mathbf{r},t), \\ \mathbf{P}(\mathbf{r},t) & = & \mathbf{P}_{\!0}(\mathbf{r},t) + \delta \mathbf{P}(\mathbf{r},t). \end{aligned}\] Here, \(\psi_{10}(\mathbf{r},t)\) and \(\mathbf{P}_{\!0}(\mathbf{r},t)\) correspond to the solutions obtained from directly solving numerically the dynamical equations ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}). These solutions describe, on the finite size of the numerical calculation grid, a single crystalline structure as illustrated for example in Fig. [\[snapshots\]](#snapshots){reference-type="ref" reference="snapshots"}. In our case, the converged steady-state solutions \(\psi_{10}(\mathbf{r},t)\) and \(\mathbf{P}_{\!0}(\mathbf{r},t)\), around which we perturb the system, are spatially inhomogeneous. This significantly complicates the analysis as will be shown below. From the linearization we obtain \[\begin{aligned} \partial_{t}\delta\psi_1 &=& \nabla^2\Big\{\!\left[\varepsilon+(1+\nabla^2)^2+3\bar{\psi}^2\right]\delta\psi_1 + 6\bar{\psi}\psi_{10}\delta\psi_1 \nonumber\\ & & {} \qquad+3\psi_{10}^2\delta\psi_1 \Big\}-v_0\,\nabla\cdot\delta\mathbf{P}, \label{eqdeltapsi1}\\[.1cm] \partial_{t}\delta\mathbf{P} &=& \nabla^2 C_1\delta\mathbf{P}-D_r C_1\delta\mathbf{P}-v_0\nabla\delta\psi_1. \label{eqdeltaP} \end{aligned}\] Since the coefficients of the perturbations in the second equation are constant, the system of equations can be significantly simplified in Fourier space. We apply the transformations in the form \[\begin{aligned} \delta\psi_1(\mathbf{q},\omega) & = & \int\! d^2r\, e^{-i\mathbf{q}\cdot\mathbf{r}} \int\! dt\, e^{-i\omega t}\; \delta\psi_1(\mathbf{r},t),\\ \delta\mathbf{P}(\mathbf{q},\omega) & = & \int\! d^2r\, e^{-i\mathbf{q}\cdot\mathbf{r}} \int\! dt\, e^{-i\omega t}\; \delta\mathbf{P}(\mathbf{r},t). \end{aligned}\] Eqs. ([\[eqdeltapsi1\]](#eqdeltapsi1){reference-type="ref" reference="eqdeltapsi1"}) and ([\[eqdeltaP\]](#eqdeltaP){reference-type="ref" reference="eqdeltaP"}) are transformed accordingly. Then the second equation can be solved for \(\delta\mathbf{P}(\mathbf{q},\omega)\), which is inserted into the first equation. In this way, the system of Eqs. ([\[eqdeltapsi1\]](#eqdeltapsi1){reference-type="ref" reference="eqdeltapsi1"}) and ([\[eqdeltaP\]](#eqdeltaP){reference-type="ref" reference="eqdeltaP"}) is reduced to a single equation on the scalar order parameter field \(\delta\psi_1(\mathbf{q},\omega)\), which reads \[\begin{aligned} \lefteqn{i\omega\,\delta\psi_1(\mathbf{q},\omega) =} \nonumber\\ && {}-q^2\Big\{\, \big[\varepsilon+(1-q^2)^2+3\bar{\psi}^2\big]\,\delta\psi_1(\mathbf{q},\omega) \nonumber\\ && {} +\frac{6\bar{\psi}}{(2\pi)^3}\int\!d^2q'\int\!d\omega'\; \psi_{10}(\mathbf{q}-\mathbf{q}',\omega-\omega')\delta\psi_1(\mathbf{q}',\omega') \nonumber\\ && {} +\frac{3}{(2\pi)^3}\int\!d^2q'\int\!d\omega'\; \psi_{10}^2(\mathbf{q}-\mathbf{q}',\omega-\omega')\delta\psi_1(\mathbf{q}',\omega') \,\Big\} \nonumber\\ && {}-\frac{q^2v_0^2}{C_1(D_r+q^2)+i\omega}\,\delta\psi_1(\mathbf{q},\omega). \label{stabaneq} \end{aligned}\] As we can see, the analysis becomes nonlocal in Fourier space. The reason is the spatial modulation of the traveling or resting unperturbed structures, as they are for example depicted in Fig. [\[snapshots\]](#snapshots){reference-type="ref" reference="snapshots"}. In principle, Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) represents an infinite system of equations in which all modes \(\delta\psi_1(\mathbf{q},\omega)\) of different \(\mathbf{q}\) and \(\omega\) are coupled. A similar situation occurs in the study of microphase-separated diblock copolymer systems Ref. . There, however, an equilibrium situation is investigated, which allows a different treatment of the problem. Fortunately the situation can be simplified by exploiting the regular nature of the unperturbed states \(\psi_{10}(\mathbf{r},t)\). First, the structures form a steady state in the sense that they migrate as a single object with a constant migration velocity \(\mathbf{v_m}\). In the co-moving frame, the single crystals are at rest. Second, the spatial modulation of the unperturbed structures is periodic. Using these observations, Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) can be split into discrete sets of equations in Fourier space, each of them still being infinite, however. A numerical evaluation is required to handle the problem. Since the unperturbed structures \(\psi_{10}(\mathbf{r},t)\) form a steady state, we can write them in the form \[\label{steadystate} \psi_{10}(\mathbf{r},t) = \psi_{10}(\mathbf{r}-\mathbf{v_m}t).\] This relation translates itself into Fourier space in the following way: \[\begin{aligned} \label{psi10co} \psi_{10}(\mathbf{q},\omega) &=& 2\pi\delta(\omega+\mathbf{q}\cdot\mathbf{v_m})\psi_{10}(\mathbf{q}), \\ \label{psi10sqco} \psi_{10}^2(\mathbf{q},\omega) &=& 2\pi\delta(\omega+\mathbf{q}\cdot\mathbf{v_m})\psi_{10}^2(\mathbf{q}). \end{aligned}\] Introducing these expressions into Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) decouples the modes in the frequencies \(\omega\), which significantly simplifies the problem. For the second simplification, we use the fact that the unperturbed structures are spatially periodic. They are resting or traveling crystals. In Fourier space they are therefore characterized by a reciprocal lattice of a discrete set of wave vectors. Translating in reciprocal space the whole structure by a reciprocal lattice vector maps the positions of the reciprocal lattice points onto themselves. In this sense, the reciprocal lattice is closed by itself under such translations. The subtractions \(\mathbf{q}-\mathbf{q}'\) in Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) correspond to these translations. As a consequence, the set of equations ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) splits into different decoupled discrete subsets of equations. Within one such set, all modes are connected via addition or subtraction of the reciprocal lattice vectors of the unperturbed structure. Still each set generally contains infinitely many modes, but the discretization already implies a major reduction. In the following we investigate three different structures: a traveling hexagonal crystal, a traveling rhombic crystal, and a traveling quadratic crystal. As mentioned above, the textures are obtained in this sequence when increasing the active drive \(v_0\). On the left-hand side, Fig. [\[FTstructure\]](#FTstructure){reference-type="ref" reference="FTstructure"} shows snapshots of the steady-state structures in real space, i.e. \(\psi_{10}(\mathbf{r})\). On the right-hand side, the corresponding power spectra \(|\psi_{10}(\mathbf{q})|\) in two-dimensional Fourier space are included. The different textures--hexagonal, rhombic, and quadratic--can be identified by eye from the real-space plots. Also the spatial arrangements of the inner intensity peaks on the right-hand side of Fig. [\[FTstructure\]](#FTstructure){reference-type="ref" reference="FTstructure"} reflect these symmetries. However, when we take into account the magnitude of the intensities, we see that the collective migration breaks the symmetry. Most obviously, there is no four-fold symmetry in the intensity magnitudes of Fig. [\[FTstructure\]](#FTstructure){reference-type="ref" reference="FTstructure"} (f) for the quadratic texture. Likewise, the single real-space density peaks in Fig. [\[FTstructure\]](#FTstructure){reference-type="ref" reference="FTstructure"} (c) appear a little distorted. The direction of collective migration marks the orientation of the remaining symmetry axis. We can see that the magnitude of the peaks in the power spectrum, coded by their brightness in Fig. [\[FTstructure\]](#FTstructure){reference-type="ref" reference="FTstructure"}, significantly decays with increasing wave number. As a consequence we introduce a positive threshold value \(\psi_{th}\). \(\psi_{10}(\mathbf{r})\) is approximated by only those modes that satisfy \(|\psi_{10}(\mathbf{q})|>\psi_{th}\). The same procedure is applied to \(\psi_{10}^2(\mathbf{q})\), i.e. the Fourier transform of \(\psi_{10}^2(\mathbf{r})\). For each investigated structure, the above procedure must be performed only once in practice. We numerically iterate Eqs. ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}) forward in time starting from random initial conditions. When the system has reached the steady state, \(\psi_{1}\) is used as an input for the unperturbed states \(\psi_{10}(\mathbf{r})\) and \(\psi_{10}^2(\mathbf{r})\). After performing the Fourier transform, the values \(\psi_{10}(\mathbf{q})\) and \(\psi_{10}^2(\mathbf{q})\) that satisfy the above threshold criterion together with their locations \(\mathbf{q}\) in the reciprocal space are stored in an ordered list structure for the further analysis. We also numerically measure \(\mathbf{v_m}\) during this initial numerical iteration and use this value in our further evaluation. (It is a challenging task to calculate the collective migration velocity \(\mathbf{v_m}\) analytically. A corresponding attempt in one spatial dimension that is restricted to the threshold vicinity is included in the appendix.) Our intention is to perform a linear stability analysis of the traveling single crystal at long wave lengths, i.e. at small wave numbers \(\|\mathbf{q}\|\). Direct numerical solution of Eqs. ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}) showed that the structures are stable on the length scale of the size of the numerical box. But the question of what happens on larger length scales has so far remained unanswered. In other words, we are interested in the question whether the single crystal will break up into a multidomain texture for large system sizes. Below, we denote the reciprocal lattice vectors of the unperturbed structure as \(\mathbf{G}_i\), with \(i\) a labeling index. They localize the positions of the spectral points in the right column of Fig. [\[FTstructure\]](#FTstructure){reference-type="ref" reference="FTstructure"}. The reciprocal lattice vector of smallest magnitude is called \(\mathbf{G}_1\). We probe a square region in reciprocal space bordered at two opposing edges by \(-\mathbf{G}_1\) and \(+\mathbf{G}_1\). This region covers the first Brillouin zone. All wave vectors \(\mathbf{q}\) lying in the first Brillouin zone belong to a different discrete coupled subset of equations spanned by the nonlocal terms in Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}). We cannot connect two different wave vectors in this region by adding or subtracting reciprocal lattice vectors. Still, these subsets of equations are infinite. We have seen, however, that the coupling strengths \(|\psi_{10}(\mathbf{q}-\mathbf{q'})|\) and \(|\psi_{10}^2(\mathbf{q}-\mathbf{q'})|\) that are responsible for the nonlocal coupling in Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) decay with increasing distance \(\|\mathbf{q}-\mathbf{q'}\|\). Therefore we introduce a cutoff distance \(q_{max}>0\) and only include modes that satisfy \(\|\mathbf{q}-\mathbf{q'}\|<q_{max}\) in the subset of equations for each wave number \(\mathbf{q}\). In this way, the discrete subsets of equations become finite and can be handled numerically. In practice, we sample the square region between \(-\mathbf{G}_1\) and \(+\mathbf{G}_1\) in small wave vector steps. At each \(\mathbf{q}\) in this region, we build up the discrete set of equations described above. For this purpose, we numerically search for all modes \(\delta\psi_1(\mathbf{q}',\omega')\) with wave vectors \(\mathbf{q}'\) that can be obtained by iteratively adding or subtracting the reciprocal lattice vectors that predominantly contribute to \(\psi_{10}\) and \(\psi_{10}^2\) (those were previously stored in a list when calculating the spectra, see above). Only modes \(\delta\psi_1(\mathbf{q}',\omega')\) that satisfy the condition \(\|\mathbf{q}-\mathbf{q'}\|<q_{max}\) are included. Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) leads for each wave vector \(\mathbf{q}\) and frequency \(\omega\) to a linear system of equations for all modes \(\delta\psi_1(\mathbf{q}',\omega')\) that are coupled to \(\delta\psi_1(\mathbf{q},\omega)\). In practice, our cut-off values are chosen such that the coupling to several hundred modes is included for each wave vector \(\mathbf{q}\). The coefficients of this system of equations are calculated and stored in matrix form. For practical reasons, to obtain a matrix of real coefficients, we consider the real and imaginary parts of the perturbations \(\delta\psi_1(\mathbf{q}',\omega')=\Re\{\delta\psi_1(\mathbf{q}',\omega')\}+i\Im\{\delta\psi_1(\mathbf{q}',\omega')\}\) as independent variables. Each equation in the system of equations ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) is split into its real and imaginary part, which doubles the number of equations. Finally, the determinant \(\mathcal{D}(\mathbf{q},\omega)\) of the resulting matrix is calculated by standard numerical procedures. At the end of this procedure, a vanishing determinant \(\mathcal{D}(\mathbf{q},\omega)\) signals a linear instability of the single crystal with respect to the perturbation \(\delta\psi_1(\mathbf{q},\omega)\). Before continuing with the results, we add a technical remark to explain the appearance of the plots included below. Due to the many modes that are coupled in each case, the number of multiplications during the process of obtaining the determinant \(\mathcal{D}(\mathbf{q},\omega)\) is very large. In fact, \(\mathcal{D}(\mathbf{q},\omega)\) often exceeds the size that can be processed on the computer. To circumvent this problem, we rescale the factors in the final multiplication by an adjusted constant positive number. Since we are only interested in the question whether \(\mathcal{D}(\mathbf{q},\omega)\) vanishes (or changes sign) this is a legal procedure. Nevertheless, the absolute values of \(\mathcal{D}(\mathbf{q},\omega)\) should not be interpreted any more. Furthermore, we calculate and plot the logarithm of the determinant \(\ln[\mathcal{D}(\mathbf{q},\omega)]\). In a first step, we consider static instabilities in the co-moving frame. That is, we set \(\omega=-\mathbf{q}\cdot\mathbf{v_m}\) (and likewise \(\omega'=-\mathbf{q}'\cdot\mathbf{v_m}\)) in Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}). Thus \(\omega\) does not describe an additional degree of freedom but is determined by the wave vector \(\mathbf{q}\). The results corresponding to the three textures in Fig. [\[FTstructure\]](#FTstructure){reference-type="ref" reference="FTstructure"} are depicted in Fig. [\[logdeterminantco\]](#logdeterminantco){reference-type="ref" reference="logdeterminantco"}. The determinant naturally vanishes at \(\mathbf{q}=\mathbf{0}\) and at \(\mathbf{q}=\pm\mathbf{G}_1\), which cannot be shown in the logarithmic plot and leads to the white wholes in the projected shadow plots below each surface plot. We do not find any other position in the first Brillouin zone where the determinant vanishes. The small dips visible in the surfaces of \(\ln[\mathcal{D}(\mathbf{q})]\) were found to decrease with increasing size of the numerical calculation box that was used to obtain \(\psi_{10}\). They were not observed to indicate a linear instability. We thus conclude that the investigated structures are linearly stable against static instabilities in the co-moving frame. After excluding static linear instabilities in the co-moving frame, we ask for dynamical linear instabilities. For example, on larger length scales, there might appear domains migrating in different directions. This question adds a further degree of freedom to our analysis, namely the frequency \(\omega\). Previously, it was fixed by the relation \(\omega=-\mathbf{q}\cdot\mathbf{v_m}\) in the co-moving frame. Now we can in principle choose any value independently of the selected wave vector \(\mathbf{q}\). Fortunately, for each perturbation \(\delta\psi_1(\mathbf{q},\omega)\), the coupled system of equations has the same size as above in the analysis within the co-moving frame. The additional degree of freedom \(\omega\) does not lead to further couplings. We can understand this result from Eqs. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}), ([\[psi10co\]](#psi10co){reference-type="ref" reference="psi10co"}), and ([\[psi10sqco\]](#psi10sqco){reference-type="ref" reference="psi10sqco"}). On the one hand, Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) implies that the modes \(\delta\psi_1\) are coupled via the unperturbed steady-state fields \(\psi_{10}(\mathbf{q},\omega)\) and \(\psi_{10}^2(\mathbf{q},\omega)\). On the other hand, due to the steady-state form of \(\psi_{10}(\mathbf{r},t)\) and \(\psi_{10}^2(\mathbf{r},t)\), Eqs. ([\[psi10co\]](#psi10co){reference-type="ref" reference="psi10co"}) and ([\[psi10sqco\]](#psi10sqco){reference-type="ref" reference="psi10sqco"}) contain a rigid relation between \(\mathbf{q}\) and \(\omega\) in the \(\delta\)-functions. In this way, each coupling wave vector \(\mathbf{q}-\mathbf{q'}\) in Eq. ([\[stabaneq\]](#stabaneq){reference-type="ref" reference="stabaneq"}) uniquely fixes one coupling frequency \(\omega-\omega'\). The complete \(\mathbf{q}\)-\(\omega\) space is too large to be systematically tested. We therefore particularly focused on the representative directions given by the reciprocal wave vectors of each structure, and the directions perpendicular to these. In neither case did we find a vanishing determinant that would indicate a dynamical linear instability. This provides evidence that the investigated traveling active single crystals are also dynamically linearly stable. A few examples of our results are depicted in Fig. [\[logdeterminantdynamic\]](#logdeterminantdynamic){reference-type="ref" reference="logdeterminantdynamic"}. One might argue that there are sources of uncertainty in our analysis. For example, we obtain our initial reciprocal lattice vectors from a numerical calculation on a grid of finite size. However, if this approximation were problematic, it would rather show up as an instability of the unperturbed state. In contrast to that, we here find that the system is linearly stable despite such approximations. We are therefore confident that our analysis reflects the real behavior of the structures, and that the active single crystals are linearly stable against perturbations. # Destabilization via hydrodynamic interactions In the previous section, we have seen that the active single crystals are linearly stable even at very large system sizes. The underlying stabilization mechanism is provided by the effective elastic interactions between the density peaks. It is introduced by the phase field crystal free energy functional that acts like a chemical potential for the density peaks in our dynamical equations ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}). This picture is further supported by an analysis of related particle simulations. On the contrary the question arises, whether in real experimental systems there might be a counteracting mechanism that could destroy this long-ranged order. A natural candidate is given by hydrodynamic interactions. It turns out that hydrodynamic interactions can destabilize the system already at relatively small system sizes. This can be studied by direct numerical simulation of our dynamical equations, so that a complicated analysis as in the previous section is not necessary. In the situation that we investigate, the self-propelled particles can still directly exchange momentum with the ground. This requirement is satisfied by particles that directly migrate on a substrate, or approximately by swimmers that propel in a low Reynolds number environment close to a surface with no-slip boundary conditions. However, we assume that the density peaks and lamellae can additionally interact with each other through a surrounding background fluid film. Our goal in the following is a simple qualitative estimate of the effect of hydrodynamic interactions. We assume that the particles and the fluid have the same mass density, and that the system is incompressible. In this way the dynamical equation of mass continuity is trivially satisfied and does not need to be considered explicitly in the following. The location of the particles is described by the concentration or density field \(\psi_1(\mathbf{r})\). We introduce a velocity field \(\mathbf{u}(\mathbf{r})\) to parameterize the fluid flow that is induced by the particle motion. On the one hand, additional convective terms appear in the dynamical equations ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}) that now read \[\begin{aligned} \hspace{-.7cm}\partial_{t}\psi_1 &=& \nabla^2\frac{\delta\mathcal{F}}{\delta\psi_1}-v_0\,\nabla\cdot\mathbf{P}-\mathbf{u}\cdot\nabla\psi_1, \label{eqpsi1hydro}\\[.1cm] \hspace{-.7cm}\partial_{t}\mathbf{P} &=& \nabla^2\frac{\delta\mathcal{F}}{\delta\mathbf{P}}-D_r\frac{\delta\mathcal{F}}{\delta\mathbf{P}}-v_0\nabla\psi_1-\mathbf{u}\cdot\nabla\mathbf{P} +\mathbf{\Omega}\cdot\mathbf{P}. \label{eqPhydro} \end{aligned}\] Here the tensor \(\mathbf{\Omega}\) with components \(\Omega_{ij}=(\nabla_iu_j-\nabla_ju_i)/2\) describes rotations due to convection. We do not consider flow alignment of the polarization field in this qualitative picture. On the other hand, the additional dynamical equation for the velocity field \(\mathbf{u}(\mathbf{r})\) in the rescaled form reads \[\partial_t\mathbf{u} =-\mathbf{u}\cdot\nabla\mathbf{u} + \mathbf{F} + \nabla\cdot\nu\{\psi_1\}\nabla\mathbf{u}-\alpha\mathbf{u}. \label{equ}\] It is supplemented by the incompressibility condition \(\nabla\cdot\mathbf{u}=0\), so that an explicit pressure term is not necessary. We can neglect the convective term on the right-hand side in the low Reynolds number regime. When the self-propelled particles migrate on the substrate, they push the surrounding fluid and set it into motion. This effect is included by the force term \(\mathbf{F}\) that we introduce as \[\mathbf{F} = \gamma \sum_{i=1}^{N_p} (\mathbf{v}_i-\mathbf{u})\,\delta(\mathbf{r}-\mathbf{R}_i)\] with a sufficiently large coefficient \(\gamma>0\). Again \(N_p\) is the number of density peaks in the sample. \(\mathbf{R}_i\) and \(\mathbf{v}_i\) denote, respectively, the current positions and velocities of the density peaks for \(i=1,\dots,N_p\). In practice we track the positions and motion of the centers of the density peaks as described above. At each peak center \(\mathbf{R}_i\), the fluid flow velocity \(\mathbf{u}\) is then adjusted to the peak velocity \(\mathbf{v}_i\). To mimic the presence of the particles in the fluid, the viscosity \(\nu\) varies with the density \(\psi_1\). We set \(\nu=10\) at positions of lowest density \(\psi_1\) and let it linearly increase by a factor of \(2\)--\(3\) when moving to the regions of high density \(\psi_1\). Due to the smooth density profiles in our numerical samples the effect of the contribution \((\nabla\nu\{\psi_1\})\cdot\nabla\mathbf{u}\) is mostly negligible. Overall, the volume elements of high density \(\psi_1\), which represent the presence of the self-propelled particles, are considered as part of the fluid flow. This approach is similar to the "fluid particle dynamics" introduced by Tanaka and Araki. The last term \(-\alpha\mathbf{u}\) in Eq. ([\[equ\]](#equ){reference-type="ref" reference="equ"}) represents a simple way to model the friction of the thin fluid film with its environment, with \(\alpha>0\) the friction constant. This term was used previously to investigate the complex flow behavior in quasi-two-dimensional systems of fluid films. Two limits for the magnitude of \(\alpha\) are obvious. On the one hand, if \(\alpha\) is small, the friction between the fluid film and its surfaces is low and the migrating particles can easily set the surrounding fluid into motion. On the other hand, if \(\alpha\) is very large, we see from Eq. ([\[equ\]](#equ){reference-type="ref" reference="equ"}) that the fluid remains practically at rest with \(\mathbf{u}(\mathbf{r})\approx\mathbf{0}\). Then the convective terms in Eqs. ([\[eqpsi1hydro\]](#eqpsi1hydro){reference-type="ref" reference="eqpsi1hydro"}) and ([\[eqPhydro\]](#eqPhydro){reference-type="ref" reference="eqPhydro"}) are negligible and we reobtain the previous Eqs. ([\[eqpsi1\]](#eqpsi1){reference-type="ref" reference="eqpsi1"}) and ([\[eqP\]](#eqP){reference-type="ref" reference="eqP"}). Thus hydrodynamic interactions are not important in the regime of large \(\alpha\). We illustrate the flow behavior at different values of the friction parameter \(\alpha\) in Fig. [\[fluidflow\]](#fluidflow){reference-type="ref" reference="fluidflow"}. For high \(\alpha\), corresponding to strong fluid friction, most parts of the fluid film cannot be set into motion as illustrated in panels (a) and (d). The fluid flow remains very localized around the density peaks and does not extend between different peaks. Consequently the hydrodynamic interactions between the density peaks are low. Around the density peaks some "backflow" of the fluid oppositely to the peak migration directions occurs. This is stressed in a rescaled version of panel (d), included as panel (e); a "zoom" into the region around one single density peak is provided for illustration. With decreasing fluid friction \(\alpha\), the "backflow" vanishes and the fluid flow more and more extends between the density peaks, as depicted in panels (b) and (f). At very low friction parameters \(\alpha\), the whole fluid can be set into motion nearly homogeneously, as shown in panels (c) and (g). Since the fluid flow here extends between the density peaks, hydrodynamic interactions are strong in this case. As for Fig. [\[vmvsv0\]](#vmvsv0){reference-type="ref" reference="vmvsv0"}, we measured again the collective migration speed \(v_m\) in the form of the sample-averaged peak velocity magnitude, \(v_m=\sum_{i=1}^{N_p} \|\mathbf{v}_i\|/{N_p}\), with \(N_p\) the number of the density peaks and \(\mathbf{v}_i\) the velocity of the single density peaks, \(i=1,\dots,N_p\). A result of \(v_m\) as a function of the active drive \(v_0\) of the individual self-propelled particles is plotted in Fig. [\[alphadep\]](#alphadep){reference-type="ref" reference="alphadep"} for two different values of the fluid friction parameter \(\alpha\). In this case \(C_1>0\), i.e. there is no alignment mechanism that would lead to a spontaneous orientational ordering of the self-propulsion directions. As a first result, we found that, within our numerical resolution, the critical self-propulsion velocity \(v_{0,c}\), at which collective motion of the active crystal sets in, remains unchanged by the hydrodynamic interactions. Second, the hydrodynamic interactions speed up the collective motion of the active single crystal. We can see this from the two data curves plotted in Fig. [\[alphadep\]](#alphadep){reference-type="ref" reference="alphadep"}. At high values of \(\alpha\) the fluid friction is large, and the self-propelling particles cannot set the surrounding fluid into motion. However, at low fluid friction for low values of \(\alpha\), the particles can set the whole surrounding fluid film into motion, as was also demonstrated in Fig. [\[fluidflow\]](#fluidflow){reference-type="ref" reference="fluidflow"}. In this way, one density peak can push the preceding peak via the fluid between them. The density peaks support each other in migration via hydrodynamic interactions. At fixed active drive \(v_0\) of the individual self-propelled particles, the collective migration speed \(v_m\) increases with decreasing fluid friction \(\alpha\) and increasing hydrodynamic interactions. However, the hydrodynamic interactions disturb the order of the active single crystals. This is illustrated in Fig. [\[destabilizedcrystal\]](#destabilizedcrystal){reference-type="ref" reference="destabilizedcrystal"}. The figure includes two identical systems that only differ by the fluid friction parameter \(\alpha\). At high \(\alpha\) and therefore little hydrodynamic interaction (a), the sample shows a perfect active single crystal that migrates collectively into one direction. On the contrary, the crystal is broken up into different domains at low \(\alpha\) corresponding to strong hydrodynamic interactions (b). Also a partial transition to lamellar textures is induced by the hydrodynamic interactions. In the case without hydrodynamic interactions this transition would typically occur only at higher magnitudes of the active drive \(v_0\). In summary, we can say that--at least within the qualitative picture presented--hydrodynamic interactions have a destabilizing effect on the active crystalline structures considered here. # Conclusions In this paper we focused on the stability of traveling active single crystals. An active phase field crystal model was introduced for this purpose. We outlined its derivation via dynamical density functional theory from a microscopic consideration of self-propelled particles that feature an active drive. Besides resting hexagonal crystals, traveling hexagonal, rhombic, and quadratic crystals, as well as traveling lamellar textures were obtained from this description. Also the inverted structures like resting and traveling honeycomb textures, as well as inverted traveling rhombic and quadratic lattices were observed. A linear stability analysis did not indicate a linear instability of the collectively migrating single crystals. In particular, long-wave-length instabilities beyond the size of the numerical calculation grids did not show up, indicating linear stability also for large system sizes. We found, however, that hydrodynamic interactions can destabilize the single-crystalline textures. The linear stability analysis was significantly complicated by the spatial modulation of the unperturbed steady state. As a consequence of this spatial modulation, our equations became nonlocal in Fourier space. We could reduce the problem to manageable size by exploiting the periodicity of the unperturbed state and its steady-state migration. This made the linear stability analysis possible. The problem of a spatial modulation of the unperturbed state occurred already in a previous equilibrium stability study of micro-phase separated diblock copolymer melts. There, however, the equilibrium conditions allowed a different strategy to perform the analysis. Our results on the stability of the traveling crystals are important when we think of the design of new active materials. It is expected that, for example, features of traveling crystals like phononic properties are different from their passive counterparts. We have seen that the structure of the migrating crystals, i.e. their hexagonal, rhombic, or quadratic form, can be tuned by the activity of the constituent self-propelled particles. It will be interesting to test whether other features of active crystals, such as their phononic behavior, can be similarly tuned by the properties of the individual active components. We shall investigate this interesting question in the future.

{'timestamp': '2014-01-22T02:09:56', 'yymm': '1401', 'arxiv_id': '1401.5332', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5332'}

# Introduction {#sec:1} The gauge/gravity duality has been proven to be a very powerful tool in studying strongly coupled phenomena using dual gravitational systems where the coupling is weak. This duality which is well founded in string theory has many interesting applications and among them one which from the first sight it seems unexpected, in condensed matter physics (for a review, see ). A condensed matter system that is well studied using the gauge/gravity duality is the holographic superconductor. The simplest holographic superconductor model which is extensively studied is described by an Einstein-Maxwell-scalar field theory with a negative cosmological constant. Its gravity sector is described by an Abelian-Higgs model with a stationary black hole metric which below a certain critical temperature the black hole acquires scalar hair. Its dual boundary field theory is described by a theory which is similar to the standard Landau-Ginzburg theory in which the scalar field corresponds to an operator, the order parameter, which condenses below the critical temperature, signalizing the onset of superconductivity. This phenomenological approach has the virtue of simplicity, but does not capture all the underlying features of the gauge/gravity duality including quantum effects. This is hard to implement, as it remains a challenge to embed this model in a quantum system (string/M-theory). In spite of that, this holographic principle has been applied to many other condensed matter systems like the conventional and unconventional superfluids and superconductors, Fermi liquid behavior, non-linear hydrodynamics, quantum phase transitions and transport. In all these studies it is crucial to understand the low temperature limit. In the gravity sector this requires to go beyond the probe limit and to consider fully back-reacted gravitational systems. In the boundary theory the \(T\rightarrow 0\) limit leads to strongly coupled systems in the quantum physics regime where new superfluid phenomena arise like for example the generation of inhomogeneous FFLO phases. Recently it was proposed that this low temperature regime can be probed by the entanglement entropy. The entanglement entropy can be considered as a measure of how a given quantum system is strongly correlated (entangled). It was introduced as a tool to describe different phases and their corresponding phase transitions of a quantum system as the temperature goes to zero. The entanglement entropy is directly related to the degrees of freedom of a system, keeping track of them. It can also play the role of an order parameter of a phase transition at very low temperatures. The most important property of the entanglement entropy is that it is non-vanishing at zero temperature. For these reasons it was employed as a probe of quantum properties of the ground state for a given quantum system. Recently the entanglement entropy was used to study various properties of holographic superconductors at low temperatures. In a model coming from \(\mathcal{N}=8\) gauged supergravity the entanglement entropy across the superconducting phase transition was studied. It was found that the entanglement entropy is lower in the superconducting phase than in the normal phase. This behavior was attributed to some kind of reorganization of the degrees of freedom of the system: because electrons are bounded in the superconducting phase to form Cooper pairs less degrees of freedom remain comparable to the normal phase. The behavior of entanglement entropy across the holographic p-wave superconductor phase transition in an Einstein-Yang-Mills theory with a negative cosmological constant was studied in . In a holographic p-wave superconductor/insulator model was considered and it was found that as the back reaction increases, the transition is changed from second order to first order. Also in the gravitational backreaction of the non-Abelian gauge field on the gravity dual to a 2+1 \(p\)-wave superconductor was studied. It was found that the \(p\)-wave superconductor has lower entanglement entropy. The behavior of holographic entanglement entropy for imbalanced holographic superconductor was considered in. It was found that entanglement entropy for this imbalanced system decreases with the increase of imbalance in chemical potentials. In this work we will use the entanglement entropy as a probe of the proximity effect in a holographic superconductor. In condensed matter physics the proximity effect describes the dynamics of a system near the superconductor-normal metal interface where the superconducting electrons (Cooper pairs) may penetrate from the superconducting to normal phase. The leakage of the Cooper pairs weakens the superconductivity near the interface with a normal metal. This phenomenon can appear even in the absence of a magnetic field. One of our motivations in this work is to construct a computationally tractable gravity model and show that it can reproduce basic properties of proximity effect in superconductivity. The phenomenological analysis of the proximity effect was given by the generalized Ginzburg-Landau theory. In this theory a complex scalar field \(\Psi\) is considered as the superconducting order parameter which contributes to the free energy in the functional F_G=a(T)\|\|\^2+\|\|\^4+(T)\|\|\^2 , where the coefficient \(a\) vanishes at the transition temperature \(T_c\). At \(T<T_c\), the coefficient \(a\) is negative and the minimum of \(F_G\) occurs for a uniform superconducting state with \(|\Psi|^2=-a/b\). The coefficient \(a\) is given by \(a=\alpha(T-T_c)\), where \(T_c\) is the critical temperature of the transition into the uniform superconducting state. The appearance of the proximity effect can simply be interpreted as the effect of the gradient term in the Ginzburg-Landau functional. Let us consider the decay of the order parameter in the normal phase, i.e., at \(T>T_c\) assuming that our system is in contact with another superconductor with a higher critical temperature, and the \(x\) axis is chosen perpendicular to the interface between the superconductor and the normal phases. The induced superconductivity is weak and, we use the linearized Ginzburg-Landau equation for the order parameter, \(a\Psi-\gamma\frac{\partial^2\Psi}{\partial x^2}=0\), with \(\gamma \neq 0\). The decaying solution is \(\Psi=\Psi_0\exp[-x/ where\) \(\gamma>0\), the order parameter decays exponentially as \(\Psi(\gamma>0)=\Psi_0\exp[-x/ to the interface, the order parameter\)(\<0)\~\_0(x/)\(, which is bigger than\)(\>0)\(in the normal phase close to the interface of the superconductor. This shows that due to the gradient term, the superconducting properties can be induced in the normal phase. This phenomenon is called the proximity effect. Simultaneously the leakage of the Cooper pairs weakens the superconductivity in the superconductor phase near the interface with a normal metal, which results in a decrease of the superconducting transition temperature in a thin superconducting layer in contact with a normal metal. Using the generalized Ginzburg-Landau functional, we see that the gradient term plays an important role in inducing the proximity effect. Our aim here is to build a holographic superconductor with a dual boundary field described by a theory similar to the generalized Ginzburg-Landau theory. Such a theory, in the probe limit, was discussed in in which a\)U(1)\(gauge field was introduced along with a complex scalar field coupled to a charged AdS black hole and a higher-derivative coupling between the\)U(1)\(gauge field and the scalar with coupling constant\[. This coupling is provided by a potential term which in a covariant form reads\)V()=m\^2\|\|\^2+\|F\^D\_\|\^2,[\[holpot\]]{#holpot label="holpot"} \(where\)F=dA\(is the strength of a U(1) gauge field\)A\_\(,\]is a charged complex scalar field of charge\)q\(and mass\)m\(and\)D\_=\_-iqA\_\(. However, this covariance is broken on the boundary, because the coefficient\[plays the role of\]of the boundary Ginzburg-Landau theory and both of them indicate the strength of the gradient derivatives of the scalar field. What we had found in is that, in the probe limit, large positive values of the coupling\[make easier the transition to a superconducting phase, while when\]becomes negative, it is more difficult for the condensation to be formed and this happens because the energy gap in the probe limit is larger for\)\<0\(than the energy gap in the conventional case (\[). One can expect that such high-derivatives terms can arise in string theory. Indeed, there are models based on exact solutions of\)D=11\(and type IIB supergravity in which after consistent Kaluza-Klein truncations to four spacetime dimensions, fully back-reacted solutions describing holographic superconductors in three spacetime dimensions have been found. These models, contain a large number of scalar, gauge fields and high derivatives of them, which need to be constrained in order to make the models tractable. It is of great interest to generalize our previous study to the fully back-reacted theory. More interestingly, we would like to construct a tractable gravity model to reproduce the proximity effect in the holographic superconductor. Our\], which is analogous to\[in the general Ginzburg-Landau functional, gives us the hope to build a holographic description of the proximity effect in the gravity model. Using the fact that the entanglement entropy plays the role of an order parameter, we will calculate it near the interface of superconducting/normal phase of a holographic superconductor in low temperatures and we will show that it gives us important information on the behavior of system. To verify this behavior we will also calculate the conductivity near the boundary interface. The work is organized as follows. In Section~\ref{sec:2} we present the gravity sector of the model. In Section~\ref{sec:3} we present the fully back-reacted solution of the holographic system. In Section~\ref{sec:4} we discuss the entanglement entropy. In Section~\ref{sec:5} we calculate the conductivity and finally in Section~\ref{sec:6} are our conclusions. \section{The gravitational sector} \label{sec:2} We will consider a scalar field coupled to a\)U(1)\(gauge field with an action discussed in \begin{equation}\label{action} S=\int d^4x \sqrt{-g}\Big[\frac{R+6/L^2}{16\pi G}-\frac{1}{4}F_{\mu\nu} F^{\mu\nu} -|D_{\mu}\Psi|^2-V(\Psi)\Big]~, \end{equation} where the potential term is given by the expression above. The field equations are given by: \begin{itemize} \item the Einstein equations \begin{eqnarray}\label{0eqEinstein} R_{\mu\nu}-\frac{1}{2}Rg_{\mu\nu}-\frac{3}{L^2}g_{\mu\nu}=8\pi GT_{\mu\nu}~, \end{eqnarray} \item the Maxwell equations \begin{eqnarray} \label{0eqMaxwell} &\nabla_{\mu}&F^{\mu\nu}+\frac{\eta}{\sqrt{-g}}\partial_\mu\Big[\sqrt{-g}(D_\kappa \Psi)(D_\lambda \Psi)^*\Big(g^{\kappa\nu}F^{\mu\lambda}-g^{\kappa\mu}F^{\nu\lambda}+g^{\nu\lambda}F^{\mu\kappa}- g^{\mu\lambda}F^{\nu\kappa}\Big)\Big]\nonumber\\ &=&iq\Big[\Psi^* (D^\nu \Psi)-\Psi(D^\nu \Psi)^*\Big]+iq\eta g_{\mu\rho}F^{\rho\nu}\Big[F^{\mu\kappa}\Psi^* (D_\kappa\Psi)-F^{\mu\lambda}\Psi (D_\lambda\Psi)^*\Big], \end{eqnarray} \item and the scalar field equation \begin{eqnarray} \label{0eqPsi} &-&\frac{1}{\sqrt{-g}}\partial_\mu \bigl[\sqrt{-g}g^{\mu\nu}\bigl(\partial_\nu\Psi-iqA_\nu \Psi\bigr)\bigr]+iqg^{\mu\nu}A_\nu\Big(\partial_\mu\Psi-iqA_\mu\Psi\Big)+ m^2\Psi\nonumber \\&=&\frac{\eta}{\sqrt{-g}}\partial_\mu \Big[\sqrt{-g} g_{\kappa\lambda}F^{\kappa\nu}F^{\lambda\mu}\Big(\partial_\nu\Psi-iqA_\nu \Psi\Big)\Big]-iq\eta g_{\kappa\lambda}F^{\kappa\nu}F^{\lambda\mu}A_\mu \Big(\partial_\nu\Psi-iqA_\nu \Psi\Big)~. \end{eqnarray} \end{itemize} In this work, we will set\)L=1\(,\)`<!-- -->`{=html}8G=1\(,\)q=1\(. We note that the presence of the coupling constant\]adds new terms in the field equations which makes the system of the differential equations highly non-trivial. We have to find numerical solutions of the fully back-reacted system. \section{The solutions of the holographic system} \label{sec:3} We will generalize the probe limit discussion in to a full back-reacted formalism by taking the ansatz of metric and matter fields as \begin{equation}\label{metric1} d s^2 =-\frac{1}{z^2} f(z) e^{-\chi(z) } dt^2 + \frac{1}{z^2} \Big( dx^2 + dy^2 \Big) + \frac{1}{z^2} \frac{d z^2}{f(z)} \ , ~~~A_{\mu} = A_t(z)dt \, ~~~\Psi = \psi(z) \. \end{equation} The temperature can be expressed as \begin{equation}\label{temperature} T=\frac{-e^{\chi/2}\partial_zf}{4\pi}|_{z=z_h}~. \end{equation} Then, the independent Einstein-Maxwell equations reduced from (\ref{0eqEinstein}) and (\ref{0eqMaxwell}) become \begin{eqnarray} \label{EE1} 0&=&\chi'-\frac{z \Big(1+2 e^{\chi } z^4 \eta \text{At}'^2\Big) \Big(e^{\chi }q^2 \text{At}^2 \psi ^2+f^2 \psi '^2\Big)}{f^2}~,\\ \label{EE2} 0&=&f'+\frac{3(1-f)}{z}-\frac{m^2\psi^2}{2z}-\frac{e^{\chi}z^3A_t^{'2}}{4}-\frac{e^{\chi}zq^2A_t^{2} (2+5e^{\chi}z^4\eta A_t^{'2})\psi^2}{4f}-\frac{zf(2+3e^{\chi}\eta z^4A_t^{'2})}{4}\psi^{'2}~,\\ \label{1eqMax} 0&=&A_t''+\Big[\frac{z^2f^2\chi'+4e^{\chi}\eta z^3\psi q^2A_t^2(z\psi f'-2f(\psi+z\psi'+\frac{3}{4}z\psi\chi'))} {2 a_0 f}\no\\&+&\frac{2\eta z^3f\psi'(z f'\psi'+2f(\psi'(1+\frac{1}{4}z\chi')+z\psi''))}{a_0}\Big]A_t'-\Big[\frac{2\psi^2q^2(1+e^{\chi}\eta z^4A_t'^2)}{a_0}\Big]A_t~, \end{eqnarray} while the scalar field equation (\ref{0eqPsi}) takes the form \begin{eqnarray} \label{1eqpsi} 0&=&\psi''+\Big[\frac{f'}{f}+\frac{(4+z\chi')(1+e^{\chi}z^4\eta A_t^{'2})+4z^5e^{\chi}\eta A_t^{'}A_t^{''}}{2z(e^{\chi} z^4\eta A_t^{'2}-1)}\Big]\psi'+\Big[\frac{e^{\chi}q^2A_t^2}{f^2}+\frac{m^2}{z^2f(e^{\chi} z^4\eta A_t^{'2}-1)}\Big]\psi~, \end{eqnarray} with\)a_0=z\^2f(1+2z\^2f'\^2)-2e\^z\^4\^2q\^2A_t\^2\(. Here (\ref{EE1}) and (\ref{EE2}) are the combination of\)tt\(and\)zz\(component of Einstein's equation and the xx component can be obtained from differentiating the two Einstein equations above. Near the boundary\)z\(, we need\)(z=0)\(=0 to recover the pure AdS boundary. The matter fields should behave as \begin{eqnarray}\label{asypsiAt1} \psi&=&\psi^{(1)}z^{\Delta_{-}}+\psi^{(2)}z^{\Delta_{+}}+\cdots,\\ A_t&=&\mu-\rho z+\cdots, \end{eqnarray} where according to the AdS/CFT dictionary,\)\^(i)=O\_i/, i=1,2\(and\)O\_i\(with the conformal dimensions\)\_=\(are the corresponding dual operators of\)\^(i)\(in the field theory side.\[and\]are the corresponding chemical potential and charge density in the dual boundary field theory, respectively. In this paper, we focus on the case\)m\^2=-2\(and set\)\^(2)=0\(to consider\)\^(1)\(as the vacuum expectation value of the operator\)O\_1\(. At the horizon\)z=z_h\(, the regular condition implies\)A_t(z_h)=0\(and\)f(z_h)=0\(. Then we can expand all the fields near the horizon and use the scaling symmetries \begin{eqnarray}\label{scalesymmetry} &&t\rightarrow a t,\quad z \rightarrow az ,\nonumber\\ &&z \rightarrow az, \quad (t,x,y) \rightarrow (t,x,y)/a, \quad A_t \rightarrow a A_t,\quad f\rightarrow f,\quad \psi\rightarrow \psi, \end{eqnarray} to set\)z_h=1\(. At high temperature, the scalar field will vanish and there is no condensation. The solution is the Reissner-Nordstr\"{o}m AdS black hole with \begin{eqnarray} \chi = \psi = 0 \,,\qquad A_t = \mu \Big(1-\frac{z}{z_h}\Big)\,, \qquad f = 1-\frac{z^3}{z_h^2} \Big(1 + \frac{\mu^2z_h^2}{4} \Big) + \frac{z^4}{z_h^4} \Big(\frac{\mu^2z_h^2}{4} \Big)~.\ \end{eqnarray} The temperature is give by \begin{eqnarray} T=\frac{1}{4\pi z_h}\Big(3-\frac{\mu^2z_h^2}{4}\Big)~. \end{eqnarray} When the temperature decreases to be lower than a critical value, a new type of charged black hole with non-vanishing charged scalar profile is numerically available. This corresponds to a hairy phase with\)O_1\(non-vanishing. The explicit dependence of the critical temperature for\)O_1\(on the coupling is presented in Fig.~\ref{fig-eta-Tc}. We see that the critical temperature increases as the coupling\[becomes larger, which is consistent with the results in the probe limit in our previous work. Moreover, we show the vacuum expectation values for\)O_1\(in Fig.~\ref{fig-condensation}. It is observed that as the coupling increases, the condensation gap is lower which agrees well with the property of the critical temperature we discussed above. Thus, from the gravitational side, we have found that the greater strength of the interaction will make the condensation easier to form in the back-reacted background. In the dual boundary field theory, this means that with stronger coupling between the\)U (1)\(gauge field and the scalar field, the gauge symmetry can be broken more easily. Having the solutions of the normal phase and the hairy phase below the critical temperature, we will study the holographic entanglement entropy (HEE) from high to lower temperatures in the next Section. \section{The entanglement entropy of the holographic system} \label{sec:4} Let us now discuss how we can incorporate the notion of entanglement entropy in the AdS/CFT correspondence. Imagine that we have a system\)A\(in the boundary CFT which has a gravity dual. Since the information included in a subsystem\)B\(is evaluated by the entanglement entropy\)S_A\(, we can ask which part of AdS space is responsible for the calculation of\)S_A\(in the dual gravity side. In a formula was proposed \begin{equation}\label{RTF} S_A=\frac{\mbox{Area}(\gamma_A)}{4G^{(d+2)}_N}\ , \end{equation} where\)\_A\(is the\)d\(-dimensional minimal surface whose boundary is given by the\)(d-1)\(-dimensional manifold\)\_A=A\(. The constant\)G\^(d+2)\_N\(is the Newton constant of the general gravity in AdS\)\_d+2\(. Before studying the holographic entanglement entropy (HEE) in this holographic model, we will geometrize the HEE of (\ref{RTF}) in terms of AdS/CFT duality. Following, we will consider the subsystem\)A\(with a straight strip geometry described by\)-x  , 0y L ,\(where\)l\(is defined as the size of region\)A\(and\)L\(is a regulator which can be set to be infinity. The induced metric of the hypersurface\)\_A\(whose boundary is the same as the stripe and has a profile like (\ref{metric1}) reads as \begin{equation} ds_{induced}^2=\frac{1}{z^2}\Big[(\frac{1}{f}+x'(z))dz^2+dy^2\Big]~. \end{equation} Thus, the HEE connecting with the area of the surface can be expressed as \begin{equation}\label{Area} 4G_4S=Area(\gamma_A)=L\int_{-l/2}^{l/2}\frac{dx}{z^2}\sqrt{1+\frac{z'(x)^2}{f}}~. \end{equation} The above expression can be treated as the Lagrangian with\)x\(direction thought of as time. The corresponding Hamiltonian is conserved because the Lagrangian does not explicitly depend on\)x\(. So we can get a constant of motion as \begin{equation}\label{zstar} \frac{1}{z_*^2}=\frac{1}{z^2\sqrt{1+\frac{z'(x)^2}{f}}}~, \end{equation} with\)z\_\*\(satisfying the condition\)dz/dx\|\_z=z\_\*=0\(. From (\ref{zstar}), we can write the width\)l\(in terms of\)z\_\*\(\begin{equation}\label{length} \frac{l}{2}=\int_{\epsilon}^{z_{*}}dz\frac{z^2}{\sqrt{(z_{*}^4-z^4)f}}~. \end{equation} Substituting (\ref{zstar}) into (\ref{Area}), we can obtain the entanglement entropy \begin{equation}\label{HEE} 4G_4S=2L\int_{\epsilon}^{z_{*}}dz\frac{z_{*}^2}{z^2}\frac{1}{\sqrt{(z_{*}^4-z^4)f}}=2L(s+\frac{1}{\epsilon})~. \end{equation} Here the term\)`<!-- -->`{=html}1/\(is divergent. While the term\)s\(is a finite term, so it is physically important. Now, we can calculate the entanglement entropy\)s\(based on the solution discussed in the last section. We will see the behavior of HEE from normal phase to hairy phase and investigate the effect of the\]coupling. Note that according to the scaling symmetries in (\ref{scalesymmetry}), the dimensionless quantities are\)T/\(,\) l\(and\)s/\(. Fixing the temperature, we see in Fig.~\ref{fig-l-s_ChangT} how the HEE changes as the width of stripe\)l\(changes. The green line is for the normal phase with the RN-AdS black hole background. We see that below the critical temperature when the scalar field starts to condensate, the entanglement entropy becomes smaller and it drops when the temperature becomes lower. This is consistent with the expectation that in the superconducting phase the degrees of freedom decrease due to the formation of Cooper pairs. This property holds for different values of the coupling\[. To illustrate the influence of the\]coupling, we present the HEE in change of temperature for a fixed\)l\(in Fig.~\ref{fig-T-s}. The light green line is the HEE for the normal state with RN-AdS black hole background. As the temperature decreases, the slope of HEE presents a discontinuous change at a critical temperatures\)T_c\(denoted by vertical dashed lines in the figure for different strength of the coupling\[. The discontinous change of the HEE marks the phase transition point from the normal state to the superconducting state. We again observe that the superconducting phase always have smaller HEE after the phase transition. At low temperature, the HEE of a superconductor with a smaller\]is smaller. Physically this happens because at low temperature the condensation becomes stronger with higher condensation gap for smaller coupling\[so the number of Cooper pairs is increased, which results in less degrees of freedom available. However this property does not hold when the temperature increases near the critical value, which is close to the interface with the normal phase, marked in the ellipse in Fig.~\ref{fig-T-s}, which is enlarged in Fig.~\ref{fig-T-sL}. When the temperature is increased above\)T/ = 0.1793\(, we see that the HEE for\]becomes higher than the case with\)= 0\(. When the temperature is above\)T/ = 0.188\(, the HEE for\[surpasses the value for\)= 0.3\(. The sharp change of the HEE which is related to the coupling\]in the vicinity of the contact interface, is not trivial. This behavior can be attributed to the proximity effect. For the smaller coupling holographic superconductor, it is more likely that the Cooper pairs of the supercontacting state can penetrate the normal state, which effectively results in the increase of the HEE in the superconductor phase. This is a similar behavior to the generalized Ginzburg-Landau theory on the boundary, where the phenomenological coefficient\[is related to the correlation length\) discussed in the introduction. A more negative \(\gamma\) for a fixed \(a\) leads more Cooper pairs to penetrate and this results in a larger leftover order parameter in the normal phase. On the other hand, for larger positive values of the coupling \(\eta\) coupling, Cooper pairs remain in the superconductor phase, thus leads the HEE to relatively smaller values. Again this agrees to the generalized Ginzburg-Landau theory description that more positive \(\gamma\) leads to smaller penetration of the Cooper pairs so that the leftover order parameter in the normal phase is weaker. # conductivity {#sec:5} In this section we will discuss the conductivity in an attempt to understand better the behavior of the system near the critical temperature. In the superconductor, the conductivity possesses certain distinguishing properties which are largely dependent on the microscopic details. To support the results we found studying the entanglement entropy, we will calculate the real part of conductivity for two characteristic temperatures at which we observed two different behaviors of the entanglement entropy. The first one is a low temperature which characterizes the superconducting phase where the entanglement entropy decreases with the decrease of the coupling constant \(\eta\), at which more Cooper pairs are formed for more negative coupling \(\eta\). For a high temperature we choose a temperature at which the superconducting state had just been formed and effectively it is close to the transition temperature of the superconducting to normal state. At the high temperature we found that the entanglement entropy is higher for negative \(\eta\). We consider the perturbation of metric and \(U(1)\) field as \(\delta g_{\mu\nu}=g_{tx}(z)e^{-i\omega t}\) and \(\delta A_{\mu}=A_x(z)e^{-i\omega t}\). Then, the first order perturbation equation of \(g_{tx}\) can be deduced as \]g_{tx}'-\frac{2 g_{tx}}{z}-A_x A_t'\Big[1-z^2 \eta\Big(f \psi'^2-\frac{q^2 e^{\chi} A_t^2 \psi^2}{f}\Big)\Big]=0~.\[ Having this equation we can deduce the linearised perturbative Maxwell equation which decouples from \(g_{tx}\) \](1+2z^2\eta f\psi^{'2})A_x^{''}+\Big(\frac{f'}{f}-\frac{\chi'}{2}+a_1\eta\Big)A_x'+\Big[ \Big(\frac{\omega^2}{f^2}-\frac{z^2A_t^{'2}}{f}\Big)e^{\chi}-\frac{2q^2\psi^2}{z^2f}+a_2\eta+a_3\eta^2\Big]A_x=0\[ with \]\begin{aligned} a_1&=&z\psi'\Big[z\psi'(4f'-f\chi')+4f(\psi'+z\psi'')\Big]~,\nonumber\\ a_2&=&\frac{e^{2\chi}q^2z^2A_t^2\psi^2}{f^2}\Big(z^2A_t^{'2}-\frac{2\omega^2}{f}\Big)\nonumber\\&&- \frac{e^{\chi}}{f}\Big[2q^2z^2A_t\psi A_t^{''}+zA_t^{'2}(2q^2\psi^2+z^2\psi'f)+q^2zA_t\psi A_t^{'} (4\psi+4z\psi'+z\psi\chi')\Big]~,\nonumber\\ a_3&=&2e^{\chi}z^6A_t^{'2}f\Big[\psi'^4-2e^{\chi}\Big(\frac{qA_t\psi\psi'}{f}\Big)^2 +e^{2\chi}\Big(\frac{qA_t\psi}{f}\Big)^4\Big]~.\nonumber\\ \nonumber \end{aligned}\[ Although that above equation seems very complicated, it can be solved by imposing the ingoing boundary condition near the horizon \]\label{Axz1} A_x(z\rightarrow z_h)\propto f^{\frac{-i\omega}{4\pi T}}\[ where the temperature \(T\) is defined in ([\[temperature\]](#temperature){reference-type="ref" reference="temperature"}). Near the asymptotic AdS boundary, the behavior of perturbation \(A_x\) is \]\label{Axz0} A_x(z\rightarrow 0)=A_x^{(0)}+z A_x^{(1)}~.\[ Then, by invoking the AdS/CFT duality, the conductivity of holographic superconductivity can be expressed as  \]\label{conductivity} \sigma=-\frac{iA_x^{(1)}}{\omega A_x^{(0)}}~.\(\) After solving the equation with the boundary condition ([\[Axz1\]](#Axz1){reference-type="ref" reference="Axz1"}), we can numerically extract the asymptotic value of \(A_x\) to calculate the conductivity of our system. We concentrate on the real part of the conductivity, since it is the dissipative part of the conductivity and measures the presence of charged states as a function of energy. At the low temperature, for example when \(T/\sqrt{\rho}=0.0574\), which is lower than the smallest crossing temperature in Fig. [\[fig-T-sL\]](#fig-T-sL){reference-type="ref" reference="fig-T-sL"}, we see in the left panel of Fig. [\[fig-sigma1\]](#fig-sigma1){reference-type="ref" reference="fig-sigma1"} that at low frequencies the holographic superconducting system with smaller coupling \(\eta\) has lower real part of conductivity and a larger frequency gap. The gap in the conductivity depends on the condensation, \(\omega_g\sim <O>\), which indicates a gap in the spectrum of charged excitations. For the smaller coupling, the higher condensation leads to the larger superconducting gap and to a smaller real part of the conductivity. The drop in the real part of the conductivity corresponds to a drop in the density of excitations at energies below the chemical potential. Thus for smaller \(\eta\), more 'electrons' are bounded in Cooper pairs which explains the smaller HEE at low temperature. At high temperature near the critical point, where the dependence of HEE on the coupling \(\eta\) is reversed compared with the low temperature case, the behavior of the real part of the conductivity is plotted in the right plot of Fig. [\[fig-sigma1\]](#fig-sigma1){reference-type="ref" reference="fig-sigma1"} with \(T/\sqrt{\rho}=0.20407\) for example, which is higher than the largest crossing temperature in Fig. [\[fig-T-sL\]](#fig-T-sL){reference-type="ref" reference="fig-T-sL"}. In the low frequency, we see that the real part of conductivity is bigger when the coupling is smaller. Fixing the frequency, we show the dependence of the real part of the conductivity on the temperature in Fig. [\[fig-sigma2\]](#fig-sigma2){reference-type="ref" reference="fig-sigma2"}. The sharp different dependence of the real part of the conductivity on the coupling \(\eta\) at low and high temperatures are clearly shown. In the contact interface, we see that for the smaller \(\eta\), the real part of the conductivity is higher. The bigger real part of the conductivity in the contact interface supports the proximity effect argument that less Cooper pairs are left in the superconducting phase with smaller \(\eta\) coupling. # Conclusions {#sec:6} We have studied the behavior of the entanglement entropy of a superconducting system. Motivated by the generalized Ginzburg-Landau theory we have introduced a higher-derivative coupling between the \(U(1)\) gauge field and the scalar field with coupling constant \(\eta\). In the boundary theory this coupling corresponds to one of the phenomenological constants of the Landau-Ginzburg theory. We have solved numerically the fully back-reacted gravitational system and found that as the coupling \(\eta\) increases the critical temperature increases and the energy gap decreases. This result suggests that the system with larger coupling is easier to enter the superconducting phase. Knowing the behavior of the system below the critical temperature, we have calculated the entanglement entropy using a stripe geometry. We have found that the entanglement entropy is less than the entanglement entropy of the normal phase. This agrees with the known results that the entanglement entropy plays the role of the order parameter of the superconducting system counting the degrees of freedom. In the superconducting phase less degrees of freedom are available because of the formation of Cooper pairs, therefore the entanglement entropy is less than the normal phase. We also found that a larger coupling gives a larger entanglement entropy. This means that for a fixed temperature for a larger coupling, less Cooper pairs are available in the superconducting system. We have found that near the boundary superconducting/normal interface the entanglement entropy has reversed behavior. More negative coupling gives higher entanglement entropy. This can be explained because of the proximity effect. For the more negative coupling, more Cooper pairs have leaked to the normal phase, so that less electrons are bounded in the superconductor phase which leads to the higher value of the entanglement entropy. This behavior of the entanglement entropy near the boundary contact interface is very interesting. It further supports the previous finding that the entanglement entropy plays the role of the order parameter, measuring the degrees of freedom of a superconducting system. To support further this behavior of the entanglement entropy we have calculated the real part of conductivity for two characteristic temperatures, one low temperature and the other near the critical temperature which effectively close to the interface of the normal phase. For the low frequency at the low temperature, smaller coupling has lower conductivity. But when the temperature is near the critical temperature, the behavior of the conductivity is reversed due to the proximity effect. In conclusion we have built a holographic superconductor and presented a holographic description of the proximity effect in superconductivity. It would be interesting to extend this study in the presence of an external magnetic field. For an inhomogeneous magnetic field the entanglement entropy can give us more information on the phase structure of the holographic superconducting system at low temperatures and possibly of the formation of FFLO states.

{'timestamp': '2014-05-08T02:11:37', 'yymm': '1401', 'arxiv_id': '1401.5720', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5720'}

# Introduction Massive young clusters are responsible for giant regions of ionized gas. These are photoionized by the clusters themselves, as well as shock-ionized by supernovae and supersonic stellar winds from the high-mass stars. Here, I will focus on the photoionized gas, which provides the emission-line diagnostics that are widely used to evaluate a variety of phenomena across the universe, for example, star-formation rate and ionizing stellar populations, as well as gas properties such as metallicity, density, pressure, and kinematics. Many papers presented in this conference exploit this technique, in particular, work by E. Telles and M. Rodriguez in this session. When considering global properties of galaxies, the  region luminosity function provides a quantitative parameterization of star formation. And finally, stellar radiation from optically thin  regions is responsible for ionizing the interstellar medium and thereby generating the diffuse, warm ionized medium (WIM). It is therefore apparent that photoionization by massive stars in clusters not only provides essential diagnostics of physical conditions, but also is itself a fundamental process. What is the fate of ionizing photons? We clearly see  regions in the immediate vicinity of massive clusters, but a significant proportion of these regions must be optically thin, and thus ionize the WIM. Similarly, if the ISM is itself optically thin in some galaxies, then Lyman continuum radiation will escape into the circumgalactic medium and perhaps, the intergalactic medium. It is widely believed that this was the case in early cosmic times, and that massive stars are responsible for the reionization of the universe. Thus, photon path lengths likely range across many orders of magnitude. It is also important to keep in mind that when the photons are absorbed, whether near or far from their source, they not only cause ionization of matter, but also impart momentum, generating radiation pressure that may be significant to gas kinematics. This topic is explored by S. Silich in this session. # Ionization-Parameter Mapping A central issue is therefore quantifying which  regions are optically thin. Many studies have examined this problem using different strategies. While some authors model globally-integrated emission-lines (e.g., Iglesias-Páramo & Muñoz-Tuñon 2002; Giammanco et al. 2004), others model the WIM surface brightness generated by its opacity and scattering (e.g., Zurita et al. 2002; Seon 2009). Here, we demonstrate an elegant technique to evaluate the optical depth of  regions based on ionization-parameter mapping (IPM), exploiting the fact that the nebular ionization structure varies between optically thin and optically thick objects. Figure [\[f_neb\]](#f_neb){reference-type="ref" reference="f_neb"} shows the / emission-line ratio map for some  regions in the Small Magellanic Cloud (SMC). In the case of an optically thick, Strömgren sphere, the ionization state will be high in the central regions near the ionizing source, and will decrease at large radii, as it transitions to the neutral environment at the nebular boundary. Thus, the nebular structure shows more highly ionized species, represented in Figure [\[f_neb\]](#f_neb){reference-type="ref" reference="f_neb"} by O\(^{++}\), in the center, surrounded by an envelope of lower-ionization species like S\(^+\). The round  regions marked with crosses in Figure [\[f_neb\]](#f_neb){reference-type="ref" reference="f_neb"} demonstrate this classic, Strömgren sphere ionization structure. In contrast, the object with irregular morphology to the east in Figure [\[f_neb\]](#f_neb){reference-type="ref" reference="f_neb"} appears to be completely dominated by  and is missing the low-ionization envelope. This implies that the object is density-bounded and therefore optically thin. The irregular morphology is also consistent with radiation hydrodynamic simulations by Arthur et al. (2011) showing that objects with the highest ionization parameters, which are the ones that become optically thin, develop instabilities leading to irregular morphology such as that seen in the eastern object of Figure [\[f_neb\]](#f_neb){reference-type="ref" reference="f_neb"}. However, exploration of the parameter space with simple photoionization models shows that this elegant picture is a bit more complicated. Using CLOUDY (Ferland et al. 1998) models, we demonstrated that for lower ionizing effective temperatures, low-ionization envelopes may be present even for quite optically thin conditions (Pellegrini et al. 2012; hereafter P12). Thus, for IPM based on only two ionic species, it is the *absence* of these low-ionization transition zones that implies low optical depth; whereas their presence merely implies a substantial likelihood of optically thick conditions. However, we note that optically thin objects that possess low-ionization envelopes occur for cooler stars, which tend to yield smaller and fainter  regions. And as noted above, nebular morphology is also a diagnostic criterion. We tested IPM as an estimator of optical depth by applying the method to a dozen nebulae in the Large Magellanic Cloud (LMC) with known spectral classifications of the ionizing stars. We crudely estimated escape fractions  by assigning the objects as optically thick, thin, or blister objects based on their ionization structure. These categories were simply assigned values of  \(= 0\), 0.6, and 0.3, respectively. More accurate  for these objects were then measured by comparing the predicted versus observed  luminosities. The extremely crude IPM estimates agree surprisingly well with the measured values, to 25% on average (P12). So in general, we suggest that objects that *look* like Strömgren spheres usually *are* Strömgren spheres. Note that any low-ionization envelopes should also be detected in the line of sight; thus IPM diagnostics are not limited to projected radial analysis. Furthermore, we stress that with three or more radially varying ionic species, the resulting maps of the ionization structure will constrain the optical depth far more strongly, essentially allowing its direct measurement. We have obtained such data for M33, which will permit us to quantify this capability. # Optical Depth of  Regions in the Magellanic Clouds The data presented above for individual nebulae in the LMC and SMC were generated from the Magellanic Clouds Emission-Line Survey (MCELS; Smith et al. 2005), a narrow-band imaging survey in , , and  of the entire star-forming extent of both Clouds. These data allow us to apply IPM to crudely estimate the optical depths of the entire  region population in both galaxies. It is especially interesting to quantitatively compare these, given that the LMC and SMC have strongly contrasting neutral  ISM: as seen in the  surveys of these galaxies (Kim et al. 1998; Stanimirović et al. 1999), the LMC has a highly shredded, filamentary neutral ISM with high porosity, owing to extensive mechanical feedback from massive stars and its flat, disk structure; whereas the SMC has a more three-dimensional and diffuse  distribution with much lower porosity due to its much lower specific star-formation rate (Oey 2007). Using IPM, we crudely categorize all  regions as optically thick, thin, or neither (see P12 for details). Figure [\[f_tauHI\]](#f_tauHI){reference-type="ref" reference="f_tauHI"} shows the frequency of optically thin objects as a function of  column density  measured within the nebular apertures for the LMC and SMC. As expected, the fraction of optically thin objects clearly decreases with  column. However, it is noteworthy that that the optically thin objects actually dominate at the lowest  in both galaxies. The value of  at which this occurs is higher in the SMC than in the LMC, owing to the SMC's 3-D ISM structure. Furthermore, we also see that there are still optically thin objects even at the highest  columns. While the mean  for the optically thick objects is larger in both galaxies, the  distributions for the optically thick and thin objects are similar in shape (P12). Figure [\[f_tauHa\]](#f_tauHa){reference-type="ref" reference="f_tauHa"} shows the frequency of optically thin objects as a function of  luminosity  for the LMC and SMC. In both galaxies, there is a clear trend that the frequency increases with . Even so, we again note optically thick objects exist even at the highest values of . However, optically thin objects dominate in numbers above a value of about \(\log\) \(\sim 37/\ergs\) in both galaxies. This low luminosity corresponds to objects ionized by single, mid-type O stars, implying that most of the bright  regions seen in star-forming galaxies tend to be optically thin. The  distributions for the optically thick vs thin objects differ far more strongly than the distributions of their  (P12). The statistics for these populations indicate lower limits on the total nebular escape fraction of \(=0.42\) and 0.40 in the LMC and SMC, respectively. From these, we can derive total "escape luminosities" of \(\log L_{\rm esc}/\ergs \sim 40.1\) and 39.2, respectively, representing the total potential  luminosities allowed by these . Comparing \(L_{\rm esc}\) to observed WIM luminosities in the LMC and SMC of \(\log\) /\(\ergs=40.0\) and 39.3, respectively, we find that not only are the \(L_{\rm esc}\) large enough to fully ionize the WIM, but also that the global escape fractions from these galaxies may be non-zero, when accounting for contributions from field OB stars having no associated nebulae (see P12 for details). If this is the case, IPM shows that  is likely dominated by a few objects in the galactic periphery, rather than the most luminous objects in the dominant, central star-forming regions. # Starburst Galaxies The individual Magellanic Clouds nebulae show that the most luminous objects are more likely to be optically thin. If so, then entire starburst galaxies might also plausibly have large . However, many studies have evaluated this possibility with mixed results. Only two local starbursts have confirmed detections of the Lyman continuum (e.g., Leitet et al. 2013; Grimes et al. 2009), while a minority of Lyman-break galaxies show such detections (e.g., Iwata et al. 2009; Shapley et al 2006). Absorption-line studies of local starbursts all imply optically thick conditions (e.g., Heckman et al. 2001; Leitherer et al. 1995). We therefore apply the IPM technique globally to a sample of local starburst galaxies: Haro 10 (NGC 5253), NGC 3125, Henize 2-10, NGC 1705, and NGC 178 (Zastrow et al. 2013). We carry out mapping in  and , using the Maryland-Magellan Tunable Filter (MMTF) at Magellan. The IPM technique is again proven, revealing a vivid ionization cone in Haro 10 (Zastrow et al. 2011; Figure [\[f_Haro10\]](#f_Haro10){reference-type="ref" reference="f_Haro10"}) and NGC 3125. The remaining galaxies show no significant evidence of optically thin regions. The narrow morphology of the ionization cones suggests that galaxy orientation to the line of sight plays a major role in the detection of Lyman continuum and optically thin gas. Thus, significant values of galactic  may be more common than observations thus far imply. While our sample is small, our results are also consistent with suggestions that a mininum star-formation intensity is needed to generate optically thin conditions (e.g., Fernandez & Shull 2011). It is also likely that a specific age range is necessary, corresponding to times after which mechanical feedback from the starburst has sufficiently shredded the ISM to facilitate the escape of ionizing radiation, but also before the ionizing stars have expired. This should be around 3--5 Myr, again consistent with the data from our small sample. Finally, since IPM implies that optically thin objects generate enhanced ionization parameters, we suggest that objects with extreme ionization parameters may therefore have high . The presentation in a later session by A. Jaskot demonstrates that this scenario may indeed be likely (Jaskot & Oey 2013). Hence we see that the technique of ionization-parameter mapping is a powerful tool on both local and galaxy-wide scales.

{'timestamp': '2014-01-28T02:07:55', 'yymm': '1401', 'arxiv_id': '1401.5779', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5779'}

# Introduction {#intro} Recent analytical insights into the nonperturbative nature of the confining interaction in QCD follow from the remarkable holographic correspondence between the equations of motion in AdS space and the light-front (LF) Hamiltonian equations of motion for relativistic light hadron bound-states in physical space-time . In fact, the mapping of the equations of motion  and the matching of the electromagnetic  and gravitational  form factors in AdS space  with the corresponding expressions derived from LF quantization in physical space time is the central feature of the LF holographic approach to hadronic physics. This approach allows us to establish a precise relation between wave functions in AdS space and the LF wavefunctions (LFWFs) describing the internal structure of hadrons. However the actual form of the effective potential has remained unknown until very recently and is thus model-dependent. It was been realized very recently  that the form of the effective LF confining potential can be obtained from the framework introduced in a remarkable paper by V. de Alfaro, S. Fubini and G. Furlan (dAFF) . It was shown by dAFF that in the Schrödinger representation, a scale can appear in the Hamiltonian operator while retaining the conformal invariance of the action . This remarkable result is based on the isomorphism of the algebra of the one-dimensional conformal group Conf\((R^1)\) to the algebra of generators of the group SO(2,1). One of the generators of this group, the rotation in the 2-dimensional space, is compact and has therefore a discrete spectrum with normalizable eigenfunctions. As a result, the form of the evolution operator is fixed and includes a confining harmonic oscillator potential, and the time variable has a finite range. In fact, it was shown in Ref. that there exists a remarkable holographic connection between the one-dimensional semiclassical approximation to light-front dynamics with gravity in a higher dimensional AdS space, and the constraints imposed by the invariance properties under the full conformal group in one dimension Conf\((R^1)\). Other approaches to emergent holography are discussed in . # Light-Front Holography and Conformal Quantum Mechanics For a hadron with four-momentum \(P^\mu\), the generators \(P^\mu = (P^-, P^+, \vec{P}_\perp)\), \(P^\pm = P^0 \pm P^3\), are constructed canonically from the QCD Lagrangian by quantizing the system on the light-front at fixed LF time \(x^+\), \(x^\pm = x^0 \pm x^3\) . The LF Hamiltonian \(P^-\) generates the LF time evolution \(P^-\vert \phi \rangle = i \frac{\partial}{\partial x^+} \vert \phi \rangle\), whereas the LF longitudinal \(P^+\) and transverse momentum \(\vec P_\perp\) are kinematical generators. In the limit of zero quark masses the longitudinal modes decouple from the invariant LF Hamiltonian equation \(H_{LF} \vert \phi \rangle = M^2 \vert \phi \rangle\) with \(H_{LF} = P_\mu P^\mu = P^-P^+-\vec{P}_\perp^2\). We obtain the wave equation  \[\label{LFWE} \left(-\frac{d^2}{d\zeta^2}-\frac{1-4L^2}{4\zeta^2} + U\left(\zeta, J\right) \right) \phi_{n,J,L} = M^2 \phi_{n,J,L},\] a relativistic single-variable LF Schrödinger equation, where the effective potential \(U\) acts on the valence sector of the theory. The effective potential follows from the systematic expression of the higher Fock components as functionals of the lower ones . The variable \(z\) of AdS space is identified with the LF boost-invariant transverse-impact variable \(\zeta\) , thus giving the holographic variable a precise definition in LF QCD . For a two-parton bound state \(\zeta^2 = x(1-x) b^{\,2}_\perp\), where \(x\) is the longitudinal momentum fraction and \(b_\perp\) is the transverse-impact distance between the quark and antiquark. Recently we have derived wave equations for hadrons with arbitrary spin starting from an effective action in AdS space . An essential element is the mapping of the higher-dimensional equations to the LF Hamiltonian equation found in Ref. [@deTeramond:2008ht]. This procedure allows a clear distinction between the kinematical and dynamical aspects of the LF holographic approach to hadron physics. Accordingly, the non-trivial geometry of pure AdS space encodes the kinematics, and the additional deformations of AdS encode the dynamics, including confinement , and determine the form of the LF effective potential from the precise holographic mapping to light-front physics . For \(d=4\) one finds from the dilaton-modified AdS action the LF potential  \[\label{U} U(\zeta, J) = \frac{1}{2}\varphi''(\zeta) +\frac{1}{4} \varphi'(\zeta)^2 + \frac{2J-3}{2 \zeta} \varphi'(\zeta),\] provided that the product of the AdS mass \(m\) and the AdS curvature radius \(R\) are related to the total and orbital angular momentum, \(J\) and \(L\) respectively, according to \((m R)^2 =-(2-J)^2 + L^2\). The critical value \(J=L=0\) corresponds to the lowest possible stable solution, the ground state of the LF Hamiltonian, in agreement with the AdS stability bound \((m R)^2 \ge-4\) , where \(R\) is the AdS radius. The classical Lagrangian of QCD is, in the limit of massless quarks, invariant under conformal transformations. Since we are interested in a semiclassical approximation to the nonperturbative domain of QCD, analogous to the quantum mechanical wave equations in atomic physics, it is natural to have a closer look at conformal quantum mechanics, a conformal field theory in one dimension, following dAFF . While leaving the action invariant, the dAFF procedure leads to a redefinition of the quantum mechanical evolution operator, and consequently to a redefinition of the corresponding 'time' evolution parameter \(\tau\), the range of which is finite. In the Schrödingier representation \[i \frac{\partial}{\partial \tau} \psi(x,\tau) = H_\tau \Big(x, -i \frac{d}{d x} \Big) \psi(x,\tau),\] the dAFF Hamiltonian \(H_\tau\) is given by  \[\label{Htaux} H_\tau = \frac{1}{2} u \left(-\frac{d^2}{d x^2} + \frac{g}{x^2}\right) + \frac{i}{4} v \left(x \, \frac{d}{d x}+ \frac{d}{d x} \, x \right) +\frac{1}{2} w x^2, \\ \] which is at \(\tau = 0\), the superposition of the 'free' Hamiltonian \(H\), the generator of dilatations \(D\) and the generator of special conformal transformations \(K\) in one dimension, the generators of Conf\((R^1)\); namely \[H_\tau = u H + v D + w K.\] The conformal group Conf\((R^1)\) is locally isomorphic to SO(2,1), the Lorentz group in 2+1 dimensions. Since the generators of Conf\((R^1)\) have different dimensions, their relations with the generators of SO(2,1) imply a scale, which here plays a fundamental role, as already conjectured in . Comparing the dAFF Hamiltonian with the light-front wave equation and identifying the variable \(x\) with the light-front invariant variable \(\zeta\), we have to choose \(u=2, \; v=0\) and relate the dimensionless constant \(g\) to the LF orbital angular momentum, \(g=L^2-1/4\), in order to reproduce the light-front kinematics. Furthermore \(w = 2 \lambda^2\) fixes the confining light-front potential to a quadratic \(\lambda^2 \, \zeta^2\) dependence. The choice of the dilaton profile \(\varphi(z) = \lambda z^2\) introduced in  thus follows from the requirements of conformal invariance. This specific form for \(\varphi(z)\) leads through to the effective LF potential \[U(\zeta, J) = \lambda^2 \zeta^2 + 2 \lambda (J-1),\] and corresponds to a transverse oscillator in the light-front. The term \(\lambda^2 \zeta^2\) is determined uniquely by the underlying conformal invariance of classical QCD incorporated in the one-dimensional effective theory, and the constant term \(2 \lambda (J-1)\) by the embedding space . ## A Light-Front Holographic Model for Mesons From one obtains for the effective LF potential \(U(\zeta, J) = \lambda^2 \zeta^2 + 2 \lambda (J-1)\) for \(\lambda >0\) a mass spectrum for mesons characterized by the total angular momentum \(J\), the orbital angular momentum \(L\) and orbital excitation \(n\) given by \[M_{n, J, L}^2 = 4 \lambda \left(n + \frac{J+L}{2} \right),\] an important result also found in Ref.. This result not only implies linear Regge trajectories, but also a massless pion and the relation between the \(\rho\) and \(a_1\) mass usually obtained from the Weinberg sum rules . The model also predicts hadronic LFWFs which underlie form factors  and other dynamical observables, as well as vector meson electroproduction . The spectral predictions for the \(J = L + S\) light pseudoscalar and vector meson families are compared with experimental data in Fig. [\[pionspec\]](#pionspec){reference-type="ref" reference="pionspec"}. The data are from PDG . As we will discuss in the next section, a spin-orbit effect is only predicted for mesons not baryons, as observed in experiment; it thus becomes a crucial test for any model which aims to describe the systematics of the light hadron spectrum. Using the spectral formula for \(M^2\) given above, we find  \[M_{a_2(1320)} > M_{a_1(1260)} > M_{a_0(980)}.\] The predicted values are 0.76, 1.08 and 1.32 GeV for the masses of the \(a_0(980)\), \(a_1(1260)\) and \(a_2(1320)\), compared with the experimental values 0.98, 1.23 and 1.32 GeV respectively. The prediction for the mass of the \(L=1\), \(n=1\) state \(a_0(1450)\) is 1.53 GeV, compared with the observed value 1.47 GeV. The LF holographic model with \(\lambda>0\) accounts for the mass pattern observed in the radial and orbital excitations of the light mesons, as well as for the triplet splitting for the \(L=1\), \(J = 0, 1, 2\), vector meson \(a\)-states. The slope of the Regge trajectories gives a value \(\sqrt \lambda \simeq 0.5 ~\rm GeV\). The value of \(\lambda\) required for describing the pseudoscalar sector is slightly higher that the value of \(\lambda\) extracted from the vector sector. The prediction for the observed spin-orbit splitting for the \(L=1\) \(a\)-vector mesons is overestimated by the model. Note that the solution for \(\lambda < 0\) leads to a pion mass heavier than the \(\rho\) meson in clear disagreement with observations. ## A Light-Front Holographic Model for Baryons The analytical exploration of the baryon spectrum using light-front holographic ideas is not as simple or as well understood as the meson case. However, as we shall discuss is this section, even a relatively simple approach provides a framework for a useful description of the baryon spectrum which gives important insights into its systematics. In a chiral spinor component representation, the light-front wave equations for baryons are given by the coupled linear differential equations  \[\begin{aligned} \label{LFDE} -\frac{d}{d\zeta} \psi_--\frac{\nu+{\textstyle{\frac{1}{2}}}}{\zeta}\psi_--V(\zeta) \psi_- &=& M \psi_+, \nonumber \\ \frac{d}{d\zeta} \psi_+-\frac{\nu+{\textstyle{\frac{1}{2}}}}{\zeta}\psi_+-V(\zeta) \psi_+ &=& M \psi_-, \end{aligned}\] where \(\nu\) is given in terms of the mass appearing in the Dirac equation in AdS space: \(\nu = \vert \mu R \vert-{\textstyle{\frac{1}{2}}}\). One can also show that the effective potential \(V\) is \(J\)-independent . This is a remarkable result, since it implies that independently of the specific form of the potential, the value of the baryon masses along a given Regge trajectory depends only on the LF orbital angular momentum \(L\), and thus, in contrast with the vector mesons, there is no spin-orbit coupling, in agreement with the observed near-degeneracy in the baryon spectrum. We choose an effective linear confining potential \(V = \lambda \, \zeta\) which also leads to linear Regge trajectories in both the orbital and radial quantum numbers for baryon excited states. The linear potential also leads to the LF oscillator form \(\lambda ^2 \zeta^2\) in the second order version of Eqs. [\[LFDE\]](#LFDE){reference-type="ref" reference="LFDE"}. For \(\lambda > 0\) we find the result \[M^2 = 4 \, \lambda \left( n + \nu + 1 \right).\] For \(\lambda <0\) no solution is possible . To determine the internal spin, internal orbital angular momentum and radial quantum number assignment of the \(N\) and \(\Delta\) excitation spectrum from the total angular momentum-parity PDG assignment, it is convenient to use the conventional \({\rm SU(6)} \supset {\rm SU(3)}_{flavor} \times {\rm SU(2)}_{spin}\) multiplet structure. The lowest possible stable state, the proton, corresponds to \(n=0\) and \(\nu = 0\). This fixes the scale \(\sqrt \lambda \simeq 0.5\) GeV. The resulting predictions for the positive-parity spin-\({\textstyle{\frac{1}{2}}}\) nucleons are shown in Fig. [\[baryonspec\]](#baryonspec){reference-type="ref" reference="baryonspec"} (a) for the parent Regge trajectory for \(n =0\) and \(\nu = 0, 2, 4, \cdots, L\), where \(L\) is the relative LF angular momentum between the active quark and the spectator cluster. The predictions for the daughter trajectories for \(n=1\), \(n = 2, \cdots\) are also shown in this figure. Only confirmed PDG states are shown. The Roper state \(N(1440)\) and the \(N(1710)\) are well accounted for as the first and second radial excited states of the proton. The newly identified state, the \(N(1900)\)  is depicted here as the first radial excitation of the \(N(1720)\). The model is successful in explaining the parity degeneracy observed in the light baryon spectrum, such as the \(L=2\), \(N(1680)-N(1720)\) pair in Fig. [\[baryonspec\]](#baryonspec){reference-type="ref" reference="baryonspec"} (a). In Fig. [\[baryonspec\]](#baryonspec){reference-type="ref" reference="baryonspec"} (b) we compare the positive parity spin-\({\textstyle{\frac{1}{2}}}\) parent nucleon trajectory with the negative parity spin-\({\frac{3}{2}}\) nucleon trajectory. It is remarkable that the gap scale \(4 \lambda\) determines not only the slope of the trajectories, but also the gap in the spectrum between the plus-parity spin-\({\textstyle{\frac{1}{2}}}\) and the minus-parity spin-\({\frac{3}{2}}\) nucleon families, as indicated by arrows in this figure. This means the respective assignment \(\nu = L\) and \(\nu = L+1\) for the lower and upper trajectories in Fig. [\[baryonspec\]](#baryonspec){reference-type="ref" reference="baryonspec"} (b). We also note that the degeneracy of states with the same orbital quantum number \(L\) is also well described, as for example the degeneracy of the \(L=1\) states \(N(1650)\), \(N(1675)\) and \(N(1700)\) in Fig. [\[baryonspec\]](#baryonspec){reference-type="ref" reference="baryonspec"} (b). We have also to take into account baryons with negative parity and internal spin \(S = {\textstyle{\frac{1}{2}}}\), as well as baryon states with positive parity and internal spin \(S = {\frac{3}{2}}\) such as the \(\Delta(1232)\). Those states are well described by the assignment \(\nu = L + {\textstyle{\frac{1}{2}}}\). This means, for example, that \(M^{2 \,(+)}_{n, L, S = \frac{3}{2}} = M^{2 \,(-)}_{n, L, S = \frac{1}{2}}\) and consequently the positive and negative-parity \(\Delta\) states lie in the same trajectory consistent with the experimental results, as depicted in Fig. [\[baryonspec\]](#baryonspec){reference-type="ref" reference="baryonspec"} (d). The newly found state, the \(N(1875)\), depicted in Fig. [\[baryonspec\]](#baryonspec){reference-type="ref" reference="baryonspec"} (c) is well described as the first radial excitation of the \(N(1520)\), and the near degeneracy of the \(N(1520)\) and \(N(1535)\) is also well accounted. Likewise, the \(\Delta(1600)\) corresponds to the first radial excitation of the \(\Delta(1232)\) as shown in Fig. [\[baryonspec\]](#baryonspec){reference-type="ref" reference="baryonspec"} (d). The model explains the important degeneracy of the \(L=2\), \(\Delta(1905),\) \(\Delta(1910),\) \(\Delta(1920),\) \(\Delta(1950)\) states which are degenerate within error bars. Our results for the \(\Delta\) states agree with those of Ref. . "Chiral partners\" such as the \(N(1535)\) and the \(N(940)\) with the same total angular momentum \(J = {\textstyle{\frac{1}{2}}}\), but with different orbital angular momentum are non-degenerate from the onset. To recapitulate, the parameter \(\nu\) has the internal spin \(S\) and parity \(P\) assignment given in the table below. [\[nuT\]]{#nuT label="nuT"} [\[asig\]]{#asig label="asig"} The assignment \(\nu = L\) for the lowest trajectory, the proton trajectory, is straightforward and follows from the mapping of AdS to light-front physics. The assignment for other spin and parity baryons states given in Table [1](#asig){reference-type="ref" reference="asig"} is phenomenological. It is expected that further analysis of the different quark, or quark--diquark, configurations and symmetries of the baryon wave function will indeed explain the actual assignment in Table [1](#asig){reference-type="ref" reference="asig"}, which successfully describes the full light baryon orbital and radial excitation spectrum, and in particular the gap between trajectories with different parity and internal spin. The holographic variable \(\zeta\) has a cluster decomposition and thus labels a system of \(n\)-partons as an active quark plus a system of \(n-1\) spectators . From this perspective, a baryon with \(n=3\) looks in light-front holography as a quark--diquark system. # Conclusions We have followed the remarkable results of De Alfaro, Fubini and Furlan  which, combined with light-front holographic QCD , give important insights into the QCD confining mechanism. It turns out that it is possible to introduce a scale by modifying the variable of dynamical evolution and nonetheless the underlying action stays conformally invariant. This procedure determines uniquely the form of the LF effective potential, corresponding to a quadratic dilaton modification of AdS space to include confinement. We obtain an effective one-dimensional quantum field theory which retains the fundamental conformal symmetry of the classical QCD Lagrangian in the limit of massless quarks. As a result the mass scale for confinement and holographic mapping to AdS space are emergent properties. The group theoretical arguments based on the underlying conformality fix the quadratic form of the effective dilaton profile and thus the corresponding form of the confinement potential of the LF QCD Hamiltonian. The result is a relativistic LF wave equation for bound states which leads to a remarkable description of the spectroscopy and dynamics of the light mesons and baryons .

{'timestamp': '2014-01-23T02:02:48', 'yymm': '1401', 'arxiv_id': '1401.5531', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5531'}

# Introduction A major theme that has unified exoplanet searches for more than 20 years is the question of how common (or rare) our own Solar system is. The *Kepler* spacecraft, which continuously monitored over 100,000 stars for tiny eclipses caused by orbiting planets, has provided exquisite data which have revolutionised our understanding of the frequency of Earth-size planets in short-period orbits . However, *Kepler* alone cannot give us a complete picture of the occurrence rate of planetary systems like our own, with rocky inner planets and one or more gas giant planets ("Jupiter analogs") at orbital distances \(a\gtsimeq\)`<!-- -->`{=html}3 AU. There is more to a system being Solar System-like than having a single planet in a potentially habitable orbit. The detection of a Jupiter analog is a second key component in determining whether an exoplanetary system is Solar system-like. Over the years, many arguments have been put forth to suggest that such external giant planets might be a necessity for a potentially habitable exo-Earth to be a promising location for the development of life. Although the role of such planets acting as a shield from an otherwise damaging impact regime has come into question, a number of other potential benefits are thought to accrue from the presence of Jupiter-analogs. For example, Jupiter-like planets have been proposed as a solution to the question of the origin of Earth's water. Current models of planetary formation suggest that the Earth formed in a region of the proto-planetary disk that was far too warm for water to condense from the gas phase. As such, it is challenging to explain the origin of our planet's water without invoking an exogenic cause. The formation and evolution of the giant planets, beyond the snow line, offers a natural explanation for the delivery of volatiles from the cold depths of a planetary system to planets that move on potentially habitable orbits (e.g. Horner et al. 2009; Horner & Jones 2010b, and references therein). The detection of a Jupiter-analog is therefore both a second key component in determining whether an exoplanetary system is Solar system-like, and a potential marker that the planets in that system might be promising targets for the future search for life beyond the Solar system. The Anglo-Australian Planet Search (AAPS) has been in operation for 15 years, and has achieved a long-term radial-velocity precision of 3 m s\(^{-1}\) or better since its inception, which is enabling the detection of long-period giant planets. To date, the AAPS has discovered six Jupiter analogs: HD 70642b, HD 160691c, HD 30177b, GJ 832b, HD 134987c , HD 142c. Here, we have defined a Jupiter analog as a giant planet which has ended up near its formation location, beyond the ice line, with \(a>3\) AU. Recently, the AAPS has shifted its priority to the detection of these Jupiter analogs. The observing strategy and target list have been modified, with the aim of producing an accurate and precise determination of the frequency of Jupiter-like planets in Jupiter-like orbits. The modified target list includes stars with long-term velocity stability such that Jupiter analogs can be robustly excluded , as well as those stars with as-yet-incomplete orbits suggestive of long-period giant planets. In this paper, we report the discovery of two such Jupiter analogs with complete orbits. HD 154857 is already known to host a 1.8M\(_{\rm Jup}\) planet with an orbital period of about 400 days; a residual velocity trend indicated a much longer-period object, as noted in the discovery work and in. This paper is organized as follows: Section 2 briefly describes the observational details and stellar parameters, and Section 3 details the orbit fitting process and gives the parameters of the two new planets. In Section 4, we present a dynamical stability analysis of the HD 154857 two-planet system, and we give our conclusions in Section 5. # Observations and Stellar Parameters AAPS Doppler measurements are made with the UCLES echelle spectrograph . An iodine absorption cell provides wavelength calibration from 5000 to 6200 Å. The spectrograph point-spread function (PSF) and wavelength calibration are derived from the iodine absorption lines embedded on the spectrum by the cell . The result is a precise Doppler velocity estimate for each epoch, along with an internal uncertainty estimate, which includes the effects of photon-counting uncertainties, residual errors in the spectrograph PSF model, and variation in the underlying spectrum between the iodine-free template and epoch spectra observed through the iodine cell. All velocities are measured relative to the zero-point defined by the template observation. For HD 114613, a total of 223 AAT observations have been obtained since 1998 Jan 16 (Table [\[114613vels\]](#114613vels){reference-type="ref" reference="114613vels"}) and used in the following analysis, representing a data span of 5636 days (15.4 yr). The mean internal velocity uncertainty for these data is 0.94 m s\(^{-1}\). HD 154857 has been observed 42 times since 2002 April (Table [\[154857vels\]](#154857vels){reference-type="ref" reference="154857vels"}), for a total time span of 4109 days (11.3 yr) and a mean internal uncertainty of 1.71 m s\(^{-1}\). HD 114613 (HR 4979; HIP 64408) is an inactive G-type star, listed as a dwarf by, though its surface gravity is more indicative of a slightly evolved subgiant (Table [\[114star\]](#114star){reference-type="ref" reference="114star"}). It is a nearby and bright star (\(V=4.85\)) with a somewhat super-solar metallicity \([Fe/H]\sim\)`<!-- -->`{=html}0.19. HD 154857 has been classified as a G5 dwarf. However, all recent measurements of its surface gravity show that this star is a subgiant (Table [\[154star\]](#154star){reference-type="ref" reference="154star"}). There is some confusion as to the mass: give two disparate mass estimates, 2.10\(\pm\)`<!-- -->`{=html}0.31M\(_{\odot}\) derived from spectroscopic analysis, and 1.27\(^{+0.35}_{-0.29}\)M\(_{\odot}\) from interpolation on a grid of Yonsei-Yale isochrones. For most stars in their sample, the two mass estimates agreed within \(\sim\)`<!-- -->`{=html}10%, but for HD 154857, they differ by almost a factor of 2. The more recent analysis by yields an intermediate value of 1.718\(^{+0.03}_{-0.022}\)M\(_{\odot}\), which we adopt in this paper. # Orbit Fitting and Planetary Parameters ## HD 114613 HD 114613 has been observed by the AAPS for the full 15 years of its operation. A long-period trend had been evident for several years, and in 2011, the trend resolved into a complete orbital cycle. We have since continued to observe HD 114613 to verify that the \(\sim\)`<!-- -->`{=html}11 yr orbit was indeed turning around. Figure [\[pgram1\]](#pgram1){reference-type="ref" reference="pgram1"} shows the Generalized Lomb-Scargle periodogram of the 223 AAT observations. This type of periodogram weights the input data by their uncertainties, whereas the traditional Lomb-Scargle method assumes uniform, Gaussian distributed uncertainties. To assess the significance of any signals appearing in these periodograms, we performed a bootstrap randomization process . This randomly shuffles the velocity observations while keeping the times of observation fixed. The periodogram of this shuffled data set is then computed and its highest peak recorded. The longest-period peak near 4000 days is well-defined and highly significant, with a bootstrap false-alarm probability less than \(10^{-5}\). The next-highest peaks are at 122 and 1400 days, respectively. We fit these data with a single, long-period Keplerian using the *GaussFit* nonlinear least-squares minimization routine. Jitter of 3.42 m s\(^{-1}\) was added in quadrature to the uncertainties at each epoch prior to orbit fitting. A single-planet fit yields a period \(P=3825\pm\)`<!-- -->`{=html}106 d, \(K=5.4\pm\)`<!-- -->`{=html}0.4 m s\(^{-1}\), and \(e=0.25\pm\)`<!-- -->`{=html}0.08 (Table [\[planetparams\]](#planetparams){reference-type="ref" reference="planetparams"}), making this planet a Jupiter analog, with a minimum mass m sin \(i\) of 0.5M\(_{\rm Jup}\), and an orbital period of 10.7 years (Figure [\[114fit\]](#114fit){reference-type="ref" reference="114fit"}). The rms about the one-planet fit is 3.9 m s\(^{-1}\), and the periodogram of the residuals to this fit is shown in the right panel of Figure [\[pgram1\]](#pgram1){reference-type="ref" reference="pgram1"}, showing a number of peaks ranging from 28 to \(\sim\)`<!-- -->`{=html}1500 d. There is structure evident in this residual periodogram (Figure [\[pgram1\]](#pgram1){reference-type="ref" reference="pgram1"}, right panel), so we examined the residuals for additional Keplerian signals. One way of determining the veracity of such signals is to examine the data by seasons or subsets. This can disentangle true planetary signals (which would consistently appear in all subsets) from stochastic signals such as stellar rotational modulation. We divided the residuals to the one-planet fit into two eight-season chunks. HD 114613 was observed intensely in 2007 and 2009 as part of the AAT "Rocky Planet Search" campaigns, in which 24-30 bright stars were observed nightly for 48 continuous nights in search of short-period planets. It is possible that such a density of observational data may skew the false-alarm probabilities when evaluating potential additional signals. We thus removed the 66 epochs from the two "Rocky Planet Search" campaigns--this resulted in the 8-year halves containing 77 and 80 observations, respectively[^1]. Periodograms of the two halves are shown in Figure [\[split\]](#split){reference-type="ref" reference="split"}; visual inspection reveals that they are markedly different. Table [\[boot\]](#boot){reference-type="ref" reference="boot"} shows the false-alarm probabilities obtained from 10,000 such realizations on each half of the 1-planet residuals. No periodicity is consistently significant in both subsets, with the possible exception of that near 27-29 days--however, this is worryingly close to both the 33-day rotation period of the star and the lunar month (at which the sampling of radial-velocity observations is well-known to impart spurious periodicities e.g. Dawson & Fabrycky 2010, Wittenmyer et al. 2013b). While it is tempting to consider a second planet near 1400 days, as found in the raw-data periodogram (Figure [\[pgram1\]](#pgram1){reference-type="ref" reference="pgram1"}), we see that this signal is simply not evident in the first 8 years of observations. As suggested by for the proposed planet orbiting Alpha Centauri B, we rephrase his sentiments to express that any shorter period periodic signal which is evident in one subset of our data should also be evident in other subsets or seasons of the data. As both HD 114613 8-year subsets have ample time coverage and data quantity (\(N=77\)) to sample the candidate periods listed in Table [\[boot\]](#boot){reference-type="ref" reference="boot"}, we can use these results to conclude that there is not yet sufficient evidence for additional planetary signals in our data for HD 114613. ## HD 154857 The presence of a planet orbiting HD 154857 was first reported by , who noted that the AAPS data were best fit with the \(\sim\)`<!-- -->`{=html}400-day planet and a linear trend, indicating a more distant body. Additional data presented in refined the planet's parameters and attempted to constrain the outer object's orbit since the residual velocity trend had begun to show curvature. They determined a minimum orbit with period 1900 days and \(K\sim\)`<!-- -->`{=html}23 m s\(^{-1}\). Now, with a further 6 years of AAT data, the outer planet has completed an orbit and a double-Keplerian model converges easily. We used *GaussFit* as described above to fit the two planets, first adding jitter of 2.6 m s\(^{-1}\) in quadrature to the uncertainties (after O'Toole et al. 2007). The best-fit parameters are given in Table [\[planetparams\]](#planetparams){reference-type="ref" reference="planetparams"}; the outer planet is a Jupiter analog moving on an essentially circular orbit with \(P=9.5\) yr (\(a=5.36\)AU) and m sin \(i=2.6\)M\(_{\rm Jup}\). The rms about the two-planet fit is 3.2 m s\(^{-1}\), and there are no significant residual periodicities. The data and two-planet model are shown in Figure [\[154fit\]](#154fit){reference-type="ref" reference="154fit"}, and the orbital fits for the individual planets are shown in Figure [\[154eachone\]](#154eachone){reference-type="ref" reference="154eachone"}. # Dynamical Stability Testing Recent work has shown that any claim of multiple orbiting bodies must be checked by dynamical stability testing to ensure that the proposed planetary orbits are feasible on astronomically relevant timescales. Such testing can support the orbit fitting results [@texas1; @142paper; @NNSer], place further constraints on the planetary system configurations, or show that the proposed planets cannot exist in or near the nominal best-fit orbits . While HD 114613 appears to be a single-planet system, HD 154857 hosts two planets which are so widely separated (1.3 and 5.4 AU) that their dynamical interactions might be expected to be negligible. However, for completeness, we performed the dynamical analysis as in our previous work . We tested the dynamical stability of the HD 154857 system using the Hybrid integrator within the n-body dynamics package [Mercury]{.smallcaps} . Holding the initial orbit of the innermost planet fixed, we tested 41x41x15x5 grid of "clones" spaced evenly across the 3\(\sigma\) range in the outer planet's semi-major axis \(a\), eccentricity \(e\), periastron argument \(\omega\), and mean anomaly \(M\), respectively. In each integration, the orbital evolution of each planet was followed until it was either ejected from the planetary system (by reaching a barycentric distance of 10 AU), or collided with the central body or one of the other planets. The times at which such events occurred was recorded, which allowed us to construct a map of the stability of the HD 154857 planetary system as a function of the semi-major axis and eccentricity of the outer planet. As expected, the entire 3\(\sigma\) region exhibited stability for the full \(10^8\)yr. Indeed, not a single ejection or collision occurred in any of the 126,075 trial systems. # Discussion and Conclusions We have described the detection of two Jupiter-analog planets from the 15-year Anglo-Australian Planet Search program. Our new data confirm the planetary nature of the previously unconstrained outer body in the HD 154857 system. These results highlight the importance of continuing "legacy" programs such as the AAPS, which is among the world's longest-running radial-velocity planet searches. The planets detailed in this work bring the total number of AAPS-discovered Jupiter analogs to 8. With three Jupiter analogs confirmed in the past two years (HD 142c, Wittenmyer et al. 2012b; HD 114613b and HD 154857c, this work), the AAPS has nearly doubled its discoveries of these objects in years 14 and 15 of operation. We expect further discoveries of Jupiter analogs over the next few years as additional candidates complete orbits. The AAPS has shifted its primary focus to the search for Jupiter analogs. Central to this strategy is the selection of a subset of \(\sim\)`<!-- -->`{=html}120 targets (from the original 250-star AAPS sample) which satisfy two criteria: (1) sufficient observational baseline to detect a Jupiter analog, and (2) a sufficiently small velocity scatter to enable the robust detection of the \(\sim\)`<!-- -->`{=html}5-15 m s\(^{-1}\) signal produced by a Jupiter analog. Criterion (1) eliminates those stars added to the AAPS list well after its inception in 1998, and criterion (2) eliminates those stars which have high levels of intrinsic activity noise which would severely degrade the achievable detection limit. Using the detections and stringent limits from the *non-detections*, for every target we will be able to detect or exclude Jupiter analogs with high confidence. The result will be a direct measurement of the frequency of such objects, without suffering from significant incompleteness, which adds substantial uncertainty to this measurement . There is an emerging correlation between debris disks and low-mass planets, first noted by. They used *Herschel* to detect debris disks around 4 of 6 stars known to host only low-mass planets; no debris disks were found in the 5 systems hosting giant planets. One of the stars discussed here, HD 154857, has been observed for infrared excess (indicative of debris disks akin to the Solar system's Edgeworth-Kuiper Belt). No excess was found from *Spitzer* and *Herschel* observations . The HD 154857 system, hosting two giant planets and no detectable debris, is consistent with the pattern noted by. To obtain a complete picture of the nature of the planet candidates we have presented here, it would be ideal to determine true masses, rather than the minimum mass derived from radial-velocity measurements. Direct imaging offers a way forward: for stars known to host a long-period radial-velocity planet candidate, imaging can determine whether that object is stellar (i.e. detectable by imaging) or substellar. This type of characterization has been done for some planet candidates, such as 14 Herculis c (\(a>7\) AU--Wittenmyer et al. 2007), for which AO imaging by established an upper limit of 42M\(_{\rm Jup}\). The TRENDS survey is currently using this strategy to target stars with known radial-velocity trends. The Gemini Planet Imager (GPI), now installed on the 8m Gemini South telescope, has been specifically designed for the detection of these giant planets. It will provide not only the high contrasts needed to detect them, but also low-resolution spectra for each planet found, which can be used for their characterisation. We are now at a convergence of two developments in exoplanetary science: (1) radial-velocity data now extend comfortably into the range of Jupiter-analog orbital periods, and (2) direct imaging techniques have improved to the point where it is possible to detect Jupiter-like planets orbiting Sun-like stars at orbital distances approaching that of our own Jupiter (\(\sim\)`<!-- -->`{=html}5 AU). These complementary techniques can bridge the detectability gap, enabling direct measurements of the occurrence rate of Jupiter analogs orbiting Sun-like stars. We thank the referee, William Cochran, for a timely report which improved this manuscript. This research is supported by Australian Research Council grants DP0774000 and DP130102695. This research has made use of NASA's Astrophysics Data System (ADS), and the SIMBAD database, operated at CDS, Strasbourg, France. This research has also made use of the Exoplanet Orbit Database and the Exoplanet Data Explorer at exoplanets.org. [^1]: Fitting a single planet with this shortened dataset gives parameters within 1\(\sigma\) of those given in Table [\[planetparams\]](#planetparams){reference-type="ref" reference="planetparams"}, showing that the exclusion of those data do not affect our conclusions about the long-period planet.

{'timestamp': '2014-01-23T02:02:25', 'yymm': '1401', 'arxiv_id': '1401.5525', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5525'}

# Introduction One of the most severe constraints of the equation of state (EoS) of QCD at extreme densities comes from the recent \(2M_\odot\) mass determinations of PSR J1614-2230 and PSR J0348-0432 having important consequences for the existence of quark matter in compact stars. In this work we assume compact stars are the so-called hybrid stars, composed of a nuclear mantle and a quark core. In order to obtain heavy hybrids stars, quark matter EoS should be stiff, with a low onset. This can be achieved in the MIT bag model by perturbative corrections to the EoS and small bag pressures, while within the Nambu--Jona-Lasinio (NJL) model large vector channel is used for stiffness, while shift in the vacuum energy or introduction of superconductivity ensures low onset. An influence on the maximum mass is expected by the nature of the phase transition itself. We propose an alternative scenario by taking into account the fact that NJL is a non-renormalizable effective model, valid up to some scale \(\Lambda\). As we approach this scale, say by increasing the chemical potential, higher dimensional operators should become important. It has been pointed out that including higher scalar interactions can reduce the critical temperature in the NJL model, bringing it in closer agreement with lattice results at finite temperature. For further work on higher dimensional operators in the context of the NJL model see. Our aim is to make an initial study of the effect of multiquark interactions on occurrence of quark matter in heavy stars. In this work we will use the NJL model parametrization of Ref.  where the critical temperature at zero chemical potential is fitted to lattice QCD. In addition, we will also introduce the 8-quark vector channel, as a natural candidate for playing a relevant role at large densities reached in the cores of compact stars. Our main results are that with scalar 8-quark interactions provided by Ref.  we are able to obtain stable hybrid stars with small vector coupling in the 4-quark and zero vector coupling in the 8-quark channel. Second, we demonstrate that the mass of the star can be increased up to and above \(2M_\odot\) with the 8-quark vector interaction, while still keeping the 4-quark vector interaction low. # NJL model with 8-quark interactions We work within the framework of a \(N_f=2\) NJL model defined as follows \[\mathcal{L}=\bar{q}(i\slashchar{\partial}-m)q+\mu_q\bar{q}\gamma^0 q+ \mathcal{L}_4+\mathcal{L}_8~, \label{eq:njllag}\] where \(\mu_q\) is the quark chemical potential and \(m\) is the current mass. The interaction terms are \[\mathcal{L}_4 =\frac{g_{20}}{\Lambda^2} [(\bar{q}q)^2 + (\bar{q}i\gamma_5\boldsymbol{\tau} q)^2]-\frac{g_{02}}{\Lambda^2}(\bar{q}\gamma_\mu q)^2~,\] \[\begin{split} \mathcal{L}_8 &= \frac{g_{40}}{\Lambda^8}[(\bar{q}q)^2 + (\bar{q}i\gamma_5\boldsymbol{\tau} q)^2]^2-\frac{g_{04}}{\Lambda^8}(\bar{q}\gamma_\mu q)^4 \\ &-\frac{g_{22}}{\Lambda^8}(\bar{q}\gamma_\mu q)^2[(\bar{q}q)^2 + (\bar{q}i\gamma_5\boldsymbol{\tau} q)^2]~, \end{split} \label{eq:lag8}\] where \(\Lambda\) is the model cutoff. The Lagrangian ([\[eq:njllag\]](#eq:njllag){reference-type="ref" reference="eq:njllag"}) may be motivated by the large \(N_c\) counting: while the 4-quark couplings scale as \(1/N_c\), the 8-quark couplings scale as \(1/N_c^3\). In the mean-field approximation the Lagrangian becomes \[\mathcal{L}_\mathrm{MF} = \bar{q}(i\slashchar{\partial}-M)q+\tilde{\mu}_q\bar{q}\gamma^0 q-U~,\] where \[M=m+2\frac{g_{20}}{\Lambda^2}\langle\bar{q}q\rangle+4\frac{g_{40}}{\Lambda^8}\langle\bar{q}q\rangle^3-2\frac{g_{22}}{\Lambda^8}\langle\bar{q}q\rangle\langle q^\dag q\rangle^2~,\] \[\tilde{\mu}_q = \mu_q-2\frac{g_{02}}{\Lambda^2}\langle q^\dag q\rangle-4\frac{g_{04}}{\Lambda^8}\langle q^\dag q\rangle^3-2\frac{g_{22}}{\Lambda^8}\langle\bar{q}q\rangle^2 \langle q^\dag q\rangle~, \label{eq:consmass}\] and the classical potential \[\begin{split} U &= \frac{g_{20}}{\Lambda^2}\langle\bar{q}q\rangle^2 + 3 \frac{g_{40}}{\Lambda^8}\langle\bar{q}q\rangle^4-3\frac{g_{22}}{\Lambda^8}\langle\bar{q}q\rangle^2 \langle q^\dag q\rangle^2\\ &-\frac{g_{02}}{\Lambda^2}\langle q^\dag q\rangle^2-3\frac{g_{04}}{\Lambda^8}\langle q^\dag q\rangle^4 ~. \end{split}\] Integrating out the quark degrees of freedom, the full thermodynamic potential takes the following form \[\begin{split} \Omega &= U-2 N_f N_c\int \frac{d^3 p}{(2\pi)^3} \Bigl\{E + T\log[1+e^{-\beta(E-\tilde{\mu}_q)}]\\ &+ T\log[1+e^{-\beta(E+\tilde{\mu}_q)}]\Bigr\} + \Omega_0~, \end{split} \label{eq:pot}\] where \(E = \sqrt{\mathbf{p}^2+M^2}\) and \(\beta=1/T\) and where \(\Omega_0\) ensures zero pressure in the vacuum. The model is solved by minimizing the thermodynamic potential with respect to the mean-fields \(X=\langle\bar{q}q\rangle, \langle q^\dag q\rangle\), i. e. \[\frac{\partial \Omega}{\partial X} = 0~.\] In this work we use the parameter set of Ref.  \(g_{20} = 1.864\), \(g_{40} = 11.435\), \(m=5.5\) MeV \(\Lambda=631.5\) MeV. The vector channel strengths are treated as free parameters, quantified by \[\eta_2 = \frac{g_{02}}{g_{20}} \,, \qquad \eta_4 = \frac{g_{04}}{g_{40}}~.\] We are interested in a particular region of the vector channel couplings where the \(g_{02}\) coupling is kept small, while \(g_{04}\) is increased. The reason behind our choice is as following. Since heavy hybrid stars require a stiff EoS, a repulsive vector coupling should be present. As the vector coupling renormalizes the chemical potential, it delays the onset of quark matter. This leads to a scenario where the hadronic mantle becomes too large for the pressure in the quark core to be able to hold the star against gravitational collapse. Therefore, the appearance of quark matter in such a scenario usually makes the star unstable. An attractive channel like e. g. superconductivity needs to be invoked in order to lower the onset. We point out that an alternative microscopic picture is possible with multiquark interactions: whereas stiffening of the EoS is provided by the 8-quark vector channel interaction, lowering of the onset is accomplished by introducing a sizeable 8-quark scalar channel interaction and keeping the 4-quark vector channel interaction small. Our choice of the phenomenologically interesting parameter space will have an effect of stiffening the quark matter at higher densities, while at the same time keeping the transition density low. Due to this restriction we will also put \(g_{22}=0\) by hand. Namely, the operator controlled by the size of \(g_{22}\) can be important at moderate \(\mu_q\) only if the vector mean-field is sizeable, which will not be the case for the parameter region of small \(g_{02}\) in which we are interested in. In addition, it will not be important at high \(\mu_q\) since there the scalar mean-field is zero. # Equation of state By evaluating the full thermodynamical potential ([\[eq:pot\]](#eq:pot){reference-type="ref" reference="eq:pot"}) at the minimum, we obtain the quark matter EoS as \(p_q =-\Omega\). The quark number density \(n_q\) and energy density \(\epsilon_q\) are defined as follows \[n_q=-\frac{\partial p_q}{\partial \mu_q}\,, \qquad \epsilon_q =-p_q+n_q\mu_q~.\] Beta equilibrium is taken into account by the weak processes \(d\to u+e^-+\bar{\nu}_e\), \(\mu^-\to e^-+ \nu_\mu +\bar{\nu}_e\), implying \(\mu_u = \mu_d + \mu_e\), \(\mu_\mu=\mu_e\), where the neutrino chemical potential is set to zero. Finally, the baryon chemical potential, and the baryon density are \(\mu_B = 3\mu_q = 2\mu_d+\mu_u\) and \(n_B=n_q/3\), respectively. For the nuclear matter we choose the DD2 EoS. The transition from nuclear to quark matter is provided by the traditional Maxwell construction. Therefore, the EoS is obtained by requiring local charge neutrality. The full pressure in the quark phase takes into account the contribution of electrons and muons \[p(\mu_q) = p_u(\mu_u)+p_{d}(\mu_d) + p_e(\mu_e)+p_\mu(\mu_\mu)~,\] where \(p_{e,\mu}\) is the pressure of a free electron (muon) gas. We emphasize that the Maxwell construction is merely the limit of large surface tension at the quark-hadron interface in a more elaborate approach that takes into account finite size effects, see. The following results ultimately depend also on the choice of the construction of the phase transition. The hybrid EoS are given on Fig. [\[fig:eos1\]](#fig:eos1){reference-type="ref" reference="fig:eos1"} for \(\eta_2=0.05\) and a range of \(\eta_4\). Owing to 8-quark scalar interactions, and small 4-quark vector interactions, onset of quark matter is rather low: for \(\eta_2=0.05\) and \(\eta_4=0.0\) it is around \(\epsilon\simeq 500\) MeV/fm\(^3\). Even a drastic increase of \(\eta_4\) has barely any influence on the onset: this is only natural, since there is extra suppression due to high dimensionality of the corresponding operator. The influence of \(\eta_4\) is best seen by inspecting the speed of sound: while small or almost vanishing vector interactions yield the relativistic value \(c_s^2 \simeq 1/3\), the speed of sound significantly increases with \(\epsilon\) already for \(\eta_4=1.0\), see right panels of Fig.  [\[fig:eos1\]](#fig:eos1){reference-type="ref" reference="fig:eos1"}. This gradual stiffening of the EoS can also be seen as a microscopic mechanism of a postulated scenario of a medium-dependent parameter \(\eta_2\), which is able to provide a very stiff quark EoS. Assuming that the vector mean-field is given by the density of massless fermions, an approximative expansion in \(g_{02}\) and \(g_{04}\) for the speed of sound \(c_s^2=\partial p/\partial \epsilon\) as a function of the quark chemical potential can be shown to hold \[\begin{split} c_s^2 &\simeq \frac{1}{3}+\frac{32g_{02}}{6\pi^2}\frac{\mu_q^2}{\Lambda^2}+ \frac{512g_{02}^2}{4\pi^4}\frac{\mu_q^4}{\Lambda^4}+ \frac{75776 g_{02}^3}{24\pi^6}\frac{\mu_q^6}{\Lambda^6}\\ &+\frac{3768320 g_{02}^4}{48\pi^8}\frac{\mu_q^8}{\Lambda^8}+ \frac{8192g_{04}}{48\pi^8}\frac{\mu_q^8}{\Lambda^8}~, \end{split} \label{eq:sos}\] illustrating that the \(g_{04}\) term starts to be important only at higher chemical potentials. While the 4-quark vector interaction respects causality, strong 8-quark vector interaction may violate the causal limit, see Ref.  for an explicit example in the nucleonic NJL model. This is the reason why the EoS with \(\eta_4=5.0\) turns acausal already at \(\epsilon\sim 2000\) MeV/fm\(^3\). In addition, for such high densities, the quark chemical potential is approaching \(\Lambda\) where the model should not be trusted anymore. Therefore, results obtained for this extreme scenario are given for illustrative purposes. # Hybrid stars The static, spherically symmetric stars are obtained as solutions of the Tolman-Oppenheimer-Volkoff (TOV) equations for the EoS shown in the previous section. On Fig. [\[fig:mrec1\]](#fig:mrec1){reference-type="ref" reference="fig:mrec1"} we display the resulting mass-radius and mass-central energy density diagrams for \(\eta_2=0.05\). Due to the early onset of quark matter, on Fig. [\[fig:mrec1\]](#fig:mrec1){reference-type="ref" reference="fig:mrec1"} we are able to obtain stable stars with pure quark matter in their cores even for very small vector coupling. Such a scenario is not easy to achieve in the NJL model with only fourth order scalar and vector operators, see however. The strong vector interaction increases the speed of sound and makes quark matter stiff, but at the same time it appears at too high energy densities. Higher dimensional vector operator stiffens the EoS and without influencing the onset significantly, gives a mechanism for \(2 M_\odot\) hybrid stars, see Fig. [\[fig:mrec1\]](#fig:mrec1){reference-type="ref" reference="fig:mrec1"}. In order to cover the present experimental window provided by PSR J1614-2230 and PSR J0348-0432 we have found that values \(\eta_4\) up to \(\eta_4\sim 2.0\) are sufficient. The extreme case with \(\eta_4=5.0\) leads to a significant increase in the mass, yielding \(M\sim 2.25 M_\odot\). # Conclusions Observing heavy compact stars offers a promising perspective on constraining the cold, dense EoS beyond saturation density. A compelling discrimination of many possible scenarios of dense matter will be possible once precise measurements of also star radii become available. The data of both masses and radii will enable Bayesian inversion of the TOV equation leading to the EoS. We have studied one possible scenario where multi-quark interactions coming from higher dimensional operators in the NJL model might play an important role at large densities. A sizeable 8-quark scalar channel is introduced in order to achieve a low onset. With a small 4-quark vector coupling onset is still low, while the relatively large 8-quark coupling is used to gradually stiffen the EoS at high densities. Within this parameter space we were able to fulfill and go beyond the \(2M_\odot\) constraint. The scenario of small 4-quark vector coupling is in accordance with the results of Ref. , but a more thorough study is needed to reveal how would the introduction of the 8-quark vector coupling influence their results. In addition, we stress that the consensus concerning the status of vector interactions in quark matter is yet to be reached in the community, as some other studies advocate a different scenario where the 4-quark vector coupling is sizeable. Using the results from the lattice QCD measurements at finite \(T\) and at imaginary chemical potential a more thorough study of the model presented here is needed to achieve better constraints on the three vector channel couplings, and to investigate the consequences at large densities. Finally, it remains to be explored what can such a setup say about the existence of strangeness in compact stars.

{'timestamp': '2014-07-14T02:01:06', 'yymm': '1401', 'arxiv_id': '1401.5380', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5380'}

# Three basic states of a compact star A magnetized, rotating compact star can be observed in one of the following three basic states: *ejector* (spin-powered pulsar), *propeller*, and *accretor*. All of these states are observationally confirmed for neutron stars (see Table [\[tb1\]](#tb1){reference-type="ref" reference="tb1"}). The white dwarfs in the accretor state are observed in Cataclysmic Variables (CVs). The low and moderately magnetized white dwarfs (with the surface field \(B \leq 1\) MG) in interacting close binary systems usually accrete material from a Keplerian disk while strongly magnetized (\(B_* \sim 10-1000\) MG) white dwarfs, which are refereed to as Polars, accrete material from an accretion channel. The only white dwarf resembling a spin-powered pulsar is currently identified with the degenerate companion of a peculiar CV AE Aquarii[@Ikhsanov-etal-2004; @Terada-etal-2008; @Ikhsanov-Beskrovnaya-2012]. [\[tb1\]]{#tb1 label="tb1"} # White dwarfs in the ejector state Observations show strong magnetization (up to a few \(\times 10^9\) G) and fast rotation (up to a period of 30 s) of white dwarfs not to be very unusual. This indicates that an existence of a strongly magnetized fast rotating white dwarf cannot be excluded. One of these objects (a white dwarf with the spin period of 33 s and the surface field of \(\sim 50\) MG) is discovered in AE Aquarii. The spin-down power of this white dwarf exceeds its bolometric luminosity by a factor of 300 and can be explained in terms of the spin-powered pulsar energy-loss mechanism provided the dipolar magnetic moment of the star is \(\sim 10^{34}\,{\rm G\,cm^3}\). The fast rotating extended magnetosphere in this case prevents the surrounding material from reaching the stellar surface. The interaction between the surrounding material and the stellar magnetic field leads to a formation of a stream which is flowing out from the binary system. The spin-down power of the white dwarf is released predominantly in a form of the wind of relativistic particles[@Ikhsanov-1998; @Ikhsanov-Beskrovnaya-2012]. Thus, the white dwarfs in the ejector state do exist but how could they forme? # Accretion-driven spin-up As the surface temperature of the white dwarf in AE Aquarii is limited to \(T_{\rm av} \leq 16\,000\) K its age \(t_{\rm cool} > 10^8\) yr turns out to be significantly larger than the spin-down timescale \(t_{\rm sd} \leq 2 \times 10^7\) yr. This indicates that the ejector white dwarf is a product of the binary evolution which included an epoch of its rapid spin-up caused by intensive accretion. The parameters of the spin-up epoch are the following:\ i) accretion from a Keplerian disk at a rate \(\dot{M}_{\rm pe} \geq 10^{-7}\,{\rm M_{\hbox{\)\odot\(}}\,yr^{-1}}\) on a timescale of a few million years;\ ii) screening of the magnetic field of the white dwarf by the accreted material by a factor of 50--100;\ iii) re-emerging of the magnetic field due to diffusion of the stellar magnetic field through the layer of the accreted material on a timescale \(t_{\rm em} < t_{\rm sd}\), where \(t_{\rm sd} = P_{\rm s}/2 \dot{P}\) is the spin-down timescale which for the parameters of AE Aquarii is \(\sim 2\) Myr. For this condition to be satisfied the total mass accreted by the white dwarf during the spin-up epoch should not exceed \(\Delta M_{\rm a} \leq 0.009\,M_{\hbox{\)\odot\(}}\) and the mass of the white dwarf itself is \(M_{\rm wd} \geq 1.1\,M_{\hbox{\)\odot\(}}\). This implies that the inclination angle of the binary system is \(\leq 54^{\hbox{\)^\circ\(}}\). An analysis of this accretion scenario suggests that the more massive the white dwarf is the shorter is the spin period which it can reach at the end of the spin-up epoch. Furthermore, the more massive a white dwarf is the smaller amount of the accreted material is required to spin it up to the shortest possible spin period. The timescale of the magnetic field re-emerging of the white dwarf after the spin-up epoch is \(\propto \Delta M_{\rm a}^{7/5}\). Therefore, the more massive the white dwarf is the faster its magnetic field is re-generated. This increases a probability for the white dwarf to appear as a star in the ejector state.

{'timestamp': '2014-01-22T02:10:38', 'yymm': '1401', 'arxiv_id': '1401.5374', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5374'}

# Introduction {#Sect:1} There are various ways of obtaining information about the presence of magnetic fields in stellar atmospheres like Ca II H & K or X-ray measurements. But the most direct way is certainly provided via the Zeeman effect, i.e., the magnetically induced line splitting and the associated line polarization. This has opened the way for essentially two different approaches to measuring the magnetic fields of cool stars. Unpolarized Zeeman broadening measurements from Stokes \(I\) profiles has led to a large amount of information about the magnetic fields of cool stars over the past decades. Because the Zeeman splitting (and broadening) is directly proportional to the magnetic field strength, Zeeman broadening measurements provide good estimates of the underlying average surface field. In the stellar astrophysical literature, the term magnetic flux for the product of surface filling factor and magnetic field strength (\(f \cdot B\)) is often used, but we prefer to denote it more accurately as the mean or average magnetic field strength, since \(f \cdot B\) is neither the (relative) magnetic flux through the stellar surface nor a measure of the magnetic flux density. Polarization or spectropolarimetric measurements, on the other hand, provide additional information about the orientation of the field vector and therefore allow, in principle, measurement of real flux densities. There are at least two obstacles that may prevent the full applicability of spectropolarimetric measurements to cool stars. First, the magnetic fields are relatively small in size and magnitude, which leads to very weak circular polarization (Stokes \(V\)) and even lower linear polarization (Stokes Q and U) signals. Second, the contributions from different polarities tend to mutually cancel out the circular polarization signal. A circumstance that mitigates the situation to some extent is rapid rotation, which partly lifts the mutual cancellation of the Stokes \(V\) spectra. This has led to the so-called Zeeman-Doppler imaging (ZDI) or Magnetic-Doppler imaging (MDI) methods, which allows reconstructing, in an inversion approach, the entire surface distribution of the magnetic field vector from phase-resolved spectropolarimetric observations. This approach has contributed significantly to our understanding of cool stars magnetic fields and stellar activity in general in recent years . However, what is good for the ZDI approach, namely rapid rotation, has an adverse effect on the Zeeman broadening approach and vice versa, such that both approaches are complementary to some degree; see also for a comprehensive overview of measuring stellar magnetic fields on cool stars. The present paper focuses on spectropolarimetric observations that do not deliver a series of phase-resolved spectra but rather single snapshot-like observations from either rapidly or slowly rotating active stars. This is the typical situation for stars with long rotation periods and/or in situations where the amount of telescope time is limited. Such observations are responsible for a number of exciting magnetic field detection in recent times. Because many of the observed polarimetric line profiles are buried in noise, the investigations rely on a prior line profile reconstruction technique, such as the least-square-deconvolution (LSD) , principal-component-analysis (PCA), or singular-value-decomposition (SVD). However, even after these preprocessing steps, the extracted line profiles exhibit a fraction of noise that makes their interpretation not always straightforward in terms of a reliable detection and magnetic field estimation. Inspired by the work of who propose a compressive sensing framework for spectropolarimetry, we utilize the compressibility of polarimetric signals to approximate the observed Stokes \(V\) profiles by a sparse decomposition with a wavelet frame. This sparse linear representation facilitates a simple and noise-robust magnetic detection and estimation method.[^1] For rapidly rotating stars, the sparse approximation of the observed signal with basis functions allows not only the effective mean longitudinal magnetic field to be measured but also the absolute value of the resolved (i.e. apparent) longitudinal magnetic field. The paper is organized as follows. In Sect. [2](#Sect:2){reference-type="ref" reference="Sect:2"} we describe the weak-field approximation and derive the center-of-gravity method for disk-integrated observations to define the effective longitudinal magnetic field. Furthermore, we also define the apparent longitudinal magnetic field that describes the disk-integrated absolute value of the longitudinal magnetic field component as measured from the net absolute circular polarization of the Stokes \(V\) profile. In Sect. [3](#Sect:3){reference-type="ref" reference="Sect:3"} we introduce the concept of a sparse Stokes profile approximation and magnetic field detection. We describe in some detail the orthogonal matching pursuit (OMP) algorithm and the signal dictionary, which provides the building blocks for the Stokes profile approximation. At the end of this section, we combine the approximation method with a detection algorithm to obtain reliable estimates for the magnetic quantities. Section [4](#Sect:4){reference-type="ref" reference="Sect:4"} gives an illustration of the magnetic OMP method and highlights the benefit of the complementary definition of the effective and apparent longitudinal magnetic field. In Sect.[5](#Sect:5){reference-type="ref" reference="Sect:5"} we investigate the sparsity of Stokes profiles and demonstrate that with the described dictionary a sparse approximation of the Stokes profile is possible for a wide parameter range. An extensive numerical assessment and evaluation of the accuracy of the magnetic OMP method are also performed in Sect. [5](#Sect:5){reference-type="ref" reference="Sect:5"}, along with a comparison with the conventional center-of-gravity method. A summary is presented in Sect. [6](#Sect:6){reference-type="ref" reference="Sect:6"}. # Magnetic field estimation {#Sect:2} Our starting point will be the so-called center-of-gravity (COG) method. Before we describe the COG method, we briefly introduce the concept of the weak-field approximation from which we later derive the effective and apparent longitudinal magnetic field. ## Disk-Integrated weak-field approximation {#Sect:2.1} The *local* weak-field-approximation (WFA) makes the assumption that the atmosphere of a resolved region is permeated by a homogeneous, weak, and depth-independent magnetic field. Then, a proportionality exists between the observed local Stokes \(V\) profile and the spectral derivative of the observed local Stokes \(I\) profile, which can be derived from the polarized transfer equation using a first-order Taylor expansion of the line profile function. This proportionality or linear relation can be written as \[V(\lambda) =-g \lambda_B \cos\gamma \frac{dI_{0}(\lambda)}{d\lambda} \: , \label{Eq:2.1.1}\] where \(\gamma\) is the angle between magnetic field vector and the line-of-sight (LOS), \(g\) the effective Landé factor, and \(\lambda_B\) the Zeeman splitting expressed in wavelength units is given by \[\lambda_B \: = \: 4.67 \times 10^{-13} g_{\rm eff}\: \lambda_0^2 \: B \:. \label{Eq:2.1.1b}\] For the sake of brevity, we denote by \(\lambda\) the relative wavelength shift from the line center \(\lambda_0\) hereafter and assume that the Stokes profiles are normalized by the local continuum intensity. We moreover introduce the variable \(\mu = \cos\gamma\) and \(\alpha =-4.67 \times 10^{-13} g_{\rm eff} \lambda_0^2\) to write Eq. ([\[Eq:2.1.1\]](#Eq:2.1.1){reference-type="ref" reference="Eq:2.1.1"}) as \[V(\lambda) = \alpha B \mu \frac{dI_{0}(\lambda)}{d\lambda} \: = \: \alpha B_l \frac{dI_{0}(\lambda)}{d\lambda}\: , \label{Eq:2.1.2}\] where \(B_l\) denotes the longitudinal component of the magnetic field vector. The proportionality between the observed local Stokes \(V\) and the observed local Stokes \(I\) profile depends on the wavelength. In the local case, i.e. resolved (solar) observations, this dependency is usually neglected since the same scaling factor, i.e. \(\alpha B \mu\), applies for all wavelengths. This condition is not generally true for unresolved stellar observations. Given that the assumptions for the weak-field approximation are applicable, it is readily seen that for the local and the resolved case, one can obtain an estimate for the longitudinal component \(B_l\) of the magnetic field through \[B_l = \frac{V(\lambda)}{\alpha \; dI_{0}(\lambda) / d\lambda} \;. \label{Eq:2.1.3}\] For each spectral line, a good estimate up to which magnetic field strength the WFA remains a valid approximation, can be deduced from the so-called Zeeman saturation. This is the regime from which on the Stokes \(V\) amplitudes no longer grow linearly but instead the individual sigma components begin to shift apart according to Eq. [\[Eq:2.1.1b\]](#Eq:2.1.1b){reference-type="ref" reference="Eq:2.1.1b"}. The Zeeman saturation regime is reached for the majority of Zeeman sensitive spectral lines in the optical at around one kilo-Gauss. In stellar spectropolarimetric observations, the only observables are the disk-integrated Stokes profiles. The obvious requirement for applying the weak-field approximation to stellar spectra is the existence of a wavelength independent linear scaling between the disk-integrated Stokes \(V\) profile, denoted hereafter as \(\tilde{V}\), and the spectral derivative of the disk-integrated Stokes \(\tilde{I}\). For the following disk integration, we choose a Cartesian coordinate system where the z-axis points along the LOS, the y-axis is in the plane defined by the LOS and the rotation axis, and the x-axis is perpendicular to that plane. Using \(\Delta \lambda_{Rot} = (\lambda_0 \textrm{\textsl{v}} \sin i) / c\), the *disk-integrated* weak-field approximation for a rotating star can then be written in its general form as \[\begin{aligned} \tilde{V}(\lambda)= \int\limits_{-1}^{+1} \int\limits_{-\sqrt{1-x^2}}^{+\sqrt{1-x^2}} V(B;\lambda-\Delta\lambda_{Rot}x;x,y) \: dx \: dy \: = \hspace{1.3cm} \nonumber \\ = \int\limits_{-1}^{+1} \int\limits_{-\sqrt{1-x^2}}^{+\sqrt{1-x^2}} \alpha B_l(x,y) \: \frac{dI_{0}(\lambda-\Delta \lambda_{Rot}x;x,y)}{d\lambda} \: dx \: dy , \label{Eq:2.1.4} \end{aligned}\] where we introduced the spatial-dependent longitudinal magnetic field strength \(B_l(x,y) = B(x,y) \; \mu(x,y)\). The observed Stokes \(\tilde{V}\) vector depends on the magnetic field, whereas the local intensity has no explicit dependency on the magnetic field under the weak-field assumption. The intensity profile originates over the entire observable disk. This disconnection, in general, prevents applying the WFA to unresolved stellar observation because the disk-integrated profile shape of the spectral derivative of Stokes \(\tilde{I}\) does not coincide with the observed Stokes \(\tilde{V}\) profile. However, the assumption that the thermodynamic parameters in the atmosphere and across the observable disk do not change between magnetic and field-free regions, as well as the assumption that the longitudinal magnetic field on the stellar surface has no explicit spatial dependence, makes the WFA applicable also to disk-integrated spectra. In this case one can obtain a similar relation to the one in Eq. ([\[Eq:2.1.3\]](#Eq:2.1.3){reference-type="ref" reference="Eq:2.1.3"}) by defining a *mean* longitudinal field strength \(\langle B_l \rangle,\) which allows one to separate the magnetic field from the integral in Eq. ([\[Eq:2.1.4\]](#Eq:2.1.4){reference-type="ref" reference="Eq:2.1.4"}). The disk-integrated WFA can then compactly written as \[\tilde{V}(\lambda) \: = \: \alpha \: \langle B_l \rangle \: \frac{d\tilde{I}(\lambda)}{d\lambda}\:. \label{Eq:2.1.5}\] This linear scaling relation must hold for all wavelengths \(\lambda\). If this is not the case, owing to a spatial dependence of the underlying magnetic field and rapid rotation, the WFA can no longer be applied. The applicability of the disk-integrated WFA is also limited to fields below the Zeeman saturation regime. Because the disk integration is a linear operation, the Zeeman saturation regime, in general, is determined not by the *mean* longitudinal field but by the *local* field strengths on the surface of the star. ## The effective longitudinal magnetic field {#Sect:2.2} The center-of-gravity method does in fact bypass the limitation that the observable Stokes \(\tilde{V}\) profile must be proportional to the spectral derivative of the Stokes \(\tilde{I}\) profile by using the first-order moment of the Stokes \(\tilde{V}\) profile. This first-order moment of the Stokes \(V\) signal can be obtained by using Eq. ([\[Eq:2.1.4\]](#Eq:2.1.4){reference-type="ref" reference="Eq:2.1.4"}), \[\begin{aligned} \int\limits_{\Lambda} \tilde{V}(\lambda) \: \lambda \: d\lambda \: = \hspace{6.0cm} \nonumber \\ = \int\limits_{\Lambda} \int\limits_{-1}^{+1} \int\limits_{-\sqrt{1-x^2}}^{+\sqrt{1-x^2}} \alpha B_l(x,y) \: \frac{dI_{0}(\lambda-\Delta \lambda_{Rot}x;x,y)}{d\lambda} \: dx \: dy \: \lambda \: d\lambda \: , \label{Eq:2.2.1} \end{aligned}\] where \(\Lambda\) denotes the integration limits that cover the entire line profile. Instead of deriving the COG method from the weak-line limit we use the WFA as a starting point. Under the WFA, the profile of a spectral line does not alter its shape, and we may introduce a line-dependent intensity distribution function \(\eta\), which accounts for geometrically induced radiative transfer effects (e.g., limb darkening and other possible temperature effects). We then write, for the Stokes \(I\) profile at position \(x,y\) on the stellar disk, \[I(\lambda;x,y) \; = \; \eta(x,y) \; I_0(\lambda;\mu = 1) \;. \label{Eq:2.2.2}\] Substituting this into Eq. ([\[Eq:2.2.1\]](#Eq:2.2.1){reference-type="ref" reference="Eq:2.2.1"}) gives \[\begin{aligned} \int\limits_{\Lambda} \tilde{V}(\lambda) \: \lambda \: d\lambda \: = \hspace{6.0cm} \nonumber \\ = \int\limits_{-1}^{+1} \int\limits_{-\sqrt{1-x^2}}^{+\sqrt{1-x^2}} \alpha \: B_l(x,y) \: \eta(x,y) \: \frac{dI_{0}(\lambda-\Delta \lambda_{Rot}x)}{d\lambda} \: dx \: dy \: \lambda \: d\lambda \:. \label{Eq:2.2.3} \end{aligned}\] By taking advantage of the Cartesian coordinate system, where values along the x-axis experience different degrees of Doppler shifts \(\Delta \lambda\), we can perform the following change of variables, \(x = \Delta \lambda / \Delta \lambda_{Rot}\), to write \[\begin{aligned} \int\limits_{\Lambda} \tilde{V}(\lambda) \: \lambda \: d\lambda \: = \: \hspace{6.0cm} \nonumber \\ = \: \alpha \int\limits_{\Lambda} \int\limits_{-\Delta\lambda_{Rot}}^{+\Delta\lambda_{Rot}} \int\limits_{-\sqrt{1-(\Delta\lambda/ \Delta\lambda_{Rot})^2}}^{+\sqrt{1-(\Delta\lambda/ \Delta\lambda_{Rot})^2}} B_l(\Delta \lambda / \Delta \lambda_{Rot},y) \times \hspace{1.0cm} \nonumber \\ \times \: \eta(\Delta \lambda / \Delta \lambda_{Rot},y) \frac{dI_{0}(\lambda-\Delta\lambda)}{d\lambda} \: dy \: d(\Delta \lambda) \: \lambda\: d\lambda \:.\hspace{1.0cm} \label{Eq:2.2.4} \end{aligned}\] We define the total flux-weighted longitudinal magnetic field distribution \(B_l^{\eta}\) along a small a strip \(dx\) or \(d(\Delta \lambda)\) of equal velocity as \[\begin{aligned} B_l^{\eta}(\Delta \lambda) = \int\limits_{-\sqrt{1-(\Delta\lambda/ \Delta\lambda_{Rot})^2}}^{+\sqrt{1-(\Delta\lambda/ \Delta\lambda_{Rot})^2}} B_l(\Delta \lambda / \Delta \lambda_{Rot},y) \: \eta(\Delta \lambda / \Delta \lambda_{Rot},y) \: dy. \label{Eq:2.2.5} \end{aligned}\] This allows us to write Eq. ([\[Eq:2.2.4\]](#Eq:2.2.4){reference-type="ref" reference="Eq:2.2.4"}) as a convolution integral \[\begin{aligned} \int\limits_{\Lambda} \tilde{V}(\lambda) \: \lambda \: d\lambda \: = \hspace{5.5cm} \nonumber \\ = \alpha \int_{\Lambda} \int\limits_{-\Delta\lambda_{Rot}}^{+\Delta\lambda_{Rot}} B_l^{\eta}(\Delta \lambda) \: \frac{dI_{0}(\lambda-\Delta\lambda)}{d\lambda} \: d(\Delta \lambda) \: \lambda \: d\lambda \: , \label{Eq:2.2.6} \end{aligned}\] or in a more compact way \[\begin{aligned} \int\limits_{\Lambda} \tilde{V}(\lambda) \: \lambda \: d\lambda \: = \: \alpha \int\limits_{\Lambda} \frac{d}{d\lambda} \left ( B_l^{\eta} \ast I_{0} \right ) \: \lambda \: d\lambda \:. \label{Eq:2.2.7} \end{aligned}\] The disk-integrated Stokes \(\tilde{I}\) profile is also given by the convolution \[\tilde{I} \: = \: \int\limits_{-\Delta\lambda_{Rot}}^{+\Delta\lambda_{Rot}} \eta(\Delta \lambda) \: I_0(\lambda-\Delta\lambda) \: d(\Delta \lambda) \: , \label{Eq:2.2.8}\] where \(\eta(\Delta \lambda)\) is defined by \[\begin{aligned} \eta(\Delta \lambda) = \int\limits_{-\sqrt{1-(\Delta\lambda/ \Delta\lambda_{Rot})^2}}^{+\sqrt{1-(\Delta\lambda/ \Delta\lambda_{Rot})^2}} \eta(\Delta \lambda / \Delta \lambda_{Rot},y) \: dy. \label{Eq:2.2.9} \end{aligned}\] With this definition we may separate a mean longitudinal magnetic field from Eq. ([\[Eq:2.2.5\]](#Eq:2.2.5){reference-type="ref" reference="Eq:2.2.5"}) that we call the effective longitudinal magnetic field \(B_{\rm eff}\). Using Eqs. ([\[Eq:2.2.8\]](#Eq:2.2.8){reference-type="ref" reference="Eq:2.2.8"}) and ([\[Eq:2.2.9\]](#Eq:2.2.9){reference-type="ref" reference="Eq:2.2.9"}), we can rearrange Eq. ([\[Eq:2.2.6\]](#Eq:2.2.6){reference-type="ref" reference="Eq:2.2.6"}) to obtain the following expression for the effective longitudinal magnetic field \[B_{\rm eff} = \frac{\int_{\Lambda} \tilde{V}(\lambda) \: \lambda \:d\lambda} {\alpha \int_{\Lambda} \left ( d\tilde{I}(\lambda) / d\lambda \right ) \: \lambda \: d\lambda} \;. \label{Eq:2.2.10}\] Performing the integration by parts in the denominator and keeping in mind that we implicitly deal with normalized profiles, we obtain \[B_{\rm eff} = \frac{\int_{\Lambda} \tilde{V}(\lambda) \: \lambda \:d\lambda} {\alpha W_I} \; , \label{Eq:2.2.11}\] where \(W_I\) is the equivalent width of the disk-integrated Stokes \(I\) profile. This is the COG method for disk-integrated observations introduced for solar observations by. Making a transformation of the *relative* wavelength according to \(\lambda \rightarrow \textrm{\textsl{v}} \lambda_0 / c\), we obtain the following equation for the velocity domain \[B_{\rm eff} = \frac{\lambda_0 \: \int_{V} \tilde{V}(\textrm{\textsl{v}}) \: \textrm{\textsl{v}} \: d\textrm{\textsl{v}}} {c \: \alpha \: W_I} =-2.14 \times 10^{12} \: \frac{\int_{V} \tilde{V}(\textrm{\textsl{v}}) \: \textrm{\textsl{v}} \: d\textrm{\textsl{v}}} {c \lambda_0 g \: W_I} \; , \label{Eq:2.2.12}\] where the central wavelength \(\lambda_0\) is given in \(\AA\). This form is extensively used in spectropolarimetric observations where the spectral line profiles are often preprocessed by transforming them into the velocity domain during a reconstruction process. ## The apparent longitudinal magnetic field {#Sect:2.3} If we now consider the case where the projected rotational velocity of the star is sufficiently large and the effective longitudinal magnetic field is zero owing to a balance of positive and negative polarities, the disk-integrated Stokes \(V\) signal of the star does not necessarily cancel out and vanishes. In fact, the net absolute circular polarization; i.e., the integral of the absolute value of the observed Stokes \(V\) signal over wavelength is in general not zero. Even though we measure no effective field, the polarized profile is clearly detectable and *apparently* a magnetic field must be present. Of course, this is one of the effects that is utilized by ZDI to resolve surface magnetic fields on rapidly rotating stars. Having such a nonvanishing net absolute circular polarization means that the magnetic field exhibits a flux imbalance for each or some of the resolved surface areas on the stellar disk, even though the overall disk-averaged magnetic field is flux balanced. To characterize this imbalance, we define the apparent longitudinal magnetic field. We start by defining the integrated Stokes \(V^*\) along a small iso-radial velocity strip \(d(\Delta \lambda)\) as \[V^*(\Delta \lambda) \: = \: \int\limits_{-\sqrt{1-(\Delta\lambda/ \Delta\lambda_{Rot})^2}}^{+\sqrt{1-(\Delta\lambda/ \Delta\lambda_{Rot})^2}} V(\Delta \lambda / \Delta \lambda_{Rot},y) \: dy \:. \label{Eq:2.3.1}\] The disk-integrated first-order moment of the Stokes \(V\) profile can then be written as \[\int\limits_{\Lambda} \tilde{V}(\lambda) \: \lambda \: d\lambda \: = \: \int\limits_{\Lambda} \int\limits_{-\Delta\lambda_{Rot}}^{+\Delta\lambda_{Rot}} V^*(\Delta \lambda) \: d(\Delta \lambda) \: \lambda \: d\lambda \:. \label{Eq:2.3.2}\] The total amount of the effective longitudinal magnetic field can be expressed by \[B^{tot}_{\rm eff} \: = \: \frac{1}{\alpha \: W_I} \: \int\limits_{-\Delta\lambda_{Rot}}^{+\Delta\lambda_{Rot}} \left | \int\limits_{\Lambda} V^*(\Delta \lambda) \: \lambda \: d\lambda \right | \: d(\Delta \lambda)\:. \label{Eq:2.3.3}\] This is, of course, not an observable, but assuming that we have a number of \(n\) discrete and spectroscopically resolvable radial-velocity strips across the surface of the star that produce separated Stokes \(V\) profiles, we may write \[\begin{aligned} \frac{1}{\alpha \: W_I} \int\limits_{-\Delta\lambda_{Rot}}^{+\Delta\lambda_{Rot}}\left | \int\limits_{\Lambda} V^*(\Delta \lambda) \: \lambda \: d\lambda \right | \: d(\Delta \lambda)\: \: \geq \: \hspace{2.5cm} \nonumber \\ \geq \: \frac{1}{\alpha \: W_I} \sum\limits_{i=0}^{n} \left | \int\limits_{\Lambda} \hat{V}^*_i(\lambda) \: \lambda \: d\lambda \right | \: , \label{Eq:2.3.4} \end{aligned}\] where \(\hat{V}^*_i\) is the Stokes \(V\) profile originating in the \(i\)-th resolved radial-velocity strip. In this case the observed Stokes \(V\) profile is the composition of \(n\) individual contributions coming from different velocity-resolved regions. Using Eqs. ([\[Eq:2.3.4\]](#Eq:2.3.4){reference-type="ref" reference="Eq:2.3.4"}) and ([\[Eq:2.3.2\]](#Eq:2.3.2){reference-type="ref" reference="Eq:2.3.2"}), we can finally define the apparent longitudinal magnetic field as \[B_{\rm app} = \frac{\sum\limits_{i=0}^{n} \left | \int\limits_{\Lambda} \hat{V}^*_i(\lambda) \: \lambda \: d\lambda \right |} {\alpha W_I} \; , \label{Eq:2.3.5}\] or in velocity coordinates \[B_{\rm app} = \frac{\lambda_0 \: \sum\limits_{i=0}^{n} \left | \int_{V} \tilde{V}_i(\textrm{\textsl{v}}) \: \textrm{\textsl{v}} \: d\textrm{\textsl{v}} \right |} {c \: \alpha \: W_I} \:. \label{Eq:2.3.6}\] The apparent longitudinal magnetic field measures the maximum resolved field that can be obtained from the observation. In general, we have \(B^{tot}_{\rm eff} \ge B_{\rm app}\), and the amount of the \(B_{\rm app}\) depends on the intrinsic line width of the local Stokes \(I\) profile. However, it can give an estimate for the distribution of the absolute effective field over the stellar disk. Although it is not clear in the first place how to obtain the apparent longitudinal field and the flux density distribution from the disk-integrated Stokes \(V\) profile, we present an easy and fast way for estimating this quantity in the next section. # Sparse approximation of Stokes profiles and magnetic field detection {#Sect:3} The basic idea of our proposed approach is to find an accurate approximation of the observed Stokes \(V\) profile by decomposing the original signal with a set of well chosen elementary waveforms. The stagewise approximation by these elementary basis functions allows us to apply the equations for the effective and apparent longitudinal magnetic field to each of these elementary functions and eventually obtain the desired magnetic quantities. The decomposition we describe in the following is based on a prescribed overcomplete set of elementary functions, called signal or profile atoms. Overcomplete here means that the number of elements provided by the dictionary is redundant, i.e. larger than the actual dimension of the signal space. In contrast to many nonredundant (e.g., orthogonal) transformations, a linear expansion of a given signal profile with an overcomplete set of signal atoms often facilitates a more efficient and sparse approximation. By using a suitable and redundant set of profile atoms, we show in the following that a sparse approximation of observed Stokes line profiles, allows a resolved analysis of the effective and apparent longitudinal magnetic flux density. ## Orthogonal matching pursuit {#Sect:3.1} The matching pursuit (MP) algorithm is an iterative algorithm for adaptive signal reconstruction and approximation. A given signal or line profile is decomposed into a linear expansion of dictionary elements. The actual realization of the dictionary is discussed in Sect. [3.2](#Sect:3.2){reference-type="ref" reference="Sect:3.2"}. For the moment it suffices to consider the dictionary as a collection of possibly linear-dependent elementary waveforms. We begin with a description of the original MP algorithm introduced by and then of its improvement called OMP, which we use for our analysis. Following let us define a dictionary \(D(\vec{x})\) as a collection of a large number of elementary wave functions \([\vec{x}_1,\vec{x}_2...,\vec{x}_n]\) formally defined in Hilbert space \(\mathcal{H}\) where each vector is of unit norm \(\| \vec{x}_i \| = 1\). A given signal \(f \in \mathcal{H}\) is approximated by MP in a first step by projecting \(f\) onto a vector \(x_l \in D\) such that \[f \: = \: \langle f,x_l \rangle x_l \: + \: Rf \: , \label{Eq:3.1.1}\] where \(Rf\) is the current residual of the approximation. Since the residual \(Rf\) of the current estimate is orthogonal to \(x_l\), we have the following energy conservation \[\| f \|^2 \: = \: | \langle f,x_l \rangle |^2 \: + \: \|Rf\|^2 \:. \label{Eq:3.1.2}\] To minimize \(\|Rf\|^2\) the vector \(x_l \in D\) is chosen such that the maximum projection (i.e., correlation) is found between the residual vector and all dictionary atoms, \[x_l \: = \: \arg\max_{x_k \in D} | \langle R f,x_k \rangle | \:. \label{Eq:3.1.3}\] The algorithm now iteratively decomposes the residual \(Rf\) by repeated projections onto the dictionary atoms. The recursive algorithm can be written in a compact way if we set the initial values for the residual \(R^0f = f\) and the initial approximation to \(f_0 = 0\), \[\begin{aligned} (I) \: \: \: x_{l_n} \: = \: \arg\max_{x_k \in D} | \langle R^nf,x_k \rangle | \:, \hspace{3.5cm} \nonumber \\ (II) \: \: f_{n+1} \: = \: f_n \: + \: \langle R^nf,x_{l_n} \rangle x_{l_n} \: , \hspace{3.3cm}\nonumber \\ (III) \: \: R^{n+1}f \: = \: R^nf-\: \langle R^nf,x_{l_n} \rangle x_{l_n} \:.\hspace{2.8cm} \label{Eq:3.1.4} \end{aligned}\] An increasingly closer approximation to the signal is obtained by repeating steps (I) to (III). A proof of convergence for the MP algorithm can be found in. The MP algorithm can be improved by realizing that the steps taken by the MP are not optimal in the sense that a newly chosen dictionary vector is not orthogonal to the previously selected vectors. This can be resolved by orthogonalizing the newly selected vector relative to the \(n-1\) previously selected vectors. The orthogonalization, which is essentially a Gram-Schmidt procedure, uses an auxiliary vector \(z\) that expresses the yet unexplained component of the newly selected vector \(x_{l}\). For the \(n\)-th iteration we may write \[z_{n} \: = \: x_{l_n} \:-\sum_{p=0}^{n-1} \frac{\langle x_{l_n},z_{p} \rangle z_{p}}{\| z_{p} \|^2} \: , \label{Eq:3.1.5}\] where the initial condition is set to \(z_{0} = x_{l_0}\). The new residual \(R^{n+1}f\) can then be expressed by the projection of the current residual \(R^{n}f\) on \(z_{n}\), \[R^{n+1}f \: = \: R^nf-\: \frac{\langle R^nf,z_{n} \rangle z_{n}}{\| z_{n} \|^2} \: , \label{Eq:3.1.6}\] using the relation in Eq. ([\[Eq:3.1.5\]](#Eq:3.1.5){reference-type="ref" reference="Eq:3.1.5"}). The latter can also be written as \[R^{n+1}f \: = \: R^nf-\: \frac{\langle R^nf,x_{l_n} \rangle z_{n}}{\| z_{n} \|^2} \:. \label{Eq:3.1.7}\] Compared to the ordinary MP algorithm a component is now subtracted in a direction that is orthogonal to all previously selected vectors. Beginning with the initial conditions \(R^0f = f\) and \(f_0 = 0\) as well as \(z_{0} = x_{l_0}\) the recursive algorithm for the OMP can then be expressed as \[\begin{aligned} (Ia) \: \:x_{l_n} \: = \: \arg\max_{x_k \in D} | \langle R^nf,x_k \rangle | \:, \hspace{3.8cm} \nonumber \\ (Ib) \: z_{n} \: = \: x_{l_n} \:-\sum_{p=0}^{n-1} \frac{\langle x_{l_n},z_{p} \rangle z_{p}}{\| z_{p} \|^2} \: ,\hspace{3.7cm} \nonumber \\ (II) \: \:f_{n+1} \: = \: f_n \: + \: \frac{\langle R^nf,x_{l_n} \rangle z_{n}}{\| z_{n} \|^2} \: ,\hspace{3.7cm} \nonumber \\ (III) \: \: R^{n+1}f \: = \: R^nf-\: \frac{\langle R^nf,x_{l_n} \rangle z_{n}}{\| z_{n} \|^2} \:. \hspace{3.2cm} \label{Eq:3.1.8} \end{aligned}\] After \(n\) iterations we obtain the following sparse approximation of our signal, \[f \: \approx \: \sum_{n} \frac{\langle R^nf,x_{l_n} \rangle \: z_{n}}{\| z_{n} \|^2} \:. \label{Eq:3.1.8}\] A convergence analysis and more results for the OMP algorithm can be found in. The set of selected signal atoms \(\{ x_{l_0},x_{l_1}...,x_{l_n} \}\), its corresponding projection weights, and auxiliary vectors give us the opportunity to analyze any given signal in terms of a small number of dictionary basis functions. ## The wavelet dictionary for Stokes profile approximation {#Sect:3.2} To facilitate a sparse and compact representation of a given Stokes profile, we need a set of basis functions that closely resemble the individual building blocks of a disk-integrated circular polarized profile. These building blocks are the local Stokes \(V\) profiles originating in a small resolved surface area of the star. From the definition of a general spectral line profile, we know that a Gaussian function can provide a reasonably good approximation of a spectral absorption profile in cases where damping is not too strong. Under the regime of the WFA, where the Stokes \(V\) profile is proportional to the derivative of the intensity profile, it seems obvious that an appropriate and easy way to approximate a composite (i.e., disk-integrated) Stokes \(V\) profile can be achieved by using a dictionary of the derivatives of Gaussian functions. In fact, we use a wavelet frame constructed from scaled and translated versions of first derivatives of a Gaussian. We define the dictionary \(D\) as a set of elementary functions or atoms \(\Psi_{j,k}\) of unit norm as \[D \: = \: \{ \Psi_{j,k}, i,k \in Z \} \:. \label{Eq:3.2.1}\] The atoms are constructed by a set of scaled and translated versions of a mother wavelet \(\Psi\). For the mother wavelet we choose, as mentioned above, the first derivative of a Gaussian function, \[\Psi(x) \: = \:-\: \left ( \frac{2}{\sqrt{\pi}} \right )^{1/2} \: x \: e^{x^2/2} \:. \label{Eq:3.2.2}\] The wavelet function obtained by a scaled and translated version of the mother wavelet can be expressed in the wavelength domain as \[\Psi_{j,k}(\lambda) \: = \: s_{j}^{-\frac{1}{2}} \: \Psi \left ( \frac{\lambda-\lambda_k}{s_j} \right ) \: , \label{Eq:3.2.3}\] where \(s_j\) and \(\lambda_k\) denote the scaling and translation parameter. To obtain a discrete set of wavelet functions, we use the following discretization of the scale parameter \[s_j \: = \: s_0 2^{j \Delta_r } \:, \mbox{where} \: j=0,1...,L \: , \label{Eq:3.2.4}\] where \(s_0\) defines the smallest resolvable scale, which we set to \(2\delta \lambda\), twice the wavelength step of the observations or synthetic calculations. The parameter \(\Delta_r\) is a variable value that provides the resolution of the transform and which is set throughout this paper to \(\Delta_r = 0.125\). This provides a reasonable trade-off between resolution and computational cost. The largest scale \(L\) is determined according to \[L \: = \: \frac{ log_2(N \: \delta \lambda / j_0) } { \Delta_r } \: , \label{Eq:3.2.5}\] where \(N\) is the number of wavelength points in the line profile. With the wavelet function Eq. ([\[Eq:3.2.3\]](#Eq:3.2.3){reference-type="ref" reference="Eq:3.2.3"}), we can express the wavelet coefficients of an observed spectrum \(V(\lambda)\) as the dot product between the Stokes \(V\) profile and the wavelet function, \[w_{j,k} \: = \: \int_R V(\lambda) \: \Psi_{j,k}(\lambda) \: d\lambda \: = \: \langle V,\Psi_{j,k} \rangle \: , \label{Eq:3.2.5}\] where the integration limits \(R\) are large enough to cover the entire profile. The projections (i.e., dot products) of Eq. ([\[Eq:3.1.8\]](#Eq:3.1.8){reference-type="ref" reference="Eq:3.1.8"}) used in the OMP algorithm can be computed easily by calculating the convolution between the current residual and the wavelet function for all scales \(s_j\). The decomposition by a restricted set of wavelets with different scales provides a multiscale representation of our original signal profile. We use this multiscale decomposition to apply an approach called multiresolution support to complement the OMP algorithm with an efficient detection procedure. ## Magnetic field detection {#Sect:3.3} We first start with the assumption that our observation vector (i.e., spectral line profile) \(V\) contains the true signal vector \(S\) and additive noise \(N\) such that we may write, for the observed signal, \[V(\lambda) \: = \: S(\lambda) \: + \: N(\lambda) \:. \label{Eq:3.3.1}\] Separating the signal from noise can be cast in the framework of nonparametric signal estimation. One very successful way of reconstructing an unknown signal from noisy observation is thresholding introduced and theoretically investigated, e.g., by. For our detection analysis we take a similar thresholding approach that uses the multiscale decomposition of the signal to test the significance of every wavelet coefficient individually. The basic strategy of shrinkage or thresholding methods is to consider the observed signal in a transformed (e.g., wavelet) domain rather than in the original data domain. Similar to Fourier filtering, wavelet-based detections methods rest on the idea that the noise contribution exhibits a different statistical behavior in the transformed domain. The OMP algorithm approximates the observed line profile in an iterative process by projecting the dictionary atoms onto the current residual of the observation while picking the most correlated signal atom in each iterative cycle. Using a wavelet dictionary, the projections can be efficiently calculated by a convolution, i.e., wavelet transform of the observed line profile. These projections between an individual wavelet function of scale \(j\) at position \(k\) and the current residual of the observation \(V\) in the presence of noise can be written thanks to the linearity of the wavelet transform as \[\langle V,\Psi_{j,k} \rangle \: = \: \langle S,\Psi_{j,k} \rangle \: + \: \langle N,\Psi_{j,k} \rangle \: , \label{Eq:3.3.2}\] which we may write according to Eq. ([\[Eq:3.2.4\]](#Eq:3.2.4){reference-type="ref" reference="Eq:3.2.4"}) as the wavelet coefficient \(w_{j,k}\) \[w_{j,k}^V \: = w_{j,k}^S \: + \: w_{j,k}^N \:. \label{Eq:3.3.3}\] When including the position index \(k\) into a vector notation, this can be written in a compact way by using an operator or matrix notation \[\vec{W}_V(j) = \vec{W}_S(j) + \vec{W}_N(j) \: , \label{Eq:3.3.4}\] where \(\vec{W}_V\),\(\vec{W}_S\), and \(\vec{W}_N\) represent the convolution over the wavelength index \(k\) on a specific scale \(j\). The expectation value for the wavelet transform of the noise contribution is \(E\{\vec{W}_N\}=0,\) and the covariance is given by \(\Sigma = \vec{W}_N\vec{W}_N^T\), where T denotes the transpose. Uncorrelated white noise reduces the covariance matrix \(\Sigma\) to a diagonal matrix. Unless the matrix is orthogonal (i.e., by using an orthogonal set of basis functions), we obtain for each scale \(j\) a different noise level \(\sigma_j\). Before we address the problem of estimating the noise level on each resolution scale \(j\), we take a closer look at the detection problem. The wavelet transform yields a level-or resolution-dependent representation of the observed spectral line profile in terms of the wavelet coefficients \(w_{j,k}\). A decision about whether this wavelet coefficient is significant, i.e., whether it includes signal information or not, can be cast in a hypothesis-testing framework. For this, we can state the null hypothesis \(H_0\) such that \(w_{j,k}\) contains only noise and no signal information. Whether the null hypothesis will be rejected or not depends on probability \(P_n\), \[P_n \: = \: Prob(\| w_{j,k} \| < \tau | H_0) \: , \label{Eq:3.3.5}\] where \(\tau\) is a detection threshold. The null hypothesis will be rejected if \(P_n\) is smaller than a given significance level \(\epsilon\); i.e. \(P_n(\tau) < \epsilon\). For a Gaussian noise distribution with zero mean and standard deviation \(\sigma\), we may write the probability density for \(w_{j,k}\) as \[p(w_{j,k}) \: = \: \frac{1}{\sqrt{2\pi} \sigma} \: e^{\frac{-w_{j,k}^2}{2\sigma^2_j}} \:. \label{Eq:3.3.6}\] The probability \(P_n\) of rejecting \(H_0\) can then be calculated by integrating over all weights \(w_{j,k}\) that are greater than a threshold \(\tau\), \[P_n^{+} \: = \: \frac{1}{\sqrt{2\pi} \sigma} \: \int_{\tau}^{+\infty} \: e^{\frac{-w_{j,k}^2}{2\sigma^2_j}} \: dw_{j,k} \:. \label{Eq:3.3.7}\] Since the wavelet transform of the Stokes profile can have both positive and negative values, we also need to integrate over all negative weights \(w_{j,k}\) up to the threshold value \(-\tau\) \[P_n^{-} \: = \: \frac{1}{\sqrt{2\pi} \sigma} \: \int^{-\tau}_{-\infty} \: e^{\frac{-w_{j,k}^2}{2\sigma^2_j}} \: dw_{j,k} \:. \label{Eq:3.3.8}\] When assuming stationary noise and choosing a specific significance level \(\epsilon\), this reduces to the following decision or threshold detection rule \[\begin{aligned} \|w_{j,k} \| \: \geq k\sigma_j \: \mbox{then \(w_{j,k}\) is significant} \nonumber \hspace{0.7cm} \\ \|w_{j,k} \| \: \leq k\sigma_j \: \mbox{then \(w_{j,k}\) is not significant} \: , \label{Eq:3.3.9} \end{aligned}\] where \(k\) depends on the value of \(\epsilon\). Choosing \(\epsilon = 0.0027\) results in \(k=3,\) the well known \(3\sigma\) detection threshold. This thresholding detection scheme can be introduced readily into the OMP algorithm of Sect. [3.1](#Sect:3.1){reference-type="ref" reference="Sect:3.1"} by implementing the threshold detection rule of Eq. ([\[Eq:3.3.9\]](#Eq:3.3.9){reference-type="ref" reference="Eq:3.3.9"}) as an additional step (Ib) after calculating the maximum projection (i.e. inner product) in step (I). If the currently selected wavelet coefficient is not significant, it will be deleted from the dictionary for this iteration cycle, and the algorithm restarts the current iteration cycle with step (I). The noise level \(\sigma_j\) for each scale \(j\) can be obtained by a wavelet transform of simulated noise. This process yields level-dependent thresholds and has the advantage that different types of noise (e.g., Gaussian or Poisson), as well as correlated noise, can be modeled. # Application to synthetic observations {#Sect:4} We now use the OMP algorithm to estimate the effective and apparent longitudinal magnetic field. Before we begin with a statistical analysis of the accuracy of our proposed method, we give an illustrative example to highlight the functionality of the algorithm. The decomposition of an observed Stokes \(V\) profile into the building blocks of the wavelet dictionary provides the opportunity for each selected signal or profile atom to be individually analyzed in terms of the effective and apparent longitudinal magnetic field. Thanks to the linearity of the algorithm, we can evaluate Eqs. ([\[Eq:2.2.11\]](#Eq:2.2.11){reference-type="ref" reference="Eq:2.2.11"}) and Eq. ([\[Eq:2.3.5\]](#Eq:2.3.5){reference-type="ref" reference="Eq:2.3.5"}) for each signal atom and iteratively approximate the desired magnetic quantities. Each contributing signal atom found by the OMP algorithm identifies a resolved structure in the wavelength (or velocity) domain. The overall number of iterations then also determines the summation index \(n\) for the apparent longitudinal field in Eq. ([\[Eq:2.3.5\]](#Eq:2.3.5){reference-type="ref" reference="Eq:2.3.5"}). The stopping criterion is given by the detection threshold as soon as the approximation enters the noise and no more significant signal atoms are detected. In the following example we have built a stellar surface model with our *iMap* code. We created a random distribution of a radial magnetic field; i.e., each surface segment has a Gaussian random distribution of its radial magnetic field strength with a mean value centered at zero Gauss and a standard deviation of 500 Gauss. The effective temperature of the test star is 5250 K, and it has solar abundance and a gravity of log\((g)=4.0\). The projected rotational velocity \(v \sin i\) is \(35\) km s\(^{-1}\), and micro-and macroturbulence were set to 2 km s\(^{-1}\). The inclination of the rotational axis is 90\(^\circ\). Because the original surface-grid resolution with a 5\(^\circ\) by 5\(^\circ\) segmentation creates a strong cancellation of the Stokes \(V\) signal we decided, for this example case, to additionally smear out the distribution over the surface with a Gaussian filter that has an angular spread of 15\(^\circ\). This smoothing leads to large random clusters of magnetic fields over the surface. The resulting magnetic field surface distribution is shown in Fig. [\[Fig:1\]](#Fig:1){reference-type="ref" reference="Fig:1"}. For the synthetic calculation, we used the magnetically sensitive iron line at 6173 \(\AA\). Line parameters were taken from the VALD line database. The resulting synthetic Stokes \(V\) profile for phase 0.0 is shown in Fig. [\[Fig:2\]](#Fig:2){reference-type="ref" reference="Fig:2"}. To highlight the performance of the method, we used noise-free Stokes profiles in this first test. The stopping criterion for the noise-free case was substituted by a convergence criterion where the iteration was stopped as soon as there was no significant improvement in the root-mean-squared (RMS) error of the approximation. The approximations are illustrated in Fig. [\[Fig:3\]](#Fig:3){reference-type="ref" reference="Fig:3"} where four snapshots at different stages in the approximation process are shown. The magnetic OMP algorithm gradually identifies the most coherent signal atoms on different scales and positions before it finally approximates the entire observed Stokes \(V\) profile with a linear combination of all selected dictionary atoms. As can be seen in Fig. [\[Fig:3\]](#Fig:3){reference-type="ref" reference="Fig:3"}, the algorithm finds the one best-matching signal atom in the first iteration (upper left). In the approximations 5 (upper right), 10 (lower left), and 20 (lower right) one can follow the rapid improvement after adding more and more signal atoms to the residuals. Already at iteration cycle 10 (i.e., 10 dictionary atoms), the differences between the original synthetic observation and the approximation is hardly visible. In Fig. [\[Fig:4\]](#Fig:4){reference-type="ref" reference="Fig:4"} the RMS error of the approximation is plotted over the iteration cycle, which also demonstrates the rapid convergence of the magnetic OMP algorithm. However, the approximation of the pure Stokes \(V\) profile is not the main task of the magnetic OMP method here, because it is supposed to give accurate estimates of the effective and apparent magnetic field as well. The effective and apparent longitudinal magnetic field is estimated during the approximation process from each contributing signal atom. To compare the estimated values with the *true* values, we need to extract the true effective longitudinal magnetic field from the model star. This is straightforward and only requires projecting the magnetic field vector of each surface segment onto the line-of-sight and eventually to sum the area-weighted contribution up from all visible surface segments. The extraction of the true apparent longitudinal magnetic field value from the model deserves some more explanation. In principle we could use the same process as for the effective longitudinal field and sum the absolute value up instead of the plain values from the stellar model, but this would not take the cancellation of the Stokes \(V\) signals coming from within iso-radial velocity strips into account. We would therefore overestimate the apparent field; in fact, we would retrieve the total absolute effective longitudinal magnetic field. To obtain an adequate value for the apparent longitudinal magnetic field, we need to bring the surface distribution of the longitudinal magnetic-flux density into an appropriate iso-radial velocity binning before we can add up their absolute contributions. [\[Table:1\]]{#Table:1 label="Table:1"} The resolution of this iso-radial velocity binning depends on the width of the used (local) spectral line; i.e., the broader the spectral line, the less surface flux can be deduced, and vice versa. Therefore we first translate the surface distribution of the longitudinal field of the model star onto a fine-binned rotational velocity coordinate axis. The resolution of this one-dimensional coordinate axis has the same resolution as our synthetic spectral line profile, i.e. 0.5 km s\(^{-1}\). The distribution of the longitudinal magnetic flux density along the rotation velocity is then convolved with an area-normalized *local* Stokes \(I\) profile to obtain the *measurable* longitudinal flux-density spectrogram over the rotational velocity which is depicted in Fig. [\[Fig:5\]](#Fig:5){reference-type="ref" reference="Fig:5"}. We then sum over the absolute values of the longitudinal flux-density spectrogram to finally obtain the apparent longitudinal magnetic field that is compared to the estimates of the OMP method. In our test case (phase=0.0), the true effective longitudinal magnetic field inferred from the model is 4.67 Gauss and the apparent longitudinal magnetic field is 16.43 Gauss. From the magnetic OMP algorithm, we obtain a value of 4.53 Gauss for the effective longitudinal magnetic field and 16.78 for the apparent longitudinal magnetic field. An exhaustive statistical evaluation is given in the next section, but one can already see that besides the good approximation and the rapid convergence, the method also yields accurate values for both magnetic quantities. There is another interesting effect that highlights the complementary nature of the effective and apparent longitudinal magnetic field. At phase 0.25 of our test star, the polarities in the longitudinal magnetic field are almost perfectly balanced over the visible hemisphere. The resulting profile (again synthesized for the iron line FeI 6173) is shown in Fig. [\[Fig:6\]](#Fig:6){reference-type="ref" reference="Fig:6"}. The almost perfect balancing of polarities is not obvious from the line profile itself. However, if we estimate the longitudinal magnetic field by the COG method's Eq. ([\[Eq:2.2.10\]](#Eq:2.2.10){reference-type="ref" reference="Eq:2.2.10"}) and the magnetic OMP algorithm, we obtain 0.005 G or 0.007 G, respectively, for the effective longitudinal magnetic field. We thus have an extremely small longitudinal magnetic field or none at all. This is not what one would expect from the mere visual inspection of the Stokes \(V\) profile in Fig. [\[Fig:6\]](#Fig:6){reference-type="ref" reference="Fig:6"}, which *apparently* shows a clear net absolute polarization. In fact, the amplitude of the Stokes \(V\) signal is even greater than the one of phase=0.0, shown in Fig. [\[Fig:2\]](#Fig:2){reference-type="ref" reference="Fig:2"}. This demonstrates the strength of the definition of the apparent longitudinal magnetic field; the magnetic OMP algorithm detects a clear apparent longitudinal field of 18.63 G (true value 17.98 G), which is slightly stronger even than in phase 0.0. The quantity of the apparent magnetic longitudinal field is therefore of particular interest in cases where balanced field distributions cause the first-order moment of the Stokes \(V\) profile to vanish. # Numerical experiments and statistical evaluation {#Sect:5} In this section we assess the accuracy of the OMP algorithm under a broad range of model conditions and for different spectral lines. For that reason we synthesized a large number of Stokes \(I\) and Stokes \(V\) profiles for different magnetic surface distributions and atmospheric conditions. The model parameters are the effective temperature, metallicity, surface gravity, projected rotational velocity, and the surface magnetic field and its distribution. For each set of the atmospheric parameters, we created a random surface distribution in the same way as for the previous example in Fig. [\[Fig:1\]](#Fig:1){reference-type="ref" reference="Fig:1"}. The atmospheric parameters were randomly chosen from within an interval given in Table [1](#Table:1){reference-type="ref" reference="Table:1"}. For the model atmospheres, we chose Kurucz/Atlas-9 models. Because the model atmospheres are provided on a fixed grid of step sizes of 250 K for the effective temperature, 0.5 for the logarithmic abundance and gravity, the atmospheres are interpolated for each randomly chosen set of atmospheric parameters. As Zeeman-sensitive spectral lines we chose three magnetically sensitive lines; the iron lines FeI 5497 (\(g_{\rm eff} = 2.22\)), FeI 6173 (\(g_{\rm eff} = 2.50\)), and FeI 8468 (\(g_{eff} = 2.50\)). All line parameters were again taken from the VALD line database. For each spectral line, we created a sample of 5,000 randomly chosen atmospheric parameters, rotational velocities, and magnetic surface distributions to eventually synthesize a set of corresponding Stokes \(I\) and Stokes \(V\) profiles with the forward module of our *iMap* code. In total, we calculated a set of 15,000 Stokes \(I\) and Stokes \(V\) profile, which were then analyzed by our magnetic OMP algorithm. The true effective longitudinal magnetic field is directly extracted from the magnetic surface distribution of the synthetic model star. The true apparent longitudinal magnetic field is obtained from the convolved magnetic spectrograms as described above. Because the individual trials, sets of atmospheric parameters, and surface magnetic field values have different scales, we chose a relative performance metric to describe the accuracy. The overall performance is judged by the mean absolute percentage error, while for illustrating the error distribution we use the relative percentage error. The relative percentage error is defined as \]D_i \: = \: \frac{B^T_i-B^{OMP}_i}{B^T_i} \: * \: 100 \: ,\[ where \(B^T_i\) is the true magnetic field, and \(B^{OMP}_i\) the calculated magnetic field of the \(i\)-th sample. Finally, the mean absolute percentage error (MAPE) is defined by \]M_B \: = \: \frac{1}{n} \: \sum_{i=1}^n \: \left | D_i \right | \:.\[ ## Sparsity of Stokes profiles {#Sect:5.0} Before we evaluate the performance of the OMP algorithm, we look more closely at the general sparsity of Stokes profiles. A sparse representation is often achieved by transforming a vector from the original data domain into a domain where the transformed coefficients (i.e., expansion coefficients) allow a more compact representation of the information (e.g., Fourier or wavelet transform). A vector is called sparse if there is a representation where most of the vector entries are zero or close to zero. A sparse approximation is performed by restricting the sparse transformation to the largest expansion coefficients such that there is no great loss of signal information. Recently, have shown that Stokes profiles are sparse and compressible under various transformations (e.g., wavelets and empirical basis functions). Compressible here means that the values of the sorted expansion coefficients exhibit an exponential decay that in turn results in a small approximation error. Although we saw from the numerical simulation of the last section that a sparse representation of a synthetic Stokes profile is possible with our dictionary elements, we also want to test empirically if this also holds for the entire set of our 15,000 Stokes profiles. To quantify the approximation error, we use the relative error, \(\sum_i^n \frac{1}{n} \frac{\|\hat{\vec{S}}_i-\vec{S}^*_i \|}{\| \vec{S}^*_i \|}\) where \(\hat{\vec{S}}\) is the approximation and \(\vec{S}^*\) the original Stokes vector. To compare the performance of the OMP expansion relative to a known sparse decomposition, we also calculated a principal component analysis (PCA) of the entire synthetic data set. A PCA decomposes a set of observed Stokes profiles into an empirical set of orthogonal eigenprofiles. As has been shown by and, a PCA expansion can result in a compact and sparse representation of Stokes profiles. Figure [\[Fig:7\]](#Fig:7){reference-type="ref" reference="Fig:7"} shows how the dictionary used with the OMP algorithm performs against a PCA decomposition of our synthetic data set. Both curves (solid OMP, dashed PCA) show a rapid decline of the mean approximation error with increasing numbers of signal atoms or eigenvectors, respectively. The performance of our dictionary of Gaussian derivatives is better than that of the empirical basis functions (i.e., eigenprofiles) obtained by the PCA. The reason for that is two-fold: First, the signal atoms of our dictionary are selected a priori to match the building blocks (i.e., elementary waveforms) of our problem and second the redundancy of our dictionary facilitates a more efficient expansion then that of the orthogonal eigenvectors of the PCA. In contrast to the empirical basis formed by the PCA, the OMP dictionary is not required to be a set of orthogonal basis functions, and this generally provides better flexibility and adaptivity to approximate a given signal. To reduce the relative approximation error with the OMP algorithm below 5%, it takes on average only nine signal atoms. Using 22 signal atoms the relative error is on average smaller than 1 %, which demonstrates that our dictionary allows a sparse representation of Stokes profiles for the parameter regime given in Table [1](#Table:1){reference-type="ref" reference="Table:1"}. ## The noise-free case {#Sect:5.1} To obtain an idea about the principal accuracy of the magnetic OMP algorithm and its performance relative to the conventional COG method, we began with a noise-free test case where the entire test sample of 15,000 Stokes \(I,\) and Stokes \(V\) profiles were used without any noise contribution. In this simulation the overall error value for the effective longitudinal magnetic field obtained from the magnetic OMP algorithm yielded a MAPE of 1.87 %. For comparison the effective longitudinal magnetic field calculated by the COG method shows a similar MAPE of 1.79 %. The same error for the apparent longitudinal magnetic field is 4.67 %. The error for the three individual spectral lines show no apparent deviation from the overall error. For the subset of synthetic Stokes \(V\) profiles of the FeI 5497, we obtain a MAPE for the effective (apparent) longitudinal magnetic field of 1.80 % (4.57 %), for the FeI 6173 lines a MAPE of 1.93 % (4.69 %), and for the FeI 8468 lines a MAPE of 1.88 % (4.72 %). The error distribution for the relative percentage error of the effective longitudinal field calculated by the OMP and the COG method is shown in Fig. [\[Fig:8\]](#Fig:8){reference-type="ref" reference="Fig:8"}. Both the OMP method and the conventional COG method provide equally good results. The error distribution of the apparent longitudinal magnetic field is shown in Fig. [\[Fig:9\]](#Fig:9){reference-type="ref" reference="Fig:9"}. Given that the apparent field is even harder to determine due to possible cancellation effects of the Stokes \(V\) signals, the error is still remarkably low, and the OMP method is able to recover the apparent field from the Stokes \(V\) profiles with good accuracy. ## The noise case {#Sect:5.2} In the following we test the accuracy of the magnetic OMP method and the COG method for the more relevant case when the Stokes \(V\) profiles are contaminated with noise. For that reason we have added an increasing amount of white noise to the 15,000 Stokes \(V\) profiles. Since the Stokes \(V\) profiles generated from the random surface distribution are of different magnitudes, we define a relative noise level \(\eta\) that is given by the relative variance of the Stokes \(V\) profile and that of the noise according to \]\eta_{\rm rel} \: = \: \frac{\sigma(N)}{\sigma(S)} \: ,\[ where \(\sigma(S)\) and \(\sigma(N)\) are the standard deviations of the Stokes \(V\) profile and the noise contribution, respectively. Figure [\[Fig:10\]](#Fig:10){reference-type="ref" reference="Fig:10"} shows how the relative error increases with higher noise levels. It is interesting how rapidly the accuracy of the COG method deteriorates (upper curve) compared to the relative robustness of the magnetic OMP method (lower curve). Already for a noise level of \(\eta_{\rm rel}\) = 0.3, the accuracy of the COG method is significantly affected, whereas the OMP method still provides good accuracy with errors of less than 10 %. Even for a noise level of unity, the OMP method has an error of just 18 % where the COG method with an error of almost 50 % can no longer be considered a meaningful tool for quantifying the magnetic field. For the apparent longitudinal magnetic field, the OMP method shows a very robust performance as well. The error distribution over the relative noise level is illustrated in Fig. [\[Fig:11\]](#Fig:11){reference-type="ref" reference="Fig:11"}. The good performance and robustness of the magnetic OMP method compared to the COG method can be understood by considering the noise-free Stokes \(V\) profile in Fig. [\[Fig:11\]](#Fig:11){reference-type="ref" reference="Fig:11"}. Because the OMP algorithm gradually approximates the noisy profile with a linear combination of dictionary profile atoms, it calculates the effective and apparent longitudinal magnetic field for each of the selected profile atoms separately. The detection mechanisms of Sect.[3.3](#Sect:3.3){reference-type="ref" reference="Sect:3.3"} prevent the algorithm from approximating insignificant profile features and thus largely avoids the quantification of the contributing noise. Figure [\[Fig:12\]](#Fig:12){reference-type="ref" reference="Fig:12"} shows one of the sample profiles generated for a random surface distribution of the magnetic field. The noiseless profile is approximated well by the OMP method, and the estimated effective and apparent longitudinal magnetic field is approximated with--41.04 G and 59.88 G, respectively, which is very close to the true values of--40.11 G and 61.65 G. The noiseless profile is approximated by the maximum number of 40 iterations, which means that 40 dictionary profile atoms are used to approximate the example profile. However, already the first three profile atoms are able to approximate the profile to such a degree that 95 % of the signal variance are described. In Fig. [\[Fig:13\]](#Fig:13){reference-type="ref" reference="Fig:13"} the same profile can be seen with varying degrees of noise contributions overplotted again by the approximation from the OMP algorithm. Despite the increasing relative noise level, the OMP algorithm finds a good approximation to the original Stokes \(V\) profile. The iteration for a noise level of \(\eta_{\rm rel}\) = 0.25 stops already at seven iterations (i.e., 7 profile atoms). For \(\eta_{\rm rel}\) = 0.5, 0.75, and 1.0, the iteration is stopped at cycles 6, 4, and 3, respectively. At a relative noise level of unity only three signal atoms are used to approximate the Stokes \(V\) profile. However, as one can see from the noiseless case, these three signal atoms already capture the essential profile shape and amplitude of the original Stokes \(V\) signal. Since the magnetic OMP algorithm interprets and analyzes each signal atom separately, the good approximation directly translates to a good estimates of the underlying effective and apparent longitudinal magnetic field. In this particular case, the estimations for the effective (apparent) longitudinal field from the magnetic OMP algorithm is--42.54 G (64.27 G) for a noise level of 0.25, --44.26 G (65.71 G) for a noise level 0.5,--36.04 G (68.85 G) for a noise level 0.75, and--34.77 G (72.54 G) for noise level of 1.0. ## Absolute limits of the COG and OMP method {#Sect:5.3} Until now we have analyzed the performance of the magnetic OMP and COG method by using *relative* noise levels. To shed more light on the absolute performance limits of the two methods, we take a closer look on the absolute values of the noise. Multiline reconstruction techniques like LSD or SVD allow us to drastically lower the noise levels of spectropolarimetric observations. The thus increased signal-to-noise (S/N) levels allows detecting faint signal profiles that were otherwise deeply buried in the noise. A natural question that arises in the context of magnetic field estimation is how much of the residual noise is falsely interpreted as magnetic field and up to which limits we can reliably estimate a true underlying magnetic field. To address this problem, we could simply perform a first-order perturbation to the COG method and apply standard error propagation to estimate the influence of the noise. However, for weak magnetic fields, the noise contribution relative to the true signal amplitude is in general not small anymore. We therefore simulate the process with a large and statistically significant number of synthetic profiles to determine the impact of typical absolute noise levels on the magnetic field estimation. We assume additive Gaussian noise with zero mean. One might expect that this would lead to a vanishing value in the field estimation by the COG method. This is, however, not true for a finite spectral resolution. The random fluctuations and the unequal weighting in the velocity (or wavelength) domain can lead to relatively large spurious contributions to the first-order moment, hence to the field estimation as shown in the following simulation. Note that any form of correlated noise would even exacerbate the problem since it would introduce a systematic bias in the evaluation of the first-order moment. With our assumption of white noise, a Stokes \(\tilde{V}^*\) profile can be written as the sum of the true noiseless Stokes \(\tilde{V}\) profile and a noise vector \(N\), \]\tilde{V}^*(\textrm{\textsl{v}}) \: = \: \tilde{V}(\textrm{\textsl{v}}) \: + \: N(\textrm{\textsl{v}}) \:.\[ The effective or mean longitudinal field as measured by the COG method can then be written as \]B_{\rm eff} = \frac{\lambda_0 \: \int_{V} \left ( \tilde{V}(\textrm{\textsl{v}}) + N(\textrm{\textsl{v}}) \right ) \: \textrm{\textsl{v}} \: d\textrm{\textsl{v}}} {c \: \alpha \: W_I} \:.\[ We assume that the noise level is small enough that the estimation of the Stokes \(I\) equivalent width remains largely unaffected. The first-order moment of a noisy Stokes \(\tilde{V}^*\) profile, as well as the magnetic field estimation, can then be split into a signal (\(B^S_{\rm eff}\)) and a noise (\(B^N_{\rm eff}\)) contributing part, \]B_{\rm eff} \: = \: B^S_{\rm eff} \: + \: B^N_{\rm eff} \: = \: \frac{\lambda_0} {c \: \alpha \: W_I} \: \left ( \int_{V} \tilde{V}(\textrm{\textsl{v}}) \: \textrm{\textsl{v}} \: d\textrm{\textsl{v}} + \int_{V} N(\textrm{\textsl{v}}) \: \textrm{\textsl{v}} \: d\textrm{\textsl{v}} \right ) \:. \label{Eq:5.3.1}\(\) Written in this form we may ask to which noise level the following relation is valid, \(B^S_{\rm eff} > B^N_{\rm eff}\), or at which point does a noise-induced magnetic field interfere with the magnetic field estimation from the true signal profile. For the simulations we defined the absolute noise level \(\eta_{abs}\) as the standard noise deviation \(\sigma_N\) relative to the continuum of the normalized intensity profile, i.e., \(\eta_{\rm abs} = \sigma_N,\) which is also commonly used to describe the S/N level (\(1/\sigma_N\)) in stellar polarimetry. To mimic a multiline, reconstructed line profile (e.g., LSD profile), we created a fictitious iron line with a mean Landé factor of 1.2 at a rest wavelength of 5000 Å. The oscillator strength and van der Waals damping were adjusted to give an equivalent width for the Stokes \(I\) profile of 100 mÅ. The Stokes \(I\) and Stokes \(V\) noise profiles were calculated with a resolution of \(\lambda/\Delta\lambda=\)`<!-- -->`{=html}100,000, which is common to current high-resolution observations. The spectral range covers a velocity of plus/minus 50 km s\(^{-1}\) around the line center, and integration limits were set accordingly. We chose noise levels from 10\(^{-3}\) down to 10\(^{-5}\) that correspond to a polarimetric sensitivity of 0.001 %, a level that is reached by current reconstruction methods. We selected 20 noise levels between 10\(^{-5}\) and 10\(^{-3}\), equally spaced on a logarithmic scale. For *each* absolute noise level, we synthesized 10 000 noisy line profiles such that we have a total set of 200 000 simulated profiles. Owing to the linear superposition of the signal and noise contribution in Eq. ([\[Eq:5.3.1\]](#Eq:5.3.1){reference-type="ref" reference="Eq:5.3.1"}), we do not depend on an underlying magnetic field and only need to simulate, besides the Stokes \(I\) profile, the noise part of the profile. To quantify the magnetic response to the noise, we calculated the effective longitudinal magnetic field for each synthetic noise profile with the COG and the OMP methods. To obtain a statistical meaningful average response for each of 20 noise levels, we computed the mean of the absolute effective field from each of the corresponding 10 000 field values. In Fig. [\[Fig:14\]](#Fig:14){reference-type="ref" reference="Fig:14"}, we plotted the magnetic response of the COG and of the OMP method over the various noise levels. From the response curve, we can immediately identify how much of the noise is falsely interpreted as a magnetic field. As the noise level increases, the COG method is increasingly affected by the noise, and more and more of the content of each noise profile is interpreted as an effective longitudinal magnetic field. For the COG method, a noise level of 10\(^{-3}\) results, on average, in an effective longitudinal magnetic field of 101 G. On the right side of Fig. [\[Fig:14\]](#Fig:14){reference-type="ref" reference="Fig:14"}, we expand the region between 10\(^{-5}\) to 10\(^{-4}\). A few times 10\(^{-5}\) is the region where many of the recently detected weak magnetic fields are reported . Despite the low noise level, we see that the COG method interprets a non-negligible amount of the noise as a magnetic field (between 2 and 10 G). The shaded areas in Fig. [\[Fig:14\]](#Fig:14){reference-type="ref" reference="Fig:14"} mark the region of magnetic field values that can no longer be properly interpreted by the COG method, i.e., the region where \(B^S_{\rm eff} < B^N_{\rm eff}\). In that sense the response curves set the lower limit or a noise threshold for the estimation of weak magnetic fields. However, even when the true field strength is above these threshold values, the *relative* contribution from the noise can severely compromise the field estimation of the COG method as was shown in Sect.[5.3](#Sect:5.2){reference-type="ref" reference="Sect:5.2"}. For the magnetic OMP method, on the other hand, we see from Fig. [\[Fig:14\]](#Fig:14){reference-type="ref" reference="Fig:14"} that the response is largely flat. The magnetic field estimation from the OMP method is almost unaffected by the noise and shows only a small increase with the noise level. The incorrectly attributed field values are no higher than 0.7 G. This has its origin in the multiresolution thresholding scheme implemented in the OMP algorithm (see Sect.[3.3](#Sect:3.3){reference-type="ref" reference="Sect:3.3"}). Only those signal or noise features that reach a certain significance are interpreted and quantified. This again results in the good performance and robustness against the contributing noise. A dark gray area that shows the incorrect response region of the OMP method is not visible in Fig. [\[Fig:14\]](#Fig:14){reference-type="ref" reference="Fig:14"} due to the low noise response of the OMP method. The specific values for the noise response of the COG method also depend of course on the line parameters, such as wavelength, equivalent width, and Landé factor. However, the deviations from our results are expected to be small because the spread in the derived average parameters for multiline reconstruction techniques that use line lists of many hundreds or thousands of lines are relatively small. The spectral resolution is also a contributing factor to the error of the COG method since the statistical fluctuations that enter into the evaluation of the first-order moment directly depend on the number of available wavelength points, such that lower spectral resolutions on average lead to larger errors (i.e., to a stronger noise response). However, a simulation and critical assessment of all these contributing factors in the COG estimation, are not subjects of the current paper. # Summary {#Sect:6} This work presents a novel technique for detecting and quantifying stellar magnetic fields. By employing a sparse profile approximation in terms of an orthogonal matching pursuit algorithm, we were able to provide a robust method for detecting and estimating the effective longitudinal magnetic field. By introducing the concept of the apparent longitudinal magnetic field, which describes the maximum of the resolvable absolute longitudinal surface field over the stellar disk, we provided a complementary measure to the effective or mean longitudinal magnetic field. For rapidly rotating stars, the definition of the apparent longitudinal magnetic field allows a quantification of the field even in situations where a small-scale and balanced surface field would otherwise result in a vanishing mean longitudinal magnetic field. It was shown by an extensive numerical simulation with our *iMap* code that the accuracy of the OMP method is remarkably good with a mean absolute percentage error of only 1.87 % for the effective longitudinal magnetic field and 4.67 % for the apparent longitudinal magnetic field. However, the real strength of the new approach is its robustness against noise. Even for relative noise levels of unity, i.e., when the variance of the noise has the same magnitude as the actual Stokes \(V\) signal, the OMP method has a mean error of only 18 %, whereas the conventional estimation by the COG method yields a mean error of almost 50 %. In an effort to understand the limitations of the conventional COG method for estimating weak magnetic fields from faint and noisy Stokes \(V\) signals, we simulated noise profiles in a low-noise regime from an absolute noise level of 10\(^{-3}\) down to 10\(^{-5}\). This is a regime that is currently reached by multiline reconstruction techniques such as LSD or SVD and will be reached by the next generation of spectropolarimeters like PEPSI at the 2\(\times\)`<!-- -->`{=html}8.4m Large Binocular Telescope (LBT). In these low-noise regimes, however, a non-negligible fraction of the noise was incorrectly interpreted as a magnetic field by the COG method. Depending on the noise level, these falsely identified field values are between 2 and 100 G, making any attempt to estimate weak fields of a similar strength extremely difficult. The magnetic OMP method, on the other hand, again shows a remarkable robustness over the entire range of noise levels, where only 0.1 to 0.7 G are incorrectly attributed to a magnetic field. This makes the magnetic OMP method the method of choice for estimating weak magnetic fields from noisy Stokes profiles. From a numerical perspective, the algorithm can be easily implemented and the iterative process is fast to evaluate.[^2] The magnetic OMP algorithm provides a viable tool for detecting and quantifying magnetic fields from spectropolarimetric observations. Although the focus in this work has been placed on disk-integrated Stokes profiles of stellar magnetic fields, this method is also directly applicable to resolved solar magnetic field observations. [^1]: An IDL-Code of the magnetic OMP method is available under http://www.aip.de/People/tcarroll and http://www.aip.de/Members/tcarroll [^2]: An IDL-Code of the magnetic OMP method is available under http://www.aip.de/People/tcarroll/ and http://www.aip.de/Members/tcarroll/

{'timestamp': '2014-01-23T02:11:18', 'yymm': '1401', 'arxiv_id': '1401.5749', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5749'}

# Introduction Large amounts of cosmic dust at high redshift are inferred from the reddening of background quasars and damped Ly\(\alpha\) systems () and pose the problem of identifying the providers of dust grains in primeval galaxies and at cosmic time less than one billion years. Type II Supernovae (SNe) are strong contenders as dust factories at high-redshift because of the short evolution time-scale of their massive stellar progenitors (\(\sim\) \(10^6\) yrs). Furthermore, dust and molecules (namely carbon monoxide, CO and silicon monoxide, SiO) have been detected at infrared (IR) wavelengths in several local SNe, including SN1987A in the Large Magellanic Cloud, a few hundreds of days after the explosion (, 2006, 2009, ). The dust mass derived is small and ranges from between \(10^{-5}\) \(M_{\odot}\) and \(10^{-2}\) \(M_{\odot}\). These values fall short of the 1 \(M_{\odot}\) of dust required per SN explosion to explain the large dust mass in primeval galaxies, as estimated by Dwek et al. (2007). Recently, cool dust grains have been detected with the submillimetre (submm) space telescope Herschel in several SN remnants, several decades or centuries after the SN explosion. The derived masses are larger than the masses observed in the IR at early times, and vary between \(\sim\) 0.08 \(M_{\odot}\) for the Cas A remnant () and 0.7 \(M_{\odot}\) for the young remnant SN1987A (). These large values are surprising because the physical conditions are harsh in SN remnants. The SN ejecta gas is re-processed by a reverse shock (RS) created when the mass of gas swept up by the explosion blast-wave exceeds the ejecta mass (). The RS travels inwards and therefore induces the partial destruction of the molecules and dust synthesised and expelled in the SN ejecta phase. Owing to its proximity (3.4 kpc) and angular size (\(\sim\) 5') (), the SN remnant Cas A has been extensively studied observationally. Cas A results from the explosion of a SN of type IIb 330 years ago. The progenitor was a massive supergiant star having lost most of its hydrogen envelope and with a ZAMS mass comprises between 15 and 25 \(M_{\odot}\) (). The explosion blast wave travelled in a diffuse medium, leading to a low density ejecta compared to regular Type II SNe (). The Cas A ejecta is now crossed by the RS and warm dust has been detected with the IR space telescope Spitzer in a mass range \(0.02-0.054\) \(M_{\odot}\) at the RS position (). Cool dust interior to the RS position has also been detected with Herschel (), with an estimated mass of \(\sim\) 0.075 \(M_{\odot}\). Of prime importance is the detection of molecules in the remnant. Warm carbon monoxide, CO, has been observed at the RS position with Spitzer (, 2012) and in one dense clump with Herschel (). In the young remnant SN1987A, cool CO has been recently observed with ALMA along with the partial spectroscopic detection of SiO (). This detection points to chemical species, CO and SiO, formed in the ejecta after the SN explosion. As in other evolved circumstellar environments, molecules appear to be strongly coupled to the formation of dust grains in SNe (). Once more, the presence of dust and molecules in Cas A and other remnants confirms this inherent aspect of the synthesis of cosmic dust and indicates a strong chemical reprocessing of the material ejected during the SN phase by the RS. Theoretical studies on both the formation of dust in SNe and the reprocessing by the RS have been carried out. Models of dust production in SN ejecta lead to contradictory results on the dust quantity produced after the SN explosion. Early studies based on classical nucleation theory (CNT) indicate large dust masses () that cannot be reconciled with IR observations of SNe. Models based on chemical kinetics and considering the coupling between molecules and dust grains result in lower amount of dust formed in the nebular phase (). More recently, Sarangi & Cherchneff (2013, hereafter SC13) show that the molecular clusters precursors to dust (hereafter, dust clusters) gradually grow from low to high masses on a time-span of a few years after outburst and provide a genuine explanation to the discrepancy on dust mass derived from IR data of SNe and submm data of SN remnants. The reprocessing of dust by the RS in SN remnants has been modelled assuming pre-shock dust distributions derived from CNT and in the context of a homogeneous SN ejecta (). None of these studies addresses the chemistry of the RS, the survival of molecules and the possibility to reform dust after the passage of the RS. Therefore, whether SNe and SN remnants are dust makers and/or destroyers is still unclear on the basis of both available theoretical predictions and observational data. In order to advance our understanding of the net dust budget produced by SNe, we carry out a global study of the synthesis of molecules and dust in SN ejecta and their reprocessing by the RS in the remnant phase, focusing on Cas A. In this paper (pap. I), we study the formation of molecules and dust clusters in the Type IIb SN that led to Cas A, and the later processing of the produced gas-phase material by the RS. In a forthcoming publication (pap. II), we will address the condensation of dust clusters, and derive grain size distributions for the Type IIb SN. We will then study the dust reprocessing in clumps for various RS models. Presently, we model the non-equilibrium chemistry of the SN ejecta and that of the post-RS gas, assuming that the processed ejecta material resides in dense, fast-moving knots, as observed by Fesen et al. (2001, 2006) and Docenko & Sunyaev (2010, hereafter DS10). In Section 2, we present the chemistry considered in this study for both the SN type IIb and the dense, shocked, remnant knots. The physical models used for the SN ejecta and the knot shocked by the RS are described in Section 3. We present the results in Section 4 and discuss our findings in section 5. # Chemical model {#chem} The synthesis of dust in SN environments stems from the chemical ability of the elements produced via nucleosynthesis during the stellar evolution of the massive progenitor and the SN explosion to assemble into molecules and dust molecular clusters. SN ejecta are characterised by extreme physical conditions including high gas temperature and velocity in the back of the explosion blast wave, radioactivity induced by the decay of \(^{56}\)Ni, and \(\gamma\)-rays and ultraviolet (UV) fields. For both environments studied in this paper, we use similar processes as in the study of dust formation in primeval and Type II-P SNe by Cherchneff & Dwek (2009, 2010) and SC13, respectively. A chemical kinetic approach is based on a chemical reaction network that includes the thermal and non-thermal processes summarised in Table [\[tab1\]](#tab1){reference-type="ref" reference="tab1"}. Thermal processes include unimolecular, bimolecular, and termolecular reactions. Because of the high gas temperatures present a few months after the SN explosion, neutral exchange reactions, which often have activation energy barriers reflecting the energy required to break and re-arrange chemical bonds, are prevalent formation processes in the build-up of chemical complexity and the formation of dust clusters (, SC13). In addition to these reactions, we consider non-thermal processes that include the destruction of molecules and ionisation of atoms by energetic Compton electrons created when the \(\gamma\)-ray photons generated by the decay of \(^{56}\)Ni and \(^{56}\)Co degrade via collisions with electrons. Rates are estimated following Cherchneff & Dwek (2009) and we assume an average \(\gamma\)-ray optical depth for the ejecta with a 19 \(M_{\odot}\) stellar progenitor (as shown by SC13, the use of a \(\gamma\)-ray optical depth for each ejecta zone as calculated by Kozma & Fransson (1992) for a 20 \(M_{\odot}\) progenitor has little impact on the trends and results derived for an average \(\gamma\)-ray optical depth). Photo-ionisation and-dissociation by UV radiation are considered in the chemistry of the shocked clump. The various atoms, molecules, and ions considered to participate in the SN ejecta chemistry are summarised in Table [\[tab2\]](#tab2){reference-type="ref" reference="tab2"}. Small molecular clusters form according to the processes described in SC13, where the description of the growth pathways of small silicate clusters, namely forsterite dimers (Mg\(_2\)SiO\(_4\))\(_2\) and enstatite dimers (MgSiO\(_3\))\(_2\), is based on the work by Goumans & Bromley (2012). This study indicates possible chemical routes to the formation of the silicate dimers, which involve the formation of the SiO dimer ring, (SiO)\(_2\), and its growth to Si\(_2\)O\(_3\) through the reaction with O\(_2\) and SO. The subsequent pathway involves the addition of a magnesium atom to the Si\(_2\)O\(_3\) structure. The later growth of clusters is described by one O-addition step followed by one Mg inclusion as a recurrent growth scenario. We have also considered a new pathway for the formation of the alumina dimer, (Al\(_2\)O\(_3\))\(_2\), that was not considered in SC13. Because the structure of the molecule Al\(_2\)O\(_3\) is akin to that of Si\(_2\)O\(_3\), we have considered a similar growth route starting from the formation of AlO, and followed by its dimerisation to Al\(_2\)O\(_2\) and the formation of Al\(_2\)O\(_3\) via oxidation. The alumina dimers are then formed from the recombination of two Al\(_2\)O\(_3\) molecules. Oxidising agents for the formation of silicates and alumina include atomic O, O\(_2\), AlO and SO, but reactions with O\(_2\) and SO are prevalent for silicate clusters while reaction with AlO dominates in the case of alumina. The chemical modelling of the Type IIb SN that led to Cas A includes two temperature regimes and chemical networks, because the initial gas temperature and density quickly drop with time from 100 days to 3000 days (\(\sim\) 8 years) post-explosion (see next Section and Table [\[tab4\]](#tab4){reference-type="ref" reference="tab4"}). These values are lower than in the case of Type II-P SNe because the ejecta quickly expands at high velocities owing to the lack of a progenitor circumstellar wind that would slow down the explosion blast wave through interaction. We thus consider a high temperature reaction network from day 100 (\(\sim\) 7000 K) to day 1000 (300K) and a low temperature network from day 1001 until day 3000 (\(\sim\) 60 K), for which the reaction rates of relevant reactions have been adjusted according to available, low-temperature data. The chemical kinetic description is applied to the gas parameters that characterise the SN ejecta and shock models (see Section § [3](#phys){reference-type="ref" reference="phys"}). The variation of the number density of the chemical species \(i\) with time is described by the following rate equation \[\label{eq1} \frac{dn_i}{dt} = P_i-L_i = {\sum}_j k_{ij}n_jn_i-{\sum}_k k_{ik}n_in_k\] where \(P_i\) and \(L_i\) are the total chemical production and loss processes for species \(i\), \(n_i\) is the number density for species \(i\), and \(k_{ij}\) the rate for the reaction of \(i\) with \(j\). This rate is expressed in Arrhenius form as \[\label{eq2} k_{ij}(T)=A_{ij} \times \left( \frac{T}{300} \right)^{\nu} \times exp(-E_{ij}/T)\] where T is the gas temperature in Kelvin, \(\nu\) the temperature dependence exponent, and \(E_{ij}\) the activation energy barrier in Kelvin. The coefficient A\(_{ij}\) has the units of s\(^{-1}\) for unimolecular processes, cm\(^3\) s\(^{-1}\) for bimolecular reactions, and cm\(^6\) s\(^{-1}\) for termolecular reactions. The system includes 93 species, 412 chemical reactions, and a set of 93 stiff, coupled, ordinary differential equations is solved using a Gear method (). # Physical models {#phys} We study the chemistry of two distinct environments: 1) the ejecta of the SN Type IIb that led to the remnant Cas A, and 2) one oxygen-rich knot in Cas A that is crossed by the reverse shock. The aim is to derive a comprehensive picture of the production and processing of molecules and dust in Cas A and to assess whether the chemical species synthesised in SNe survive to reverse shock processing in the remnant phase. First, we present our physical model for the ejected material of a Type IIb SN and, second, we describe our model of a shocked clump or knot in Cas A. ## The type IIb SN ejecta {#ejec} Stars with mass in the range \(8-30\) M\(_{\odot}\) explode as core-collapse SNe of Type II, showing strong hydrogen lines in their spectrum. Type IIb SNe show a weak hydrogen line, that subsequently disappears with time, and are characterised by higher expansion velocities, resulting in a more diffuse ejected material compared to regular Type II SNe. The Cas A remnant ensues from the explosion of a blue supergiant of mass \(15-25\) \(M_{\odot}\) as a type IIb SN (). We consider a stratified ejecta whose elemental composition is given by the 19 \(M_{\odot}\) SN progenitor model of Rauscher et al. (2002). The ejecta consists of mass zones of specific chemical composition summarised in Table [\[tab3\]](#tab3){reference-type="ref" reference="tab3"}. Each zone is microscopically mixed and we assume no chemical leakage between zones. The number density and temperature profiles are taken from explosion models for Type IIb SNe presented by Nozawa et al. (2010), where a homologous expansion is assumed for the ejecta gas, at constant velocity \(v_{ej}=4000\) km s\(^{-1}\) (i.e., the expansion velocity of the oxygen-rich core in Nozawa's model). The gas number density is given by \[\label{eq3} n (M_r, t) = n(M_r, t_0) \times {\left(\frac{t}{t_0}\right)}^{-3},\] where \(M_r\) is the mass zone position, and \(t\) the time after explosion with \(t_0=100\) days. The temperature profile is given by \[\label{eq4} T(t) = T_0 \times {\left(\frac{t}{t_0}\right)}^{3(1-\gamma)}\] where \(T_0\) is the gas temperature at 100 days, and \(\gamma\) is the adiabatic index. Values are: \(T_0=6663.95\) K and \(\gamma = 1.433\). A flat density profile was assumed across the He-core with a value at 100 days equal to \(\rho(100)= 5.1\times 10^{-15}\) g cm\(^{-3}\). All values were taken from Nozawa et al. (2010). The SN Type IIb parameters for the Cas A progenitor and the values for the ejecta gas number density and temperature as a function of time are summarised in Table [\[tab4\]](#tab4){reference-type="ref" reference="tab4"}. ## The reverse shock models {#revsh} The Cas A remnant shows evidence for several high-density knots or clumps embedded in a rarefied inter-clump medium. Fesen et al. (2001) have observed fast-moving knots (FMKs) in the optical while Rho et al. (2009, 2012) have detected ro-vibrational transitions of CO in emission in several small knots located at the RS position. High energy line detection of shocked CO with Herschel confirmed the existence of high-density gas regions in the remnant (). In Cas A, the RS velocity relative to the ejecta is \(\sim\) 2000 km s\(^{-1}\) (). When encountering a dense clump with a density contrast with respect to the inter-clump medium \(\chi= n_c/n_{ic}\) (\(n_c\) and \(n_{ic}\) are the number density of the clump and the interclump medium, respectively), the RS velocity decreases by a factor \(\sqrt\chi\) owing to the conservation of energy. For clump density contrast ranging from between \(100\) and \(1000\) and RS velocities of \(1000-2000\) km s\(^{-1}\), the RS velocity in the clump spans \(30-200\) km s\(^{-1}\). Such velocities have been confirmed by DS10 from their analysis of fine-structure far-IR atomic lines coming from an oxygen-rich FMK in Cas A. In the present study, we want to assess the fate of the chemical species, including molecules and dust clusters, within ejecta clumps that are processed by the RS. We thus investigate various RS velocities, derive the time-variation of the post-RS gas parameters, based on existing shock models, and apply our chemical kinetic formalism to study the post-RS chemistry. Borkowski & Shull (1990) (hereafter BS90) model steady state radiative shocks in a pure oxygen gas with velocities that range between 35 km s\(^{-1}\) and 170 km s\(^{-1}\). This study provides the adequate range of conditions for the case of the RS impacting a dense knot in Cas A. Considering a knot of gas with an initial density contrast with respect to the inter-clump medium, the parameters pertaining to a description of the RS, as illustrated in Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}, are the shock velocity relative to the knot velocity \(V_s\), the pre-shock number density and temperature of the knot, \(n_0, T_0\), the gas density and temperature in the post-shock hot region (HR), \(n_1, T_1\), and those for the photoionised region (PIR), \(n_2, T_2\), and the time length of the PIR, \(t_{PIR}\). In the HR, the compression induces an equilibration between ions and electrons which draw their temperature through Coulomb collisions with ions when they cool via inelastic collisions with ions. The cooling time being shorter than the recombination time for ions, a constant ionisation state prevails in the HR until the gas quickly cools down and recombines. Ionisation is still sustained in the PIR by the UV flux coming form the HR so that thermal equilibrium is reached over a time \(t_{PIR}\) (or length \(L_{PIR}\)) defined by the depletion of the HR UV flux. From our SN ejecta model, we can constrain the knot physical properties as follows: considering the oxygen-rich zones of the ejecta (e.g., zone 2) and assuming the ejecta expansion velocity of Table [\[tab4\]](#tab4){reference-type="ref" reference="tab4"}, the volume of the homogenous ejected material at time \(t\) (in days) after explosion, assuming a spherical expansion, is \[\label{eq5} V(t) = {{4\times \pi}\over {3}} \times (v_{ej}\times t \times 8.64\times 10^4)^3,\] where \(v_{ej}\) is the ejecta velocity. For our chosen values at day 100, \(V(100) = 1.7 \times 10^{47}\) cm\(^{-3}\). Values for the clump volume filling factor, \(f_c\), derived from radiative transfer studies of the dust emission in SN ejecta typically range from 0.05 to 0.2 (). Assuming \(f_c = 0.05\), we derive that the total volume in the form of clumps in the SN ejecta is \(V_{tot, c}(100)= f_c \times V(100) = 8.6 \times 10^{45}\) cm\(^{-3}\). For the ejecta mass given in Table [\[tab4\]](#tab4){reference-type="ref" reference="tab4"}, we derive a typical clump mass ranging from \(10^{-3}\) \(M_{\odot}\) to \(3 \times 10^{-3}\) \(M_{\odot}\), where the number of clumps \(N_c\) is 1000 and 3000, respectively. These \(N_c\) values are usually used in radiative transfer models. Our derived clump mass is also consistent with values obtained from 3D hydrodynamic models of SN explosion (). The number density of one oxygen-rich clump at day 100 post explosion is then given by \[\label{eq6} n_c(100) = {M_c \over{V_{tot, c}(100)/ N_c}},\] with \(n_c(100) = 3.4 \times 10^{10}\) cm\(^{-3}\) for \(N_c=1000\). This gas number density is roughly a factor of 200 larger than the values listed in Table [\[tab4\]](#tab4){reference-type="ref" reference="tab4"} for the O-rich zones. Using Eq. [\[eq3\]](#eq3){reference-type="ref" reference="eq3"}, we recover a typical oxygen-rich clump number density 330 years after explosion of \(\sim\) \(20\) cm\(^{-3}\), which is consistent with the value of \(100\) cm\(^{-3}\) inferred by DS10 for the clump density not yet shocked by the RS. If we assume an inter-clump medium number density of \(1\) cm\(^{-3}\) and \(0.1\) cm\(^{-3}\) (, BS90), we recover a clump/inter clump density contrast \(\chi\) value of between \(\sim\) 20 and 200, respectively. The density contrast used by DS10 is \(100\) and that assumed by Sylvia et al. (2012) is \(1000\). For a range of initial unattenuated RS shock speeds of \(1000-2000\) km s\(^{-1}\), the RS velocity through the clump ranges from \(\sim\) \(35\) km s\(^{-1}\) to \(200\) km s\(^{-1}\). These RS velocities have been modelled by BS90 and we use their results as a basis for our RS analytical models. The RS parameters are listed in Table [1](#tab5){reference-type="ref" reference="tab5"}. We assume five velocities for the shock relative to the clump, an initial pre-shock density in the clump of \(n_0=100\) cm\(^{-3}\) and a PIR length \(L_{PIR} = 2\times 10^{11}\) cm for all models, as derived by DS10. In the PIR, we assume the oxygen column densities listed in Table 10 of BS90 for the various shock speeds listed in Table [1](#tab5){reference-type="ref" reference="tab5"}. The time length of the PIR is derived assuming the conservation of momentum through the shock front and is given by \[\label{eq7} t_{PIR} = {N(O) \over{n_2\times v_2}} = {N(O) \over{n_0\times V_s}},\] where \(N(O)\) is the oxygen column density, and \(v_2\) the gas velocity in the PIR. The density in the PIR \(n_2\) is derived from the definition of the PIR length \(L_{PIR}\) given by \[\label{eq8} L_{PIR} = v_2 \times t_{PIR} = {n_0 \over n_2} \times V_s \times t_{PIR}.\] The temperature in the PIR is that derived by DS10 for the high velocity shocks and is fixed to \(1500\) K. We also test the impact on the PIR chemistry of the higher PIR temperature derived by BS90 (\(T_2=4500\) K). For the slowest shock, we take the PIR temperature derived for that shock speed by BS90. Non-thermal processes include the penetration of the PIR by UV photons coming from the HR. These UV photons partially sustain the level of ionisation in the PIR via the photoionisation of oxygen. BS90 estimate the photon fluxes emergent from the HR as a function of the shock velocity and energy band. We consider the ionisation of OI in various levels of OII involving UV photons with energy from between 13.62 and 35.12 eV as the main source of electrons in the PIR. UV flux values are listed in Table 5 of BS90 for the shock speeds considered in this study. The unattenuated photoionisation rate for oxygen \(\zeta_{0,\nu}\) at frequency \(\nu\) is given by \[\label{eq9} \zeta_{0,\nu} = \int_{\nu} \sigma_{\nu} \times 4 \pi \frac{J_{\nu}}{h\nu}\times d\nu,\] where 4\(\pi\)J\(_{\nu}\) is the radiation intensity averaged over the solid angle and \(\sigma_{\nu}\) is the photoionisation cross section of oxygen taken from at the frequency \(\nu\). The attenuated rate due to extinction in the post-shock gas is calculated as follows: we omit potential attenuation from dust grains present in the PIR (this will be considered in paper II.). We estimate the unattenuated rate for the various frequency bands \(k\) and the integrated fluxes given in Tables 3 and 5 of BS90, respectively, and calculate the frequency-dependent optical depth \(\tau_k\) for each frequency band \(k\), which is given by \[\label{eq10} \tau_k(t) = \sigma_k \times v_2\times t \times n_2.\] where \(t\) is the time in the PIR. The total attenuated photoionisation rate at time \(t\) in the PIR is then given by \[\label{eq11} \zeta(t) = \sum_k \zeta_{0,k}\times exp [-\tau_k(t)],\] where \(\zeta_{0, k}\) is the unattenuated photoionisation rate in the frequency band \(k\). The unattenuated photoionisation of oxygen is then included as a non-thermal process in the chemical scheme described in §[2](#chem){reference-type="ref" reference="chem"} with the rate calculated from Equation [\[eq9\]](#eq9){reference-type="ref" reference="eq9"} for the frequency bands \(k\), and the attenuated rate is calculated at each integration time step according to Equation [\[eq11\]](#eq11){reference-type="ref" reference="eq11"}. # Results {#res} We study the variation of the abundances and masses of chemical species and dust clusters that form in the ejecta of the Cas A supernova precursor and present the results in §[4.1](#rescasa){reference-type="ref" reference="rescasa"}. We then focus on an oxygen clump in Cas A whose chemical composition has been derived in §[4.1](#rescasa){reference-type="ref" reference="rescasa"}, and study the impact of the reverse shock on such a dense clump. Various reverse shock velocities that correspond to clump/interclump density enhancement of \(100-1000\) are considered and results are presented in §[4.2](#resRS){reference-type="ref" reference="resRS"}. ## Molecules and dust clusters in the Cas A supernova progenitor {#rescasa} Type IIb SNe have low-density ejecta compared to their type II-P counterparts. In §[4.1.1](#lowden){reference-type="ref" reference="lowden"}, we present the results for a stratified ejecta whose parameters are given in Table [\[tab4\]](#tab4){reference-type="ref" reference="tab4"}. We define this low-density Type IIb SN ejecta as our \"standard case\". Results on the impact of increasing the gas density on the chemical composition of the gas and the dust are presented in §[4.1.2](#highden){reference-type="ref" reference="highden"}. ### Low density Type IIb ejecta {#lowden} The masses of the prevalent molecules and dust clusters that form in all the ejecta zones as a function of post-explosion time are shown in Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} for the \"standard case\". The dominant species is SiS which efficiently forms in zone 1A/B of the ejecta from the radiative association reaction \[\label{R1} Si + S \longrightarrow SiS + h\nu.\] This reaction has already been identified as the prevalent formation process for SiS in the innermost zones of the denser ejecta of Type II-P SNe (SC13). This reaction is temperature-independent and can therefore proceed at late post-outburst time and ensure the growth of SiS mass with time. SiS destruction is provided by the reaction with Ar\(^+\). CO forms in the oxygen-rich zones 3A/B and 2 via the radiative association reaction \[\label{R2} C + O \longrightarrow CO + h\nu,\] and the reaction with O\(_2\), \[\label{R3} C + O_2 \longrightarrow CO + O.\] CO destruction stems from charge-exchange reaction with O\(^+\) and with reaction with Ne\(^+\). Silicon monoxide, SiO, primarily forms in zones 1B and 2, with a small contribution from zones 3A/B, from the radiative association process \[\label{R4} Si + O \longrightarrow SiO + h\nu.\] and is destroyed by the reaction with Ne\(^+\). Its abundance increases until \(\sim\) day 1300 when SiO starts to be depleted in SiO dimers, Si\(_2\)O\(_2\), and trimers. Finally, SO and O\(_2\) also form from radiative association reactions in zone 1A and 1B/2/3, respectively. Inert gas ions do not efficiently recombine at early time because of the low ejecta gas density and their abundance gradually decreases owing to the decrease in \(^{56}\)Ni mass and \(\gamma\)-ray photons with time. The formation of the prevalent molecules is then postponed to \(t> 1000\) days, in contrast with Type II-P SNe (SC13). The masses of the most abundant synthesised dust clusters are also shown in Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. A small amount of Si\(_5\)O\(_5\) is formed in the O-rich zones 1B, 2, and 3A/B out of SiO and its polymerisation. The dimerisation of SiO is possible after day 1000 but further growth is inhibited owing to the very low ejecta gas density. No significant amount of silicate dust clusters is formed in the O-rich zones of the ejecta because the first chemical steps in the synthesis of forsterite and enstatite dimers involve neutral-neutral processes with a moderate activation barrier (, SC13), and higher gas temperatures than those found at day \(\sim\) 1000 in the ejecta are thus required. The low gas density at day \(> 1000\) also hampers the efficient nucleation of silicate clusters. Some SiC dimers form in the carbon-rich outermost zones 4A/B after day 1800, but as for SiO polymers, their later growth is hampered by the low gas density at late times. The masses of molecules and dust clusters at 3000 days after explosion and for the standard case are summarised in Table [2](#tab6){reference-type="ref" reference="tab6"}. We see that the mass fraction of the ejected material in the form of molecules only amounts to \(0.2\) % and that dust clusters form in small amount. Therefore, the low-density, homogeneous ejecta of a Type IIb SN seems to be inefficient at forming significant quantities of dust clusters and grains. These results are in contrast with the theoretical study of dust formation in the SN progenitor of Cas A presented by Nozawa et al. (2010), where the authors derive a total dust mass of \(0.167\) \(M_{\odot}\) for similar ejecta conditions and a complex chemical composition for the dust, including silicates, carbon and metal oxides. Their study of dust synthesis is based on a CNT formalism and assumes that all available carbon and silicon atoms are locked up in CO and SiO when the C and Si elemental abundances are less than that of oxygen. The formalism fails to include the chemistry of the ejecta gas phase, and the formation and destruction of molecules and dust clusters from chemical kinetics. We find that only \(\sim\) \(1\times 10^{-5}\) \(M_{\odot}\) of silicon carbide, silica, and alumina dust clusters can form for the standard case. Our results point to the importance of the nucleation phase, which describes the formation of molecules and dust clusters out of the elements comprised in the ejecta zones, as a bottleneck to dust condensation. This issue has already been addressed by Cherchneff & Lilly (2008), Cherchneff & Dwek (2009, 2010) and SC13 and will be discussed in §[5](#dis){reference-type="ref" reference="dis"}. On the observational front, Spitzer data of Cas A reveal that the IR excess due to dust in the remnant can be reproduced by an ensemble of dust grains of various composition, including silicates, metal oxides and sulphides, pure metal grains and some amorphous carbon grains (). Dust is clearly present in the diffuse remnant phase, implying that it could form in the denser ejecta of the supernova. This is in contrast with our present findings and forces us to reconsider the impact of the gas density on the nucleation of dust. ### Impact of gas density on the chemical composition of the ejecta {#highden} We now increase the ejecta number density of our standard case by various factors \(x\) ranging from between 10 and \(2000\) (the standard case corresponds to \(x=1\)), and study the nucleation phase and the chemical composition of the gas. This increase is applied to all zones of the ejecta at day 100 and the gas number density follows the time dependence given by Equation [\[eq3\]](#eq3){reference-type="ref" reference="eq3"}. We do not aim at modelling a clumpy SN ejecta but want to study the impact of density enhancements that characterise clumps on the ejecta chemistry. According to §[3.2](#revsh){reference-type="ref" reference="revsh"}, clumps are characterised by a number density roughly 200 times larger than that of the homogeneous ejecta given in Table [\[tab4\]](#tab4){reference-type="ref" reference="tab4"}. Results on the species masses for a density enhancement of \(200\) are then presented in Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}. The masses of both molecules and dust clusters have significantly increased. The chemical processes responsible for the formation of the prevalent molecules vary from those reported in the previous section because of the high gas density. While radiative association reactions contribute to the synthesis of all molecules in the ejecta, bimolecular processes including reactions with O\(_2\) contributes in forming CO, SO, and AlO in the various zones. SiS is formed essentially in the innermost zones 1A/B. As shown by SC13 for Type II-P SNe, CO is mostly synthesised in the O-rich zones 2 and 3A/3B and does not trace the formation of carbon dust clusters. SiO forms essentially in zone 1B and 2 and gets quickly depleted into silicate clusters after day \(\sim\) 600. The dust clusters start to form as early as 400 days post-explosion for the pure metal clusters, and significant masses of alumina and forsterite clusters form in the O-rich zones after day 800. The masses of dust clusters at day 3000 are shown in Figure [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} as a function of gas number density increase in the ejecta. The build-up of the chemical complexity of the dust in the ejecta is clearly seen. As seen in the previous section, small amounts of SiC, silica, and alumina dust clusters form in the \"standard case\" ejecta given by the explosion model of a Type IIb SN, with upper limit on the dust mass not exceeding \(\sim\) \(1\times 10^{-5}\) \(M_{\odot}\). More complex dust types are synthesised as the density increases, with the formation of alumina, SiC, and pure metal clusters (silicon, iron, and magnesium) when the density is raised by a factor of 10. For a density enhancement factor of 200, the prevalent dust clusters are alumina and forsterite, with SiC and pure metal clusters. The composition of the dust formed for the enhancement factor of 2000 resembles that derived for a SN of type II-P and prevalently includes alumina and forsterite (SC13). The carbon dust clusters seem to be the most density-dependent clusters and start forming for the largest increase in the density. In a high density media, carbon dust clusters form in the outermost zone of the ejecta from the initial production of the C\(_2\) carbon chain, which grows via C and C\(_2\) addition to form long carbon chains and the first cyclic ring C\(_{10}\). These molecules form at late post-explosion time (\(\sim\) 1000 days) once the He\(^+\) ion abundance drops owing to ion recombination and decrease in the \(\gamma\)-ray flux (SC13). In the case of carbon chains, the synthesis of C\(_{10}\) involves many chemical steps and necessitates the high gas density to facilitate the recombination of He\(^+\). For the enhancement factor \(x= 200\), the gas density is not sufficient to weaken the He\(^+\) recombination and permits the survival of He\(^+\) ions at very late times. Therefore, almost negligible masses of carbon dust precursors are produced as seen in Figure [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}. The masses of molecules and dust for the cases \(x=200\) and \(x=2000\) are summarised in Table [2](#tab6){reference-type="ref" reference="tab6"} at day 3000 after explosion. We clearly see that the efficiency at forming molecules and dust clusters increases with gas density enhancement but species are affected in different ways. For example, O\(_2\) and SO are more responsive to a rise in gas density owing to the larger number of chemical reactions involved in their formation and the fact that SO forms from O\(_2\) at high density (SC13). More generally, we conclude that a low-density stratified ejecta as derived from non-clumpy explosion models of Type IIb SNe are not conducive to the synthesis of large amounts of molecules and are almost dust-free. High-density clumps thus need to be considered as the main sources of molecules and dust in the SN ejecta, to explain the presence of dust observed at IR wavelengths in Cas A. These clumps may retain the chemical composition of the ejecta zones from where they originate, evolve, and are shocked by the reverse shock in the remnant phase. ## Molecules and dust clusters in the reverse shock {#resRS} FMKs have been observed in Cas A and interpreted as dense ejecta clumps that experience the crossing of the RS (). Their chemical composition includes a large fraction of oxygen atoms, along with neon, silicon, sulphur, argon and iron. This specific composition coincides with that of the oxygen-rich zones of the SN ejecta. Radiative shock theoretical models are able to reproduce optical and IR lines observed in these knots and provide information on the physical conditions at the reverse shock (, BS90,, DS10). We want to assess the impact on the RS on a FMK in terms of molecule and dust cluster reprocessing, and thus consider an oxygen-rich clump whose chemical composition is derived from our SN ejecta study presented in §[4.1.2](#highden){reference-type="ref" reference="highden"}. We follow the chemistry in the post-shock gas using the analytical model for the reverse shock illustrated in Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} and described in §[3.2](#revsh){reference-type="ref" reference="revsh"}. We consider the five shock velocities of Table [1](#tab5){reference-type="ref" reference="tab5"}. According to BS90, oxygen atoms are fully ionised to different ionisation states in the HR. The high HR temperature (\(\sim\) 10\(^6\) K) ensures that all molecule and dust clusters present in the pre-shock clump are destroyed by the RS and atoms are ionised in the HR. We then model the chemistry in the PIR, assuming all atoms are singly ionised as initial conditions and study the chemistry, including the potential reformation of small dust clusters. Abundances for the prevalent atoms, ions, and molecules in the PIR are shown in Figure [\[fig5\]](#fig5){reference-type="ref" reference="fig5"} for the 200 km s\(^{-1}\) shock. This shock model corresponds to the PIR conditions derived by DS10 with a different initial chemical composition. The ionisation of O atoms by the UV field coming from the HR is effective at maintaining a high ionisation fraction until \(\sim\) 150 days, despite the rapid recombination to O atoms before \(\sim\) 25 days. O\(^+\) ions then recombine and exchange charge with other atoms to sustain an ionisation fraction of \(\sim\) \(0.03\), specifically through charge exchange reaction with Mg. Molecules chiefly reform from the gas phase in the PIR and CO has the largest abundance, followed by SiO, O\(_2\), and SiS. The formation processes involved are radiative association reactions such as Reactions [\[R1\]](#R1){reference-type="ref" reference="R1"}, [\[R2\]](#R2){reference-type="ref" reference="R2"} or [\[R4\]](#R4){reference-type="ref" reference="R4"}. Other bimolecular processes that are active in the dense ejecta of Type II-P SNe are not efficient at the low gas number density of the PIR (in this case \(n_{gas} = 10^6\) cm\(^{-3}\)), e.g., Reaction [\[R3\]](#R3){reference-type="ref" reference="R3"}. CO, O\(_2\) and SiS reform with lower abundances by a factor \(\ge 1000\) than their initial values in the clump before the passage of the RS. SiO is an exception because the reformed gas-phase abundance is higher than its pre-shock value. As seen in §[4.1.2](#highden){reference-type="ref" reference="highden"}, SiO is quickly depleted into dust clusters in the SN ejecta that are later destroyed by the RS. The PIR conditions are then favourable to the formation of SiO from the gas phase, which traces the destruction of silicate clusters by the RS. Similar trends operate for our lower shock velocity models shown in Figure [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}. A lower shock speed combined with a shorter PIR duration result in the reformation of molecules with lower abundances as the RS velocity decreases. The physical conditions pertaining to the 140 km s\(^{-1}\) shock of BS90 in Figure [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} are similar to the PIR model derived for the 200 km s\(^{-1}\) by DS10, and the species abundances for both models are akin (see Figure [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}). CO is by far the more abundant species that reforms from the gas phase via Reaction [\[R2\]](#R2){reference-type="ref" reference="R2"}. The faster the RS, the larger the molecular content after the shock passage and the greater the chemical complexity of the PIR. When we consider the high PIR temperature derived by BS90 for their shock models with conduction (\(T_2=4500\) K), no major changes in the abundance trends occur because most of the chemical routes that control the formation of molecules at the low gas number densities of the PIR are radiative association reactions. These processes are slow and temperature-independent. Dust clusters reform in the PIR from the gas phase for all shock velocities but with extremely low abundances (\(\le 10^{-20}\)) because of the low gas density involved. We explore the possibility of synthesising dust clusters by artificially increasing the gas number density in the PIR for the 140 km s\(^{-1}\) RS. We consider two PIR enhanced densities, \(10^7\) cm\(^{-3}\) and \(10^9\) cm\(^{-3}\), and results are presented in Figure [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}. We see that even the largest gas density fails at building up significant amounts of dust clusters, with a maximum abundance of \(\sim\) \(10^{-7}\) for forsterite and alumina dimers and (SiO) trimers. The PIR corresponds to a region where both the gas temperature and number density are high compared with values characterising the pre-shock gas in the clump and non-shocked gas in the remnant. This environment is thus more conducive to dust cluster formation from the gas phase. If dust clusters are unable to form from the gas phase in this region, we can then anticipate that the dust grains destroyed by the RS crossing dense gas clumps are unable to reform in the remnant from the gas phase. We have so far considered the chemistry of the PIR region after the passage of the RS. Since molecules reform but dust clusters do not for the standard physical conditions of the PIR region given by Table [1](#tab5){reference-type="ref" reference="tab5"}, we want to investigate whether molecules could keep forming over a time span \(t >> t_{PIR}\). Following Sutherland & Dopita (1995), the clump-crushing time is given by \(\tau_{cc} = {\chi}^{1/2}/v_0\), where \(v_0\) is the pre-RS velocity and \(\chi\) the density contrast defined in § [3.2](#revsh){reference-type="ref" reference="revsh"}. According to Silvia et al. (2010), typical clump-crushing times for the \(\chi\) factors and pre-RS velocities considered in this study range from 6 to 100 years. We thus consider the RS shock model with \(V_s=200\) km s\(^{-1}\), keep the PIR number density constant over a period of \(\sim\) 15 years, and decrease the gas temperature to 300 K to account for radiative cooling. The abundances of the prevalent molecules are shown in Figure [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}. Most molecules slowly keep forming from radiative association processes while dust clusters form with negligible abundances. The final abundance values for molecules at \(\tau_{cc}\) remain all below \(10^{-3}\) for \(6 < \tau_{cc}< 100\) years. We can thus conclude that molecules quickly reform in the dense and warm PIR region and continue forming in the post-RS gas till \(t = \tau_{cc}\). We also confirm that dust clusters are unable to form from the gas phase in the PIR and at later times in the shocked clump. # Discussion {#dis} We have investigated the synthesis of molecules and dust clusters in the stratified ejecta of the Cas A Type IIb SN progenitor. Focussing on the oxygen-rich material formed in the oxygen-rich zones of the ejecta, we have studied their processing by the reverse shock in a dense clump of the Cas A remnant. Our results highlight the following points - The stratified, homogeneous ejecta of a Type IIb SN is too diffuse to form large amounts of molecules and dust. For normal ejecta conditions, only \(\sim\) \(10^{-2}\) \(M_{\odot}\) of molecules and \(\sim\) \(10^{-5}\) \(M_{\odot}\) of silicon carbide and silica dust clusters form. When the gas density is raised, the masses of molecules and clusters also increase, and most importantly, the chemical complexity of the dust composition is enhanced. - In the remnant, molecules and dust clusters in the gas phase of oxygen-rich clumps are destroyed by the crossing of the reverse shock but chemical species, including CO, SiO, SiS, and O\(_2\) reform in the post-shock gas. Stronger shocks lead to denser post-shock photo-ionised regions and thus higher abundances of reformed molecules. - No clusters of silicate, silica, silicon carbide, amorphous carbon, or pure metal grains can reform in the back of the shock because the post-shock gas densities are too low to permit the chemical build-up of dust clusters and the formation of dust grains. The main results of this study point to the importance of density in the formation processes of molecules and dust in SNe and their remnants. In the homogeneous ejecta of a Type IIb SN, we find that no molecular cluster precursors to dust could form at the low gas density involved. This points to the complexity of the dust formation processes at a microscopic scale, and the fact that dust nucleation is driven by non-equilibrium chemistry in the gas-phase. Existing studies based on the CNT formalism ignore the dust nucleation bottleneck and indicate fairly large masses of grains formed in Type IIb SNe (e.g., ). Our present results are in contrast with these studies and show that the formation of dust is extremely gas density-dependent. This dependency does not pertain to the condensation phase of dust clusters but to their formation phase, i.e., nucleation out of the gas-phase. Increase in gas density by a factor of \(200-2000\) permits an efficient nucleation of dust grains and indicates the need for non-homogenous, Type IIb SN ejecta in the form of clumps or knots. The chemical complexity of the dust grains that form in SN ejecta is also density-dependent and grows with increasing gas density. In particular, carbon dust only forms at high gas number densities and traces the densest SN ejecta conditions. In Cas A, the IR spectra measured by Spitzer indicate a variety of spectral signatures ascribed to warm dust which vary according to position and observed atomic lines in the remnant (). These grains include silicates, alumina, pure metals, carbides, iron sulphide, and amorphous carbon and have been processed by the RS, but they originate from the SN ejecta that led to Cas A. According to our results, the SN ejecta gas must be in the form of dense clumps, with density contrast of at least 200 compared to the homogeneous Type IIb SN ejecta, to properly account for the dust chemical diversity in Cas A inferred from the IR data. More generally, the chemical type of dust detected in SNe and SNRs can be used as a tracer of the physical conditions of the dust birthplace, a denser gas leading to dust grains of greater chemical variety. Recent observations with Herschel of high excitation energy rotational lines of CO in one region of Cas A where CO was already detected at the RS position with Spitzer have brought evidence of hot and dense gas in the remnant (). According to the line analysis, this O/CO-rich region is characterised by warm and dense gas (\(n_{gas} \sim 10^6\) cm\(^{-3}\)), and probably corresponds to post-RS gas. Such conditions are found in the wake of our 200 and 140 km s\(^{-1}\) RS crossing a dense clump where CO reforms on a time scale of \(\sim\) 100 days. These observations therefore support the present model results which indicate that molecules, including CO and SiO, reform after the RS passage in sufficient amount to warrant detection. The molecular content of Cas A should then include two components reflecting various evolutionary phases of the remnant. A cold component would comprise species formed in the SN clumpy ejecta (e.g., SiS, CO, O\(_2\), and SO), that have not yet been processed by the RS and have large abundances. Such a cool component has been confirmed in SN1987A by the detection of low excitation rotational lines of CO with ALMA (). A warm/hot component would coincide with shocked clumps and newly reformed molecules (e.g., CO, SiO) in the post-reverse shock gas, as seen in Cas A with Herschel. The warm CO component should have a lower abundance than its cold counterpart. The amount of reformed molecules depends on the shock strength, the faster shocks producing larger quantities of chemical species in the post-shock gas. For all shock speeds, dust clusters are unable to reform in sufficient quantities in the post-shock gas to guarantee a viable dust formation pathway from the gas phase in SN remnants.

{'timestamp': '2014-01-23T02:06:29', 'yymm': '1401', 'arxiv_id': '1401.5594', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5594'}

# Introduction {#sec:intro} Over the last two decades, more than 1000 planets outside our Solar System (exoplanets) have been confirmed and several thousands are classified as candidates. Beyond the sheer detection, observation techniques based on the radial velocity and transit method also allow one to infer some planetary properties such as mass, radius, and orbit. Today, much research is devoted to the interpretation of the statistical distribution of these properties. One of the current research questions in the field of exoplanets is how different orbital and planetary parameters affect the habitability of a planet, i.e. its ability to host life. Commonly, a planet is considered habitable if its surface conditions allow for the existence of liquid water. The estimation of habitability is then usually provided as an estimation of the habitable zone (HZ), i.e. the range of distances from the host star that would allow for habitable conditions. Consequently, two different processes define the inner and outer boundaries of the HZ: at the inner boundary, the run-away greenhouse effect leads to a complete evaporation of water at the surface; at the outer boundary, a completely frozen surface limits the habitability of a planet. Climate models can provide insights to the investigation of exoplanets and their habitability. Ranging from toy models to general circulation models, they allow one to study the effect of different processes with variable degree of approximation and computational cost. Radiative-convection models (RCMs) were the first climate models applied to the investigation of exoplanets (e.g. ). In these one dimensional vertical models, the atmosphere of the planet is treated as a single column. As they do not capture effects of temporal or spatial heterogeneity, they were subsequently complemeted by latitudinally resolved energy-balance models (EBMs) (e.g. ). More recently, also general circulation models (GCMs) have been employed in habitability studies and, more generally, to study their climates and circulations. Estimates of the HZ differ among these models, as they include different processes at different degree of approximation. Simulations with RCMs yield an extent of the HZ of an Earth-like planet orbiting around a Sun-like star from 0.99 AU to 1.70 AU. Results obtained from EBMs and GCMs, however, indicate that RCMs tend to underestimate the extent of the HZ due to processes that are not or not sufficiently represented in these models. Simulations with EBMs show that the inclusion of seasonal variability due to either a non-zero planetary obliquity, an eccentric orbit or both can extend the outer boundary of the HZ relative to results from RCMs. As recent work shows, the extension of the HZ can be further increased if planetary obliquity and orbital eccentricites vary over time. Moreover, results from a GCM show that a better representation of cloud feedbacks can lead to an extension of the HZ relative to results from RCMs also at its inner boundary. Large seasonal variability might indeed be a common feature of extra-solar planets as detected planets exhibit a large diversity of orbital eccentricities. Around 35 % of the detected exoplanets are located on an orbit with an eccentricity \(e > 0.2\) and around 9 % on an orbit with \(e > 0.5\) (Figure [\[fig:ecc\]](#fig:ecc){reference-type="ref" reference="fig:ecc"}). Highly eccentric orbits feature dramatic intra-annual variability: while for \(e=0.2\) a planet at periastron receives about twice the amount of energy at apoastron, this factor increases to around 9 for \(e = 0.5\). Also the polar obliquity of a planet \(\theta\) plays an important role in determining seasonal variability. In particular the combination of a high value of \(e\) and a high value of \(\theta\) can lead to extreme seasonal variability. Multistability of the climate state further complicates estimates of the outer boundary of the HZ. Since the experience gathered with the first EBMs, it is well known that the climate of a planet can exhibit two stable states with different degree of ice-coverage, usually referred to as warm (ice-free or only partially ice-covered) and cold (completely or mostly ice-covered) climate state. Evidence suggests that Earth might have entered and exited a cold state several times in the past, and the transition to the cold state can be reproduced with state-of-the-art GCMs. While the multistability property has been extensively discussed in the context of paleoclimate (see also ), its implications for the habitability of exoplanets is a relatively new area of research. provided an investigation of bistability in a thermodynamic framework, which has recently been extended to include the parameters rotation rate and atmospheric opacity. The effect of obliquity and eccentricity has so far primarily been investigated with EBMs. Their results reveal that both parameters have substantial implications for the extent of the HZ, particularly on eccentric orbits. As downside of the numerical efficiency of these models, however, they rely on crude simplifications. One of their most critical simplifications is that ice is represented by simply assuming a temperature dependent surface albedo. Moreover, they do not include any dynamical processes apart from a simple diffusion parametrization of meridional heat transport. However show that the phenomenological transport efficiency relating temperature gradients and heat transport varies by an order of magnitude amongst earth-like planets at different obliquities. As these parameterizations used in EBMs for meridional heat transport are tuned to reproduce the climate of Earth, their validity is challenged by the very different shapes of atmospheric circulations and climates that can be expected on exoplanets with high polar obliquities, highly eccentric orbits, and variable degree of ice-coverage. Despite the need for simulations with climate models of higher complexity, only three studies assessed the effect of either obliquity or eccentricity on atmospheric circulations and climate with a GCM so far. They did, however, not provide a systematic investigation of the implications for habitability. This is the aim of this work. Using a GCM of intermediate complexity we explore the atmospheric circulations and climates of idealized Earth-like planets for different astronomical parameters and examine the implications for habitability. It is well known that silicate weathering affects the climate of Earth as a stabilizing mechanism acting on timescales of million of years: when surface temperatures are sufficiently low and rocks are not exposed to the atmosphere, the weathering is strongly suppressed, with an ensuing accumulation of CO2 in the atmosphere. In our study, we do not consider this process (see instead ) but keep the CO2 concentration at a fixed value, following and. Therefore, our study should be thought primarily as a parametric exploration of the effect of changing orbital parameters on a planet of given atmospheric composition rather than a realistic study of the HZ. We anticipate that seasonal variability leads to an increase of the the maximal distance between planet and host star that allows for habitable conditions. The combined effect of obliquity and eccentricity on multistability has to our knowledge not been investigated yet. Our results reveal that obliquity is crucial in determining the extent of both the warm and cold state. Moreover, eccentric orbits are generally associated with a narrower range of two stable climate states. Our simulations also show that seasonal variability primarily leads to temporally ice-free regions, which stresses the importance of different definitions of habitability. Sensitivity experiments that include the parameters azimuthal obliquity and ocean heat capacity test the robustness of our results. Since we neglect the effect of silicate weathering, our results of the outer boundary of the habitable zone can only be used as conservative estimates. Large intra-annual variations of temperature and ice-coverage found in many of our experiments however question traditional estimates with energy balance models that do not take these variations into account. The paper is structured as follows. In Section [2](#sec:model){reference-type="ref" reference="sec:model"} we introduce and discuss our model PlaSim, explain the experimental setup, and list all simulations. The main features of atmospheric circulations of planets at different obliquities and eccentricities are presented in Section [3](#sec:climates){reference-type="ref" reference="sec:climates"}. Section [4](#sec:results){reference-type="ref" reference="sec:results"} shows the implications of obliquity and eccentricity for multistability and the degree of habitability. A discussion of our results follows in Section [5](#sec:discuss){reference-type="ref" reference="sec:discuss"}. Finally, we draw our conclusions in Section [6](#sec:concl){reference-type="ref" reference="sec:concl"}. # Model and experimental setup {#sec:model} ## The Planet Simulator All numerical simulations are performed with the Planet Simulator (PlaSim), a general circulation model of intermediate complexity, freely available at *h*ttp://www.mi.uni-hamburg.de/plasim. Atmospheric processes included in PlaSim are: atmospheric dynamics, surface turbulent fluxes, clear-sky and cloudy-sky atmosphere-radiation interaction, and moist and dry convection. The atmospheric component of the model is run coupled to a thermodynamic sea ice model and a slab ocean model (50m). The model excludes any sea ice and ocean dynamics as well as tracer chemistry and vegetation dynamics. In the following we briefly describe the parameterizations of the atmospheric component of the model that are included in the current set of numerical simulations. - *Dynamic Solver.* PlaSim relies on a spectral atmospheric dynamical core to integrate the primitive equations for the vertical component of the vorticity \(\zeta=\mathbf k\cdot(\nabla\times\mathbf{v})\) and horizontal wind divergence \(\delta=\nabla_h\cdot\mathbf{v}\), temperature \(T\) and surface pressure \(p_s\). In spectral dynamical solvers the prognostic variables are represented as a sum of spherical harmonics truncated at order \(N\) (triangular truncation is implemented in our case). Due to truncation, numerical hyperviscosity \(\sim \nabla^8 (\zeta, \delta, T)\) is applied in PlaSim to \(\zeta, \delta, T\) to remove the enstrophy accumulating at the smallest resolved scales generating numerical instability. Sound waves are filtered by hydrostatic approximation. The Robert-Asselin time filter is used to suppress spurious computational modes associated with the leapfrog time-stepping scheme implemented to deal with the vorticity equation. More details about the numerical methods can be found in. - *Radiation.* The shortwave radiation scheme is based on for the clear-sky atmosphere. The solar spectral range is divided into a visible and ultraviolet part \(\lambda < 0.75\) \(\mu\) m with ozone absorption and Rayleigh scattering and a near infrared part \(\lambda > 0.75\) \(\mu\) m with water vapor absorption only. For the cloudy fraction, transmissivities and albedos, both depending on the cloud liquid water content, are calculated differently in the near infrared and visible-ultraviolet bands, following the ideas of and. The incoming solar flux \(F_{sw}^{toa}\) at the top of the atmosphere is \(F_{sw}^{toa}=S^\star \cos Z\), where \(S^\star\) is the solar constant and \(Z\) the zenith angle, a function of the latitude and time estimated according to . Longwave radiation for the clear sky relies on a broad band emissivity method. The transmissivities for water vapor, carbon dioxide and ozone depend on their local concentration and are determined according to the empirical relations of. For each layer with a certain fractional cloud cover, the longwave transmissivity and cloud albedo are determined from the cloud liquid water content, the clear-sky transmissivity and the fractional cloud cover, as discussed in . For more extensive details of the equations used to parametrize the longwave and shortwave transmissivities the reader is referred to. From the optical properties of the atmosphere (shortwave and longwave transmissivity and albedo), \(F_{sw}^{toa}\) and the surface temperature upwards and downwards radiative fluxes at the interfaces of each atmospheric layer and hence the radiative heating rates are computed. - *Surface and moist processes.* In all simulations the lower boundary is a flat surface with prescribed albedo and heat capacity. This is implemented with a shallow slab ocean model. The surface temperature therefore evolves in time according to \(C_{slab}\dot{T}_s=F_{sw}^{surf}+F_{lw}^{-}=\sigma_B T_s^4+F_{sens}+F_{lat}\), where \(C_{slab}\) is the heat capacity of the slab ocean, \(F_{sw}^{surf}\) is the net solar radiation at the surface, \(F_{lw}^{-}\) the downward longwave radiation at the surface and \(F_{sens}+F_{lat}\) the sum of sensible and latent heat fluxes at the surface. Cumulus convection is parameterized by a Kuo-type convection scheme . Large scale condensation occurs when the air is supersaturated. Because condensed water is instantaneously removed as precipitation, no water is stored in clouds. Cloud cover and cloud liquid water content are diagnostic quantities. The fractional cloud cover of a grid box is parameterized following using the relative humidity for the stratiform cloud amount and the convective precipitation rate for the convective cloud amount. The bulk aerodynamic formulas are employed for surface flux parameterizations of wind stress (\(\tau_x, \tau_y\)), sensible heat flux \(F_{sens}\) and latent heat flux \(F_{lat}\). Drag and transfer coefficients follow the approach by and. The thermodynamic sea ice model is based on the 0-layer version of the model of, which is a simplified version of the thermodynamic sea ice model. Its main simplifications are the exclusion of the capacity of the ice to store heat and the assumption of a constant temperature gradient in the ice sheet (note that for the computation of the sea ice surface temperature a heat capacity corresponding to an ice layer with a thickness of 10 cm is assumed). Despite these simplifications, a comparison with both a more sophisticated 3-layer version of the same model and the Maykut and Untersteiner model shows deviations in the seasonal cycle of ice thickness of only about tens of centimeters. Further simplifications of the model are the exclusion of oceanic heat transport and the exclusion of wind induced sea ice drift. The thickness of sea ice in equilibrium is governed by thermodynamic processes (freezing and melting) and dynamic processes (advection, collision and deformation). Only thermodynamic processes are included in our model. The ice thickness is then constrained by the diffusion of heat through the ice sheet, which primarily depends on the surface temperature and oceanic heat flux. In order to avoid artificial sources of energy, the oceanic heat flux is set to zero in our model. Its limiting effect on ice growth is accounted for by prescribing a maximal ice thickness \(h_{i, \rm{max}} = \SI{3}{m}\). This value is in good agreement with observation of equilibrium ice thickness between 2 and 3 m of 1-year and 2-year old sea ice on Earth. A detailed description of the sea ice model can be found in the PlaSim reference manual. A more detailed discussion of the model and also on neglecting dynamic sea ice processes can be found in. Depending on the ice model, different processes are involved in the ice-albedo feedback. The simplest form of the feedback captures only changes in the areal extent of the ice cover. An initial perturbation of surface temperature is then amplified or damped only if it leads to a change of the the area that is ice-covered and ice-free. More realistically, also further processes associated with the ice thickness, snow cover, or melt pond characteristics can constitute positive or negative feedback mechanisms. Yet these processes require more sophisticated models and are therefore not taken into account in our experiments. PlaSim offers a reasonable trade-off between numerical efficiency and the number of explicitly included processes. Even though all model elements have been tuned to simulate the Earth's climate, we expect them to remain reasonably accurate for climates close to Earth's climatic conditions. In addition to have been intensively used to study the climate of Earth, the model has already been employed in investigations of atmospheric circulations of terrestrial planets, snowball transitions for different astronomical and atmospheric characteristics and the climate of tidally-locked, water-trapped planets. ## Astronomical parameters: obliquity and eccentricity The two astronomical parameters investigated in this work are the obliquity \(\theta\) of a planet (the tilt of its spin axis with respect to a normal to the plane of the orbit) ranging from \(\SI{0}{\degree}\) to \(\SI{90}{\degree}\) and the eccentricity \(e\) of its orbit ranging from 0 to 1. Both parameters determine the temporal and spatial distribution of solar insolation (see for a comprehensive derivation of solar insolation as function of these two parameters; spatio-temporal patterns of insolation are also explored in ). On circular orbits the spatial distribution of annual mean irradiation at the top of the atmosphere is determined by the obliquity of the planet (Figure [\[fig:ins\]](#fig:ins){reference-type="ref" reference="fig:ins"}). The relation between irradiation at low and high latitudes is reversed at the critical value \(\theta_{c} \approx \SI{55}{\degree}\). For \(\theta \leq \theta_{c}\), low latitudes receive on annual average more radiation than high latitudes. The converse applies for \(\theta > \theta_{c}\) (Figure [\[fig:ins\]](#fig:ins){reference-type="ref" reference="fig:ins"}). This consequence of high obliquity has also been considered relevant for explaining geological records that hint at low-latitude glaciation in Earth's history (; see also for an overview), although no physical mechanism responsible for large changes of Earth's obliquity has been found. In our experiments we set \(\theta\) to \(\SI{0}{\degree},\SI{60}{\degree}\), and \(\SI{90}{\degree}\), representing planets with most intensive irradiation at the equator, weak latitudinal gradients, and the highest annual mean insolation at the poles, respectively. As we aim at a parametric investigation of how different values of obliquity influence climate, we do not account for the effect of tidal processes that can erode the obliquity of a planet on time scales of billions of years. On eccentric orbits, the insolation pattern is also influenced by the orbital longitude of periastron \(\omega\) measured from northern vernal equinox (also refered to as azimuthal obliquity). This third astronomical parameter determines the hemispheric asymmetry of the annual mean insolation. Most of our analysis deals with planets with \(\omega = \SI{0}{\degree}\), which feature an annual mean irradiation symmetric with respect to the equator. In other words, periastron and apoastron are aligned with vernal equinoxes and both hemispheres receive the same amount of energy per year. On Earth, \(\omega\) changes over time, with a current value of \(\omega \approx \SI{103}{\degree}\). In order to test the sensitivity of our results against changes in \(\omega\), we conduct two simulations with \(\omega = \SI{30}{\degree}\) and \(\omega = \SI{90}{\degree}\). The latter is equivalent to an alignment of periastron with solstice, thus maximizing the difference of mean irradiation between the northern and southern hemisphere. ## Experimental setup {#sec:model:planet} All simulations are performed with a horizontal resolution of 64 \(\times\) 32 grid points (T21 spectral resolution) and 10 vertical levels, which enables us to run a large ensemble of numerical simulations in a reasonable amount of time. In experiments not included in this paper the T21 horizontal resolution-corresponding approximately to \(500\) km-has proven to provide an accurate representation of the large scale circulation features as compared to higher resolutions (T42). Most importantly, the T21 horizontal resolution captures baroclinic eddies, which play a fundamental role in the transport of energy and momentum polewards [^1]. Direct comparison between runs obtained at the two horizontal resolution T21 and T42 (not shown) demonstrate that all main features (jet streams, meridional circulation, temperature structure) are well captured at T21. Similar experiments showed that the vertical resolution ensures a convergence of the radiative transfer and an adequate representation of the vertical structure of the atmosphere. Although the model PlaSim has been used to study the climate of planets with very different properties (see for instance for an investigation of the climate of Mars), in this work we focus on idealized planets that are Earth-like with respect to their radius, density, host star, and atmospheric composition. This choice of the setup is motivated by the presence of several studies on snowball dynamics and habitability for Earth-like planets, employing both simple EBMs and GCMs, which allows us to build on a broad basis and put our analysis in a wider context. All planets studied in this work are completely covered with water, i.e. aquaplanets. Therefore, surface properties such as surface roughness, moisture, and heat capacity are fixed and uniform all over the globe. Also the albedo of snow-and ice-free areas is uniformly set to \(\alpha = 0.069\), as corresponding to an ice-free ocean. The consequences of different land-ocean distributions or different surface materials are therefore not considered in our experiments. We acknowledge that these properties might however play a crucial role in determining spatial and temporal variability and therefore interact with orbital forcing in a complex way. For instance, a planet with one large North Polar continent can experience large temperature oscillations for highly tilted spin axes due to lower heat capacity of land than ocean, potentially leading to the formation of ice caps. show that the climate of such planet indeed also shows a strong dependence on obliquity and that its habitability exhibits a much more seasonal character than on a uniform planet. A list of parameters and their chosen values can be found in Table [\[tab:params\]](#tab:params){reference-type="ref" reference="tab:params"}. We modulate the intensity of the stellar irradiation \(S\) that reaches the top of the atmosphere in all our simulations, which can alternatively be interpreted as changing the planet-star distance. As our model is not able to simulate the physical properties of a thick, warm atmosphere with a substantial mass fraction of H\(_{\rm{2}}\)O, we limit our analysis to the outer edge of the HZ. Limitations of our model also imply that our definition of habitability differs from the maximum greenhouse limit that is usually considered as the outer boundary of the HZ. Our model does not account for the effects of silicate weathering, i.e. the stabilizing feedback of CO\(_{\rm{2}}\) accumulation in the atmosphere, acting on timescales of millions of years, once the temperature falls sufficiently low to affect carbon fluxes at the surface. This implies that our study, in fact, should be thought primarily as a parametric exploration of the effect of changing orbital parameters on a planet of given atmospheric composition. Since we neglect a feedback pushing climate towards conditions favourable for life, we deduce that our estimated extents of the HZ are always conservative providing a lower boundary for the outer edge of the HZ. In the following, whenever we use the term habitable zone (HZ) we refer to this conservative definition of habitability. The limitations of our investigation, in particular regarding the lack of representation of the silicate weathering cycle, are later discussed in Section 5. In the following we refer to the irradiation relative to the irradiation on Earth as \(\tilde{S} = S/S^{\star}\) with \(S^{\star} = \SI{1365}{Wm^{-2}}\). A list of all experiments whose results are presented in this work is shown in Table [\[tab:sims\]](#tab:sims){reference-type="ref" reference="tab:sims"}. # Climate states and atmospheric circulations {#sec:climates} The annual and zonal mean surface temperatures of simulations with \(\theta=0^\circ,\,60^\circ,\,90^\circ\) and \(e=0,\,0.5\) for ice-free climates at \(S=S^\star\) are shown in Figure [\[fig:gradient\]](#fig:gradient){reference-type="ref" reference="fig:gradient"}. Planets with low obliquities receive, on average, more stellar radiation at the equator than at the poles and usually feature a marked meridional temperature gradient. This situation is reversed at high obliquities (e.g. \(\theta=90^\circ\)) with polar regions receiving on average more radiation than equatorial regions. Large seasonal deviations from the annual mean temperature are typical of climates at high obliquity. Patterns of zonal-mean surface temperature and zonal-mean net incoming stellar radiation are shown for each month in Fig. [\[fig:sol_for\]](#fig:sol_for){reference-type="ref" reference="fig:sol_for"} and Fig. [\[fig:sol_for2\]](#fig:sol_for2){reference-type="ref" reference="fig:sol_for2"} for ice-free and snowball states for different values of obliquity and eccentricity. Because snowball states have low thermal inertia, the patterns of surface temperature follow closely the patterns of the net incoming stellar radiation. For ice-free climate states, the heat capacity of the slab ocean causes a delay of the temperature response to solar forcing of about two to three months. The seasonal cycle of insolation at both high obliquity and eccentricity has a complex pattern, shifting the times of maximum insolation in the northern and southern hemisphere closer to each other (from month 3 to month 5 and from month 9 to month 7 respectively). Although solstice occurs on both poles at the same distance from the star (azimuthal obliquity \(\omega=0^\circ\)), there is a slight asymmetry in the hemispheric temperature patterns at \(\theta=60^\circ\) and \(\theta=90^\circ\). This is due to the fact that in the northern hemisphere summer is followed by the planet getting closer to the star, while in the southern hemisphere summer is followed by the planet moving away from the star. That is, both poles receive the same overall amount of energy, but distributed differently over time, which due to the thermal inertia gives rise to the asymmetric shape of surface temperature observed in Fig. [\[fig:sol_for\]](#fig:sol_for){reference-type="ref" reference="fig:sol_for"} and  [\[fig:sol_for2\]](#fig:sol_for2){reference-type="ref" reference="fig:sol_for2"}. Horizontal heat transport is essential for smoothing the effect of extreme seasonal insolation at high obliquity. Since the net incoming stellar radiation features extreme seasonal variations, also the meridional energy transport shows extreme changes over one year (Fig. [\[fig:energy\]](#fig:energy){reference-type="ref" reference="fig:energy"}). For planets at high obliquity (Fig. [\[fig:energy\]](#fig:energy){reference-type="ref" reference="fig:energy"}(a)), the mean annual transport is equatorwards at \(\theta = 90^\circ\) and nearly zero at \(\theta = 60^\circ\). During the month of maximum irradiation in the northern hemisphere (Fig. [\[fig:sol_for\]](#fig:sol_for){reference-type="ref" reference="fig:sol_for"}), the meridional heat transport is southwards at almost all latitudes, substantially differing from the annual mean values. Deviations from the annual means become extreme for frozen planets (Fig. [\[fig:energy\]](#fig:energy){reference-type="ref" reference="fig:energy"}(c)). Eccentricity introduces only small variations to the annual mean meridional heat transport but changes the timing of its seasonal oscillations because of shifted and distorted insolation patterns (\(e=0.5\), Fig. [\[fig:sol_for\]](#fig:sol_for){reference-type="ref" reference="fig:sol_for"}(b, d)). The atmospheric meridional energy transports follows directly from the atmospheric circulation. Circulations of earth-like planets with low obliquity (\(\theta\rightarrow 0\)) are similar to those of the Earth's atmosphere, with equatorial regions warmer than polar regions, a direct Hadley cell confined in the Tropics and jet streams at mid-latitudes. Snowball Earth-like climates for planets at low obliquity are characterized by a very dry atmosphere (neither the condensation latent heat nor the water vapor greenhouse effect have any significant impact on climate) and by very weak vertical gradients. These climates have intensively been discussed in the literature. High obliquity causes a reversal of the atmospheric meridional temperature profile with respect to climates at low obliquity, because on annual mean equatorial regions receive more solar radiation than polar regions (Fig. [\[fig:u_T\]](#fig:u_T){reference-type="ref" reference="fig:u_T"}(a, c, e, f)). In Fig. [\[fig:u_T\]](#fig:u_T){reference-type="ref" reference="fig:u_T"}(a) the annual means of atmospheric temperatures and zonal winds are shown. The meridional streamfunction can be see in in Fig. [\[fig:stream\]](#fig:stream){reference-type="ref" reference="fig:stream"}(a) for the same temporal means. Since the circulations show pronounced seasonal variations, also the warmest month in the northern hemisphere during which there is the strongest meridional temperature contrast is included in Fig. [\[fig:stream\]](#fig:stream){reference-type="ref" reference="fig:stream"}. Westerly storm-tracks develop in the summer hemisphere in the lower troposphere, with a strongly sheared easterly jet aloft for the warm climate at high obliquity (FIg. [\[fig:u_T\]](#fig:u_T){reference-type="ref" reference="fig:u_T"}(b). The atmospheric meridional overturning circulation during the NH summer (Fig. [\[fig:stream\]](#fig:stream){reference-type="ref" reference="fig:stream"}(b)) is thermally direct (upwelling in the northern hemisphere and downwelling in the southern hemisphere) and extends over the whole hemisphere. The reversal of the direct cell and its shift in the southern hemisphere during the local summer results in the two small equatorial indirect cells seen in the annual mean (Fig. [\[fig:stream\]](#fig:stream){reference-type="ref" reference="fig:stream"}(a)), which hence are just an artefact of the annual averaging. In the winter hemisphere no strong westerly jet develops close the ground. show that the strongly sheared easterly winds in the mid-high troposphere are essential to determine baroclinic instability leading to the development of the westerly jet close to the ground, which are missing in the southern/winter hemisphere. Circulations of Earth-like planets at high obliquity and on eccentric orbits differ only slightly from the pattern discussed so far. The main difference is the asymmetric circulation between the two hemispheres (Fig. [\[fig:u_T\]](#fig:u_T){reference-type="ref" reference="fig:u_T"}(e, f) and Fig. [\[fig:stream\]](#fig:stream){reference-type="ref" reference="fig:stream"}(e, f)) reflecting the asymmetric temperature patterns already discussed and shown in Fig. [\[fig:sol_for\]](#fig:sol_for){reference-type="ref" reference="fig:sol_for"}(c) and Fig.  [\[fig:sol_for2\]](#fig:sol_for2){reference-type="ref" reference="fig:sol_for2"}(c). One of main features of snowball climates is the replacement of the ocean (in our simulations represented by a slab-ocean model) with a thick ice cover, which drastically reduces the thermal inertia of the ocean. As a consequence, surface temperatures respond almost instantaneously to the distribution of absorbed solar radiation (Fig. [\[fig:sol_for\]](#fig:sol_for){reference-type="ref" reference="fig:sol_for"},  [\[fig:sol_for2\]](#fig:sol_for2){reference-type="ref" reference="fig:sol_for2"}). As for snowball climates at low obliquity (for which a rich literature exists, ) snowball climates at high obliquity feature very weak vertical thermal gradients (Fig. [\[fig:u_T\]](#fig:u_T){reference-type="ref" reference="fig:u_T"}(c, g)). This is a direct consequence of the dryness of the atmosphere and therefore the lack of greenhouse effect. Important deviations from the annually-averaged temperature profile are however observed seasonally. Fig. [\[fig:u_T\]](#fig:u_T){reference-type="ref" reference="fig:u_T"}(d, h) refers to the northern hemisphere summer and shows a very stable temperature profile in the southern hemisphere, with an atmosphere which is almost isothermal, except close to the surface where a strong thermal inversion is found. The vertical profile is instead destabilized in the summer hemisphere by the intense irradiation at the ground. Air rises in the summer hemisphere forming a thermally direct cell which extends globally up to the south pole (Fig. [\[fig:stream\]](#fig:stream){reference-type="ref" reference="fig:stream"}(d, h)). The solstitial circulation patterns dominate the annal mean, resulting in a fictitious two-cell structure of the annually-averaged meridional streamfunction (Fig. [\[fig:stream\]](#fig:stream){reference-type="ref" reference="fig:stream"}(c, g)). As already noted above, changes in insolation due to eccentricity do not substantially impact the main circulation features of snowball climates at high obliquity, but introduce an asymmetric behaviour between the northern and southern hemisphere (Fig. [\[fig:stream\]](#fig:stream){reference-type="ref" reference="fig:stream"}(g)). # Multistability and habitability at extreme seasonality {#sec:results} Within a certain range of \(\tilde{S}\), two distinct climate states coexist, one warm (mostly ice-free) and one cold (mostly ice-covered) climate state. Building on previous work on the bistability and habitability of Earth-like planets the implications of seasonal variability on the stable intervals of both climate states are examined in the following. For each set of parameters, a hysteretic cycle is performed. We start the model at an irradiation around \(\tilde{S}=1\) and then gradually decrease it by steps of \(0.05\) until the planet is completely covered with ice. For each value of \(\tilde{S}\), the model is run for fifty years, of which the last thirty years are used for the computation of climatological quantities. Once the planet is completely ice-covered, \(\tilde{S}\) is step-wise increased again until the planet is completely ice-free. ## Effect of obliquity on ice formation and melting {#sec:results:freeze} We begin our analysis by investigating the effect of obliquity on ice-formation and ice-melt on circular orbits, where the obliquity determines both the spatial distribution of insolation and its temporal variability. Our results show that both the mean distribution and its variability have substantial consequences for the transitions between the completely ice-free and the completely ice-covered climate state. Planets with low obliquity receive most radiation at the equator, whereas planets with high obliquity receive most of their energy at the poles (Figure [\[fig:ins\]](#fig:ins){reference-type="ref" reference="fig:ins"}). Consequently, when decreasing \(\tilde{S}\), ice formation starts at the poles and the equator for \(\theta = \SI{0}{\degree}\) and \(\theta = \SI{90}{\degree}\), respectively (Figure [\[fig:trans\]](#fig:trans){reference-type="ref" reference="fig:trans"}). This can be correlated with the spatial pattern of surface temperature (not shown) that roughly follows the distribution of annual mean insolation (Figure [\[fig:ins\]](#fig:ins){reference-type="ref" reference="fig:ins"}). Planets with low obliquity are more susceptible to global freezing than planets with high obliquity (Figure [\[fig:trans\]](#fig:trans){reference-type="ref" reference="fig:trans"}). On planets with low values of \(\theta\), strong gradients of meridional temperature exist that can allow for the existence of ice at the poles despite relatively high temperatures at the equator (Figure [\[fig:gradient\]](#fig:gradient){reference-type="ref" reference="fig:gradient"}). Once ice forms at the poles at a certain value of \(\tilde{S}\), a further decrease of \(\tilde{S}\) leads to an advancement of the ice edge towards the equator. This leads to a higher planetary albedo and thus a further cooling of the planet, which finally results in a global freezing at relatively high values of \(\tilde{S}\). In contrast, temperature distributions on planets with high values of \(\theta\) are more homogeneous (Figure [\[fig:gradient\]](#fig:gradient){reference-type="ref" reference="fig:gradient"}). This implies that ice does not form for relatively low values of \(\tilde{S}\) and hence the ice-albedo feedback does not amplify the cooling due to less incoming radiation. Yet once a critical threshold of \(\tilde{S}\) is crossed and the temperature falls below the freezing point at the equator, a rather sharp transition to a completely frozen state takes place. Consequently, the transition to the completely frozen state occurs in a more gradual manner and for a higher value of \(\tilde{S}\) on the planet with \(\theta = \SI{0}{\degree}\) than for \(\theta = \SI{90}{\degree}\) (Figure [\[fig:trans\]](#fig:trans){reference-type="ref" reference="fig:trans"}). The dynamics of ice-formation and ice-melt is reversed for planets with high and low obliquity, because the geography of ice is reversed in the two cases. Melting the ice sheet of planets with \(\theta = \SI{0}{\degree}\) requires relatively high values of \(\tilde{S}\), as seasonal peak irradiation is not strong enough to trigger an initial appearance of ice-free areas at the equator. Once the critical threshold is crossed and initial ice-free areas appear, the ice-albedo feedback leads to a sharp transition. As opposed to this, planets with high values of \(\theta\) feature ice-free areas at the poles already at lower values of \(\tilde{S}\). Hence also the transition to a completely ice-free state occurs in a more gradual manner (Figure [\[fig:trans\]](#fig:trans){reference-type="ref" reference="fig:trans"}). ## Effect of obliquity and eccentricity on multistability We compute the global mean surface temperature for different values of irradiation \(\tilde{S}\) in order to summarize information about the climate state into one observable. The choice of surface temperature as observable is motivated by two aspects. First, the observable can directly be linked to usual habitability conditions requiring the surface temperature to remain between the freezing and boiling point of water. Second, the observable depends strongly enough on ice-coverage to show a clear shift at the point of transition between the two distinct climate states. Also alternative ways to compress information about the climate state in one quantity based on thermodynamics have been proposed and applied in similar studies, revealing a good correlation between global mean surface temperature, material entropy production and quantities indicative of the efficiency of the climate machine. However, because our focus is on average climate conditions, we choose surface temperature as observable in our analysis of climate stability. In Section [4.3](#sec:results:hab){reference-type="ref" reference="sec:results:hab"} we develop a further index that is more relevant in the discussion about habitability. In all experiments, the relationship between global mean surface temperature and solar irradiation exhibits two branches over a certain range of \(\tilde{S}\). Over this range of \(\tilde{S}\), the actual climate state depends on the history of the planet (Figure [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}(a, b)). This phenomenon is known as climate bistability, or multistability. Where both climate states exist we distinguish them as warm and cold (snowball) climate state. As explained above, the more homogeneous the distribution of surface temperature (Figure [\[fig:gradient\]](#fig:gradient){reference-type="ref" reference="fig:gradient"}), the lower the critical value of \(\tilde{S}\) at which the planet enters the cold state. Here, an obliquity \(\theta = \SI{60}{\degree}\) shows the largest extent of the warm state, followed by \(\theta = \SI{90}{\degree}\) (Figure [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}(a, b)). In the cold state, ice-free areas appear once seasonal insolation at any latitude is intense enough to melt the ice cap. Therefore, the extent of the cold state is minimal for high values of obliquity (Figure [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}), which are associated with very high seasonal peak irradiation. The effect of obliquity on the extent of the warm state is qualitatively the same for a circular orbit as for an orbit with \(e=0.5\). Also on such orbit, the largest extent of the warm state is shown by the planet with \(\theta = 60^{\circ}\), followed by \(\theta = 90^{\circ}\) (Figure [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}(a, b)). However, the range of bistability shrinks for all experiments with \(e=0.5\) compared to the experiments with \(e=0\). Moreover, all experiments with \(e=0.5\) show a larger extent of the warm state than the corresponding experiments with \(e=0\) (Figure [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}(a, b)). This can in part be explained with the increase of the global and annual mean insolation \(\langle I \rangle_{e}\) on an orbit with eccentricity \(e\), whose effect is investigated separately below: \[\langle I \rangle _{e} = \frac{\langle I \rangle _{e=0}}{\sqrt{1-e^2}}. \label{eq:ins_ecc}\] Also the extent of the cold state is strongly affected by the eccentricity of the orbit. In particular for \(\theta= 0^{\circ}\), the width of the cold state is much smaller for \(e=0.5\) than for \(e=0\) (Figure [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}(a, b)). This is due to the intense irradiation around periastron. While the irradiation on a planet with \(\theta= 0^{\circ}\) and \(e=0\) is not strong enough to lead to ice-free regions, the generally higher annual mean irradiation and the high peak irradiation around periastron on an orbit with \(e=0.5\) suffice to melt the ice layer. This generally leads to a sharper transition between the two states and possibly only one stable state for higher eccentricities (Figure [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}(a, b)). This bifurcation from two stable states to only one stable state was also observed in previous studies where the length of one year was varied. Meridional transport processes due to atmospheric dynamics play a crucial role in determining the behaviour of the ice-albedo feedback and therefore also affect the transitions between ice-free and ice-covered climate states. Hysteresis curves shown in Fig. [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}(c, d) are obtained after "switching off" the atmospheric dynamic core of PlaSim, leaving each atmospheric column isolated from its neighbouring columns and in radiative-convective equilibrium. Because temperature differences are not smoothed by transport processes, large gradients build up between regions with more and less intense insolation. These simulations show that the absence of horizontal transport processes has as a consequence for the disappearance of the catastrophic events in the hysteresis curves. For different obliquities and different eccentricities the planets get gradually frozen without the sudden glaciation observed in full-dynamics simulations (Fig. [\[fig:bifurc\]](#fig:bifurc){reference-type="ref" reference="fig:bifurc"}(c, d)). This is in line with the analysis of results obtained from one-dimensional EBM of Earth, showing that the sensitivity of the ice-line to changes in solar radiation, which goes to infinity during a snowball catastrophe, is identically zero if there is no meridional heat transport. As concluding remark we note that, although catastrophic events represented by sudden jumps of the mean surface temperature disappear when transport processes are excluded, hysteresis still remains with curves showing characteristics different between the two directions of the transition. ## Implications for habitability {#sec:results:hab} The existence of liquid water is the most common criterion for habitability. In estimates of the outer boundary of the habitable zone with zero dimensional EBMs or RCMs, this has been translated to the requirement of a global average surface temperature above the freezing point of water. Another sufficient condition for habitability is the existence of a warm stable climate state, because in the case of bistability the warm state always features some ice-free areas. These two conditions are however conservative and neglect that in case of spatial and temporal heterogeneity ice-free areas can also exist under both conditions, at a global mean surface temperature below the freezing point and in a cold climate state. We therefore use a slightly modified version of the definition of temporal habitability proposed by, based on sea ice cover instead of surface temperature as measure for habitability. This reflects our use of an explicit sea ice model and the fact that stability issues of PlaSim do not allow to simulate climate close to the boiling point of water, which restricts our analysis to the outer boundary of the HZ. We then compute the zonal habitability \(H(\varphi, t)\) from sea ice cover \(\sigma(\varphi, \lambda, t)\) according to \[H(\varphi, t) = 1-\frac{1}{2\pi}\int_{0}^{2\pi} \sigma(\varphi, \lambda, t) \rm{d}\lambda,\] where \(\varphi\) and \(\lambda\) are latitude and longitude respectively. The temporal habitability \(f(\varphi)\) is then defined as \[f(\varphi) \; = \; \frac{\int_{t = 0}^{T}\;H(\varphi, t)\; dt}{T}\] where \(T\) denotes the length of one year. We then classify each climate state as one of the following: - *habitable* if \(f(\varphi) = 1\) at all latitudes \(\varphi\) - *partially habitable* if \(f(\varphi) = 1\) at some latitude - *temporally habitable* if \(f(\varphi) > 0\) at some latitude but \(f(\varphi) < 1\) at all latitudes - *unhabitable* if \(f(\varphi) = 0\) at all latitudes Our experiments reveal several aspects of the effect of obliquity and eccentricity on a planet's habitability. First of all, seasonal variability can lead to a substantial extension of the habitable zone. While a planet with \(e=0\) and \(\theta = 0^{\circ}\) is habitable for \(\tilde{S} > 0.95\) (warm initial state) or \(\tilde{S} > 1.3\) (cold initial state), a planet with \(e=0.5\) and \(\theta = 90^{\circ}\) is habitable for \(\tilde{S} > 0.55\) (E05o90, Figure [\[fig:hab_degree\]](#fig:hab_degree){reference-type="ref" reference="fig:hab_degree"}). Moreover, also cold climate states can exhibit regions that are temporally and even continously ice free. While the extent of the warm state is an appropriate measure of the HZ on planets without seasonal variations (E00o00, Figure [\[fig:hab_degree\]](#fig:hab_degree){reference-type="ref" reference="fig:hab_degree"}), it leads to a substantial underestimation of the HZ on planets with large seasonal variations (E05o90, Figure [\[fig:hab_degree\]](#fig:hab_degree){reference-type="ref" reference="fig:hab_degree"}). On such planets the extension of the HZ is mainly due to regions that are temporally habitable, i.e ice free only at some time of the year. A clear distinction between partial and temporal habitability is therefore not only motivated from a biological perspective but also strongly suggested by the results of our climate simulations. Because eccentricity affects both the annual mean irradiation and its variability, these two effects have to be analysed separately. This is illustrated in the right column of Figure [\[fig:hab_degree\]](#fig:hab_degree){reference-type="ref" reference="fig:hab_degree"}. Here the results of the simulations at \(e=0.5\) are rescaled to the mean irradiation at \(e=0\) by dividing them through the factor at which the annual mean irradiation increases from \(e=0\) to \(e=0.5\) (see Equation 2). The results show that while the increase of annual mean irradiation with eccentricity is a good proxy for the extent of the warm state on an eccentric orbit, the simple scaling does neither capture the effects that lead to the extension of the HZ nor the effects that limit the extent of the cold state on an eccentric orbit. The remaining effect of eccentricity has to be attributed to the higher amplitude of the seasonal cycle and the different spatial and temporal distribution of irradiation (Figure [\[fig:hab_degree\]](#fig:hab_degree){reference-type="ref" reference="fig:hab_degree"}). Whether temporally habitable regions might indeed be suitable for life might depend on the amplitude of these seasonal variations and how long such regions might be ice-covered and ice-free. The presence of land masses (not accounted for in this study) may in turn facilitate or go against this partially habitable, ice-free regions. As illustrative examples we show the seasonal cycle of surface temperature and ice thickness of two simulations on an orbit with \(e=0.5\) and \(\theta = 0^{\circ}\) and \(\theta = 90^{\circ}\). Both are shown for \(\tilde{S} = 0.7\) (Figure [\[fig:contour:00\]](#fig:contour:00){reference-type="ref" reference="fig:contour:00"} and [\[fig:contour:90\]](#fig:contour:90){reference-type="ref" reference="fig:contour:90"}). The planet with \(\theta = 0^{\circ}\) shows largest seasonal variations at low latitudes. Here surface temperatures exceed the freezing point for a short period around periastron, but long enough for a tiny area to become temporally ice free (Figure [\[fig:contour:00\]](#fig:contour:00){reference-type="ref" reference="fig:contour:00"}). As compared to this, the climate of the planet with \(\theta = 90^{\circ}\) is extremely variable, with almost half of the planet's surface experiencing freezing and complete melting in one year (Figure [\[fig:contour:90\]](#fig:contour:90){reference-type="ref" reference="fig:contour:90"}). Also surface temperatures show considerable oscillations with a maximal amplitude of the annual cylce of about 100 K at the equator. Apart from these unhabitable regions, however, high latitudes around the South Pole show rather moderate climatic conditions with a temperature oscillation of only about 20 K (Figure [\[fig:contour:90\]](#fig:contour:90){reference-type="ref" reference="fig:contour:90"}). The experiments shown in Figure [\[fig:contour:00\]](#fig:contour:00){reference-type="ref" reference="fig:contour:00"} and Figure [\[fig:contour:90\]](#fig:contour:90){reference-type="ref" reference="fig:contour:90"} differ only in the polar obliquity \(\theta\). As both planets exhibit ice-free areas at some point of the year, they are equally classified as temporally habitable for \(\tilde{S} = 0.7\). Yet a closer look reveals substantial differences with respect to the size of the temporally habitable area, the time this area is ice-free, and the amplitude of surface temperature at that location. Although the total amount of energy per year that both planets receive is the same, its different temporal and spatial distribution leads to extremely different climatic conditions. ## Sensitivity to asymmetric mean irradiation and heat capacity In all numerical experiments with eccentric orbits we assumed that the total radiation received over one year is symmetrically distributed among the two hemispheres (\(\omega = 0^{\circ}\)). Our results show, however that the expansion of the HZ due to seasonal variability is primarily determined by the peak of the locally incoming radiation and whether this maximal radiation suffices to lead to temporally ice-free regions. We therefore expect that an uneven distribution of mean radiation and a higher peak radiation at one pole, both associated with \(\omega \neq 0^{\circ}\), leads to a further extension of the HZ. In order to quantify this effect, we consider the most extreme case \(\omega = 90^{\circ}\) (solstice aligned with periastron). In a previous study experiments with \(\omega = 30^{\circ}\) showed the maximal extent of the HZ. We also include one simulation with \(\omega = 30^{\circ}\), although we expect that the value of \(\omega\) associated with the maximal expansion of the HZ is sensitive to further parameters such as the planet's heat capacity. The results are shown in Figure [\[fig:hab_degree_mvelp\]](#fig:hab_degree_mvelp){reference-type="ref" reference="fig:hab_degree_mvelp"}. As a consequence of the very warm polar summer for \(\omega = \SI{90}{\degree}\), high latitudes becomes ice-free out to \(\tilde{S} > 0.35\). Moreover, as the incoming radiation is less evenly distributed, the temporally ice-free areas shrink compared to \(\omega = 0^{\circ}\). In contrast to, our experiments show the maximal expansion of the HZ for \(\omega = \SI{90}{\degree}\) (Figure [\[fig:hab_degree_mvelp\]](#fig:hab_degree_mvelp){reference-type="ref" reference="fig:hab_degree_mvelp"}). We conclude that an uneven irradiation with \(\omega = \SI{90}{\degree}\) leads to the most extreme climates in our experiments with small areas becoming temporary ice-free for even very low values of \(\tilde{S}\). Variations of surface temperature due to seasonal forcing are damped by the heat capacity of the planet. The mean value and the amplitude of surface temperature at \(e = 0.5\) and \(\theta = 0^{\circ}\) and \(\theta = 90^{\circ}\) are depicted in Figure [\[fig:hab_sens_oc\]](#fig:hab_sens_oc){reference-type="ref" reference="fig:hab_sens_oc"}. Both experiments show roughly the same pattern. In an ice-free state, temperature oscillations are relatively small (\(\sim\) 10-20 K, Figure [\[fig:hab_sens_oc\]](#fig:hab_sens_oc){reference-type="ref" reference="fig:hab_sens_oc"}). With sea ice at the surface, this amplitude increases up to more than 100 K as the exchange of heat between the atmosphere and ocean is reduced and the ice-albedo feedback further amplifies temperature variations (Figure [\[fig:hab_sens_oc\]](#fig:hab_sens_oc){reference-type="ref" reference="fig:hab_sens_oc"}). Choosing an ocean depth of 10 m in our sensitivity experiment hence affects the temperature oscillations and ice coverage only in warm climate states (Figure [\[fig:hab_sens_oc\]](#fig:hab_sens_oc){reference-type="ref" reference="fig:hab_sens_oc"}). In addition, lower heat capacity leads to the formation of ice at higher irradiation \(\tilde{S}\), as less heat can be stored around periastron and the surface temperature thus drops below the freezing point around apoastron. Hence, although a reduced heat capacity reduces the continuously habitable surface area in favor of temporally habitable regions, it has no effect on the lower limit of \(\tilde{S}\) that still leads to ice-free areas. The sensitivity experiment thus demonstrates that our estimate of the HZ is robust with respect to changing the heat capacity of the surface layer. # Discussion {#sec:discuss} In this work we provide for the first time a parametric investigation of the effect of seasonal variability on the climate of idealized Earth-like planets based on a general circulation model (GCM), having in mind the exploration of habitability conditions. Exploring the effects of the two parameters obliquity and eccentricity we build on previous work with energy balance models (EBMs) that focuses on the same parameters but in which simpler models are used. At the same time our work complements parametric investigations with the same model focusing on different parameters. Although the realism of our experiments regarding the determination of the HZ is constrained by both neglecting the silicate weathering cycle and the lack of a dynamic ocean and a dynamic sea ice model, we aim to provide valuable insights into the effects of seasonal variability for the climate of idealized planets on shorter time scales. The numerical experiments performed in this work are based on the assumption of idealized aquaplanets. In order to test the robustness of our results further aspects need to be taken into account. First, our idealized planets are all characterized by uniform surface characteristics. As discussed in Section [2.3](#sec:model:planet){reference-type="ref" reference="sec:model:planet"}, surface heat capacity, albedo, and moisture supply can alter externally induced temperature oscillations and substantially affect the atmospheric dynamics of a planet. The effects of different land-sea distributions are taken into account in and, but a detailed investigation with general circulation models has not been done yet. Second, our numerical experiments are performed with a slab ocean model and thus do not account for oceanic heat transport. On tidally locked planets, where strong spatial variations can imply spatially confined habitable conditions similar to our results, oceanic heat transport can lead to a substantial extension of the HZ. However, for high obliquity and low eccentricity demonstrated that slab ocean models reproduce well the surface climate of coupled simulations for an ocean depth of 50 m. Third, only thermodynamic sea ice processes are represented in our model. Any dynamical processes such as wind induced sea ice drift or the flow of sea ice due to gravity are not taken into account. In some cases we expect this to limit the realism of our results. In the experiment of a planet on an eccentric orbit with \(\theta = \SI{0}{\degree}\), for instance, the prescribed maximal sea ice thickness might bias our results towards temporally habitable conditions, because sea ice is not allowed to grow beyond a thickness of \(\SI{3}{m}\) anywhere on the planet during apoastron. Without this restriction, sea ice could grow much thicker at the poles and subsequently flow towards lower latitudes, preventing the quick melt of the sea ice layer that appears in our simulation and thus resulting in a permanently unhabitable climate. Taking Earth as reference, a sea ice thickness beyond 3 m might indeed be realistic on exoplanets, as it is estimated that during Snowball events Earth was covered by an ice layer several hundreds of meters thick . A further limitation of our sea ice model might be due to a too high amplitude of the seasonal cycle of surface temperatures, attributable to the low vertical resolution of the ice layer in our model. Because associated high peak values of surface temperature are required for the existence of ice-free areas in some of our simulations, also this limitation of our sea ice model tends to bias our results towards (temporally) habitable conditions. Fourth, the effect of a different rotation rate needs to be investigated because of its large impact on the meridional heat transport. Experiments with energy balance models where the effect of the planetary spin rate is parametrized through the meridional heat transport diffusion coefficient show that on planets with seasonal variability, the effect on habitability is comparatively small. Besides its effect on the meridional transport of energy, the planetary spin also affects climate through diurnal variability, which may have a large effect on habitability for slowly rotating and phase-locked planets. Several authors have addressed the effect of seasonal variability on the habitability of exoplanets using climate models of lower complexity. On circular orbits, our results are in good agreement with previous results based on one-dimensional latitudinal resolved EBMs (, Figure [\[fig:zone\]](#fig:zone){reference-type="ref" reference="fig:zone"}). They also agree well in a qualitative sense with recent work based on an energy balance model, showing a positive relation between obliquity and the maximal habitable zone. On eccentric orbits with \(e=0.5\), however, simulations with EBMs tend to overestimate the outer boundary of the HZ relative to our results (, Figure [\[fig:zone\]](#fig:zone){reference-type="ref" reference="fig:zone"}). Despite this overestimation, the qualitative effect of obliquity agrees well also on eccentric orbits, showing the maximal extent for planets with \(\theta = 90^{\circ}\) (Figure [\[fig:zone\]](#fig:zone){reference-type="ref" reference="fig:zone"}). Conversely to the overestimation of the HZ by EBMs at \(e=0.5\), simulations at \(e=0.2\) show a smaller extent of the HZ simulated by EBMs than observed in our experiments (Figure [\[fig:zone\]](#fig:zone){reference-type="ref" reference="fig:zone"}). This somewhat surprising observation might be attributable to slightly different experimental setups of the EBMs used in the experiments at \(e=0\) and \(e=0.2\), but can also hint at a non-linear effect of eccentricity on the extent of the HZ in the EBMs. Such effect might originate from the complex interplay between the different timescales of variability, which is dramatically simplified by assessing only its consequence for the extent of the HZ. Apart from apparent differences between the two different classes of models, such as the explicit simulation of the atmospheric circulation in our model, further explanations for the discrepancies are different definitions of habitability and the representation of ice processes. While and subsequent studies use \(T_{s} \geq \SI{273}{K}\) as condition for habitability, our definition is based on ice-coverage of the surface. Hence it suffices that the surface temperature shortly exceeds the melting point to be habitable in their model, while in our experiments it is required that the sea ice be locally completely melted. This difference becomes more severe due to the representation of the ice-albedo feedback in the two models. While the sea ice model of PlaSim considers only the areal extent of sea ice, i.e. a complete melting of the ice column is required to affect the surface albedo and therefore amplify the initial signal, the EBM incorporates a smooth dependency of surface albedo on temperature, which can be interpreted as the inclusion of melting ponds on top of the ice sheet and facilitates numerical computations. The most critical limitations of our work are shared with these previous studies. These limitations are due to the neglection of the effect of CO\(_2\) accumulation on the extent of the HZ. The most common estimates of the habitable zone rely on zero dimensional radiation convection models (RCMs). As opposed to our general circulation model, these models include feedback mechanisms of the carbon cycle: on the one hand, a decrease of surface temperature reduces the rate of chemical weathering at the surface, which leads to an accumulation of CO\(_{2}\) in the atmosphere and an enhanced greenhouse effect. On the other hand, a higher concentration of CO\(_{2}\) leads to an increased Rayleigh scattering of incoming radiation, which tends to further decrease the surface temperature. The outer boundary of the HZ, also called the maximum greenhouse limit, is then defined by the balance between these two processes. Because our aim is a parametric investigation of the effects of obliquity and eccentricity on habitability, we have to limit the number of parameters and our experiments can not provide a realistic estimate of habitability of a generic planet. In leaving aside the stabilizing effect of silicate weathering, however, we provide conservative estimates of the HZ. # Conclusions {#sec:concl} Our results show a great diversity of climates due to different obliquities and orbit eccentricities. Moreover, seasonal variability has a large effect on habitability. For planets with Earth-like atmospheres, seasonal variability can extend the maximal distance between planet and host star that still allows for habitable conditions from 1.03 AU (obliquity \(\theta = 0^{\circ}\), eccentricity \(e = 0\)) to 1.69 AU (\(\theta = 90^{\circ}\), \(e = 0.5\), \(\omega = 90^{\circ}\)). While the effect of obliquity on habitability is comparatively small on circular orbits, it becomes highly relevant on eccentric orbits. This effect of seasonal variability on habitability is primarily due to regions that are ice-free only at some time of the year. An appropriate assessment of the HZ therefore asks for a clear distinction between different degrees of habitability (Table [\[tab:hab\]](#tab:hab){reference-type="ref" reference="tab:hab"}). Moreover, our experiments show that the multistability of the climate state is strongly influenced by the obliquity of a planet, in particular on circular orbits. The largest extent of the warm state is found for an obliquity around \(\theta = \SI{55}{\degree}\), where the temperature distribution is most homogeneous. The smallest extent of the warm state and the largest extent of the cold state is found for planets with low values of obliquity. On eccentric orbits, the range of distances that allow for two stable climate states shrinks relative to circular orbits, possibly leading to monostability for planets with very large seasonal variability. While the extent of the warm climate state is a good proxy for the habitability of planets without seasonal variability, this does not hold for planets with high obliquities and on eccentric orbits, since also cold states can allow for ice-free regions. Whether life can exist in such extreme climates with strong temporal and spatial variability in surface conditions is still an open question. Our results also suggest a need for more sophisticated indicators of climate that account for the effects of strong variability. Some of the effects of seasonal variability can be attributed to an increase of the annual mean irradiation with eccentricity. Nonetheless, while this effect of eccentricity can in part explain the extension of the warm state on eccentric orbits, it does neither capture the widening effect on the HZ nor the limiting effect on the extent of the cold state. Estimates of the outer boundary of the HZ that are based on RCMs and account for silicate weathering yield a maximal planet-star distance of 1.67 AU, recently updated to 1.70 AU. These estimates, however, rely on the assumption of a constant and uniform surface temperature, which is-as show in our results-a crude simplification of the problem. Planets with high seasonal variability can indeed feature a mean surface temperature below 240 K although having large areas completely ice-free throughout the year (Figure 8). How weathering might be affected by such a high spatial and temporal variability and what this implies for estimates of the outer boundary of the habitable zone is an interesting open question. Future investigations with GCMs that combine seasonal variability, a dynamic CO\(_{2}\) concentration, and ocean and sea ice dynamics are required in order to provide further insights into the implications of seasonal variability for the climate of exoplanets and their ability to host life.

{'timestamp': '2014-10-13T02:02:50', 'yymm': '1401', 'arxiv_id': '1401.5323', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5323'}

# Overview advances in the field of distributed inference have produced several useful strategies aimed at exploiting local *cooperation* among network nodes to enhance the performance of each individual agent. However, the increasing availability of streaming data continuously flowing across the network has added the new and challenging requirement of online *adaptation* to track drifts in the data. In the adaptive mode of operation, the network agents must be able to enhance their learning abilities continually in order to produce reliable inference in the presence of drifting statistical conditions, drifting environmental conditions, and even changes in the network topology, among other possibilities. Therefore, concurrent adaptation (i.e., tracking) and learning (i.e., inference) are key components for the successful operation of distributed networks tasked to produce reliable inference under dynamically varying conditions and in response to streaming data. Several useful distributed implementations based on consensus strategies  and diffusion strategies  have been developed for this purpose in the literature. The diffusion strategies have been shown to have superior stability ranges and mean-square performance when constant step-sizes are used to enable continuous adaptation and learning . For example, while consensus strategies can lead to unstable growth in the state of adaptive networks even when all agents are individually stable, this behavior does not occur for diffusion strategies. In addition, diffusion schemes are robust, scalable, and fully decentralized. Since in this work we focus on studying *adaptive* distributed inference strategies, we shall therefore focus on diffusion schemes due to their enhanced mean-square stability properties over adaptive networks. Now, the interplay between the two fundamental aspects of cooperation and adaptation has been investigated rather extensively in the context of *estimation* problems. Less explored in the literature is the same interplay in the context of *detection* problems. This is the main theme of the present work. Specifically, we shall address the problem of designing and characterizing the performance of diffusion strategies that reconcile both needs of adaptation and detection in decentralized systems. The following is a brief description of the scenario of interest. A network of connected agents is assumed to monitor a certain phenomenon of interest. As time elapses, the agents collect an increasing amount of streaming data, whose statistical properties depend upon an *unknown* state of nature. The state is formally represented by a pair of hypotheses, say, \({\cal H}_0\) and \({\cal H}_1\). At each time instant, each agent is expected to produce a decision about the state of nature, based upon its own observations and the exchange of information with neighboring agents. The emphasis here is on *adaptation*: we allow the true hypothesis to drift over time, and the network must be able to track the drifting state. This framework is illustrated in Fig. [\[fig:adaptection\]](#fig:adaptection){reference-type="ref" reference="fig:adaptection"}, where we show the time-evolution of the actual realization of the decision statistics computed by three generic network agents. Two situations are considered. In the first case, the agents run a constant-step size diffusion strategy  and in the second case, the agents run a consensus strategy with a decaying step-size of the form \(\mu_n=1/n\) . Note from the curves in the figure that the statistics computed by different sensors are hardly distinguishable, emphasizing a certain equivalence in performance among distinct agents, an important feature that will be extensively commented on in the forthcoming analysis. Assume that high (positive) values of the statistic correspond to deciding for \({\cal H}_1\), while low (negative) values correspond to deciding for \({\cal H}_0\). The bottom panel in the figure shows how the true (unknown) hypothesis changes at certain (unknown) epochs following the sequence \({\cal H}_0\rightarrow{\cal H}_1\rightarrow{\cal H}_0\). It is seen in the figure that the adaptive diffusion strategy is more apt in tracking the drifting state of nature. It is also seen that the decaying step-size consensus implementation is unable to track the changing conditions. Moreover, the inability to track the drift degrades further as time progresses since the step-size sequence \(\mu_n=1/n\) decays to zero as \(n\rightarrow\infty\). For this reason, in this work we shall set the step-sizes to constant values to enable continuous adaptation and learning by the distributed network of detectors. In order to evaluate how well these adaptive networks perform, we need to be able to assess the goodness of the inference performance (reliability of the decisions), so as to exploit the trade-off between adaptation and learning capabilities. This will be the main focus of the paper. ## Related Work The literature on distributed detection is definitely rich, see, e.g.,  as useful entry points on the topic. A distinguishing feature of our approach is its emphasis on *adaptive* distributed detection techniques that respond to streaming data in real-time. We address this challenging problem with reference to the *fully decentralized* setting, where no fusion center is admitted, and the agents cooperate through local interaction and consultation steps. Several recent works on the subject of distributed detection focus on the performance of decentralized consensus algorithms with decaying step-size . As already mentioned, however, these strategies are inherently non-adaptive. In order to overcome this issue, diffusion algorithms with constant step-size are used. With reference to this class of algorithms, while several results have been obtained for the mean-square-error (MSE) *estimation* performance of adaptive networks , less is known about the performance of distributed detection networks (see, e.g., ). This is mainly due to the fact that results on the asymptotic distribution of the error quantities under constant step-size adaptation over networks are largely unavailable in the literature. While  argues that the error in single-agent least-mean-squares (LMS) adaptation converges in distribution, the resulting distribution is not characterized. Only recently these questions have been considered in  in the context of distributed estimation over adaptive networks. Nevertheless, these results on the asymptotic distribution of the errors are still insufficient to characterize the rate of decay of the probability of error over networks of distributed detectors. To do so, it is necessary to pursue a large deviations analysis in the constant step-size regime. Motivated by these remarks, we therefore provide a thorough statistical characterization of the diffusion network in a manner that enables detector design and analysis. ## Main Result and Ramifications Consider a connected network of \(S\) sensors performing distributed detection by means of adaptive diffusion strategies, as described in the next section. The adaptive nature of the solution allows the network to track variations in the hypotheses being tested over time. As anticipated, to enable continuous adaptation and learning, we shall employ distributed strategies with a *constant* step-size parameter \(\mu\). Now, let \(\alpha_{k,\mu}\) and \(\beta_{k,\mu}\) represent the steady-state (as \(n\rightarrow\infty\)) Type-I (false-alarm) and Type-II (miss-detection) error probabilities at the \(k\)-th sensor. One of the main conclusions established in this paper can be summarized by the following scaling laws: \[\boxed{ \alpha_{k,\mu}\stackrel{\cdot}{=}e^{-(1/\mu)\,S\,{\cal E}_0},\qquad \beta_{k,\mu}\stackrel{\cdot}{=}e^{-(1/\mu)\,S\,{\cal E}_1} } \label{eq:mainres2}\] where the notation \(\stackrel{\cdot}{=}\) means equality to the leading exponential order as \(\mu\) goes to zero . In the above expressions, the parameters \({\cal E}_0\) and \({\cal E}_1\) are solely dependent on the moment generating function of the single-sensor data \(\bm{x}\), and of the decision regions. These parameters are *independent* of the step-size \(\mu\), the number of sensors \(S\), and the network connectivity. Result ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}) has at least four important and insightful ramifications about the performance of adaptive schemes for distributed detection over networks. To begin with, Eq. ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}) reveals a fundamental scaling law for distributed detection with diffusion adaptation, namely, it asserts that as the step-size decreases, the error probabilities are driven to zero exponentially as functions of \(1/\mu\), and that the error exponents governing such a decay increase linearly in the number of sensors. These implications are even more revealing if examined in conjunction with the known results concerning the scaling law of the Mean-Square-Error (MSE) for adaptive distributed estimation over diffusion networks . Assuming a connected network with \(S\) sensors, and using sufficiently small step-sizes \(\mu\approx 0\), the MSE that is attained by sensor \(k\) obeys (see expression (32) in ): \[\textnormal{MSE}_k\; \propto \;\frac{\mu}{S}, \label{eq:MSEscales}\] where the symbol \(\propto\) denotes proportionality. Some interesting symmetries are observed. In the estimation context, the MSE decreases as \(\mu\) goes to zero, and the scaling rate improves linearly in the number of sensors. Recalling that smaller values of \(\mu\) mean a lower degree of adaptation, we observe that reaching a better inference quality costs in terms of adaptation speed. This is a well-known tradeoff in the adaptive estimation literature between tracking speed and estimation accuracy. Second, we observe from ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}) and ([\[eq:MSEscales\]](#eq:MSEscales){reference-type="ref" reference="eq:MSEscales"}) that the scaling laws governing errors of detection and estimation over distributed networks behave very differently, the former exhibiting an exponential decay proportional to \(1/\mu\), while the latter is linear with decay proportional to \(\mu\). The significance and elegance of this result for adaptive distributed networks lie in revealing an intriguing analogy with other more traditional inferential schemes. As a first example, consider the standard case of a centralized, non-adaptive inferential system with \(N\) i.i.d. data points. It is known that the error probabilities of the best detector decay exponentially fast to zero with \(N\), while the optimal estimation error decays as \(1/N\) . Another important case is that of rate-constrained multi-terminal inference . In this case the detection performance scales exponentially with the bit-rate \(R\) while, again, the squared estimation error vanishes as \(1/R\). Thus, at an abstract level, reducing the step-size corresponds to increasing the number of independent observations in the first system, or increasing the bit-rate in the second system. The above comparisons furnish an interesting interpretation for the step-size \(\mu\) as the basic parameter quantifying the cost of information used by the network for inference purposes, much as the number of data \(N\) or the bit-rate \(R\) in the considered examples. A third aspect pertaining to the performance of the distributed network relates to the potential benefits of cooperation. These are already encoded into ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}), and we have already implicitly commented on them. Indeed, note that the error exponents increase linearly in the number of sensors. This implies that cooperation offers *exponential gains* in terms of detection performance. The fourth and final ramification we would like to highlight relates to how much performance is lost by the *distributed* solution in comparison to a centralized stochastic gradient solution. Again, the answer is contained in ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}). Specifically, the centralized solution is equivalent to a fully connected network, so that ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}) applies to the centralized case as well. As already mentioned, the parameters \({\cal E}_0\) and \({\cal E}_1\) do not depend on the network connectivity, which therefore implies that, as the step-size \(\mu\) decreases, the distributed diffusion solution of the inference problem exhibits a detection performance governed by the *same error exponents* of the centralized system. This is a remarkable conclusion and it is also consistent with results in the context of adaptive distributed estimation over diffusion networks . We now move on to describe the adaptive distributed solution and to establish result ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}) and the aforementioned properties. We use boldface letters to denote random variables, and normal font letters for their realizations. Capital letters refer to matrices, small letters to both vectors and scalars. Sometimes we violate this latter convention, for instance, we denote the total number of sensors by \(S\). The symbols \(\P\) and \({\mathbb{E}}\) are used to denote the probability and expectation operators, respectively. The notation \(\P_h\) and \({\mathbb{E}}_h\), with \(h=0,1\), means that the pertinent statistical distribution corresponds to hypothesis \({\cal H}_0\) or \({\cal H}_1\). # Problem Formulation {#sec:probform} The scalar observation collected by the \(k\)-th sensor at time \(n\) will be denoted by \(\bm{x}_k(n)\), \(k=1,2,\dots,S\), which arises from a stationary distribution with mean \({\mathbb{E}} \bm{x}\) and variance \(\sigma_x^2\). Data are assumed to be spatially and temporally independent and identically distributed (i.i.d.), *conditioned* on the hypothesis that gives rise to them. The distributed network is interested in making an inference about the true state of nature. As it is well-known, for the i.i.d. data model, an optimal centralized (and non-adaptive) detection statistic is the sum of the log-likelihoods. When these are not available, alternative detection statistics obtained as the sum of some suitably chosen functions of the observations are often employed, as happens in some specific frameworks, e.g., in locally optimum detection  and in universal hypothesis testing . Accordingly, each sensor in the network will try to compute, as its own detection statistic, a weighted combination of some function of the local observations. We assume the symbol \(\bm{x}_k(n)\) represents the local statistic that is available at time \(n\) at sensor \(k\). Since we are interested in an adaptive inferential scheme, and given the idea of relying on weighted averages, we resort to the class of diffusion strategies for adaptation over networks . These strategies admit various forms. We consider the ATC form due to some inherent advantages in terms of a slightly improved mean-square-error performance relative to other forms . In the ATC diffusion implementation, each node \(k\) updates its state from \(\bm{y}_{k}(n-1)\) to \(\bm{y}_{k}(n)\) through local cooperation with its neighbors as follows: \[\begin{aligned} \bm{v}_k(n)&=&\bm{y}_k(n-1) + \mu[\bm{x}_k(n)-\bm{y}_k(n-1)],\label{eq:diff1}\\ \bm{y}_k(n)&=&\sum_{\ell=1}^S a_{k,\ell} \bm{v}_{\ell}(n), \label{eq:diff2} \end{aligned}\] where \(0<\mu\ll 1\) is a small step-size parameter. In this construction, node \(k\) first uses its local statistic, \(\bm{x}_k(n)\), to update its state from \(\bm{y}_{k}(n-1)\) to an intermediate value \(\bm{v}_k(n)\). All other nodes in the network perform similar updates simultaneously using their local statistics. Subsequently, node \(k\) aggregates the intermediate states of its neighbors using nonnegative convex combination weights \(\{a_{k,\ell}\}\) that add up to one. Again, all other nodes in the network perform a similar calculation. If we collect the combination coefficients into a matrix \(A=[a_{k,\ell}]\), then \(A\) is a right-stochastic matrix in that the entries on each of its rows add up to one: \[a_{k,\ell}\geq 0,\quad A\mathds{1}=\mathds{1},\] with \(\mathds{1}\) being a column-vector with all entries equal to \(1\). At time \(n\), the \(k\)-th sensor needs to produce a decision based upon its state value \(\bm{y}_k(n)\). To this aim, a decision rule must be designed, by choosing appropriate decision regions. The performance of the test will be measured according to the Type-I and Type-II error probabilities defined, respectively, as \[\begin{aligned} \alpha&{:=}&\P[\textnormal{choose }{\cal H}_1\textnormal{ when }{\cal H}_0\textnormal{ is true}],\\ \beta&{:=}&\P[\textnormal{choose }{\cal H}_0\textnormal{ when }{\cal H}_1\textnormal{ is true}]. \end{aligned}\] An important remark is needed at this stage. Computation of the exact distribution of \(\bm{y}_k(n)\) is generally intractable. This implies that the structure of the optimal, or even of a reasonable test, is unknown. We tackle this difficult problem by pursuing the following approach. First, we perform a thorough analysis of the statistical properties of \(\bm{y}_k(n)\). As standard in the adaptation literature , this will be done with reference to \(i)\) the steady-state properties (as \(n\rightarrow\infty\)), and \(ii)\) for small values of the step-size (\(\mu\rightarrow 0\)). Throughout the paper, the term steady-state will refer to the limit as the time-index \(n\) goes to infinity, while the term asymptotic will be used to refer to the slow adaptation regime where \(\mu\rightarrow 0\). Specifically, we will follow these steps: - We show that \(\bm{y}_k(n)\) has a *limiting distribution* as \(n\) goes to infinity (Theorem 1). - For small step-sizes, the distribution of \(\bm{y}_k(n)\) approaches a Gaussian, i.e., it is *asymptotically normal* (Theorem 2). - We characterize the *large deviations* of the steady-state output \(\bm{y}_k(n)\) in the slow adaptation regime when \(\mu\rightarrow 0\) (Theorem 3). - The results of the above steps will provide a series of tools for designing the detector and characterizing its performance (Theorem 4). We would like to mention that the detailed statistical characterization offered by Theorems 1-3 is not confined to the specific detection problems we are dealing with. As a matter of fact, these results are of independent interest, and might be useful for the application of adaptive diffusion strategies in broader contexts. # Existence of Steady-State Distribution {#sec:steady} Let \(\bm{y}_n\) denote the \(S\times 1\) vector that collects the state variables from across the network at time \(n\), i.e., \[\bm{y}_n=\mbox{\rm col}\{\bm{y}_1(n),\,\bm{y}_2(n),\,\ldots,\bm{y}_S(n)\}.\] Likewise, we collect the local statistics \(\{\bm{x}_k(n)\}\) at time \(n\) into the vector \(\bm{x}_n\). It is then straightforward to verify from the diffusion strategy ([\[eq:diff1\]](#eq:diff1){reference-type="ref" reference="eq:diff1"})--([\[eq:diff2\]](#eq:diff2){reference-type="ref" reference="eq:diff2"}) that the vector \(\bm{y}_n\) is given by: \[\boxed { \bm{y}_n=(1-\mu)^n\,A^n \bm{y}_0 + \frac{\mu}{1-\mu} \sum_{i=1}^{n} (1-\mu)^{n-i+1} A^{n-i+1} \bm{x}_i }\] By making the change of variables \(i\leftarrow n-i+1\), the above equation can be written as \[\bm{y}_n=(1-\mu)^n\,A^n \bm{y}_0 + \frac{\mu}{1-\mu} \sum_{i=1}^{n} (1-\mu)^i A^i \bm{x}_{n-i+1}. \label{eq:changeofvariable}\] We note that the term \(\bm{x}_{n-i+1}\) depends on the time index \(n\) in such a way that the most recent datum \(\bm{x}_n\) is assigned the highest scaling weight, in compliance with the adaptive nature of the algorithm. However, since the vectors \(\bm{x}_i\) are i.i.d. across time, and since we shall be only concerned with the distribution of partial sums involving these terms, the statistical properties of \(\bm{y}_n\) are left unchanged if in ([\[eq:changeofvariable\]](#eq:changeofvariable){reference-type="ref" reference="eq:changeofvariable"}) we replace \(\bm{x}_{n-i+1}\) with a random vector \(\bm{x}^\prime_i\), where \(\{\bm{x}_i^\prime\}\) is a sequence of i.i.d. random vectors distributed similarly to the \(\{\bm{x}_{n-i+1}\}\). Formally, we write \[\bm{y}_n\stackrel{d}{=}(1-\mu)^n\,A^n \bm{y}_0 + \frac{\mu}{1-\mu} \sum_{i=1}^{n} (1-\mu)^i A^i \bm{x}^\prime_i,\] where \(\stackrel{d}{=}\) denotes equality in distribution. It follows that the state of the \(k-\)th sensor is given by: \[\begin{aligned} \bm{y}_k(n)&\stackrel{d}{=}& \underbrace{(1-\mu)^n\,\sum_{\ell=1}^S b_{k,\ell} (n) \bm{y}_\ell(0)}_{\textnormal{transient}} \nonumber\\ &+& \underbrace{ \frac{\mu}{1-\mu} \sum_{i=1}^{n} (1-\mu)^i \sum_{\ell=1}^S b_{k,\ell}(i) \bm{x}^\prime_\ell(i), }_{\textnormal{steady-state}} \label{eq:mainATC} \end{aligned}\] where the scalars \(b_{k,\ell}(n)\) are the entries of the matrix power: \[B_n{:=} A^n.\] Since we are interested in reaching a *balanced* fusion of the observations, we shall assume that \(A\) is *doubly-stochastic* with second largest eigenvalue magnitude strictly less than one, which yields : \[B_n\stackrel{n\rightarrow\infty}{\longrightarrow} \displaystyle{\frac 1 S}\,\mathds{1}\mathds{1}^T. \label{eq:doublestoc}\] In order to reveal the steady-state behavior of \(\bm{y}_k(n)\), let us focus on the RHS of ([\[eq:mainATC\]](#eq:mainATC){reference-type="ref" reference="eq:mainATC"}). It is useful to rewrite it as: \[\bm{y}_k(n)\stackrel{d}{=}\textnormal{ transient }+\sum_{i=1}^{n} \bm{z}_k(i), \label{eq:sumexpress}\] where the definition of \(\bm{z}_k(i)\) should be clear. As a result, we are faced with a sum of independent, but *not* identically distributed, random variables. Let us evaluate the first two moments of the sum: \[{\mathbb{E}} \left(\sum_{i=1}^n \bm{z}_k(i)\right) = {\mathbb{E}} \bm{x} \, \sum_{i=1}^n \mu(1-\mu)^{i-1} \underbrace{\sum_{\ell=1}^S b_{k,\ell}(i)}_{=1} \stackrel{n\rightarrow\infty}{\longrightarrow} {\mathbb{E}} \bm{x}, \label{eq:oneseries}\] and (\(\textnormal{VAR}\) denotes the variance operator): \[\begin{aligned} \textnormal{VAR} \left( \sum_{i=1}^n \bm{z}_k(i) \right)&=& \sigma^2_x \, \sum_{i=1}^n \mu^2(1-\mu)^{2(i-1)} \sum_{\ell=1}^S \underbrace{b^2_{k,\ell}(i)}_{ \leq 1} \nonumber\\ &\leq& \frac{\sigma^2_x \,S\,\mu}{2-\mu}<\infty. \label{eq:twoseries} \end{aligned}\] We have thus shown that the expectation of the sum expression from ([\[eq:sumexpress\]](#eq:sumexpress){reference-type="ref" reference="eq:sumexpress"}) converges to \({\mathbb{E}} \bm{x}\), and that its variance converges to a finite value. In view of the Infinite Convolution Theorem---see , these two conditions are sufficient to conclude that the second term on the RHS of ([\[eq:sumexpress\]](#eq:sumexpress){reference-type="ref" reference="eq:sumexpress"}), i.e., the sum of random variables \(\bm{z}_k(i)\), converges in distribution as \(n\rightarrow\infty\), and the first two moments of the limiting distribution are equal to \({\mathbb{E}}\bm{x}\) and \(\sum_{i=1}^\infty \textnormal{VAR}(\bm{z}_k(i))\). The random variable characterized by the limiting distribution will be denoted by \(\bm{y}_{k,\mu}^{\star}\), where we make explicit the dependence upon the step-size \(\mu\) for later use. Since the first term on the RHS of ([\[eq:mainATC\]](#eq:mainATC){reference-type="ref" reference="eq:mainATC"}) vanishes with \(n\), by application of Slutsky's Theorem  we have that, at steady-state, the diffusion output \(\bm{y}_k(n)\) is *still* distributed as \(\bm{y}_{k,\mu}^{\star}\). The above statement can be sharpened to ascertain that the sum of random variables \(\bm{z}_k(i)\) actually converges almost surely (a.s.). This conclusion can be obtained by applying Kolmogorov's Two Series Theorem . In view of the a.s. convergence, it makes sense to define the limiting random variable \(\bm{y}_{k,\mu}^\star\) as: \[\boxed{ \bm{y}^\star_{k,\mu}{:=} \sum_{i=1}^{\infty} \sum_{\ell=1}^S \mu\,(1-\mu)^{i-1} b_{k,\ell}(i) \bm{x}^{\prime}_\ell(i) } \label{eq:zstardef}\] We wish to avoid confusion here. We are not stating that the actual diffusion output \(\bm{y}_k(n)\) converges almost surely (a behavior that would go against the adaptive nature of the diffusion algorithm). We are instead claiming that \(\bm{y}_k(n)\) converges in distribution to a random variable \(\bm{y}_{k,\mu}^\star\) that can be conveniently defined in terms of the a.s. limit ([\[eq:zstardef\]](#eq:zstardef){reference-type="ref" reference="eq:zstardef"}). The main result about the steady-state behavior of the diffusion output is summarized below (the symbol \(\rightsquigarrow\) means convergence in distribution). [Theorem 1:]{.smallcaps} *(Steady-state distribution of \(\bm{y}_k(n)\)). The state variable \(\bm{y}_k(n)\) that is generated by the diffusion strategy ([\[eq:diff1\]](#eq:diff1){reference-type="ref" reference="eq:diff1"})--([\[eq:diff2\]](#eq:diff2){reference-type="ref" reference="eq:diff2"}) is asymptotically stable in distribution, namely, \[\boxed{ \bm{y}_k(n)\stackrel{n\rightarrow\infty}{\rightsquigarrow} \bm{y}^\star_{k,\mu} } \label{eq:Theo1}\]*  \(\square\) It is useful to make explicit the meaning of Theorem 1. By definition of convergence in distribution (or weak convergence), the result ([\[eq:Theo1\]](#eq:Theo1){reference-type="ref" reference="eq:Theo1"}) can be formally stated as : \[\lim_{n\rightarrow\infty} \P[\bm{y}_k(n)\in\Gamma]=\P[\bm{y}^\star_{k,\mu} \in \Gamma], \label{eq:weakconv}\] for any set \(\Gamma\) such that \(\P[\bm{y}^\star_{k,\mu} \in \partial\Gamma]=0\), where \(\partial\Gamma\) denotes the boundary of \(\Gamma\). It is thus seen that the properties of the steady-state variable \(\bm{y}^\star_{k,\mu}\) will play a key role in determining the steady-state performance of the diffusion output. Accordingly, we state two useful properties of \(\bm{y}^\star_{k,\mu}\). First, when the local statistic \(\bm{x}_k(n)\) has an *absolutely continuous* distribution (where the reference measure is the Lebesgue measure over the real line), it is easily verified that the distribution of \(\bm{y}^\star_{k,\mu}\) is *absolutely continuous as well*. Indeed, note that we can write \(\bm{y}^\star_{k,\mu}=\bm{z}_k(1)+\sum_{i=2}^\infty \bm{z}_k(i)\). Now observe that \(\bm{z}_k(1)\), which has an absolutely continuous distribution by assumption, is independent of the other term. The result follows by the properties of convolution and from the fact that the distribution of the sum of two independent variables is the convolution of their respective distributions. Second, when the local statistic \(\bm{x}_k(n)\) is a *discrete* random variable, by the Jessen-Wintner law , we can only conclude that \(\bm{y}^\star_{k,\mu}\) is of *pure type*, namely, its distribution is pure: absolutely continuous, or discrete, or continuous but singular. An intriguing case is that of the so-called *Bernoulli convolutions*, i.e., random variables of the form \(\sum_{i=1}^\infty (1-\mu)^{i-1} \bm{x}(i)\), where \(\bm{x}(i)\) are equiprobable \(\pm 1\). For this case, it is known that if \(1/2<\mu<1\), then the limiting distribution is a *Cantor* distribution . This is an example of a distribution that is neither discrete nor absolutely continuous. When \(\mu<1/2\), which is relevant for our discussion since we shall be concerned with small step-sizes, the situation is markedly different, and the distribution is absolutely continuous for almost all values of \(\mu\). Before proceeding, we stress that we have proved that a steady-state distribution for \(\bm{y}_k(n)\) exists, but its form is not known. Accordingly, even in steady-state, the structure of the optimal test is still unknown. In tackling this issue, and recalling that the regime of interest is that of slow adaptation, we now focus on the case \(\mu\ll 1\). # The Small-\(\mu\) Regime. {#sec:smallmu} While the exact form of the steady-state distribution is generally impossible to evaluate, it is nevertheless possible to approximate it well for small values of the step-size parameter. Indeed, in this section we prove two results concerning the statistical characterization of the steady-state distribution for \(\mu\rightarrow 0\). The first one is a result of *asymptotic normality*, stating that \(\bm{y}^\star_{k,\mu}\) approaches a Gaussian random variable with known moments as \(\mu\) goes to zero (Theorem 2). The second finding (Theorem 3) provides the complete characterization for the *large deviations* of \(\bm{y}^\star_{k,\mu}\). [Theorem 2:]{.smallcaps} *(Asymptotic normality of \(\bm{y}^\star_{k,\mu}\) as \(\mu\rightarrow 0\)). Under the assumption \({\mathbb{E}}|\bm{x}_k(n)|^3<\infty\), the variable \(\bm{y}^\star_{k,\mu}\) fulfills, for all \(k=1,2,\dots,S\):* \[\boxed { \frac{\bm{y}^\star_{k,\mu}-{\mathbb{E}} \bm{x}}{\sqrt{\mu}}\stackrel{\mu\rightarrow 0}{\rightsquigarrow} {\cal N}\left(0,\frac{\sigma^2_x}{2\,S}\right) } \label{eq:CLT}\] The argument requires dealing with independent but non-identically distributed random variables, as done in the Lindeberg-Feller CLT (Central Limit Theorem) . This theorem, however, does not apply to our setting since the asymptotic parameter is *not* the number of samples, but rather the step-size. Some additional effort is needed, and the detailed technical derivation is deferred to Appendix A.  \(\square\) ## Implications of Asymptotic Normality Let us now briefly comment on several useful implications that follow from the above theorem: 1. First, note that *all sensors* share, for \(\mu\) small enough, the *same* distribution, namely, the inferential diffusion strategy equalizes the statistical behavior of the agents. This finding complements well results from  where the asymptotic equivalence among the sensors has been proven in the context of mean-square-error estimation. One of the main differences between the estimation context and the detection context studied in this article is that in the latter case, the regression data is deterministic and the randomness arises from the stochastic nature of the statistics \(\{\bm{x}_k(n)\}\). For this reason, the steady-state distribution in ([\[eq:CLT\]](#eq:CLT){reference-type="ref" reference="eq:CLT"}) is characterized in terms of the moments of these statistics and not in terms of the moments of regression data, as is the case in the estimation context. 2. The result of Theorem 2 is valid provided that the connectivity matrix fulfills ([\[eq:doublestoc\]](#eq:doublestoc){reference-type="ref" reference="eq:doublestoc"}). This condition is satisfied when the network topology is strongly-connected, i.e., there exists a path connecting any two arbitrary nodes and at least one node has \(a_{k,k}>0\) . Obviously, condition ([\[eq:doublestoc\]](#eq:doublestoc){reference-type="ref" reference="eq:doublestoc"}) is also satisfied in the fully connected case when \(a_{k,\ell}=b_{k,\ell}=1/S\) for all \(k,\ell=1,2,\dots,S\). This latter situation would correspond to a representation of the centralized stochastic gradient algorithm, namely, an implementation of the form \[\bm{y}^{(c)}(n)=\bm{y}^{(c)}(n-1)+\frac{\mu}{S}\sum_{\ell=1}^S [\bm{x}_{\ell}(n)-\bm{y}^{(c)}(n-1)], \label{eq:centstochalg}\] where \(\bm{y}^{(c)}(n)\) denotes the output by the centralized solution at time \(n\). The above algorithm can be deduced from ([\[eq:diff1\]](#eq:diff1){reference-type="ref" reference="eq:diff1"})--([\[eq:diff2\]](#eq:diff2){reference-type="ref" reference="eq:diff2"}) by defining \[\bm{y}^{(c)}(n){:=} \frac{1}{S}\sum_{\ell=1}^S \bm{y}_{\ell}(n).\] Now, since the moments of the limiting Gaussian distribution in ([\[eq:CLT\]](#eq:CLT){reference-type="ref" reference="eq:CLT"}) are independent of the particular connectivity matrix, the net effect is that each agent of the *distributed* network acts, asymptotically, as the *centralized* system. This result again complements well results in the estimation context where the role of the statistics variables \(\{\bm{x}_{k}(n)\}\) is replaced by that of stochastic regression data . 3. The asymptotic normality result is powerful in approximating the steady-state distribution for relatively small step-sizes, thus enabling the analysis and design of inferential diffusion networks in many different contexts. With specific reference to the detection application that is the main focus here, Eq. ([\[eq:CLT\]](#eq:CLT){reference-type="ref" reference="eq:CLT"}) can be exploited for an accurate threshold setting when one desires to keep under control one of the two errors, say, the false-alarm probability, as happens, e.g., in the Neyman-Pearson setting . To show a concrete example on how this can be done, let us assume that, without loss of generality, \({\mathbb{E}}_0\bm{x}<{\mathbb{E}}_1\bm{x}\), and consider a single-threshold detector for which: \[\Gamma_0=\{\gamma \in\mathbb{R}:\, \gamma\leq \eta_\mu \},\qquad \Gamma_1=\mathbb{R}\setminus\Gamma_0, \label{eq:gamma0}\] where the threshold is set as \[\eta_\mu={\mathbb{E}}_0\bm{x} + \sqrt{\frac{\mu \sigma_{x,0}^2}{2 S}}\,Q^{-1}(\bar\alpha). \label{eq:NPth1}\] Here, \(\sigma^2_{x,0}\) is the variance of \(\bm{x}\) under \({\cal H}_0\), \(Q(\cdot)\) denotes the complementary CDF for a standard normal distribution, and \(\bar\alpha\) is the prescribed false-alarm level. By ([\[eq:CLT\]](#eq:CLT){reference-type="ref" reference="eq:CLT"}), it is straightforward to check that this threshold choice ensures \[\lim_{\mu\rightarrow 0} \P_0[\bm{y}_{k,\mu}^\star>\eta_\mu]=\bar\alpha. \label{eq:NPpfa1}\] In summary, Theorem 2 provides an approximation of the diffusion output distribution for small step-sizes. At first glance, this may seem enough to obtain a complete characterization of the detection problem. A closer inspection reveals that this is not the case. A good example to understand why Theorem 2 alone is insufficient for characterizing the detection performance is obtained by examining the Neyman-Pearson threshold setting just described in ([\[eq:NPth1\]](#eq:NPth1){reference-type="ref" reference="eq:NPth1"})--([\[eq:NPpfa1\]](#eq:NPpfa1){reference-type="ref" reference="eq:NPpfa1"}) above. While we have seen that the asymptotic behavior of the false-alarm probability in ([\[eq:NPpfa1\]](#eq:NPpfa1){reference-type="ref" reference="eq:NPpfa1"}) is completely determined by the application of Theorem 2, the situation is markedly different as regards the miss-detection probability \(\P_1[\bm{y}_{k,\mu}^\star\leq\eta_\mu]\). Indeed, by using ([\[eq:NPth1\]](#eq:NPth1){reference-type="ref" reference="eq:NPth1"}) we can write: \[\begin{aligned} \lefteqn{ \P_1[\bm{y}_{k,\mu}^\star\leq\eta_\mu] = \P_1\left[ \frac{\bm{y}_{k,\mu}^\star-{\mathbb{E}}_1\bm{x}}{\sqrt{\mu}}\leq\frac{\eta_\mu-{\mathbb{E}}_1\bm{x}}{\sqrt{\mu}} \right] } \nonumber\\ &=&\P_1\left[ \frac{\bm{y}_{k,\mu}^\star-{\mathbb{E}}_1\bm{x}}{\sqrt{\mu}}\leq \frac{{\mathbb{E}}_0\bm{x}-{\mathbb{E}}_1\bm{x}}{\sqrt{\mu}}+\sqrt{\frac{\sigma_{x,0}^2}{2 S}}\,Q^{-1}(\bar\alpha) \right].\nonumber\\ \end{aligned}\] Since \({\mathbb{E}}_0\bm{x} < {\mathbb{E}}_1 \bm{x}\), the quantity \(\frac{{\mathbb{E}}_0\bm{x}-{\mathbb{E}}_1\bm{x}}{\sqrt{\mu}}\) diverges to \(-\infty\) as \(\mu\rightarrow 0\). As a consequence, the fact that \(\frac{\bm{y}_{k,\mu}^\star-{\mathbb{E}}_1\bm{x}}{\sqrt{\mu}}\) is asymptotically normal does not provide much more insight than revealing that the miss-detection probability converges to zero as \(\mu\rightarrow 0\). A meaningful asymptotic analysis would instead require to examine the way this convergence takes place (i.e., the error exponent). The same kind of problem is found when one lets *both* error probabilities vanish exponentially, such that the Type-I and Type-II detection error exponents furnish a meaningful asymptotic characterization of the detector. In order to fill these gaps, the study of the *large* deviations of \(\bm{y}_{k,\mu}^\star\) is needed. ## Large Deviations of \(\bm{y}^\star_{k,\mu}\). {#subsec:LDP} From ([\[eq:CLT\]](#eq:CLT){reference-type="ref" reference="eq:CLT"}) we learn that, as \(\mu\rightarrow 0\), the diffusion output shrinks down to its limiting expectation \({\mathbb{E}}\bm{x}\) and that the *small* (of order \(\sqrt{\mu}\)) deviations around this value have a Gaussian shape. But this conclusion is not helpful when working with *large* deviations, namely, with terms like: \[\P[|\bm{y}^\star_{k,\mu}-{\mathbb{E}}\bm{x}|>\delta] \stackrel{\mu\rightarrow 0}{\longrightarrow} 0,\quad \delta>0, \label{eq:Large}\] which play a significant role in detection applications. While the above convergence to zero can be inferred from ([\[eq:CLT\]](#eq:CLT){reference-type="ref" reference="eq:CLT"}), it is well known that ([\[eq:CLT\]](#eq:CLT){reference-type="ref" reference="eq:CLT"}) is not sufficient in general to obtain the rate at which the above probability vanishes. In order to perform accurate design and characterization of reliable inference systems  it is critical to assess this rate of convergence, which turns out to be the main purpose of a large-deviation analysis. Accordingly, we will be showing in the sequel that the process \(\bm{y}^\star_{k,\mu}\) obeys a Large Deviation Principle (LDP), namely, that the following limit exists : \[\lim_{\mu\rightarrow 0} \mu\, \ln \P[\bm{y}^\star_{k,\mu}\in \Gamma]=-\inf_{\gamma\in \Gamma} I(\gamma){:=}-I_\Gamma, \label{eq:LDPdef}\] for some \(I(\gamma)\) that is called the *rate function*. Equivalently: \[\P[\bm{y}^\star_{k,\mu}\in\Gamma]=e^{-(1/\mu)\,I_\Gamma + o(1/\mu)} \stackrel{\cdot}{=}e^{-(1/\mu)\,I_\Gamma}, \label{eq:expoexpo}\] where \(o(1/\mu)\) stands for any correction term growing slower than \(1/\mu\), namely, such that \(\mu\,o(1/\mu)\rightarrow 0\) as \(\mu\rightarrow 0\), and the notation \(\stackrel{\cdot}{=}\) was introduced in ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}). From ([\[eq:expoexpo\]](#eq:expoexpo){reference-type="ref" reference="eq:expoexpo"}) we see that, in the large deviations framework, only the dominant exponential term is retained, while discarding any sub-exponential terms. It is also interesting to note that, according to ([\[eq:expoexpo\]](#eq:expoexpo){reference-type="ref" reference="eq:expoexpo"}), the probability that \(\bm{y}^\star_{k,\mu}\) belongs to a given region \(\Gamma\) is dominated by the infimum \(I_\Gamma\) of the rate function \(I(\gamma)\) within the region \(\Gamma\). In other words, the smallest exponent (\(\Rightarrow\) highest probability) dominates, which is well explained in  through the statement: "*any large deviation is done in the least unlikely of all the unlikely ways*". In summary, the LDP generally implies an exponential scaling law for probabilities, with an exponent governed by the rate function. Therefore, knowledge of the rate function is enough to characterize the exponent in ([\[eq:expoexpo\]](#eq:expoexpo){reference-type="ref" reference="eq:expoexpo"}). We shall determine the expression for \(I(\gamma)\) pertinent to our problem in Theorem 3 further ahead---see Eq. ([\[eq:ratefunSomega\]](#eq:ratefunSomega){reference-type="ref" reference="eq:ratefunSomega"}). In the traditional case where the statistic under consideration is the arithmetic average of i.i.d. data, the asymptotic parameter is the number of samples and the usual tool for determining the rate function in the LDP is Cramér's Theorem . Unfortunately, in our adaptive and distributed setting, we are dealing with a more general statistic \(\bm{y}^\star_{k,\mu}\), whose dependence is on the step-size parameter and not on the number of samples. Cramér's Theorem is not applicable in this case, and we must resort to a more powerful tool, known as the Gärtner-Ellis Theorem , stated below in a form that uses directly the set of assumptions relevant for our purposes. [Gärtner-Ellis Theorem ]{.smallcaps}. *Let \(\bm{z}_\mu\) be a family of random variables with Logarithmic Moment Generating Function (LMGF) \(\phi_{\mu}(t)=\ln {\mathbb{E}} \exp\{t \bm{z}_\mu\}.\) If \[\phi(t){:=}\lim_{\mu\rightarrow 0} \mu\, \phi_{\mu}(t/\mu)\] exists, with \(\phi(t)<\infty\) for all \(t\in\mathbb{R}\), and \(\phi(t)\) is differentiable in \(\mathbb{R}\), then \(\bm{z}_\mu\) satisfies the LDP property ([\[eq:LDPdef\]](#eq:LDPdef){reference-type="ref" reference="eq:LDPdef"}) with rate function given by the Fenchel-Legendre transform of \(\phi(t)\), namely: \[\Phi(\gamma){:=}\sup_{t\in\mathbb{R}}[\gamma t-\phi(t)]. \label{eq:FLtransf}\]*  \(\square\) In what follows, we shall use capital letters to denote Fenchel-Legendre transforms, as done in ([\[eq:FLtransf\]](#eq:FLtransf){reference-type="ref" reference="eq:FLtransf"}). We now show how the result allows us to assess the asymptotic performance of the diffusion output in the inferential network. Let us introduce the LMGF of the data \(\bm{x}_k(n)\), and that of the steady-state variable \(\bm{y}^\star_{k,\mu}\), respectively: \[\begin{aligned} \psi(t)&{:=}& \ln {\mathbb{E}} \exp\{t \bm{x}_k(n)\},\\ \phi_{k,\mu}(t)&{:=}& \ln {\mathbb{E}} \exp\{t \bm{y}_{k,\mu}^\star\}. \end{aligned}\] [Theorem 3:]{.smallcaps} *(Large deviations of \(\bm{y}^\star_{k,\mu}\) as \(\mu\rightarrow 0\)). Assume that \(\psi(t)<+\infty\) for all \(t\in \mathbb{R}\). Then, for all \(k=1,2,\dots,S\):* - \[\boxed { \phi(t){:=}\lim_{\mu\rightarrow 0} \mu\, \phi_{k,\mu}(t/\mu) = S\, \omega(t/S) } \label{eq:LMGFlimit}\] where \[\boxed { \omega(t){:=}\int_{0}^{t}\frac{\psi(\tau)}{\tau}d\tau \label{eq:omegadef} }\] - The steady-state variable \(\bm{y}^\star_{k,\mu}\) obeys the LDP with a rate function given by: \[\boxed { I(\gamma)=S\,\Omega(\gamma) } \label{eq:ratefunSomega}\] that is, by the Fenchel-Legendre transform of \(\omega(t)\) multiplied by the number of sensors \(S\). See Appendix B.  \(\square\) ## Main Implications of Theorem 3 {#subsec:mainimplic} From Theorem 3, a number of interesting conclusions can be drawn: - The function \(\omega(t)\) in ([\[eq:omegadef\]](#eq:omegadef){reference-type="ref" reference="eq:omegadef"}) depends only upon the LMGF \(\psi(t)\) of the original statistic \(\bm{x}_k(n)\), and does *not* depend on the number of sensors. - As a consequence of the above observation, part \(ii)\) implies that the rate function (and, therefore, the large-deviation exponent) of the diffusion output depends *linearly on the number of sensors*. Moreover, the rate can be determined by knowing only the statistical distribution of the input data \(\bm{x}_k(n)\). - The rate function does not depend on the particular sensor \(k\). This implies that *all sensors are asymptotically equivalent also in terms of large-deviations*, thus strengthening what we have already found in terms of asymptotic normality---see Theorem 2 and the subsequent discussion. - Theorem 3 can be applied to the centralized stochastic algorithm ([\[eq:centstochalg\]](#eq:centstochalg){reference-type="ref" reference="eq:centstochalg"}) as well, and, again, the diffusion strategy is able to emulate, asymptotically, the *centralized* solution. Before ending this section, it is useful to comment on some essential features of the rate function \(\Omega(\gamma)\), which will provide insights on its usage in connection with the distributed detection problem. To this aim, we refer to the following convexity properties shown in Appendix C (see also , Ex. 2.2.24, and , Ex. I.16): - \(\omega^{\prime\prime}(t)>0\) for all \(t\in\mathbb{R}\), implying that \(\omega(t)\) is strictly convex. - \(\Omega(\gamma)\) is strictly convex in the interior of the set: \[{\cal D}_{\Omega}=\{\gamma\in\mathbb{R}:\;\Omega(\gamma)<\infty\}.\] - \(\Omega(\gamma)\) attains its unique minimum at \(\gamma={\mathbb{E}}\bm{x}\), with \[\Omega({\mathbb{E}}\bm{x})=0.\] In light of these properties, it is possible to provide a geometric interpretation for the main quantities in Theorem 3, as illustrated in Fig. [\[fig:RateFunGen\]](#fig:RateFunGen){reference-type="ref" reference="fig:RateFunGen"}. The leftmost panel shows a typical behavior of the LMGF of the original data \(\bm{x}_k(n)\). Using the result \(\omega^\prime(t)=\psi(t)/t\), and examining the sign of \(\psi(t)/t\), it is possible to deduce the corresponding typical behavior of \(\omega(t)\), depicted in the middle panel. As it can be seen, the slope at the origin is preserved, and is still equal to the expectation of the original data, \({\mathbb{E}}\bm{x}\). The intersection with the \(t\)-axis is changed, and moves further to the right in the considered example. Starting from \(\omega(t)\), it is possible to draw a sketch of its Fenchel-Legendre transform \(\Omega(\gamma)\) (rightmost panel), which illustrates its convexity properties, and the fact that the minimum value of zero is attained only at \(\gamma={\mathbb{E}}\bm{x}\). # The Distributed Detection Problem {#sec:detect} The tools and results developed so far allow us to address in some detail the detection problem we are interested in. Let us denote the decision regions in favor of \({\cal H}_0\) and \({\cal H}_1\) by \(\Gamma_0\) and \(\Gamma_1\), respectively. We assume that they are the same at all sensors because, in view of the asymptotic equivalence among sensors proved in the previous section, there is no particular interest in making a different choice. Note, however, that all the subsequent development does not rely on this assumption and applies, *mutatis mutandis*, to the case of distinct decision regions used by distinct agents. The Type-I and Type-II error probabilities at the \(k\)-th sensor at time \(n\) are defined as: \[\begin{aligned} \alpha_k(n)&{:=}&\P_0[\bm{y}_k(n)\in \Gamma_1 ],\\ \beta_k(n)&{:=}&\P_1[\bm{y}_k(n)\in \Gamma_0 ]. \end{aligned}\] The steady-state detection performance is: \[\lim_{n\rightarrow\infty} \alpha_k(n),\quad \lim_{n\rightarrow\infty} \beta_k(n).\] Some questions arise. Do these limits exist? Do these probabilities vanish as \(n\) approaches infinity? Theorem 1 provides the answers. Indeed, we found that \(\bm{y}_k(n)\) stabilizes in distribution as \(n\) goes to infinity. In the sequel, in order to avoid dealing with pathological cases, we shall assume that \(\P_0[\bm{y}^\star_{k,\mu}\in\partial\Gamma_1]=0\) and that \(\P_1[\bm{y}^\star_{k,\mu}\in\partial\Gamma_0]=0\). This is a mild assumption, which is verified, for instance, when the limiting random variable \(\bm{y}_{k,\mu}^\star\) has an absolutely continuous distribution, and the decision regions are not so convoluted to have boundaries with strictly positive measure. Accordingly, by invoking the weak convergence result of Theorem 1, and in view of ([\[eq:weakconv\]](#eq:weakconv){reference-type="ref" reference="eq:weakconv"}) we can write: \[\begin{aligned} \alpha_{k,\mu}&{:=}& \lim_{n\rightarrow\infty} \alpha_k(n)=\P_0[\bm{y}^\star_{k,\mu} \in \Gamma_1],\label{eq:alphabet1}\\ \beta_{k,\mu}&{:=}& \lim_{n\rightarrow\infty} \beta_k(n)=\P_1[\bm{y}^\star_{k,\mu} \in \Gamma_0], \label{eq:alphabet2} \end{aligned}\] where the dependence upon \(\mu\) has been made explicit for later use. We notice that, in the above, we work with decision regions that do not depend on \(n\), which corresponds exactly to the setup of Theorem 1. Generalizations where the regions are allowed to change with \(n\) can be handled by resorting to known results from asymptotic statistics. To give an example, consider the meaningful case of a detector with a sequence of thresholds \(\eta(n)\) that converges to a value \(\eta\) as \(n\rightarrow\infty\). Here, \[\lim_{n\rightarrow\infty}\P_h[\bm{y}_k(n) > \eta(n)]=\P_h[\bm{y}^\star_{k,\mu} > \eta],\] which can be seen, e.g., as an application of Slutsky's Theorem . From ([\[eq:alphabet1\]](#eq:alphabet1){reference-type="ref" reference="eq:alphabet1"})--([\[eq:alphabet2\]](#eq:alphabet2){reference-type="ref" reference="eq:alphabet2"}), it turns out that, as time elapses, the error probablities do not vanish exponentially. As a matter of fact, they do not vanish at all. This situation is in contrast to what happens in the case of running consensus strategies with decaying step-size studied in the literature . We wish to avoid confusion here. In the decaying step-size case, one does need to examine the effect of large deviations  for large \(n\), quantifying the rate of decay to zero of the error probabilities *as time progresses*. In the adaptive context, on the other hand, where *constant* step-sizes are used to enable continuous adaptation and learning, the large-deviation analysis is totally different, in that it is aimed at characterizing the decaying rate of the error probabilities *as the step-size \(\mu\) approaches zero.* Returning to the detection performance evaluation ([\[eq:alphabet1\]](#eq:alphabet1){reference-type="ref" reference="eq:alphabet1"})--([\[eq:alphabet2\]](#eq:alphabet2){reference-type="ref" reference="eq:alphabet2"}), we stress that the steady-state values of these error probabilities are unknown, since the distribution of \(\bm{y}^\star_{k,\mu}\) is generally unknown. However, the large-deviation result offered by Theorem 3 allows us to characterize the *error exponents* in the regime of small step-sizes. Theorem 3 can be tailored to our detection setup as follows (suffixes \(0\) and \(1\) are used to indicate that the statistical quantities are evaluated under \({\cal H}_0\) and \({\cal H}_1\), respectively):  \(\square\) [Remark I]{.smallcaps}. The technical requirement that the LMGFs \(\psi_0(t)\) and \(\psi_1(t)\) are finite is met in many practical detection problems, as already shown in . In particular, the assumption is clearly verified when the observations are discrete and supported on a finite alphabet; when they have compact support; and for shift-in-mean detection problems where the data distributions fulfill mild regularity conditions---see Remark I in  for a detailed list. [Remark II]{.smallcaps}. As typical in large-deviations analysis, we have worked with regions \(\Gamma_0\) and \(\Gamma_1\) that do not depend on the step-size \(\mu\). Generalizations are possible to the case in which these regions depend on \(\mu\). A relevant case where this might be useful is the Neyman-Pearson setup, where one needs to work with a fixed (non-vanishing) value of the false-alarm probability. An example of this scenario is provided in Sec. [6.3](#subsec:Laplace){reference-type="ref" reference="subsec:Laplace"}---see the discussion following ([\[eq:betamuNP\]](#eq:betamuNP){reference-type="ref" reference="eq:betamuNP"})---along with the detailed procedure for the required generalization. In order to visualize the main result in Theorem 4, we refer to the geometric interpretation given in Fig. [\[fig:Theo4Geo\]](#fig:Theo4Geo){reference-type="ref" reference="fig:Theo4Geo"}. Assume, without loss of generality, that \({\mathbb{E}}_0\bm{x}<{\mathbb{E}}_1\bm{x}\), and, for the sake of concreteness, consider a detector with threshold \(\eta\), amounting to the following form for the decision regions: \[\Gamma_0=\{\gamma\in\mathbb{R}:\, \gamma\leq \eta \},\qquad \Gamma_1=\mathbb{R}\setminus\Gamma_0. \label{eq:gamma0_bis}\] Let us set \({\mathbb{E}}_0\bm{x}<\eta<{\mathbb{E}}_1 \bm{x}\) since, as will be clear soon, choosing a threshold outside the range \(({\mathbb{E}}_0\bm{x},{\mathbb{E}}_1\bm{x})\) will always nullify one of the exponents. According to Theorem 4, to evaluate the exponent \({\cal E}_0\) (resp., \({\cal E}_1\)), one must consider the worst-case, i.e., the smallest value of the function \(\Omega_0(\gamma)\) (resp., \(\Omega_1(\gamma)\)), within the corresponding *error* region \(\Gamma_1\) (resp., \(\Gamma_0\)). In view of the convexity properties discussed at the end of Sec. [4.3](#subsec:mainimplic){reference-type="ref" reference="subsec:mainimplic"}, and reported in Appendix C, we see that, for the threshold detector, both minima are attained only at \(\gamma=\eta\). Certainly, this shape turns out to be of great interest in practical applications where, inspired by the optimality properties of a log-likelihood ratio test in the centralized case, a threshold detector is often an appealing and reasonable choice. On the other hand, we would like to stress that different, arbitrary decision regions can be in general chosen, and that the minima of \(\Omega_0(\cdot)\) and \(\Omega_1(\cdot)\) in Fig. [\[fig:Theo4Geo\]](#fig:Theo4Geo){reference-type="ref" reference="fig:Theo4Geo"} might be correspondingly located at two different points. In summary, Theorem 4 allows us to compute the exponents \({\cal E}_0\) and \({\cal E}_1\) as functions of \(i)\) the kind of statistic \(\bm{x}\) employed by the sensors, which determines the shape of the LMGFs \(\psi_h(t)\) to be used in ([\[eq:Omegah\]](#eq:Omegah){reference-type="ref" reference="eq:Omegah"}); and \(ii)\) of the employed decision regions relevant for the minimizations in ([\[eq:E01\]](#eq:E01){reference-type="ref" reference="eq:E01"}). Once \({\cal E}_0\) and \({\cal E}_1\) have been found, the error probabilities \(\alpha_{k,\mu}\) and \(\beta_{k,\mu}\) can be approximated using Eq. ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}). This result is then key for both detector design and analysis, so that we are now ready to illustrate the operation of the adaptive distributed network of detectors. # Examples of Application {#sec:examples} In this section, we apply the developed theory to four relevant detection problems. We start with the classical Gaussian shift-in-mean problem. Then, we consider a scenario of specific relevance for sensor network applications, namely, detection with hardly (one-bit) quantized measurements. This case amounts to testing two Bernoulli distributions with different parameters under the different hypotheses. Both the Gaussian and the finite-alphabet assumptions are removed in the subsequent example, where a problem of relevance to radar applications is addressed, that is, shift-in-mean with additive noise sampled from a Laplace (double-exponential) distribution. Finally, we examine a case where the agents have limited knowledge of the underlying data model, and agree to employ a simple sample-mean detector, in the presence of noise distributed as a Gaussian mixture. Before dwelling on the presentation of the numerical experiments, we provide some essential details on the strategy that has been implemented for obtaining them: - The network used for our experiments consists of ten sensors, arranged so as to form the topology in Fig. [\[fig:skeleton\]](#fig:skeleton){reference-type="ref" reference="fig:skeleton"}, with combination weights \(a_{k,\ell}\) following the Laplacian rule . - The decision rule for the detectors is based on comparing the diffusion output \(\bm{y}_k(n)\) to some threshold \(\eta\), namely, \[\bm{y}_k(n)\mathop{\lesseqgtr}_{{\cal H}_1}^{{\cal H}_0} \eta, \label{eq:TheTest}\] where the decision regions are the same as in ([\[eq:gamma0_bis\]](#eq:gamma0_bis){reference-type="ref" reference="eq:gamma0_bis"}). - Selecting the threshold \(\eta\) in ([\[eq:TheTest\]](#eq:TheTest){reference-type="ref" reference="eq:TheTest"}) is a critical stage of detector design and implementation. This choice can be guided by different criteria, which would lead to different threshold settings. In the following examples, we present three relevant cases, namely: \(i)\) a threshold setting that is suited to the Bayesian and the max-min criteria (Sec. [6.2](#subsec:bernoulli){reference-type="ref" reference="subsec:bernoulli"}); \(ii)\) a Neyman-Pearson threshold setting (Sec. [6.3](#subsec:Laplace){reference-type="ref" reference="subsec:Laplace"}); \(iii)\) and a threshold setting in the presence of insufficient information about the underlying statistical models (Sec. [6.4](#subsec:Gmix){reference-type="ref" reference="subsec:Gmix"}). We would like to stress that using different threshold setting rules for different statistical models has no particular meaning. These choices are just meant to illustrate different rules and different models while avoiding repetition of similar results. - The diffusion output is obtained after consultation steps involving the exchange of some local statistics \(\bm{x}_k(n)\). The particular kind of statistic used in the different examples will be detailed when needed. ## Shift-in-mean Gaussian Problem The first hypothesis testing problem we consider is the following: \[\begin{aligned} {\cal H}_0&:&\bm{d}_k(n)\sim{\cal N}(0,\sigma^2),\\ {\cal H}_1&:&\bm{d}_k(n)\sim{\cal N}(\theta,\sigma^2), \end{aligned}\] where \(\bm{d}_k(n)\) denotes the local datum collected by sensor \(k\) at time \(n\). In the above expression, \({\cal N}(a,b)\) is a shortcut for a Gaussian distribution with mean \(a\) and variance \(b\), and the symbol \(\sim\) means "distributed as\". We assume the local statistic \(\bm{x}_k(n)\) to be shared during the diffusion process is the log-likelihood ratio of the measurement \(\bm{d}_k(n)\): \[\bm{x}_k(n)=\frac{\theta}{\sigma^2}\left( \bm{d}_k(n)-\frac{\theta}{2} \right).\] Note that in the Gaussian case the log-likelihood ratio is simply a shifted and scaled version of the collected observation \(\bm{d}_k(n)\), such that no substantial differences are expected if the agents share directly the observations. In the specific case that \(\bm{x}_k(n)\) is the log-likelihood ratio, the expectations \({\mathbb{E}}_0\bm{x}\) and \({\mathbb{E}}_1\bm{x}\) assume a peculiar meaning. Indeed, they can be conveniently represented as: \[{\mathbb{E}}_0\bm{x}=-{\cal D}({\cal H}_0||{\cal H}_1),\quad {\mathbb{E}}_1\bm{x}={\cal D}({\cal H}_1||{\cal H}_0),\] where \({\cal D}({\cal H}_i||{\cal H}_j)\), with \(i,j\in\{0,1\}\), is the Kullback-Leibler (KL) divergence between hypotheses \(i\) and \(j\)---see . In particular, for the Gaussian shift-in-mean problem the distribution of the log-likelihood ratio can be expressed in terms of the KL divergences as follows: \[\bm{x}_k(n)\stackrel{{\cal H}_0}{\sim} {\cal N}(-{\cal D},2{\cal D}),\qquad \bm{x}_k(n)\stackrel{{\cal H}_1}{\sim} {\cal N}({\cal D},2{\cal D}), \label{eq:lrdist}\] where \[{\cal D}{:=} {\cal D}({\cal H}_0||{\cal H}_1)={\cal D}({\cal H}_1||{\cal H}_0)=\frac{\theta^2}{2\sigma^2},\] is the KL divergence for the Gaussian shift-in-mean case . Since the LMGF of a Gaussian random variable \({\cal N}(a,b)\) is \(a t +b t^2/2\) , we deduce from ([\[eq:lrdist\]](#eq:lrdist){reference-type="ref" reference="eq:lrdist"}) that \[\psi_0(t)={\cal D} t (t-1),\quad \psi_1(t)={\cal D} t (t+1). \label{eq:LMGFforGauss}\] Note that \(\psi_1(t)=\psi_0(t+1)\), a relationship that holds true more generally when working with the LMGFs of the log-likelihood ratio---see, e.g., . Now, applying ([\[eq:Omegah\]](#eq:Omegah){reference-type="ref" reference="eq:Omegah"}) to ([\[eq:LMGFforGauss\]](#eq:LMGFforGauss){reference-type="ref" reference="eq:LMGFforGauss"}) readily gives \[\omega_0(t)={\cal D} t\,\left(\frac t 2-1\right),\quad \omega_1(t)={\cal D} t\,\left(\frac t 2 +1\right).\] According to its definition ([\[eq:FLtransf\]](#eq:FLtransf){reference-type="ref" reference="eq:FLtransf"}), in order to find the Fenchel-Legendre transform we should maximize, with respect to \(t\), the function \(\gamma t-\omega(t)\). In view of the convexity properties proved in Appendix C, this can be done by taking the first derivative and equating it to zero, which is equivalent to writing \[\begin{aligned} \gamma&=&\omega_0^{\prime}(t_0)=\frac{\psi_0(t_0)}{t_0}\Rightarrow t_0=\frac{\gamma}{{\cal D}}+1,\\ \gamma&=&\omega_1^{\prime}(t_1)=\frac{\psi_1(t_1)}{t_1}\Rightarrow t_1=\frac{\gamma}{{\cal D}}-1. \end{aligned}\] These expressions lead to \[\Omega_0(\gamma)=\frac{(\gamma+{\cal D})^2}{2{\cal D}},\qquad \Omega_1(\gamma)=\frac{(\gamma-{\cal D})^2}{2{\cal D}}.\] Selecting the threshold \(\eta\) within the interval \((-{\cal D},{\cal D})\), the minimization in ([\[eq:E01\]](#eq:E01){reference-type="ref" reference="eq:E01"}) is easily performed---refer to Fig. [\[fig:Theo4Geo\]](#fig:Theo4Geo){reference-type="ref" reference="fig:Theo4Geo"} and the related discussion. The final result is: \[\boxed { \alpha_{k,\mu}\stackrel{\cdot}{=} e^{-(1/\mu)\,S\,\frac{(\eta+{\cal D})^2}{2{\cal D}}},\qquad \beta_{k,\mu}\stackrel{\cdot}{=} e^{-(1/\mu)\,S\,\frac{(\eta-{\cal D})^2}{2{\cal D}}} }\] These expressions provide the complete asymptotic characterization to the leading exponential order (i.e., they furnish the detection error exponents) of the adaptive distributed network of detectors for the Gaussian shift-in-mean problem, and for any choice of the threshold \(\eta\) within the interval \((-{\cal D},{\cal D})\). We have run a number of numerical simulations to check the validity of the results. Clearly, in order to show the generality of our methods, it is desirable to test them on non-Gaussian data as well. Since the interpretation of the results for both Gaussian and non-Gaussian data is essentially similar, we shall skip the numerical results for the Gaussian case to avoid unnecessary repetitions and focus on other cases. Accordingly, also the discussion on how to make a careful selection of the detection threshold \(\eta\) is postponed to the forthcoming sections. ## Hardly (one-bit) Quantized Measurements {#subsec:bernoulli} We now examine the example in which the measurements at the local sensors are hardly quantized. This situation can be formalized as the following hypothesis test: \[\begin{aligned} {\cal H}_0&:&\bm{d}_k(n) \sim{\cal B}(p_0),\label{eq:BernTest0} \\ {\cal H}_1&:&\bm{d}_k(n) \sim{\cal B}(p_1), \label{eq:BernTest1} \end{aligned}\] with \({\cal B}(p)\) denoting a Bernoulli random variable with success probability \(p\). As in the previous example, we assume that the local statistics \(\bm{x}_k(n)\) employed by the sensors in the adaptation/combination stages are chosen as the local log-likelihood ratios that, in view of ([\[eq:BernTest0\]](#eq:BernTest0){reference-type="ref" reference="eq:BernTest0"})--([\[eq:BernTest1\]](#eq:BernTest1){reference-type="ref" reference="eq:BernTest1"}), can be written as: \[\bm{x}_k(n)=\bm{d}_k(n)\ln\left(\frac{p_1}{p_0}\right) + (1-\bm{d}_k(n))\ln\left(\frac{q_1}{q_0}\right),\] where \(q_h=1-p_h\), with \(h=0,1\). Since \(\bm{d}_k(n)\in\{0,1\}\), we see that \(\bm{x}_k(n)\) is a binary random variable taking on the values \(\ln(p_1/p_0)\) or \(\ln(q_1/q_0)\). The distribution of \(\bm{x}_k(n)\) is then characterized by: \[\P_0 \left[\bm{x}_k(n)=\ln\left( \frac{p_1}{p_0} \right) \right]=p_0, ~ \P_1\left[ \bm{x}_k(n)=\ln\left( \frac{p_1}{p_0} \right) \right]=p_1,\] and, hence, the LMGFs for this example are readily computed: \[\psi_0(t)=\ln\left(\, \frac{p_1^t}{p_0^{t-1}} + \frac{q_1^t}{q_0^{t-1}}\,\right),\] \[\psi_1(t)=\ln\left(\, \frac{p_1^{t+1}}{p_0^t} + \frac{q_1^{t+1}}{q_0^t}\,\right).\] According to the relationship ([\[eq:Omegah\]](#eq:Omegah){reference-type="ref" reference="eq:Omegah"}) found in Theorem 4, these closed-form expressions are used for the evaluation of \(\omega_0(t)\) and \(\omega_1(t)\), which in turn are needed to compute the rate functions \(\Omega_0(\gamma)\) and \(\Omega_1(\gamma)\). Differently from the Gaussian example, here these tasks need to be performed numerically. The resulting rate functions are displayed in the leftmost panel of Fig. [\[fig:BernoulliTot\]](#fig:BernoulliTot){reference-type="ref" reference="fig:BernoulliTot"}, and the observed behavior reproduces what is predicted by the general properties of the rate function---see also the explanation of Fig. [\[fig:RateFunGen\]](#fig:RateFunGen){reference-type="ref" reference="fig:RateFunGen"}. Let us now examine the adaptive distributed network of detectors in operation. To do so, we must decide on how to set the detection threshold \(\eta\) in ([\[eq:TheTest\]](#eq:TheTest){reference-type="ref" reference="eq:TheTest"}). As a method for selecting the threshold, in this section we illustrate the asymptotic Bayesian criterion that prescribes maximizing the exponent of the average error probability \[p^{(e)}_{k,\mu}=\pi_0 \alpha_{k,\mu} + \pi_1 \beta_{k,\mu},\] where \(\pi_0\) and \(\pi_1\) are the prior probabilities of occurrence of hypotheses \({\cal H}_0\) and \({\cal H}_1\), respectively. It is easily envisaged that the exponent of the average error probability is determined by the worst one (slowest decay) between the Type-I and Type-II error exponents---see . As a result, optimizing the Bayesian error exponent is equivalent to a max-min approach aimed at maximizing the minimum exponent. We now apply this criterion to the considered example. To this aim, a close inspection of the rate functions in Fig. [\[fig:BernoulliTot\]](#fig:BernoulliTot){reference-type="ref" reference="fig:BernoulliTot"} is beneficial. First, as it can be seen by the close-up shown in the inset plot, setting the threshold to \(\eta=0\) would imply \[{\cal E}_0=\inf_{\gamma>0}\Omega_0(\gamma)=\Omega_0(0)=\Omega_1(0)= \inf_{\gamma\leq 0}\Omega_1(\gamma)={\cal E}_1{:=} {\cal E}. \label{eq:01exponents}\] Moreover, any other choice of the threshold \(\eta\neq 0\) makes one of the two exponents smaller than \({\cal E}\). This can be clearly visualized by varying the position of \(\eta\) in Fig. [\[fig:Theo4Geo\]](#fig:Theo4Geo){reference-type="ref" reference="fig:Theo4Geo"}, and computing the infima over the pertinent decision regions. In summary, according to whether we adopt a Bayesian or a max-min criterion, an optimal choice for the threshold in this case is \(\eta=0\). In the simulations, we refer to a sufficiently large time horizon, such that the steady-state assumption applies, and evaluate the error probabilities for different values of the step-size---see the rightmost panel in Fig. [\[fig:BernoulliTot\]](#fig:BernoulliTot){reference-type="ref" reference="fig:BernoulliTot"}. In the considered example, it is easily verified by symmetry arguments that the error probabilities (and not only the exponents) of first and second kind are equal, and therefore they equal the average error probability for any prior distribution of the hypotheses: \[\alpha_{k,\mu}=\beta_{k,\mu}=p^{(e)}_{k,\mu}.\] Accordingly, in the following description the terminologies "error probability\" and "error exponent\" can be equivalently and unambiguously referred to any of these errors. In Fig. [\[fig:BernoulliTot\]](#fig:BernoulliTot){reference-type="ref" reference="fig:BernoulliTot"}, rightmost panel, the performance of all the agents is displayed as a function of \(1/\mu\), and different agents are marked with different colors. For comparison purposes, the performance of the fully connected system is also displayed. All these probability curves have been obtained by Monte Carlo simulation. Some remarkable features are observed. First, all the different curves pertaining to different agents stay nearly parallel for sufficiently small values of the step-size \(\mu\). This is a way to visualize that \(i)\) the detection error probabilities vanish exponentially at rate \(1/\mu\); and \(ii)\) the detection error *exponents* at different sensors are equal, and further equal to that of the fully-connected system corresponding to the centralized stochastic gradient solution. This is the basic message conveyed by the large deviations analysis. Indeed, the asymptotic relationships for the error probabilities in ([\[eq:mainres2\]](#eq:mainres2){reference-type="ref" reference="eq:mainres2"}) express convergence *to the first leading order in the exponent*. It remains to show that the *exponents* of the simulated error probabilities match the *exponents* predicted by Theorem 4. This is made in the inset plot of Fig. [\[fig:BernoulliTot\]](#fig:BernoulliTot){reference-type="ref" reference="fig:BernoulliTot"}, rightmost panel, where the horizontal dashed line depicts the theoretical exponent \(S {\cal E}\), with \({\cal E}\) computed using ([\[eq:01exponents\]](#eq:01exponents){reference-type="ref" reference="eq:01exponents"}), while the solid curves represent the empirical error exponents seen at different sensors, namely the quantities \(-\mu\ln p^{(e)}_{k,\mu}\), for \(k=1,2,\dots, S\). It is observed that, as the step-size decreases, the empirical error exponents converge toward the theoretical one \(S\,{\cal E}\). A further interesting evidence seems to emerge from the numerical experiments. The error probability curves in Fig. [\[fig:BernoulliTot\]](#fig:BernoulliTot){reference-type="ref" reference="fig:BernoulliTot"}, rightmost panel, are basically ordered. Examining the relationship between this ordering and the sensor placement in Fig. [\[fig:skeleton\]](#fig:skeleton){reference-type="ref" reference="fig:skeleton"}, it is seen that the ordering reflects the degree of connectivity of each agent. For instance, sensor \(3\) has the highest number of neighbors (five), and its performance is the closest to the fully connected case. On the other hand, sensor \(8\) is the most isolated, and its error probability curve appears accordingly the highest one. Note that, since from the presented theory we learned that each agent reaches asymptotically the same detection exponent, these differences are related to higher order corrections (i.e., sub-exponential terms that are neglected in a large deviations analysis) and/or to non-asymptotic effects. A systematic and thorough analysis of the above features, as well as of their exact interplay with the network connectivity and more in general with the overall structure of the connectivity matrix \(A\), requires a refined asymptotic estimate that goes beyond the large deviations analysis carried out here. ## Shift-in-mean with Laplacian noise {#subsec:Laplace} In this section we consider another non-Gaussian example, namely, the case of a shift-in-mean detection problem with noise distributed according to a Laplace distribution. Denoting by \({\cal L}(a,b)\) a (shifted) Laplace distribution with shift parameter \(a\) and scale parameter \(b\), i.e., having the probability density function: \[f_L(\] the hypothesis test we are now interested in is formulated as follows: \[\begin{aligned} {\cal H}_0&:&\bm{d}_k(n)\sim{\cal L}(0,\sigma),\\ {\cal H}_1&:&\bm{d}_k(n)\sim{\cal L}(\theta,\sigma). \label{eq:LaplaceTest} \end{aligned}\] We assume again that the local statistics \(\bm{x}_k(n)\) are chosen as the local log-likelihood ratios: \[\bm{x}_k(n)=\frac 1 \sigma ( |\bm{d}_k(n)|-|\bm{d}_k(n)-\theta| ).\] Then, the LMGFs for this case can be computed in closed form , and are given by: \[\begin{aligned} \psi_0(t)&=&\ln\left( \frac{1-t}{1-2t}\,e^{-\rho\,t} -\frac{t}{1-2t}e^{-\rho\,(1-t)} \right), \\ \psi_1(t)&=&\ln\left( \frac{1+t}{1+2t}\,e^{\rho\,t} + \frac{t}{1+2t}e^{-\rho\,(1+t)} \right), \end{aligned}\] having defined \(\rho=\theta/\sigma\). As done before, we can use the above expressions in ([\[eq:Omegah\]](#eq:Omegah){reference-type="ref" reference="eq:Omegah"}), for performing numerical evaluation of \(\omega_0(t)\) and \(\omega_1(t)\), and of their Fenchel-Legendre transforms \(\Omega_0(\gamma)\) and \(\Omega_1(\gamma)\), which are displayed in Fig. [\[fig:LaplaceTot\]](#fig:LaplaceTot){reference-type="ref" reference="fig:LaplaceTot"}, leftmost panel. Differently from the previous section, we now consider an alternative threshold setting, which is grounded on the well-known Neyman-Pearson criterion . Its classical (asymptotic) formulation sets a maximum tolerable value for the false-alarm probability, and examines the decaying rate of the miss-detection probability (the role of the two errors can also be reversed). The main difference in relation to the setup considered so far is that we relax the condition that the Type-I error probability vanish exponentially, and this allows in general for a gain in terms of the Type-II error exponent. The procedure for the Neyman-Pearson threshold setting has been already described in Sec. [4](#sec:smallmu){reference-type="ref" reference="sec:smallmu"}---see  ([\[eq:NPth1\]](#eq:NPth1){reference-type="ref" reference="eq:NPth1"})--([\[eq:NPpfa1\]](#eq:NPpfa1){reference-type="ref" reference="eq:NPpfa1"}). Accordingly, to achieve a false-alarm probability \(\bar\alpha\), we need a threshold \[\eta=\eta_\mu={\mathbb{E}}_0\bm{x} + \sqrt{\frac{\mu \sigma_{x,0}^2}{2 S}}\,Q^{-1}(\bar\alpha). \label{eq:NPthresh2}\] It remains to evaluate the Type-II error probability \[\beta_{k,\mu}=\P_1[\bm{y}^\star_{k,\mu}\leq \eta_\mu], \label{eq:betamuNP}\] or, more precisely, the corresponding exponent \({\cal E}_1\). For this purpose, we must resort to Theorem 4. Note, however, that the threshold \(\eta=\eta_\mu\) now depends on \(\mu\) and, hence, Theorem 4 does not directly apply. As noted in Remark II, it is instructive to examine how the result of Theorem 4 can be generalized to manage similar situations. Indeed, we can work in terms of the shifted variables \[\widehat{\bm{y}}^\star_{k,\mu}=\bm{y}^\star_{k,\mu}-\sqrt{\frac{\mu \sigma_{x,0}^2}{2 S}}\,Q^{-1}(\bar\alpha),\] yielding \[\beta_{k,\mu}= \P_1[ \widehat{\bm{y}}^\star_{k,\mu}\leq {\mathbb{E}}_0\bm{x}]. \label{eq:NPexamp}\] By application of the Gärtner-Ellis Theorem to the shifted variables \(\widehat{\bm{y}}^\star_{k,\mu}\), it is easy to see that the added deterministic term (vanishing with \(\mu\)) does not alter the limiting function \(\omega_1(t)\) in ([\[eq:Omegah\]](#eq:Omegah){reference-type="ref" reference="eq:Omegah"}), and consequently the final rate function \(\Omega_1(\gamma)\). Accordingly, and based on ([\[eq:NPexamp\]](#eq:NPexamp){reference-type="ref" reference="eq:NPexamp"}), the Type-II error exponent is \[{\cal E}_1=\inf_{\gamma\leq{\mathbb{E}}_0\bm{x}}\Omega_1(\gamma)=\Omega_1({\mathbb{E}}_0\bm{x}). \label{eq:NPexpo}\] The main implication of the above result can be understood, e.g., by examining the close-up in the leftmost panel of Fig. [\[fig:LaplaceTot\]](#fig:LaplaceTot){reference-type="ref" reference="fig:LaplaceTot"}, where it is seen that: \[{\cal E}_1=\Omega_1({\mathbb{E}}_0\bm{x})> \Omega_1(0),\] the latter value being the Type-II error exponent achieved by the max-min optimal detector with zero threshold previously described. This immediately shows the gain achieved by relaxing the constraint that *both* error probabilities must vanish exponentially. We now present the numerical evidence for the Neyman-Pearson adaptive distributed detector. The middle panel in Fig. [\[fig:LaplaceTot\]](#fig:LaplaceTot){reference-type="ref" reference="fig:LaplaceTot"} shows the convergence of \(\alpha_{k,\mu}\) toward the prescribed Type-I error probability \(\bar\alpha\) as the step-size \(\mu\) goes to zero. The rightmost panel refers instead to the corresponding Type-II error probability curves. The conclusions that can be drawn are similar to those discussed in the previous example, confirming the validity of the theoretical analysis. It is also interesting to note that the ordering of the different curves, for both error probabilities, is exactly the same obtained in the Bernoulli example. Since the network employed for the simulations is unchanged, this is another clue that the ordering may be related to the structure of the connectivity matrix \(A\). ## Shift-in-mean with Gaussian mixture noise {#subsec:Gmix} As a final example, we consider the case of a shift-in-mean detection problem with noise distributed according to a zero-mean Gaussian mixture, having the probability density function \[f_{GM}( \frac{1}{\sqrt{2\pi b_1}}e^{-\frac{( + \frac{1}{\sqrt{2\pi b_2}}e^{-\frac{( \right),\] namely, a balanced mixture of normal random variables with different variances \(b_1\) and \(b_2\), and symmetric expectations \(\pm a_0\). Denoting by \({\cal N}_{mix}(a,a_0,b_1,b_2)\) a *shifted* Gaussian mixture distribution with shift parameter \(a\), we consider the following hypothesis test: \[\begin{aligned} {\cal H}_0&:&\bm{d}_k(n) \sim{\cal N}_{mix}(0,\theta_0,\sigma^2_1,\sigma^2_2),\\ {\cal H}_1&:&\bm{d}_k(n) \sim{\cal N}_{mix}(\theta,\theta_0,\sigma^2_1,\sigma^2_2). \label{eq:GMTest} \end{aligned}\] For this model, we do *not* assume that the local statistics \(\bm{x}_k(n)\) are chosen as the local log-likelihood ratios. We assume instead that the agents of the network have scarce knowledge about the underlying statistical model. They know that it is a shift-in-mean problem, and possess a rough information about the value of \(\theta\). In these circumstances, the agents decide to implement a distributed sample-mean detector, namely, they exchange the local measurements *as they are, without any additional pre-processing*. This amounts to state that \[\begin{aligned} {\cal H}_0&:&\bm{x}_k(n)\sim{\cal N}_{mix}(0,\theta_0,\sigma^2_1,\sigma^2_2),\\ {\cal H}_1&:&\bm{x}_k(n)\sim{\cal N}_{mix}(\theta,\theta_0,\sigma^2_1,\sigma^2_2). \label{eq:GMTest} \end{aligned}\] Then, the LMGFs for this case can be computed in closed form , and are given by: \[\begin{aligned} \psi_0(t)&=&\ln\left( \frac 1 2 e^{\theta_0 t +\frac{\sigma_1^2 t^2}{2}} + \frac 1 2 e^{-\theta_0 t +\frac{\sigma_2^2 t^2}{2}}\right), \\ \psi_1(t)&=&\theta t +\psi_0(t). \end{aligned}\] The above expressions are used in ([\[eq:Omegah\]](#eq:Omegah){reference-type="ref" reference="eq:Omegah"}) for evaluating numerically \(\omega_0(t)\) and \(\omega_1(t)\), and then their Fenchel-Legendre transforms \(\Omega_0(\gamma)\) and \(\Omega_1(\gamma)\). These latter are depicted in the leftmost panel of Fig. [\[fig:GmixTot\]](#fig:GmixTot){reference-type="ref" reference="fig:GmixTot"}. We assume the agents in the network are not able to optimize the choice of the detection threshold, due to their limited knowledge of the underlying statistical models. The particular value used in the simulations is \(\eta=\theta/3\), which is marked in the close-up of Fig. [\[fig:GmixTot\]](#fig:GmixTot){reference-type="ref" reference="fig:GmixTot"}, leftmost panel. It is seen that, differently from the previous examples, this choice does not correspond to a balancing of the detection error exponents, such that it is expected that the Type-I and Type-II error probabilities behave quite differently in this case. This is clearly observed in the middle (Type-I error) and rightmost (Type-II error) panels of Fig. [\[fig:GmixTot\]](#fig:GmixTot){reference-type="ref" reference="fig:GmixTot"}. The numerical evidence confirms the theoretical predictions, as well as the essential features found in all the previous examples. Moreover, it is seen that the enhanced decaying rate of the Type-II error probability arising from the unbalanced threshold setting is paid in terms of a higher Type-I error probability. ## Adaptation and detection In the simulation results illustrated so far, we focused on the system performance at steady-state. It is of great interest to consider also the *time-evolution* of the system performance, and even more to show the system at work in a *dynamic* situation where the true hypothesis is changing over time, which is truly the main motivation for an *adaptive* framework. To this aim, we return to the kind of situation described in Fig. [\[fig:adaptection\]](#fig:adaptection){reference-type="ref" reference="fig:adaptection"}, which is now re-examined in more quantitative terms by focusing on the actual error probabilities, rather than on the time-evolution of the detection statistics. Specifically, in Fig. [\[fig:finalerrprob\]](#fig:finalerrprob){reference-type="ref" reference="fig:finalerrprob"} we display the performance of three generic agents of the network, for two values of the step-size. For comparison purposes, we show also the performance of the running consensus algorithm . The underlying statistical model is the shift-in-mean with Laplacian noise detailed in Sec. [6.3](#subsec:Laplace){reference-type="ref" reference="subsec:Laplace"}, and we employ a zero-threshold detector. First, the inference/adaptation trade-off is emphasized: smaller values of \(\mu\) allow better inference (lower values of the steady-state error probabilities), at the cost of increasing the time for reliably learning that a change occurred. In this respect, the running consensus performance represents an extreme case: indeed, here the step-size is vanishing, i.e., \(\mu_n=1/n\), which explains the bad performance in terms of adaptation exhibited in Fig. [\[fig:adaptection\]](#fig:adaptection){reference-type="ref" reference="fig:adaptection"}. # Concluding Remarks and Open Issues {#sec:conclu} The asymptotic tools developed in this paper allow designing and characterizing the performance of network detectors that are *adaptive and decentralized*. We show that the steady-state detection error probabilities of each individual agent decrease exponentially with the inverse of the step-size and that cooperation among sensors makes the error exponents governing such decay equal to that of a centralized stochastic gradient solution. Closed-form expressions are derived, giving insights about the main scaling laws with respect to the fundamental system parameters. In our treatment, we studied the detection performance of the diffusion strategy, *given a certain local statistic* \(\bm{x}\). Our findings show that the steady-state observable, as well as its detection performance, in general depend upon the kind of transmitted data \(\bm{x}\). A plausible, though heuristic, choice for \(\bm{x}\) is that of the log-likelihood ratio of the measured data. However, the problem of choosing the *best* statistic is open, and we feel that the obtained results can assist in exploring the relationship between the asymptotic performance and the choice of an optimal statistic \(\bm{x}\). We would like to finally note that in order to avoid a prohibitive number of Monte-Carlo runs, the simulations in the previous section were run in the small signal-to-noise ratio regime, where the error probabilities need not be too small. In this regime, the exact rate functions could in principle be replaced by parabolic approximations (see, e.g., the leftmost plot in Fig. [\[fig:BernoulliTot\]](#fig:BernoulliTot){reference-type="ref" reference="fig:BernoulliTot"}) and a parabolic approximation is basically a Gaussian approximation. To avoid confusion, we note that the results of this work do not require any small signal-to-noise ratio assumption; they hold in greater generality. Moreover, using a Gaussian approximation will generally lead to a wrong error exponent. For the same reason of avoiding prohibitive simulation runs in the analysis of the Type-II error exponent, the Type-I error probability for the Neyman-Pearson setting of Fig. [\[fig:LaplaceTot\]](#fig:LaplaceTot){reference-type="ref" reference="fig:LaplaceTot"} was set to \(\bar\alpha=1/4\) (rather than to much smaller values) and used to illustrate the theoretical findings against the simulated curves. # Appendix A: Proof of Theorem 2 {#appendix-a-proof-of-theorem-2 .unnumbered} Motivated by the fact that the transient term in Eq. ([\[eq:mainATC\]](#eq:mainATC){reference-type="ref" reference="eq:mainATC"}) does not affect the limiting behavior of \(\bm{y}_k(n)\), we introduce the following finite-horizon variable: \[\bm{y}_k^{\star}(n){:=} \sum_{i=1}^n \bm{z}_k(i)\;=\; \sum_{i=1}^n \sum_{\ell=1}^S \mu (1-\mu)^{i-1} b_{k,\ell}(i)\bm{x}_{\ell}'(i). \label{eq:regimeterm}\] Since \(\bm{y}^{\star}_k(n)\) converges in distribution to \(\bm{y}^\star_{k,\mu}\) as \(n\rightarrow\infty\), by Lévy's continuity Theorem , the corresponding characteristic functions must converge as well. It is convenient to work in terms of the normalized variable: \[\widetilde\bm{y}^\star_{k,\mu}=\frac{\bm{y}^\star_{k,\mu}-{\mathbb{E}}\bm{x}}{\sqrt{\mu\,\sigma_x^2/(2S)}}. \label{eq:centeredy}\] Denoting by \(\varphi_{k,\mu}(t)\) the characteristic function of \(\widetilde \bm{y}^\star_{k,\mu}\), using  ([\[eq:regimeterm\]](#eq:regimeterm){reference-type="ref" reference="eq:regimeterm"}) and ([\[eq:centeredy\]](#eq:centeredy){reference-type="ref" reference="eq:centeredy"}), and taking the limit as \(n\rightarrow\infty\), we have: \[\varphi_{k,\mu}(t)={\mathbb{E}} e^{j t \widetilde\bm{y}^\star_{k,\mu}} = \prod_{i=1}^{\infty}\prod_{\ell=1}^S {\mathbb{E}} e^{j t \widetilde\bm{x}^{\prime}_{\ell}(i) \zeta_{i,\ell}},\] defined in terms of the non-random variable \[\zeta_{i,\ell}=\sqrt{2 S\mu}(1-\mu)^{i-1}b_{k,\ell}(i), \label{eq:zetadefin}\] and the centered and normalized random variable \[\widetilde{\bm{x}}_{\ell}^{\prime}(i)=\frac{\bm{x}_{\ell}^{\prime}(i)-{\mathbb{E}}\bm{x}}{\sigma_x}.\] Now, the claim of asymptotic normality in ([\[eq:CLT\]](#eq:CLT){reference-type="ref" reference="eq:CLT"}) can be proven by showing the convergence, as \(\mu\rightarrow 0\), of \(\varphi_{k,\mu}(t)\) towards the characteristic function of the standard normal distribution, \(e^{-\frac{t^2}{2}}\). It suffices to work with \(t>0\) to verify the validity of this latter property. Formally, we would like to show that the quantity: \[\left|\varphi_{k,\mu}(t)-e^{-\frac{t^2}{2}}\right| =\left|\prod_{i=1}^{\infty}\prod_{\ell=1}^S {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-e^{-\frac{t^2}{2}}\right| \label{eq:claimclaim}\] converges to zero as \(\mu\rightarrow 0\). To this aim, we start by working with a finite \(n\), and write: \[\begin{aligned} \lefteqn{ \hspace*{-20pt} \left|\prod_{i=1}^{n}\prod_{\ell=1}^S {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-e^{-\frac{t^2}{2}}\right| } \nonumber\\ &\leq& \left|\prod_{i=1}^{n}\prod_{\ell=1}^S {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-\prod_{i=1}^{n}\prod_{\ell=1}^S e^{-\frac{t^2 \zeta_{i,\ell}^2}{2}} \right|\nonumber\\ &+& \left| \prod_{i=1}^{n}\prod_{\ell=1}^S e^{-\frac{t^2 \zeta_{i,\ell}^2}{2}}-e^{-\frac{t^2}{2}}\right|. \label{eq:boundbound} \end{aligned}\] We first focus on the first term on the RHS of ([\[eq:boundbound\]](#eq:boundbound){reference-type="ref" reference="eq:boundbound"}). For complex \(w_i,z_i\), with \(|w_i|\leq 1\) and \(|z_i|\leq 1\), it is known that : \[\left|\prod_{i=1}^n w_i-\prod_{i=1}^n z_i\right| \leq\sum_{i=1}^n |w_i-z_i|. \label{eq:prodsumineq}\] Since \({\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}\) is a characteristic function, its magnitude is not greater than one , such that it is legitimate to write, in view of ([\[eq:prodsumineq\]](#eq:prodsumineq){reference-type="ref" reference="eq:prodsumineq"}): \[\begin{aligned} \lefteqn{ \hspace*{-40pt} \left|\prod_{i=1}^{n}\prod_{\ell=1}^S {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-\prod_{i=1}^{n}\prod_{\ell=1}^S e^{-\frac{t^2 \zeta_{i,\ell}^2}{2}}\right| } \nonumber\\ &&\leq \sum_{i=1}^{n}\sum_{\ell=1}^S \left| {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-e^{-\frac{t^2 \zeta_{i,\ell}^2}{2}} \right|. \end{aligned}\] The single summand on the right-hand side of the above expression is upper bounded by \[\left| {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-1+\frac{t^2\zeta^2_{i,\ell}}{2} \right| + \left| e^{-\frac{t^2 \zeta_{i,\ell}^2}{2}}-1+\frac{t^2\zeta^2_{i,\ell}}{2}\right|. \label{eq:twoterms}\] Using the fact that \({\mathbb{E}}\widetilde{\bm{x}}_{\ell}^{\prime}(i)=0\) and \({\mathbb{E}}[\widetilde{\bm{x}}_{\ell}^{\prime}(i)]^2=1\), we can further bound the first term in the above expression as \[\begin{aligned} \lefteqn{\hspace*{-5pt} \left| {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-1+\frac{t^2\zeta^2_{i,\ell}}{2} \right| } \nonumber\\ &=& \left| {\mathbb{E}} \left( e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-1-j\widetilde\bm{x}_{\ell}^{\prime}(i) t \zeta_{i,\ell} + [\widetilde\bm{x}_{\ell}^{\prime}(i)]^2\,\frac{t^2\zeta^2_{i,\ell}}{2} \right) \right|\nonumber\\ &\leq& {\mathbb{E}}|\widetilde\bm{x}_{\ell}^{\prime}(i)|^3\, \frac{t^3 \zeta_{i,\ell}^3}{6}, \end{aligned}\] where the last inequality follows from upper bounding the remainder of the Taylor expansion of the complex exponential: \[\left| e^{j t}-1-\frac{jt}{1!}-\dots-\frac{(j t)^{n-1}}{(n-1)!} \right| \leq \frac{|t|^n}{n!}.\] Likewise, the second term in ([\[eq:twoterms\]](#eq:twoterms){reference-type="ref" reference="eq:twoterms"}) is upper bounded by \(\frac{t^4\zeta^4_{i,\ell}}{8}\) since \(|e^{-s}-1+s|\leq s^2/2\) for any \(s\geq 0\). We can accordingly rewrite ([\[eq:boundbound\]](#eq:boundbound){reference-type="ref" reference="eq:boundbound"}) as: \[\begin{aligned} \lefteqn{ \hspace*{-20pt} \left|\prod_{i=1}^{n}\prod_{\ell=1}^S {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-e^{-\frac{t^2}{2}}\right| }\nonumber\\ &\leq& {\mathbb{E}}|\widetilde\bm{x}_{\ell}^{\prime}(i)|^3\, \frac{t^3}{6}\, \sum_{i=1}^{n}\sum_{\ell=1}^S \zeta_{i,\ell}^3 \nonumber\\ &+& \frac{t^4}{8}\, \sum_{i=1}^{n}\sum_{\ell=1}^S \zeta_{i,\ell}^4\nonumber\\ &+& \left|e^{-\frac{t^2}{2} \sum_{i=1}^{n}\sum_{\ell=1}^S \zeta^2_{i,\ell}}-e^{-\frac{t^2}{2}}\right|. \nonumber\\ \label{eq:takethelimits} \end{aligned}\] We now take the limit as \(n\rightarrow\infty\) in the above expression. To this aim, observe that, by the definition ([\[eq:zetadefin\]](#eq:zetadefin){reference-type="ref" reference="eq:zetadefin"}), the summation: \[\sum_{i=1}^{n}\sum_{\ell=1}^S\zeta^{m}_{i,\ell}, \qquad m=1,2,\dots \label{eq:serieszetaterms}\] is made of nonnegative terms, and is upper bounded by a convergent geometric series, since \(b_{k,\ell}\leq 1\). This implies the convergence of the series ([\[eq:serieszetaterms\]](#eq:serieszetaterms){reference-type="ref" reference="eq:serieszetaterms"}) as \(n\rightarrow\infty\). Accordingly, taking the limit as \(n\rightarrow\infty\) in ([\[eq:takethelimits\]](#eq:takethelimits){reference-type="ref" reference="eq:takethelimits"}), and using ([\[eq:claimclaim\]](#eq:claimclaim){reference-type="ref" reference="eq:claimclaim"}), we have: \[\begin{aligned} \left|\varphi_{k,\mu}(t)-e^{-\frac{t^2}{2}}\right|&=& \left|\prod_{i=1}^{\infty}\prod_{\ell=1}^S {\mathbb{E}} e^{j t \widetilde\bm{x}_{\ell}^{\prime}(i) \zeta_{i,\ell}}-e^{-\frac{t^2}{2}}\right| \nonumber\\ &\leq& {\mathbb{E}}|\widetilde\bm{x}_{\ell}^{\prime}(i)|^3\, \frac{t^3}{6}\, \sum_{i=1}^{\infty}\sum_{\ell=1}^S \zeta_{i,\ell}^3 \nonumber\\ &+& \frac{t^4}{8}\, \sum_{i=1}^{\infty}\sum_{\ell=1}^S \zeta_{i,\ell}^4\nonumber\\ &+& \left|e^{-\frac{t^2}{2} \sum_{i=1}^{\infty}\sum_{\ell=1}^S \zeta^2_{i,\ell}}-e^{-\frac{t^2}{2}}\right|. \nonumber\\ \end{aligned}\] In order to show that the first term on the RHS converges to zero as \(\mu\rightarrow 0\), it suffices to verify that: \[\begin{aligned} &&\sum_{i=1}^{\infty}\sum_{\ell=1}^S\zeta^{m}_{i,\ell}\stackrel{\mu\rightarrow 0}{\longrightarrow} 0, \qquad m=3,4, \label{eq:terms34} \\ &&\sum_{i=1}^{\infty}\sum_{\ell=1}^S\zeta^2_{i,\ell}\stackrel{\mu\rightarrow 0}{\longrightarrow} 1. \label{eq:term2} \end{aligned}\] The key for proving the above properties is Perron's Theorem, which provides a uniform bound on the convergence rate---see . Let \(\lambda_2\) be the second largest magnitude eigenvalue of \(A\). For any positive \(\lambda\) such that \(|\lambda_2|<\lambda<1\), there exists a positive constant \({\cal C}={\cal C}(\lambda,A)\), ensuring for all \(i, k\) and \(\ell\): \[\left |b_{k,\ell}(i)-\frac 1 S\right |\leq {\cal C}\lambda^i. \label{eq:Perron}\] The above result follows by noting that the largest magnitude eigenvalue of the difference matrix \(B_n-(1/S)\,\mathds{1}\mathds{1}^T\) is \(\lambda_2\), and by applying the result on the convergence rate in . According to the above discussion, let us modify the variables \(\zeta_{i,\ell}\) by replacing the matrix entries \(b_{k,\ell}(i)\) with their limit \(1/S\), namely, \[\widetilde \zeta_{i,\ell}=\sqrt{2 S\mu}(1-\mu)^{i-1}\frac 1 S,\] and introduce, for any integer \(m\geq 2\), the absolute difference: \[\begin{aligned} \lefteqn{ \left| \sum_{i=1}^{\infty}\sum_{\ell=1}^S \zeta^m_{i,\ell}-\sum_{i=1}^{\infty}\sum_{\ell=1}^S \widetilde\zeta^m_{i,\ell} \right| \leq \sum_{i=1}^{\infty}\sum_{\ell=1}^S |\zeta^m_{i,\ell}-\widetilde\zeta^m_{i,\ell}| } \nonumber\\ &=&(2S\mu)^{m/2} \,\sum_{i=1}^{\infty}\sum_{\ell=1}^S (1-\mu)^{m(i-1)}\,\left|b^m_{k,\ell}(i)-\frac{1}{S^m}\right|.\nonumber\\ \end{aligned}\] Recalling the factorization \[a^m-b^m=(a-b)\sum_{k=0}^{m-1} a^k b^{m-1-k},\] (which can be proved, for \(a\neq b\), by using the geometric sum \(\sum_{k=0}^{m-1} r^k=\frac{1-r^m}{1-r}\), and using \(r=a/b\)), along with the fact that \(b_{k,\ell}(i)\) and \(1/S\) are not greater than one, we conclude that \[\left|b^m_{k,\ell}(i)-\frac{1}{S^m}\right|\leq m \left|b_{k,\ell}(i)-\frac{1}{S}\right|,\] yielding \[\begin{aligned} \lefteqn{ \left| \sum_{i=1}^{\infty}\sum_{\ell=1}^S \zeta^m_{i,\ell}-\sum_{i=1}^{\infty}\sum_{\ell=1}^S \widetilde\zeta^m_{i,\ell} \right| }\nonumber\\ &\leq& m (2S\mu)^{m/2} \,\sum_{i=1}^{\infty}\sum_{\ell=1}^S (1-\mu)^{m(i-1)}\,\left|b_{k,\ell}(i)-\frac{1}{S}\right| \nonumber\\ &\leq& {\cal C}\lambda m (2S\mu)^{m/2} \sum_{i=1}^{\infty}\sum_{\ell=1}^S (1-\mu)^{m(i-1)} \lambda^{i-1}\nonumber\\ &=& {\cal C}\lambda m 2^{m/2}S^{m/2+1}\, \frac{\mu^{m/2}}{1-\lambda (1-\mu)^m}\stackrel{\mu\rightarrow 0}{\longrightarrow}0, \end{aligned}\] where the second inequality follows from ([\[eq:Perron\]](#eq:Perron){reference-type="ref" reference="eq:Perron"}), and the limit holds because \(\lambda<1\). In view of the above result, in order to establish ([\[eq:terms34\]](#eq:terms34){reference-type="ref" reference="eq:terms34"}) and ([\[eq:term2\]](#eq:term2){reference-type="ref" reference="eq:term2"}) it is enough to study the limiting behavior of the summation: \[\begin{aligned} \sum_{i=1}^{\infty}\sum_{\ell=1}^S \widetilde\zeta^m_{i,\ell}&=& \frac{(2\mu)^{m/2}}{S^{m/2-1}} \sum_{i=1}^{\infty} (1-\mu)^{m(i-1)}\nonumber\\ &=& \frac{2^{m/2}}{S^{m/2-1}}\, \frac{\mu^{m/2}}{1-(1-\mu)^m}. \end{aligned}\] Applying L'Hospital's rule , the limit of the RHS as \(\mu\rightarrow 0\) is: \[\left(\frac{2}{S}\right)^{m/2-1}\,\lim_{\mu\rightarrow 0} \frac{\mu^{m/2-1}}{(1-\mu)^{m-1} }, \label{eq:DelHopital}\] which converges to \(1\) for \(m=2\), and to \(0\) otherwise, completing the proof. # Appendix B: Proof of Theorem 3 {#appendix-b-proof-of-theorem-3 .unnumbered} We first list some regularity properties of \(\psi(t)\) that will be applied in the subsequent analysis---see, e.g., : 1. By assumption, \(\psi(t)<\infty\) for all \(t\in\mathbb{R}\). Since it is a LMGF, it is infinitely differentiable in \(\mathbb{R}\). Also, since \(\bm{x}\) is a non-degenerate (i.e., non deterministic) random variable, we have \[\psi^{\prime\prime}(t)>0,\qquad \forall t\in\mathbb{R}, \label{eq:psi2der}\] and, hence, \(\psi(t)\) is strictly convex in \(\mathbb{R}\). 2. With reference to the function \(\frac{\psi(t)}{t}\) appearing in ([\[eq:omegadef\]](#eq:omegadef){reference-type="ref" reference="eq:omegadef"}), we note that \[\lim_{t\rightarrow 0} \frac{\psi(t)}{t}=\psi^\prime(0), \label{eq:psiprime0}\] and, hence, \(\frac{\psi(t)}{t}\) is continuous for all \(t\in\mathbb{R}\), and the integral in ([\[eq:omegadef\]](#eq:omegadef){reference-type="ref" reference="eq:omegadef"}) is well-posed. 3. For all \(t\neq 0\), we have \[\frac{d}{dt}\frac{\psi(t)}{t}=\frac{\psi^\prime(t)\,t-\psi(t)}{t^2}, \label{eq:dertneq0}\] with \[\lim_{t\rightarrow 0}\frac{\psi^\prime(t)\,t-\psi(t)}{t^2}=\frac{\psi^{\prime\prime}(0)}{2}, \label{eq:tequal0}\] implying that \(\frac{d}{dt}\frac{\psi(t)}{t}\) is continuous for all \(t\in\mathbb{R}\). In addition, we have: \[\frac{d}{dt} \frac{\psi(t)}{t}>0,\quad \forall t\in\mathbb{R}. \label{eq:posder}\] This is immediately verified for \(t=0\) by using ([\[eq:psi2der\]](#eq:psi2der){reference-type="ref" reference="eq:psi2der"}) in ([\[eq:tequal0\]](#eq:tequal0){reference-type="ref" reference="eq:tequal0"}). For \(t\neq0\), since \(\psi(t)\) is strictly convex and differentiable in \(\mathbb{R}\), we can apply the first-order condition for strict convexity---see Eq. (3.3) in : \[\psi(a)-\psi(b)>\psi^{\prime}(b)(a-b),\quad \forall a,b\in\mathbb{R},\quad a\neq b. \label{eq:firstordcond}\] Setting \(a=0\), \(b=t\neq 0\), and using \(\psi(0)=0\), result ([\[eq:posder\]](#eq:posder){reference-type="ref" reference="eq:posder"}) now follows from ([\[eq:dertneq0\]](#eq:dertneq0){reference-type="ref" reference="eq:dertneq0"}). In the following, we denote by \(\phi^{(c)}_{\mu}(t)\) the LMGF of the steady-state variable \(\bm{y}_{k,\mu}^\star\) that would correspond to a fully connected network with uniform weights, \(a_{k,\ell}=b_{k,\ell}=1/S\) for all \(k,\ell=1,2,\dots,S\). We start by stating two lemmas (their proofs are given in the sequel). [Lemma 1]{.smallcaps} Define an auxiliary function \(f_1(t)\) whose values over the negative and positive ranges of time are scaled as follows: \[f_1(t)= t^2 \times \left\{ \begin{array}{l} \displaystyle { \max_{\tau\in[0, t]} \left(\frac{d}{d\tau}\frac{\psi(\tau)}{\tau}\right), \quad t\geq 0, } \\ \\ \displaystyle{ \max_{\tau\in[t, 0]} \left(\frac{d}{d\tau}\frac{\psi(\tau)}{\tau}\right), \quad t<0. } \end{array} \right. \label{eq:f1def}\] Then, the LMGF of \(\bm{y}_{k,\mu}^\star\) for the fully connected solution with uniform weights is: \[\boxed{ \phi_{\mu}^{(c)}(t)= \frac{S}{\mu}\, \left[ \int_{0}^{\frac\mu S \,t} \frac{\psi(\tau)}{\tau}d\tau + \sum_{i=1}^\infty c_i(t;\mu) \right] }\] where the functions \(c_i(t;\mu)\) are nonnegative and satisfy \[\sum_{i=1}^\infty c_i(t;\mu)\leq f_1\left(\frac\mu S \,t\right)\times\frac{\mu^2}{1-(1-\mu)^2}.\]  \(\square\) [Lemma 2]{.smallcaps} Let \(\lambda_2\) be the second largest eigenvalue of \(A\) in magnitude, and let \(|\lambda_2|<\lambda<1\). Define another auxiliary function as: \[f_2(t)= |t| \times \left\{ \begin{array}{l} \displaystyle{\max_{\tau\in[0, t]} |\psi^\prime(\tau)|, \quad t\geq 0,} \\ \displaystyle{\max_{\tau\in[t, 0]} |\psi^\prime(\tau)|, \quad t<0.} \end{array} \right. \label{eq:f2def}\] Then, the LMGF of the steady-state diffusion output \(\bm{y}_{k,\mu}^\star\) defined by ([\[eq:Theo1\]](#eq:Theo1){reference-type="ref" reference="eq:Theo1"}) is: \[\boxed{ \phi_{k,\mu}(t)= \phi^{(c)}_{\mu}(t) + \sum_{i=1}^\infty \sum_{\ell=1}^S c_{i,\ell}(t;\mu) }\] where the functions \(c_{i,\ell}(t;\mu)\) now satisfy \[\sum_{i=1}^\infty \sum_{\ell=1}^S |c_{i.\ell}(t;\mu)|\leq ({\cal C}\lambda S)\,\frac{f_2(\mu t)}{1-\lambda(1-\mu)},\] for a positive constant \({\cal C}\) depending on \(\lambda\) and on the combination matrix \(A\).  \(\square\) We can easily show that: \[0\leq f_1(t)<\infty,\qquad 0\leq f_2(t)<\infty,\qquad \forall t\in\mathbb{R}.\] Indeed, \(f_1(t)\geq 0\) from ([\[eq:posder\]](#eq:posder){reference-type="ref" reference="eq:posder"}), while \(f_2(t)\geq 0\) by definition. Finiteness of both functions follows from Weierstrass extreme value theorem  since, by the properties of \(\psi(t)\) discussed at the beginning of this appendix, the maxima appearing in ([\[eq:f1def\]](#eq:f1def){reference-type="ref" reference="eq:f1def"}) and ([\[eq:f2def\]](#eq:f2def){reference-type="ref" reference="eq:f2def"}) are maxima of continuous functions over compact sets for any finite \(t\). By using the above lemmas (whose proofs will be given soon), it is straightforward to prove Theorem 3. we start by proving that \[\lim_{\mu\rightarrow 0}\mu\, \phi^{(c)}_\mu(t/\mu)=S\,\int_0^{t/S} \frac{\psi(\tau)}{\tau}d\tau.\] From the above Lemma 1 we have: \[\begin{aligned} \lefteqn{ \left| \mu\, \phi^{(c)}_\mu(t/\mu)-S\,\int_0^{t/S} \frac{\psi(\tau)}{\tau}d\tau \right| } \nonumber\\ &=& S\, \,\sum_{i=1}^\infty c_i(t / \mu ;\mu) \leq S\,f_1(t/S)\times \frac{\mu^2}{1-(1-\mu)^2}\stackrel{\mu\rightarrow 0}{\longrightarrow 0}.\nonumber\\ \end{aligned}\] On the other hand, using Lemma 2, \[\begin{aligned} \lefteqn{\hspace*{-40pt} \mu\, \left| \phi_{k,\mu}(t/\mu)-\phi^{(c)}_\mu(t/\mu) \right| = \mu\, \left| \sum_{i=1}^\infty\sum_{\ell=1}^S c_{i,\ell}(t/\mu;\mu) \right| } \nonumber\\ &\leq& ({\cal C}\lambda S)\,f_2(t)\, \frac{\mu}{1-\lambda (1-\mu)}\stackrel{\mu\rightarrow 0}{\longrightarrow 0}, \end{aligned}\] and claim \(i)\) is proven. From the definition of \(\omega(t)\) in ([\[eq:omegadef\]](#eq:omegadef){reference-type="ref" reference="eq:omegadef"}) we have \(\omega^\prime(t)=\psi(t)/t\), which follows by continuity of \(\psi(t)/t\) for all \(t\in\mathbb{R}\)---see property 2) at the beginning of this appendix. Then, using the result proven in part \(i)\), since \(\omega(t)\) is differentiable in \(\mathbb{R}\), the Gärtner-Ellis Theorem  stated in Sec. [4.2](#subsec:LDP){reference-type="ref" reference="subsec:LDP"} can be applied to conclude that \(\bm{y}_{k,\mu}^\star\) must obey the LDP ([\[eq:LDPdef\]](#eq:LDPdef){reference-type="ref" reference="eq:LDPdef"}) with rate function given by the Fenchel-Legendre transform of the function \(S\,\omega(t/S)\). It is straightforward to verify that the Fenchel-Legendre transform of a function scaled in this way is simply \(S\,\Omega(\gamma)\). We now prove the two lemmas. . For the case of a fully connected network with uniform weights, the finite-horizon variable introduced in ([\[eq:regimeterm\]](#eq:regimeterm){reference-type="ref" reference="eq:regimeterm"}) reduces to \[\bm{y}_k^{\star}(n)\;=\; \sum_{i=1}^n \sum_{\ell=1}^S \mu (1-\mu)^{i-1} \frac{1}{S} \,\bm{x}_{\ell}'(i).\] Now since the LMGF is additive for sums of independent random variables, the LMGF of \(\bm{y}_k^{\star}(n)\) defined above, for any fixed time instant \(n\), is given by: \[S\,\sum_{i=1}^n \psi\left((1-\mu)^{i-1} \frac \mu S\,t\right). \label{eq:centLMGF}\] First we notice that, if we were able to show that this quantity converges as \(n\) goes to infinity, the limit will represent the LMGF, \(\phi_{\mu}^{(c)}(t)\), of the steady-state random variable \(\bm{y}^\star_{k,\mu}\) in the fully connected case, in view of the continuity theorem for the moment generating functions . Define \(g(t)=\psi(t)/t\) and let us focus initially on \(t>0\). We introduce the countably infinite partition of the interval \([0,\frac \mu S\,t]\) with endpoints \[\tau_i=(1-\mu)^{i-1} \frac \mu S\,t,\qquad i=1,2,\dots,\infty.\] Repeated application of the mean-value theorem  then gives: \[\begin{aligned} \int_{\tau_{n+1}}^{\tau_1} g(\tau)d\tau &=& \sum_{i=1}^n \int_{\tau_{i+1}}^{\tau_{i}} g(\tau) d\tau =\sum_{i=1}^n g(\bar{t}_i)\delta_i \nonumber\\ &=& \sum_{i=1}^n g( \tau_i)\delta_i + \sum_{i=1}^n g^\prime(\tilde{t}_i) (\bar{t}_i-\tau_i)\delta_i ,\nonumber\\ \end{aligned}\] where \(\bar{t}_i\in(\tau_{i+1},\tau_i)\), \(\tilde{t}_{i}\in (\bar{t}_i,\tau_i)\), and \(\delta_i=\tau_{i}-\tau_{i+1}.\) Using the explicit expressions for \(\tau_i\) and \(g(\cdot)\), we have \[\begin{aligned} \sum_{i=1}^n g(\tau_i)\delta_i&=& \sum_{i=1}^n \psi(\tau_i)\left(1-\frac{\tau_{i+1}}{\tau_i}\right)\nonumber\\ &=& \mu\,\sum_{i=1}^n \psi\left((1-\mu)^{i-1}\frac \mu S\,t\right), \end{aligned}\] and we conclude that we can write \[\mu\,\sum_{i=1}^n \psi\left((1-\mu)^{i-1}\frac \mu S\,t\right)= \int_{\tau_{n+1}}^{\tau_1} g(\tau)d\tau + \sum_{i=1}^n c_{i}(t;\mu), \label{eq:LMGF2}\] where \(c_i(t;\mu)\) is defined as: \[c_i(t;\mu)= g^{\prime}(\tilde{t}_i)(\tau_i-\bar{t}_i)\delta_i>0.\] Positiveness follows since \(\bar{t}_i\in(\tau_{i+1},\tau_i)\), and \(g^\prime(t)>0\) for all \(t\in\mathbb{R}\) in view of ([\[eq:posder\]](#eq:posder){reference-type="ref" reference="eq:posder"}). Now note that \[\sum_{i=1}^n c_i(t;\mu) \leq \sum_{i=1}^\infty \delta_i^2\,\max_{\tau \in[0,\mu t/S]} g^\prime(\tau),\] and recalling the definition of \(\delta_i\), we have \[\begin{aligned} \sum_{i=1}^\infty \delta_i^2&=& \left(\frac \mu S\,t\right)^2 \sum_{i=1}^\infty [(1-\mu)^{i-1}-(1-\mu)^i]^2\nonumber\\ &=& \left(\frac \mu S\,t\right)^2\, \frac{\mu^2}{1-(1-\mu)^2}. \end{aligned}\] The proof for the case \(t<0\) follows the same line of reasoning. We now obtain \[\sum_{i=1}^\infty c_{i}(t;\mu)\leq f_1\left(\frac \mu S\,t\right) \times \frac{\mu^2}{1-(1-\mu)^2},\] where \(f_1(\cdot)\) is defined in ([\[eq:f1def\]](#eq:f1def){reference-type="ref" reference="eq:f1def"}). As \(n\rightarrow\infty\) in ([\[eq:LMGF2\]](#eq:LMGF2){reference-type="ref" reference="eq:LMGF2"}), the first term on the RHS converges to \(\int_{0}^{\frac \mu S\,t}g(\tau)d\tau\) since the \(\tau_i\)'s define a countably infinite partition of \([0,\frac{\mu}{S}\,t]\). The second term is convergent from what was just proved. Using now ([\[eq:centLMGF\]](#eq:centLMGF){reference-type="ref" reference="eq:centLMGF"}), and letting \(n\rightarrow\infty\), we finally get \[\phi^{(c)}_\mu(t)=\frac{S}{\mu} \left[ \int_{0}^{\frac \mu S\,t} \frac{\psi(\tau)}{\tau}d\tau + \sum_{i=1}^\infty c_{i}(t;\mu) \right].\] . Using the mean-value theorem again, the LMGF of the variable \(\bm{y}_{k}^{\star}(n)\) defined earlier in ([\[eq:regimeterm\]](#eq:regimeterm){reference-type="ref" reference="eq:regimeterm"}) for diffusion networks using combination weights that are not necessarily uniform is given by \[\begin{aligned} \lefteqn{ \sum_{i=1}^n \sum_{\ell=1}^S \psi\left(\mu(1-\mu)^{i-1}b_{k,\ell}(i)t\right) } \nonumber\\ &=& S\,\sum_{i=1}^n \psi\left((1-\mu)^{i-1}\,\frac \mu S\,t\right)\nonumber\\ &+& \sum_{i=1}^n \sum_{\ell=1}^S \underbrace{\psi^{\prime}(t_{i,\ell})\mu (1-\mu)^{i-1}\,\left[b_{k,\ell}(i)-\frac 1 S\right] t}_{{:=} c_{i,\ell}(t;\mu)}, \nonumber\\ \end{aligned}\] for an intermediate variable \(t_{i,\ell}\) that, focusing first on the case \(t> 0\), must be certainly contained in the range \([0,\mu t]\), since \(b_{k,\ell}\leq 1\). This yields: \[\sum_{i=1}^\infty\sum_{\ell=1}^S |c_{i,\ell}(t;\mu)|\leq ({\cal C}\lambda S)\,\max_{\tau \in[0,\mu t]} |\psi^\prime(\tau)| \, \frac{\mu\,t}{1-\lambda(1-\mu)},\] where we used Perron's Theorem ([\[eq:Perron\]](#eq:Perron){reference-type="ref" reference="eq:Perron"}). A similar argument holds for \(t<0\). # Appendix C: Convexity properties of \(\omega(t)\) and \(\Omega(\gamma)\) {#appendix-c-convexity-properties-of-omegat-and-omegagamma .unnumbered} The following properties hold. - \(\omega^{\prime\prime}(t)>0\) for all \(t\in\mathbb{R}\), implying that \(\omega(t)\) is strictly convex. - \(\Omega(\gamma)\) is strictly convex in the interior of the set: \[{\cal D}_{\Omega}=\{\gamma\in\mathbb{R}:\;\Omega(\gamma)<\infty\}.\] - \(\Omega(\gamma)\) attains its unique minimum at \(\gamma={\mathbb{E}}\bm{x}\), with \[\Omega({\mathbb{E}}\bm{x})=0.\] \(i)\) In view of ([\[eq:omegadef\]](#eq:omegadef){reference-type="ref" reference="eq:omegadef"}) we have \(\omega^\prime(t)=\psi(t)/t\). Positivity of \(\omega^{\prime\prime}(t)\) follows now from ([\[eq:posder\]](#eq:posder){reference-type="ref" reference="eq:posder"}). \(ii)\) Consider first the following equation: \[\gamma=\omega^\prime(t). \label{eq:statpointeq}\] Since \(\omega^\prime(t)\) is strictly increasing, it makes sense to define \[\lim_{t\rightarrow+\infty} \omega^\prime(t)=\omega_{+}, \qquad \lim_{t\rightarrow-\infty} \omega^\prime(t)=\omega_{-}.\] Clearly, if \(\omega_{+}=+\infty\) and \(\omega_{-}=-\infty\), Eq. ([\[eq:statpointeq\]](#eq:statpointeq){reference-type="ref" reference="eq:statpointeq"}) will have a solution \(t\) for any \(\gamma\in\mathbb{R}\). Consider the most restrictive situation that \(\omega_-\) and \(\omega_+\) are both finite, and that \(\gamma\notin(\omega_{-},\omega_+)\). The case that only one of them is finite follows then in a straightforward manner. Recall that the Fenchel-Legendre transform \(\Omega(\gamma)\) of the function \(\omega(t)\) is defined as: \[\Omega(\gamma)=\sup_{t\in\mathbb{R}}[\gamma t-\omega(t)], \label{eq:FL}\] and let us introduce the function: \[h(t){:=}\gamma t-\omega(t). \label{eq:hdef}\] From the first-order condition for strict convexity ([\[eq:firstordcond\]](#eq:firstordcond){reference-type="ref" reference="eq:firstordcond"}) applied to the strictly convex function \(\omega(t)\), we can write, for \(t\neq 0\), \(\omega^\prime(t) t >\omega(t)\), which implies: \[h(t)>[\gamma-\omega^\prime(t)]\,t.\] If \(\gamma>\omega_+\), the term on the RHS diverges to \(+\infty\) as \(t\rightarrow + \infty\). Similarly, if \(\gamma<\omega_-\), the term on the RHS diverges to \(+\infty\) as \(t\rightarrow-\infty\). This yields: \[\sup_{t\in\mathbb{R}} h(t)=\infty,\] showing, in view of ([\[eq:FL\]](#eq:FL){reference-type="ref" reference="eq:FL"}) that the condition \(\gamma\notin(\omega_{-},\omega_+)\) implies \(\gamma\notin{\cal D}_{\Omega}\). The proof will be complete if we are able to show that \(\Omega(\gamma)<\infty\) and \(\Omega(\gamma)\) is strictly convex for \(\gamma\in(\omega_-,\omega_+)\). Now, since \(\omega(t)\) is differentiable and strictly convex in \(\mathbb{R}\), we have that, for any \(\gamma\), the function \(h(t)\) in ([\[eq:hdef\]](#eq:hdef){reference-type="ref" reference="eq:hdef"}) is differentiable and strictly concave in \(\mathbb{R}\), with \[h^{\prime}(t)=\gamma-\omega^\prime(t).\] Moreover, for \(\gamma\in(\omega_{-},\omega_+)\) the stationary-point equation \[h^{\prime}(t)=0 \Leftrightarrow \gamma=\omega^{\prime}(t) \label{eq:statpointeq2}\] admits a unique (since \(\omega^{\prime}(t)\) is strictly increasing) solution \(t(\gamma)\). The strict concavity of \(h(t)\) allows us to determine the supremum in ([\[eq:FL\]](#eq:FL){reference-type="ref" reference="eq:FL"}) as follows: \[\Omega(\gamma)=\gamma\,t(\gamma)-\omega(t(\gamma))<\infty, \label{eq:omegastarprop}\] where finiteness of \(\Omega(\gamma)\) follows by the fact that \(t(\gamma)\in\mathbb{R}\), and by finiteness of \(\omega(t)\). By further noting that \(\omega^\prime(t)\) is differentiable and \(\omega^{\prime\prime}(t)>0\), the theorem about differentiation of the inverse function  allows concluding that the derivative of the function \(t(\gamma)\) can be computed as: \[\frac{d}{d\gamma}t(\gamma)=\frac{1}{\omega^{\prime\prime}(t(\gamma))}>0.\] Then we can write \[\frac{d}{d\gamma}\Omega(\gamma)=t(\gamma)+\gamma\,\frac{d}{d\gamma}t(\gamma)-\underbrace{\omega^\prime(t(\gamma))}_{\gamma}\,\frac{d}{d\gamma}t(\gamma)=t(\gamma),\] and \[\frac{d^2}{d\gamma^2}\Omega(\gamma)=\frac{d}{d\gamma}t(\gamma)>0,\] which completes the proof. \(iii)\) We have \[\Omega(\gamma)=\sup_{t\in\mathbb{R}} [\gamma\,t-\omega(t)]\geq \gamma\,0-\omega(0)=0.\] Since \(\omega^{\prime}(0)={\mathbb{E}}\bm{x}\), from ([\[eq:omegastarprop\]](#eq:omegastarprop){reference-type="ref" reference="eq:omegastarprop"}) we conclude that \[\Omega({\mathbb{E}}\bm{x})=({\mathbb{E}}\bm{x})\,0-\omega(0)=0,\] and, hence, the minimum allowed value of zero is attained.

{'timestamp': '2014-01-24T02:09:02', 'yymm': '1401', 'arxiv_id': '1401.5742', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5742'}

# Introduction Proteins are fascinating molecules due to their ability to play many roles in biological systems. Their functions often involve complex configurational changes. Therefore the familiar aphorism that "form is function" should rather be replaced by a view of the "dynamic personalities of proteins". This is why proteins are also intriguing for theoreticians because they provide a variety of yet unsolved questions. Besides the dynamics of protein folding, the rise in the time averaged mean square fluctuation \(\langle \Delta r^2\rangle\) occurring at temperatures around \(\approx 200K\), sometimes called the "protein dynamic transition" is arguably the most considerable candidate in the search of unifying principles in protein dynamics. Protein studies lead to the concept of *energy landscape*. According to this viewpoint a protein is a system which explores a complex landscape in a highly multidimensional space and some of its properties can be related to an incomplete exploration of the phase space. The protein glass transition, in which the protein appears to "freeze" when it is cooled down to about \(200\;\)K is among them. Protein folding too can be related to this energy landscape. The famous kinetic limitation known as the Levinthal paradox, associated to the difficulty to find the native state among a huge number of possible configurations, is partly solved by the concept of a funneled landscape which provides a bias towards the native state. These considerations suggest that the dynamics of the exploration of protein phase space deserves investigation, particularly at low temperature where the dynamic transition occurs. But, in spite of remarkable experimental progress which allows to "watch protein in action in real time at atomic resolution", experimental studies at this level of detail are nevertheless extremely difficult. Further understanding from models can help in analyzing the observations and developing new concepts. However, studies involving computer modeling to study the dynamics of protein fluctuations are not trivial either because the range of time scales involved is very large. This is why many meso-scale models, which describe the protein at scales that are larger than the atom, have been proposed. Yet, their validity to adequately describe the qualitative features of a real protein glass remains to be tested. In this paper we examine a model with an intermediate level of complexity. This frustrated Gō model is an off-lattice model showing fluctuations at a large range of time scales. It is though simple enough to allow the investigation of time scales which can be up to \(10^9\) times larger than the time scales of small amplitude vibrations at the atomic level. The model, which includes a slight frustration in the dihedral angle potential which does not assume a minimum for the positions of the experimentally determined structure, exhibits a much richer behavior than a standard Gō model. Besides folding one observes a rise of fluctuations above a specific temperature, analogous to a dynamical transition, and the coexistence of two folded states. This model has been widely used and it is therefore important to assess to what extent it can describe the qualitative features of protein dynamics beyond the analysis of folding for which it was originally designed. This is why we focus our attention on its low-temperature properties in an attempt to determine if a fairly simple model can provide some insight on the protein dynamical transition. The purpose of the present article is to clarify the origin of the transition in the computer model, and to determine similarities and differences with respect to experimental observations. Although the calculations are performed with a specific model, the methods are more general and even raise some questions for experiments, especially concerning the non-equilibrium properties. This article is organized as follows. The numerical findings relating to the "dynamical transition"from previous studies are presented in Sec. [2](#sec:model){reference-type="ref" reference="sec:model"}. As a very large body of experimental studies of protein dynamics emanates from neutron scattering experiments, it is rational to seek a connection between theory and experiment by studying the most relevant experimental observable for dynamics, the incoherent structure factor (ISF). We calculate the ISF from molecular dynamics simulations of the model in Sec. [3](#sec:strfact){reference-type="ref" reference="sec:strfact"}. We show that its main features can be well reproduced by a theoretical analysis based on the one-phonon approximation, which indicates that, at low temperature, the dynamics of the protein within this model takes place in a single minimum of the energy landscape. Sec. [4](#sec:is){reference-type="ref" reference="sec:is"} proceeds to an inherent structure analysis to examine how the transitions among energy states start to play a role when temperature increases. As the freezing of the protein dynamics at low temperature is often called a "glass transition", this raises the question of the properties of the model protein in non-equilibrium situations. In Sec. [5](#sec:fdt){reference-type="ref" reference="sec:fdt"} we examine the violation of the fluctuation--dissipation theorem after a sudden cooling of the protein. We find that the effective temperature of the quenched protein, deduced from the Fluctuation-Dissipation Theorem (FDT) deviates from the temperature of the thermostat, however it relaxes towards the actual temperature with an an Arrhenius behavior as the waiting time increases. This would imply that the dynamics of the protein model is very slow but not actually glassy. This method could be useful to distinguish very slow dynamics from glassy dynamics, in experimental cases as well as in molecular dynamics simulations. Finally Sec. [6](#sec:discussion){reference-type="ref" reference="sec:discussion"} summarizes and discusses our results. # A dynamical transition in a simple protein model? {#sec:model} Following earlier studies we chose to study a small protein containing the most common types of secondary structure elements (\(\alpha\;\)helix, \(\beta\;\)sheets and loops), protein \(G\), the \(B_1\) domain of immunoglobulin binding protein (Protein Data Bank code 2GB1). It contains 56 residues, with one \(\alpha\;\)helix and four \(\beta\;\)strands forming a \(\beta\;\)sheet. We describe it by an off-lattice Gō model with a slight frustration which represents its geometry in terms of a single particle per residue, centered at the location of each \(C_{\alpha}\) carbon in the experimentally determined tertiary structure. The interactions between these residues do not distinguish between the type of amino acids. Details on the simulation process and the parametrization of the model are presented in the Appendix. In spite of its simplicity, this model appears to exhibit properties which are reminiscent of the protein dynamical transition. This shows up when one examines the temperature dependence of its mean-squared fluctuations by calculating the variance \(\Delta r^2\) of the residue distances to the center of mass as a function of temperature, defined by \[\begin{aligned} \Delta r^2 &=& \frac{1}{N}\sum_{i=1}^{N}\left( \overline{r_{i0}^2}- \overline{r_{i0}}^2 \right) \ \ \. \end{aligned}\] Here, \(N\) denotes the number of residues, and \(r_{i0}\) is distance of residue \(i\) with respect to the instantaneous center of of mass. The average \(\overline{A}\) of the observable \(A(t)\) is the time average \(\overline{A}=\frac{1}{T}\int_{0}^{T}dt\ A(t)\). The variances of 20 trajectories (Langevin dynamics simulations, each \(3\cdot 10^{7}\) time units long) were averaged for each temperature point (\(\langle \cdot\rangle\) denotes the average over independent initial conditions). Figure [\[fig:deltar2\]](#fig:deltar2){reference-type="ref" reference="fig:deltar2"} shows the evolution of \(\langle \Delta r^2 \rangle\) as a function of temperature. It exhibits a crossover in the fluctuations in the temperature range \(T/T_f = [0.4,0.5]\) resembling the transition observed for hydrated proteins in neutron scattering and Mössbauer spectroscopy experiments, the so-called *dynamical transition*. Above \(T/T_f = 0.4\), the fluctuations increase quickly with temperature whereas a smaller, linear, growth is observed below. One may wonder whether the complexity of the protein structure, reflected by the Gō model, is sufficient to lead to a dynamical transition or whether, notwithstanding the resemblance of the onset of fluctuations in the present model and the experimentally determined transition, different physical and not necessarily related events may contribute to the curves which by coincidence look similar. One can already note that, for a folding temperature in the range \(330-350K\), the range \(0.4-0.5\ T/T_f\) corresponds to \(132-175K\), lower than the experimentally observed transition occurring around \(180-200K\). If it had been confirmed the observation of a dynamical transition in a fairly simple protein model would have been very useful to shine a new light on this transition which is still not fully understood. It is generally agreed that it is hydration-dependent, but still different directions for a microscopic interpretation are being pursued, suggesting the existence or non-existence of a transition in the solvent coinciding at the dynamical transition temperature. Recently, a completely different mechanism based on percolation theory for the hydration layer has been proposed. The precise nature of the interaction between the solvent and proteins, and the driving factor behind the transition, hence still remain to be understood. The dynamical transition has often been called the *protein glass transition* due its similarity with some physical properties of structural glasses at low temperatures. In particular, it was pointed out that, for both glasses and protein solutions, the transition goes along with a crossover towards non-exponential relaxation rates at low temperatures. The comparison is however vague since the glass transition itself and notably its mechanism are ongoing subjects of research and debate. Our goal in this paper is to clarify the origin of the numerically observed transition, which moreover gives hints on the possibilities and limits of protein computer models. # Analysis of the incoherent structure factor {#sec:strfact} If computer models of proteins are to be useful they must go beyond a simple determination of the dynamics of the atoms, and make the link with experimental observations. This is particularly important for the "dynamical transition" because its nature in a real protein is not known at the level of the atomic trajectories. It is only observed indirectly through the signals provided by experiments. Therefore a valid analysis of the transition observed in the computer model must examine it in the same context, i.e. determine its consequences on the experimental observations. Along with NMR and Mössbauer spectroscopy, neutron scattering methods have been among the most versatile and valuable tools to provide insight on the internal motion of proteins. Indeed, the thermal neutron wavelength being of the order of Ångströms and the kinetic energy of the order of \(meV\)s, neutrons provide an adequate probe matching the length-and frequency scales of atomic motion in proteins. An aspect brought forward in the discussion of the dynamical transition in view of the properties of glassy materials is the existence of a *boson peak* at low frequencies in neutron scattering spectra. Such a broad peak appears to be a characteristic feature of unstructured materials as compared to the spectra of crystals. ## Incoherent structure factor from molecular dynamics trajectories In neutron scattering the vibrational and conformational changes in proteins appear as a quasielastic contribution to the dynamic structure factor \(S(\mathbf{q},\omega)\) which contains crucial information about the dynamics on different time-and length scales of the system. In scattering experiments one measures the double-differential scattering cross section \(d^2 \sigma/(d\Omega dE)\) which gives the probability of finding a neutron in the solid angle element \(d\Omega\) with an energy exchange \(dE\) after scattering. The total cross-section of the experiment is obtained by integration over all angles and energies. Neglecting magnetic interaction and only considering the short-range nuclear forces, the isotropic scattering is characterized by a single parameter \(b_i\), the scattering length of the atomic species \(i\), which can be a complex number with a non-vanishing imaginary part accounting for absorption of the neutron. If one defines the average over different spin states \(b^{coh}=\left|\langle b \rangle\right|\) as the coherent scattering length, and the root mean square deviation \(b^{inc}=\sqrt{\langle |b|^2 \rangle-\left|\langle b \rangle\right|^2}\) as the incoherent scattering length, the double-differential cross section arising from the scattering of a monochromatic beam of neutrons with incident wave vector \(\mathbf{k}_0\) and final wave vector \(\mathbf{k}\) by \(N\) nuclei of the sample can be expressed as \[\begin{aligned} \frac{d^2 \sigma}{d\Omega dE} &=& \frac{N}{\hbar} \frac{|\mathbf{k}|}{|\mathbf{k}_0|} \left(b^{coh}\right)^2 S_{coh}(\mathbf{q},\omega) \nonumber \\ & &+ \frac{N}{\hbar}\frac{|\mathbf{k}|}{|\mathbf{k}_0|} \left(b^{inc}\right)^2 S_{inc}(\mathbf{q},\omega) \ \ \ \, \end{aligned}\] where \(\mathbf{q}=\mathbf{k}-\mathbf{k}_0\) is the wave vector transfer in the scattering process and \(\mathbf{r}_i\) denote the time-dependent positions of the sample nuclei and the coherent and incoherent dynamical structure factors are \[\begin{aligned} S_{coh}(\mathbf{q},\omega)&=\frac{1}{2\pi N}\sum_{i,j}\int_{-\infty}^{\infty}dt\ e^{-i\omega t} \langle e^{-i\mathbf{q}(\mathbf{r}_i(t)-\mathbf{r}_j(0))} \rangle, \\ S_{inc}(\mathbf{q},\omega)&=\frac{1}{2\pi N}\sum_{i}\int_{-\infty}^{\infty}dt\ e^{-i\omega t} \langle e^{-i\mathbf{q}(\mathbf{r}_i(t)-\mathbf{r}_i(0))} \rangle. \end{aligned}\] The coherent structure factor contains contributions from the position of all nuclei. The interference pattern of \(S_{coh}(\mathbf{q},\omega)\) contains the average (static) structural information on the sample, whereas the incoherent structure factor \(S_{inc}(\mathbf{q},\omega)\) monitors the average of atomic motions as it is mathematically equivalent to the Fourier transform in space and time of the particle density autocorrelation function. In experiments on biological samples, incoherent scattering from hydrogens dominates the experimental spectra unless deuteration of the molecule and/or solvent are used. Since the Gō-model represents a reduced description of the protein and the locations of the individual atoms in the residues are not resolved, we use "effective" incoherent weights of equal value for the effective particles of the model located in the position of the \(C_{\alpha}\)-atoms. Such a coarse grained view assumes that the average number of hydrogens atoms and their location in the residues is homogeneous, which is of course a crude approximation in particular in view of the extension and the motion of the side chains. These approximations are nevertheless acceptable here as we do not intend to provide a quantitative comparison with experimental results considering the simplifications and the resulting limitations of the model. We generated Langevin and Nosé-Hoover dynamic trajectories of length \(t=10^5\) time units, i.e. about 1000 periods of the slowest vibrational mode of the protein, after an equilibration of equal length for protein G at temperatures in the interval \(T/T_f = [0.0459,0.9633]\). To compute the incoherent structure factor for the Gō-model of protein G, we use *nMoldyn* to analyze the molecular dynamics trajectories generated at different temperatures. The data are spatially averaged over \(N_q=50\) wave vectors sampling spheres of fixed modules \(|\mathbf{q}|=2,3,4\) Å\(^{-1}\), and the Fourier transformation is smoothed by a Gaussian window of width \(\sigma=5\%\) of the full length of the trajectory. Prior to the analysis, a root-mean square displacement alignment of the trajectory onto the reference structure at time \(t=0\) is performed using virtual molecular dynamics (VMD). Such a procedure is necessary in order to remove the effects of global rotation and translation of the molecule. Figure [\[fig425\]](#fig425){reference-type="ref" reference="fig425"} shows the frequency dependence of the incoherent structure factor \(S(q,\omega)\) for a fixed wave vector \(q=4\) Å\(^{-1}\) for a simulation with the Nosé-Hoover thermostat. On Fig. [\[fig425\]](#fig425){reference-type="ref" reference="fig425"}-a, the evolution of the low frequency range of the structure is shown for a range of temperatures including the supposed dynamic transition region \(T/T_f = [0.4,0.5]\). At low temperatures up to \(T/T_f \approx 0.51\), individual modes are clearly distinguishable and become broadened as temperatures increases. The slowest mode, located around \(4\;{\mathrm{cm}}^{-1}\) is also the highest in amplitude. It has a time constant of about \(\tau=80\) in reduced units (\(\approx 8\;\)ps). These well-defined lines are observed to be shifting towards lower frequencies with increasing temperature, similar to the phonon frequency shifts that are frequently observed in crystalline solids. As we show in the following section, the location of these lines can be calculated from a harmonic approximation associated to a single potential energy minimum. Therefore, the shift in frequency and the appearance of additional modes can be seen as a signature of increasingly anharmonic dynamics involving several minima associated to different conformational substates. If, instead of the Nosé-Hoover thermostat, we consider the results obtained with Langevin dynamics and a friction constant \(\gamma=0.01\), the stronger coupling to the thermostat leads to low energy modes which are significantly broader than with the Nosé-Hoover thermostat, so that they can hardly be resolved anymore. However the location of the peaks in the spectra remains the same as the one shown on Fig. [\[fig425\]](#fig425){reference-type="ref" reference="fig425"}-b. Besides the larger damping, Langevin calculations pose additional technical difficulties because Langevin dynamics does not preserve the total momentum of the system. The center of mass of the protein diffuses on the time scale of the trajectories. At low temperatures when fluctuations are small, the alignment procedure can efficiently eliminate contributions from diffusion as the center of mass is well defined for a rigid structure. At high temperatures however, it cannot be excluded that the alignment procedure adds spurious contributions to the structure factor calculations as the fluctuations grow in amplitude and the structure becomes flexible. The analysis of the incoherent structure factor has shown that the low temperature dynamics of the Gō-model is dominated by harmonic contributions. An increase of temperature leads to a broadening and a shift of these modes until they eventually become continuously distributed. However, for both strong and weak coupling to the heat bath, no distinct change of behavior can be detected within the temperature range \(T/T_f = [0.4,0.5]\) in which Fig. [\[fig:deltar2\]](#fig:deltar2){reference-type="ref" reference="fig:deltar2"} shows an apparent dynamical transition. Instead, the numerical results suggest a continuous increase of anharmonic dynamics, and the absence of a dynamical transition in this model, even though, in the range \(T/T_f = [0.4,0.5]\), the peaks of the structure factor in the Nosé-Hoover simulations broaden significantly. In the lowest temperature range the structure factor does not show any contribution reminiscent of a Boson peak. ## Structure factor from normal mode analysis in the one phonon approximation A further analysis can be carried out to determine whether the low temperature behavior of the protein model shows a complex glassy behavior or simply the properties of an harmonic network made of multiple bonds. The picture of a rough energy landscape of a protein with many minima separated by barriers of different height does not exclude the possibility that, in the low temperature range, the system behaves as if it were in thermal equilibrium in a single minimum of this multidimensional space. This would be the case if the time scale to cross the energy barrier separating this minimum from its neighbor basins were longer than the observation time (both in numerical or real experiments). In this case, it should be possible to describe the low temperature behavior of the protein in terms of a set of normal modes. To determine if this is true for the Gō model that we study, one can compare the spectrum obtained from thermalized numerical simulations at low temperature (low temperature curves on Fig. [\[fig425\]](#fig425){reference-type="ref" reference="fig425"}) with the calculation of the structure factor in terms of phonon modes, in the spirit of the study performed in ref.  for the analysis of inelastic neutron scattering data of staphylococcal nuclease at \(25\;\)K on an all-atom protein model. The theoretical basis for a quantitative comparison is an approximate expression of the quantum-mechanical structure factor \(S(\mathbf{q},\omega)\) in the so-called *one-phonon limit* which only accounts for single quantum process in the scattering events assuming harmonic dynamics of the nuclei. In this approximation, the incoherent structure factor can be written as \[\begin{aligned} S_{inc}(\mathbf{q},\omega)&=\sum_{i}\sum_{\lambda} b_{i}^2 e^{\hbar \omega_{\lambda}\beta/2} e^{-2 W_{i}(\mathbf{q})} \hbar |\mathbf{q.}\mathbf{e}_{\lambda,i}|^2 \nonumber \\ & \times \left(4 m_i \omega_{\lambda} \sinh(\beta\hbar\omega_{\lambda}/2) \right)^{-1} \delta(\omega-\omega_{\lambda})\. \label{nmsf} \end{aligned}\] Here, the indices \(i\) and \(\lambda\) denote the atom and normal modes indices respectively. \(\mathbf{e}_{\lambda,i}\) is the subvector relating to the coordinates of particle \(i\) of the normal mode vector associated to index \(\lambda\). \(W_{i}(\mathbf{q})\) denotes the Debye-Waller factor, which in the quantum calculation of harmonic motion reads \[\begin{aligned} W_{i}(\mathbf{q})&=&\sum_{\lambda} \frac{\hbar |\mathbf{q.}\mathbf{e}_{\lambda,i}|^2}{ m_i \omega_{\lambda} } \left[2n(\omega_{\lambda})+1 \right]\ \ \, \end{aligned}\] \(n(\omega)\) being the Bose factor associated to the energy level \(\omega\). For the calculations of the structure factor in the Gō-model within this approximation, we average \(S_{inc}(\mathbf{q},\omega)\) on a shell of \(\mathbf{q}\)-vectors by transforming the Cartesian coordinate vector \((q_x,q_y,q_z)\) into spherical coordinates \(\mathbf{q}=q\cdot \left(sin(\theta)cos(\phi),sin(\theta)sin(\phi),cos(\theta)\right)\), and generate a grid with \(N_q\) points for the interval \(\phi=[0,2\pi]\), and \(N_q\) points for \(\theta=[0,\pi]\). With this shell of vectors, we can evaluate the isotropic average \(S_{inc}(q,\omega)\). In equation ([\[nmsf\]](#nmsf){reference-type="ref" reference="nmsf"}), \(\hbar\) appears as a prefactor to the Debye-Waller factor \(W_i(\mathbf{q})\) in the exponentials and in the inverse hyperbolic function. In order to evaluate the structure factor in reduced units of the Gō-model, we therefore need to estimate the order of \(\hbar\) in a similar way as we did for the energy scale (see Appendix) by comparing the fractions \[\begin{aligned} \frac{\hbar \omega}{k_BT_f}&=&\frac{\hbar' \omega'}{(k_BT_f)'} \ \ \, \end{aligned}\] the non-primed variables denoting quantities in reduced units. In the numerical evaluation of equation ([\[nmsf\]](#nmsf){reference-type="ref" reference="nmsf"}), we discretize the spectrum of frequencies from the smallest eigenvalue to the largest mode into \(10000\) grid points to evaluate the \(\delta\)-function. We use \(N_q=225\) vectors to average on a shell of modulus \(|\mathbf{q}|=4\) Å\(^{-1}\). The summation runs over all eigenvectors except for the six smallest frequencies which are numerically found to be close to zero, and result from the invariance to overall translation and rotation of the potential energy function. In a first step, we use the coordinates of the global minimum of the Gō-model for protein G corresponding to the inherent structure with index \(\alpha_0\) to calculate the Hessian of the potential energy function. The second derivatives are calculated by numerically differentiating the analytical first derivatives at the minimum. As discussed in the Appendix, due to the presence of frustration in the potential, the experimental structure does not correspond to the global minimum of the model. The difference between the minimum and the experimental structure is however small, with root-mean square deviation \(0.16\) Å and notable changes in position occurring only for a small number of residues located in the second turn. To estimate the normal mode frequencies in absolute units, we use the conversion of the time unit of \(0.1\;\)ps introduced in the Appendix. The conversion into wave numbers, which is convenient for the comparison to experimental data and to the results from all-atom calculations, is achieved by noting that, from \(c k = f\), we can assign the conversion \(1\;{\mathrm{ps}}^{-1} \rightarrow 33.3\;{\mathrm{cm}}^{-1}\) and multiply the frequencies by this scaling factor. Figure [\[fig481\]](#fig481){reference-type="ref" reference="fig481"}-a shows the results of the calculation of the incoherent structure factor \(S(q=4\)Å\(,\omega)\) in the one phonon approximation at the temperature \(T=0.0459 \, T_f\). Since in this approximation the normal mode frequencies enter with a delta function into equation ([\[nmsf\]](#nmsf){reference-type="ref" reference="nmsf"}), there is no line width associated to these modes unless the structure factor is convoluted with an instrumental resolution function or a frictional model. Comparing to the structure factor calculated from a molecular dynamics trajectory at the same temperature (Fig. [\[fig481\]](#fig481){reference-type="ref" reference="fig481"}-b), we find a good correspondence of the location of the lines and their relative amplitude with respect to each other. Therefore the analysis of the incoherent structure factor using a harmonic approximation quantitatively confirms the dominant contribution of harmonic motion at low temperatures. In particular, the motion at very low temperatures occurs in a single energy well associated to one conformational substate. To see how this behavior changes with increasing temperature, in the next section we analyze the distribution of inherent structures with temperature. # Inherent structure analysis in the dynamic transition region {#sec:is} The freezing of the dynamics of a protein at temperatures below the "dynamic transition" is also described as a "glass transition". This leads naturally to consider an energy landscape with many metastable states, also called "inherent states" in the vocabulary of glass transitions. In refs.  we showed that the thermodynamics of a protein can be well described in terms of its inherent structure landscape, i.e. a reduced energy landscape which does not describe the complete energy surface but only its minima. This picture is valid at all temperatures, including around the folding transition and above. For our present purpose of characterizing the low temperature properties of a protein and probe its possible relation with a glassy behavior, it is therefore useful to examine how the protein explores its inherent structure landscape in the vicinity of the dynamic transition. Here, we shall try to find how the number of populated minima changes with temperature around the transition region \(T/T_f = [0.4,0.5]\) for the Gō-model of protein G, and which conformational changes can be associated to these inherent structures. For three selected temperatures \(T_1=0.275\ T_f\), \(T_2=0.39\ T_f\), \(T_3=0.482\ T_f\) shown on Fig. [\[fig:deltar2\]](#fig:deltar2){reference-type="ref" reference="fig:deltar2"}, we generated \(10\) trajectories from independent equilibrated initial conditions for \(2\cdot 10^7\) reduced time units using the Nosé-Hoover thermostat. Along each trajectory, a minimization was performed every \(2\cdot 10^4\) time units such as to yield \(N_m=20000\) minima for each temperature point. In the classification of these minima and their graphical representation, we only keep those minima which have been visited at least \(2\) times within the \(N_m\) minima, which lead to discard less than \(10\) events from the total number of counts. Most of the counts are concentrated on a small number of inherent structures. In figure [\[figihs\]](#figihs){reference-type="ref" reference="figihs"}, we show the relative populations of the inherent structures on a two-dimensional subspace spanned by the inherent structure energy and the structural difference with the experimental structure measured by the dissimilarity factor. The radius of the circles centered at the location of the minima on this plane is set proportional to \(1/2\ \log(w)\) where \(w\) is the absolute number of counts of this minimum along the trajectories. This definition is necessary to allow the graphical representation on the plane, however, it may visually mask that linear differences in the radii translate into exponential differences of the frequency of visit of the minimum. As an example, the minima \(\alpha_0,\alpha_1,\alpha_2,\alpha_3\) have the occupation probabilities \(p(\alpha_0)=w(\alpha_0)/N_m\approx\ 92\%\), \(p(\alpha_1)\approx \ 8\%\), \(p(\alpha_2)\approx\ 0.1\%\) and \(p(\alpha_3)\approx\ 0.02\%\) at \(T=T_1\). From figure [\[figihs\]](#figihs){reference-type="ref" reference="figihs"}, we notice that already at \(T_1\) more than one minimum is populated though the global minimum \(\alpha_0\) is dominant. In these figures, lines are drawn between minima that are connected along the trajectory, i.e. that form a sequence of events. It should however be noted that since the sampling frequency is low, it cannot be excluded that an intermediate corresponding to an additional connection line is skipped. Connections between all minima may therefore exist even though they did not appear in the sequences observed in this study. Moving to higher temperatures \(T_2\) and \(T_3\), a larger number of minima which are both higher in energy and structural dissimilarity appear. As the temperature rises, their population numbers become more important, as can be seen e.g. by inspection of the radii of the block \(\alpha_4\)-\(\alpha_7\) on figure [\[figihs\]](#figihs){reference-type="ref" reference="figihs"}. To obtain a physical picture of the conformational changes associated with these minima, it is useful to align their coordinates onto the coordinates of the global minimum. The results of such an alignment are shown in figure [\[figihs2\]](#figihs2){reference-type="ref" reference="figihs2"}. In this figure, the coordinates of the effective Gō-model particles located at the positions of the \(C_{\alpha}\)-atoms for each amino-acid are drawn in red color. One notices that the conformational changes associated to \(\alpha_1\)-\(\alpha_3\) are small. It is interesting to notice that these small changes already appear in the range of temperatures where the rise in fluctuations seems to grow still linearly with the temperature. The next higher minima involve in particular a reorientation of a turn within the \(\beta\)-sheets of a protein. The temperature range at which these minima start to be populated coincides with the transition region revealed by the mean-distance displacement \(\langle \Delta r^2\rangle\), suggesting that the anharmonic motion required to make transitions between the basins of these minima is at its origin. We again observe that the dynamic transition region does not exhibit any particular change of behavior that could deserve the name of "transition", but rather a gradual evolution which gets noticeable in the range \(T/T_f = [0.4,0.5]\). In the next section we use a non-equilibrium approach to reveal whether the dynamics below the transition range can be characterized as "glassy" or not. # Test of the fluctuation-dissipation theorem (FDT)-a non-equilibrium approach {#sec:fdt} An alternative approach to study the low temperature transition, for which equilibrium simulations take a significant amount of computer time, consists in the test of the response of the protein to external perturbations. Rather than waiting a long time to see rare fluctuations dominating the average fluctuation at low temperatures, the system is driven out of equilibrium on purpose to either observe the relaxation back to equilibrium and its associated structural changes, or the response to a continuous perturbation to be compared to fluctuations at equilibrium. The fluctuation-dissipation theorem (FDT) relates the response to small perturbations and the correlations of fluctuations at thermal equilibrium for a given system. In the past years, the theorem and its extensions have become a useful tool to characterize glassy dynamics in a large variety of complex systems. For glasses below the glass transition temperature, the equilibrium relaxation time scales are very large so that thermal equilibrium is out of reach. Consequently, the FDT cannot be expected to hold in these situations, and the response functions and correlation functions in principle provide distinct information. In this section, we test the FDT for the Gō-model of protein G at various temperatures to see whether a signature of glassy dynamics is present in the system. To this aim, we first recall the basic definitions and notations for the theorem. In our studies we start from a given initial condition and put the system in contact with a thermostat during a waiting time \(t_w\). The end of the waiting time is selected as the origin of time (\(t=0\)) for our investigation. If \(t_w\) is large enough (strictly speaking \(t_w \to \infty\)) the system is in equilibrium at \(t=0\). We denote the Hamiltonian of the unperturbed system \(H_0\), which under a small linear perturbation of the order \(\epsilon(t)\) acting on an observable \(B(t)\) becomes \[\begin{aligned} H&=&H_0-\epsilon(t) B(t) \ \ \, \end{aligned}\] where for \(\epsilon=0\) we recover the unperturbed system. For any observable \(A(t)\), we accordingly define the two ensemble averages \(\langle A(t) \rangle_0^{t_w}\) and \(\langle A(t) \rangle_{\epsilon}^{t_w}\) where the index references the average with respect to the unperturbed/perturbed system respectively and the exponent \(t_w\) indicates how long the system was equilibrated before the start of the investigation. The correlation function in the unperturbed system relating the observables \(A(t)\), \(B(t')\) at two instances of time \(t,t'\) is defined by \[\begin{aligned} C_{AB}(t,t')&=&\langle A(t) B(t')\rangle_0^{t_w}-\langle A(t) \rangle_0^{t_w} \;\cdot \langle B(t')\rangle_0^{t_w} \ \ \. \label{eq:defcorrel} \end{aligned}\] The susceptibility \(\chi_{AB}(t)\), which measures the time-integrated response of the of the observable \(A(t)\) at the instant \(t\) to the perturbation \(\epsilon(t')\) at the instant \(t'\), reads \[\begin{aligned} \chi_{AB}(t)&=&\int_{t_0}^{t}dt'\ \frac{\delta \langle A(t) \rangle_{\epsilon}^{t_w}}{\delta \epsilon(t')} \ \;. \end{aligned}\] The index \(B\) in the susceptibility indicates that the response is measured with respect to the perturbation arising from the application of \(B(t)\), and the lower bound \(t_0\) of the integral indicates the instant of time at which the perturbation has been switched on. The integrated form of the FDT states that the correlations and the integrated response are proportional and related by the system temperature at equilibrium \[\begin{aligned} & \chi_{AB}(t) =\frac{1}{k_BT} \; \Delta C \quad \mathrm{with} \nonumber \\ & \Delta C = \left( C_{AB}(t,t)-C_{AB}(t,0) \right) \ \. \label{fldis} \end{aligned}\] In the linear response regime for a sufficiently small and constant field \(\epsilon\), the susceptibility can be approximated as \[\begin{aligned} \chi_{AB}(t)&\approx&\frac{\langle A (t) \rangle_{\epsilon}^{t_w}-\langle A (t) \rangle_{0}^{t_w}}{\epsilon} \label{eq:chilin} \end{aligned}\] such that in practice, verifying the FDT accounts for the comparison of observables on both perturbed and unperturbed trajectories.\  \ The basic steps for a numerical experiment aiming to verify the FDT can be summarized as follows:\   - Initialize two identical systems 1 and 2; 1 to be simulated with and 2 without perturbation. - Equilibrate both systems without perturbation during \(t_w\). - At time \(t_0\) (in practice \(t_0=0\), i.e. immediately after the end of the equilibration period) switch on the perturbation for system 1 and acquire data for both systems for a finite time \(t_{\mathrm{FDT}}\). - Repeat the calculation over a large number of initial conditions to yield the ensemble averages \(\langle\cdot\rangle_0^{t_w}\) and \(\langle\cdot\rangle_{\epsilon}^{t_w}\); combine the data according to equation ([\[fldis\]](#fldis){reference-type="ref" reference="fldis"}). The protocol may be modified to include an external perturbation which break the translational invariance in time. For instance the initial state can result from a quench from a high to a low temperature. Then the system is only equilibrated for a short time \(t_w\) before the perturbation in the Hamiltonian is switched on. In this case the distribution of the realizations of the initial conditions is *not the equilibrium distribution* so that the correlation function defined by Eq. ([\[eq:defcorrel\]](#eq:defcorrel){reference-type="ref" reference="eq:defcorrel"}) depends on the two times \(t\) and \(t'\) and not only on their difference. ## Simulation at constant temperature This case corresponds to \(t_w \to \infty\). In our calculations we start from an initial condition which as been thermalized for at least \(5000\;\) time units. The first step is to make an appropriate choice for the perturbative potential \(\epsilon(t) B(t)\). An earlier application of the FDT to a protein model has used the perturbative term \[\label{eq:pothayashi} \epsilon(t) B(t) = \epsilon \sum_{i=1}^N \cos(k \; y_i) \;,\] where \(k\) is a scalar, \(y_i\) the \(y\) coordinate of amino-acid \(i\) and \(\epsilon \not= 0\) for \(t>t_0\) a constant. This perturbation is invariant neither by a translation of the system nor by its rotation. Although this does not invalidate the FDT, this choice poses some problems for the accuracy of the calculations because, even in the absence of internal dynamics of the protein, the perturbation varies as the molecules diffuse in space or rotate. To avoid this difficulty we selected the perturbation \[\label{eq:potpert} W :=-\epsilon B(t) =-\epsilon \sum_{i=1, i\not= 28}^N \cos(k \; r_{i,28}) \;,\] where \(r_{i,28}\) is the distance between amino-acid \(i\) and the amino-acid \(28\) which has been chosen as a reference point within the protein because it is located near the middle of the amino-acid chain. Such a potential only depends on the internal state of the molecule, while it remains unaffected by its position in space. To test the FDT for the Gō-model of protein G using Eqs. ([\[fldis\]](#fldis){reference-type="ref" reference="fldis"}) and ([\[eq:chilin\]](#eq:chilin){reference-type="ref" reference="eq:chilin"}), we add this potential \(W\) to the potential energy \(V\) of the model and we select \(A(t) = B(t)\). The thermal fluctuations are described with the same Langevin dynamics as previously. We switch on the perturbation for the equilibrated protein model and record \(50000\) to \(400000\) trajectories (depending on the value of \(\epsilon\)) of duration \(2000\) time units for temperatures in the range \(T/T_f = [0.275,0.826]\) covering both the low temperature domain and the approach of the folding transition of the protein. The perturbation prefactors chosen in this first set of simulations were \(\epsilon=0.05\) and \(\epsilon = 0.005\), and the wave number of the cosine-term was \(k=2 \pi / 10\). Figure [\[fig:fdtequilibrium\]](#fig:fdtequilibrium){reference-type="ref" reference="fig:fdtequilibrium"} shows the evolution of the relation between the susceptibility and the variation of the correlation function \(\Delta C = C_{AB}(t,t)-C_{AB}(t,0)\). The straight lines represent the slopes expected from the FDT. One notices that, for \(\epsilon = 0.05\), at \(T = 0.826 \, T_f\), in the long term the value of \(\chi / \Delta C\) stabilizes around a value which is away from the expected value \(1/k_B T\). From a first glance, this result is reminiscent of the properties of a glass driven out of equilibrium. In this context, the deviation from the slope expected from the FDT is interpreted as the existence of an "effective temperature" for non-equilibrium systems. For the case studied here, finding an effective temperature would be surprising as the results are obtained from measurements on a thermalized protein model, i.e. a system in a state of thermal equilibrium. How is it then possible to explain the apparent deviation from the FDT? The calculations performed with \(\epsilon = 0.005\) give the clue because they show that the deviation appeared because the perturbation was too large and outside of the linear response regime assumed to calculate the susceptibility because for this lower value of \(\epsilon\) the deviation has vanished. If one computes the average value of the perturbation energy \(\langle W \rangle\) and compares it to the protein average energy \(\langle E(T) \rangle\), for the case shown on Fig. [\[fig:fdtequilibrium\]](#fig:fdtequilibrium){reference-type="ref" reference="fig:fdtequilibrium"}, \(\epsilon = 0.05\), one finds \(\langle W \rangle / \langle E(T) \rangle = 1.3\cdot 10^{-2}\). This is small, but, at temperatures which approach \(T_f\) the protein is a highly deformable object and even a small perturbation can bring it out of the linear response regime. This shows up by the the a rise of \(\chi\) versus time for \(\epsilon = 0.05\). At low temperatures the protein is more rigid an therefore more resilient to perturbations. The insets on Fig. [\[fig:fdtequilibrium\]](#fig:fdtequilibrium){reference-type="ref" reference="fig:fdtequilibrium"} show that for \(\epsilon = 0.05\) the calculations find that the fluctuation-dissipation relation at \(T = 0.275 \, T_f\) is almost perfectly verified although a very small deviation can still be detected for this value of \(\epsilon = 0.05\). Therefore a careful choice of parameters is necessary to test the FDT under controlled conditions. In particular, the perturbation needs to be carefully chosen to only probe the internal dynamics and not to dominate them. ## Simulation of quenching A typical signature of a glassy system is its aging after a perturbation. In the context of the protein "glass transition", one can therefore expect to detect a slow evolution of the system as a function of the time after which it has been brought to the glassy state. This is usually tested in quenching experiments, which can be investigated by a sharp temperature drop in the numerical simulations. Our calculations start from an equilibrium state at high temperature \(T =1.40\,T_f\), which is abruptly cooled at a temperature \(T_q\) below the temperature of the dynamical transition studied in the previous sections. The model protein is then maintained at this temperature \(T_q\) by a Langevin thermostat. After a waiting time \(t_w\) we start recording the properties of the system over a time interval \(t_{\mathrm{FDT}} = 25000\;\)t.u. to probe the fluctuation dissipation relation. In order to avoid nonlinear effects we use a small value of \(\epsilon = 0.005\). For such a weak perturbation, the response is weak compared to thermal fluctuations and a large number of realizations (\(50000\) or more) is necessary to achieve reliable statistical averages. To properly probe the phase space of the model, these averages must be made over different starting configurations before quenching. This is achieved by starting the simulations from a given initial condition properly thermalized at \(T = 1.40\,T_f\) in a preliminary calculation. Then we run a short simulation at this initial temperature, during which the unfolded conformations change widely from a run to another with different random forces because at high temperatures the fluctuations of the protein are very large. The conformations reached after this short high-\(T\) thermalization are the conformations which are then quenched to \(T_q\), for the FDT analysis. Typical results are shown on Figs. [\[fig:timequench\]](#fig:timequench){reference-type="ref" reference="fig:timequench"} and [\[fig:fdquench\]](#fig:fdquench){reference-type="ref" reference="fig:fdquench"} for two values of \(t_w\). The time evolution of the energy shows that, after the waiting time \(t_w\), even for the largest value \(t_w = 50000\;\)t.u. the model protein is still very far from equilibrium because its energy is well above the ground state energy (chosen as the reference energy \(0\)). This non-equilibrium situation sometimes leads to rapid energy drops, generally accompanied by a decrease of the dissimilarity with the native state, which superimpose to random fluctuations which have to be expected for this system in contact with a thermal bath. As expected the sharp variations of the conformations are more noticeable for the shortest waiting time. Figure [\[fig:fdquench\]](#fig:fdquench){reference-type="ref" reference="fig:fdquench"} shows that, while for small values of \(\Delta C = \left( C_{AB}(t,t)-C_{AB}(t,0) \right)\), which also correspond to shorter times after we start to collect the data for the FDT test, the variation of \(\chi_{AB}(t)\) versus \(\Delta C\) follows the curve given by the FDT relation, then at larger \(\Delta C\) the curve shows a significant deviation from the slope \(1/T\), which defines an effective temperature \(T_{\mathrm{eff}} > T\). The effective temperature is larger for short waiting times after the quench and decreases when \(t_w\) increases. This should be expected because, in the limit \(t_w \to \infty\) we should have \(T_{\mathrm{eff}} \to T\) when the system reaches equilibrium. It is not surprising to find a deviation from the FDT behavior after a strong quench of the protein model because we put the system very far from equilibrium. Therefore the observation of an effective temperature that differs from the actual temperature of the system is *not* a sufficient indication to conclude at the existence of a glassy state of the protein model. What is important is the timescale at which the system tends to equilibrium and how it depends on temperature. To study this we have performed a systematic study of the variation of \(T_{\mathrm{eff}}(t_w;T)\) as a function of the waiting time \(t_w\) and temperature \(T\), at the temperatures \(T= 0.1875\,T_f\), \(T= 0.2752\,T_f\), \(T= 0.3670\,T_f\), \(T= 0.4128\,T_f\) and \(T = 0.4817\,T_f\). The temperature domain that we can study numerically is limited both from below and from above. At the lowest temperatures the relaxation of the system is very slow so that \(t_w\) must be strongly increased. Moreover the speed at which the protein model explores its phase space by moving from an inherent structure to another becomes very low and statistically significant data cannot be obtained without a large increase of \(t_{\mathrm{FDT}}\). Running enough calculations to get a good average on the realizations becomes unpractical. As discussed above, at high temperatures the protein becomes "soft" so that one quickly leaves the linearity domain of the FDT, unless the applied perturbation becomes very small. But then the large thermal fluctuations reduce the signal to noise ratio. Therefore the advantage of faster relaxation times at high temperature is whipped out by the need to make statistical averages over a much larger number of realizations. However the temperature range over which one can get statistically significant results overlaps the temperature \(T \approx 0.45\,T_f\) above which the fluctuations of the model appear to grow faster (Fig. [\[fig:deltar2\]](#fig:deltar2){reference-type="ref" reference="fig:deltar2"}) so that one could expect to observe a change in the properties of the system at this temperature, if it existed. At a given temperature \(T\) we have defined a relaxation time \(\tau(T)\) by assuming that the effective temperature relaxes exponentially towards the actual temperature according to \[\label{eq:teffT} \left( \frac{T_{\mathrm{eff}}}{T}-1 \right) \propto \exp \left(- \frac{t_w}{\tau(T)} \right) \;.\] Figure [\[fig:tefftw\]](#fig:tefftw){reference-type="ref" reference="fig:tefftw"} shows that this assumption is well verified by the numerical calculations. It should however be noticed that, for the longest waiting times, we may observe a large deviation from the exponential decay, as shown in Fig. [\[fig:tefftw\]](#fig:tefftw){reference-type="ref" reference="fig:tefftw"} for the results at \(T=0.4128\,T_f\). We attribute this to the limitations of our observations because, when the system has sufficiently relaxed so that its effective temperature approaches the actual temperature, all subsequent relaxations become extremely slow and may exceed the observation time \(t_{\mathrm{FDT}} = 25000\;\)t.u. so that the test of the FDT no longer properly probes the phase space. Increasing \(t_{\mathrm{FDT}}\) by an order of magnitude might allow us to observe the relaxation further but is beyond our computing possibilities as we have to study at least \(50000\) realizations or more to achieve a reasonable accuracy. A fit of the values of \(\left( {T_{\mathrm{eff}}}/{T}-1 \right)\) versus \(t_w\), which takes into account the statistical weight of each point according to its standard deviation obtained from the uncertainties on \(T_{\mathrm{eff}}\) determines the relaxation time \(\tau(T)\) and its corresponding standard deviation. Figure [\[fig:relaxtime\]](#fig:relaxtime){reference-type="ref" reference="fig:relaxtime"} shows the variation of \(\tau(T)\) with temperature, in logarithmic scale, versus \(T_f/T\). It shows that, except for the value at the lowest temperature \(T = 0.1875\,T_f\), i.e. \(T_f/T = 5.45\), within the numerical errors evaluated at each stage of our calculation, the relaxation time obeys a standard Arrhenius relation \[\label{eq:arrheniustau} \tau(T) = \tau_0 \exp\left( \frac{E_a}{T}\right)\] with an activation energy \(E_a/(k_B T_f) = 0.5668 \pm 0.1799\). At the lowest temperature, the relaxation temperature estimated from the Arrhenius law is \(\tau(T=0.1875\, T_f) \approx 100000\,\)t.u. so that calculations with \(t_w \ge 100000\) as well as \(t_{\mathrm{FDT}} \gg 100000\) would be necessary, which is unpractical. However the observed deviation from Arrhenius law at low \(T\) cannot be attributed to a low-temperature glass transition because such a transition would lead to a relaxation time larger than predicted by the Arrhenius relation, while we observe the opposite. In any case the Arrhenius relation is well verified for a temperature range which overlaps the temperature \(T \approx 0.45\,T_{f}\), i.e. \(T_f/T \approx 2.22\), above which dynamical simulations suggested a possible increase of the fluctuations (Fig. [\[fig:deltar2\]](#fig:deltar2){reference-type="ref" reference="fig:deltar2"}). Therefore the relaxation of the protein model after a quench appears to follow a standard activated process, with an activation energy of the order of \(0.57 \;k_B T_f\), without any sign of a glassy behavior. These observations can be compared with other studies of the fluctuations of the same Gō model of protein G. Paper \[\] investigated the fluctuation *in equilibrium* below the folding temperature \(T_f\). In these conditions, the numerical simulations of a protein which is near its native state detect small, up and down, jumps of the dissimilarity factor, which, in any case, stays very low (\(d \approx 0.06\) for the equilibrium temperature \(T=0.55 \, T_f\)) but switches between values that differ by \(\approx 0.01\). In its equilibrium state the protein may jump from an inherent structure to another but these fluctuations are much slower than the one that we observed shorter after a temperature quench because they occur on a time scale of the order of \(10^7\;\)t.u. Their activation energy had been found to be \(E_B = 6.2\,T_f\), i.e. much higher than the activation energy \(E_a\) that characterizes the relaxation of the effective temperature that we measured. Those results are not in contradiction because they correspond to fundamentally different phenomena. The non-equilibrium fluctuations that we discuss in the present paper appear because the potential energy surface has minima on the side of the "funnel" that leads to the native state. Such minima, corresponding to protein structures which are not fully folded, can temporarily trap the protein in intermediate states. However the lifetime of these high-energy minima is only of the order of \(10^5\;\)t.u. i.e. they are short lived compared to the residence time of the protein in an inherent structure close to the native state. When the protein escapes from one of these high-energy minima we observe an energy drop, as shown in Fig. [\[fig:timequench\]](#fig:timequench){reference-type="ref" reference="fig:timequench"}. The study that we presented here is neither a study of the protein near equilibrium, nor an investigation of the full folding process which also occurs on much longer time scales (typically \(10^7\)--\(10^8\;\)t.u.) as observed in Refs. \[\] and \[\]. Therefore the activation energy \(E_a\) is also different from the energy barrier for folding. For the same reason the effective temperature after the quench \(T_{\mathrm{eff}}\) should not be confused with the configurational temperature \(T_{\mathrm{cnf}}\) defined in Ref. \[\] which relates the entropy and energy of the inherent structures during folding. \(T_{\mathrm{cnf}}\) gives a global view of the phase space explored during folding, and it evolves with a characteristic time of \(10^7\)--\(10^8\;\)t.u. as the folding itself. Compared to these scales, the FDT analysis that we presented here appears as a snapshot of the strongly non-equilibrium state created after a fast quench. It offers a new view of the time evolution of the protein model, which completes the ones which had been presented earlier. # Discussion {#sec:discussion} The starting point of our study has been the numerical observation of a low temperature transition in a simplified protein model resembling the experimentally observed dynamical transition in hydrated protein samples. This suggested that, in spite of its simplicity, the frustrated Gō model could be used not only to study the folding of proteins but also their dynamics in the low temperature range, opening the way for an exploration of the glassy behavior of proteins. We have therefore used different approaches to further characterize the properties of the protein model in the low temperature range. Thermalized molecular dynamics simulations have been used to calculate the incoherent structure factor than one could expect to observe for the protein. It shows peaks that broaden as temperature increases, suggesting that the dynamics of the protein model is dominated by harmonic or weakly anharmonic vibrations. This has been confirmed by the calculation of the structure factor in the one-phonon approximation. All the main features of the structure factor obtained by simulations, such as the peak positions and even the power law decay of the amplitude of the modes with frequency, are well reproduced. In the low temperature range, the dynamics of the protein appears to occur in a single energy well of its highly multidimensional energy landscape. Of course this is no longer true when the temperature increases. By analyzing the population of the inherent structures, we have shown that, in the temperature range of the dynamical transition, there is a continuous increase of the number of states which are visited by the protein. The transition seems to be continuous, and it is likely that numerical observations suggesting a sudden increase may have the origin in the limited statistics due to finite time observation. Indeed, as shown for the example of protein G, the conformational transitions become extremely slow at low temperatures, such that the waiting time between the jumps between conformations may exceed the numerical (or the experimental) observation time. It is this breakdown of the ergodic hypothesis together with the observation of non-exponential relaxation rates which may have lead to the emergence of the terminology "protein glass transition" in analogy to the phenomenology of glasses. Non equilibrium studies allowed us to systematically probe a possible glassy behavior by searching for violations of the fluctuation--dissipation theorem. First we have shown that these calculations must be carried out with care because apparent violations are possible, even when the system is in equilibrium, due to nonlinearities in the response. Except at very low temperatures they can be observed even for perturbations as low as 1% of the potential energy of the system. Once this artifact is eliminated by the choice of a sufficiently small perturbative potential, we have shown that, after a sudden quench from an unfolded state to a very low temperature \(T_q\), *one does observe* a violation of the FDT in the protein model, analogous to what is found in glasses. The quenched protein is characterized by an effective temperature \(T_{\mathrm{eff}} > T_q\). But *the relaxation of the model towards equilibrium*, deduced from the evolution of the effective temperature \(T_{\mathrm{eff}}\) as a function of the waiting time after the quench, *follows a standard Arrhenius behavior*, even when the temperature crosses the value \(T \approx 0.45 T_f\) at which dynamical simulations appeared to show a change in the amplitude of the fluctuations. Although one cannot formally exclude that the results could be different for other protein structures or other simplified protein models, this work concludes that a coarse-grain model such as the Gō model is too simple to describe the complex behavior of protein G and particularly its glass transition. Indeed such a model does not include a real solvent, which plays an important role in experiments. The thermostat used in the molecular dynamics simulations only partially models the effect of the surrounding of the protein. The apparent numerical transition previously observed for protein G may simply be related to finite-time observations of the activation of structural transitions which appears in a particularly long time scale for proteins. This is an obvious limitation of molecular dynamics calculations, but this could also sound as a warning to experimentalists. Indeed experiments can access much longer time scales. But they also deal with real systems which are much more complex than the Gō model. In these systems relaxations may become very long, so that the experimental observation of a transition could actually face the same limitations as the numerical simulations. Such a "time window" interpretation has also been brought forward for the experimentally observed dynamical transition, suggesting that the transition may in fact depend on the energy-, and thus, on the time-resolution of the spectrometer. In this respect, as shown by our non-equilibrium studies to test the validity of the FDT, such measurements, if they could be performed for a protein should tell us a lot about the true nature of the "glass transition" of proteins.

{'timestamp': '2014-01-24T02:07:28', 'yymm': '1401', 'arxiv_id': '1401.5954', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5954'}

# Introduction Secret sharing schemes were proposed independently by Shamir and Blakley. In a \((k,n)\)-threshold secret sharing (\((k,n)\)-SS for short) scheme (e.g. see ), a dealer shares a secret among all participants, and then, \(k\) participants can reconstruct the secret while any \(k-1\) participants obtain no information on the secret. Since Shamir and Blakley proposed secret sharing schemes, various research on them have been reported. On the other hand, "time" is intimately related to our lives. We get up, eat something, do a job, and get asleep at a time of our (or someone's) choice. From the above reason, it appears that cryptographic protocols associated with "time" are useful and meaningful. Actually, as those protocols, *timed-release cryptographic protocols* introduced in are well-known. From the above discussion, it is useful and important to consider a secret sharing scheme with timed-release security. Therefore, we study such a scheme, which we call a *timed-release secret sharing* (TR-SS) scheme, in this paper. **Timed-Release Security.** Informally, the goal of timed-release cryptography is *to securely send a certain information into the future*. For instance, in timed-release encryption, a sender transmits a ciphertext so that a receiver can decrypt it when the time which the sender specified has come, and the receiver cannot decrypt it before the time. The timed-release cryptography was first proposed by May in 1993, and after that, Rivest et al. developed it in a systematic and formal way. Since Rivest et al. gave a formal definition of timed-release encryption (TRE) in, various research on timed-release cryptography including timed-release signatures (e.g., ) and timed-release encryption have been done based on computational security. In particular, TRE in the public-key setting has been recently researched on intensively (e.g., ). Recently, information-theoretically (or unconditionally) secure timed-release cryptography was proposed by Watanabe et al.. In addition, they investigated not only an encryption but also a key-agreement and an authentication code with information-theoretic timed-release security. To the best of our knowledge, however, there is no paper which reports on the study of secret sharing schemes with (information-theoretic) timed-release security. **Our Contribution.** In adding timed-release functionality to secret sharing schemes, we conceive the following two types of schemes. One is a secret sharing scheme such that information associated with time (called *time-signals*) is required whenever a secret is reconstructed, which means a secret sharing scheme with a simple combination of traditional secret sharing functionality and timed-release functionality. For realizing it, we propose \((k,n)\)-TR-SS in this paper. In \((k,n)\)-TR-SS, a dealer can specify positive integers \(k, n\) with \(k \le n\), where \(n\) is the number of participants and \(k\) is a threshold value, and future time when a secret can be recovered; and the secret can be reconstructed from at least \(k\) shares and a time-signal at the specified time. On the other hand, participants cannot reconstruct the secret without the time-signal even if they can obtain all shares. Specifically, we define a model and security notions of \((k,n)\)-TR-SS, and we derive lower bounds on the sizes of shares, time-signals, and entities' secret-keys required for \((k,n)\)-TR-SS. Moreover, we provide a direct construction of \((k,n)\)-TR-SS, which is constructed by using polynomials over finite fields and provably secure in our security definition. In addition, we show that the direct construction meets the lower bounds on the sizes of shares, time-signals, and entities' secret-keys with equalities. Therefore, it turns out that our lower bounds are tight, and that the direct construction is optimal. Another one is a *hybrid* TR-SS, which means a secret sharing scheme in which traditional secret sharing functionality and timed-release functionality are simultaneously realized. In our hybrid TR-SS, a secret can be reconstructed, if one of the following condition is satisfied: a secret can be reconstructed from \(k_1\) shares and a time-signal at a specified time as in the \((k_1,n)\)-TR-SS; or a secret can be reconstructed from \(k_2\) shares as in the traditional \((k_2,n)\)-SS. Hence, we consider two threshold values \(k_1,k_2\) to define a model of the hybrid TR-SS, and we propose \((k_1,k_2,n)\)-TR-SS as such a model, where \(k_1\le k_2 \le n\). Specifically, in \((k_1,k_2,n)\)-TR-SS, a dealer can specify future time, and arbitrarily chooses \(k_1\), \(k_2\) and \(n\). At least \(k_1\) (and less than \(k_2\)) participants can reconstruct a secret with a time-signal at the specified time, and at least \(k_2\) participants can reconstruct a secret *without* any time-signal (i.e. they can reconstruct from *only* their shares). Specifically, we define a model and security notions of \((k_1,k_2,n)\)-TR-SS, and we derive *tight* lower bounds on the sizes of shares, time-signals, and entities' secret-keys required for \((k_1,k_2,n)\)-TR-SS. Moreover, we provide two direct constructions of \((k_1,k_2,n)\)-TR-SS: One is a naive construction, which is very simple, however, does not meet the above lower bounds with equalities; The other is an *optimal* construction, which meets the above lower bounds with equalities. In particular, a theoretically-interesting point in our results includes that the timed-release security can be realized without any additional redundancy on the share-size in both schemes. **Applications of TR-SS.** Our TR-SS is a secret sharing scheme with timed-release property. We consider one of particular applications of TR-SS. Recently, in a real world setting, secret sharing schemes have been considered as applications, especially for cloud computing (e.g., secure data storage services). As represented by big data, information sharing via cloud computing has been developing over recent years. By applying TR-SS, an information provider can specify arbitrary time when the information is shared. Actually, the following case is known: Some companies share their big data, which usually includes sensitive data, and that each company uses shared data for its own business. Then, by using TR-SS, each company can specify future time when other companies can use such sensitive information. Therefore, we can say that TR-SS can provide more *flexible* security than traditional secret sharing schemes. Furthermore, TR-SS can also provide cryptographic protocols with timed-release functionality. For example, we can construct information-theoretically secure TRE in the two-user setting from \((1,1)\)-TR-SS and the one-time pad as follows. For a plaintext \(M\) and a shared key \(K\), a sender chooses a random number \(r\) whose length is equal to the plaintext-length, and computes a cipertext \(C:=M\oplus r \oplus K\). Then, the sender specifies future time, and he generates one share from the secret \(r\) by \((1,1)\)-TR-SS. A receiver can compute \(C\oplus K=M\oplus r\) by using the shared key \(K\) in advance, however, he cannot obtain \(M\) until the specified time comes since he can get \(r\) only after the specified time. In a similar way, it is expected that TR-SS is useful for building other timed-release cryptographic protocols such as timed-release authentication code in the two-user setting, and that TR-SS might be able to provide some new timed-release cryptographic protocols, e.g., timed-release threshold encryption. **Organization of this paper.** The rest of this paper is organized as follows. In Sections [2](#Sec_TR-SS1){reference-type="ref" reference="Sec_TR-SS1"} and [3](#Sec_TR-SS2){reference-type="ref" reference="Sec_TR-SS2"}, we describe \((k,n)\)-TR-SS and \((k_1,k_2,n)\)-TR-SS, respectively, which are based on the ideas according to. Specifically, in each section, we define a model and security of each scheme, and derive lower bounds on the sizes of shares, time-signals and secret-keys required for each scheme, respectively. Furthermore, we propose a direct construction of each scheme, and show it is provably secure and optimal. Finally, in Section [4](#Conclusion){reference-type="ref" reference="Conclusion"}, we give concluding remarks of this paper. **Notation.** Throughout this paper, we use the following notation. Generally speaking, \(X\) indicates a random variable which takes values in \(\mathcal{X}\) (e.g., \(A, B,\) and \(C\) are random variables which take values in \(\mathcal{A}, \mathcal{B},\) and \(\mathcal{C}\), respectively). For any finite set \(\mathcal{Z}\) and arbitrary non-negative integers \(z_1,z_2\), let \(\mathcal{PS}(\mathcal{Z}, z_1, z_2) := \{ Z \subset \mathcal{Z} | z_1 \le |Z| \le z_2 \}\) be the family of all subsets of \(\mathcal{Z}\) whose cardinality is at least \(z_1\) but no more than \(z_2\). # \((k,n)\)-Timed-Release Secret Sharing Scheme {#Sec_TR-SS1} In this section, we propose a model and a security definition of \((k,n)\)-TR-SS. In \((k,n)\)-TR-SS, a time-signal at the specified time is always required when a secret is reconstructed. In other words, a secret cannot be reconstructed without a time-signal at the specified time even if there are all shares. ## The Model and Security Definition {#Sec_Model1} First, we introduce the model of \((k, n)\)-TR-SS. Unlike traditional secret sharing schemes, we assume that there is a trusted authority (or a trusted initializer) \(TA\) whose role is to generate and to distribute secret-keys of entities. We call this model the *trusted initializer model* as in In \((k, n)\)-TR-SS, there are \(n + 3\) entities, a dealer \(D\), \(n\) participants \(P_1, P_2, \ldots, P_n\), a time-server \(TS\) for broadcasting time-signals at most \(\tau\) times and a trusted initializer \(TA\), where \(k\), \(n\) and \(\tau\) are positive integers. In this paper, we assume that the identity of each user \(P_i\) is also denoted by \(P_i\). Informally, \((k, n)\)-TR-SS is executed as follows. First, \(TA\) generates secret-keys on behalf of \(D\) and \(TS\). After distributing these keys via secure channels, \(TA\) deletes it in his memory. Next, \(D\) specifies future time, as \(D\) wants, when a secret is reconstructed by participants, and he generates \(n\) shares from the secret by using his secret-key. And, \(D\) sends each share to each participant respectively via secure channels. The time-server \(TS\) periodically broadcasts a time-signal which is generated by using his secret-key. When the specified time has come, at least \(k\) participants can compute the secret by using their shares and the time-signal of the specified time. Formally, we give the definition of \((k, n)\)-TR-SS as follows. In this model, let \(\mathcal{P}:=\{P_1, P_2, \ldots, P_n\}\) be a set of all participants. And also, \(\mathcal{S}\) is a set of possible secrets with a probability distribution \(P_S\), and \(\mathcal{SK}\) is a set of possible secret-keys. \(\mathcal{T} := \{ 1, 2, \ldots, \tau \}\) is a set of time. Let \(\mathcal{U}_{i}^{(t)}\) be the set of possible \(P_i\)'s shares at the time \(t\in {\cal T}\). Also, \(\mathcal{U}_i:=\bigcup_{t=1}^{\tau}\mathcal{U}_i^{(t)}\) is a set of possible \(P_i\)'s shares for every \(i\in \{1,2,\ldots,n \}\), and let \(\mathcal{U} := \bigcup^{n}_{i=1}\mathcal{U}_i\). In addition, \(\mathcal{TI}^{(t)}\) is a set of time-signals at time \(t\), and let \(\mathcal{TI}:=\bigcup_{t=1}^{\tau}\mathcal{TI}^{(t)}\). Furthermore, for any subset of participants \(\mathcal{J} = \{P_{i_1},\ldots,P_{i_j}\} \subset \mathcal{P}\), \(\mathcal{U}_{\mathcal{J}}^{(t)} := \mathcal{U}_{i_1}^{(t)} \times \cdots \times \mathcal{U}_{i_j}^{(t)}\) denotes the set of possible shares held by \(\mathcal{J}\). In the above model, we assume that \(\Pi\) meets the following *correctness* property: If \(D\) correctly completes the phase *Share* and \(TS\) correctly completes the phase *Extract*, then, for all possible \(i \in \{1,2,\ldots,n\}\), \(t \in \mathcal{T}\), \(s \in \mathcal{S}\), \(u_{i}^{(t)} \in \mathcal{U}_{i}\), and \(mk^{(t)} \in \mathcal{TI}^{(t)}\), it hold that any \(\mathcal{A} \in \mathcal{PS}(\mathcal{P}, k, n)\) will correctly reconstruct the secret \(s\) at the end of phase *Reconstruct*, namely, \[\begin{aligned} H(S \mid U_{\mathcal{A}}^{(t)}, TI^{(t)}) = 0. \end{aligned}\] Next, we formalize a security definition of \((k, n)\)-TR-SS based on the idea of the information-theoretic timed-release security and secret sharing schemes (e.g. see ). In \((k, n)\)-TR-SS, we consider the following two kinds of security. The first security which we consider is basically the same as that of the traditional \((k,n)\)-SS: less than \(k\) participants cannot obtain any information on a secret. In addition to this, as the second security we want to require that even at least \(k\) participants cannot obtain any information on a secret before the specified time comes (i.e., before a time-signal at the specified time is received), since we consider timed-release security in this paper. Therefore, we formally define secure \((k, n)\)-TR-SS as follows. Intuitively, the meaning of two conditions (i) and (ii) in Definition [\[Security1\]](#Security1){reference-type="ref" reference="Security1"} is explained as follows. (i) No information on a secret is obtained by any set of less than \(k\) participants, even if they obtain time-signals at all the time; (ii) No information on a secret is obtained by any set of more than \(k-1\) participants, even if they obtain time-signals at all the time except the specified time.[^3] Note that the condition (iii) is stronger than (i). However, we do not consider (iii) in this paper because of the following two reasons: first, the condition (i) is more natural than (iii), since it does not seem natural to consider the situation that any set of less than \(k\) participants colludes with \(TS\) in the real world; and secondly, our lower bounds in Theorem [\[Bounds\]](#Bounds){reference-type="ref" reference="Bounds"} are still valid even under the conditions (ii) and (iii), in other words, even if we consider the conditions (ii) and (iii), we can derive the same lower bounds in Theorem [\[Bounds1\]](#Bounds1){reference-type="ref" reference="Bounds1"} since Definition [\[Security1\]](#Security1){reference-type="ref" reference="Security1"} is weaker. Interestingly, our direct construction in Section [2.3](#Direct){reference-type="ref" reference="Direct"} also satisfies (iii), and *tightness* of our lower bounds and *optimality* of our direct construction will be valid not depending on the choice of the condition (i) or (iii). Furthermore, we do not have to consider an attack by dishonest \(TS\) only, since \(TS\)'s master-key is generated independently of a secret. The proof follows from the following lemmas. *Proof.* The proof of this lemma can be proved in a way similar to the proof in. For arbitrary \(i\in \{1,2,\dots,n\}\), we take a subset \(\mathcal{B}_i \in \mathcal{PS}(\mathcal{P}\setminus\{P_i\},k-1,k-1)\) of participants. Then, for any \(t \in \mathcal{T}\), we have where ([\[LemmaU1_01\]](#LemmaU1_01){reference-type="ref" reference="LemmaU1_01"}) follows from the correctness of \((k,n)\)-TR-SS and ([\[LemmaU1_02\]](#LemmaU1_02){reference-type="ref" reference="LemmaU1_02"}) follows from the condition (i) in Definition [\[Security1\]](#Security1){reference-type="ref" reference="Security1"}. *Proof.* For any \(\mathcal{A} \in \mathcal{PS}(\mathcal{P},k,n)\) and any \(t \in \mathcal{T}\), we have \[\begin{aligned} H(TI^{(t)}) \ge&H(TI^{(t)} \mid TI^{(1)}, \dots, TI^{(t-1)}) \nonumber \\ \ge& H(TI^{(t)} \mid U_{\mathcal{A}}^{(t)}, TI^{(1)}, \dots, TI^{(t-1)}) \nonumber \\ \ge& I(S;TI^{(t)} \mid U_{\mathcal{A}}^{(t)}, TI^{(1)}, \dots, TI^{(t-1)}) \nonumber\\ =& H(S \mid U_{\mathcal{A}}^{(t)}, TI^{(1)}, \dots, TI^{(t-1)}) \label{LemmaTI1_01}\\ =& H(S), \label{LemmaTI1_02} \end{aligned}\] where ([\[LemmaTI1_01\]](#LemmaTI1_01){reference-type="ref" reference="LemmaTI1_01"}) follows from the correctness of \((k,n)\)-TR-SS and ([\[LemmaTI1_02\]](#LemmaTI1_02){reference-type="ref" reference="LemmaTI1_02"}) follows from the condition (ii) in Definition [\[Security1\]](#Security1){reference-type="ref" reference="Security1"}. *Proof.* We have where the last inequality follows from Lemma [\[LemmaTI1\]](#LemmaTI1){reference-type="ref" reference="LemmaTI1"}. *Proof of Theorem [\[Bounds1\]](#Bounds1){reference-type="ref" reference="Bounds1"}:* From Lemmas [\[LemmaU1\]](#LemmaU1){reference-type="ref" reference="LemmaU1"}-[\[LemmaSK1\]](#LemmaSK1){reference-type="ref" reference="LemmaSK1"}, the proof of Theorem [\[Bounds1\]](#Bounds1){reference-type="ref" reference="Bounds1"} is completed. As we will see in Section [2.3](#Direct){reference-type="ref" reference="Direct"}, the above lower bounds are tight since our construction will meet all the above lower bounds with equalities. Therefore, we define optimality of constructions of \((k,n)\)-TR-SS as follows. ## Direct Construction {#Direct} We propose a direct construction of \((k, n)\)-TR-SS. In addition, it is shown that our construction is optimal. The detail of our construction of \((k, n)\)-TR-SS \(\Pi\) is given as follows. 1. *Initialize.* Let \(q\) be a prime power, where \(q > \max(n, \tau)\), and \(\mathbb{F}_q\) be the finite field with \(q\) elements. We assume that the identity of each participant \(P_i\) is encoded as \(P_i \in \mathbb{F}_q \backslash \{0\}\). Also, we assume \(\mathcal{T} = \{1, 2, \dots, \tau\} \subset \mathbb{F}_q \backslash \{ 0 \}\) by using appropriate encoding. First, \(TA\) chooses uniformly at random \(\tau\) distinct numbers \(r^{(j)} (1 \le j \le \tau)\) from \(\mathbb{F}_q\). \(TA\) sends a secret-key \(sk:=(r^{(1)},\ldots,r^{(\tau)})\) to \(TS\) and \(D\) via secure channels, respectively. 2. *Share.* First, \(D\) chooses a secret \(s \in \mathbb{F}_q\). Also, \(D\) specifies the time \(t\) at which participants can reconstruct the secret. Next, \(D\) randomly chooses a polynomial \(f(x) := c^{(t)} + \sum^{k-1}_{i=1}\) \(a_{i}x^i\) over \(\mathbb{F}_q\), where \(c^{(t)}\) is computed by \(c^{(t)}:=s+r^{(t)}\) and each coefficient \(a_i\) is randomly and uniformly chosen from \(\mathbb{F}_q\). Finally, \(D\) computes \(u_i^{(t)} := f(P_i) (i = 1,2,\ldots,n)\) and sends \((u_i^{(t)}, t)\) to \(P_i (i=1, 2, \ldots, n)\) via a secure channel. 3. *Extract.* For \(sk\) and time \(t \in \mathcal{T}\), \(TS\) broadcasts \(t\)-th key \(r^{(t)}\) as a time-signal at time \(t\) to all participants via a (authenticated) broadcast channel. 4. *Reconstruct.* First, a set of at least \(k\) participants \(\mathcal{A}=\{P_{i_1},P_{i_2},\ldots,P_{i_k}\} \in \mathcal{PS}(\mathcal{P}, k,k)\) computes \(c^{(t)}\) by Lagrange interpolation: \[\begin{aligned} c^{(t)}=\sum_{j=1}^{k}(\prod_{l\neq j}\frac{P_{i_j}}{P_{i_j}-P_{i_l}})f(P_{i_j}), \end{aligned}\] from their \(k\) shares. After receiving \(ts^{(t)}=r^{(t)}\), they can compute and get \(s =c^{(t)}-r^{(t)}\). The security and optimality of the above construction is stated as follows. *Proof.* First, we show the proof of (i) in Definition [\[Security1\]](#Security1){reference-type="ref" reference="Security1"}. Assume that any \(k-1\) participants \(\mathcal{F}=\{P_{i_1},\ldots,P_{i_{k-1}}\}\in\mathcal{PS}(\mathcal{P},k-1,k-1)\) try to guess \(c^{(t)}\) by using their shares. Note that they know \(r^{(t)} = c^{(t)}-s\) and \[\begin{aligned} f(P_{i_j})=(1,P_{i_j},\ldots,P_{i_j}^{k-1}) \left( \begin{array}{c} c^{(t)} \\ a_1 \\ \vdots \\ a_{k-1}\\ \end{array} \right), \end{aligned}\] for \(j=1,\ldots,k-1\). Thus, they can know the following matrix: \[\begin{aligned} \left( \begin{array}{cccc} 1 & P_{i_1} & \cdots & P_{i_1}^{k-1} \\ 1 & P_{i_2} & \cdots & P_{i_2}^{k-1} \\ \vdots & \vdots & \ddots & \vdots \\ 1 & P_{i_{k-1}} & \cdots & P_{i_{k-1}}^{k-1}\\ \end{array} \right) \left( \begin{array}{c} c^{(t)} \\ a_1 \\ \vdots \\ a_{k-1}\\ \end{array} \right). \label{direct01} \end{aligned}\] However, from ([\[direct01\]](#direct01){reference-type="ref" reference="direct01"}), they cannot guess at least one element of \((c^{(t)}, a_1, \ldots,a_{k-1})\) with probability larger than \(1/q\). Therefore, \(H(S \mid U_{\mathcal{F}}, TI^{(1)},\ldots,TI^{(\tau)}) = H(S)\) for any \(\mathcal{F}\in\mathcal{PS}(\mathcal{P},1,k-1)\) and any \(t\in\mathcal{T}\). Next, we show the proof of (ii) in Definition [\[Security1\]](#Security1){reference-type="ref" reference="Security1"}. Without loss of generality, we suppose that \(\tau\) is a specified time, and that all participants try to guess \(r^{(\tau)}\) by using \(c^{(\tau)}\) and time-signals at all the time except the time \(\tau\), since they obtain \(c^{(\tau)}=s+r^{(\tau)}\) from their shares. They get \(\tau-1\) time-signals \(r^{(1)},\ldots,r^{(\tau-1)}\). However, since each time-signal is chosen uniformly at random from \(\mathbb{F}_q\), they can guess \(r^{(\tau)}\) only with probability \(1/q\). By the security of one-time pad, we have \(H(S \mid U_{1},\ldots,U_{n}, TI^{(1)},\ldots,T^{(\tau-1)}) = H(S)\). Hence, for any \(\mathcal{A} \in \mathcal{PS}(\mathcal{P}, k,n)\) and for any \(t \in \mathcal{T}\), we have \(H(S \mid U_{\mathcal{A}}^{(t)}, TI^{(1)},\ldots,TI^{(t-1)},TI^{(t+1)},\ldots,T^{(\tau)}) = H(S)\). Finally, it is straightforward to see that the construction satisfies all the equalities of lower bounds in Theorem [\[Bounds1\]](#Bounds1){reference-type="ref" reference="Bounds1"}. Therefore, the above construction is optimal. # \((k_1,k_2,n)\)-Timed-Release Secret Sharing Scheme {#Sec_TR-SS2} We propose \((k_1,k_2,n)\)-TR-SS, where \(k_1\) and \(k_2\) are threshold values with \(1 \le k_1\le k_2 \le n\). \((k_1,k_2,n)\)-TR-SS can realize timed-release functionality---a secret can be reconstructed from at least \(k_1\) shares and a time-signal at the specified time---and traditional secret sharing functionality---a secret can be also reconstructed from only at least \(k_2\) shares---simultaneously. In the case that \(k=k_1=k_2\), \((k,k,n)\)-TR-SS can be considered as traditional \((k,n)\)-SS (for details, see Remark [\[Adequacy\]](#Adequacy){reference-type="ref" reference="Adequacy"}). ## Model and Security Definition {#Model} In this section, we propose a model and a security definition of \((k_1,k_2,n)\)-TR-SS. First, we introduce a model of \((k_1,k_2,n)\)-TR-SS. In \((k_1,k_2,n)\)-TR-SS, there are same entities and sets as those of \((k,n)\)-TR-SS. The main difference from \((k,n)\)-TR-SS is that a dealer \(D\) can specify two kinds of threshold values, \(k_1\) and \(k_2\) with \(k_1 \le k_2 \le n\): \(k_1\) indicates the number of participants who can reconstruct a secret \(s\) with the time-signal at the time specified by the dealer; and \(k_2\) indicates the number of participants who can reconstruct \(s\) without any time-signals. We give the definition of \((k_1,k_2,n)\)-TR-SS as follows. In the above model, we assume that \(\Theta\) meets the following *correctness* properties: 1. If \(D\) correctly completes the phase *Share* and \(TS\) correctly completes the phase *Extract*, then, for all possible \(i \in \{1,2,\ldots,n\}\), \(t \in \mathcal{T}\), \(s \in \mathcal{S}\), \(u_{i}^{(t)} \in \mathcal{U}_{i}^{(t)}\), and \(ts^{(t)} \in \mathcal{TI}^{(t)}\), it holds that any \(\mathcal{A} \in \mathcal{PS}(\mathcal{P},k_1,k_2-1)\) will correctly reconstruct the secret \(s\) at the end of phase *Reconstruct with time-signals*, namely, \[\begin{aligned} H(S \mid U_{\mathcal{A}}^{(t)}, TI^{(t)}) = 0. \end{aligned}\] 2. If \(D\) correctly completes the phase *Share*, then, for all possible \(i \in \{1,2,\ldots,n\}\), \(t \in \mathcal{T}\), \(s \in \mathcal{S}\), and \(u_{i}^{(t)} \in \mathcal{U}_{i}^{(t)}\), it holds that any \(\hat{\mathcal{A}} \in \mathcal{PS}(\mathcal{P},k_2,n)\) will correctly reconstruct the secret \(s\) at the end of phase *Reconstruct without time-signals*, namely, \[\begin{aligned} H(S \mid U_{\hat{\mathcal{A}}}^{(t)}) = 0. \end{aligned}\] Next, we formalize a security definition of \((k_1,k_2,n)\)-TR-SS in a similar way to that of \((k,n)\)-TR-SS as follows. In Definition [\[Security2\]](#Security2){reference-type="ref" reference="Security2"}, intuitively, the meaning of (i) is the same as that of \((k,n)\)-TR-SS (Definition [\[Security1\]](#Security1){reference-type="ref" reference="Security1"}), and the meaning of the condition (ii) is explained that no information on a secret is obtained by any set of at least \(k_1\) but *no more than \(k_2\)* participants, even if they obtain time-signals at all the time except the specified time. We also consider a more strong security notion in a similar to \((k,n)\)-TR-SS, however, we do not consider such a strong notion for the same reason as in the case of \((k,n)\)-TR-SS. ## Lower Bounds {#Sec_Bounds2} In this section, we show lower bounds on sizes of shares, time-signals, and secret-keys required for secure \((k_1,k_2,n)\)-TR-SS as follows. The proof follows from the following lemmas. *Proof.* The proof of this lemma can be proved in a way similar to the proof in. For arbitrary \(i\in \{1,2,\dots,n\}\), we take a subset \(\mathcal{B}_i \in \mathcal{PS}(\mathcal{P}\setminus\{P_i\},k_2-1,k_2-1)\) of participants. Then, for any \(t \in \mathcal{T}\), we have where ([\[LemmaU2_01\]](#LemmaU2_01){reference-type="ref" reference="LemmaU2_01"}) follows from the correctness of \((k_1,k_2,n)\)-TR-SS and ([\[LemmaU2_02\]](#LemmaU2_02){reference-type="ref" reference="LemmaU2_02"}) follows from the condition (ii) in Definition [\[Security2\]](#Security2){reference-type="ref" reference="Security2"}. *Proof.* The statement is true in the case that \(k_1=k_2\), since Shannon entropy is non-negative. Therefore, in the following, we assume \(k_1<k_2\). For arbitrary \(i\in \{1,2,\dots,n\}\), we take a subset \(\mathcal{B}_i \in \mathcal{PS}(\mathcal{P}\setminus\{P_i\},k_2-1,k_2-1)\) of participants. For any \(t \in \mathcal{T}\), we have \[\begin{aligned} &H(TI^{(t)}) \nonumber \\ \ge&H(TI^{(t)} \mid TI^{(1)}, \dots, TI^{(t-1)}) \nonumber \\ \ge&I(TI^{(t)};U_1^{(t)},U_2^{(t)},\ldots,U_n^{(t)}\mid TI^{(1)},\ldots,TI^{(t-1)}) \nonumber \\ =&H(U_1^{(t)},U_2^{(t)},\ldots,U_n^{(t)}\mid TI^{(1)},\ldots,TI^{(t-1)}) \nonumber \\ &\qquad-H(U_1^{(t)},U_2^{(t)},\ldots,U_n^{(t)}\mid TI^{(1)},\ldots,TI^{(t)}) \nonumber \\ =&H(U_1^{(t)},\ldots,U_{k_1}^{(t)}\mid TI^{(1)},\ldots,TI^{(t-1)}) \nonumber \\ &+H(U_{k_1+1}^{(t)},\ldots,U_{k_2}^{(t)}\mid TI^{(1)},\ldots,TI^{(t-1)},U_1^{(t)},\ldots,U_{k_1}^{(t)}) \nonumber \\ &\quad +H(U_{k_2+1}^{(t)},\ldots,U_n^{(t)}\mid TI^{(1)},\ldots,TI^{(t-1)},U_1^{(t)},\ldots,U_{k_2}^{(t)}) \nonumber \\ &-H(U_1^{(t)},\ldots,U_{k_1}^{(t)}\mid TI^{(1)},\ldots,TI^{(t)}) \nonumber \\ &\quad-H(U_{k_1+1}^{(t)},\ldots,U_{k_2}^{(t)}\mid TI^{(1)},\ldots,TI^{(t)},U_1^{(t)},\ldots,U_{k_1}^{(t)}) \nonumber \\ &\qquad-H(U_{k_2+1}^{(t)},\ldots,U_n^{(t)}\mid TI^{(1)},\ldots,TI^{(t)},U_1^{(t)},\ldots,U_{k_2}^{(t)}) \nonumber \\ \ge&H(U_1^{(t)},\ldots,U_{k_1}^{(t)}\mid TI^{(1)},\ldots,TI^{(t)}) \nonumber \\ &+H(U_{k_1+1}^{(t)},\ldots,U_{k_2}^{(t)}\mid TI^{(1)},\ldots,TI^{(t-1)},U_1^{(t)},\ldots,U_{k_1}^{(t)}) \nonumber \\ &\quad +H(U_{k_2+1}^{(t)},\ldots,U_n^{(t)}\mid TI^{(1)},\ldots,TI^{(t)},U_1^{(t)},\ldots,U_{k_2}^{(t)}) \nonumber \\ &-H(U_1^{(t)},\ldots,U_{k_1}^{(t)}\mid TI^{(1)},\ldots,TI^{(t)}) \nonumber \\ &\quad-H(U_{k_1+1}^{(t)},\ldots,U_{k_2}^{(t)}\mid TI^{(1)},\ldots,TI^{(t)},U_1^{(t)},\ldots,U_{k_1}^{(t)}) \nonumber \\ &\qquad-H(U_{k_2+1}^{(t)},\ldots,U_n^{(t)}\mid TI^{(1)},\ldots,TI^{(t)},U_1^{(t)},\ldots,U_{k_2}^{(t)}) \nonumber \\ =&H(U_{k_1+1}^{(t)},\ldots,U_{k_2}^{(t)}\mid TI^{(1)},\ldots,TI^{(t-1)},U_1^{(t)},\ldots,U_{k_1}^{(t)}) \nonumber \\ &\quad-H(U_{k_1+1}^{(t)},\ldots,U_{k_2}^{(t)}\mid TI^{(1)},\ldots,TI^{(t)},U_1^{(t)},\ldots,U_{k_1}^{(t)}) \nonumber \\ \ge&\sum_{i=k_1+1}^{k_2}H(U_{i}^{(t)}\mid TI^{(1)},\ldots,TI^{(t-1)},U_{\mathcal{B}_i}^{(t)}) \nonumber \\ &-\sum_{i=k_1+1}^{k_2}H(U_{i}^{(t)}\mid TI^{(1)},\ldots,TI^{(t)},U_1^{(t)},\ldots,U_{i-1}^{(t)}) \nonumber \\ =&(k_2-k_1)H(S), \label{LemmaTI2_01} \end{aligned}\] where ([\[LemmaTI2_01\]](#LemmaTI2_01){reference-type="ref" reference="LemmaTI2_01"}) follows from ([\[LemmaU2_00\]](#LemmaU2_00){reference-type="ref" reference="LemmaU2_00"}) in the proof of Lemma [\[LemmaU2\]](#LemmaU2){reference-type="ref" reference="LemmaU2"}, the assumption of \(H(U_i^{(t)})=H(S)\), and the following claim. *Proof.* First, for arbitrary \(i\in \{1,2,\ldots,n\}\), we take subsets \(\mathcal{B}_i:=\mathcal{PS}(\mathcal{P}\setminus\{P_i\},k_1-1,k_1-1)\) and \(\mathcal{A}_i:=\mathcal{PS}(\mathcal{P}\setminus\{P_i\},k_1,k_2-1)\) of participants such that \(\mathcal{B}_i \subset \mathcal{A}_i\). Then, for any \(t\in\mathcal{T}\), we have \[\begin{aligned} &H(U_i^{(t)}) \nonumber \\ \ge&H(U_i^{(t)}\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)}) \nonumber \\ \ge&H(U_i^{(t)}\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)})-H(U_i^{(t)}\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)},S) \label{ClaimTI_00} \\ =&I(U_i^{(t)};S\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)}) \nonumber \\ =&H(S\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)})-H(S\mid U_{\mathcal{B}_i}^{(t)},U_i^{(t)},TI^{(t)}) \nonumber \\ =&H(S\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)}) \label{ClaimTI_01} \\ =&H(S), \label{ClaimTI_02} \end{aligned}\] where ([\[ClaimTI_01\]](#ClaimTI_01){reference-type="ref" reference="ClaimTI_01"}) follows form the correctness of \((k_1,k_2,n)\)-TR-SS and ([\[ClaimTI_02\]](#ClaimTI_02){reference-type="ref" reference="ClaimTI_02"}) follows from the condition (i) in Definition [\[Security2\]](#Security2){reference-type="ref" reference="Security2"}. From ([\[ClaimTI_00\]](#ClaimTI_00){reference-type="ref" reference="ClaimTI_00"}) and the assumption of \(H(U_i^{(t)})=H(S)\), we have \[\begin{aligned} &H(U_i^{(t)}\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)}) \\ =& H(U_i^{(t)}\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)})-H(U_i^{(t)}\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)},S). \end{aligned}\] Therefore, we have \[\begin{aligned} H(U_i^{(t)}\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)},S)=0. \end{aligned}\] Hence, we have \[\begin{aligned} H(U_i^{(t)}\mid U_{\mathcal{A}_i}^{(t)},TI^{(t)}) =&H(U_i^{(t)}\mid U_{\mathcal{A}_i}^{(t)},TI^{(t)},S)\\ \le& H(U_i^{(t)}\mid U_{\mathcal{B}_i}^{(t)},TI^{(t)},S)\\ =&0. \end{aligned}\] Since \(H(U_i^{(t)}\mid U_{\mathcal{A}_i}^{(t)},TI^{(t)})\ge 0\), we have \(H(U_i^{(t)}\mid U_{\mathcal{A}_i}^{(t)},TI^{(t)})=0\). *Proof of Lemma [\[LemmaTI2\]](#LemmaTI2){reference-type="ref" reference="LemmaTI2"}:* From the above claim, the proof of Lemma [\[LemmaTI2\]](#LemmaTI2){reference-type="ref" reference="LemmaTI2"} is completed. *Proof.* We can prove in a similar way to the proof of Lemma [\[LemmaSK1\]](#LemmaSK1){reference-type="ref" reference="LemmaSK1"}. We have where the last inequality follows from Lemma [\[LemmaTI2\]](#LemmaTI2){reference-type="ref" reference="LemmaTI2"}. *Proof of Theorem [\[Bounds2\]](#Bounds2){reference-type="ref" reference="Bounds2"}:* From Lemmas [\[LemmaU2\]](#LemmaU2){reference-type="ref" reference="LemmaU2"}-[\[LemmaSK2\]](#LemmaSK2){reference-type="ref" reference="LemmaSK2"}, the proof of Theorem [\[Bounds2\]](#Bounds2){reference-type="ref" reference="Bounds2"} is completed. As we will see in Section [3.3.2](#Optimal){reference-type="ref" reference="Optimal"}, the lower bounds in Theorem [\[Bounds2\]](#Bounds2){reference-type="ref" reference="Bounds2"} are tight since our construction will meet all the above lower bounds with equalities. Therefore, we define optimality of constructions of \((k_1,k_2,n)\)-TR-SS as follows. ## Construction {#Construction} We propose a direct construction of \((k_1,k_2,n)\)-TR-SS. In addition, it is shown that our construction is optimal. Before that, we show a naive construction based on \((k_1,n)\)-TR-SS and \((k_2,n)\)-SS, which is not optimal. ### Naive Construction Our idea of a naive construction is a combination of \((k_1,n)\)-TR-SS (Section [2.3](#Direct){reference-type="ref" reference="Direct"}) and Shamir's \((k_2,n)\)-SS. 1. *Initialize.* Let \(q\) be a prime power, where \(q > \max(n, \tau)\), and \(\mathbb{F}_q\) be the finite field with \(q\) elements. We assume that the identity of each participant \(P_i\) is encoded as \(P_i \in \mathbb{F}_q \backslash \{0\}\). Also, we assume \(\mathcal{T} = \{1, 2, \dots, \tau\} \subset \mathbb{F}_q \backslash \{ 0 \}\) by using appropriate encoding. First, \(TA\) chooses uniformly at random \(\tau\) distinct numbers \(r^{(j)} (1 \le j \le \tau)\) from \(\mathbb{F}_q\). \(TA\) sends a secret-key \(sk:=(r^{(1)},\ldots,r^{(\tau)})\) to \(TS\) and \(D\) via secure channels, respectively. 2. *Share.* First, \(D\) chooses a secret \(s \in \mathbb{F}_q\). Also, \(D\) specifies the time \(t\) when at least \(k_1\) participants can reconstruct the secret and chooses \(t\)-th key \(r^{(t)}\). Next, \(D\) randomly chooses two polynomials \(f_1(x) := s + r^{(t)} + \sum^{{k_1}-1}_{i=1}a_{1i}x^i\) and \(f_2(x) := s + \sum^{{k_2}-1}_{i=1}a_{2i}x^i\) over \(\mathbb{F}_q\), where each coefficient is randomly and uniformly chosen from \(\mathbb{F}_q\). Then, \(D\) computes \(u_i^{(t)} := (f_1(P_i),f_2(P_i))\). Finally, \(D\) sends \((u_i^{(t)}, t)\) to \(P_i (i=1, 2, \ldots, n)\) via a secure channel. 3. *Extract.* For \(sk\) and time \(t \in \mathcal{T}\), \(TS\) broadcasts \(t\)-th key \(r^{(t)}\) as a time-signal at time \(t\) to all participants via a (authenticated) broadcast channel. 4. *Reconstruct with time-signals.* First, \(\mathcal{A}=\{P_{i_1},P_{i_2},\ldots,P_{i_{k_1}}\} \in \mathcal{PS}(\mathcal{P},k_1, k_1)\) computes \(s + r^{(t)}\) by Lagrange interpolation: \[\begin{aligned} s + r^{(t)}=\sum_{j=1}^{k_1}(\prod_{l\neq j}\frac{P_{i_j}}{P_{i_j}-P_{i_l}})f_1(P_{i_j}), \end{aligned}\] from \((f_1(P_{i_1}),\ldots,f_1(P_{i_{k_1}}))\). After receiving \(ts^{(t)}=r^{(t)}\), they can compute and get \(s =s+ r^{(t)}-ts^{(t)}\). 5. *Reconstruct without time-signals.* any \(\hat{\mathcal{A}}=\{P_{i_1},P_{i_2},\ldots,P_{i_{k_2}}\} \in \mathcal{PS}(\mathcal{P},k_2,k_2)\) computes \[\begin{aligned} s=\sum_{j=1}^{k_2}(\prod_{l\neq j}\frac{P_{i_j}}{P_{i_j}-P_{i_l}})f_2(P_{i_j}), \end{aligned}\] by Lagrange interpolation from \((f_2(P_{i_1}),\ldots,f_2(P_{i_{k_2}}))\). It is easy to see that the above construction is secure, since this construction is a simple combination of \((k_1,n)\)-TR-SS and Shamir's \((k_2,n)\)-SS. Also, the above construction is simple, however not optimal since the resulting share-size is twice as large as that of secrets. ### Optimal (but Restricted[^4]) Construction {#Optimal} To achieve an optimal construction, we use the technique as in : In the phase *Share*, the dealer computes public parameters, and the public parameters are broadcasted to participants or else stored on a publicly accessible authenticated bulletin board. The detail of our construction is given as follows. 1. *Initialize.* Let \(q\) be a prime power, where \(q > \max(n, \tau)\), and \(\mathbb{F}_q\) be the finite field with \(q\) elements. We assume that the identity of each participant \(P_i\) is encoded as \(P_i \in \mathbb{F}_q \backslash \{0\}\). Also, we assume \(\mathcal{T} = \{1, 2, \dots, \tau\} \subset \mathbb{F}_q \backslash \{ 0 \}\) by using appropriate encoding. First, \(TA\) chooses \(\ell\), which is the maximum difference between \(k_2\) and \(k_1\). Note that \(k_1\) and \(k_2\) will be determined by a dealer \(D\) in the phase *Share*. Then, \(TA\) chooses \(\tau\ell\) numbers \(r^{(t)}_{i} \ (1\le i \le \ell)\) and \((1\le t\le \tau)\) from \(\mathbb{F}_q\) uniformly at random. \(TA\) sends a secret-key \(sk:=\{(r^{(t)}_1,r^{(t)}_2,\ldots,r^{(t)}_{\ell})\}_{1\le t \le \tau}\) to \(TS\) and \(D\) via secure channels, respectively. 2. *Share.* First, \(D\) randomly selects a secret \(s \in \mathbb{F}_q\), and chooses \(k_1\), \(k_2\) and \(n\) such that \(k_2-k_1\le\ell\). Also, \(D\) specifies the time \(t\) when at least \(k_1\) participants can reconstruct the secret. Next, \(D\) randomly chooses a polynomial \(f(x) := s + \sum^{{k_2}-1}_{i=1}a_{i}x^i\) over \(\mathbb{F}_q\), where each coefficient \(a_i\) is randomly and uniformly chosen from \(\mathbb{F}_q\). Then, \(D\) computes a share \(u_i^{(t)} := f(P_i)\) and a public parameter \(p_i^{(t)}:=a_{k_1-1+i}+r_i^{(t)} \ (i = 1,2,\ldots,k_2-k_1)\). Finally, \(D\) sends \((u_i^{(t)}, t)\) to \(P_i (i=1, 2, \ldots, n)\) via a secure channel and discloses \((p_1^{(t)},\ldots,p_{k_2-k_1}^{(t)})\). 3. *Extract.* For \(sk\) and time \(t \in \mathcal{T}\), \(TS\) broadcasts a time-signal at time \(t\), \(ts^{(t)} :=(r_1^{(t)},r_2^{(t)},\ldots,r_{\ell}^{(t)})\) to all participants via a (authenticated) broadcast channel. 4. *Reconstruct with time-signals.* Suppose that all participants receive \(ts^{(t)}=(r_1^{(t)},r_2^{(t)},\ldots,r_{\ell}^{(t)})\). Let \(\mathcal{A}=\{P_{i_1},P_{i_2},\ldots,P_{i_{k_1}}\} \in \mathcal{PS}(\mathcal{P},k_1,k_1)\) be a set of any \(k_1\) participants. First, each \(P_{i_j} \in \mathcal{A}\) computes \(a_{k_1-1+i}^{(t)}=p_i^{(t)}-r_i^{(t)} \ (i=1,2,\ldots,k_2-k_1)\) and constructs \(g(x):=\sum^{k_2-1}_{k_1}a_{i}x^{i}\). Then, each \(P_{i_j}\) computes \(h(P_{i_j}):=f(P_{i_j})-g(P_{i_j}) \ (j=1,\ldots,k_1)\) such that \(h(x):=s +\sum^{{k_1}-1}_{i=1}a_{i}x^i\). Then, they compute \[\begin{aligned} s=\sum_{j=1}^{k_1}(\prod_{l\neq j}\frac{P_{i_j}}{P_{i_j}-P_{i_l}})h(P_{i_j}), \end{aligned}\] by Lagrange interpolation from \((h(P_{i_1}),\ldots,h(P_{i_{k_1}}))\). 5. *Reconstruct without time-signals.* any \(\hat{\mathcal{A}}=\{P_{i_1},P_{i_2},\ldots,P_{i_{k_2}}\} \in \mathcal{PS}(\mathcal{P},k_2,k_2)\) computes \[\begin{aligned} s=\sum_{j=1}^{k_2}(\prod_{l\neq j}\frac{P_{i_j}}{P_{i_j}-P_{i_l}})f(P_{i_j}), \end{aligned}\] by Lagrange interpolation from their \(k_2\) shares. The security and optimality of the above construction is stated as follows. *Proof.* First, we show the proof of (i) in Definition [\[Security2\]](#Security2){reference-type="ref" reference="Security2"}. Assume that \(k_1-1\) participants \(\mathcal{F}=\{P_{i_1},\ldots,P_{i_{{k_1}-1}}\}\in\mathcal{PS}(\mathcal{P},k_1-1,k_1-1)\) try to guess \(s\) by using their shares, public parameters, and all time-signals. \(\mathcal{F}\) can compute \(g(x)\) from public parameters and the time-signal at the specified time, hence they can get \(h(P_{i_l})=f(P_{i_l})-g(P_{i_l}) \ (l=1,\ldots,{k_1}-1)\). Thus, they can know the following matrix: \[\begin{aligned} \left( \begin{array}{cccc} 1 & P_{i_1} & \cdots & P_{i_1}^{k_1-1} \\ 1 & P_{i_2} & \cdots & P_{i_2}^{k_1-1} \\ \vdots & \vdots & \ddots & \vdots \\ 1 & P_{i_{k_1-1}} & \cdots & P_{i_{k_1-1}}^{k_1-1}\\ \end{array} \right) \left( \begin{array}{c} s \\ a_1 \\ \vdots \\ a_{k-1}\\ \end{array} \right). \label{optimal01} \end{aligned}\] However, from ([\[direct01\]](#direct01){reference-type="ref" reference="direct01"}), they cannot guess at least one element of \((s, a_1, \ldots,a_{k_1-1})\) with probability larger than \(1/q\). Therefore, for any \(\mathcal{F} \in \mathcal{PS}(\mathcal{P}, 1,k_1-1)\) and any \(t \in \mathcal{T}\), we have \(H(S \mid U_{\mathcal{F}}^{(t)}, TI^{(1)},\ldots,TI^{(\tau)}) = H(S)\). Next, we show the proof of (ii) in Definition [\[Security2\]](#Security2){reference-type="ref" reference="Security2"}. Without loss of generality, we suppose that \(\tau\) is a specified time, that \(k_2-k_1=\ell\), and that \({k_2}-1\) participants try to guess \(s\) by using their shares, public parameters, and time-signals at all the time except the time \(\tau\). First, they cannot guess at least one coefficient of \(f(x)\) with probability larger than \(1/q\) since the degree of \(f(x)\) is at most \(k_2-1\). Therefore, they attempt to guess one of \(a_{k_1},\ldots,a_{k_2-1}\) by using their \(k_2-1\) shares, public parameters and \(\tau-1\) time-signals, since if they obtain any one of these coefficient, they can get \(f^*(P_{i_l}) \ (l=1,\ldots,{k_2}-1)\) such that the degree of \(f^*(x)\) is \(k_2-2\) and reconstruct \(s\) by Lagrange interpolation. They know \(\tau-1\) time-signals, however, these time-signals \(\{(r_1^{(j)},\ldots,r_{\ell}^{(j)})\}_{1\le j \le \tau-1}\) are independent of the time-signal \((r_1^{(\tau)},\ldots,r_{\ell}^{(\tau)})\) at \(\tau\). Hence, by the security of one-time pad, they cannot guess each \(a_{k_1-1+i} \ (=p_i^{(\tau)}-r_i^{(\tau)}) \ (1\le i \le k_2-k_1)\) with probability larger than \(1/q\) since each \(r_i^{(\tau)}\) is chosen from \(\mathbb{F}_q\) uniformly at random. Therefore, we have \(H(S \mid U_{l_1}^{(\tau)},\ldots,U_{l_{{k_2}-1}}^{(\tau)}, TI^{(1)},\ldots,T^{(\tau-1)}) = H(S)\). Hence, for any \(\mathcal{A} \in \mathcal{PS}(\mathcal{P}, k_1,k_2-1)\) and any \(t \in \mathcal{T}\), we have \(H(S \mid U_{\mathcal{A}}^{(t)}, TI^{(1)},\ldots,TI^{(t-1)},TI^{(t+1)},\ldots,T^{(\tau)}) = H(S)\). Finally, if \(k_2-k_1=\ell\), it is straightforward to see that the construction satisfies all the equalities of lower bounds in Theorem [\[Bounds2\]](#Bounds2){reference-type="ref" reference="Bounds2"}. Therefore, the above construction is optimal if \(k_2-k_1=\ell\). # Concluding Remarks {#Conclusion} In this paper, we first studied two kinds of secret sharing schemes with timed-release security in the information-theoretic setting, \((k,n)\)-TR-SS and \((k_1,k_2,n)\)-TR-SS. Specifically, we defined a model and security for each scheme, and derived tight lower bounds on sizes of shares, time-signals, and secret-keys required for each scheme. Moreover, we respectively proposed optimal direct constructions of both schemes. These results showed that information-theoretic timed-release security can be realized in secret sharing schemes without any redundancy on share-sizes. In a similar way, it is expected that information-theoretic timed-release security can be realized for secret sharing schemes with any *access structure* without any redundancy on share-sizes. It would be also interesting to extend our results to timed-release verifiable secret sharing schemes, and furthermore, to multiparty computation schemes with timed-release security. [^1]: If we consider a situation in which \(TS\) is trusted and \(TS\) has functionality of generating keys and distributing them to participants by secure private channels, we can identify \(TA\) with \(TS\) in the situation. However, there may be a situation in which the roles of \(TA\) and \(TS\) are quite different (e.g., \(TA\) is a provider of secure data storage service and \(TS\) is a time-signal broadcasting server). Therefore, we assume two entities \(TA\) and \(TS\) in our model to capture various situations. [^2]: More precisely, there is no need to keep the specified time confidential (\(D\) only has to send shares via secure channels). [^3]: In this sense, we have formalized the security notion stronger than the security that any set of more than \(k-1\) participants cannot obtain any information on a secret *before* the specified time, as is the same approach considered in. [^4]: In this optimal construction, a dealer is only allowed to choose \(k_1\) and \(k_2\) such that \(k_2-k_1\le\ell\), where \(\ell\) is determined by \(TA\) in the phase *Initialize*. In this sense, this construction is restricted.

{'timestamp': '2014-05-13T02:10:42', 'yymm': '1401', 'arxiv_id': '1401.5895', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5895'}

# Introduction H in metals has been of considerable interest for many decades due to possible uses of metals as H storage devices but often more commonly because of the deleterious effect H has on metals through processes such as H embrittlement. There have also been numerous studies both by experiment and by first principles methods of the formation energies and stability of phases especially the precipitation of hydrides. One important system subject to hydriding is that of Zr alloys, often used for fuel cladding in light water nuclear reactors, due to their good mechanical and corrosion properties and low capture cross-section for thermal neutrons. The formation of hydrides in Zr alloys is a key issue in the ongoing development of fuel cladding for water cooled reactors. The rate of production, absorption, and precipitation of hydrides is one of the most important safety factors, as it is a limiting factor determining how long fuel assemblies may be kept in a reactor; it can cause complications with intermediary spent fuel storage. Zr corrodes in an aqueous environment, producing an oxide layer and H. Once produced, some H is released into the water, and some diffuses through the protective oxide barrier, and into the Zr metal. At reactor operating temperatures, H is soluble up to around 100ppm and extremely mobile in \(\alpha\)-Zr. However, if sufficient H is present or if the solubility limit of H in \(\alpha\)-Zr is lowered (commonly due to a temperature drop during reactor transients ), then H will precipitate out, forming a hydride. The hydrides are brittle and liable to cause failure of the fuel cladding; this is especially likely if they align along the radial direction of the fuel pin. Current understanding suggests that there are five main ways that H can be sequestrated in Zr metal; these are, solid solution and four hydrides: \(\zeta\), \(\gamma\), \(\delta\) and \(\varepsilon\). The structures of these different hydrides are shown in Fig [\[fig:Hydrides\]](#fig:Hydrides){reference-type="ref" reference="fig:Hydrides"}. There is a general relationship of decreasing c/a ratio with increasing H content. Of particular interest is that the most commonly observed \(\delta\) phase is often reported to have a formula of ZrH\(_{1.66}\) and a disordered fluorite structure, while simulators often use a simplified ersatz of periodic ZrH\(_{1.5}\). The \(\gamma\) hydride has been assumed to be meta-stable, as it is less readily observed than the \(\delta\) hydride. However, other investigations have observed the \(\gamma\)-hydride at room temperature conditions, under both slow cooling and fast cooling regimes. Overall, it appears that the stability and occurrence of these phases is a complex phenomenon, where the H concentration, thermal treatments, alloying additions and stress states all have a part to play in determining which phases are observed. The focus of this study is on modelling hydrides using atomistic simulation techniques based on density functional theory (DFT). DFT has previously been used to successfully model thermodynamics in H-metal systems. In this system in particular, it has been found that H preferentially occupies tetrahedral sites in the Zr lattice, rather than other sites. Some works have focused on the octahedral site as the main location for H atom solution, however such works appear to be in the minority. It has been theoretically predicted that the FCT structure of ZrH\(_{2}\) can have another stable phase for c/a \> 1. Zhu et al. studied the ordered hydride phases using DFT; they concluded that \(\delta\)-ZrH\(_{1.5}\) is thermodynamically less stable than the other phases at high temperature. Zhong and MacDonald, who published previous DFT results combined with new calculations of the \(\gamma\) phase, suggest the \(\gamma\) phase is stable at temperatures below about 523 K. The simulation of hydrides is complicated due to the random distribution of the H atoms in some phases. Although hydrides have been simulated in the past, few studies have attempted to examine hydrides whilst taking into account the disorder. One of the most successfully used techniques for simulating disordered atomic structures is the special quasi-random structures (SQS) method developed by Zunger et al., which has been used to simulate a range of non-stoichiometric materials and structures. It has also recently been applied to this system. In this study SQS techniques are combined with a statistical analysis of a large number of randomly generated cells, in order to examine the impact of disorder on hydride precipitation. Phonon calculations are also used to calculate thermodynamic properties such as the vibrational entropy and the sensible enthalpy changes during precipitation. Previous studies on other systems by DFT have shown the importance of vibrational entropy on the solubility limit in precipitation reactions from solid solution, and its importance in creating a temperature dependent understanding of hydride precipitation. Thus, a comprehensive view of the enthalpic and entropic contributions towards hydride precipitation in Zr is developed. # Methodology ## Simulation Parameters For this investigation, CASTEP 5.5 was used to simulate the different structures. As a plane-wave pseudopotential code, it is particularly appropriate for modelling crystals. Ultrasoft pseudopotentials were generated \"on-the-fly\", under the formalisation of Vanderbilt et al..Valence electrons for Zr were modelled as \(4s^{2}4p^{6}5s^{2}4d^{2}\). Convergence with respect to basis-set cut-off energy and k-point grid density was tested in a series of electronic self consistency calculations. It was found that the simulations were accurate to 2 d.p. for a cut-off energy of 400 eV and a k-point grid spaced at 0.03 \(\angstrom^{-1}\). k-points were arranged in a gamma-centred Monkhorst-Pack grid. As the system displays metallic characteristics, the integration of the Brillouin zone is achieved via a Methfessel-Paxton scheme, with a band smearing width of 1 eV. All cells used in this work are geometry relaxed in order to approach their minimum energy configuration. Cells were considered relaxed when the difference between two successively modified iterations were below all of the following criteria: - Energy derivative \(<\) 0.001 eV - Force on ions \(<\) 0.05 \(\textrm{eV\angstrom}^{-1}\) - Displacement of ions derivative \(<\) 0.001 \(\textrm{\angstrom}\) - Total stress derivative \(<\) 0.1 GPa Relaxation of atomic positions was carried out under the quasi-Newtonian BFGS scheme. Both atomic positions within the cell, the lattice constants and cell aspect ratios were unconstrained during relaxation. This means that volume and cell distortions due to H accommodation are fully accounted for. ## Thermodynamic Considerations With static simulations it is usual to consider the energy change from a set of reactants to a set of products. Relating the calculated energy changes to a real system is often difficult, as there are numerous different energy components, while static calculations can directly only evaluate the ground state changes (ie. the latent enthalpy of a reaction). Thus, many factors present in a real system are not represented in a simulation. These different components and factors are often poorly communicated between experimentalists and simulators, leading to inappropriate assumptions when using terms such as energy, enthalpy and entropy. To ensure a clear understanding, the terms that were calculated here are now discussed. The fundamental measure of the driving force behind a reaction is the Gibbs free energy change: \[\Delta G = \Delta H-T\Delta S \label{eq:gibbs}\] where \(\Delta H\) represents the enthalpy change of the system and \(\Delta S\) represents the entropy change. There are different contributors to the enthalpic and entropic terms, thus Eq. [\[eq:gibbs\]](#eq:gibbs){reference-type="ref" reference="eq:gibbs"} can be expanded into: \[\Delta G = (\Delta H_{l}+ \Delta H_{s T})-T (\Delta S_{v} + \Delta S_{cl}) \label{eq:gibbs2}\] where: - \(\Delta H_{l}\) is the latent enthalpy associated with the reaction. It is due simply to the formation or destruction of bonds and is independent of external conditions. When quoting results from DFT simulations, this is the most commonly reported number. If the other terms are ignored this is often quoted as the energy. - \(\Delta H_{s}\) is the sensible enthalpy and is related to the heat capacities of the different species involved in the reaction. In order to calculate this, the amount of thermal energy that can be stored in the lattice needs to be known. This is achieved by calculating the phonon density of states of the materials in question, and integrating the different acoustic modes in a quasi-harmonic approximation, for a temperature \(T\), to yield \(H_{s T}\). - \(\Delta S_{v}\) is the vibrational entropy and is a function of the number of discrete vibrational energy levels that exist amongst the different acoustic modes of the system. This is determined by the application of the Boltzmann entropy equation to a quasi-harmonic approximation of the phonon distribution. - \(\Delta S_{cl}\) is the lattice component of the configurational entropy. This only applies to solids with a disordered lattice, which in this system consists of the H sub-lattice. An approximation of this is reported in section [\[sec:Entropy\]](#sec:Entropy){reference-type="ref" reference="sec:Entropy"}. It is clear that the distribution of phonons must be calculated in order to determine two of the terms included here. Analysis of phonon distributions has been previously examined in this system by Blomqvist et al.. In the present study, this was achieved by means of the finite displacement method, in the (direct) supercell approach. The application of the phonon data to the creation of thermodynamic parameters has been described previously. In a real precipitation event, there would be stress related effects, as well as contributions arising from the creation of a hydride-Zr interface. There would also be entropy created by intrinsic defects on the Zr lattice. This study cannot comment on such effects, as they would require simulations that are either significantly larger than the scale that DFT presently allows, or an exceedingly large number of simulations. However, the numbers produced in this work represent a large portion of the overall driving forces for precipitation, and are a significant advance on what has been achieved for this system previously. ## Cell Configurations This study examines both ordered and disordered models of hydrides and Zr-H solid solutions. The ordered models are performed by a straight-forward geometry optimisation of hydride structures including the commonly accepted structures of \(\zeta\)-Zr\(_{2}\)H, \(\gamma\)-ZrH, \(\delta\)-ZrH\(_{1.5}\) and \(\varepsilon\)-ZrH\(_{2}\). In addition, dilute solid solutions were also simulated in which one hydrogen atom was placed in \(2\times2\times2\), \(3\times3\times2\) and \(4\times4\times3\) supercells of \(\alpha\)-Zr, giving concentrations of 5.8 at%H, 2.7 at%H and 1.0 at%H respectively. A further calculation was carried out in which one atom of H was placed in a \(2\times2\times2\) supercell of FCC Zr. All of the solid solution calculations described so far were carried out with H occupancy being investigated on both octahedral and tetrahedral sites. A different method was used to simulate more disordered structures. As a starting point for the generation of off-stoichiometric hydride phases, a \(\delta\) hydride structure is built, formed from a \(2\times2\times2\) supercell of the primitive cell, with all tetrahedral H sites occupied (giving a formula of ZrH\(_{2}\)). This is similar to the \(\varepsilon\) hydride, except that the \(c/a\) ratio is 1. A new cell is generated from this input, by giving each H atom a random chance to be removed of 0.166 (or \(1-\frac{1.\dot{6}}{2}\)). This is repeated to generate a large number of cells, with the only constraint being that new cells must be unique from previously generated cells. These cells then undergo geometry optimisation to provide energies for each configuration. No constraints are placed upon the number of H atoms in any given cell, meaning some cells will have more H than the 1.66 ratio, and some will have less. Providing the number of cells is large enough, the formulae of the different cells will follow a normal distribution with a mean of 1.66. A similar method was used with a \(2\times2\times2\) supercell of \(\alpha\)-Zr, and a removal probability of 0.89 with single H atom cells discounted. This provides a selection of solid solution cells where the H exists in small clusters, modally containing 2 H atoms, which are used to bridge the stoichiometry gap between a dilute solid solution and the \(\zeta\)-phase. These sets will have a number of different possible configurations, which allows examination of how configurationally sensitive the properties of the hydrides and solid solutions are. Taken together, an arbitrarily large set of these cells provide a more complete description of the \(\delta\)-hydride and a concentrated solid solution than any single periodic calculation. The advantages of random structure generation have been discussed previously. In order to qualify this method against established techniques, the random \(\delta\)-hydride set is compared with SQS generated cells. The SQS technique, as detailed in reference, works on the assumption that in any random arrangement of atoms on a pre-defined set of atomic sites, some clusters of atoms will be more common than others. The more common clusters are defined by the structure and stoichiometry of the simulated crystal. Thus, a number of \"special\" cells can be constructed that comprise the more common configurations, and less common configurations can be discounted. In this study, 3 different SQS cells are simulated containing 48, 52 and 56 H atoms, along with 32 Zr atoms. The SQS method has been recently applied to this system in order to model bulk parameters. The 3 most representative SQS cells in that paper are used again here. Some phases, however, exhibit more order than the \(\delta\) phase, such as the \(\gamma\) and \(\zeta\) phases. Generally, these phases are simulated as ordered phases with a defined structure and no disorder. However, any real hydride will likely exhibit a degree of disorder, particularly when a disordered phase can transform to an ordered phase and vice versa. Thus, a method has been devised to introduce disorder into these structures, whilst biasing the results towards the known structures. In this \"skew-random\" method, the starting point is a cell of the ordered structure. For each occupied H site, a small removal probability of the H atom is introduced. Likewise, for each unoccupied site, a small chance is created to have a H atom inserted. In both of these cases the probability of 0.05 was used, as it created a reasonable spread of random structures, whilst remaining close to the crystallographic form of the hydride. As before, a large number of these cells were generated and only unique cells are simulated to ensure that the full ensemble of simulations is relevant to a (partially) disordered hydride. A smaller number of random cells was also generated with no structure biasing and a probability for removal of 0.5. These were used purely as a comparison with the skew-random cells. Ultimately, all possible arrangements of H on interstitial sites in Zr exist as points in the configuration space. Each of these different methods develops a different sampling of that configuration space, aiming to ensure that a valid distribution is identified. Such sampling methods are required, as a single point is not sufficient to model all the important aspects of the system, while a full census of the configuration space carries a computationally prohibitive cost. # Results ## Elements and Ordered Crystals In order to ensure the validity of the simulations performed, it is important to confirm that the simulation results can reproduce experimental data. As much of the more useful thermodynamic data makes use of the pure elements as reference states, it is particularly important to ensure that the reference states are well modelled. In Table [\[tab:elements\]](#tab:elements){reference-type="ref" reference="tab:elements"}, properties of reference states are reported compared with established literature values. Excellent agreement is achieved on all counts, to within a maximum discrepancy of 1.65%. As most of the thermodynamic values in this study are contingent on the presence and position of H atoms in a Zr cell, the values obtained for H are of critical importance. For this reason, great care has been taken to ensure that the simulation parameters can accurately produce the properties of the H\(_{2}\) gas reference state. The enthalpies of formation of different ordered crystals are given in Table [\[tab:ordered\]](#tab:ordered){reference-type="ref" reference="tab:ordered"}. The formation energies were calculated via the equation: \[\Delta E^{F} = \frac{1}{x+y}\left[E(\mathrm{Zr}_{x}\mathrm{H}_{y})-\left(E(x\mathrm{Zr}) + y\frac{1}{2} E(\mathrm{H}_{2})\right)\right] \label{eq:Formation Enthalpy}\] where \(E(Zr_{x}H_{y})\) represents the energy of the hydride (or solid solution containing H) and the other terms refer to the pure elements. As the number of atoms in each simulation differs, the formation energy must be normalised with respect to the number of atoms, in order to not bias the formation energies towards the larger cells. Solution enthalpies of H atoms in a HCP and an FCC matrix are also presented. Although the latter configuration is un-physical (pure FCC Zr is not a stable phase, nor observed in real alloys) it does provide useful comparison points. Although this phase assumes the absorption of H into an FCC lattice, the reference state is still taken to be HCP Zr. This is done to ensure fair comparison with other results and is based on the assumption that any starting point that could lead to this configuration would still be based on HCP Zr. As formation energies for FCC solutions must also contain the energy associated with a HCP \(\rightarrow\) FCC phase change, it is reasonable that the formation energies for the FCC solutions are higher than their HCP counterparts. The dilute tetrahedral solid solution is of particular importance as it represents a reference point for comparison in further calculations. The number reported here of-0.60 eV compares favourably with-0.52 eV from,-0.604 eV from and-0.464 from. The tetrahedral site for H occupancy remains the most favourable, in agreement with most prior work. For the remainder of this work, when considering sites for H occupancy, only the tetrahedral site is considered. More exotic configurations such as H\(_{2}\) dimers on interstitial sites have also previously been found to be unfavourable. With regards to the approximate \(\delta\) phase, ZrH\(_{1.5}\), a conventional unit cell of FCC Zr offers 8 sites for H occupancy, 6 of which must be filled with the other two vacant. If the system is cubic, then symmetry reduces the number of configurations to three different arrangements. These arrangements are where the vacancies are both in the \[100\], \[110\] and \[111\] directions. These are referred to in Table [\[tab:ordered\]](#tab:ordered){reference-type="ref" reference="tab:ordered"} by these directions. The configuration with the lowest enthalpy of formation is the one in the \[111\] orientation, where the vacancies are separated by the longest distance. The energy difference between these states is relatively small, and similar (but slightly larger) than the average of the energy calculated from using the three SQS configurations. ## Statistical Analysis In order to ensure that the simulations are representative of the disordered system it is important that a large enough sample of the configuration space is achieved. With this in mind, the statistical parameters generated in the sets used are shown in Table [\[tab:Stats\]](#tab:Stats){reference-type="ref" reference="tab:Stats"}. The \(\delta\) and solid solution series rely on a random distribution about the selected stoichiometry, while the \(\zeta\) and \(\gamma\) phases use the skew-random method. A large enough sample has been made when the data set forms a normal distribution centred on the target stoichiometry. A simple convention for determining normality is a plot of the cumulative distribution probabilities of the data against theoretical cumulative distribution probabilities generated by a standard normal distribution, as created from the parameters in Table [\[tab:Stats\]](#tab:Stats){reference-type="ref" reference="tab:Stats"}. A straight line fit would represent perfectly normal data. Normality tests were performed on sets of increasing sample size until a high degree of confidence in normality was achieved. Figs [\[fig:Statform\]](#fig:Statform){reference-type="ref" reference="fig:Statform"} and [\[fig:Statenth\]](#fig:Statenth){reference-type="ref" reference="fig:Statenth"} show normality tests for each set of data, generated from both the stoichiometry distribution and the formation enthalpy distribution. In both cases, we see all series display a good linear fit. We report that for sample sizes of 50 cells per set, all datasets showed high coefficients of linear regression with the lowest R\(^{2}\) being 0.9602. The average stoichiometry for each set is extremely close to the experimental formula value considered representative for that hydride structure. This gives confidence to the hypothesis that this set of randomly generated structures approximates a disordered material when taken as a whole. ## Enthalpies[\[sec:Enthalpy\]]{#sec:Enthalpy label="sec:Enthalpy"} Relative thermodynamic phase stability can be determined by plotting the formation energy across the range of compositions. These results are presented in Fig. [\[fig:FormationEnthaply\]](#fig:FormationEnthaply){reference-type="ref" reference="fig:FormationEnthaply"}. Specifically, this represents a latent formation enthalpy, as opposed to a free energy. In addition to a presentation of the various data sets, a convex hull has been drawn on the plot. Here, a convex hull is defined as the smallest convex path to contain all of the available data points, when viewed from below the plot. It is useful, because any mixture with an enthalpy less negative than the convex hull would be more stable as a mixture of the two configurations which bound that segment of the hull. In Fig. [\[fig:FormationEnthaply\]](#fig:FormationEnthaply){reference-type="ref" reference="fig:FormationEnthaply"}, all enthalpies are negative indicating that there is a general thermodynamic driving force for formation, which becomes stronger with greater H-content phases. However, the majority of configurations lie above the convex hull, indicating they are less stable than a mixture of other phases. The configurations which lie on the convex hull are the 1 at%H solid solution, the stoichiometric and ordered \(\gamma\)-hydride, and the \(\varepsilon\)-phase with a c/a ratio of less than 1. The enthalpies in the stoichiometry range ZrH\(_{0}\) to ZrH\(_{0.6}\) agree with a similar plot produced by Hollinger et al. in terms of the range of formation enthalpies of the different structures. However, whereas that work noted stable structures in this range, none are found in the present work. This is almost certainly due to the fact that the work of Hollinger et al. was focused on hexagonal phases, whereas the present work shows that cubic phases out-compete hexagonal structures in terms of stability. Domain et al., provide a similar plot with no convex hull, however adding one demonstrates the same phases (\(\varepsilon\) and \(\gamma\)) as stable and by similar energies. This result has also been found by Zhong and MacDonald, who used this data to suggest that the \(\gamma\) phase is thermodynamically stable below \(\approx\) 523 K. This contrasts greatly with Zhu et al., who claimed that the \(\delta\)-hydride is by far the most stable hydride, by nearly 8 eV more than the other phases. It is, however, difficult to understand that result, since the \"convex hull\" presented was not actually convex, and the magnitude of this number is out of line with other results. Ultimately, hydrides are formed by the precipitation of H from solid solution in the \(\alpha\)-Zr matrix. This reaction is given by the expression: \[\Delta E^{P} = \left[E(Zr_{x}H_{y})+(y)E(Zr_{R})\right]-\left[yE(Zr_{R}H)+ xE(Zr)\right] \label{eq:Precipitation Enthalpy}\] This equation forms the basis of calculating the change in different thermodynamic parameters such as the latent enthalpy of precipitation. The term \(R\) is the number of Zr atoms in the solid solution reference cell. Precipitation enthalpies have been calculated using reference solutions containing 96, 36 or 16 atoms of Zr to one atom of H. This equation is balanced with free Zr on both sides because it ensures that the reaction maintains reversibility in situations where the concentration is different. As with the formation enthalpies, these precipitation enthalpies must be normalised to ensure that larger simulations are not shown as having larger enthalpies purely based on their size, and not on changes in composition and thermodynamic behaviour. To this end, all simulations are divided by the total number of H atoms present in the hydride phase, and then converted into kJmol\(^{-1}\). Thus, the enthalpies presented hence forth are in units of kJmolH\(^{-1}\), representing the enthalpy change required for one mole of H atoms to precipitate from a solid solution. The latent enthalpies of precipitation are presented in Fig. [\[fig:LatentEnthaply\]](#fig:LatentEnthaply){reference-type="ref" reference="fig:LatentEnthaply"}. As with the formation enthalpies, there is a general trend that the precipitation of H-rich hydrides is more preferable than H-poor hydrides. On the H poor side of the graph, solid solutions have more negative enthalpies when they are less concentrated than the reference solid solution for that series, suggesting a trend towards dilution of H atoms. Moving across towards products with a greater H content, there is then a peak of unfavourable H clusters around ZrH\(_{0.2}\), followed by a steady return to the more preferable hydride phases. In particular, the \(\gamma\)-phase exhibits the strongest preference for precipitation, with the most favourable configuration being the structure typically modelled in other simulation studies, shown in Fig. [\[fig:Hydrides\]](#fig:Hydrides){reference-type="ref" reference="fig:Hydrides"}. There is a notable discontinuity in all series at \(\approx\) ZrH\(_{0.75}\), corresponding to the point where the series switched from modelling HCP hydrides to FCC hydrides. There are no negative latent enthalpies of precipitation for the precipitations from 1 at%H and 2.7 at%H solid solutions. However there are for the 5.9 at%H solid solution. It is sensible that increasing the H content in the Zr lattice increases the impetus for the rearrangement of the H atoms into a hydride, as is evidenced by the existence of a terminal solubility limit for H in Zr. Overall, this plot is consistent with H having a bimodal distribution in Zr, preferring to exist either as a sparsely distributed solid solution, or as a concentrated hydride. A middle-ground between these two modes is unfavourable. So far, this only describes the latent enthalpy with no regards for the effects of temperature. The sensible enthalpy of precipitation is related to the heat capacities of the products and reactants of the precipitation reaction. Heat capacities calculated at 298 K are 23.00 kJmol\(^{-1}\) K\(^{-1}\) for \(\alpha\)-Zr and 28.64 kJmol\(^{-1}\) K\(^{-1}\) for \(\varepsilon\) ZrH\(_{2}\) (compared with the available experimental values of 25.45 kJmol\(^{-1}\) K\(^{-1}\) and 31.08 kJmol\(^{-1}\) K\(^{-1}\) ). Fig. [\[fig:AbsSensibleEnthalpy\]](#fig:AbsSensibleEnthalpy){reference-type="ref" reference="fig:AbsSensibleEnthalpy"} gives the absolute sensible enthalpies for reference simulations, as they vary with temperature. The values for the three \(\delta\)-phase stoichiometries are calculated using the SQS generated cells. There is a general trend for increasing sensible enthalpy with increasing H content. The \(\zeta\)-phase has a substantially lower sensible enthalpy than the other hydride phases. As the temperature increases, the variance in sensible enthalpies decreases. The enthalpy calculated at 0 K represents the zero point energy contribution to the enthalpy of this system. The absolute sensible enthalpy is of less interest than the change in sensible enthalpy which may drive precipitation. Using Fig. [\[fig:AbsSensibleEnthalpy\]](#fig:AbsSensibleEnthalpy){reference-type="ref" reference="fig:AbsSensibleEnthalpy"}, a surface is generated to describe the relationship between composition, temperature and sensible enthalpy. Using this surface, values are interpolated for sensible enthalpies for all the cells examined in this work. The enthalpy data is sufficiently close that a simple linear interpolation does not introduce unreasonable variance. Feeding this interpolation into Eq. [\[eq:Precipitation Enthalpy\]](#eq:Precipitation Enthalpy){reference-type="ref" reference="eq:Precipitation Enthalpy"} (replacing the latent enthalpy term), the sensible enthalpy change during precipitation for a variety of different structures and temperatures is generated and plotted in Fig. [\[fig:SensibleEnthalpy\]](#fig:SensibleEnthalpy){reference-type="ref" reference="fig:SensibleEnthalpy"}. This information is presented only for the precipitation from the 16 atom solid solution. Given that the sensible enthalpy is added to the latent enthalpy, in Eq. (2), a negative value of sensible enthalpy represents a driving force for precipitation, while a positive value represents a driving force for solution. The sensible enthalpy appears to drive the system towards precipitation for all product stoichiometries greater than \(\approx\) ZrH\(_{0.08}\). As temperature increases, the driving force for precipitation also increases. There is a relative increase in this driving force for stoichiometries of \(\approx\) ZrH\(_{0.4}\), which corresponds roughly with the stoichiometries found in the \(\zeta\)-phase hydrides. The sensible enthalpy then becomes less negative for stoichiometries appropriate to \(\gamma\)-hydrides before reducing slightly for hydrides with even greater H content. ## Entropy [\[sec:Entropy\]]{#sec:Entropy label="sec:Entropy"} As described previously, computing the free energy of a reaction requires a description of the entropy as well as the enthalpy. In this study, we examine two sources of entropy-the vibrational and the configurational. Configurational entropy stems from the disorder available when the structure may have multiple different forms. It is quantified by the Boltzmann entropy equation: \[S_{c} = k \ln\Omega \label{eq:Boltzmann}\] where \(\Omega\) is defined as the number of different configurations or micro-states in which the system may be arranged and \(k\) is Boltzmann's constant. In an atomistic context, the number of different structure configurations is given by adapting the standard permutations expression: \[\Omega= \frac{(N_{V}+N_{H})!}{N_{V}!N_{H}!} \label{eq:Permutation}\] where \(N_{V}\) is the number of potential H sites which are vacant, while \(N_{H}\) is the number of H atoms. As before, the primary concern is not the absolute entropy, but the change in entropy during precipitation. Using the entropy calculated in Eq. [\[eq:Boltzmann\]](#eq:Boltzmann){reference-type="ref" reference="eq:Boltzmann"} in the precipitation equation (4) (with the H coming from the 16 atom Zr cell), the change in configurational entropy is determined across a range of stoichiometries, and displayed in Fig. [\[fig:ConfigEntropy\]](#fig:ConfigEntropy){reference-type="ref" reference="fig:ConfigEntropy"}. These entropies are presented as a \(T \Delta S\) product. As entropies are subtracted from enthalpies to generate a free energy, a negative value indicates a driving force towards solution while a positive value drives towards precipitation. For non-zero temperatures, we see that the configurational entropy represents a driving force for solution, with increasing temperatures, of course, leading to more negative values. There is a notable discontinuity when the simulated series shift to modelling FCC hydrides. This is because the FCC structure has more tetrahedral sites per Zr atom, which are considered as possible sites for H occupancy (i.e. it offers greater configurational options). Thus, the shift from HCP to FCC is favoured by the configurational entropy and this driving force increases with temperature. The final contribution examined in this work is the vibrational entropy. Vibrational entropies are shown in Fig. [\[fig:AbsVibrationalEntropy\]](#fig:AbsVibrationalEntropy){reference-type="ref" reference="fig:AbsVibrationalEntropy"} for the same reference cells as used in calculating the sensible enthalpy. Vibrational entropies at 298 K are 37.52 Jmol\(^{-1}\) K\(^{-1}\) for \(\alpha\)-Zr and 31.387 Jmol\(^{-1}\) K\(^{-1}\) for \(\varepsilon\) ZrH\(_{2}\), compared with the available experimental values of 39.144 Jmol\(^{-1}\) K\(^{-1}\) and 35.154 Jmol\(^{-1}\) K\(^{-1}\) respectively. It should be noted that experimental results will include other forms of entropy (such as that generated by intrinsic defects on the Zr lattice), hence it is reasonable that the theoretical results are slightly smaller than experimental values. Fig. [\[fig:AbsVibrationalEntropy\]](#fig:AbsVibrationalEntropy){reference-type="ref" reference="fig:AbsVibrationalEntropy"} demonstrates a decreasing vibrational entropy with increasing H content. Similar to the calculation of sensible enthalpies, this plot is used to interpolate values from a temperature-composition-entropy surface. Applied across the range of compositions, the vibrational entropy is given in Fig. [\[fig:VibrationalEntropy\]](#fig:VibrationalEntropy){reference-type="ref" reference="fig:VibrationalEntropy"} as a T\(\Delta\)S product. This plot shows negative values for all compositions above \(\approx\) ZrH\(_{0.08}\), and temperatures above 0 K. This is consistent with the vibrational entropy driving the reaction towards solution, with the effect becoming stronger with increasing temperature. The vibrational entropy change during precipitation is positive for dilute solid solutions, becomes negative for non-dilute solutions, and becomes more negative as H content increases. There is a decrease in the magnitude of the entropy change for hydrides of around \(\approx\) ZrH\(_{1.5}\), suggesting vibrational entropy may contribute to stabilising the \(\delta\) phase. ## Free Energy With the change in both the enthalpy and entropy terms calculated for the precipitation reaction, the overall free energy change can be calculated from Eq. [\[eq:gibbs2\]](#eq:gibbs2){reference-type="ref" reference="eq:gibbs2"}. It is sometimes stated that vibrational entropies and sensible enthalpies are too small to be important in this system and other hexagonal metals. Although this may be true for predicting if hydrides occur at all, and for determining energies when one reactant is in a different state (eg. H\(_{2}\) gas), in a system with multiple phases, containing subtle interactions, the magnitude of these other terms may be important. Given that the sensible enthalpy, configurational entropy, and vibrational entropy are all within the range of-20 to 40 kJmolH\(^{-1}\), none of these variables can be discounted and all have a part to play in determining phase stability. In Figs. [\[fig:FreeEnergy1\]](#fig:FreeEnergy1){reference-type="ref" reference="fig:FreeEnergy1"}-[\[fig:FreeEnergy3\]](#fig:FreeEnergy3){reference-type="ref" reference="fig:FreeEnergy3"}, the lowest energy configuration from each data set is plotted as a free energy, with respect to the stoichiometry. Fig. [\[fig:FreeEnergy1\]](#fig:FreeEnergy1){reference-type="ref" reference="fig:FreeEnergy1"} represents the free energy of precipitation from the 96 atom cell, Fig. [\[fig:FreeEnergy2\]](#fig:FreeEnergy2){reference-type="ref" reference="fig:FreeEnergy2"}, is from the 36, and Fig. [\[fig:FreeEnergy3\]](#fig:FreeEnergy3){reference-type="ref" reference="fig:FreeEnergy3"} is from the 16. In the first of these plots, Fig. [\[fig:FreeEnergy1\]](#fig:FreeEnergy1){reference-type="ref" reference="fig:FreeEnergy1"}, the free energy remains positive across the entire stoichiometry range. Temperature raises the energy by over three times the 0 K values. As with all these plots, there appear to be five distinct regions, defined by stoichiometry. The first occurs between ZrH\(_{0}\) \(\rightarrow\) ZrH\(_{0.1}\). Here, increasing the stoichiometry drastically increases the free energy of precipitation, suggesting that concentrating the H in the lattice is energetically unfavourable. This reaches a relatively flat region 2, made up of clusters of H atoms. This is particularly unfavourable, suggesting that H prefers to remain distributed. There is then a significant drop in free energy entering into region 3. The start of region 3 contains both H clusters, and sub-stoichiometric \(\zeta\)-hydrides as modelled by the skew-random technique. The \(\zeta\)-hydrides are more energetically favourable, and have a minimum energy point at ZrH\(_{0.5}\), for the expected structure of the \(\zeta\)-hydrides. However, as H content continues to increase, the free energy rises again and is out-competed by the sub-stoichiometric \(\gamma\)-phase at the start of region 4. This phase remains competitive until region 5 is entered, where stoichiometries are closer to that of the \(\delta\)-phase than \(\gamma\). Beyond this, energies remain relatively flat until the terminating \(\varepsilon\)-phase is reached. As temperatures increase, region 5 begins to show a slight upwards slope, signifying it is more preferable to precipitate larger quantities of H-poor hydrides, than smaller quantities of H-rich hydrides. The results are similar with an initial concentration of 2.7 at%H in Fig. [\[fig:FreeEnergy2\]](#fig:FreeEnergy2){reference-type="ref" reference="fig:FreeEnergy2"}. Although the same region structure follows, the 1.0 at%H solid solution, \(\gamma\) and \(\varepsilon\)-phases are now negative, showing precipitation is now favourable at 0 K. At higher temperatures, the reaction is still driven towards solid solution. Thus, the kinetics of H transport will mean that H is maintained in isolated solid solution. The free energies are lower overall, and the difference brought about by increasing temperature is smaller. Finally, the free energies drop significantly when moving to an initial concentration of 5.9 at%H. Fig. [\[fig:FreeEnergy3\]](#fig:FreeEnergy3){reference-type="ref" reference="fig:FreeEnergy3"}, shows negative precipitation energies across the full range of stoichiometry and temperatures, with the exception of stoichiometries in the range of ZrH\(_{0.1}\) to ZrH\(_{0.4}\) (region 2). Below this, there is still a thermodynamic driving force for keeping H dispersed, but above this, there is impetus for hydride formation. As in all prior plots, increasing temperature cause free energies to become more positive and drives solution. At the higher temperatures, the energy of the \(\zeta\)-phase increases to the point where its precipitation is no longer thermodynamically favourable. The tendency for regions 4 and 5 to slope upwards at higher temperatures is even greater when precipitation occurs from a more concentrated initial solution. The most favourable phases are \(\gamma\) and \(\varepsilon\), although higher temperatures seem to favour \(\gamma\)-hydrides. Overall, this remains consistent with a bimodal H distribution, as described previously. # Discussion ## Hydrogen in zirconium It is well established that there is a strong thermodynamic impetus for H to become sequestrated in a Zr lattice. This is corroborated by the energy values given for H solution in Table [\[tab:ordered\]](#tab:ordered){reference-type="ref" reference="tab:ordered"}, which are all substantially negative. If we compare this for any of the values for precipitation, we see that the impetus of adding H to the Zr lattice is substantially greater than the energy of rearranging or precipitating the H once it is already in the lattice. This means that if the thermodynamic values for H\(_2\) gas are less affected by temperature than the solid solution values are, it would be expected that H will continue to be added to Zr over the life of the cladding, steadily driving up concentration. In Figs. [\[fig:FreeEnergy1\]](#fig:FreeEnergy1){reference-type="ref" reference="fig:FreeEnergy1"} and [\[fig:FreeEnergy2\]](#fig:FreeEnergy2){reference-type="ref" reference="fig:FreeEnergy2"}, the lowest free-energy configuration is the 1.0 at%H solid solution, suggesting that H will preferentially form a dilute solid solution if possible. Temperature effects drive this behaviour further, in that the energy of the 1.0 at%H solid solution becomes more negative and the more H rich solid solutions become more positive. In order to produce this behaviour, there must be some sort of interaction between H atoms that raises the energy of the system. Given that electron interactions have been previously demonstrated to be extremely localised to H atoms in the H-Zr system, it is unlikely that the chemistry of H is driving this response. This leaves geometrical factors and most notably stress. It is possible that the stress fields created by the insertion of H atoms into nearby interstitial positions in the Zr lattice are mutually repulsive. ## Implications for hydride precipitation If there is an initial impetus for H atoms to remain in solid solution, then given that hydrides have been noted to form experimentally, at some point conditions *must* change to favour hydride formation. As more and more H atoms are absorbed by the Zr, the barrier for H atoms to congregate must be overcome. In Fig. [\[fig:FreeEnergy3\]](#fig:FreeEnergy3){reference-type="ref" reference="fig:FreeEnergy3"}, a high starting concentration of 5.9 at%H provides this condition. This suggests at some point between 307 ppm and 690 ppm, that H atoms will be so numerous that they will be pushed past their mutually repulsive behaviour and will start to form hydrides. These values are significantly higher than those measured globally in actual alloys, implying Zr has a much higher local H carrying capacity than suggested experimentally. If this is the case, in order for precipitation to occur, forces beyond those predicted in these DFT simulations must generate a driving force for concentration, and hence precipitation. Given this case, it is possible that larger scale stress states, such as those provided by defects or cracks, could lessen this mutually repulsive force, allowing H atoms to diffuse together more easily. The idea that stress impacts diffusion is not new, and has been suggested as a key part of the mechanism behind DHC. It is reasonable that the Zr lattice around an interstitial H atom is in a state of compression. Given this, a tensile stress field would provide a nullifying effect on the repulsive interaction. Coupled with areas that are not under stress, there would be an impetus for H atoms to move away from regions where H atoms are in close proximity and not under tension, towards areas where they can congregate more favourably. This argument is based upon the existence of the aforementioned barrier to association, and provides an area which can be investigated in future studies. Temperature also has the effect of driving the system towards solution, by raising the composition of the first point where hydriding may occur. In the room temperature series (300 K), the first hydride with a negative free energy of precipitation has a composition of about ZrH\(_{0.43}\), while in the operating temperature series, this is raised to over ZrH\(_{0.7}\). The main reason for this increase is entropic, in that both the configurational and vibrational entropy drive the reaction towards solution. ## Zirconium Hydrides The free energy curves produced in this study can be used to predict which hydride phase will precipitate, should precipitation occur. At 0 K, the most stable hydrides in all free energy figures are either the \(\gamma\) or \(\varepsilon\) phases. The commonly found \(\delta\)-phase rests somewhere between these two structures. Fig. [\[fig:FreeEnergy3\]](#fig:FreeEnergy3){reference-type="ref" reference="fig:FreeEnergy3"} suggests that as temperature increases, the stability of the \(\gamma\)-phase relative to \(\varepsilon\) increases. The primary driver of this is the configurational entropy. Thus, in reactor, we would expect to see \(\gamma\)-phase hydrides, while at room temperature or lower, hydrides with a greater H content are favoured. Due to the difficulties present in observing hydrides during operation, the exact morphology of hydrides present in a reactor are unknown and are generally inferred from low temperature experiments. However, it has been noted that \(\gamma\)-phase hydrides are present in both slowly and rapidly cooled samples, suggesting that the \(\gamma\)-phase can remain stable under low temperature equilibrium conditions. There are, however, other factors in the precipitation of hydrides that are difficult to investigate using DFT. For example, although the work here provides the thermodynamic driving force for precipitation of a quantity of hydride from a quantity of Zr-H solid solution, there is also the issue of forming an interface, which may or may not be coherent, resulting in additional parameters. If such parameters increased the energy of the \(\gamma\) or \(\varepsilon\)-phases, then the \(\delta\)-phase could be more preferable. The recently found \(\zeta\) phase is normally considered to be a metastable phase. The precipitation energies calculated here support this view, as the free energy curve forms a local minima about the ZrH\(_{0.5}\) composition. The sensible enthalpy plays a large part in the stabilisation of this phase over others. Interestingly, it is the only phase in Fig. [\[fig:FreeEnergy3\]](#fig:FreeEnergy3){reference-type="ref" reference="fig:FreeEnergy3"} that shows a transition from negative to positive precipitation energy with increasing temperature. The results suggest the \(\zeta\) phase should not be observed in samples at high temperature. # Conclusion This investigation has used DFT to investigate the thermodynamics of the precipitation of Zr hydrides over a range of temperatures, compositions and starting solid solution concentrations. The use of statistically significant numbers of randomly generated configurations has been coupled with SQS cells to ensure that disordered cells are modelled accurately. This investigation has lead to the following conclusions: - H favours a bimodal distribution within the Zr lattice. At low concentration, it prefers to maintain a dilute, non-clustered configuration, with a high energy barrier to hydride formation. As more H is absorbed by the Zr this barrier is overcome and hydride precipitation become energetically favourable. - The predicted concentration of the H solutions required to initiate precipitation is greater than observed experimentally, suggesting their may be additional mechanisms needed to enhance local H concentration to drive precipitation. Stress may play a part in this. - The calculation of latent enthalpies alone are insufficient to fully describe this system. Vibrational entropy, configurational entropy and sensible enthalpy are important for dealing with phase stabilities of precipitates and solid solutions. - Sensible enthalpies drive the reaction towards precipitation and are particularly significant for the \(\zeta\)-hydride. - Configurational entropy drives the system towards solution. They are particularly significant when contemplating the difference between HCP and FCC based hydrides. - Vibrational entropy and thus temperature drives the system towards solution. - Generally, the \(\gamma\) phase is the most stable, suggesting other mechanisms, (such as precipitate interface lattice strain) may be responsible for the observed presence of \(\delta\) hydrides.

{'timestamp': '2014-01-23T02:08:12', 'yymm': '1401', 'arxiv_id': '1401.5637', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5637'}

# Conclusions In summary, Co/graphene and Rh/graphene are revealed to be excellent Chern half metals, through DFT calculations and TB model analyses. We found that the Co/graphene and Rh/graphene have huge type-III TI gaps (about 50 and 100 meV) and large perpendicular magnetic anisotropy energies (5.3 and 11.5 meV). Therefore, they are ideal model systems for the observation of the QAHE at elevated temperature. We expect that the TI phase found here may be generalized to other materials. The pure and dissipationless spin current at the edges of Chern half metals may find significant applications in spintronics and quantum computating. # AUTHOR INFORMATION {#author-information .unnumbered} ## Corresponding Authors {#corresponding-authors .unnumbered} \(^*\)E-mail: jhu\@suda.edu.cn (J.H.). \(^*\)E-mail: wur\@uci.edu (R.W.).

{'timestamp': '2015-03-13T01:05:48', 'yymm': '1401', 'arxiv_id': '1401.5453', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5453'}

# Introduction {#Introduction} The long standing problem in nebular physics related to the abundance and temperature discrepancy between the results obtained from optical recombination lines (ORL) and collisionally excited lines (CEL) has been visited by many researchers in the last few decades. Several explanations for the general trend of obtaining higher abundances and lower temperatures from ORL than from CEL of the same ions have been put forward. One explanation is a multi-component model of planetary nebulae in which low temperature, high metallicity components (clumps) produce the ORL and are embedded in a high temperature, relatively low metallicity, hydrogen-rich component that produces the CEL. It has been suggested recently that this discrepancy may be largely explained by the assumption of a different electron energy distribution, a \(\kappa\)-distribution, rather than the  (MB) distribution which is traditionally accepted as the dominant distribution in the low density plasma found in the planetary nebulae. The debate about the electron energy distribution in astronomical objects, including planetary nebulae, is relatively old and dates back to at least the 1940s when Hagihara suggested that the distribution of free electrons in gaseous assemblies deviates considerably from the MB distribution. This was refuted by who, on the basis of a detailed quantitative balance analysis, concluded that any deviation from the Maxwellian equilibrium distribution is very small. The essence of their argument is that for typical planetary nebulae conditions, the thermalization process of elastic collisions is by far the most frequent event and typically occurs once every second, while other processes that shift the system from its thermodynamic equilibrium, like inelastic scattering with other ions that leads to metastable excitation or recapture, occur at much larger time scales estimated to be months or even years. Bohm and Aller also indicated the significance of any possible deviation from a Maxwellian distribution on derived elemental abundances. However, nobody seems to have considered non-MB electron distributions as a possible solution for the ORL/CEL discrepancy problem until the recent proposal of. In the light of this suggestion, it seems appropriate to ask whether there is any empirical evidence from observation for departures from MB energy distributions in nebulae. A preliminary investigation of this possibility has been given recently by and who proposed a method based on using atomic dielectronic recombination (DR) theoretical data in conjunction with observational line fluxes to sample the free electron distribution and compare to the theoretical distributions, i.e.  and \(\kappa\). For all the objects studied by the limited observational data analyzed were found to be better described by a single MB distribution than a \(\kappa\)-distribution but the uncertainties were sufficient that a \(\kappa\)-distribution could not be conclusively excluded. A two-component model incorporating low temperature material was also found to not fit the data as well as a single MB model. The current paper approaches this problem by modeling the magnitude of the Balmer discontinuity and nearby Balmer continuum of the planetary nebula . This nebula is an extreme example of the ORL/CEL discrepancy trend since its abundance discrepancy factor (ADF), defined as the ratio of ORL abundance to CEL abundance, reaches an exceptionally high value between 68-84. An exceptionally high temperature discrepancy between the CEL and ORL results, which differ by about an order of magnitude, has also been obtained for this nebula. The shape of the Balmer continuum is principally determined by a convolution of the \(n=2\) hydrogen recombination cross-section and the free electron energy distribution. The shape of the continuum therefore offers the possibility of gaining information about the free electron energy distribution. We shall see below that in the case of  the continuum shape is primarily determined by the free electron energy distribution. In section [2](#Hf2-2){reference-type="ref" reference="Hf2-2"} we outline the properties of  and in section [3](#Theory){reference-type="ref" reference="Theory"} we deal with the theory describing the continuum processes and new results relating to recombination using a \(\kappa\)-distribution. The fitting procedure and the resulting fits to the spectral data with MB and \(\kappa\) distributions are described in section [4](#Method){reference-type="ref" reference="Method"} and our conclusions in section [5](#Conclusions){reference-type="ref" reference="Conclusions"}. # {#Hf2-2}  is a faint southern highly symmetric low density planetary nebula with a central cavity and interior disc-shaped multi-shell structure. Its spectrum includes carbon, nitrogen, oxygen, neon, sulphur and argon lines as well as helium.  seems to have a high abundance in He and C with a high C/O ratio and an exceptionally strong C \(\lambda\)`<!-- -->`{=html}4267 Å signature. The nebula, which is in the galactic bulge at a distance of about 4.0-4.5 kpc, has a spectrally-varying close binary central system with a period of about 0.40 day.  has common spectral features with old novae like DQ Her , and therefore being a planetary nebula may be disputed. This unusual nebula is marked by a number of exceptional features.  exhibits a very high ADF, possibly the highest recorded for a planetary nebula, with very strong ORL emission.  also has the very large C/O abundance ratio of about 19. A third unusual feature is a very large Balmer jump associated with a rapid fall in the Balmer continuum intensity towards shorter wavelengths. An estimation based on the magnitude of the Balmer discontinuity indicates an electron temperature of about 780-1000 K in sharp contrast to the forbidden lines estimation of about 8800 K from an \[O\] line. The deduced Balmer discontinuity temperature is one of the lowest observed for a planetary nebula. These findings led to conclude that a multi-component model of cold hydrogen-deficient knots embedded into a metal-poor nebula may be inevitable to explain these exceptional observations. With regard to ORL electron temperature derivations for , obtained a mean electron temperature of \(<600\) K from C lines, while derived an electron temperature of about 630 K and a temperature of about 3160 K from O lines. As for the stellar temperature of this nebula, it is estimated by between 50000--67000 K, while set the temperature of the two stars (assuming a binary system which seems to be largely accepted) in their model to 67000 K and 7500 K. The observational data of  which is used in the present paper were obtained in 2001 by a Boller and Chivens long-slit spectrograph mounted on the ground-based 1.52 m European Southern Observatory telescope located in La Silla Chile. More details about this data set can be found in. The observed spectrum is shown in Figure [\[ContFitSeg\]](#ContFitSeg){reference-type="ref" reference="ContFitSeg"}, with the measured flux normalized to Balmer H11, an apparently unblended line close to the Balmer discontinuity. # Theory {#Theory} To model the magnitude of the Balmer discontinuity and the shape of the Balmer continuum we consider the contributions to the continuum spectrum from recombination of H\(^+\) and He\(^+\), H 2-photon emission and the underlying scattered stellar continuum. Recombination of He\(^{2+}\) is neglected due to the very low He\(^{2+}\)/H\(^+\) fraction in  which is estimated to be about 0.002. The recombination of an atomic ion \(X^+\) with an electron, \(e^-\) of energy \(E\) to a state \(X^*\) of the recombined ion, \[X^+ + e^-\rightarrow X^* +h\nu,\label{genrecomb}\] gives rise to continuous emission with emission coefficient (energy emitted per unit time per unit frequency and per unit particle densities) \[\gamma^*(\nu) = \frac{h^3 \alpha^3 c}{32\pi^2 m_e a_0^2} \frac{\omega^*}{\omega^+} \left( \frac{h\nu}{R} \right)^3 \left( \frac{R}{E} \right)^{\frac{1}{2}} f(E)\ \sigma^*_{\nu} \label{gamma}\] where, \(\nu\) is the frequency of the emitted photon, \(\omega^+\) and \(\omega^*\) are the statistical weights of the initial and final states respectively, \(f(E)\) is the free electron energy distribution function, \(\sigma^*_{\nu}\) is the photoionization cross-section for the inverse process of Equation [\[genrecomb\]](#genrecomb){reference-type="ref" reference="genrecomb"} and \(R\) is the Rydberg energy constant. The other symbols have their usual meanings. In the present paper, the photoionization cross-sections for states of H and He are taken from and respectively as described in more detail by. In this paper we consider two possible forms for \(f(E)\), the Maxwell-Boltzmann distribution \[f_{_{\rm MB}}(E,T) = \frac{2}{\left(kT\right)^{3/2}}\sqrt{\frac{E}{\pi}}e^{-\frac{E}{kT}}\label{ElDiEq2},\] where \(T\) is the electron temperature, and the \(\kappa\)-distribution given by \[f_{\kappa}\left(E, T_{\kappa}\right)=\frac{2\sqrt{E}}{\sqrt{\pi(kT_{\kappa})^3}}\frac{\Gamma\left(\kappa+1\right)}{(\kappa-\frac{3}{2})^{\frac{3}{2}}\Gamma\left(\kappa-\frac{1}{2}\right)}\left(1+\frac{E}{(\kappa-\frac{3}{2}) kT_{\kappa}}\right)^{-\left(\kappa+1\right)} \label{kappaEq}\] where \(\kappa\) is a parameter defining the distribution, \(\Gamma\) is the gamma-function for the given arguments, and \(T_{\kappa}\) is a characteristic temperature. Note that \(f_{\kappa}\) becomes a MB distribution with \(T_{\kappa} \rightarrow T\) as \(\kappa \rightarrow \infty\). The normalization of the continuum flux to H11 flux requires that the effective recombination coefficients for H11 should be calculated both with a Maxwell-Boltzmann distribution and a \(\kappa\)-distribution. The hydrogen line emissivities tabulated by and were calculated assuming that the free electron energy distribution is described by a Maxwell-Boltzmann distribution for all physical processes between bound and continuum states involving free electrons. Here we use the techniques and computer codes described in the last two references and extend them to include a \(\kappa\)-distribution. At the electron number densities typical of photoionized nebulae (\(10^2-10^5\) cm\(^{-3}\)) the populations of the low-lying states of H are determined primarily by recombination and radiative cascading and are relatively insensitive to the ambient electron density. Hence any error introduced into the calculation of collision rates between high-\(n\) states caused by the use of a Maxwell-Boltzmann distribution rather than a \(\kappa\)-distribution should have minimal effect at nebular densities. We therefore make the approximation of computing the direct recombination coefficients to all the individual levels using a \(\kappa\)-distribution but retain a Maxwell-Boltzmann distribution for the energy and angular momentum changing collisions among the higher-\(n\) states. This should provide a good approximation for the H11 emissivity. For a general free electron energy distribution, the recombination coefficient to a state X\(^*\) is given by \[\alpha^*_{RC} = \frac{R^{\frac{5}{2}}}{\sqrt{2}c^2 m_e^{\frac{3}{2}}} \frac{\omega^*}{\omega^+} \int_0^\infty \left( \frac{h\nu}{R} \right)^2 \left( \frac{R}{E} \right)^{\frac{1}{2}} \sigma^*_{\nu}\ f(E)\ {\rm d}\left(\frac{E}{R} \right) \label{alphaEq}\] On solving the collisional-radiative recombination problem for hydrogen we obtain the recombination coefficients to all levels and the effective recombination coefficients \(\alpha_{\rm eff}(\lambda)\) for a transition of wavelength \(\lambda\). We define a line emission coefficient \(\epsilon(\lambda)\) as the energy emitted per unit volume per unit time for unit ion and electron density, so that \[\epsilon(\lambda) = \alpha_{\rm eff}(\lambda)\ \frac{hc}{\lambda}.\] derive the following approximate expression for converting recombination coefficients calculated with a Maxwell-Boltzmann electron energy distribution to those applicable with a \(\kappa\)-distribution, \[x(\kappa) \equiv \frac{\alpha_{\kappa}(\lambda)}{\alpha_{\rm MB}(\lambda)} = \frac{\epsilon_{\kappa}(\lambda)}{\epsilon_{\rm MB}(\lambda)} = \frac{(1-\frac{3}{2\kappa})\Gamma(\kappa+1)}{(\kappa-\frac{3}{2})^{\frac{3}{2}}\Gamma(\kappa-\frac{1}{2})}.\label{MBkappa}\] To obtain this result, use the fact that the photoionization cross-section falls approximately as \((\nu_0/\nu)^3\) for \(\nu > \nu_0\), where \(\nu_0\) is the threshold frequency for photoionization. so that the integrand in Equation [\[alphaEq\]](#alphaEq){reference-type="ref" reference="alphaEq"} contains a \(1/\nu\) term. simplify the integral by moving the \(1/\nu\) term outside of the integral, making it possible to carry out the integration analytically. In practice the integral is a convolution of the free-electron energy distribution with the \(1/\nu\) term, which is energy dependent. At low temperatures the rapid fall in the electron energy distribution function with increasing energy means that it is a good approximation to neglect the energy dependence of the frequency term. At higher temperatures, both terms are essential in the calculation of the recombination coefficient. The expression for \(x(\kappa)\) in Equation [\[MBkappa\]](#MBkappa){reference-type="ref" reference="MBkappa"} is independent of temperature, density and transition. This approximation will fail at sufficiently high temperatures where the effects of the free-electron energy distribution and the frequency dependent term become comparable. Using the approximate function \(x(\kappa)\), we can express the results of our more complete collisional-radiative treatment for \(\epsilon(\lambda)\) in terms of a correction factor \(y(\lambda,T,\kappa)\) as follows \[\epsilon_{\kappa}(\lambda) = \epsilon_{\rm MB}(\lambda)\ x(\kappa)\ y(\lambda,T,\kappa). \label{epsEq}\] The values of \(\epsilon_{\rm MB}\) can be obtained from and. In Table [1](#H11MBkappa){reference-type="ref" reference="H11MBkappa"} we tabulate values of \(y(\lambda,T,\kappa)\) for H11 corresponding to various values of \(\kappa\) at a range of temperatures. We also tabulate in the last line \(x\) as a function of \(\kappa\). Note that, in principle, \(y\) also depends on electron density but in practice it is very insensitive to electron density for densities up to \(10^5\) cm\(^{-3}\), so we only tabulate values calculated at \(10^3\) cm\(^{-3}\). ::: ## Fits with Maxwell-Boltzmann distributions It is clear from Figure [\[ContFitSeg\]](#ContFitSeg){reference-type="ref" reference="ContFitSeg"} that the magnitude of the Balmer discontinuity and the slope of the Balmer continuum shortward of the discontinuity cannot be explained by emission from hydrogen and helium at the forbidden line temperature. This has already been discussed by who estimated that the shape of the continuum implies that the emitting region has a temperature of \(\approx 900\) K. In Figure [\[SinglefreeMB\]](#SinglefreeMB){reference-type="ref" reference="SinglefreeMB"} we show the result of making a fit to all three spectral segments shown in Figure [\[ContFitSeg\]](#ContFitSeg){reference-type="ref" reference="ContFitSeg"} using a single MB electron energy distribution in which the temperature is a free parameter. The background continuum is computed using the parameters from fit (a) in Table [3](#confit){reference-type="ref" reference="confit"}. The temperature of best fit is 1334 K and although the continuum model is improved compared to Figure [\[ContFitSeg\]](#ContFitSeg){reference-type="ref" reference="ContFitSeg"}, it matches neither the magnitude of the discontinuity nor the slope adequately. As discussed earlier, one resolution of the ORL/CEL abundance and temperature discrepancy in nebulae that has been proposed is that there are relatively cold high-metallicity knots embedded in hotter lower abundance gas. We therefore attempt to match the continuum with two components having different temperatures. We set the temperature of one component equal to a typical forbidden line temperature for this object of 8800 K and allow the other to vary. The continuum emissivity has only weak density dependence so we choose representative electron densities for the two components of 1000 cm\(^{-3}\) for the forbidden line region and 5000 cm\(^{-3}\) for the cold material. The relative emission measure of the two components is then allowed to vary by making the volume ratio of the two components a free parameter. In Figure [\[2MB1free\]](#2MB1free){reference-type="ref" reference="2MB1free"} we show the resulting fit, which is excellent. The optimum temperature of the cold component is 540 K and the fraction of the total volume occupied by the cold component is 0.00706. Hence the cold component has an emission measure, which is proportional to the product of the volume fraction and particle density squared, that is 15.0% of the total recombination line and continuum emission measure. ## Fit with \(\kappa\)-distribution An alternative possible resolution of the ORL/CEL problem is that there is no separation of the emitting material into high and low temperature components with significantly different abundances but rather a single medium in which the electron energy distribution is non-Maxwellian, with a \(\kappa\)-distribution being a possible candidate. The Balmer continuum close to the Balmer discontinuity provides a means of testing the validity of this proposal for the very low energy part of the distribution function. In Figures [\[T100003kappa\]](#T100003kappa){reference-type="ref" reference="T100003kappa"} and [\[T10003kappa\]](#T10003kappa){reference-type="ref" reference="T10003kappa"} we illustrate how the magnitude of the Balmer discontinuity and shape of the Balmer continuum change as \(\kappa\) is varied for two different temperatures. Compared to a MB, \(\kappa\)-distributions have more particles at the lowest and highest energies and less at intermediate energies. At low energies the \(\kappa\)-distribution function thus falls more rapidly with increasing energy than a MB distribution. This can be seen in Figure [\[T10003kappa\]](#T10003kappa){reference-type="ref" reference="T10003kappa"}, for example, with the magnitude of the Balmer discontinuity increasing and the slope of the Balmer continuum becoming steeper as \(\kappa\) decreases. We now fit the chosen three spectral segments to a \(\kappa\)-distribution of which the temperature and value of \(\kappa\) are free parameters and the final values that we obtained for these parameters are \(\kappa=2.11\) and \(T=4640\) K. Figure [\[1kappa\]](#1kappa){reference-type="ref" reference="1kappa"} shows a plot of this fit. On visual inspection the best-fit \(\kappa\) distribution where \(\kappa\) and temperature are allowed to vary is significantly less good than the two-component MB fit. It should be remarked that the optimized \(\kappa\) obtained from this fitting exercise is very different to values proposed by which usually fall in the range 10-20. For comparison, we include in Figure [\[1kappa\]](#1kappa){reference-type="ref" reference="1kappa"} a fit carried out with a fixed value of \(\kappa=10\). The temperature was allowed to vary and the best fit value was \(1665\) K. In the next section we attempt to quantify the relative quality of these fits. ## Statistical Analysis {#Discussion} We wish to address the question of whether the fit to the continuum using a \(\kappa\)-distribution is significantly worse than the two-component MB fit. A visual inspection of Figures [\[2MB1free\]](#2MB1free){reference-type="ref" reference="2MB1free"} and [\[1kappa\]](#1kappa){reference-type="ref" reference="1kappa"} shows that while the-two-component fit provides a very good match to the three segments of the continuum used for fitting, the best \(\kappa\)-distribution fit underestimates the magnitude of the Balmer discontinuity and also underestimates the steepness of the decline at shorter wavelengths. We will use a comparison of \(\chi^2\) values for these two fits to quantify the relative goodness of fit. To estimate \(\chi^2\) we need an estimate of the uncertainties on the observational data, which by inspection and from instrumental considerations are not independent of wavelength, being much larger at the shortest wavelengths. We therefore compute separate rms deviations of the data from the two-component MB fit for each of the three wavelength segments. Since we use deviations from the two-component fit to define the data uncertainties it follows that the value of \(\chi^2\) for the fit, summed over the three wavelength segments, will be given by \[\chi^2 = N_{df} \equiv N_{dp}-N_{fp} = 237,\] where \(N_{df}\) is the number of degrees of freedom, \(N_{dp}\) is the number of data points in the three segments and \(N_{fp}\) is the number of free parameters. We now use the estimates of the data uncertainties from the two-component MB fit to compute \(\chi^2\) for the best fit \(\kappa\)-distribution and we find \(\chi^2=433\). We wish to test the null hypothesis that the \(\kappa\)-distribution is an equally good fit to the data as the two-component MB fit. The following square-root transformation of \(\chi^2\) , \[t = \sqrt{2\chi^2}-\sqrt{2N_{df}},\] is such that for large \(N_{df}\), the transformed distribution is approximately normal with zero mean and unit standard deviation. For the best fit \(\kappa\)-distribution we find that \(t=7.67\) standard deviations from the zero mean, indicating that the probability that the null hypothesis is true is vanishingly small. Thus, a two-component  model is highly favored compared to the single-component \(\kappa\) model. The single-component MB model is even less likely to be correct than either of the two-parameter models based on the same \(\chi^2\) analysis, giving \(t=14.4\) standard deviations. # Conclusions {#Conclusions} The analysis of the Balmer continuum spectrum of  seems to indicate the signature of a two-component nebula with two different temperatures. Assuming that one component has a temperature of 8800 K, we find that the best fit to the data is obtained by adding a second component with temperature of 540 K and emission measure 15.0% of the total. We also modeled the continuum with a single component with \(\kappa\)-distributed free electron energies, but on the basis of a \(\chi^2\) test we found this model to be significantly less likely to be correct. In the case of \(\kappa\)-distribution models, it is important to note that modeling the Balmer continuum only samples the free electron energy distribution at the lowest energies, below the mean energy, and therefore gives no information about departures from MB at energies above the mean energy which might affect the excitation of nebular forbidden lines, for example. The extreme nature of  and the similarity of some of its spectral features to old novae like DQ Her may also cast some doubt on the applicability of the conclusions reached to the properties of planetary nebulae in general.

{'timestamp': '2014-01-23T02:07:55', 'yymm': '1401', 'arxiv_id': '1401.5630', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5630'}

# Introduction When two thermodynamic systems are placed in contact, the expected result is that the systems equilibrate with one other. In this Letter, we show that the opposite may also occur: starting from systems which individually equilibrate and placing them in contact may instead drive the combined system *out* of equilibrium, leading to a *nonequilibrium stationary state* (NESS) with a nontrivial internal circulation flow. We show that this counterintuitive behavior occurs already in one of the simplest diffusive systems, biased random walkers, where this effect is amenable to an exact analysis. It is well known that a diffusive process subject to a constant drift, or bias, towards a reflecting wall reaches an equilibrium state, characterized by an exponential distribution decaying away from the wall, known as the barometric formula. Being an equilibrium distribution, this can be interpreted as the configuration factor of a Boltzmann distribution, where the constant bias is due to a force derived from a linear potential, e.g., gravity, or an electric field, and the diffusion constant is related to the temperature through the Einstein relation. In particular, these results form the basis of the experimental results of Perrin showing the "objective reality of molecules" via Brownian motion under gravity. On the other hand, the nature of nonequilibrium steady states has attracted much attention; see, e.g., Ref. . Although the statistical properties of these systems are independent of time, currents driven by external forces are sustained within the systems and detailed balance is not satisfied. In this Letter, we study the effect of coupling together two biased diffusive systems with reflecting boundaries. Each of these systems individually equilibrates, and coupling them at a single site produces a global equilibrium state. However, when the systems are coupled appropriately, we show that the combined system in general reaches a nonequilibrium steady state (NESS) with a non-vanishing current, i.e., a constant particle circulation, with particles traveling against the force in one of the columns and returning through the other. For definiteness, we present a discrete-time model, in which all walkers jump at each time step; however, our results can easily be transposed to a continuous-time version. Results for periodic models have been obtained by Kolomeisky and co-workers. # Model and equilibrium states ## Model Our model consists of two vertical, semi-infinite discrete columns, with sites at height \(i\) having concentrations \(\cc{1}{i}{n}\) and \(\cc{2}{i}{n}\), respectively, at time step \(n\); see Fig. [\[fig:sketch\]](#fig:sketch){reference-type="ref" reference="fig:sketch"}. Reflecting boundaries are placed at the bottom of each column, which we take to be at \(i=0\). In each column \(\nu=1,2\), we consider biased random walkers in discrete time, with probabilities \({p}_\nu\) to jump upwards and \({q}_\nu=1-{p}_\nu\) to jump downwards at each time step. To enforce the reflecting boundary condition, walkers which try to jump downwards from site \(0\) remain at site \(0\). We believe that this is the simplest model in which the effect is found, allowing a complete analytical solution. However, as we discuss below, similar phenomena can be expected to arise whenever two or more systems with spatially varying equilibrium states are coupled at several places. When the columns are not coupled, the following master equation describes the time evolution of the concentrations: \[c_\nu(i ,n+1) = {p}_\nu c_\nu(i-1,n) + {q}_\nu c_\nu(i+1,n), \quad j \neq 0\] with boundary conditions \[c_\nu(0,n+1) = {q}_\nu c_\nu(0,n) + {q}_\nu c_\nu(1,n).\] If \({q}_\nu \geq {p}_\nu\), i.e., if the bias is directed towards the reflecting boundary at the origin, then the physical (i.e., normalizable) stationary solutions of these equations are: \[c_\nu(i) = {K}_\nu \left({p}_\nu/{q}_\nu\right)^i,\] where the \({K}_\nu\) are normalization constants. The crucial feature of these solutions is that the vertical current \(J_\nu(i) := {p}_\nu c_\nu(i-1)-{q}_\nu c_\nu(i)\) between sites within either column vanishes identically, so that the system is in *equilibrium*. ## Coupling We now couple the columns together so that they may exchange particles at a *single* height \(j\), where particles have the same probability \(\alpha\) to jump "horizontally" from one column to the other in either direction (see Fig. [\[fig:sketch\]](#fig:sketch){reference-type="ref" reference="fig:sketch"}). The equations describing the system are then modified in the vicinity of site \(j\) to take account of the additional current due to the coupling: \[\begin{aligned} c_\nu(j+1,n+1) = (1-\alpha){p}_\nu c_\nu(j,n) + {q}_\nu c_\nu(j+2,n); \nonumber \\ c_\nu(j-1,n+1) = {p}_\nu c_\nu(j-2,n) + (1-\alpha){q}_\nu c_\nu(j,n); \nonumber \\ c_1(j,n+1) = {p}_1 c_1(j-1,n) + {q}_1 c_1(j+1,n) +\alpha c_2(j,n); \nonumber \\ c_2(j,n+1) = {p}_2 c_2(j-1,n) + {q}_2 c_2(j+1,n) +\alpha c_1(j,n). \nonumber \end{aligned}\] There is then a *unique* stationary solution for the whole system, given by \[c_\nu(i)=K \cases{ (1-\alpha)\left({p}_\nu/{q}_\nu\right)^{i-j} & if \(i\neq j\); \\ 1 & if \(i=j\),}\] for \(\nu=1, 2\), with a single normalization constant \(K\). Note that the concentrations in the two sites at height \(j\), where interchange is allowed, are equal, giving a zero total horizontal current there. Thus, this is again an equilibrium distribution with vanishing current everywhere, showing that such a coupling procedure is natural, and allows the systems to equilibrate with each other. # Non-equilibrium stationary state However, coupling at *more than one* site instead forces the system to settle into a nonequilibrium stationary state, with non-vanishing currents. To show this, in what follows we consider the case of horizontal coupling at every height, with symmetrical probability \(\alpha\) of jumping between columns. In the stationary state, the concentrations satisfy the following set of coupled linear equations at sites \(i \neq 0\): \[\begin{aligned} c_1(i) = \left(1-\alpha\right)\left[ {p}_1 c_1(i-1) + {q}_1 c_1(i+1) \right] + \alpha c_2(i), \label{eq:ci1} \label{eq:system-first}\\ c_2(i) = \left(1-\alpha\right) \left[ {p}_2 c_2(i-1) + {q}_2 c_2(i+1) \right] + \alpha c_1(i) \label{eq:ci2}. \end{aligned}\] At height \(0\), the reflecting boundary conditions give the following: \[\begin{aligned} c_1(0) = \left(1-\alpha\right)\left[{q}_1 c_1(0) + {q}_1 c_1(1)\right] + \alpha c_2(0) \label{eq:boundary_1},\\ c_2(0) = \left(1-\alpha\right)\left[{q}_2 c_2(0) + {q}_2 c_2(1)\right] + \alpha c_1(0)\label{eq:boundary_2}. \end{aligned}\] To solve the system [\[eq:system-first\]](#eq:system-first){reference-type="eqref" reference="eq:system-first"}--[\[eq:boundary_2\]](#eq:boundary_2){reference-type="eqref" reference="eq:boundary_2"}, we introduce *generating functions*, a discrete version of the Laplace transform: \[\hat{C}_\nu(z) := \sum_{i=0}^{\infty} c_\nu(i) z^i, \qquad \nu=1,2. \label{eq:generating-fns}\] Multiplying [\[eq:ci1\]](#eq:ci1){reference-type="eqref" reference="eq:ci1"} and [\[eq:ci2\]](#eq:ci2){reference-type="eqref" reference="eq:ci2"} by \(z^i\), summing over \(i\) from \(1\) to \(\infty\), and using the boundary conditions leads to two simultaneous linear equations for the generating functions in terms of the values of the concentrations \(c_1(0)\) and \(c_2(0)\) at the boundary, which have yet to be determined: \[\left( \begin{array}{cc} f_1(z) &-\alpha\\-\alpha & f_2(z) \end{array} \right) \left( \begin{array}{c} \hat{C}_1(z) \\ \hat{C}_2(z) \end{array} \right) = \left( \begin{array}{c} g_1(z) \\ g_2(z) \end{array} \right),\] where \(f_\nu(z) := 1-(1-\alpha)[{p}_\nu z + \frac{{q}_\nu}{z}]\) and \(g_\nu(z) := (1-\alpha){q}_\nu (1-\frac{1}{z}) c_\nu(0)\). Solving yields \[\begin{aligned} \fl \hat{C}_1(z)=\frac{ [z-(1-\alpha) ({p}_2z^2+{q}_2)] {q}_1 c_1(0) + \alpha z {q}_2 c_2(0)} {\alpha z\left[ {q}_1+{q}_2-z({p}_1+{p}_2) \right] -(1-\alpha) (1-z) ({q}_1-z{p}_1) ({q}_2-z{p}_2)}, \label{eq:chachan} \end{aligned}\] and an analogous equation for \(\hat{C}_2(z)\) with the subscripts \(1\) and \(2\) interchanged in the numerator (the denominator being unchanged). This gives (in principle) an exact solution for the stationary concentration distributions, obtained by inverting Eq. [\[eq:generating-fns\]](#eq:generating-fns){reference-type="eqref" reference="eq:generating-fns"}: \(c_\nu(i) = \frac{1}{i!} \hat{C}^{(i)}_\nu(0)\), where \(\hat{C}_\nu^{(i)}\) denotes the \(i\)th derivative; or, equivalently, by expanding in formal power series in \(z\) and equating coefficients. The vertical current in column \(\nu\) between site \(j\) and site \(j+1\) is given by \(J_\nu(j) := (1-\alpha)[{p}_\nu c_\nu(j)-{q}_\nu c_\nu(j+1)]\). The corresponding generating functions follow from Eq. [\[eq:chachan\]](#eq:chachan){reference-type="eqref" reference="eq:chachan"}: \[\begin{aligned} \fl \hat{J}_1(z)=\alpha(1-\alpha) \frac{({q}_2-z{p}_2){q}_1 c_1(0)-({q}_1 -z{p}_1) {q}_2 c_2(0)}{\alpha z\left[{q}_1+{q}_2-z({p}_1+{p}_2)\right] -(1-z)(1-\alpha)[{q}_1-z{p}_1][{q}_2-z{p}_2]} \label{eq:chan} \end{aligned}\] and \(\hat{J}_2(z) =-\hat{J}_1(z)\), reflecting the fact that the system is in a steady state. However, neither current (in either column) can be identically zero unless the corresponding hopping probabilities are equal, \(p_1 = p_2\), in which case the steady state is an equilibrium state. Rather, there is a circulation through the system, moving upwards in the column with larger \(p_{\nu}\) and downwards in the other column. Figures [\[fig:concs\]](#fig:concs){reference-type="ref" reference="fig:concs"} and [\[fig:currents\]](#fig:currents){reference-type="ref" reference="fig:currents"} show this phenomenon for a representative set of parameters, giving the concentration profiles, and vertical and horizontal current, respectively. The stationary solution was obtained by numerically iterating the master equation for the time evolution of the \(c_\nu(i)\), using a finite system with a reflecting boundary condition also at the top, confirming that the concentration profiles have exponential tails. It may be checked that this numerical solution agrees with the exact solution described below when the corresponding parameter values are substituted in the exact expressions Other relevant quantities can also be easily calculated. For example, since there is a sustained current in the system, it must be driven by thermodynamic forces. These forces can be related to the respective hopping probabilities \({p}_\nu\) and \({q}_\nu\) in each column via \(f_\nu = \ln ({p}_\nu / {q}_\nu)\); this identification arises from the fact that when \(\alpha=0\), the concentrations reach equilibrium distributions of the form \(c_\nu(j) {\sim} ({p}_\nu/{q}_\nu)^j=\exp [-j \, \ln ({q}_\nu/{p}_\nu) ]\), so that \(j \, \ln({q}_\nu/{p}_\nu)\) is a potential from which the force \(f_\nu\) is derived. Then, the amount of "work" dissipated can be calculated from the generating function \(\hat{W}(z)=\ln({p}_1 /{q}_1) \hat{J}_1(z)+\ln({p}_2 /{q}_2) \hat{J}_2(z)\), giving that the total work dissipated within the system is \[\begin{aligned} \fl \hat{W}(z=1) = (1-\alpha)\left[\ln({q}_2 {p}_1/{p}_2 {q}_1)\right] \frac{\left [({q}_2-{p}_2){q}_1 c_1(0)-({q}_1 -{p}_1) {q}_2 c_2(0) \right]} {\left[ {q}_1+{q}_2-({p}_1+{p}_2) \right] }. \end{aligned}\] # Exact solution and asymptotics To evaluate and make use of all of the above analytical expressions, we must determine the hitherto unknown constants \(c_1(0)\) and \(c_2(0)\). Taking the limit \(z \to1\) in Eq. [\[eq:chachan\]](#eq:chachan){reference-type="eqref" reference="eq:chachan"} gives \(\lim_{z \to 1} \hat{C}_1(z) = \lim_{z \to 1} \hat{C}_2(z)\), i.e., the total probabilities in each column are equal in the stationary state. This is a result of the nature of the coupling used, and expresses the fact that the *total* current from one column to the other must be \(0\) in the stationary state. (Note that the (horizontal) current from one column to the other, at height \(i\), is \(\alpha [c_1(i)-c_2(i)]\).) In what follows, we assume the total normalization to be equal to one, so that \(\hat{C}_1(z=1) = \hat{C}_2(z=1) = \frac{1}{2}\). Evaluating this equality gives one relation between the unknown constants \(c_1(0)\) and \(c_2(0)\): \[{q}_1 c_1(0)+ {q}_2 c_2(0)= 1-(p_1 + p_2) = {q}_1-{p}_2.\] To determine another relation between them, and hence find their precise values, we use the face that the generating functions \(\hat{C}_{\nu}\) be analytic in the region \(|z|\leq 1\). Indeed, \(c_\nu(j)\) are probabilities, and thus take values in the interval \([0,1]\), so that generating functions \(\hat{C}_\nu(z)\) can have no poles in the interval \(z\in [0,1]\). To impose this constraint, we must determine the zeros of the denominator \(\Delta(z)\) in the expression [\[eq:chachan\]](#eq:chachan){reference-type="eqref" reference="eq:chachan"} for the \(\hat{C}_{\nu}(z)\), given by \[\begin{aligned} \fl \Delta(z) := \alpha z\left[{q}_1+{q}_2-z({p}_1+{p}_2)\right] -(1-z)(1-\alpha)({q}_1-z{p}_1)({q}_2-z{p}_2). \end{aligned}\] Since \(\Delta(0) =-{q}_1 {q}_2 (1-\alpha) < 0\) and \(\Delta(1) = 2 \alpha ({q}_1 + {q}_2-1)\), a sufficient condition for \(\Delta(z)\) to have a zero in the interval \((0,1)\) is that \({q}_1 + {q}_2 > 1\). We denote the zeros of \(\Delta\) by \(z_0 \in (0,1)\) and \(z_1, z_2 \ge 1\). For \(\hat{C}_1(z)\) and \(\hat{C}_2(z)\) to have no poles in \((0,1)\), the numerator in Eq. ([\[eq:chachan\]](#eq:chachan){reference-type="ref" reference="eq:chachan"}) must thus also vanish at \(z_0\). Since \(\Delta(z)\) is a cubic polynomial, one can find exact explicit expressions for \(z_0\) in terms of the parameters of the system. However, the general resulting expressions are cumbersome and unenlightening, and we do not give them here. Nonetheless, simple perturbative expressions can be found, for example, in the limit of small \(\alpha\), for which we have, to first order in \(\alpha\), \[z_0 \simeq 1-\alpha\frac{({q}_1+{q}_2)-({p}_1+{p}_2)}{({q}_1-{p}_1)({q}_2-{p}_2)},\] provided \({q}_1 > {p}_1\) and \({q}_2 > {p}_2\). If, instead, \({p}_1 > {q}_1\), say, and \({q}_2\) is large enough to ensure that \(\Delta(1) > 0\), then to first order \[z_0 \simeq \frac{{q}_1}{{p}_1}-\alpha\frac{{q}_1}{{p}_1-{q}_1},\] whereas if \({q}_1 = {p}_1\) then \(z_0 \simeq 1-\sqrt{2\alpha}\). Imposing the condition of analyticity on, say, \(\hat{C}_1(z)\) then provides the missing condition to determine \(c_1(0)\) and \(c_2(0)\): \[{q}_1 c_1(0) + \alpha {q}_2 c_2(0)=0.\] Note that, given the form of the denominator, had we chosen instead to impose the requirement of analyticity at \(z_0\) on \(\hat{C}_2(z)\), we would have arrived at the same condition. Equivalently, we could impose that the generating function for the current, Eq. ([\[eq:chan\]](#eq:chan){reference-type="ref" reference="eq:chan"}), be analytic at \(z_0\). This leads to \[c_1(0)=\frac{1}{2}\left[\frac{({q}_1+{q}_2)-({p}_1+{p}_2)}{({q}_1+{q}_2)-z_0 ({p}_1+{p}_2)}\right]({q}_1-z_0{p}_1)\] and a corresponding equation for \(c_2(0)\), which provide the complete solution to the problem in terms of the root \(z_0\). Then, for example, substitution of these expressions in the equation for the total work dissipated in the system gives \[\hat{W}(z=1)= \frac{\alpha z_0}{2} \left[\frac{{p}_1{q}_2-{p}_2{q}_1}{({q}_1-z_0{p}_1)({q}_2 -z_0{p}_2)}\right] \ln \left ( \frac{{q}_2{p}_1}{{q}_1{p}_2} \right).\nonumber\] The generating functions \(\hat{C}_\nu(z)\) of the concentrations, as shown in Eq. [\[eq:chachan\]](#eq:chachan){reference-type="eqref" reference="eq:chachan"}, are rational functions of \(z\), which correspond to concentration profiles formed by a linear combination of decaying exponentials \(z_1^{-j}\) and \(z_2^{-j}\). As is the case for \(z_0\), approximate expressions for these other roots at small values of \(\alpha\) can easily be calculated. For example, when \({p}_1 < {q}_1\) and \({p}_2 < {q}_2\), to first order in \(\alpha\) we have \[z_\nu \simeq \frac{{q}_\nu}{{p}_\nu}+\alpha\frac{{q}_\nu}{{q}_\nu-{p}_\nu},\qquad \nu=1,2.\] A special case occurs when one of the \({p}_\nu\) vanishes, say \(p_{2} = 0\). In this case, the cubic denominator \(\Delta(z)\) simplifies to a quadratic, and the exact expressions for its zeros become significantly simpler: \[z_{0,1} = \frac{1+\alpha q_{1} \pm \sqrt{(1+\alpha q_{1} )^2-4 p_{1} q_{1} (1-\alpha) } }{2 p_{1}},\] where \(z_{0} \in [0,1]\) is given by taking the minus sign for the radical. The \(\hat{C}_{\nu}(z)\) must vanish at \(z=z_{0}\), giving a single factor \(z-z_{1}\) in the denominator. The two concentration profiles are then exactly exponential for \(i>0\), differing only by a multiplicative factor and the concentration at height \(0\). # Conclusions In conclusion, by furnishing a complete analytical solution, we have shown that coupling together two simple biased diffusive systems can give rise to nonequilibrium steady state with a sustained current throughout the combined system, even though each of them separately reaches equilibrium. This effect arises due to the fact that the equilibrium profiles can be "matched" when the systems are coupled at just a single site, so that the current across that site vanishes, whereas if the coupling occurs at various sites, then the equilibrium distributions can no longer be made to match at all sites simultaneously. The same phenomenon thus occurs even with coupling at only two heights, \(j_1\) and \(j_2\). The density profiles and currents obtained numerically for a specific realization of this minimal case are shown in Fig. [\[fig:two-site-coupling\]](#fig:two-site-coupling){reference-type="ref" reference="fig:two-site-coupling"}. This case can also be solved analytically: below \(j_1\) and above \(j_2\), the profiles are exactly exponential, due to the absence of current there; these exponentials are matched with the exact solution between \(j_1\) and \(j_2\). Physical realizations of this phenomenon may be obtained, for example, with particles suspended in a fluid under the combined effect of gravity and an applied electric field, as in electrophoresis. If the particles are capable of capturing or losing a charge (or can be induced to do so), then they can transition from one charge state to the other. The drift induced by the field will then differ depending on the charge state of the particle. Note that our model can also be interpreted in this way, where the subindex describes the charge state of the particle, rather than its physical position. These and other realizations may require generalizing the analysis to the case in which there is asymmetric coupling, i.e., in which the probability of jumping from state 1 to state 2 differs from the reverse transition probability. The effect described in this work will nonetheless persist. Financial support from SEP-CONACYT grant CB-101246 and PAPIIT grants IN109111, IN116212 and IN117214 from DGAPA-UNAM is acknowledged.

{'timestamp': '2014-01-23T02:02:29', 'yymm': '1401', 'arxiv_id': '1401.5526', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5526'}

# Introduction As is well known, gravitational radiation induces relative displacements in a system of inertial test particles. Zel'dovich and Polnarev first noted that, within linearized gravity, the passage of a pulse of gravitational radiation can cause a permanent change in the relative displacement these particles. This effect is known as the *memory effect*. The net relative displacement, \(\Delta D^a\), after passage of the pulse for test masses with initial separation \(D^a\) can be expressed as \[\Delta D_a = \frac{1}{2}\Delta h^{\textnormal{TT}}_{ab} D^b \,,\] where \(\Delta h^{\textnormal{TT}}_{ab}\) is the net change in the metric perturbation in a transverse-traceless gauge. In linearized gravity, for gravitational radiation produced by a change in the motion of particle-like sources occurring in a localized region of spacetime, to leading order in \(1/r\) we have, \[\Delta h^{\textnormal{TT}}_{ab}= \frac{1}{r} \Delta\sum_A \frac{4M_A}{\sqrt{1-v_A^2}}\left[\frac{(v_A)_a (v_A)_b}{1-v_A\cos\theta_A}\right]^{\textnormal{TT}} \,, \label{nonrev}\] where \(A\) is an index labeling the source particles, which have mass \(M_A\) and velocity \(v_A\) at angle \(\theta_A\) with respect to the direction of the detector. The brackets \([...]^{\textnormal{TT}}\) represent the transverse-traceless part of the object within, and \(\Delta\) means the difference between the quantity at late and early times. In the case where there is an emission of a null-particle-like source of energy \(E\) moving in the \(z\)-direction and all other sources (before and after the emission) are non-relativistic, the memory effect for a detector at \((\theta, \phi)\) becomes \[\Delta D_a=\frac{E}{r}\frac{\sin^2\theta}{1-\cos\theta}\left(\theta_a \theta_b-\phi_a \phi_b \right)D^b \,, \label{nullmem}\] where \(\theta^a\) and \(\phi^a\) denote unit vectors in the \(\theta\) and \(\phi\) directions, respectively. Making use of a careful analysis of the nonlinear Einstein equation, Christodoulou found that there can be significant nonlinear contributions to the memory effect. Christodoulou's formula for the nonlinear contribution to the memory effect associated with the passage of a gravitational radiation to future null infinity is expressed in terms of an integral of the Bondi flux over future null infinity. For the case where the Bondi flux is highly localized in the \(z\)-direction and the integrated flux is \(E\), Christodoulou's formula reduces to [\[nullmem\]](#nullmem){reference-type="eqref" reference="nullmem"}. Thorne and Wiseman and Will soon interpreted Christodoulou's nonlinear memory effect as simply corresponding to the linear memory effect, but with the nonlinear effective stress-energy of gravitational waves replacing the "particle" sources. Further support for this interpretation can be found in the fact that a similar nonlinear memory effect occurs when a flux of electromagnetic radiation reaches null infinity, thereby showing that the nonlinear memory effect is not special to gravitational waves. Very recently, Bieri and Garfinkle have shown that the linear memory effect for null matter can be derived in close parallel to Christodoulou's derivation, thus further confirming that the nonlinear memory effect can be interpreted as being the same as the linear memory effect, with the effective stress energy of gravitational radiation replacing the ordinary stress-energy of null matter. However, there remains a puzzling aspect of the alternative derivations of the memory effect. In the derivation of the formula [\[nonrev\]](#nonrev){reference-type="eqref" reference="nonrev"} for the linear memory effect, one considers the retarded solution associated with sources. The linear memory effect is thereby seen to be simply an aspect of the gravitational radiation emitted by the sources. In particular, for slowly moving sources, what is relevant for producing a nontrivial memory effect is the net *change* in the time derivative of the quadrupole moment of the sources. If the time derivative of the quadrupole moment does not vary---such as for the case of a single particle moving on a timelike geodesic of Minkowski spacetime---there is no gravitational radiation and no memory effect. Similar results hold if one does not assume slow motion of the sources. By contrast, in Christodoulou's derivation of the nonlinear memory effect and in the Bieri and Garfinkle derivation of the linear memory effect for null matter, there is no allusion to emission by sources. If one examines these derivations, it would appear that all that is relevant to the memory effect is that there be a flux of gravitational radiation or null matter to future null infinity. This flux could just as well have originated from past null infinity as have been emitted by sources at some finite time. This would suggest[^1] that the null memory effect should be interpreted as being associated with simply the passage of null (effective) stress-energy to infinity---perhaps as a "tidal effect"---rather than as being caused by a burst of radiation associated with some "emission event" within the spacetime. We will attempt to gain insight into this issue by considering the simple problem of obtaining retarded solution to the wave equation in Minkowski spacetime with source given by a particle moving on a null geodesic. This directly yields the retarded solution of a scalar field with a null particle scalar charge source, and it also yields the corresponding retarded solution for the electromagnetic and linearized gravitational cases for appropriate choices of gauge. We now summarize our main results, after which we explain the relevance of our results to the interpretation of the memory effect. First, we find in section 2 that the retarded solution to the wave equation in Minkowski spacetime with a delta-function source on a complete null geodesic does not exist as a distribution. This difficulty is not due to the "nullness\" of the source but rather to its non-compactness. However, if we "cut off\" the source at a time \(t_0\) in the past, i.e., if we consider the scalar charge source \[\begin{aligned} S_0=q\delta(x)\delta(y)\delta(z-t)\Theta(t-t_0) \,, \label{scalsource} \end{aligned}\] the retarded solution is \[\varphi_0=\frac{q}{u}\Theta\left((t-t_0)-\sqrt{x^2+y^2+(z-t_0)^2}\right) \,, \label{phiret}\] where \[u \equiv t-z\] The right side of [\[phiret\]](#phiret){reference-type="eqref" reference="phiret"} is well defined as a distribution. This scalar field associated with a null source created at time \(t_0\) will produce a force on a test particle of scalar charge \(Q\) given by \(f^a=Q\partial^a\varphi_0\). Differentiation of the \(\Theta\)-function in [\[phiret\]](#phiret){reference-type="eqref" reference="phiret"} will yield a \(\delta\)-function term in \(f^a\) that will give the test particle a \(4\)-momentum "kick," which can be understood as being due to the radiation produced by the creation of the source at time \(t_0\). Although the limit as \(t_0 \to-\infty\) of expression [\[phiret\]](#phiret){reference-type="eqref" reference="phiret"} does not exist as a distribution, we shall show in section 2 that \(\lim_{t_0 \to-\infty} k^{[a} \partial^{b]} \varphi\) is well defined as a distribution, where \(k^a = t^a + z^a\) is the vector field that is parallel to the tangent to the source worldline. Thus, although the retarded field produced by a scalar charge moving on a (past and future complete) null geodesic is ill defined, we obtain a well defined \(4\)-momentum kick, modulo multiples of \(k^a\), from such a source. However, this kick differs from the kick resulting from a creation event at a finite time \(t_0\). As we shall see in section 3, the situation in the electromagnetic case is similar. Maxwell's equations in Lorenz gauge reduces to four scalar wave equations, so we can immediately write down the retarded solution for the vector potential \(A_a\) in terms of the retarded solution to the scalar wave equation. However, one important difference is that Maxwell's equations require conservation of charge, so one cannot simply create a charge at a finite time \(t_0\), as in [\[scalsource\]](#scalsource){reference-type="eqref" reference="scalsource"}. Nevertheless, one can consider a charge that sits "at rest" until time \(t_0\) and thereafter moves on a null geodesic. To order \(1/r\), the radiation from this sharp change in the \(4\)-velocity of the source produces an electromagnetic field of the form \[F^{ab}=-2q\frac{1}{r}\frac{\sin\theta}{1-\cos\theta}\theta^{[a}K^{b]}\delta(U) \,, \label{EMcutoff}\] where \(K^a=t^a+r^a\) and \(U\) is given by eq. [\[upsilon\]](#upsilon){reference-type="eqref" reference="upsilon"} below. This field, in turn, will produce a "velocity kick" on a distant test particle, in agreement with recent results of Bieri and Garfinkle. In section 3, we also consider the limit as \(t_0 \to-\infty\) in the electromagnetic case. The limit of the vector of the vector potential in Lorenz gauge does not exist, but the limit of the field tensor \(F^{ab} = 2 \nabla^{[a} A^{b]}\) does exist, and we find \[F^{ab}=-4q\frac{1}{\rho}\rho^{[a}k^{b]}\delta(u) \,, \label{EMnocutoff}\] where \(\rho = (x^2 + y^2)^{1/2}\) and \(\rho^a = \nabla^a \rho\). This agrees with results of Jackiw, Kabat and Oritz. This retarded field of a charged particle that moves on a null geodesic forever also gives rise to a velocity kick on a distant test particle, but this kick is very different from the velocity kick produced by the field [\[EMcutoff\]](#EMcutoff){reference-type="eqref" reference="EMcutoff"}. In section 4, we treat the linearized gravitational case. Again, we can immediately write down the solution to the linearized Einstein equation in Lorenz gauge in terms of the retarded solution to the scalar wave equation. However, the linearized Einstein equation requires conservation of \(4\)-momentum of the source, so we can neither create a mass at time \(t_0\) nor have a mass initially "at rest" suddenly start moving on a null geodesic. Nevertheless, we can start with a particle of mass \(M\) "at rest" and, at time \(t_0\), have it "emit" a particle of energy \(E\) that moves on a null geodesic, with the original particle then losing mass and recoiling so as to conserve total \(4\)-momentum. As we shall see in section 4, to order \(1/r\) and to leading order in \(E/M\), the Riemann tensor of the retarded solution is \[\begin{aligned} R_{abcd}=4E&\frac{1}{r}\big[\frac{2}{1-\cos\theta}k_{[a}K_{b]}K_{[c}k_{d]}\nonumber\\ &-\left(2K_{[a}(t_{b]}z_{[c}+z_{b]}t_{[c})K_{d]}+(1+\cos\theta)\left(2K_{[a}t_{b]}t_{[c}K_{d]}+K_{[a}\eta_{b][c}K_{d]}\right)\right)\big]\delta'(U) \,. \label{Cutoff} \end{aligned}\] This Riemann tensor produces a "relative displacement kick" on test particles of the form \[\Delta D_a=E\frac{1}{r}\frac{\sin^2\theta}{1-\cos\theta}\left(\theta_a \theta_b-\phi_a\phi_b\right)D^b \,,\] in agreement with the form of the memory effect for null matter. In section 4, we also take the limit as \(t_0 \rightarrow-\infty\). Although the metric perturbation in Lorenz gauge does not exist as a distribution in this limit, the linearized Riemann tensor has the limit \[R_{abcd}=16E\frac{1}{\rho^2}k_{[a}(\rho_{b]}\rho_{[c}-\phi_{b]}\phi_{[c})k_{d]}\delta(u) \,-\, 16\pi Ek_{[a}q_{b][c}k_{d]}\delta(x)\delta(y) \delta(u) \,, \label{NoCutoff}\] where \(q_{ab}\) is the projection of the metric into the "\(x-y\)" plane. Thus, the Riemann tensor [\[NoCutoff\]](#NoCutoff){reference-type="eqref" reference="NoCutoff"} corresponds to the retarded linearized curvature produced by a null particle. Eq. [\[NoCutoff\]](#NoCutoff){reference-type="eqref" reference="NoCutoff"} agrees with the Aichelberg-Sexl solution (modulo what appear to be some sign misprints in their eq.(3.12)). The Riemann tensor [\[NoCutoff\]](#NoCutoff){reference-type="eqref" reference="NoCutoff"} has no "derivative of a \(\delta\)-function" piece, so unlike [\[Cutoff\]](#Cutoff){reference-type="eqref" reference="Cutoff"}, it provides no "relative displacement kick" to test particles. However, it does provide a "relative velocity kick" to test particles, which falls off as \(1/r^2\). Interestingly, as we shall show in section 4, this instantaneous relative velocity kick agrees, up to a factor of \(2\), with the integrated relative velocity change of test particles that would occur in Newtonian gravity due to tidal effects produced by the passage of a particle of mass \(m=E\) moving with velocity \(v=c\). Thus, the Aichelberg-Sexl Riemann tensor may be thought of as corresponding to a "special relativistic compression" of the Newtonian tidal effects of a particle moving at the speed of light into the null hyperplane containing the null particle source. Returning, finally, to the questions that motivated our investigations, we see that the Riemann tensor [\[NoCutoff\]](#NoCutoff){reference-type="eqref" reference="NoCutoff"} represents the retarded field produced by a null particle source in linearized gravity. As just noted above, this is a "pure tidal field" and there is no memory effect associated with this tidal field. We conclude that the memory effect should not be interpreted as being "caused by" the passage of null (effective) stress-energy to infinity. Conversely, the fact that there is a memory effect associated with the passage of null (effective) stress-energy to infinity is directly related to the fact that the Riemann tensor [\[NoCutoff\]](#NoCutoff){reference-type="eqref" reference="NoCutoff"} is not physically acceptable: It fails to be asymptotically flat at spatial infinity (even if we "smooth out" the source, as we can in linearized gravity), since the Riemann tensor vanishes in all non-equatorial directions and falls off too slowly (as \(1/r^2\)) in equatorial directions near spatial infinity. In order to have a solution with a null source that is asymptotically flat at spatial infinity, one must either "emit" the null source at a finite time or have "additional radiation" incoming from infinity. The memory effect should be thought of as being produced by the gravitational radiation resulting from such an emission event or such additional radiation. # Scalar Field As discussed in the previous section, we are interested in obtaining the retarded solution to the massless scalar wave equation in Minkowski spacetime \[\nabla^a \nabla_a \varphi =-4\pi S \label{waveeq}\] where the source, \(S\), corresponds to a (scalar) charged particle moving on a null geodesic, which we take to be moving in the "\(z\)-direction" in some global inertial coordinates \((t,x,y,z)\) \[S(t,x,y,z) = q \delta(x)\delta(y)\delta(z-t) \,. \label{source}\] We will denote events in spacetime by capital letters (i.e., \(X\) and \(X'\)), so as not to confuse spacetime points with the \(x\)-coordinate of our global inertial coordinates. We note, first, that although there appears to be a widespread belief that charged particle sources that move at the speed of light (or faster than light) should somehow be "illegal" (see, e.g., the remark below eq.(2.7) of ), there is, in fact, no difficulty in obtaining the retarded solution (as a distribution) to the wave equation [\[waveeq\]](#waveeq){reference-type="eqref" reference="waveeq"} for any distributional source, \(S\), of compact support. This can be seen as follows: If \(S\) is of compact support, the problem of obtaining the (distributional) retarded solution, \(\varphi_R(X)\), to [\[waveeq\]](#waveeq){reference-type="eqref" reference="waveeq"} is essentially the same as defining the product of the distributions \(G_R(X,X')\) and \(S(X')\), where \(G_R\) denotes the retarded Green's function, \[G_R(t,\vec{x};t',\vec{x}')=\frac{1}{2\pi}\delta \left[-(t-t')^2+|\vec{x}-\vec{x}'|^2 \right] \Theta(t-t') \,,\] since for any test function, \(f\), we have \(\varphi_R(f) = G_R S(F)\), where \(F(X,X') = f(X) h(X')\), with \(h\) being any test function with \(h=1\) on the support of \(S\). As a distribution on \({\bf R}^4 \times {\bf R}^4\), the wavefront set of \(G_R(X,X')\) is known to be of the form, \[{\rm WF} [G_R] = \{(X,K;X',-K')\}\] where \(X\) lies on a future-directed null geodesic starting from \(X'\), \(K\) is a (future-or past-directed) (co-)tangent to this geodesic[^2] at \(X\), and \(K'\) is the parallel transport of \(K\) to \(X'\). As a distribution on \({\bf R}^4 \times {\bf R}^4\), the wavefront set of \(S(X')\) is of the form \(\{(X,0;X',K')\}\) where \((X',K')\) is in the wavefront set of \(S\) as a distribution on \({\bf R}^4\). Hence, we cannot get a zero cotangent vector in \({\bf R}^4 \times {\bf R}^4\) by adding cotangent vectors in \({\rm WF} [G_R]\) to those in \({\rm WF} [S]\). It follows (see, e.g., ) that \(G_R S\) is well defined as a distribution for any distribution \(S\), and the retarded solution is well defined as a distribution for any \(S\) of compact support. Note that this argument generalizes straightforwardly to an arbitrary globally hyperbolic curved spacetime. On account of the support properties of \(G_R\), it is obvious that the requirement that \(S\) be of compact support can be replaced by the requirement that \(S\) vanish to the past of some Cauchy surface. However, the source [\[source\]](#source){reference-type="eqref" reference="source"} does not have this property, so it is not obvious that the retarded solution exists. Consequently, we will, instead, consider the source \[S_0 (t,x,y,z) = q \delta(x)\delta(y)\delta(z-t)\Theta(t-t_0) \,. \label{source2}\] corresponding to the "creation" of a scalar charged particle at time \(t_0\), which subsequently moves on a null geodesic. By the above general arguments, the retarded solution with source [\[source2\]](#source2){reference-type="eqref" reference="source2"} exists as a distribution. We will then consider the limit \(t_0 \to-\infty\). The retarded solution with source [\[source2\]](#source2){reference-type="eqref" reference="source2"} is \[\begin{aligned} \varphi_0(X)&=&4\pi\int\textnormal{d}^4x'\;G_R(X; X')S_0(X')\nonumber\\ &=&2q\int\;\textnormal{d}^4x'\;\delta\left[-(t-t')^2+ |\vec{x}-\vec{x}'|^2\right]\Theta(t-t') \delta(x')\delta(y')\delta(z'-t')\Theta(t'-t_0) \,. \end{aligned}\] Carrying out the \(\delta\)-function integrations over \(x',y',z'\), we obtain \[\begin{aligned} \varphi_0(X) &=&2q\int\;\textnormal{d}t'\;\delta\left[-(t-t')^2+x^2+y^2+(z-t')^2\right]\Theta(t-t')\Theta(t'-t_0)\nonumber\\ &=&2q\int\;\textnormal{d}t'\;\delta\left[2(t-z)t'-t^2+x^2+y^2+z^2 \right]\Theta(t-t')\Theta(t'-t_0)\nonumber\\ &=&\frac{q}{t-z}\Theta\left(t-\frac{t^2-x^2-y^2-z^2}{2(t-z)}\right)\Theta\left(\frac{t^2-x^2-y^2-z^2}{2(t-z)}-t_0\right) \,. \end{aligned}\] The two step functions can be combined into a single step function to produce our final result \[\varphi_0=\frac{q}{t-z}\Theta\left((t-t_0)-\sqrt{x^2+y^2+(z-t_0)^2}\right) \,. \label{retsol}\] Note that although \(\varphi_0\) is unbounded (since it diverges as \(t \downarrow z\) at \(x=y=0\)) it is locally in \(L^1\) and thus is well defined as a distribution. The "\(4\)-force," \(f^a\), exerted by the field \(\varphi_0\) on a test particle of charge \(Q\) is \[\begin{aligned} f^a=Q\nabla^a\varphi_0 \,. \end{aligned}\] Unlike the case of electromagnetism, \(f^a\) is not automatically orthogonal to the \(4\)-velocity of the test particle, and hence will, in general, produce a change in the rest mass of the particle as well as a change in its momentum. From [\[retsol\]](#retsol){reference-type="eqref" reference="retsol"}, we obtain \[f^a =\frac{qQ}{u^2}k^a\Theta(U)-\frac{qQ}{u}\left(t^a+\frac{rr^a-t_0z^a}{\sqrt{x^2+y^2+(z-t_0)^2}}\right)\delta(U) \,, \label{scalarforce}\] where \[u = t-z\] \[U=(t-t_0)-\sqrt{x^2+y^2+(z-t_0)^2} \,, \label{upsilon}\] and \(t^a\) and \(r^a\) are unit vectors in the time and radial directions. Note that \(U = 0\) corresponds to the future light cone of the event occurring at \(t = z= t_0, x=y=0\), where the source [\[source2\]](#source2){reference-type="eqref" reference="source2"} was created. To leading order in \(1/r\), our expression [\[scalarforce\]](#scalarforce){reference-type="eqref" reference="scalarforce"} becomes \[f^a =-\frac{qQ}{r}\frac{K^a}{1-\cos\theta} \delta(U) + O(1/r^2)\,,\] where \[K^a = t^a + r^a \,.\] This \(\delta\)-function contribution to \(f^a\) will give rise to an instantaneous "kick" in the \(4\)-momentum of the test particle. If the test particle is initially "at rest" and its motion remains non-relativistic, then the change in \(4\)-momentum due to this instantaneous kick is given by \[\Delta P^a =-\frac{qQ}{r}\frac{K^a}{1-\cos\theta}\,. \label{momkick}\] Note that this expression for the net kick is independent of \(t_0\), i.e., a change in \(t_0\) affects the kick only to higher order in \(1/r\) (although, of course, a change of \(t_0\) affects the time at which the kick is felt). Since the kick arises from the \(\delta(U)\) term in the force, the kick can be understood as being produced by a burst of radiation emitted when the source was created. Note that the kick diverges as \(\theta \to 0\). Let us now take the limit as \(t_0 \to-\infty\). Naively taking the limit of [\[retsol\]](#retsol){reference-type="eqref" reference="retsol"}, we obtain \[\varphi = \lim_{t_0 \to-\infty}\varphi_0=\frac{q}{t-z}\Theta(t-z) \,.\] However, the right side of this equation is not locally in \(L^1\) and does not make sense as a distribution. Indeed, it is easy to see that for any fixed, non-negative test function \(f\) with \(f \neq 0\) at some point at which \(t=z\) we have \[\lim_{t_0 \to-\infty} \int \varphi_0 f = \infty \,,\] so the weak distributional limit of \(\varphi_0\) does not exist as \(t_0 \to-\infty\). We conclude, therefore, that for the scalar wave equation, it does not make sense to talk about the retarded field of a charged particle[^3] source that moves forever on a null geodesic. As our derivation has indicated, the problem with obtaining a distributional solution arises from the "forever\" (i.e., non-compactness) character of the source rather than its "null\" character. Nevertheless, although \(\lim_{t_0 \to-\infty} \varphi_0\) does not exist as a distribution, some aspects of this limit do exist. Specifically, let \(k^a = t^a + z^a\) be the vector field on Minkowski spacetime that is everywhere parallel to the tangent to the null geodesic source [\[source2\]](#source2){reference-type="eqref" reference="source2"}. Then we claim that the weak distributional limit \(\lim_{t_0 \to-\infty} k^{[a} \nabla^{b]}\varphi_0\) does exist. To see this, let \(\alpha^{ab}\) be a smooth, antisymmetric tensor field of compact support. We wish to evaluate \[\lim_{t_0 \to-\infty}-\int \varphi_0 k_a \nabla_b \alpha^{ab}\ = \lim_{t_0 \to-\infty}-\int_{U > 0} {\rm d}^4 x \,\, \frac{1}{u} k_a \nabla_b \alpha^{ab} \,. \label{smearing1}\] Integrating by parts, we obtain \[-\int_{U > 0} {\rm d}^4 x \,\, \frac{1}{u} k_a \nabla_b \alpha^{ab} = -\int_{U > 0} {\rm d}^4 x \,\, \frac{1}{u^2} k_a\nabla_b u \alpha^{ab}-\int_{U = 0} \frac{1}{u} k_a n_b \alpha^{ab}\] where \(n^a\) is the normal to the \(U=0\) surface, \[n^a=t^a+\frac{\rho \rho^a+(z-t_0)z^a}{\sqrt{\rho^2+(z-t_0)^2}} \,.\] (Here \(\rho = (x^2 + y^2)^{1/2}\) and \(\rho^a = \nabla^a \rho\).) The bulk integral vanishes because \(\alpha^{ab}\) is antisymmetric and \(k_{[a}\nabla_{b]}u = 0\). The surface term is \[-\int_{U = 0} \frac{1}{u} k_a n_b \alpha^{ab} =-\int_{U = 0} \rho\,\textnormal{d}\rho\,\textnormal{d}\phi\,\textnormal{d}z \frac{\sqrt{\rho^2+(z-t_0)^2}k_{[a}t_{b]}+\rho k_{[a}\rho_{b]}+(z-t_0)k_{[a}z_{b]}}{\left(\sqrt{\rho^2+(z-t_0)^2}-(z-t_0)\right)\sqrt{\rho^2+(z-t_0)^2}} \alpha^{ab}\] As \(t_0 \to-\infty\), the numerator in this expression converges uniformly on compact sets to \(\rho k_{[a}\rho_{b]}\), whereas the denominator converges uniformly on compact sets to \(\rho^2/2\). Furthermore, as \(t_0 \to-\infty\), we have \(U\to u\). From this it can be seen that the (weak) limit of \(k^{[a} \nabla^{b]}\varphi_0\) as \(t_0 \to-\infty\) exists and is given by \[\lim_{t_0\rightarrow-\infty}k^{[a}\nabla^{b]}\varphi_0=2q\frac{1}{\rho}\rho^{[a}k^{b]}\delta(u) \,. \label{finalscalarforce}\] Thus, in the limit \(t_0 \to-\infty\), the force exerted on a test particle is well defined modulo addition of multiples of \(k^a\). Since this force also has a \(\delta\)-function character, it gives rise to a \(4\)-momentum kick of the form \[\Delta P_\infty^a =-2qQ\frac{1}{\rho}\rho^a\] modulo multiples of \(k^a\). This \(4\)-momentum kick is very different in form from the kick [\[momkick\]](#momkick){reference-type="eqref" reference="momkick"} produced by the burst of radiation arising from a "creation event." # Electromagnetic Field In this section, we wish to obtain the retarded solution to Maxwell's equations with a charged particle source moving on a null geodesic. As in the case of the scalar wave equation, in order to have a well defined solution, we would like to "create" the source at a finite time \(t_0\) and then consider the limit \(t_0 \to-\infty\). However, unlike the scalar case, we cannot "create" a charge at a finite time because Maxwell's equations require conservation of charge. Therefore, we consider, instead, a situation where a charge sits "at rest" until time \(t=t_0\) and thereafter moves on a null geodesic, i.e., we take the \(4\)-current to be \[j_0^a = q \delta(x)\delta(y)\left[\delta(z-t_0)\Theta(t_0-t)t^a +\delta(z-t)\Theta(t-t_0)k^a\right] \,,\] where \(k^a = t^a + z^a\) is tangent to the null geodesic \(x=y=0, t=z\). Maxwell's equations for the vector potential, \(A^a\), in Lorenz gauge, \(\nabla_a A^a = 0\), take the form of a wave equation [\[waveeq\]](#waveeq){reference-type="eqref" reference="waveeq"} for each global inertial component of \(A^a\). Therefore, we can immediately write down the retarded solution in Lorenz gauge using the well known Coulomb solution for the source for \(t < t_0\) and using [\[retsol\]](#retsol){reference-type="eqref" reference="retsol"} for \(t \geq t_0\). We obtain \[A_0^a=\frac{q}{\sqrt{x^2+y^2+(z-t_0)^2}} \Theta(-U) t^a +\frac{q}{t-z} \Theta(U) k^a \label{A0}\] where \(U\) was defined in [\[upsilon\]](#upsilon){reference-type="eqref" reference="upsilon"} above. The electromagnetic field tensor is given in terms of the vector potential by \(F^{ab} = 2 \nabla^{[a} A^{b]}\). From [\[A0\]](#A0){reference-type="eqref" reference="A0"}, we obtain \[(F_0)^{ab} =-2q\frac{1}{r}\frac{\sin\theta}{1-\cos\theta}\theta^{[a}K^{b]}\delta(U) + O(1/r^2)\,. \label{F0}\] The force on a test particle of charge \(Q\) and \(4\)-velocity \(u^a\) is \(f_a = Q F_{ab} u^b\). As in the scalar case, the leading order in \(1/r\) contribution to \(f_a\) is a \(\delta\)-function term, which will give the particle an instantaneous momentum kick. In the case of electromagnetism, \(f^a\) is automatically orthogonal to \(u^a\) and, hence, does not change the rest mass of the test particle, i.e., the particle gets only a "velocity kick." For a test particle that is initially "at rest" and whose motion remains non-relativistic, the instantaneous kick in \(4\)-momentum is given by \[\Delta P^a =qQ\frac{1}{r}\frac{\sin\theta}{1-\cos\theta}\theta^a \,. \label{momkickEM}\] This agrees with the velocity kick obtained by Bieri and Garfinkle. Let us now take the limit \(t_0 \to-\infty\). The contribution of the first (Coulomb) term in [\[A0\]](#A0){reference-type="eqref" reference="A0"} clearly goes to zero in this limit. However, apart from the factor of \(k^a\), the contribution of the second term in [\[A0\]](#A0){reference-type="eqref" reference="A0"} is identical to the scalar case, and hence it does *not* have a distributional limit. We conclude that the retarded solution for the vector potential of a charged particle that moves forever on a null geodesic does not exist in Lorenz gauge. Nevertheless, since the Coulomb contribution vanishes in the limit, we see that that \[F_{ab} \equiv \lim_{t_0 \to-\infty}(F_0)_{ab} =-2 \lim_{t_0 \to-\infty} k_{[a} \nabla_{b]} \varphi_0\] with \(\varphi_0\) given by [\[retsol\]](#retsol){reference-type="eqref" reference="retsol"}. As we showed in the previous section, the limit on the right side of this equation *does* exist as a distribution, and we obtain \[F_{ab} =-4q\frac{1}{\rho}\rho_{[a}k_{b]}\delta(u) \,. \label{Fret}\] Equation [\[Fret\]](#Fret){reference-type="eqref" reference="Fret"} may thus be interpreted as providing the retarded field[^4] of a charged particle that moves on a null geodesic for all time, in agreement with Jackiw, Kabat and Oritz (see also problem 11.18 of the third edition of Jackson ). The field [\[Fret\]](#Fret){reference-type="eqref" reference="Fret"} produces an instantaneous momentum kick on a test particle of charge \(Q\) (assumed to be initially at rest) given by \[\Delta P_\infty^a = 2qQ\frac{1}{\rho}\rho^a \,.\] Again, this differs in form from the momentum kick [\[momkickEM\]](#momkickEM){reference-type="eqref" reference="momkickEM"} produced by the burst of radiation associated with the instantaneous change of motion of the source at time \(t_0\). # Linearized Gravitational Field We now turn to the case of linearized gravity, with a source \(T_{ab}\) corresponding to a particle moving on a null geodesic. As in the scalar and electromagnetic cases, we would like to "create" this particle at time \(t_0\) and then take the limit \(t_0 \to-\infty\). However, the linearized Einstein equation requires conservation of stress-energy, which, for particle sources, requires conservation of \(4\)-momentum. Thus, the simplest case to consider would be a particle of mass \(M\) which is at rest until time \(t_0\), at which time it emits a null particle of energy, \(E\), and then loses mass and recoils so as to conserve \(4\)-momentum. Thus, we consider a stress-energy source of the form \[\begin{aligned} T_0^{ab} &=& \delta(x)\delta(y)\big[M \delta(z-t_0)\Theta(t_0-t)t^at^b + \nonumber \\ &&+ M' \delta(z'-t_0) \Theta(t-t_0) t'^a t'^b + E \delta(z-t)\Theta(t-t_0)k^a k^b \big] \,, \label{gravsource} \end{aligned}\] where \(M'\) and \(t'^a\) are chosen so as to conserve \(4\)-momentum and \(z'\) is the global inertial "\(z\)-coordinate" in the frame in which the recoiling particle is at rest. We denote the metric perturbation by \(h_{ab}\). As is well known, in Lorenz (harmonic) gauge, the linearized Einstein equation for \(\bar{h}_{ab} \equiv h_{ab}-\frac{1}{2} h \eta_{ab}\) (where \(\eta_{ab}\) is the Minkowski metric and \(h = \eta^{ab} h_{ab}\)) takes the form \[\nabla^c \nabla_c\bar{h}_{ab}=-16\pi T_{ab} \,,\] yielding a wave equation for each of its global inertial components. We can therefore immediately obtain the retarded solution for \(\bar{h}_{ab}\) with source [\[gravsource\]](#gravsource){reference-type="eqref" reference="gravsource"}---and, hence, obtain the retarded solution for \(h_{ab}\)---as a sum of 3 pieces: (I) a linearized Schwarzschild piece arising from the first ("particle at rest") term in the source, \[(h^{\rm I}_0)_{ab}=\frac{2M}{\sqrt{x^2+y^2+(z-t_0)^2}}\left(\eta_{ab}+2t_at_b\right)\Theta(-U) \,, \label{hI}\] (II) a boosted Schwarzschild piece arising from the second ("recoiling particle") term in the source \[(h^{\rm II}_0)_{ab}=\frac{2M'}{\sqrt{x^2+y^2+(z'-t_0)^2}}\left(\eta_{ab}+2t'_at'_b\right)\Theta(U) \,, \label{hII}\] and (III) a piece arising from the third ("null particle") term in the source \[(h^{\rm III}_0)_{ab}=\frac{4E}{t-z}k_ak_b\Theta(U) \,. \label{hIII}\] The linearized Riemann tensor, \(R_{abcd}\), associated with metric perturbation \(h_{ab}\) is \[R_{abcd} = 2\nabla_{[a} \nabla_{|[d} h_{c]|b]} \,.\] The leading order in \(1/r\) contribution to the linearized Riemann tensor will arise from differentiation of the \(\Theta\)-functions appearing in [\[hI\]](#hI){reference-type="eqref" reference="hI"}-[\[hIII\]](#hIII){reference-type="eqref" reference="hIII"}. Assuming \(E \ll M\) and keeping only the leading order term in \(E/M\), we find \[\begin{aligned} R_{abcd}=4E&\frac{1}{r}\big[\frac{2}{1-\cos\theta}k_{[a}K_{b]}K_{[c}k_{d]}\nonumber\\ &-\left(2K_{[a}(t_{b]}z_{[c}+z_{b]}t_{[c})K_{d]}+(1+\cos\theta)\left(2K_{[a}t_{b]}t_{[c}K_{d]}+K_{[a}\eta_{b][c}K_{d]}\right)\right)\big]\delta'(U) \,, \label{RI+II+III} \end{aligned}\] where the first term in the square brackets arises from \((h^{\rm III}_0)_{ab}\) and the remaining terms arise from \((h^{\rm II}_0)_{ab}\). Although we have chosen a particular decay/recoil process in order to do these calculations, the details of the process are irrelevant at \(O(E/M)\) provided that all of the particles apart from the null particle are non-relativistic, i.e., the details of the decay process would affect [\[RI+II+III\]](#RI+II+III){reference-type="eqref" reference="RI+II+III"} only at higher orders in \(E/M\). The linearized Riemann tensor will produce a relative acceleration (i.e., geodesic deviation) for nearby freely falling test particles. If the particles are initially "at rest" (i.e., \(4\)-velocity parallel to \(t^a\)) and separated by spatial displacement \(D^a\), then \[t^e \nabla_e t^f \nabla_f D^a =-{R_{bcd}}^a t^b t^d D^c\] i.e., in terms of components \[\frac{d^2 D^\mu}{dt^2}=-{R_{t \nu t}}^\mu D^\nu \,. \label{relacc2}\] The "derivative of a \(\delta\)-function" terms in the linearized Riemann tensor will therefore produce an instantaneous relative *displacement kick* to the test particles. *This is precisely the memory effect.* For the Riemann tensor [\[RI+II+III\]](#RI+II+III){reference-type="eqref" reference="RI+II+III"}, we obtain \[(\Delta D_0)_a =\frac{E}{r}\frac{\sin^2\theta}{1-\cos\theta}\left(\theta_a\theta_b-\phi_a\phi_b\right)D^b \,.\] This agrees with the memory effect formulas of Christodoulou and Bieri and Garfinkle. Let us now take the limit as \(t_0 \to-\infty\) of the metric perturbation [\[hI\]](#hI){reference-type="eqref" reference="hI"}-[\[hIII\]](#hIII){reference-type="eqref" reference="hIII"}. It is clear from eqs. [\[hI\]](#hI){reference-type="eqref" reference="hI"} and [\[hII\]](#hII){reference-type="eqref" reference="hII"} that \[\lim_{t_0 \to-\infty} (h^{\rm I}_0)_{ab} = \lim_{t_0 \to-\infty} (h^{\rm II}_0)_{ab} = 0 \,.\] On the other hand, \((h^{\rm III}_0)_{ab} = 4\varphi_0 k_a k_b\), so the limit as \(t_0 \to-\infty\) of \((h^{\rm III}_0)_{ab}\) does not exist. We conclude that, as in the electromagnetic case, the retarded solution for the metric perturbation of a particle that moves forever on a null geodesic does not exist in the Lorenz gauge. On the other hand, the contribution of \((h^{\rm III}_0)_{ab}\) to the linearized Riemann tensor is \[(R^{\rm III}_0)_{abcd} = 8\nabla_{[a} \nabla_{|[d} \varphi_0 k_{c]|} k_{b]}\] and it follows that the limit as \(t_0 \to-\infty\) of \((R^{\rm III}_0)_{abcd}\) *does* exist. In fact, we obtain \[R_{abcd} = \lim_{t_0 \to-\infty} (R^{\rm III}_0)_{abcd} = 4 k_{[a} \nabla_{b]} F_{cd} \label{Rnull0}\] where \(F_{ab}\) is given by [\[Fret\]](#Fret){reference-type="eqref" reference="Fret"} (with \(q\) replaced by \(E\)) and the derivative is taken in the distributional sense. To calculate this distributional derivative more explicitly, let \(\beta^{abcd}\) be smooth and of compact support and have the tensor symmetries of the linearized Riemann tensor. We wish to evaluate \[-16\int \nabla_b \beta^{abcd} k_a \left(-\frac{1}{\rho}\rho_c k_d\right) \rho d\rho d\phi dz \,.\] To do so, we exclude a disc of radius \(\epsilon\) about \(\rho = 0\), integrate by parts with respect to \(\rho\), and then let \(\epsilon \to 0\). We thereby obtain \[R_{abcd}=16E\frac{1}{\rho^2}k_{[a}(\rho_{b]}\rho_{[c}-\phi_{b]}\phi_{[c})k_{d]}\delta(u) \,-\, 16\pi Ek_{[a}q_{b][c}k_{d]}\delta(x)\delta(y) \delta(u) \,, \label{Rnull}\] where \(q_{ab}\) is the projection of the metric into the "\(x\)-\(y\)" plane. Equation [\[Rnull\]](#Rnull){reference-type="eqref" reference="Rnull"} agrees with the Riemann curvature tensor of the Aichelburg-Sexl solution---apart from several sign discrepancies, which are undoubtedly misprints in eq.(3.12) of their paper[^5]. Equation [\[Rnull\]](#Rnull){reference-type="eqref" reference="Rnull"} may be interpreted as the linearized curvature[^6] of the retarded field of particle of energy \(E\) that moves on a null geodesic forever. Although, as we have seen, the retarded solution for the perturbed metric in the Lorenz gauge, does not exist as a distribution, it should be possible to find other gauges in which a distributional metric perturbation giving rise to [\[Rnull\]](#Rnull){reference-type="eqref" reference="Rnull"} does exist. Unlike [\[RI+II+III\]](#RI+II+III){reference-type="eqref" reference="RI+II+III"}, the Riemann tensor [\[Rnull\]](#Rnull){reference-type="eqref" reference="Rnull"} does not have a derivative of a \(\delta\)-function term. Furthermore, its effects fall off at large distances like \(1/r^2\) rather than \(1/r\). Consequently, *we conclude there is no memory effect associated with the retarded field of a particle that moves on a null geodesic forever.* However, the \(\delta\)-function in [\[Rnull\]](#Rnull){reference-type="eqref" reference="Rnull"} will produce an instantaneous "relative velocity kick" to a system of test particles moving on geodesics. Integrating [\[Rnull\]](#Rnull){reference-type="eqref" reference="Rnull"}, we find that if the particles have initially separation \(D^a\), the relative velocity kick will be \[\Delta v_a= 4E\frac{1}{\rho^2}\left(\rho_{a}\rho_{b}-\phi_{a}\phi_{b}\right)D^b \,. \label{ASkick}\] This velocity kick can be given a simple interpretation in terms of Newtonian tidal effects. Consider, in Newtonian gravity, a particle of mass \(E\) traveling with velocity \(c\) along the \(z\)-axis. The Newtonian potential produced by such a particle at time \(t\) is \[\chi=-\frac{E}{\sqrt{x^2+y^2+(z-ct)^2}} \,.\] The tidal tensor associated with this potential is \[\Phi_{a b}=\frac{E}{r'^3}(3r'_a r'_b-\delta_{ab}) \,,\] where \(r'=\sqrt{x^2+y^2+(z-ct)^2}\) and \(r'_a = \nabla_a r'\). We can integrate the tidal tensor once to get the net relative velocity change of two neighboring test particles over all time. For test particles initially separated by the displacement \(D^j\), we obtain \[\begin{aligned} \Delta v_i = \int_{-\infty}^\infty\textnormal{d}t'\Phi_{ij}(t',x,y,z)D^j\nonumber\\ =2E\frac{1}{\rho^2}(\rho_i \rho_j-\phi_i \phi_j)D^j \,. \end{aligned}\] Thus, apart from a factor of \(2\), the net relative velocity change in the Newtonian case produced by a particle of mass \(E\) that moves forever along the \(z\)-axis at velocity \(c\) agrees with the relative velocity kick in linearized gravity produced by a particle of energy \(E\) that moves forever on a corresponding null geodesic. The only difference is that in Newtonian gravity, these tidal effects occur over all time, whereas in linearized gravity, the tidal effects are "compressed" into a null plane traveling along with the source. Thus, the Newtonian tidal acceleration is gradual and continuous, whereas in linearized gravity, one obtains an instantaneous velocity kick. # Summary and Conclusions We have have investigated the retarded solution for a scalar field, an electromagnetic field, and a linearized gravitational field associated with the creation of a null particle at time \(t_0\) in Minkowski spacetime. In the scalar case, we can simply create a charged null particle; in the electromagnetic and linearized gravitational cases, other sources must also be present in order to conserve, respectively, charge and \(4\)-momentum. There are then two distinct limits of this retarded solution that we can take. The first is to fix \(t_0\) and extract the leading order in \(1/r\) behavior of the solution. In all three cases, there are effects produced on distant test particles at order \(1/r\) caused by the creation of the null particle. In the scalar and electromagnetic cases, they give rise to an instantaneous "kick" to the \(4\)-momentum of a test particle. In the linearized gravitational case, the \(O(1/r)\) effect is to produce an instantaneous relative displacement of test particles---the memory effect. The alternative limit is to fix the observation point and let \(t_0 \to-\infty\). This limit can be thought of as providing the retarded field of a null particle that moves on a null geodesic forever. In the scalar case, we found that this limit does not exist as a distribution. However, in the electromagnetic and linearized gravitational cases, although the limits of the Lorenz gauge vector potential and Lorenz gauge metric perturbation similarly do not exist, the limits of the electromagnetic field tensor and linearized Riemann tensor do exist. In the electromagnetic case, the limiting electromagnetic field tensor gives rise to a velocity kick on distant test particles at order \(1/r\), but the form of this velocity kick is very different from the \(O(1/r)\) velocity kick produced by the creation of a null charge at finite time \(t_0\). In the linearized gravitational case, the limiting linearized Riemann tensor yields the Aichelberg-Sexl solution. It falls off as \(1/r^2\) and thus produces no effects of any kind at order \(1/r\). In particular, there is no memory effect. The leading order (\(1/r^2\)) effect of this linearized Riemann tensor is to produce an instantaneous relative velocity kick on test particles, of exactly the same form as the integrated Newtonian tidal force would produce. We conclude that in linearized gravity, the "radiation field" (retarded solution) produced by a particle moving on a null geodesic forever is the Aichelberg-Sexl solution, which is a pure "tidal field" that produces no associated memory effect. Thus, the memory effect should not be interpreted as being caused merely by the passage of (effective) stress-energy to null infinity. However, as already noted in the Introduction, the Aichelberg-Sexl solution (even with a smoothed out source) fails to be asymptotically flat at spatial infinity, and thus is not physically acceptable. One way of producing a physically acceptable solution is to create the null particle at a finite time \(t_0\) via an "emission event," as we have considered. In that case, there will be a burst of radiation associated with the emission event that produces a nontrivial memory effect, in agreement with previous results. More generally, the requirement of asymptotic flatness at spatial infinity implies either the finite time creation of the null particle or the presence of additional "incoming radiation" from past null infinity that is not directly associated with the null particle. We believe that the memory effect is most naturally interpreted as being caused by either the emission event or by the additional incoming radiation from past null infinity, rather than by the passage of the particle to future null infinity. We wish to thank Lydia Bieri and David Garfinkle for helpful discussions. This research was supported in part by NSF grant PHY12-02718 to the University of Chicago. Some of this research was carried out while one of us (R.M.W.) was in residence during September, 2013 at the Mathematical Sciences Research Institute in Berkeley, California, supported by the National Science Foundation under Grant No. 0932078 000. [^1]: Neither Christodoulou nor Bieri and Garfinkle propose an interpretation of the "cause" of the memory effect. [^2]: For \(X=X'\), the wavefront set is \(\{X,K;X,-K\}\) for all \(K \neq 0\). [^3]: Since, in Minkowski spacetime, averging over the observation point is equivalent to averaging over the source, the failure to obtain a distributional solution for a particle source moving forever on a null geodesic implies the failure to have any retarded solution at all for a smooth, null fluid source with everywhere parallel \(4\)-velocity. [^4]: Note that although we showed above that the retarded vector potential in Lorenz gauge does not exist for this solution, one can find other gauges in which a distributional vector potential for the field ([\[Fret\]](#Fret){reference-type="ref" reference="Fret"}) can be found; see and. [^5]: In particular, the Ricci component \(R_{00}\) is easily computed by adding together the first two lines of their eq.(3.12) and does not agree with the (correct) expression they give below eq.(3.12); their eq.(3.12) also fails to be rotationally invariant in the plane orthogonal to the direction of the particle. [^6]: As Aichelburg and Sexl have argued, this solution may be interpreted as a solution to the full, nonlinear Einstein equation, not merely the linearized Einstein equation. Indeed, Aichelberg and Sexl obtained their solution by taking an infinite boost limit of the exact Schwarszchild solution.

{'timestamp': '2014-01-24T02:02:09', 'yymm': '1401', 'arxiv_id': '1401.5831', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5831'}

# Introduction {#sec:Introduction} The ability to accurately reproduce the experimentally determined baryon electromagnetic form factors stands as a fundamental test for any theoretical description of baryon structure. This is a test which quantum chromodynamics (QCD), our theory of the strong interactions, has not yet definitively passed . In particular, with ever-improving experimental determinations of the form factors revealing slight deviations from the phenomenological dipole form , a precise determination of these and related quantities from first-principles QCD is essential. Lattice simulation remains the only rigorous method to quantitatively probe the non-perturbative domain of QCD. As well as facilitating a comparison of theory with experimental data, lattice simulations provide a great deal of physical insight and valuable information for model building by revealing the dependence of hadron properties on quark mass . Recent years have seen a progression from quenched to fully dynamical lattice QCD studies of the electromagnetic form factors. Despite this significant advance, operator self-contractions (disconnected quark diagrams) are still often omitted from simulations as they are notoriously noisy and expensive to calculate. While in general this omission produces a systematic effect (shown to be small in Ref. ), exact results may be obtained for isovector quantities, where contributions from disconnected loops cancel. We present new dynamical \(2+1-\)flavor lattice QCD simulation results for the electromagnetic form factors of the octet baryons. This data set includes results for \(G_{E/M}\) for all outer-ring octet baryons at a range of discrete \(Q^2\) values up to 1.3 GeV\(^2\). As chiral extrapolations are different for the electric and magnetic form factors, we present here an analysis of the magnetic form factor only. \(G_E\) will be considered in future work. We extrapolate the lattice results for \(G_M\), at each value of \(Q^2\), as a function of quark mass to the physical point. As the lattice simulations neglect disconnected quark contractions, this extrapolation is performed using a variation of partially-quenched chiral perturbation theory. The distinguishing feature of this formalism is that valence and sea quarks are treated separately. For example, one may set the electric charge of the sea quarks to zero, removing the same disconnected quark contractions omitted in the lattice simulations . This is termed 'connected chiral perturbation theory'. Finite-volume effects are estimated by using the leading one-loop results of the chiral effective field theory. By carrying out the lattice simulations over a range of light and strange quark masses it is possible to tightly constrain the chiral extrapolation on the relevant parameter space and obtain surprisingly accurate results for the form factors at the physical point. Those results compare quite favourably with the experimental values. The details of the lattice simulation are given in Sec. [2](#sec:LatticeSimulation){reference-type="ref" reference="sec:LatticeSimulation"}, while Sec. [3](#sec:ChiPTextrap){reference-type="ref" reference="sec:ChiPTextrap"} presents the effective field theory formalism. Fits to the lattice simulation results are described in Sec. [4](#sec:Fits){reference-type="ref" reference="sec:Fits"}, followed by results for the magnetic isovector form factors, octet baryon magnetic moments and magnetic radii in Sec. [5](#sec:Results){reference-type="ref" reference="sec:Results"}. The appendices provide further details, including tables of lattice results and functional forms for the chiral expansions. # Lattice simulation {#sec:LatticeSimulation} ## Magnetic radii {#sec:radii} The magnetic radii of the octet baryons are defined by \[\langle r^2_M \rangle^B =-\frac{6}{G_M^B(0)}\frac{d}{dQ^2}G_M^B(Q^2)\bigg|_{Q^2=0}.\] To evaluate this expression from the lattice simulation results, we use the dipole-like fits (Eq. ([\[eq:gendipfit\]](#eq:gendipfit){reference-type="ref" reference="eq:gendipfit"})) shown in Appendix [\[app:FFPics\]](#app:FFPics){reference-type="ref" reference="app:FFPics"}. Results, compared with available experimental data, are given in Table [\[tab:MagRadii\]](#tab:MagRadii){reference-type="ref" reference="tab:MagRadii"}. It is notable that we find consistently smaller values for the magnetic radii than those determined experimentally (for the nucleon) or predicted in chiral quark models (for the octet baryons) . This is perhaps not unexpected; comparing Figs. [\[fig:pKelly\]](#fig:pKelly){reference-type="ref" reference="fig:pKelly"} and [\[fig:nKelly\]](#fig:nKelly){reference-type="ref" reference="fig:nKelly"} with Figs. [\[fig:MMp\]](#fig:MMp){reference-type="ref" reference="fig:MMp"} and [\[fig:MMn\]](#fig:MMn){reference-type="ref" reference="fig:MMn"} shows that although our results are quite consistent with the experimental parameterization of the nucleon form factors *where they are calculated*, the best-fit dipole function has slightly less curvature. As noted in the previous sections, greater curvature in the \(Q^2\) fit forms would improve consistency with the experimental magnetic moments for all of the octet baryons. To improve the extraction of the magnetic radii, we consider a second functional form in \(Q^2\), inspired by the Kelly-style parameterizations of experimental results with a more general polynomial in the denominator: \[G^B_M(Q^2)=\frac{\mu_B}{1+c Q^2 + d Q^4 + f Q^6}. \label{eq:GenFit}\] We now fix \(\mu_B\) to the experimental magnetic moment, so there are again three free parameters, \(c\), \(d\) and \(f\). As illustrated for the proton in Fig. [\[fig:GPComp\]](#fig:GPComp){reference-type="ref" reference="fig:GPComp"}, the quality of fit using this functional form is entirely comparable with that for the dipole-like fit. The shift in the extracted value of the magnetic radius, however, is significant, as shown in Table [\[tab:MagRadii\]](#tab:MagRadii){reference-type="ref" reference="tab:MagRadii"}. This example confirms that truly robust predictions for the hyperon magnetic radii from lattice QCD will require results at much lower \(Q^2\) values to eliminate the significant dependence on the functional form chosen for the \(Q^2\) extrapolation. Nevertheless, the level of agreement of the extracted nucleon magnetic radii with experimental values indicates that, by taking the experimental magnetic moments as additional input, we have achieved the first accurate calculation of the magnetic radii of the entire outer ring of the baryon octet from lattice QCD, extrapolated to the physical pseudoscalar masses. ## Quark form factors The chiral extrapolations discussed in previous sections are in fact performed for the individual doubly and singly represented quark contributions to the magnetic form factors. Inspecting these contributions can give insight into the environmental sensitivity of the distribution of quarks inside a hadron. Chiral extrapolations for the connected part of these quark contributions, shown along the trajectory which holds the singlet pseudoscalar mass \((m_K^2+m_\pi^2/2)\) fixed to its physical value, are presented in Figs. [\[fig:DoublyRep\]](#fig:DoublyRep){reference-type="ref" reference="fig:DoublyRep"} and [\[fig:SinglyRep\]](#fig:SinglyRep){reference-type="ref" reference="fig:SinglyRep"}. The figures show the lowest \(Q^2\) result, at approximately 0.26 GeV\(^2\). Of course, the fits shown are simultaneously constrained by the lattice simulation results for all of the octet baryons at that \(Q^2\). Comparison of the \(u\) quark contributions to the proton and \(\Sigma^+\) in Fig. [\[fig:DoublyRep\]](#fig:DoublyRep){reference-type="ref" reference="fig:DoublyRep"} shows the relative suppression of \(G_M^{\Sigma,u}\) caused by the heavier spectator quark in the sigma. This effect is replicated, and is more significant, when probing the singly represented quark, as can be seen by the relative suppression (in magnitude) of the \(u\) contribution to the cascade baryon compared to the \(d\) in the proton in Fig. [\[fig:SinglyRep\]](#fig:SinglyRep){reference-type="ref" reference="fig:SinglyRep"}. Changing the mass of the probed quark--doubly represented in the proton compared with the cascade, or singly represented in the proton compared with the sigma--causes a similar effect. # Conclusion We have presented the results of a 2+1-flavor lattice QCD study of the electromagnetic form factors of the octet baryons. Calculations are performed on one volume with a single lattice spacing, six different sets of pseudoscalar masses and six values of \(Q^2\) in the range 0.2--1.3 GeV\(^2\). The Dirac and Pauli radii of the nucleon, extracted using generalized dipole fits, are in line with other recent 2+1 and 2+1+1 flavor lattice calculations with similar values of the pion mass. By performing lattice simulations on configurations which 'map out' the \(m_l-m_s\) plane, rather than following a single trajectory in this space, we are able to robustly constrain a chiral extrapolation of the magnetic Sachs form factor \(G_M\) to the physical pseudoscalar masses at each simulation \(Q^2\). Systematic uncertainties are controlled by performing finite-volume corrections. The uncertainties inherent in the determination of the lattice scale \(a\), the shape of the ultra-violet cutoff and the value of the cutoff parameter \(\Lambda\) in the finite-range regularization scheme are found to be negligible. As such, the single most significant limitation of this calculation is that disconnected quark loops are omitted from the lattice simulations. For this reason isovector combinations, where contributions from disconnected quark loops cancel, are of significant interest. The nucleon isovector form factor extracted from this work compares well with the experimental results over the entire range of \(Q^2\) values considered. It is notable that the precision of these results is such that it is foreseeable that this generation of lattice QCD simulations will rival experiment in terms of precision. The proton and neutron magnetic form factors from this work, which include only the 'connected' quark loop contributions, agree rather well with the experimental determinations at all simulation \(Q^2\) values. The comparison with experiment is also favourable for the magnetic moments and magnetic radii of the rest of the outer-ring baryon octet, extracted using a dipole-like form in \(Q^2\). This suggests that the omitted disconnected quark loop contributions are small relative to the uncertainties of this calculation. We point out that a pure dipole form in \(Q^2\) does not, in general, provide a good fit to the lattice simulation results. A dipole-like function with a more general polynomial in the denominator is significantly better, as described above. A comparison of nucleon observables extracted using both fit forms indicates that the dipole yields significantly poorer predictions for the magnetic moments and radii, despite the form factors matching the experimental values at larger \(Q^2\). This suggests that meaningful extractions of the magnetic moments and radii from lattice QCD require a more careful analysis than the standard procedure using a dipole fit in \(Q^2\), unless simulations are performed for very small \(Q^2\) values much less than 0.2 GeV\(^2\). Analyses similar to that performed here may reveal that other existing lattice simulations are in fact more compatible with experiment than the results of the standard calculations indicate.

{'timestamp': '2014-04-22T02:08:19', 'yymm': '1401', 'arxiv_id': '1401.5862', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5862'}

# Introduction Cryptography underpins all secure communications, whether it is used for transferring credit card information between a buyer and a seller through the Internet, or relaying classified information over a military network. Classical cryptography is unsatisfactory in several respects. Its security is generally based on assumptions about computation which are believed true but for which no absolute proof exists. That is, the assumptions depend on relative results which say that if certain computational problems are not efficiently solvable, then specific cryptosystems are secure. In practical terms, classical cryptosystems frequently suffer from difficulties in secure key distribution and often need to be revised and updated to maintain their security in response to strong advances in the capabilities of modern computer systems. In contrast, quantum cryptography was devised in the mid-1980's in order to create cryptographic protocols that are guaranteed to be secure by the laws of nature \[1\] (i.e. physics, and specifically quantum mechanics). The BB84 protocol was the first practical quantum cryptography protocol to be proven secure using the laws of quantum mechanics \[2, 3\], and it has provided the framework for most implementations of quantum cryptography. One major shortcoming of BB84 is its inefficient use of quantum mechanical information in that the key generation rate of a shared secret key is less than half of the repetition rate of the single photon source used for the protocol \[1\]. In 1995, Goldenberg and Vaidman proposed the first orthogonal states-based protocol which combined principles of special relativity with quantum mechanics to allow for secure communication \[4\]. Since then, variations of their protocol have been proposed \[5, 6\]. The way in which a relativistic quantum key distribution protocol exploits the laws of quantum mechanics is by creating a situation in which an eavesdropper must cause a detectable time delay if she wishes to obtain information about the secret key. The eavesdropper cannot undo this delay without using faster-than-light communication-a task which is impossible according to special relativity. In this paper we present a related but more efficient and flexible cryptographic protocol for which security is similarly guaranteed by nature and which expands the reach of provably secure cryptography. Our QKD protocol has the unique feature that it can achieve a key generation rate greater than the repetition rate of the single photon source used for the protocol. # The Protocol ## Procedure According to common nomenclature, the sender of a cryptographic message is referred to as Alice, the receiver as Bob, and the eavesdropper as Eve. The proposed protocol generates a secret key between Alice and Bob by encoding information on the phase between four states of a coherent superposition of single photons. The state of each photon carries two bits of information. The protocol is designed so that if Eve attempts to gain information about the secret key by manipulating the coherent superposition of any photon, either the photon will reach Bob detectably late, or the photon will be in a detectably different state of superposition. In order to describe the protocol, we will follow the path of a single photon on its journey from Alice to Bob. We will assume that communications are noiseless and lossless, meaning that no additional photons are introduced into the system and no photons are taken out of the system. Further, we will assume that the channel between Alice and Bob is the shortest path between them (we will revisit this assumption later). Let us also say that Alice and Bob share synchronized clocks. Consider the setup in Figure 1. In Alice's domain, a single photon is emitted from Alice's single photon source. Since single photon sources emit photons probabilistically, the photon is emitted at a random time. Essential to the protocol is the window of time or "time bin" in which the photon will ultimately be received by Bob at one of his detectors. The amount of time it takes for a photon to reach Bob's detectors depends upon how much the photon is delayed during its travels. Next, the photon will pass through a 50-50 beam-splitter and will become a coherent superposition of a state on two separate paths. Afterwards, the states on each of the paths pass through 50-50 beamsplitters, putting the photon into a coherent superposition of a state on four separate paths. We will denote the component of the state on the bottom path by \(|a\rangle\), the component on the second to bottom path by \(|b\rangle\), the component on the second to top path by \(|c\rangle\), and the component on the top path by \(|d\rangle\). The components of the photon are then phase-shifted randomly by Alice so that the photon is in one of the four states \[|\Psi_0\rangle=\frac{1}{2}\left[ \left(|a\rangle+|b\rangle\right)+\left(|c\rangle+|d\rangle\right) \right]\] \[|\Psi_1\rangle=\frac{1}{2}\left[ \left(|a\rangle-|b\rangle\right)+\left(|c\rangle-|d\rangle\right) \right]\] \[|\Psi_2\rangle=\frac{1}{2}\left[ \left(|a\rangle+|b\rangle\right)-\left(|c\rangle+|d\rangle\right) \right]\] \[|\Psi_3\rangle=\frac{1}{2}\left[ \left(|a\rangle-|b\rangle\right)-\left(|c\rangle-|d\rangle\right) \right]\] Note that there are loops in Fig. 1 which symbolize additional lengths in the path of each component of the photon. If the distance between Alice and Bob is \(D\), we will take the length of each loop to be \(D\). We will later show that the length of each loop can be made very short. If the length is equal to \(D\), no two components of the state are accessible to Eve at the same time. On Bob's receiving end of the setup, he adds length to the paths of the components of the state so that he can bring them all to one point in space and time. Bob then interferes the \(|a\rangle\) and \(|b\rangle\) components with a 50-50 beamsplitter, and interferes the \(|c\rangle\) and \(|d\rangle\) components with a second 50-50 beamsplitter. If Alice sent the state \(|\Psi_1\rangle\) or \(|\Psi_3\rangle\), the two interfered components of the state will end up on Branch 1. Alternatively, if Alice sent the state \(|\Psi_0\rangle\) or \(|\Psi_2\rangle\), the two interfered components of the state will end up on Branch 2. Whichever branch the components end up on, the two interfered components of the state are then interfered with one another by a final 50-50 beamsplitter. Note that if Alice sent either the state \(|\Psi_2\rangle\) or \(|\Psi_3\rangle\), the photon will be received by one of the two detectors marked Detector 1, and if Alice sent either the state \(|\Psi_0\rangle\) or \(|\Psi_1\rangle\) the photon will be received by one of the two detectors marked Detector . Once Bob has received all of the states that Alice has sent, he broadcasts to Alice the time at which he received each state (i.e., the time that each photon reached one of his detectors). Additionally, for each state that he received, Bob decides randomly whether to communicate on which branch he received the state (either Branch 1 or Branch 2) or at which detector he received the state (either Detector 1 or Detector 2). Note that since Alice is randomly choosing the states which she sends to Bob, there is no correlation between the "which-branch\" information and the "which-detector\" information. As a result, if for a particular state Bob broadcasts to Alice the which-branch information, they will use the undisclosed which-detector information as a bit for their shared secret key. On the other hand, if for a particular state Bob broadcasts to Alice the which-detector information, they will use the undisclosed which-branch information as a bit for their shared secret key. If Alice tells Bob that he either received a state at an incorrect time or that some which-branch or which-detector information was incorrect, then Bob knows that an eavesdropper must have been present. In the protocol, Alice and Bob can achieve a key generation rate equal to the repetition rate of Alice's single photon source since two bits are encoded onto the spatio-temporal modes of a single photon, and only one of those bits is discarded for security. To achieve a key generation rate greater than the repetition rate of the source, the protocol may be modified by increasing the number of paths into which the photon is split. In the modified protocol, Alice will put each photon in an equal superposition of a state on \(2^k\) separate paths where one component at a time is sent to Bob's domain. Alice can choose to add a relative phase of \(\pi\) to combinations of components of the state, leading to \(2^k\) states that she can send to Bob. Once Bob has received all of the components of the photon, he interferes them such that the photon will end up at one of \(2^k\) detectors, each one corresponding to one of the \(2^k\) states that Alice can send. It follows that in this modified protocol, the superposition state of each photon encodes \(k\) bits, and some subset of those bits can be crosschecked in order to detect Eve. Setting \(k=2\), we recover the protocol which is the subject of this paper. In practice, single photon sources are expensive and very difficult to fabricate while channels, such as fiber optic cables, are an abundant resource. Therefore, it is most efficient for quantum key distribution protocols to optimize the number of secure bits that can be encoded per each emitted photon. Our protocol provides a significant practical advantage over other existing protocols since it optimizes the number of bits that are encoded per each photon by using higher-dimensional channels. ## Security proof The security proof which follows is inspired by the structure of a proof given by Goldenberg and Vaidman for their own protocol \[4\]. We will prove that if Eve tries to gain any information about the secret key from any photon sent from Alice to Bob, then either Bob will receive the photon at the wrong time, or the photon will be in a detectably different state of superposition. Eve cannot add or subtract photons from the system without inducing detectable noise or loss. The only way that Eve could attempt to avoid detection is by preserving both the phase and the timing of a photon's wavefunction. We will consider two times \(t_1\) and \(t_2\), the first of which is before any component of the photon has left Alice's domain, and the second of which is after all components of the photon have entered Bob's domain. The most general operation that Eve can perform on the state is that of a superoperator. In this case, we take \(\textbf{U}\) to be one of the Kraus components of Eve's superoperator which takes a state at time \(t_1\) to a state at time \(t_2\). For all four possible states that Alice can send, the free time-evolution with no eavesdropper is \[|\Psi_0(t_1)\rangle \longrightarrow |\Psi_0(t_2)\rangle =\frac{1}{2}\left[ \left(|a(t_2)\rangle+|b(t_2)\rangle\right)+\left(|c(t_2)\rangle+|d(t_2)\rangle\right) \right]\] \[|\Psi_1(t_1)\rangle \longrightarrow |\Psi_1(t_2)\rangle = \frac{1}{2}\left[ \left(|a(t_2)\rangle-|b(t_2)\rangle\right)+\left(|c(t_2)\rangle-|d(t_2)\rangle\right) \right]\] \[|\Psi_2(t_1)\rangle \longrightarrow |\Psi_2(t_2)\rangle =\frac{1}{2}\left[ \left(|a(t_2)\rangle+|b(t_2)\rangle\right)-\left(|c(t_2)\rangle+|d(t_2)\rangle\right) \right]\] \[|\Psi_3(t_1)\rangle \longrightarrow |\Psi_3(t_2)\rangle = \frac{1}{2}\left[ \left(|a(t_2)\rangle-|b(t_2)\rangle\right)-\left(|c(t_2)\rangle-|d(t_2)\rangle\right) \right]\] Since Eve does not know whether Bob will communicate to Alice the which-branch or which-detector information, she must prepare for either scenario. In the case Bob tells Alice the which-branch information for a given state, Eve must make the absolute squares of the amplitudes of the \(|a(t_2)\rangle\) and \(|b(t_2)\rangle\) components have the same value, and the absolute squares of the amplitudes of the \(|c(t_2)\rangle\) and \(|d(t_2)\rangle\) components also have the same value. Additionally, the relative phase between the \(|a(t_2)\rangle\) and \(|b(t_2)\rangle\) components as well as the relative phase between the \(|c(t_2)\rangle\) and \(|d(t_2)\rangle\) components must remain unchanged. In the case Bob shares the which-detector information for a given state, Eve must make sure that the sum of the absolute squares of the amplitudes of the \(|a(t_2)\rangle\) and \(|b(t_2)\rangle\) components equals the sum of the absolute squares of the amplitudes of the \(|c(t_2)\rangle\) and \(|d(t_2)\rangle\) components. Further, the relative phase between the \(|a(t_2)\rangle\) and \(|c(t_2)\rangle\) components as well as the relative phase between the \(|b(t_2)\rangle\) and \(|d(t_2)\rangle\) components must remain unchanged. So as to account for both scenarios, it follows that Eve must make the magnitude of each of the final amplitudes equal to a constant \(C\). Eve must also preserve the relative phases between any two components of the state. Let \(|\Phi(t)\rangle\) be the state of an auxiliary system which Eve uses to extract information. It follows that for \(i=0,1,2,3\), the general form of the evolution from time \(t_1\) to \(t_2\) is \[\textbf{U}\,|\Psi_0(t_1)\rangle \, |\Phi(t_1)\rangle= 2C |\Psi_0(t_2)\rangle \, |\Phi_0(t_2)\rangle\] \[\textbf{U}\,|\Psi_1(t_1)\rangle \, |\Phi(t_1)\rangle= 2C |\Psi_1(t_2)\rangle \, |\Phi_1(t_2)\rangle\] \[\textbf{U}\,|\Psi_2(t_1)\rangle \, |\Phi(t_1)\rangle= 2C |\Psi_2(t_2)\rangle \, |\Phi_2(t_2)\rangle\] \[\textbf{U}\,|\Psi_3(t_1)\rangle \, |\Phi(t_1)\rangle= 2C |\Psi_3(t_2)\rangle \, |\Phi_3(t_2)\rangle\] where \(|\Phi_i (t_2)\rangle\) is the state of Eve's auxiliary system at time \(t_2\) after it has been acted upon by \(\textbf{U}\) in the case that Alice sends the state \(|\Psi_i \rangle\). Note that if \[|\Phi_0(t_2)\rangle=|\Phi_1(t_2)\rangle=|\Phi_2(t_2)\rangle=|\Phi_3(t_2)\rangle\] it is impossible for Eve to extract information. Consider a photon at time \(t_1\) in the state \[|d(t_1)\rangle =\frac{1}{2}\left(|\Psi_0(t_1)\rangle-|\Psi_1(t_1)\rangle-|\Psi_2(t_1)\rangle+|\Psi_3(t_1)\rangle \right)\] It follows that \(\textbf{U}\,|d(t_1)\rangle \, |\Phi(t_1)\rangle\) can be written as \[\begin{aligned} \textbf{U}\,|d(t_1)\rangle \, |\Phi(t_1)\rangle =C \,[ |a\rangle \,(|\Phi_0(t_2)\rangle-|\Phi_1(t_2)\rangle-|\Phi_2(t_2)\rangle+|\Phi_3(t_2)\rangle)\,\nonumber\\ +|b\rangle \,(|\Phi_0(t_2)\rangle+|\Phi_1(t_2)\rangle-|\Phi_2(t_2)\rangle-|\Phi_3(t_2)\rangle)\,\,\nonumber\\ +|c\rangle \,(|\Phi_0(t_2)\rangle-|\Phi_1(t_2)\rangle+|\Phi_2(t_2)\rangle-|\Phi_3(t_2)\rangle)\,\,\nonumber\\ +|d\rangle \,(|\Phi_0(t_2)\rangle+|\Phi_1(t_2)\rangle+|\Phi_2(t_2)\rangle+|\Phi_3(t_2)\rangle)] \end{aligned}\] This equation expresses that unless \[\begin{aligned} |\Phi_0(t_2)\rangle-|\Phi_1(t_2)\rangle-|\Phi_2(t_2)\rangle+|\Phi_3(t_2)\rangle=0 \nonumber \\ \nonumber \\ |\Phi_0(t_2)\rangle+|\Phi_1(t_2)\rangle-|\Phi_2(t_2)\rangle-|\Phi_3(t_2)\rangle=0 \nonumber\\ \nonumber\\ |\Phi_0(t_2)\rangle-|\Phi_1(t_2)\rangle+|\Phi_2(t_2)\rangle-|\Phi_3(t_2)\rangle=0 \end{aligned}\] the photon could be measured in a state \(|a(t_2)\rangle\), \(|b(t_2)\rangle\), or \(|c(t_2)\rangle\). However, the photon cannot be measured to be in any of these three states. The reason is that \(\textbf{U}\) is an operator applied by Eve that takes a photon in a state \(|d(t_1)\rangle\) to a coherent superposition of the states \(|a(t_2)\rangle\), \(|b(t_2)\rangle\), \(|c(t_2)\rangle\) and \(|d(t_2)\rangle\). But a photon in a state \(|a(t_2)\rangle\), \(|b(t_2)\rangle\), or \(|c(t_2)\rangle\) must have exited Eve's domain before the component in the state \(|d(t_1)\rangle\) has entered it. Thus, for Eve to preserve causality, the coefficients of \(|a\rangle\), \(|b\rangle\) and \(|c\rangle\) in Eq. (15) must be zero which immediately leads to Eq. (16), or equivalently \[|\Phi_0(t_2)\rangle=|\Phi_1(t_2)\rangle=|\Phi_2(t_2)\rangle=|\Phi_3(t_2)\rangle\] and so Eve cannot extract any information. This completes the proof. Note that a security proof of the modified protocol presented at the end of Section II.A is a straightforward extension of the above proof. # Implementation and Variations ## Shortening the Loop Length In the above explanation of the protocol, we assumed that the length of each loop in the setup was \(D\) where \(D\) is the distance between Alice and Bob. In practice, having such long loops would decrease the key generation rate dramatically, and would also lead to photon loss due to the large distances that each component of the photon would have to travel. The solution to this problem was originally proposed by Goldenberg and Vaidman in the context of their own protocol \[4\]. Let us say that the accuracy of the time of Alice and Bob's measurements is \(\Delta t\). Therefore, if Bob receives a photon that has been delayed by any amount of time greater than or equal to \(\Delta t\) relative to Alice's sending time, he and Alice will be able to detect this delay. Let us now say that the length of each loop is \(c\,(\Delta t+\epsilon)\) for some positive \(\epsilon\). It is easy to see that the terms \(|a(t_2)\rangle\), \(|b(t_2)\rangle\), or \(|c(t_2)\rangle\) still cannot appear in \(\textbf{U}\,|d(t_1)\rangle \, |\Phi(t_1)\rangle\). Therefore, the security proof will still hold for loop lengths of \(c\,(\Delta t+\epsilon)\), and it is practical to incorporate this change into the protocol. ## Mode Coupling and Sub-optimal Paths In implementing the protocol experimentally, it is not practical to have components of a photon travel along four separate paths over the large distances that may separate the domains of Alice and Bob. Optical fibers tend to expand and contract with small temperature changes, which can change the relative phase between the components of the photon thereby affecting how they will eventually interfere with one another in Bob's domain. Atmospheric turbulence provides similar difficulties for free-space transmission of photons. Both problems can seriously degrade the effectiveness of the protocol, and potentially reduce the secret key generation rate. The problem can be solved by using switches to couple the four paths to four temporal modes of one fiber or one free-space transmission line. The photon then travels in a single spatial mode from Alice's domain to Bob's domain. By analogy, imagine four trains traveling on four separate tracks. By using track switches to bring the trains onto a single track, one can align the trains so that they follow one another. The process can later be reversed, allowing the trains to be switched back onto four separate tracks. Similarly, the four components of the photon traveling along the four paths are coupled to one spatial mode using switches S\(1\), S\(2\), and S\(3\) in Alice's domain (Figure 2). The photons then travel one after the other in a single fiber or free-space path until reaching Bob's domain where switches S\(4\), S\(5\), and S\(6\) couple their temporal modes back into four spatial modes. Recall that the protocol assumes that the path of the photons between the domains of Alice and Bob is a straight line. In satellite communication, this assumption is correct since the communication is line-of-sight. However, in fiber optic communications, fibers are never laid down in straight lines. If the path of the photons is not a straight line, an eavesdropper could create a shorter straight-line path between Alice and Bob thereby reducing the travel time of the photons. As a result, the eavesdropper can allow herself extra time in which to measure and delay Alice's photons and still send them on to Bob such that he will receive the photons at the expected time. Therefore, using a suboptimal path could mask the presence of an eavesdropper. The protocol can be modified to address the problem of a suboptimal path. For such a modification, Alice and Bob must take into account the difference in length between their suboptimal path and an optimal straight-line path between them. The difference in distance between the two paths will be designated as \(\mu\). To make the protocol secure and account for this extra distance, each loop in the setup must be increased in length by at least \(\mu\). However, this change need not decrease the efficiency of the protocol since a setup in which the temporal modes are coupled to the spatial mode of a single fiber could be designed such that components of one photon are spatially interlaced with the components of other photons. It is important to note that Alice and Bob's switches and setup of interferometers have to be designed to function in accordance with the way that the photon components are interlaced. # Conclusion In our information age, large amounts of data need to be transmitted securely and rapidly for commercial and military purposes. The increasing power of computers makes conventional non-quantum cryptography protocols susceptible to attack. Quantum cryptography offers benefits over standard methods of cryptography because it provides fundamental security. The protocol presented in this paper is a new method for relativistic orthogonal states quantum key distribution which provides security and achieves a key generation rate higher than other known QKD protocols. The efficiency of the proposed protocol makes it an attractive alternative to existing QKD protocols for both secure fiber optic as well as secure satellite communication.

{'timestamp': '2014-06-12T02:11:15', 'yymm': '1401', 'arxiv_id': '1401.5493', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5493'}

# Introduction Nowadays practitioners often face very large data sets, which are difficult to analyze directly. In this context, clustering is an important tool providing a partition among individuals. Other approaches cluster both individuals and variables simultaneously. Furthermore, with the increasing number of variables at hand, the risk of observing correlated descriptors, even within the same class, is often high. In view of these difficulties, the practitioner can choose between two approaches. The first approach consists in performing a selection among the observed variables in order to extract uncorrelated data, thereby losing some potentially crucial information. The second approach consists in applying a method for modeling the conditional dependencies on the whole set of variables. Clustering methods can be divided into two kinds of approaches: geometric ones based on the distances between individuals and probabilistic ones modelling the data generation process. Geometric methods are generally faster than probabilistic ones, but they are often quite sensitive to the choice of distance between individuals. Furthermore, as the probabilistic tools are not available for these methods, difficult questions, e.g. the selection of the number of clusters cannot be addressed rigorously. For categorical data, geometric methods either define a metric in the initial variables space e.g. *k-means*, or compute their metric on the axes of the multiple correspondence analysis. Many geometric approaches can also be interpreted in a probabilistic way. Thus, for the continuous data, the classical *k-means* algorithm can be identified as an homoscedastic Gaussian mixture model with equal proportions. For the categorical variables, show that the CEM algorithm, applied to a classical latent class model, maximizes a classical information criteria close to a \(\chi^2\) metric. Other links between the two approaches are described in, Chapter 9. Let us now introduce our proposal for this problem. In the categorical case, the *latent class model* also known as naive Bayes belongs to the folklore. In this article, we refer to this model as the conditional independence model (further denoted by [cim]{.smallcaps}). Classes are explicitly described by the probability of each modality for each variable under the conditional independence assumption. The sparsity of the model, implied by this assumption, is a great advantage since it restricts the curse of dimensionality. [cim]{.smallcaps} has obtained quite good results in practice in different areas, e.g. behavioral science and medicine. However, [cim]{.smallcaps} may suffer from severe biases when the data are intra-class correlated. For instance, an application presented by shows that [cim]{.smallcaps} over-estimates the number of clusters when the conditional independence assumption is violated. For a long time, people have tried to relax the conditional independence assumption by modeling conditional interactions between variables using an additive model. The main drawback of this approach is that its number of parameters becomes huge, making their estimation intractable. Some other methods take into account the intra-class dependencies as *mixtures of Bayesian networks*. Conditionally on each class, a directed acyclic graph is built with a set of nodes representing each variable. However, if no constraint is added, the network estimation is also quite complex. By constraining the network to be a tree, the model selection and the parameter estimation can be easily performed. Moreover the correlation model enjoys great flexibility. The extension of the dependency tree of was done by for the supervised classification and by for the clustering. However the main problem of these models is that they too often require too many parameters. When covariates are available, the conditional dependencies between the categorical ones can be modeled by a logistic function. By assuming that these covariates are unobserved, the *multilevel latent class model* naturally incorporates the intra-class dependencies. This model has connections with the approach of where the intra-class dependencies are modeled by a latent continuous variable with a probit function. The *hybrid model* is more general approach in which, for each class, a factor analysis model is fitted to either all categorical variables or to those categorical variables having dependencies. Recently, proposed the *mixture model of latent traits analyzers* which assumes that the distribution of the categorical variables depends on both a categorical latent variable (the class) and many continuous latent traits variables. The parameter estimation is also a difficult point which is solved via a variational approach. All these models consider the intra-class dependencies, but their main drawback is that these dependencies must be interpreted among relations with a latent variable. As a result, pertinent interpretation can be difficult. The log-linear models were originally proposed to model the log-probability of each individual by selecting interactions between variables. The most general mixture model is the *log-linear mixture model* as it is able to incorporate many forms of interactions. It has been used at least since. used it to cluster radiographic cross-diagnostics and in a market segmentation problem. However this model family is huge and the model selection is a real challenge. In the literature, authors often require the modeled interactions ahead of time. Another option is to perform a deterministic search e.g. the *forward* method which is sub-optimal. Furthermore, the number of parameters to estimate increases with the conditional modalities interactions, thus implying potential over-fitting and more difficult interpretation. The latent class model ([cim]{.smallcaps}) can be seen as a particular log-linear mixture model, in which interactions are discarded. Our aim is to present a version of the log-linear mixture model which takes into account the interactions of order one or more while keeping the number of unknown parameters to a reasonable amount. We propose to extend the classical latent class model ([cim]{.smallcaps}) for categorical data, by a new latent class model which relaxes the conditional independence assumption. We refer to this new model as the *conditionally correlated model* (denoted by [cmm]{.smallcaps}). This model is a parsimonious version of the log-linear mixture model, and thus benefits from its interpretative power. Furthermore, we propose a Bayesian approach to automatically perform model selection. The [ccm]{.smallcaps} model groups the variables into conditionally independent blocks given the class. The main intra-class dependencies are thus shown by the repartition of the variables into blocks. This approach, allowing modeling of the main conditional interactions, was first proposed by in order to cluster continuous and categorical data. For [cmm]{.smallcaps}, each block follows a particular dependency distribution which corresponds to our main contribution. This distribution consists in a bi-component mixture of an *independence* and a *maximal dependency* distribution according to the Cramer's V criterion. This specific distribution of the blocks allows summarizing the conditional dependencies of the variables with only one continuous parameter: the maximum dependency distribution proportion. Thus, the model underlines the main conditional dependencies and their strength. The new model is a two-degree parsimonious version of a log-linear mixture model. A first degree of parsimony is introduced by grouping the variables which are conditionally dependent into the same block. This repartition of the variables per blocks defines the interactions considered by the model for each class. Moreover, the strength of the correlation is reflected by the proportion of maximum dependency distribution. A second degree of parsimony is induced by the specific distribution of the blocks. As for all log-linear mixture models, the selection of the pertinent interactions is a combinatorial problem. We propose to perform this model selection via a Gibbs sampler in order to overcome the enumeration of all the models. Thus, this general approach could also select the interactions of any log-linear mixture model. This paper is organized as follows. Section [\[latentclass\]](#latentclass){reference-type="ref" reference="latentclass"} reviews the principles of the latent class model. Section [\[ourmodel\]](#ourmodel){reference-type="ref" reference="ourmodel"} presents the new mixture model taking into account the intra-class correlations. Section [\[estimation\]](#estimation){reference-type="ref" reference="estimation"} is devoted to parameter estimation in the case where the number of classes and the blocks of variables are supposed to be known. Section [\[modelsel\]](#modelsel){reference-type="ref" reference="modelsel"} presents a Gibbs algorithm for avoiding the combinatorial difficulties inherent to block selection. Section [\[simulations\]](#simulations){reference-type="ref" reference="simulations"} presents results on simulated data. Section [\[applications\]](#applications){reference-type="ref" reference="applications"} presents a comparison between two main model-based clustering approaches and our proposal on a classical medical data set, and presents another application on a larger real data set. A tutorial of the R package `Clustericat`[^1] which performs the model selection and the estimation of the parameters of [cmm]{.smallcaps} is given with the first application (see Appendix [\[tuto\]](#tuto){reference-type="ref" reference="tuto"}). Section [\[conclu\]](#conclu){reference-type="ref" reference="conclu"} presents our conclusions. # Classical models[\[latentclass\]]{#latentclass label="latentclass"} ## Latent class model: intra-class independence of variables Observations to be classified are described with \(d\) discrete variables \(\boldsymbol{x}=(\boldsymbol{x}^1,\ldots,\boldsymbol{x}^d)\) defined on the probabilistic space \(\mathcal{X}\). Each variable \(j\) has \(m_{j}\) response levels with \(m_j\geq 2\) and is written \(\boldsymbol{x}^j=(x^{j1},\ldots,x^{jm_j})\) where \(x^{jh}=1\) if variable \(j\) takes modality \(h\) and \(x^{jh}=0\) otherwise. In the standard latent class model ([cim]{.smallcaps}), the variables are assumed to be *conditionally independent* knowing the latent cluster. Furthermore, data are supposed to be drawn independently from a mixture of \(g\) multivariate multinomial distributions with probability distribution function (pdf) \[p(\textbf{\textit{x}}; \boldsymbol{\theta})=\sum_{k=1}^{g} \pi_{k} \mathring{p}(\textbf{\textit{x}} ;\boldsymbol{\alpha}_{k})\qquad \text{ with } \qquad \mathring{p}(\textbf{\textit{x}};\boldsymbol{\alpha}_{k})=\prod_{j=1}^{d}\prod_{h=1}^{m_{j}}(\alpha_{k}^{jh})^{x^{jh}},\] with \(\boldsymbol{\theta}=(\pi_1,\ldots,\pi_g,\boldsymbol{\alpha}_1,\ldots,\boldsymbol{\alpha}_g)\), \(\pi_k\) being the proportion of the component \(k\) in the mixture where \(\pi_k>0\) and \(\sum_{k=1}^g \pi_k=1\), and \(\boldsymbol{\alpha}_{k} =(\alpha_{k}^{jh};j = 1, \ldots, d;h = 1,\ldots, m_{j})\) where \(\alpha_{k}^{jh}\) denotes the probability that variable \(j\) has level \(h\) if the object is in cluster \(k\) and satisfies the two following constraints: \(\alpha_{k}^{jh}>0\) and \(\sum_{h=1}^{m_j}\alpha_k^{jh}=1\). The classical latent class model is much more parsimonious than the saturated log-linear model, which requires \((\prod_j m_j)-1\) parameters, since it only requires \(\nu_{\textsc{cim}}\) parameters with \[\nu_{\textsc{cim}}=(g-1)+g\sum_{k=1}^g (m_j-1). \label{nbparamcim}\] Its maximum likelihood estimator is easily computed via an EM algorithm. For the cluster analysis, the mixture identifiability up to a permutation of the class is generally necessary. However, there are mixtures, such as the products of Bernoulli distributions, which are not identifiable but which produce good results in several applications. In order to relax this too stringent concept of identifiability, the notion of generic identifiability was introduced by @All09: a model is generically identifiable if it is identifiable except for a subset of the parameter space with Lebesgue measure zero. ## Latent class model extension: intra-class independence of blocks {#introex} Despite its simplicity, the latent class model s to good results in many situations. However, in the case of intra-correlated variables, it can entail severe biases in the partition estimation and it may also overestimate the number of clusters. In order to reduce these biases, a classical extension of the latent class model was introduced by for conditionally correlated mixed data. This model is implemented in the Multimix software. It considers that *conditionally* on the class \(k\), variables are grouped into \(\textsc{b}_k\) *independent blocks* and that each block follows a specific distribution. The repartition in blocks of the variables determines a partition \(\boldsymbol{\sigma}_k=(\boldsymbol{\sigma}_{k1},\ldots,\boldsymbol{\sigma}_{k\textsc{b}_k})\) of \(\{1,\ldots,d\}\) in \(\textsc{b}_k\) disjoint non-empty subsets where \(\boldsymbol{\sigma}_{kb}\) represents the subset \(b\) of variables in the partition \(\boldsymbol{\sigma}_k\). This partition defines \(\boldsymbol{x}^{\{kb\}}=\boldsymbol{x}^{\boldsymbol{\sigma}_{kb}}=(\boldsymbol{x}^{\{kb\}j};j=1,\ldots,d^{\{kb\}})\) which is the subset of \(\boldsymbol{x}\) associated to \(\boldsymbol{\sigma}_{kb}\). The integer \(d^{\{kb\}}=\text{card}(\boldsymbol{\sigma}_{kb})\) is the number of variables in block \(b\) of component \(k\) and \(\boldsymbol{x}^{\{kb\}j}=({x}^{\{kb\}jh};h=1,\ldots,m_j^{\{kb\}})\) corresponds to variable \(j\) of block \(b\) for component \(k\) with \(x^{\{kb\}jh}=1\) if the individual takes modality \(h\) for variable \(\boldsymbol{x}^{\{kb\}j}\) and \(x^{\{kb\}jh}=0\) otherwise and where \(m_j^{\{kb\}}\) represents the number of modalities of \(\boldsymbol{x}^{\{kb\}j}\). Note that different repartitions of the variables into blocks are allowed for each component and they are grouped into \(\boldsymbol{\sigma}=(\boldsymbol{\sigma}_1,\ldots,\boldsymbol{\sigma}_g)\). For each component \(k\), each block \(b\) follows a specific parametric distribution denoted as \(p(\textbf{\textit{x}}^{\{kb\}};\boldsymbol{\theta}_{kb})\) where \(\boldsymbol{\theta}_{kb}\) groups the parameters of this distribution. The model pdf is written as \[p(\textbf{\textit{x}};\boldsymbol{\sigma},\boldsymbol{\theta})= \sum_{k=1}^{g} \pi_{k} p(\textbf{\textit{x}};\boldsymbol{\sigma}_k,\boldsymbol{\theta}_{k})\qquad \text{ with } \qquad p(\textbf{\textit{x}};\boldsymbol{\sigma}_{k},\boldsymbol{\theta}_{k})=\prod_{b=1}^{B_{k}}p(\textbf{\textit{x}}^{\{kb\}};\boldsymbol{\theta}_{kb}),\] where \(\boldsymbol{\theta}\) is redefined as \(\boldsymbol{\theta}=(\pi_1,\ldots,\pi_g,\boldsymbol{\theta}_1,\ldots,\boldsymbol{\theta}_g)\) with \(\boldsymbol{\theta}_k = (\boldsymbol{\theta}_{k1},\ldots,\boldsymbol{\theta}_{k\textsc{b}_k})\). Figure [\[exeblo\]](#exeblo){reference-type="ref" reference="exeblo"} presents an example of the distribution with conditional independent blocks for a mixture with two components described by five variables. Blank cells indicate that the intra-class correlation is neglected and black cells indicate that this correlation is taken into account. For instance, Figure [\[exeblo\]](#exeblo){reference-type="ref" reference="exeblo"}.a illustrates the distribution of the class 1 where two blocks (\(\textsc{b}_1=2\)) are considered. More precisely, the first block is composed by the first two variables (\(\boldsymbol{\sigma}_{11}=\{1,2\}\)) and the second block is composed by the last three variables (\(\boldsymbol{\sigma}_{12}=\{3,4,5\}\)). Note that the classical latent class model with conditional independence, would be represented by white cells off the diagonal and black cells on the diagonal. This approach is very general, since any distribution can be chosen for each block as soon as it is different from the distribution of independence. The mixture model by conditional independent blocks is a parsimonious version of the log-linear mixture model. Indeed, the distribution of variables in blocks determines which interactions are modeled. The interactions between variables of different blocks will be zero and those between variables of the same block can be modeled by the specific distribution of the block. The limiting case of this model where \(\textsc{b}_k=d\) for each class is equivalent to the latent class model with the conditional independence assumption. The generic identifiability of the mixture model with conditionally independent blocks follows, under specific constraints, from Theorem 4 of by assuming that the distribution of each block is itself identifiable. This proof is given in Appendix A of. # Intra-block parsimonious distribution[\[ourmodel\]]{#ourmodel label="ourmodel"} The goal is now to define a parsimonious distribution for each block that takes into account the dependency between variables. Furthermore, the parameters of the distribution inside the block must be meaningful for the practitioner. In this context, we propose to model the distribution of each block by a mixture of the extreme distributions according to the Cramer's V criterion computed on all the variable couples. The model results in a bi-component mixture between an independence distribution and a maximum dependency distribution which can be easily interpreted by the user. The maximum dependency distribution is introduced first. The resulting conditional correlated model (CCM) is also defined as a block model extension of the latent class model where the distribution inside the block is modeled by this bi-component mixture. Remark: Without loss of generality, the variables are considered as ordered by decreasing number of modalities in each block: ∀(k, b) mkb j ≥ mkb j The goal is now to define a parsimonious distribution for each block that takes into account the dependency between variables. Furthermore, the parameters of the distribution inside the block must be meaningful for the practitioner. In this context, we propose to model the distribution of each block by a mixture of the extreme distributions according to the Cramer's V criterion computed on all the variable couples. The model results in a bi-component mixture between an independence distribution and a maximum dependency distribution which can be easily interpreted by the user. The maximum dependency distribution is introduced first. The resulting conditional correlated model ([ccm]{.smallcaps}) is also defined as a block model extension of the latent class model where the distribution inside the block is modeled by this bi-component mixture.\ **Remark:** Without loss of generality, the variables are considered as ordered by decreasing number of modalities in each block: \(\forall(k,b)\; m_j^{\{kb\}}\geq m_{j+1}^{\{kb\}}\) where \(j=1,\ldots,d^{\{kb\}}-1\). ## Maximum dependency distribution The maximum dependency distribution is defined as the "opposite" distribution of independence according to the Cramer's V criterion computed on all the variable couples. Indeed, this latter minimizes this criterion while the maximum dependency distribution maximizes it. Under this distribution, the modality knowledge of one variable provides the maximum information on all the subsequent variables. Note that it is a non-reciprocal functional link between variables. Indeed, if \(\boldsymbol{x}^{\{kb\}}\) arises from this distribution, the knowledge of the variable having the largest number of modalities determines exactly the others but the reverse does not necessarily apply. So, this distribution defines successive surjections from the space of \(x^{\{kb\} j}\) to the space of \(x^{\{ kb \} j+1}\) with \(j=1,\ldots,d^{\{kb\}}-1\) (recall that the variables are ordered by decreasing number of modalities in each block). In fact, it is a reciprocal functional link only when \(m_j^{\{kb\}}=m_{j+1}^{\{kb\}}\). Since the first variable determines the other ones, this distribution is defined by a product between the multinomial distribution of the first variable parametrized by \(\boldsymbol{\tau}_{kb}=(\tau_{kb}^{h};h=1,\ldots,m_1^{\{kb\}})\) with \(\tau_{kb}^h\geq 0\) and \(\sum_{h=1}^{m^{\{kb\}}_1}\tau_{kb}^h =1\), and the product between the conditional distributions defined as specific multinomial distributions. So, conditionally on \(x^{\{kb\}1h}=1\), \(\boldsymbol{x}^{\{kb\}j}\) is deterministic for each \(j=2,\ldots,d^{\{kb\}}\). Indeed, in such a case, \(\boldsymbol{x}^{\{kb\}j}\) follows a specific multinomial distribution whose parameters are 0 and 1. More precisely, this distribution is parametrized by \(\boldsymbol{\delta}_{kb}^{hj}=(\delta_{kb}^{hjh'};h'=1,\ldots,m_j^{\{kb\}})\) with the following constraints defining the successive surjections: \(\delta_{kb}^{hjh'} \in \{0,1\}\), \(\sum_{h'=1}^{m_j^{\{kb\}}} \delta_{kb}^{hjh'}=1\) (multinomial distribution) and \(\sum_{h=1}^{m_1^{\{kb\}}} \delta_{kb}^{hjh'}\geq 1\) (surjections). By denoting \(\boldsymbol{\delta}_{kb}=(\boldsymbol{\delta}_{kb}^{hj};h=1,\ldots,m_1^{\{kb\}};j=2,\ldots,d^{\{kb\}})\), the pdf of maximum dependency distribution is defined as: \[\begin{aligned} \acute{p}(\boldsymbol{x}^{\{kb\}};\boldsymbol{\tau}_{kb},\boldsymbol{\delta}_{kb})&= p(\boldsymbol{x}^{\{kb\}1};\boldsymbol{\tau}_{kb}) \prod_{j=2}^{d^{\{kb\}}} p(\boldsymbol{x}^{\{kb\}j}|\boldsymbol{x}^{\{kb\}1} ;\{\boldsymbol{\delta}_{kb}^{hj}\}_{h=1,\ldots ,m^{\{kb\}}_1}) \nonumber \\ &=\prod_{h=1}^{m_1^{\{kb\}}} \Big( \tau_{kb}^h \prod_{j=2}^{d^{\{kb\}}} \prod_{h'=1}^{m_{j}^{\{kb\}}} (\delta_{kb}^{hjh'} )^{x^{\{kb\}jh'} } \Big)^{x^{\{kb\}1h}}. \end{aligned}\] Figure [\[exem\]](#exem){reference-type="ref" reference="exem"} shows two examples of the maximum dependency distributions. The probabilities of the joint distribution are represented by the area of dark boxes. Notice that \(\boldsymbol{\delta}_{kb}\) defines the position where the probabilities are non-zero (location of a dark boxes) and \(\boldsymbol{\tau}_{kb}\) defines the probabilities of this non-zero cells (area of the dark boxes). For the example illustrated in Figure [\[exem\]](#exem){reference-type="ref" reference="exem"}.a, the probability that the first variable takes the modality one is 0.1 (\(\tau_{11}^1=0.1\)). Moreover, conditionally on \(x^{\{11\}11}=1\), variable \(\boldsymbol{x}^{\{11\}2}\) is deterministic since the probability that \(x^{\{11\}21}=1\) given \(x^{\{11\}11}=1\) is one. A sufficient condition of identifiability is to impose \(\tau^{h}_{kb}>0\) for all \(h=1,\ldots,m_1^{\{kb\}}\). This distribution has very limited interest because it is so unrealistic that it can almost never be used alone. We will see in the next section how to use it in a more efficient way. ## A new block distribution: mixture of two extreme distributions[\[blockdistribution\]]{#blockdistribution label="blockdistribution"} We proposed to model the distribution of each block by a bi-components mixture between an *independence* distribution and a *maximum dependency* distribution. For block \(b\) of component \(k\), the block distribution is modeled by: \[p(\textbf{\textit{x}}^{\{kb\}};\boldsymbol{\theta}_{kb})=(1-\rho_{kb})\mathring{p}(\textbf{\textit{x}}^{\{kb\}};\boldsymbol{\alpha}_{kb})+\rho_{kb}\acute{p}(\textbf{\textit{x}}^{\{kb\}};\boldsymbol{\tau}_{kb},\boldsymbol{\delta}_{kb}),\label{DDI}\] where \(\boldsymbol{\theta}_{kb}=(\rho_{kb},\boldsymbol{\alpha}_{kb},\boldsymbol{\tau}_{kb},\boldsymbol{\delta}_{kb})\) and where \(\rho_{kb}\) is the proportion of the maximum dependency distribution in this mixture with \(0\leq\rho_{kb} \leq 1\). The proposed model requires few additional parameters compared with the conditional independence model. In addition, it is easily interpretable as explained in the next paragraph. Note that the limiting case where \(\rho_{kb}=0\) considers that the block follows an independence distribution. In this particular case, the parameters of the maximum dependency distribution are no longer defined. Under this distribution, the proportion of the maximum dependency distribution reflects the deviation from independence under the assumption that the other allowed distribution is the maximum dependency distribution. The parameter *\(\rho_{kb}\)* gives an indicator of the *inter-variables correlation* of the block. It is not here a pairwise dependency among variables but a dependency between all variables of the block. Furthermore, it stays bounded when the number of variables is larger than two while the Cramer's V is non-upper-bounded in this case. The *intra-variables dependencies* between the variables are defined by \(\boldsymbol{\delta}_{kb}\). The strength of these dependencies is explained by \(\boldsymbol{\tau}_{kb}\) since it gives the *weight of the over-represented modalities crossing* compared with the independence distribution. Above, we interpreted the distribution by conditionally independent blocks as a parsimonious version of the log-linear mixture model because it determines the interactions to be modeled for each class. By choosing the proposed distribution for blocks, a second level of parsimony is added. Indeed, among the interactions allowed by this distribution with independent blocks, only those corresponding to the maximum dependency distribution will be modeled. Other interactions are considered as null. #### Properties: {#properties .unnumbered} - The [ccm]{.smallcaps}, stays parsimonious compared with [cim]{.smallcaps} since, for each block with at least two variables, the number of the additional parameters depends only on the number of modalities of the first variable of the block and not on the number of variables in the block. By using \(\nu_{\textsc{cim}}\) defined in Equation [\[nbparamcim\]](#nbparamcim){reference-type="eqref" reference="nbparamcim"}, the number of parameters of [ccm]{.smallcaps} is denoted \(\nu_{\textsc{ccm}}\) by: \[\nu_{\textsc{ccm}}=\nu_{\textsc{cim}}+\sum_{\{(k,b)|d^{\{kb\}}>1\}}m_1^{\{kb\}}.\] - The proposed distribution is identifiable under the condition that the block is composed by at least three variables (\(d^{\{kb\}}>2\)) or that the modality number of the last variable of the block is more than two (\(m_2^{\{kb\}}>2\)). This result is demonstrated in Appendix B of. The parameter \(\rho_{kb}\) is a new indicator allowing measuring the correlation between variables, not limited to correlation between variable couples. In the case where the identifiability conditions cannot be met, we distinguish two cases. If \(d^{\{kb\}}=1\), then block \(b\) contains only one variable, and the proposed model is reduced to model a multinomial distribution, \(\rho_{kb}=0\) and the maximum dependency distribution is not defined. If \(d^{\{kb\}}=2\) and \(m_2^{\{kb\}}=2\) then a new constraint is added. In order to have the most meaningful parameters, the chosen value of \(\rho_{kb}\) is the largest value maximizing the log-likelihood. This additional constraint does not falsify the definition of \(\rho_{kb}\) as an indicator of the dependency strength between the variables of the same block. Furthermore, this constraint is natural since blocks with the biggest dependencies are wanted. Note that \(\rho_{kb}\) seems to be correlated with the Cramer's V. An example is given in Section 3 of. - Note that the marginal probabilities given class can be straightforwardly deduced from the parameters of the model. # Estimation of the parameters [\[estimation\]]{#estimation label="estimation"} For a fixed model \((g,\boldsymbol{\sigma})\), the parameters must be estimated. Since the proposed distribution [ccm]{.smallcaps} has two latent variables (the class membership and the intra-block distribution membership), two algorithms derived from the EM algorithm perform the estimation of the associated continuous parameters. The combinatorial problems arising from the consideration of the discrete parameters are avoided by using a Metropolis-Hastings algorithm. ## Global GEM algorithm [\[EMglo\]]{#EMglo label="EMglo"} The whole data set consisting of \(n\) independent and identically distributed individuals is denoted by \(\textbf{x} = (\textbf{\textit{x}}_{1},\ldots\textbf{\textit{x}}_{n} )\) where \(\textbf{\textit{x}}_{i} \in \mathcal{X}\). The objective is to obtain the maximum log-likelihood estimator \(\hat{\boldsymbol{\theta}}\) defined as (\(g\) is now implicit in each expression) \[\hat{\boldsymbol{\theta}}=\text{argmax}_{\boldsymbol{\theta}} L(\boldsymbol{\theta}; \mathbf{x},\boldsymbol{\sigma}) \quad \text{with} \quad L(\boldsymbol{\theta}; \mathbf{x},\boldsymbol{\sigma})=\sum_{i=1}^{n} \ln \Big( p(\textbf{\textit{x}}_{i};\boldsymbol{\sigma},\boldsymbol{\theta}) \Big). \label{vrai}\] The search for maximum likelihood estimates for mixture models entails solving equations having no analytical solutions. For the mixture models, the assignments of the individuals into the classes can be considered as missing data. So, the tool generally used is the Expectation-Maximization algorithm (denoted EM algorithm) and its extensions [@Dem77; @Mcl97]. We denote the unknown indicator vectors of the \(g\) clusters by \(\textbf{z} = (z_{ik};i=1,\ldots,n;k=1,\ldots,g)\) where \(z_{ik} = 1\) if \(\textbf{\textit{x}}_{i}\) arises from cluster \(k\), \(z_{ik} = 0\) otherwise. Thus, the mixture model distribution corresponds to the marginal distribution of the random variable \(\textbf{X}\) obtained from the couple distribution of the random variables \((\textbf{X},\textbf{Z})\). In order to maximize the log-likelihood, the EM algorithm uses the complete-data log-likelihood which is defined as \[L_{\text{c}}(\boldsymbol{\theta}; \mathbf{x},\textbf{z},\boldsymbol{\sigma})=\sum_{i=1}^{n}\sum_{k=1}^{g} z_{ik} \ln \Big( \pi_{k} p(\textbf{\textit{x}}_{i};\boldsymbol{\sigma}_k,\boldsymbol{\theta}_{k}) \Big). \label{vraic}\] The EM algorithm is an iterative algorithm which alternates between two steps: the computation of the complete-data log-likelihood conditional expectation (E step) and its maximization (M step). Many algorithms are derived from the EM algorithm and among them the Generalized EM algorithm (GEM) is of interest to us. It works on the same principle as the EM algorithm, but the maximization step is replaced by a GM step where the proposed parameters increase the expectation of the complete-data log-likelihood according to its previous value without necessarily maximizing it. We prefer to use the GEM algorithm, since the maximization step in the EM algorithm requires estimating the continuous parameters for too many possible values of the discrete parameters in order to warrant the maximization of the complete-data log-likelihood expectation. Indeed, exhaustive enumeration for estimating the discrete parameters is generally impossible when a block contains variables with many modalities and/or many variables, as detailed below. If \(S(a,b)\) is the number of possible surjections from a set of cardinal \(a\) into a set of cardinal \(b\), then \(\boldsymbol{\delta}_{kb}\) is defined in the discrete space of dimension \(\prod_{j=1}^{d^{\{kb\}}-1} S(m_j^{\{kb\}},m_{j+1}^{\{kb\}})\). For example, a block with three variables and \(m^{\{kb\}}=(5,4,3)\) implies \(51,840\) possibilities for \(\boldsymbol{\delta}_{kb}\). Thus, a stochastic approach is proposed in Section [\[discestim\]](#discestim){reference-type="ref" reference="discestim"} to overcome this problem. Then, the estimation of the continuous parameters conditionally on the discrete parameters is performed via the classical EM algorithm presented in Section [\[continuestim\]](#continuestim){reference-type="ref" reference="continuestim"} since their estimation cannot be obtained in closed form. At iteration \((r)\), the steps of the global GEM can be written as: - **\(\text{E}_{\text{global}}\) step:** \({z}_{ik}^{(r)}=\frac{\pi_k^{(r)}p(\boldsymbol{x}_i;\boldsymbol{\sigma}_k,\boldsymbol{\theta}_k^{(r)})}{\sum_{k'=1}^g \pi_{k'}^{(r)}p(\boldsymbol{x}_i;\boldsymbol{\sigma}_{k'},\boldsymbol{\theta}_{k'}^{(r)})}\), - **\(\text{GM}_{\text{global}}\) step:** \(\pi^{(r+1)}_{k}=\frac{n_k^{(r)}}{n}\) where \(n_k^{(r)}=\sum_{i=1}^n {z}_{ik}^{(r)}\) and \(\forall (k,b)\) \(\boldsymbol{\theta}_{kb}^{(r+1)}\) is updated under the constraint that the conditional expectation of complete-data log-likelihood increases (see Sections [\[discestim\]](#discestim){reference-type="ref" reference="discestim"} and [\[continuestim\]](#continuestim){reference-type="ref" reference="continuestim"}). #### Initialization of the algorithm: {#initialization-of-the-algorithm .unnumbered} Since this algorithm is performed in an stochastic algorithm used for the model selection (see Section [\[modelsel\]](#modelsel){reference-type="ref" reference="modelsel"}) and since this latter has an influence on the GEM initialization, this point will be detailed in Section [\[MCMCapprox\]](#MCMCapprox){reference-type="ref" reference="MCMCapprox"}. #### Stopping criterion: {#stopping-criterion .unnumbered} The GEM algorithm is stopped after \(r_{\max}\) iterations and we fix \(\hat{\boldsymbol{\theta}}=\boldsymbol{\theta}^{(r_{\max})}\). ## Details of the \(\text{GM}_{\text{global}}\) step of the GEM[\[discestim\]]{#discestim label="discestim"} The maximization of the expected complete-data log-likelihood is done by optimizing its terms for each \((k,b)\). Thus, the determination of \(\boldsymbol{\theta}_{kb}^{(r+1)}\) is performed independently of the parameters of the other blocks. A Metropolis-Hastings algorithm is also performed, for each \((k,b)\), to avoid the combinatorial problems induced by the detection of the discrete parameters \(\boldsymbol{\delta}_{kb}\). It performs a random walk over the discrete parameters space and computes the maximum likelihood estimators of the continuous parameters \((\rho_{kb},\boldsymbol{\alpha}_{kb},\boldsymbol{\tau}_{kb})\) associated with them. This stochastic algorithm allows finding the estimator maximizing the expected complete-data log-likelihood of block \(b\) for component \(k\): \[\underset{\boldsymbol{\theta}_{kb}}{\text{argmax }} \sum_{i=1}^n z_{ik}^{(r)} \ln p(\boldsymbol{x}^{\{kb\}}_i;\boldsymbol{\theta}_{kb}).\] At each iteration \((s)\) of this Metropolis-Hastings algorithm, a discrete parameter denoted by \(\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})}\) is sampled with a uniform distribution in a neighborhood of \(\boldsymbol{\delta}_{kb}^{(r,s)}\) denoted as \(\Delta(\boldsymbol{\delta}_{kb}^{(r,s)})\). Then the continuous parameters \((\rho_{kb}^{(r,s+\frac{1}{2})},\boldsymbol{\alpha}_{kb}^{(r,s+\frac{1}{2})},\) \(\boldsymbol{\tau}_{kb}^{(r,s+\frac{1}{2})})\) are computed, conditionally on the value of \(\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})}\), in order to maximize the expected complete-data log-likelihood of block \(b\) for component \(k\): \[\sum_{i=1}^n z_{ik}^{(r)} \ln p(\boldsymbol{x}^{\{kb\}}_i;\rho_{kb},\boldsymbol{\alpha}_{kb},\boldsymbol{\tau}_{kb},\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})} ).\] The candidate parameters are now denoted by \(\boldsymbol{\theta}_{kb}^{(r,s+\frac{1}{2})} =(\rho_{kb}^{(r,s+\frac{1}{2})} ,\boldsymbol{\alpha}_{kb}^{(r,s+\frac{1}{2})} ,\boldsymbol{\tau}_{kb}^{(r,s+\frac{1}{2})} ,\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})} )\). The whole block parameters \(\boldsymbol{\theta}_{kb}^{(r,s+1)}\) of the next step are then defined as \(\boldsymbol{\theta}_{kb}^{(r,s+\frac{1}{2})}\) with the acceptance probability \(\mu^{(r,s+1)}\) and \(\boldsymbol{\theta}_{kb}^{(r,s)}\) otherwise, where: \[\mu^{(r,s+1)}=\min\left\{\frac{ \prod_{i=1}^n p(\boldsymbol{x}^{\{kb\}}_i;\boldsymbol{\theta}_{kb}^{(r,s+\frac{1}{2})} )^{z_{ik}^{(r)}}|\Delta(\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})} )|}{ \prod_{i=1}^n p(\boldsymbol{x}^{\{kb\}}_i;\boldsymbol{\theta}_{kb}^{(r,s)})^{z_{ik}^{(r)}}|\Delta(\boldsymbol{\delta}_{kb}^{(r,s)})| },1\right\},\] \(|\Delta(\boldsymbol{\delta}_{kb}^{(r,s)})|\) denoting the cardinal of \(\Delta(\boldsymbol{\delta}_{kb}^{(r,s)})\). Thus, at iteration \((s)\), the algorithm performs the three following steps: - **Stochastic step on \(\boldsymbol{\delta}_{kb}\):** generate \(\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})}\) with a uniform distribution among the elements of \(\Delta(\boldsymbol{\delta}_{kb}^{(r,s)})\), - **Maximization step on the continuous parameters (\(\text{M}_{\boldsymbol{\theta}}\) step):** compute the continuous parameters of \(\boldsymbol{\theta}_{kb}^{(r,s+\frac{1}{2})}\) (see Section [\[continuestim\]](#continuestim){reference-type="ref" reference="continuestim"}), - **Stochastic step on \(\boldsymbol{\theta}_{kb}\):** sample \(\boldsymbol{\theta}_{kb}^{(r,s+1)}=\left\{ \begin{array}{rl} \boldsymbol{\theta}_{kb}^{(r,s+\frac{1}{2})} & \text{ with probability } \mu^{(r,s+1)} \\ \boldsymbol{\theta}_{kb}^{(r,s)} & \text{ otherwise. } \end{array}\right.\) Neighborhood \(\Delta(\boldsymbol{\delta}_{kb}^{(r,s)})\) is defined as the set of parameters where at most two surjections are different from that of \(\boldsymbol{\delta}_{kb}^{(r,s)}\). Figure [\[voisindelta\]](#voisindelta){reference-type="ref" reference="voisindelta"} illustrates this definition. #### Initialization of the algorithm: {#initialization-of-the-algorithm-1 .unnumbered} The initialization of the algorithm is done by \(\boldsymbol{\theta}_{kb}^{(r+1,0)}=\boldsymbol{\theta}_{kb}^{(r)}\). #### Stopping criterion: {#stopping-criterion-1 .unnumbered} This algorithm is stopped after a number of iterations \(s_{\max}\). The parameter \(\boldsymbol{\theta}_{kb}^{(r+1)}=\boldsymbol{\theta}_{kb}^{(r+1,\tilde{s})}\) is returned with \(\tilde{s}=\underset{s}{\text{argmax}} \sum_{i=1}^n z_{ik}^{(r)} \ln p(\boldsymbol{x}^{\{kb\}}_i;\boldsymbol{\theta}_{kb}^{(r,s)})\). Thus, the proposed initialization ensures an increased likelihood at each iteration of the GEM algorithm. #### Remark: {#remark .unnumbered} When the space of possible \(\boldsymbol{\delta}_{kb}\) is small (for example when the block groups a small number of binary variables), an exhaustive approach obtains the same results as the proposed algorithm with less computation time. Thus, the retained approach (exhaustive or stochastic) depends on the number of variables and modalities. ## Details of the \(\text{M}_{\boldsymbol{\theta}}\) step of the \(\text{GM}_{\text{global}}\) step[\[continuestim\]]{#continuestim label="continuestim"} As there is a second level of mixing, another EM algorithm can be performed for the continuous parameters \((\rho_{kb},\boldsymbol{\alpha}_{kb},\boldsymbol{\tau}_{kb})\) estimation by introducing other unknown vectors corresponding to the indicator of the blocks distributions conditionally on \(\textbf{z}\). These vectors are denoted by \(\textbf{y}=(\textbf{y}^{\{kb\}};k=1,\ldots,g;b=1,\ldots,\textsc{b}_k)\) with \(\textbf{y}^{\{kb\}}=(y_{1}^{\{kb\}}...,y_{n}^{\{kb\}})\) where \(y_{i}^{\{kb\}}=1\) if \(\textbf{\textit{x}}_{i}^{\{kb\}}\) arises from the *maximum dependency* distribution for block \(b\) of cluster \(k\) and \(y_{i}^{\{kb\}}=0\) if \(\textbf{\textit{x}}_{i}^{\{kb\}}\) arises from the *independence* distribution for block \(b\) of cluster \(k\). The whole mixture model distribution corresponds to the marginal distribution of the random variable \(\textbf{X}\) obtained from the triplet distribution of the random variables \((\textbf{X},\textbf{Y},\textbf{Z})\). Since the blocks are independent conditionally on \(\textbf{Z}\), the *full* complete-data log-likelihood (both in \(\textbf{Y}\) and \(\textbf{Z}\)) is defined as: At iteration \((t)\), the local EM algorithm estimates the continuous parameters of block \(b\), with fixed values of \(\textbf{z}^{(r)}\) and \(\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})}\), by the following two steps: - **\(\text{E}_{\text{local}}\) step:** \(y_{i}^{\{kb\}(r,s+\frac{1}{2},t)}=\frac{\rho_{kb}^{(r,s+\frac{1}{2},t)}\acute{p}(\textbf{\textit{x}}_{i}^{\{kb\}};\boldsymbol{\tau}_{kb}^{(r,s+\frac{1}{2},t)},\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})})}{p(\textbf{\textit{x}}^{\{kb\}}_i;\rho_{kb}^{(r,s+\frac{1}{2},t)},\boldsymbol{\alpha}_{kb}^{(r,s+\frac{1}{2},t)},\boldsymbol{\tau}_{kb}^{(r,s+\frac{1}{2},t)},\boldsymbol{\delta}_{kb}^{(r,s+\frac{1}{2})} )}\), - **\(\text{M}_{\text{local}}\) step:** \(\rho_{kb}^{(r,s+\frac{1}{2},t+1)}=\frac{n_{kb}^{(r,s+\frac{1}{2},t)}}{n_{k}^{(r)}},\) \(\tau_{kb}^{(r,s+\frac{1}{2},t+1)}=\frac{\sum_{i=1}^{n}z_{ik}^{(r)}y_{i}^{\{kb\}(r,s+\frac{1}{2},t)} x_{i}^{\{kb\}1h}}{n_{kb}^{(r,s+\frac{1}{2},t)}}\),\ \(\alpha_{kb}^{(r,s+\frac{1}{2},t+1)}=\frac{\sum_{i=1}^{n}z_{ik}^{(r)}(1-y_{i}^{\{kb\}(r,s+\frac{1}{2},t)})x_{i}^{\{kb\}jh}}{n_{k}^{(r)}-n_{kb}^{(r,s+\frac{1}{2},t)}},\) where \(n_{kb}^{(r,s+\frac{1}{2},t)}=\sum_{i=1}^{n}z_{ik}^{(r)}y_{i}^{\{kb\}(r,s+\frac{1}{2},t)}\). #### Conjecture: {#conjecture .unnumbered} During our numerous experiments, we empirically noticed that the log-likelihood function of the mixture between the independence and the maximum dependency distributions had a unique optimum. We conjecture that this function has indeed a unique maximum. #### Initialization of the algorithm: {#initialization-of-the-algorithm-2 .unnumbered} The previous conjecture allows to perform only one initialization of the EM algorithm fixed to: \((\rho_{kb}^{(r,s+\frac{1}{2},0)},\boldsymbol{\alpha}_{kb}^{(r,s+\frac{1}{2},0)},\boldsymbol{\tau}_{kb}^{(r,s+\frac{1}{2},0)})=(\rho_{kb}^{(r,s)},\boldsymbol{\alpha}_{kb}^{(r,s)},\) \(\boldsymbol{\tau}_{kb}^{(r,s)})\). #### Stopping criterion: {#stopping-criterion-2 .unnumbered} This algorithm is stopped after a number of iterations denoted by \(t_{\max}\) and returns the value of block parameters \(\boldsymbol{\theta}_{kb}^{(r,s+\frac{1}{2})}\) defined as \(\boldsymbol{\theta}_{kb}^{(r,s+\frac{1}{2})}=\boldsymbol{\theta}_{kb}^{(r,s+\frac{1}{2},t_{\max})}\). #### Remark: {#remark-1 .unnumbered} In the specific case where \(\boldsymbol{\delta}_{kb}\) are known for each \((k,b)\), the estimation of all the continuous parameters could be performed by a unique EM algorithm where, at iteration \((r)\), the E step would compute both \(\textbf{z}^{(r)}\) and \(\textbf{y}^{(r)}\) while the M step would estimate all the parameters maximizing the expectation of the *full* complete-data log-likelihood. # Model selection [\[modelsel\]]{#modelsel label="modelsel"} ## Gibbs algorithm for exploring the space of models Since the number of components \(g\) determines the dimension of \(\boldsymbol{\sigma}\), the model construction is done in two steps. Firstly, the selection of the number of components and, secondly, the determination of the variable repartition per blocks for each component. In a Bayesian context, the best model \((\hat{g},\hat{\boldsymbol{\sigma}})\) is defined as: \[(\hat{g},\hat{\boldsymbol{\sigma}})=\underset{g,\boldsymbol{\sigma}}{\text{argmax }}p(g,\boldsymbol{\sigma}|\textbf{x}).\] Thus, by considering that \(p(g)=\frac{1}{g_{\max}}\) if \(g\leq g_{\max}\) and \(0\) otherwise, where \(g_{\max}\) is the maximum number of classes allowed by the user, and by assuming that \(p(\boldsymbol{\sigma} | g)\) follows a uniform distribution, the best model is also defined as: \[(\hat{g},\hat{\boldsymbol{\sigma}})=\underset{g}{\text{argmax }} \Big[ \underset{\boldsymbol{\sigma}}{\text{argmax }} p(\textbf{x}|g,\boldsymbol{\sigma}) \Big].\] To obtain \((\hat{g},\hat{\boldsymbol{\sigma}})\), a Gibbs algorithm is used for estimating \(\text{argmax}_{\boldsymbol{\sigma}}\; p(\textbf{x}|g,\boldsymbol{\sigma})\), for each value of \(g\in\{1,\ldots,g_{max}\}\). Indeed, this method limits the combinatorial problem involved by the detection of the block structure of variables. A reversible jump method could be used, but this approach is rarely performed with mixed parameters (continuous and discrete). Indeed, in such a case, it is difficult to define a mapping between the parameters space of two models. So, we propose to use an easier Gibbs sampler-type having \(p(\boldsymbol{\sigma}|\textbf{x},g)\) as stationary distribution. It alternates between two steps: the generation of a stochastic neighborhood \(\Sigma^{[q]}\) conditionally on the current model \(\boldsymbol{\sigma}^{[q]}\) by a proposal distribution and the generation of a new pattern \(\boldsymbol{\sigma}^{[q+1]}\) included in \(\Sigma^{[q]}\) with a probability proportional to its posterior probability. At iteration \([q]\), it is written as: - **Neighborhood step:** generate a stochastic neighborhood \(\Sigma^{[q]}\) by a proposal distribution given below conditionally on the current model \(\boldsymbol{\sigma}^{[q]}\), - **Pattern step:** ::: {.small} \(\boldsymbol{\sigma}^{[q+1]} \sim p(\boldsymbol{\sigma}|\textbf{x},g, \Sigma^{[q]})\) with \(p(\boldsymbol{\sigma}|\textbf{x},g, \Sigma^{[q]})=\left\{\begin{array}{rl} \frac{p(\textbf{x}|g,\boldsymbol{\sigma})}{\sum_{\boldsymbol{\sigma '}\in \Sigma^{[q]}}p(\textbf{x}|g,\boldsymbol{\sigma '})} & \text{ if } \boldsymbol{\sigma}\in \Sigma^{[q]}\\ 0 & \text{ otherwise.} \end{array}\right.\) ::: A possible deterministic neighborhood of \(\boldsymbol{\sigma}^{[q]}\) could be defined as the set of models where, at most one variable is affected, for one component, in another block (possibly creating a new block): \(\left\{ \boldsymbol{\sigma}: \exists ! (k,b,j) \; j \in \boldsymbol{\sigma}_{kb}^{[q]} \text{ and } j \notin \boldsymbol{\sigma}_{kb} \right\} \cup \left\{\boldsymbol{\sigma}^{[q]} \right\}\). However, as this deterministic neighborhood can be very large, our proposal distribution allows reducing it to a stochastic neighborhood \(\Sigma^{[q]}\) by reducing the number of \((k,b)\) where \(\boldsymbol{\sigma}_{kb}\) could be different to \(\boldsymbol{\sigma}_{kb}^{[q]}\). Thus, one component \(k^{[q]}\) is randomly sampled in \(\{1,\ldots,g\}\) then one block \(b_{from}^{[q]}\) is randomly sampled in \(\{1,\ldots,B_{k^{[q]}}^{[q]}\}\). Another block \(b^{[q]}\) is randomly sampled in \(\{1,\ldots,B_{k^{[q]}}^{[q]}\} \setminus b_{from}^{[q]}\) and the set \(b_{to}^{[q]}=\{b^{[q]},B_{k^{[q]}}^{[q]}+1\}\) is built. The stochastic neighborhood \(\Sigma^{[q]}\) is then defined as: We denote the elements of \(\Sigma^{[q]}\) as \(\boldsymbol{\sigma}^{[q+\varepsilon(e)]}\) where \(\varepsilon(e)=\frac{e}{| \Sigma^{[q]} |+1}\) and \(e=1,\ldots,| \Sigma^{[q]} |\). Figure [\[voisinsigma\]](#voisinsigma){reference-type="ref" reference="voisinsigma"} shows an illustration of this definition. At the generation pattern step, the previous algorithm needs the value of \(p(\textbf{x}|g,\boldsymbol{\sigma})\) \(\forall \boldsymbol{\sigma} \in \Sigma^{[q]}\). By using the BIC approximation, this probability is approximated by: \[\ln p(\textbf{x}|g,\boldsymbol{\sigma}) \simeq L(\hat{\boldsymbol{\theta}};\textbf{x},g,\boldsymbol{\sigma})-\frac{\nu_{\text{\textsc{ccm}}}}{2}\log(n),\] \(\hat{\boldsymbol{\theta}}\) being the maximum likelihood estimator obtained by the GEM algorithm previously described in Section [\[estimation\]](#estimation){reference-type="ref" reference="estimation"}. Thus, at the iteration \([q]\), for each \(e=1,\ldots,| \Sigma^{[q]} |\), estimator \(\hat{\boldsymbol{\theta}}^{[q+\varepsilon(e)]}\) associated to element \(\boldsymbol{\sigma}^{[q+\varepsilon(e)]}\) is computed by the GEM algorithm. #### Initialization: {#initialization .unnumbered} Whatever the initial value selected for \(\boldsymbol{\sigma}^{[0]}\), the algorithm converges to the same value of \(\boldsymbol{\sigma}\). However, this convergence can be very slow when the initialization is poor. Since blocks consist of the most correlated variables, a Hierarchical Ascendant Classification (HAC) is used on the matrix of Cramer's V distances on the variable couples. The partition produced by the HAC minimizing the block number without blocks consisting of more than four variables is chosen for each \(\boldsymbol{\sigma}^{[0]}_k\). For the initialization, the variables number of a block is limited to four, because very few blocks having more than four variables were exhibited in the course of our experiments. Obviously, the Gibbs algorithm can then violate this initial constraint if necessary. #### Stopping criterion: {#stopping-criterion-3 .unnumbered} The algorithm is stopped when \(q_{\max}\) successive iterations have not discovered a better model. ## Consequences of the Gibbs algorithm on the GEM algorithm [\[MCMCapprox\]]{#MCMCapprox label="MCMCapprox"} #### Initialization of the GEM algorithm: {#initialization-of-the-gem-algorithm .unnumbered} At iteration \([q]\) of the Gibbs algorithm, the GEM algorithm estimates \(\hat{\boldsymbol{\theta}}^{[q+\varepsilon(e)]}\) associated to model \(\boldsymbol{\sigma}^{[q+\varepsilon(e)]}\) for \(e=1,\ldots,| \Sigma^{[q]} |\). Since these models are close to \(\boldsymbol{\sigma}^{[q]}\), their maximum likelihood estimators should be closed to \(\hat{\boldsymbol{\theta}}^{[q]}\). The GEM algorithm initialization is also done by the value of \(\hat{\boldsymbol{\theta}}^{[q]}\) for the not modified blocks. Thus, \(\boldsymbol{\theta}_{kb}^{[q+\varepsilon(e)](0)}=\hat{\boldsymbol{\theta}}_{kb}^{[q]}\) if the blocks are not modified (\(\boldsymbol{\sigma}_{kb}^{[q+\varepsilon(e)]}=\boldsymbol{\sigma}_{kb}^{[q]}\)). For the other blocks, the continuous parameters are randomly sampled. For those blocks, in order to avoid the combinatorial problems, we use a sequential method to initialize \(\boldsymbol{\delta}_{kb}^{[q+\varepsilon(e)](0)}\): the surjections from \(\boldsymbol{x}^{\{kb\}1}\) to \(\boldsymbol{x}^{\{kb\}j}\) are sampled, according to \(\textbf{x}\) and to the continuous parameters previously sampled \((\rho_{kb}^{[q+\varepsilon(e)](0)},\boldsymbol{\alpha}_{kb}^{[q+\varepsilon(e)](0)},\boldsymbol{\tau}_{kb}^{[q+\varepsilon(e)](0)})\), for each \(j=2,\ldots,d^{\{kb\}}\) as follows: \[\begin{small} \boldsymbol{\delta}_{kb}^{.j[q+\varepsilon(e)](0)} \propto \prod_{i=1}^{n} p(x_i^{\{kb\}1},x_i^{\{kb\}j};\rho_{kb}^{[q+\varepsilon(e)](0)},\boldsymbol{\alpha}_{kb}^{1[q+\varepsilon(e)](0)},\boldsymbol{\alpha}_{kb}^{j[q+\varepsilon(e)](0)},\boldsymbol{\tau}_{kb}^{[q+\varepsilon(e)](0)},\boldsymbol{\delta}_{kb}^{.j})^{z_{ik}^{[q]}}, \label{initdelta} \end{small}\] \(\text{ where } \boldsymbol{\delta}_{kb}^{.j[q+\varepsilon(e)]}=(\boldsymbol{\delta}_{kb}^{hj[q+\varepsilon(e)]};h=1,\ldots,m_1^{\{kb\}})\) and where \(z_{ik}^{[q]}=E \big[Z_{ik} | \boldsymbol{x}_i, \boldsymbol{\theta}^{[q]} \big]\). #### Remark about \(r_{\max}\): {#remark-about-r_max .unnumbered} As said in Section [\[EMglo\]](#EMglo){reference-type="ref" reference="EMglo"}, the algorithm is stopped after a fixed number of iterations \(r_{\max}\). If the algorithm is stopped before its convergence, the proposed initialization limits the problems. Indeed, if the model has a high *a posteriori* probability, it will stay in neighborhood \(\Sigma^{[q]}\) during some successive iterations, so its log-likelihood will increase. # Simulations [\[simulations\]]{#simulations label="simulations"} Table [1](#valparam){reference-type="ref" reference="valparam"} presents the adjustment parameters values used for all the simulations. Table [2](#simulkull){reference-type="ref" reference="simulkull"} shows the mean and the standard deviation of the Kullback-Leibler divergence between the parameters used for the dataset generation and the estimated parameters according to the number of variables. When \(n\) increases, the Kullback-Leibler divergence converges to zero. It confirms the good behavior of the proposed algorithm. For the [cim]{.smallcaps}, the BIC criterion selects a high number of classes, since it selected eight classes. The interpretation of the clusters is also difficult and we can assume that the estimator's quality is very poor. Figure [\[resumGD\]](#resumGD){reference-type="ref" reference="resumGD"} helps the interpretation for the [ccm]{.smallcaps} with five components (best model according to the BIC criterion). Its interpretation is the same as the interpretation of Figure [\[resumdentists\]](#resumdentists){reference-type="ref" reference="resumdentists"}. For example, this figure shows that the first class has a proportion of \(0.29\) and that it is composed of four blocks. The most correlated block of the first class has \(\rho_{kb}\simeq 0.80\) and the strength of the biggest modalities link is also close to \(0.85\). This block consists of the variables *TDC* and *TRM*. Here is now a possible interpretation of Class 1 (note that the others classes are also meaningful; see details in ): - **General:** This class has a proportion equal to 0.29 and consists of three blocks of dependency and one block of independence. - **Block 1:** There is a strong correlation (\(\rho_{11}\)) between the variables diarrhea treatment of the calf and mother preventive treatment against respiratory disease, especially between the modality no treatment against the calf diarrhea and the absence of preventive treatment against respiratory disease of its mother (\(\boldsymbol{\tau}_{11}\) and \(\boldsymbol{\delta}_{11}\)). - **Block 2:** There is a strong correlation (\(\rho_{12}\)) between the variables treatment against respiratory illness of the calf and mother preventive treatment against diarrhea, especially between the modality preventive treatment against respiratory illness of the calf and the presence of diarrhea preventive treatment of its mother (\(\boldsymbol{\tau}_{12}\) and \(\boldsymbol{\delta}_{12}\)). - **Block 3:** There exists another strong link between the behavior of the mother, the emptying of the umbilical and its disinfection (\(\boldsymbol{\tau}_{13}\) and \(\boldsymbol{\delta}_{13}\)). - **Block 4:** This block is characterized by absence of preventive treatment against omphalitis and having \(50\%\) of the calves infected by this illness (\(\boldsymbol{\alpha}_{14}\)). # Conclusion [\[conclu\]]{#conclu label="conclu"} By using the block extension of the latent class model, a new mixture model is proposed for clustering categorical data by taking into account the intra-class correlation. The block distribution is defined as a mixture between an independent distribution and a maximum dependency distribution. This specific distribution, which remains parsimonious, is compared to the full latent class model and allows different levels of interpretation. The blocks of variables detect the conditional dependencies between variables while their strengths are reflected by the proportions of maximum dependency distribution. The parameters of the block distribution reflect the links and the strength between modalities. The parameter estimation and the model selection are simultaneously performed via a Gibbs sample-type algorithm. It allows reducing the combinatorial problems of the block structure detection and the links between modalities search for the estimation of the maximum dependency distribution. The results are good when the number of modalities is small for each variable. For more than six modalities, the detection of other links encounters some persistent difficulties. So the algorithm can be slow in this case. The proposed approach to estimate the block structure is not adapted for data sets with many variables. A deterministic but sub-optimal solution could be used to perform a forward algorithm. The R package *Clustericat* allows clustering categorical data sets by using [cmm]{.smallcaps}. This package is available on Rforge at the following url *https://r-forge.r-project.org/R/?group_id=1803*. The proposed model can be easily extended to the case of ordinal data. For this purpose, some additional constraints on the dependency structure of each distribution of maximum dependency need to be added.

{'timestamp': '2014-07-11T02:09:31', 'yymm': '1401', 'arxiv_id': '1401.5684', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5684'}

# Introduction High-dimensional data have become commonplace in broad application areas, and there is an exponentially increasing literature on statistical and computational methods for big data. In such settings, it is well known that classical methods such as maximum likelihood estimation break down, motivating a rich variety of alternatives based on penalization and thresholding. There is a rich theoretical literature justifying the optimality properties of such penalization approaches, with fast algorithms and compelling applied results leading to routine use of \(L_1\) regularization in particular. The overwhelming emphasis in this literature has been on rapidly producing a point estimate with good empirical and theoretical properties. However, in many applications, it is crucial to obtain a realistic characterization of uncertainty in the estimates of parameters and functions of the parameters, and in predictions of future outcomes. Usual frequentist approaches to characterize uncertainty, such as constructing asymptotic confidence regions or using the bootstrap, can break down in high-dimensional settings. For example, in regression when the number of subjects \(n\) is much less than the number of predictors \(p\), one cannot naively appeal to asymptotic normality and resampling from the data may not provide an adequate characterization of uncertainty. Most penalization approaches have a Bayesian interpretation as corresponding to the mode of a posterior distribution obtained under a shrinkage prior. For example, the wildly popular Lasso/\(L_1\) regularization approach to regression is equivalent to maximum *a posteriori* (MAP) estimation under a Gaussian linear regression model having a double exponential (Laplace) prior on the coefficients. Given this connection, it is natural to ask whether we can use the entire posterior distribution to provide a probabilistic measure of uncertainty. In addition to providing a characterization of uncertainty, a Bayesian perspective has distinct advantages in terms of tuning parameter choice, allowing key penalty parameters to be marginalized over the posterior distribution instead of relying on cross-validation. In addition, by inducing penalties through shrinkage priors, important new classes of penalties can be discovered that may outperform usual \(L_q\)-type choices. From a frequentist perspective, we would like to be able to choose a default shrinkage prior that leads to similar optimality properties to those shown for \(L_1\) penalization and other approaches. However, instead of showing that a particular penalty leads to a point estimator having a minimax optimal rate of convergence under sparsity assumptions, we would rather like to show that the entire posterior distribution concentrates at the optimal rate, i.e., the posterior probability assigned to a shrinking neighborhood of the true parameter value converges to one, with the neighborhood size proportional to the frequentist minimax rate. An amazing variety of shrinkage priors have been proposed in the Bayesian literature; however with essentially no theoretical justification for the performance of these priors in the high-dimensional settings for which they were designed. and provided conditions on the prior for asymptotic normality of linear regression coefficients allowing the number of predictors \(p\) to increase with sample size \(n\), with requiring a very slow rate of growth and assuming \(p \le n\). These results required the prior to be sufficiently flat in a neighborhood of the true parameter value, essentially ruling out shrinkage priors. considered shrinkage priors in providing simple sufficient conditions for posterior consistency in linear regression where the number of variables grows slower than the sample size, though no rate of contraction was provided. In studying posterior contraction in high-dimensional settings, it becomes clear that it is critical to understand several aspects of the prior distribution on the high-dimensional space, including (but not limited to) the prior concentration around sparse vectors and the implied dimensionality of the prior. Specifically, studying the reduction in dimension induced by shrinkage priors is challenging due to the lack of exact zeros, with the prior draws being sparse in only an approximate sense. This substantial technical hurdle has prevented any previous results (to our knowledge) on posterior concentration in high-dimensional settings for shrinkage priors. In fact, investigating these properties is critically important not just in studying frequentist optimality properties of Bayesian procedures but for Bayesians in obtaining a better understanding of the behavior of their priors and choosing associated hyperparameters. Without such technical handle, it becomes an art to use intuition and practical experience to indirectly induce a shrinkage prior, while focusing on Gaussian scale families for computational tractability. Some beautiful classes of priors have been proposed by among others, with showing that essentially all existing shrinkage priors fall within the Gaussian global-local scale mixture family. One of our primary goals is to obtain theory that can allow evaluation of existing priors and design of novel priors, which are appealing from a Bayesian perspective in allowing incorporation of prior knowledge and from a frequentist perspective in leading to minimax optimality under weak sparsity assumptions. # A new class of shrinkage priors {#sec:prior_prop} ## **Bayesian sparsity priors in normal means problem** For concreteness, we focus on the widely studied normal means problem (see, for example, and references therein); although most of the ideas developed in this paper generalize directly to high-dimensional linear and generalized linear models. In the normal means setting, one aims to estimate a \(n\)-dimensional mean[^1] based on a single observation corrupted with i.i.d. standard normal noise: \[\begin{aligned} \label{eq:norm_means} y_i &= \theta_i + \epsilon_i, \quad \epsilon_i \sim \mbox{N}(0, 1), \quad 1 \leq i \leq n. \end{aligned}\] Let \(l_0[q; n]\) denote the subset of \(\mathbb{R}^n\) given by \[l_0[q; n] = \{ \theta \in \mathbb{R}^n ~:~ \#(1 \leq j \leq n: \theta_j \neq 0) \leq q\}.\] For a vector \(x \in \mathbb{R}^r\), let \(\norm{x}_2\) denote its Euclidean norm. If the true mean \(\theta_0\) is \(q_n\)-sparse, i.e., \(\theta_0 \in l_0[q_n; n]\), with \(q_n = o(n)\), the *squared minimax rate* in estimating \(\theta_0\) in \(l_2\) norm is known to be \(2q_n \log(n/q_n)(1 + o(1))\), i.e.[^2] \[\begin{aligned} \label{eq:minmax} \inf_{\hat{\theta}} \sup_{\theta_0 \in l_0[q_n; n]} E_{\theta_0} \| \hat{\theta}-\theta_0 \|_2^2 \asymp q_n \log(n/q_n). \end{aligned}\] In the above display, \(E_{\theta_0}\) denotes an expectation with respect to a \(\mbox{N}_n(\theta_0, \mathrm{I}_n)\) density. In the presence of sparsity, one thus only looses a logarithmic factor (in the ambient dimension) as a penalty for not knowing the locations of the zeroes. Moreover, [\[eq:minmax\]](#eq:minmax){reference-type="eqref" reference="eq:minmax"} implies that one only needs a number of replicates in the order of the sparsity to consistently estimate the mean. Appropriate thresholding/penalized estimators achieve the minimax rate [\[eq:minmax\]](#eq:minmax){reference-type="eqref" reference="eq:minmax"}; for a somewhat comprehensive list of *minimax-optimal* estimators, refer to. For a subset \(S \subset \{1, \ldots, n\}\), let \(|S|\) denote the cardinality of \(S\) and define \(\theta_S = (\theta_j: j \in S)\) for a vector \(\theta \in \mathbb{R}^n\). Denote \(\mbox{supp}(\theta)\) to be the *support* of \(\theta\), the subset of \(\{1, \ldots, n\}\) corresponding to the non-zero entries of \(\theta\). For a high-dimensional vector \(\theta \in \mathbb{R}^n\), a natural way to incorporate sparsity in a Bayesian framework is to use point mass mixture priors: \[\begin{aligned} \label{eq:point_mass} \theta_j \sim (1-\pi) \delta_0 + \pi g_{\theta}, \quad j = 1, \ldots, n, \end{aligned}\] where \(\pi = \mbox{Pr}(\theta_j \neq 0)\), \(\bbE\{ |\mbox{supp}(\theta)| \mid \pi\} = n \pi\) is the prior guess on model size (sparsity level), and \(g_{\theta}\) is an absolutely continuous density on \(\mathbb{R}\). These priors are highly appealing in allowing separate control of the level of sparsity and the size of the signal coefficients. If the sparsity parameter \(\pi\) is estimated via empirical Bayes, the posterior median of \(\theta\) is a minimax-optimal estimator which can adapt to arbitrary sparsity levels as long as \(q_n = o(n)\). In a fully Bayesian framework, it is common to place a beta prior on \(\pi\), leading to a beta-Bernoulli prior on the model size, which conveys an automatic multiplicity adjustment. In a beautiful recent paper, established that prior ([\[eq:point_mass\]](#eq:point_mass){reference-type="ref" reference="eq:point_mass"}) with an appropriate beta prior on \(\pi\) and suitable tail conditions on \(g_{\theta}\) leads to a minimax optimal rate of *posterior contraction*, i.e., the posterior concentrates most of its mass on a ball around \(\theta_0\) of squared radius of the order of \(q_n \log (n/q_n)\): \[\begin{aligned} \label{eq:post_conc} E_{\theta_0} \bbP(\norm{\theta-\theta_0}_2 < M s_n \mid y) \to 1, \, \text{as} \, n \to \infty, \end{aligned}\] where \(M > 0\) is a constant and \(s_n^2 = q_n \log(n/q_n)\). obtained consistency in model selection using point-mass mixture priors with appropriate data-driven hyperparameters. ## **Global-local shrinkage rules** Although point mass mixture priors are intuitively appealing and possess attractive theoretical properties, posterior sampling requires a stochastic search over an enormous space, leading to slow mixing and convergence. Computational issues and consideration that many of the \(\theta_j\)s may be small but not exactly zero has motivated a rich literature on continuous shrinkage priors; for some flavor of the vast literature refer to. noted that essentially all such shrinkage priors can be represented as global-local (GL) mixtures of Gaussians, \[\begin{aligned} \label{eq:lg} \theta_j \sim \mbox{N}(0, \psi_j \tau), \quad \psi_j \sim f, \quad \tau \sim g, \end{aligned}\] where \(\tau\) controls global shrinkage towards the origin while the local scales \(\{ \psi_j \}\) allow deviations in the degree of shrinkage. If \(g\) puts sufficient mass near zero and \(f\) is appropriately chosen, GL priors in ([\[eq:lg\]](#eq:lg){reference-type="ref" reference="eq:lg"}) can intuitively approximate ([\[eq:point_mass\]](#eq:point_mass){reference-type="ref" reference="eq:point_mass"}) but through a continuous density concentrated near zero with heavy tails. GL priors potentially have substantial computational advantages over point mass priors, since the normal scale mixture representation allows for conjugate updating of \(\theta\) and \(\psi\) in a block. Moreover, a number of frequentist regularization procedures such as ridge, lasso, bridge and elastic net correspond to posterior modes under GL priors with appropriate choices of \(f\) and \(g\). For example, one obtains a double-exponential prior corresponding to the popular \(L_1\) or lasso penalty if \(f\) is an exponential distribution. However, unlike point mass priors ([\[eq:point_mass\]](#eq:point_mass){reference-type="ref" reference="eq:point_mass"}), many aspects of shrinkage priors are poorly understood, with the lack of exact zeroes compounding the difficulty in studying basic properties, such as prior expectation, tail bounds for the number of large signals, and prior concentration around sparse vectors. Hence, subjective Bayesians face difficulties in incorporating prior information regarding sparsity, and frequentists tend to be skeptical due to the lack of theoretical justification. This skepticism is somewhat warranted, as it is clearly the case that reasonable seeming priors can have poor performance in high-dimensional settings. For example, choosing \(\pi=1/2\) in prior ([\[eq:point_mass\]](#eq:point_mass){reference-type="ref" reference="eq:point_mass"}) leads to an exponentially small prior probability of \(2^{-n}\) assigned to the null model, so that it becomes literally impossible to override that prior informativeness with the information in the data to pick the null model. However, with a beta prior on \(\pi\), this problem can be avoided. In the same vein, if one places i.i.d. \(\mbox{N}(0, 1)\) priors on the entries of \(\theta\), then the induced prior on \(\norm{\theta}\) is highly concentrated around \(\sqrt{n}\) leading to misleading inferences on \(\theta\) almost everywhere. Although these are simple examples, similar *multiplicity problems* can transpire more subtly in cases where complicated models/priors are involved and hence it is fundamentally important to understand properties of the prior and the posterior in the setting of [\[eq:norm_means\]](#eq:norm_means){reference-type="eqref" reference="eq:norm_means"}. There has been a recent awareness of these issues, motivating a basic assessment of the marginal properties of shrinkage priors for a single \(\theta_j\). Recent priors such as the horseshoe and generalized double Pareto are carefully formulated to obtain marginals having a high concentration around zero with heavy tails. This is well justified, but as we will see below, such marginal behavior alone is not sufficient; it is necessary to study the joint distribution of \(\theta\) on \(\mathbb{R}^n\). With such motivation, we propose a class of Dirichlet-kernel priors in the next subsection. ## **Dirichlet-kernel priors** Let \(\phi_0\) denote the standard normal density on \(\mathbb{R}\). Also, let \(\mbox{DE}(\tau)\) denote a zero mean double-exponential or Laplace distribution with density \(f(y) = (2 \tau)^{-1} e^{-\abs{y}/\tau}\) for \(y \in \mathbb{R}\). Let us revisit the global-local specification ([\[eq:lg\]](#eq:lg){reference-type="ref" reference="eq:lg"}). Integrating out the local scales \(\psi_j\)'s, [\[eq:lg\]](#eq:lg){reference-type="eqref" reference="eq:lg"} can be equivalently represented as a global scale mixture of a kernel \(\mathcal{K}(\cdot)\), \[\begin{aligned} \label{eq:scaledkernel} \theta_j \stackrel{\text{i.i.d.}}{\sim} \mathcal{K}(\cdot \;, \tau), \quad \tau \sim g, \end{aligned}\] where \(\mathcal{K}(x) = \int \psi^{-1/2} \phi_0(x/\sqrt{\psi}) g(\psi) d\psi\) is a symmetric unimodal density (or kernel) on \(\mathbb{R}\) and \(\mathcal{K}(x, \tau) := \tau^{-1/2} \mathcal{K}(x/\sqrt{\tau})\). For example, \(\psi_j \sim \mbox{Exp}(1/2)\) corresponds to a double-exponential kernel \(\mathcal{K} \equiv \mbox{DE}(1)\), while \(\psi_j \sim \mbox{IG}(1/2,1/2)\) results in a standard Cauchy kernel \(\mathcal{K} \equiv \mbox{Ca}(0, 1)\). These traditional choices lead to a kernel which is *bounded* in a neighborhood of zero. However, if one instead uses a half Cauchy prior \(\psi_j^{1/2} \sim \mbox{Ca}_+(0,1)\), then the resulting horseshoe kernel is unbounded with a singularity at zero. This phenomenon coupled with tail robustness properties leads to excellent empirical performance of the horseshoe. However, the joint distribution of \(\theta\) under a horseshoe prior is understudied and further theoretical investigation is required to understand its operating characteristics. One can imagine that it concentrates more along sparse regions of the parameter space compared to common shrinkage priors since the singularity at zero potentially allows most of the entries to be concentrated around zero with the heavy tails ensuring concentration around the relatively small number of signals. The above class of priors rely on obtaining a suitable kernel \(\mathcal{K}\) through appropriate normal scale mixtures. In this article, we offer a fundamentally different class of shrinkage priors that alleviate the requirements on the kernel, while having attractive theoretical properties. In particular, our proposed class of Dirichlet-kernel (Dk) priors replaces the single global scale \(\tau\) in [\[eq:scaledkernel\]](#eq:scaledkernel){reference-type="eqref" reference="eq:scaledkernel"} by a vector of scales \((\phi_1\tau, \ldots, \phi_n \tau)\), where \(\phi = (\phi_1, \ldots, \phi_n)\) is constrained to lie in the \((n-1)\) dimensional simplex \(\mathcal{S}^{n-1} = \{ x = (x_1, \ldots, x_{n})^{{ \mathrm{\scriptscriptstyle T} }}: x_j \geq 0, \sum_{j=1}^{n} x_j = 1\}\) and is assigned a \(\mbox{Dir}(a, \ldots, a)\) prior: \[\begin{aligned} \label{eq:kd} \theta_j \mid \phi_j, \tau \sim \mathcal{K}(\cdot \;, \phi_j \tau), \quad \phi \sim \mathrm{Dir}(a, \ldots, a). \end{aligned}\] In [\[eq:kd\]](#eq:kd){reference-type="eqref" reference="eq:kd"}, \(\mathcal{K}\) is any symmetric (about zero) unimodal density with exponential or heavier tails; for computational purposes, we shall restrict attention to the class of kernels that can be represented as scale mixture of normals. While previous shrinkage priors in the literature obtain marginal behavior similar to the point mass mixture priors [\[eq:point_mass\]](#eq:point_mass){reference-type="eqref" reference="eq:point_mass"}, our construction aims at resembling the *joint distribution* of \(\theta\) under a two-component mixture prior. Constraining \(\phi\) on \(\mathcal{S}^{n-1}\) restrains the degrees of freedom of the \(\phi_j\)'s, offering better control on the number of dominant entries in \(\theta\). In particular, letting \(\phi \sim \mbox{Dir}(a, \ldots, a)\) for a suitably chosen \(a\) allows ([\[eq:kd\]](#eq:kd){reference-type="ref" reference="eq:kd"}) to behave like [\[eq:point_mass\]](#eq:point_mass){reference-type="eqref" reference="eq:point_mass"} jointly, forcing a large subset of \((\theta_1, \ldots, \theta_n)\) to be *simultaneously* close to zero with high probability. We focus on the Laplace kernel from now on for concreteness, noting that all the results stated below can be generalized to other choices. The corresponding hierarchical prior given \(\tau\), \[\begin{aligned} \label{eq:laplace_dir} \theta_j \mid \phi, \tau \sim \mathrm{DE}(\phi_j \tau), \quad \phi \sim \mathrm{Dir}(a, \ldots, a), \end{aligned}\] is referred to as a Dirichlet--Laplace prior, denoted \(\theta \mid \tau \sim \mbox{DL}_{a}(\tau)\). To understand the role of \(\phi\), we undertake a study of the marginal properties of \(\theta_j\) conditional on \(\tau\), integrating out \(\phi_j\). The results are summarized in Proposition [\[propWG\]](#propWG){reference-type="ref" reference="propWG"} below. Thus, marginalizing over \(\phi\), we obtain an unbounded kernel \(\mathcal{K}\), so that the marginal density of \(\theta_j \mid \tau\) has a singularity at 0 while retaining exponential tails. A proof of Proposition [\[propWG\]](#propWG){reference-type="ref" reference="propWG"} can be found in the appendix. The parameter \(\tau\) plays a critical role in determining the tails of the marginal distribution of \(\theta_j\)'s. We consider a fully Bayesian framework where \(\tau\) is assigned a prior \(g\) on the positive real line and learnt from the data through the posterior. Specifically, we assume a \(\mbox{gamma}(\lambda, 1/2)\) prior on \(\tau\) with \(\lambda = n a\). We continue to refer to the induced prior on \(\theta\) implied by the hierarchical structure, \[\begin{aligned} \label{eq:laplace_dir_tau} \theta_j \mid \phi, \tau \sim \mathrm{DE}(\phi_j \tau), \quad \phi \sim \mathrm{Dir}(a, \ldots, a), \quad \tau \sim \mbox{gamma}(na, 1/2), \end{aligned}\] as a Dirichlet--Laplace prior, denoted \(\theta \sim \mbox{DL}_a\). There is a recent frequentist literature on including a local penalty specific to each coefficient. The adaptive Lasso relies on empirically estimated weights that are plugged in. instead propose to sample the penalty parameters from a posterior, with a sparse point estimate obtained for each draw. These approaches do not produce a full posterior distribution but focus on sparse point estimates. ## **Posterior computation** {#sec:postcomp} The proposed class of DL priors leads to straightforward posterior computation via an efficient data augmented Gibbs sampler. Note that the \(\mbox{DL}_a\) prior [\[eq:laplace_dir_tau\]](#eq:laplace_dir_tau){reference-type="eqref" reference="eq:laplace_dir_tau"} can be equivalently represented as \[\begin{aligned} \theta_j \sim \mbox{N}(0, \psi_j \phi_j^2 \tau^2), \, \psi_j \sim \mathrm{Exp}(1/2), \, \phi \sim \mbox{Dir}(a, \ldots, a), \, \tau \sim \mbox{gamma}(na, 1/2). \end{aligned}\] We detail the steps in the normal means setting noting that the algorithm is trivially modified to accommodate normal linear regression, robust regression with heavy tailed residuals, probit models, logistic regression, factor models and other hierarchical Gaussian cases. To reduce auto-correlation, we rely on marginalization and blocking as much as possible. Our sampler cycles through (i) \(\theta \mid \psi, \phi, \tau, y\), (ii) \(\psi \mid \phi, \tau, \theta\), (iii) \(\tau \mid \phi, \theta\) and (iv) \(\phi \mid \theta\). We use the fact that the joint posterior of \((\psi, \phi, \tau)\) is conditionally independent of \(y\) given \(\theta\). Steps (ii)-(iv) together give us a draw from the conditional distribution of \((\psi, \phi, \tau) \mid \theta\), since \[\begin{aligned} = [\psi \mid \phi, \tau, \theta] [\tau \mid \phi, \theta] [\phi \mid \theta]. \end{aligned}\] Steps (i)--(iii) are standard and hence not derived. Step (iv) is non-trivial and we develop an efficient sampling algorithm for jointly sampling \(\phi\). Usual one at a time updates of a Dirichlet vector leads to tremendously slow mixing and convergence, and hence the joint update in Theorem [\[thm:post_comp\]](#thm:post_comp){reference-type="ref" reference="thm:post_comp"} is an important feature of our proposed prior; a proof can be found in the Appendix. Consider the following parametrization for the three-parameter generalized inverse Gaussian (giG) distribution: \(Y \sim \mbox{giG}(\lambda, \rho, \chi)\) if \(f(y) \propto y^{\lambda-1} e^{-0.5 (\rho y + \chi/y)}\) for \(y > 0\). The summary of each step are finally provided below. **(i)** : To sample \(\theta \mid \psi, \phi, \tau, y\), draw \(\theta_j\) independently from a \(\mbox{N}(\mu_j, \sigma_j^2)\) distribution with \[\begin{aligned} \sigma_j^2 = \{1 + 1/(\psi_j \phi_j^2 \tau^2)\}^{-1}, \quad \mu_j = \{1 + 1/(\psi_j \phi_j^2 \tau^2)\}^{-1} y. \end{aligned}\] **(ii)** : The conditional posterior of \(\psi \mid \phi, \tau, \theta\) can be sampled efficiently in a block by independently sampling \(\psi_j \mid \phi, \theta\) from an inverse-Gaussian distribution \(\mbox{iG}(\mu_j, \lambda)\) with \(\mu_j = \phi_j \tau/|\theta_j|, \lambda = 1\). **(iii)** : Sample the conditional posterior of \(\tau \mid \phi, \theta\) from a \(\mbox{giG}(\lambda-n, 1, 2 \sum_{j=1}^n | \theta_j |/\phi_j )\) distribution. **(iv)** : To sample \(\phi \mid \theta\), draw \(T_1, \ldots, T_n\) independently with \(T_j \sim \mbox{giG}(a-1, 1, 2 | \theta_j| )\) and set \(\phi_j = T_j/T\) with \(T = \sum_{j=1}^n T_j\). # Concentration properties of Dirchlet--Laplace priors In this section, we study a number of properties of the joint density of the Dirichlet--Laplace prior \(\mbox{DL}_a\) on \(\mathbb{R}^n\) and investigate the implied rate of posterior contraction [\[eq:post_conc\]](#eq:post_conc){reference-type="eqref" reference="eq:post_conc"} in the normal means setting [\[eq:norm_means\]](#eq:norm_means){reference-type="eqref" reference="eq:norm_means"}. Recall the hierarchical specification of \(\mbox{DL}_a\) from [\[eq:laplace_dir_tau\]](#eq:laplace_dir_tau){reference-type="eqref" reference="eq:laplace_dir_tau"}. Letting \(\psi_j = \phi_j \tau\) for \(j = 1, \ldots, n\), a standard result (see, for example, Lemma IV.3 of ) implies that \(\psi_j \sim \mbox{gamma}(a,1/2)\) independently for \(j = 1, \ldots, n\). Therefore, [\[eq:laplace_dir_tau\]](#eq:laplace_dir_tau){reference-type="eqref" reference="eq:laplace_dir_tau"} can be alternatively represented as[^3] \[\begin{aligned} \label{eq:mingyuan} \theta_j \mid \psi_j \sim \mbox{DE}(\psi_j), \quad \psi_j \sim \mbox{Ga}(a, 1/2). \end{aligned}\] The formulation [\[eq:mingyuan\]](#eq:mingyuan){reference-type="eqref" reference="eq:mingyuan"} is analytically convenient since the joint distribution factors as a product of marginals and the marginal density can be obtained analytically in Proposition [\[propBes\]](#propBes){reference-type="ref" reference="propBes"} below. The proof follows from standard properties of the modified Bessel function; a proof is sketched in the Appendix. Figure [\[fig:plot_marg\]](#fig:plot_marg){reference-type="ref" reference="fig:plot_marg"} plots the marginal density [\[eq:dens_marg\]](#eq:dens_marg){reference-type="eqref" reference="eq:dens_marg"} to compare with other common shrinkage priors. We shall continue to denote the joint density of \(\theta\) on \(\mathbb R^n\) by \(\Pi\), so that \(\Pi(\theta) = \prod_{j=1}^n \Pi(\theta_j)\). For a subset \(S \subset \{1, \ldots, n\}\), let \(\Pi_S\) denote the marginal distribution of \(\theta_S = \{\theta_j: j \in S\} \in \mathbb R^{|S|}\). For a Borel set \(A \subset \mathbb R^n\), let \(\bbP(A) = \int_{A} \Pi(\theta) d\theta\) denote the prior probability of \(A\), and \(\bbP(A \mid y^{(n)})\) the posterior probability given data \(y^{(n)} = (y_1, \ldots, y_n)\) and the model [\[eq:norm_means\]](#eq:norm_means){reference-type="eqref" reference="eq:norm_means"}. Finally, let \(E_{\theta_0}/P_{\theta_0}\) respectively indicate an expectation/probability w.r.t. the \(\mbox{N}_n(\theta_0, \mathrm{I}_n)\) density. We now establish that under mild restrictions on \(\norm{\theta_0}\), the posterior arising from the \(\mbox{DL}_a\) prior [\[eq:laplace_dir_tau\]](#eq:laplace_dir_tau){reference-type="eqref" reference="eq:laplace_dir_tau"} contracts at the minimax rate of convergence for an appropriate choice of the Dirichlet concentration parameter \(a\). A proof of Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"} can be found in Section [6](#sec:main_pf){reference-type="ref" reference="sec:main_pf"}. To best of our knowledge, Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"} is the first result obtaining posterior contraction rates for a continuous shrinkage prior in the normal means setting or the closely related high-dimensional regression problem. Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"} posits that when the parameter \(a\) in the Dirichlet--Laplace prior is chosen, depending on the sample size, to be \(n^{-(1+\beta)}\) for any \(\beta > 0\) small, the resulting posterior contracts at the minimax rate [\[eq:minmax\]](#eq:minmax){reference-type="eqref" reference="eq:minmax"}, provided \(\norm{\theta_0}_2^2 \leq q_n \log^4 n\). Using the Cauchy--Schwartz inequality, \(\norm{\theta_0}_1^2 \leq q_n \norm{\theta_0}_2^2\) for \(\theta_0 \in l_0[q_n; n]\) and the bound on \(\norm{\theta_0}_2\) implies that \(\norm{\theta_0}_1 \leq q_n (\log n)^2\). Hence, the condition in Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"} permits each non-zero signal to grow at a \((\log n)^2\) rate, which is a fairly mild assumption. Moreover, in a recent technical report, the authors showed that a large subclass of global-local priors [\[eq:lg\]](#eq:lg){reference-type="eqref" reference="eq:lg"} including the Bayesian lasso lead to a sub-optimal rate of posterior convergence; i.e., the expression in [\[eq:dl_cons\]](#eq:dl_cons){reference-type="eqref" reference="eq:dl_cons"} converges to \(0\) whenever \(\norm{\theta_0}_2^2/q_n \to \infty\). Therefore, Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"} indeed provides a substantial improvement over a large class of GL priors. The choice \(a_n = n^{-(1+\beta)}\) will be evident from the various auxiliary results in Section [3.1](#sec:aux){reference-type="ref" reference="sec:aux"}, specifically Lemma [\[lem:prior_m\]](#lem:prior_m){reference-type="ref" reference="lem:prior_m"} and Theorem [\[thm:compress\]](#thm:compress){reference-type="ref" reference="thm:compress"}. The conclusion of Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"} continues to hold when \(a_n=1/n\) under an additional mild assumption on the sparsity \(q_n\). In Table [1](#tab:1){reference-type="ref" reference="tab:1"} of Section [4](#sec:sims){reference-type="ref" reference="sec:sims"}, detailed empirical results are provided with \(a_n = 1/n\) as a default choice. The lower bound result alluded to in the previous paragraph precludes GL priors with polynomial tails, such as the horseshoe. We hope to address the polynomial tails case elsewhere, though based on strong empirical performance, we conjecture that the horseshoe leads to the optimal posterior contraction in a much broader domain compared to the Bayesian lasso and other common shrinkage priors. ## **Auxiliary results** {#sec:aux} In this section, we state a number of properties of the DL prior which provide a better understanding of the joint prior structure and also crucially help us in proving Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"}. We first provide useful bounds on the joint density of the DL prior in Lemma [\[lem:prior_b\]](#lem:prior_b){reference-type="ref" reference="lem:prior_b"} below; a proof can be found in the Appendix. It is evident from Figure [\[fig:plot_marg\]](#fig:plot_marg){reference-type="ref" reference="fig:plot_marg"} that the univariate marginal density \(\Pi\) has an infinite spike near zero. We quantify the probability assigned to a small \(\delta\)-neighborhood of the origin in Lemma [\[lem:prior_m\]](#lem:prior_m){reference-type="ref" reference="lem:prior_m"} below. A proof of Lemma [\[lem:prior_m\]](#lem:prior_m){reference-type="ref" reference="lem:prior_m"} can be found in the Appendix. In case of point mass mixture priors [\[eq:point_mass\]](#eq:point_mass){reference-type="eqref" reference="eq:point_mass"}, the induced prior on the model size \(|\mbox{supp}(\theta)|\) follows a \(\mbox{Binomial}(n, \pi)\) prior (given \(\pi\)), facilitating study of the multiplicity phenomenon. However, \(\bbP(\theta = 0) = 1\) for any continuous shrinkage prior, which compounds the difficulty in studying the degree of shrinkage for these classes of priors. Letting \(\mbox{supp}_{\delta}(\theta) = \{ j: |\theta_j| > \delta\}\) to be the entries in \(\theta\) larger than \(\delta\) in magnitude, we propose \(|\mbox{supp}_{\delta}(\theta)|\) as an approximate measure of model size for continuous shrinkage priors. We show in Theorem [\[thm:compress\]](#thm:compress){reference-type="ref" reference="thm:compress"} below that for an appropriate choice of \(\delta\), \(|\mbox{supp}_{\delta}(\theta)|\) doesn't exceed a constant multiple of the true sparsity level \(q_n\) with *posterior probability* tending to one, a property which we refer to as *posterior compressibility*. The choice of \(\delta_n\) in Theorem [\[thm:compress\]](#thm:compress){reference-type="ref" reference="thm:compress"} guarantees that the entries in \(\theta\) smaller than \(\delta_n\) in magnitude produce a negligible contribution to \(\norm{\theta}\). Observe that the prior distribution of \(|\mbox{supp}_{\delta_n}(\theta)|\) is \(\mbox{Binomial}(n, \zeta_n)\), where \(\zeta_n = \bbP(|\theta_1| > \delta_n)\). When \(a_n = n^{-(1 + \beta)}\), \(\zeta_n\) can be bounded above by \(\log n/n^{1+\beta}\) in view of Lemma [\[lem:prior_m\]](#lem:prior_m){reference-type="ref" reference="lem:prior_m"} and the fact that \(\Gamma(x) \geq 1/x\) for \(x\) small. Therefore, the prior expectation \(\bbE \abs{\mathrm{supp}_{\delta_n}(\theta)} \leq \log n/n^{\beta}\). This actually implies an exponential tail bound for \(\bbP(\abs{\mathrm{supp}_{\delta_n}(\theta)} > A q_n)\) by Chernoff's method, which is instrumental in deriving Theorem [\[thm:compress\]](#thm:compress){reference-type="ref" reference="thm:compress"}. A proof of Theorem [\[thm:compress\]](#thm:compress){reference-type="ref" reference="thm:compress"} along these lines can be found in the Appendix. The posterior compressibility property in Theorem [\[thm:compress\]](#thm:compress){reference-type="ref" reference="thm:compress"} ensures that the dimensionality of the posterior distribution of \(\theta\) (in an approximate sense) doesn't substantially overshoot the true dimensionality of \(\theta_0\), which together with the bounds on the joint prior density near zero and infinity in Lemma [\[lem:prior_b\]](#lem:prior_b){reference-type="ref" reference="lem:prior_b"} delivers the minimax rate in Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"}. # Simulation Study {#sec:sims} To illustrate the finite-sample performance of the proposed DL prior (with \(a= 1/n\)), we show the results from a replicated simulation study with various dimensionality \(n\) and sparsity level \(q_n\). In each setting, we have \(100\) replicates of a \(n\)-dimensional vector \(y\) sampled from a \(\mbox{N}_{n}(\theta_0, \mathrm{I}_n)\) distribution with \(\theta_0\) having \(q_n\) non-zero entries which are all set to be a constant \(A > 0\). We chose two values of \(n\), namely \(n = 100, 200\). For each \(n\), we let \(q_n = 5, 10, 20 \%\) of \(n\) and choose \(A = 7, 8\). This results in \(12\) simulation settings in total. The simulations were designed to mimic the setting in Section 3 where \(\theta_0\) is sparse with a few moderate-sized coefficients. # Prostate data application {#sec:real} We consider a popular dataset from a microarray experiment consisting of expression levels for \(6033\) genes for \(50\) normal control subjects and \(52\) patients diagnosed with prostate cancer. The data takes the form of a \(6033 \times 102\) matrix with the \((i, j)\)th entry corresponding to the expression level for gene \(i\) on patient \(j\); the first \(50\) columns correspond to the normal control subjects with the remaining \(52\) for the cancer patients. The goal of the study is to discover genes whose expression levels *differ* between the prostate cancer patients (treatment) and normal subjects (control). A two sample \(t\)-test with \(100\) degrees of freedom was implemented for each gene and the resulting t-statistic \(t_i\) was converted to a \(z\)-statistic \(z_i = \Phi^{-1} ( T_{100}(t_i))\). Under the null hypothesis \(H_{0i}\) of no difference in expression levels between the treatment and control group for the \(i\)th gene, the null distribution of \(z_i\) is \(\mbox{N}(0, 1)\). Figure [\[fig:hist\]](#fig:hist){reference-type="ref" reference="fig:hist"} shows a histogram of the \(z\)-values, comparing it to a \(\mbox{N}(0,1)\) density with a multiplier chosen to make the curve integrate to the same area as the histogram. The shape of the histogram suggests the presence of certain interesting genes. The classical Bonferroni correction for multiple testing flags only \(6\) genes as significant, while the two-group empirical Bayes method of found \(139\) significant genes, being much less conservative. The local Bayes false discovery rate (fdr) control method identified \(54\) genes as non-null. For detailed analysis of this dataset using existing methods, refer to. To apply our method, we set up a normal means model \(z_i = \theta_i + \epsilon_i, i=1, \ldots, 6,033\) and assign \(\theta\) a \(\mbox{DL}_a\) prior. Instead of fixing \(a\), we use a discrete uniform prior on \(a\) supported on the interval \([1/6,000, 1/2]\), with the support points of the form \(10(k + 1)/6,000, k = 0, 1, \ldots, K\). Such a fully Bayesian approach allows the data to dictate the choice of the tuning parameter \(a\) which is only specified up to a constant by the theory and also avoids potential numerical issues arising from fixing \(a = 1/n\) when \(n\) is large. Updating \(a\) is straightforward since the full conditional distribution of \(a\) is again a discrete distribution on the chosen support points. We implemented the Gibbs sampler in Section [2.4](#sec:postcomp){reference-type="ref" reference="sec:postcomp"} for 10,000 draws discarding a burn-in of 5,000. Mixing and convergence of the Gibbs sampler was satisfactory based on examination of trace plots, with the 5,000 retained samples having an effective sample size of 2369.2 averaged across the \(\theta_i\)'s. The computational time per iteration scaled approximately linearly with the dimension. The posterior mode of \(a\) was at \(1/20\). In this application, we expect there to be two clusters of \(| \theta_i |\)s, with one concentrated closely near zero corresponding to genes that are effectively not differentially expressed and another away from zero corresponding to interesting genes for further study. As a simple automated approach, we cluster \(| \theta_i |\)s at each MCMC iteration using *kmeans* with 2 clusters. For each iteration, the number of non-zero signals is then estimated by the smaller cluster size out of the two clusters. A final estimate (\(M\)) of the number of non-zero signals is obtained by taking the mode over all the MCMC iterations. The \(M\) largest (in absolute magnitude) entries of the posterior median are identified as the non-zero signals. Using the above selection scheme, our method declared 128 genes as non-null. Interestingly, out of the 128 genes, 100 are common with the ones selected by EBMed. Also all the 54 genes obtained using FDR control form a subset of the selected \(128\) genes. Horseshoe is overly conservative; it selected only \(1\) gene (index: 610) using the same clustering procedure; the selected gene was the one with the largest *effect size* (refer to Table 11.2 in ). # Proof of Theorem 3.1 {#sec:main_pf} We prove Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"} for \(a_n = 1/n\) in details and note the places where the proof differs in case of \(a_n = n^{-(1+\beta)}\). Recall \(\theta_{S}: = \{ \theta_j, j \in S\}\) and for \(\delta \geq 0\), \(\mathrm{supp}_{\delta}(\theta) := \{ j: \abs{\theta_j} > \delta \}\). Let \(E_{\theta_0}/P_{\theta_0}\) respectively indicate an expectation/probability w.r.t. the \(\mbox{N}_n(\theta_0, \mathrm{I}_n)\) density. For a sequence of positive real numbers \(r_n\) to be chosen later, let \(\delta_n = r_n/n\). Define \(\mathcal{D}_n = \int \prod_{i=1}^n f_{\theta_{0i}}(y_i)/f_{\theta_i}(y_i) \, d\Pi(\theta)\). Let \[\begin{aligned} \mathcal{A}_n = \{\mathcal{D}_n \geq e^{-4r_n^2} \bbP(\norm{\theta-\theta_0}_2 \leq 2r_n) \} \end{aligned}\] be a subset of \(\sigma(y^{(n)})\), the sigma-field generated by \(y^{(n)}\), as in Lemma 5.2 of such that \(P_{\theta_0}(\mathcal{A}_n^c) \leq e^{-r_n^2}\). Let \(\mathcal{S}_n\) be the collection of subsets \(S \subset \{1, 2, \ldots, n\}\) such that \(\abs{S} \leq Aq_n\). For each such \(S\) and a positive integer \(j\), let \(\{\theta^{S, j, i}: i = 1, \ldots, N_{S,j}\}\) be a \(2j r_n\) net of \(\Theta_{S, j, n}=\{\theta \in \mathbb{R}^n: \mathrm{supp}_{\delta_n}(\theta) = S, 2 jr_n \leq \norm{\theta-\theta_0}_2 \leq 2(j+1)r_n \}\) created as follows. Let \(\{ \phi^{S, j, i}: i = 1, \ldots, N_{S,j}\}\) be a \(j r_n\) net of the \(|S|\)-dimensional ball \(\{ \| \phi-\theta_{0S} \| \leq 2(j+1) r_n\}\); we can choose this net in a way that \(N_{S, j} \leq C^{|S|}\) for some constant \(C\) (see, for example, Lemma 5.2 of ). Letting \(\theta^{S, j, i}_S = \phi^{S, j, i}\) and \(\theta^{S, j, i}_k = 0\) for \(k \in S^c\), we show this collection indeed forms a \(2 j r_n\) net of \(\Theta_{j,S,n}\). To that end, fix \(\theta \in \Theta_{S, j, n}\). Clearly, \(\| \theta_S-\theta_{0S} \| \leq 2 (j+1) r_n\). Find \(1 \leq i \leq N_{S, j}\) such that \(\| \theta^{S, j, i}_S-\theta_S \| \leq j r_n\). Then, \[\begin{aligned} \norm{\theta^{S, j, i}-\theta}_2^2 = \norm{\theta^{S, j, i}_S-\theta_S }_2^2 + \norm{\theta_{S^c}}_2^2 \leq (jr_n)^2 + (n-q_n)r_n^2/n^2 \leq 4j^2r_n^2, \end{aligned}\] proving our claim. Therefore, the union of balls \(B_{S, j, i}\) of radius \(2 j r_n\) centered at \(\theta^{S, j, i}\) for \(1 \leq i \leq N_{S, j}\) cover \(\Theta_{S, j, n}\). Since \(E_{\theta_0} \bbP(\abs{\mathrm{supp}_{\delta_n}(\theta)} > A q_n \mid y^{(n)}) \to 0\) by Theorem [\[thm:compress\]](#thm:compress){reference-type="ref" reference="thm:compress"}, it is enough to work with \(E_{\theta_0} \bbP( \theta: \norm{\theta-\theta_0}_2 > 2 M r_n, \mathrm{supp}_{\delta_n}(\theta) \in \mathcal{S}_n \mid y^{(n)} )\). Using the standard testing argument for establishing posterior convergence rates (see, for example, the proof of Proposition 5.1 in ), we arrive at \[\begin{aligned} \label{eq:interior} E_{\theta_0} \bbP( \theta: \norm{\theta-\theta_0}_2 > 2 M r_n, \mathrm{supp}_{\delta_n}(\theta) \in \mathcal{S}_n \mid y^{(n)} ) \leq \sum_{S \in \mathcal{S}_1} \sum_{j \geq M} \sum_{i =1}^{N_{S,j}} 2 \sqrt{\beta_{j,S,i}} e^{-C j^2 r_n^2}, \end{aligned}\] where \[\begin{aligned} \beta_{S, j, i} = \frac{\bbP(B_{S, j, i})}{e^{-4 r_n^2} \bbP(\theta: \norm{\theta-\theta_0}_2 < 2 r_n )}. \end{aligned}\] Let \(r_n^2 = q_n \log n\). The proof of Theorem [\[thm:minimaxDL\]](#thm:minimaxDL){reference-type="ref" reference="thm:minimaxDL"} is completed by deriving an upper bound to \(\beta_{S, j, i}\) in the following Lemma [\[lem:betaSj\]](#lem:betaSj){reference-type="ref" reference="lem:betaSj"} akin to Lemma 5.4 in. Substituting the upper bound for \(\beta_{S,j,i}\) obtained in Lemma [\[lem:betaSj\]](#lem:betaSj){reference-type="ref" reference="lem:betaSj"}, and noting that \(\abs{S} \leq Aq_n\) and \(\abs{N_{S,j}} \leq e^{C\abs{S}}\), the expression in the left hand side of [\[eq:interior\]](#eq:interior){reference-type="eqref" reference="eq:interior"} can be bounded above by \[\begin{aligned} & 2\sum_{S \in \mathcal{S}_1} \sum_{j \geq M} \sum_{i =1}^{N_{S,j}} \exp\{Aq_n \log(2j)/2 + C/2 (A+1)q_n \log n + C'2 r_n^2/2\} e^{-C j^2 r_n^2} \\ & \leq 2\sum_{S \in \mathcal{S}_1} \sum_{j \geq M} \exp\{C q_n + Aq_n \log(2j)/2 + C/2 (A+1)q_n \log n + C' r_n^2/2\} e^{-C j^2 r_n^2}. \end{aligned}\] Since \(\abs{\mathcal{S}_1} \leq Aq_n{n \choose Aq_n} \leq Aq_ne^{Aq_n \log (ne/Aq_n)}\), it follows that \(E_{\theta_0} \bbP( \theta: \norm{\theta-\theta_0}_2 > 2 M r_n, \mathrm{supp}_{\delta_n}(\theta) \in \mathcal{S}_n \mid y^{(n)} ) \to 0\) for large \(M > 0\). When \(a_n = n^{-(1+ \beta)}\), the conclusion of Lemma [\[lem:betaSj\]](#lem:betaSj){reference-type="ref" reference="lem:betaSj"} remains unchanged and the proof of Theorem [\[thm:compress\]](#thm:compress){reference-type="ref" reference="thm:compress"} does not require \(q_n \gtrsim \log n\). The rest of the proof remains exactly the same. [^1]: Following standard practice in this literature, we use \(n\) to denote the dimensionality and it should not be confused with the sample size. [^2]: Given sequences \(a_n, b_n\), we denote \(a_n = O(b_n)\) or \(a_n \lesssim b_n\) if there exists a global constant \(C\) such that \(a_n \leq C b_n\) and \(a_n = o(b_n)\) if \(a_n/b_n \to 0\) as \(n \to \infty\). [^3]: This formulation only holds when \(\tau \sim \mbox{gamma}(na, 1/2)\) and is not true for the general \(\mbox{DL}_a(\tau)\) class with \(\tau \sim g\). [^4]: The EBMed procedure was implemented using the package. [^5]: Given a draw for \(s\), a subset \(S\) of size \(s\) is drawn uniformly. Set \(\theta_j = 0\) for all \(j \notin S\) and draw \(\theta_j, j \in S\) i.i.d. from standard Laplace. The beta-bernoulli priors in [\[eq:point_mass\]](#eq:point_mass){reference-type="eqref" reference="eq:point_mass"} induce a similar prior on the subset size.

{'timestamp': '2014-01-22T02:11:24', 'yymm': '1401', 'arxiv_id': '1401.5398', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5398'}

# Introduction Displacive solid-to-solid phase transformations (also termed as martensitic phase transformations) are accompanied by remarkable changes in the material behavior making these materials important for the design of intelligent structures. It has been highlighted that these changes are strongly dependent on the austenite-martensite (A-M) interface energy which in the end controls the material's applicability for sensors and damping elements in structures . Therefore, the role of the interface energy in the transformation kinetics and consequently in the material behavior has drawn an increasing interest in condensed matter physics. In the past, mainly continuum based models have been developed. Recently, a phase field approach has been applied to characterize the effect of geometric incompatibilities and crystal symmetries (twins and habit plane formation) on the A-M interface energy. A consistent framework to incorporate surface tension and energy in the Ginzburg-Landau theory for multivariant martensitic phase transformation has been developed by  for modeling coherent interfaces. Lately, an approach based on statistical mechanics of lattice systems linking the microscopic state and the macroscopic behavior has been applied to describe consequences of martensitic transformations on the mechanical behavior of materials . In this approach, the phase interface energy can easily be taken into consideration. The interface energy contains chemical and mechanical contributions, where the latter are due to elastic energy stored in the vicinity of the interface between the austenitic and martensitic phases. So the interface energy considered here does not present a microscopical (local) quantity, but should be considered at a mesoscopic scale as an integral measure of the potential interface between austenite and martensite phase. This interpretation implies higher values of the interface energy than is usually used in micromechanical models. A straight atomically perfect interface has a minimum energy. The less perfect the interface, the higher both the stored energy and the transformation barrier. In the current article we employ a lattice Monte-Carlo simulation for martensitic phase transformations in analogy to the techniques used in magnetic systems . But in contrast to the latter, we use a specific Hamiltonian proposed in . We discuss the influence of the interface energy on the width of the transformation zone and material's compliance depending both on stress and temperature paths. # Theory and statistical model In the model we assume that the center of each single crystal (microscale) occupies a site of an \(L\)-component one-dimensional lattice (macroscale). The microstate specifying variable is a vector \(\boldsymbol{\mu}\), each component \(\mu_i, i \in [1,L]\) can be either A or M. The space of microstates for the prescribed stress-temperature state is labeled by \(\wedge\). We use the constant pressure (isobaric) distribution, the pressure is interchanged by the Cauchy stress tensor \(\boldsymbol\sigma\). The partition function \(Z\) and the Hamiltonian \(\,\hat{\!\mathcal{H}}\) are defined as \[\begin{split} &Z = \sum_{\wedge} \exp\left[-\beta \,\hat{\!\mathcal{H}} (T,\mathbf{\boldsymbol{\mu}})\right],\\ &\,\hat{\!\mathcal{H}} (T,\mathbf{\boldsymbol{\mu}}) = E(T,\mathbf{\boldsymbol{\mu}})-\boldsymbol\sigma: \,\tilde{\!\boldsymbol\varepsilon}( \mathbf{\boldsymbol{\mu}} ), \end{split} \label{eq:1}\] where \(\beta\) is a Lagrange multiplier \(\beta=1/(kT)\), where \(k\) is the modified Boltzmann constant which determines the relative size of the cells, and \(T\) is the temperature; the total internal energy \[E(T, \mathbf{\boldsymbol{\mu}}) = \sum_{i}^{L} U_{\mu_i}(T) + 1/2\!\!\!\!\sum_{i=1,|i-j|=1}^{L} U_{\mu_i,\mu_j}^I. \label{eq:2}\] is a combination of the internal \(U_{\mu_i}(T)\) and the interface \(U_{\mu_i,\mu_j}^I\) energy densities between neighbor sites. In a first assumption, the interface energy is a temperature independent quantity. It is worth noting that the internal energy densities of the martensite and austenite have only one intersection point on the considered temperature interval \(T\in[T_0; T_1]\), moreover, for \(T=T_0\), \(U_{_{A}} > U_{_{M}}\) and for \(T=T_1\) \(U_{_{A}} < U_{_{M}}\), in our case \(T_0 \approx 2\) K, \(T_1 \to \infty\). In a single martensitic variant system the interface energy between sites is given by \[U_{\mu_i,\mu_j}^I = U_{_{AM}} \left( 1-\delta_{\mu_i}^{\mu_j}\right), \label{eq:2a}\] where \(U_{_{AM}}\) is the austenite-martensite interface energy per unit surface and \(\delta_{\mu_i}^{\mu_j}\) is the Kronecker delta. The term \(\boldsymbol\sigma: \tilde\boldsymbol\varepsilon(\mathbf{\boldsymbol{\mu}})\) in Eq. ([\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"}) takes into account the energy through an external mechanical stress \(\boldsymbol\sigma\) and the strain density tensor \(\;\tilde{\!\boldsymbol\varepsilon}( \mathbf{\boldsymbol{\mu}} )\) given by \[\tilde{\!\boldsymbol\varepsilon}( \mu_{_A} ) = \mathbf{S}_{_A}\!:\boldsymbol\sigma / \eta_{\mu_A},\quad \tilde{\!\boldsymbol\varepsilon}( \mu_{_M} ) = \left(\mathbf{S}_{_M}\!:\boldsymbol\sigma + \boldsymbol\varepsilon^{\mbox{\tiny tr}}\right) / \eta_{\mu_M} \label{eq:3}\] for austenite \(\mu_{_A}\) and martensite \(\mu_{_M}\) sites, respectively, where \(\boldsymbol\varepsilon^{\mbox{\tiny tr}}\) is the transformation strain tensor for the given variant of the martensite, \(\mathbf{S}\) is the fourth-order compliance tensor from the theory of linear elasticity and \(\eta_{\mu_i}\) is the molar density of a pure state. The decomposition of the total strain is made according to the framework of small strain theory. In a first assumption, the model omits thermal strains and stress fluctuations. Further information about the effective Hamiltonian formulation can be found in . Here, we employ a Monte-Carlo simulation using the Metropolis algorithm  for a one-dimensional A-M lattice system with fully periodic boundary conditions. The lattice site interfaces are simulated as soft fluctuating objects within the lattice. # Description of computation One of the properties of interest in martensitic phase transformation is the evolution of the martensite fraction \(\langle f_{_M}\rangle\). In a Monte-Carlo simulation, this quantity can be computed as follows: by reaching the equilibrium at time \(t_0\) for a given stress-temperature state, one obtains the martensite fraction as an average of the instantaneous martensite fractions \(f_{_M}(t^{\mbox{\tiny MC}},T,\boldsymbol\sigma)\) over interval \(\Delta t\) of the discrete and dimensionless Monte-Carlo time \(t^{\mbox{\tiny MC}}\): \[\langle f_{_M}(T,\boldsymbol\sigma)\rangle = \frac{1}{\Delta t} \sum\limits_{t^{\mbox{\tiny MC}}=t_0}^{t_0+\Delta t} f_{_M}(t^{\mbox{\tiny MC}},T,\boldsymbol\sigma)\] We found that the results converge with an increasing number of lattice sites \(L\), i.e. \[\forall\varepsilon > 0, \exists\,L: \forall l > L: \left\|\langle f_{_M}(T,\sigma)\rangle_{{l}}-\langle f_{_M}(T,\sigma)\rangle_{{L}}\right\| \le \varepsilon.\] For \(\varepsilon=1\%\) it is found that \(L\approx7\). The symbol \(\| \dots \|\) denotes here the maximum norm \(L^\infty\). The lattice system under consideration does not exhibit a critical slowing down, thus the simple Metropolis algorithm is applicable for a proper prediction of the phase transition, in contrast to magnetic systems which require cluster algorithms to avoid this slowing down phenomenon (e.g., Swendson-Wang algorithm ). The input data for our model include molar densities \(\eta_{_A},\eta_{_M}\), elastic properties \(\mathbf S_{_A},\mathbf S_{_M}\), evolution of the internal energy densities \(U_{A}(T), U_{M}(T)\) of the pure phases (which can be found from specific heat capacities \(c_{_A}(T),c_{_M}(T)\) ), interface energy densities between austenite and martensite phases \(U^I_{_{AM}}\) as well as between different martensitic variants (if they are present) and the crystallographic transformation strain \(\boldsymbol\varepsilon^{\mbox{\tiny tr}}\) of the martensitic variants. Herein, typical material data for a shape memory alloy are taken from : Young's moduli \(E_{_A} = 56\) GPa, \(E_{_M}=26\) GPa, Poisson's ratios \(\nu_{_A} = \nu_{_M} = 0.3\), molar densities \(\eta_{_A} = \eta_{_M} = 9.7\times 10^4\) mol/m\(^3\), transformation strain \(\varepsilon_{11}^{\mbox{\tiny tr}} = 2\%\), internal energy densities \(U_{A}(T) = 10 + 7.66 T + 0.033 T^2\) J/m\(^3\), \(U_{M}(T) = 20-0.25 T + 0.067 T^2\) J/m\(^3\) and the interface energy density \(U_{_{AM}}\) varies up to \(50\) J/m\(^2\), whereas the Lagrange multiplier \(\beta = 7.24/T\). For the lattice simulation we prescribe the number of sites \(L=10\) and the Monte-Carlo time period \(\Delta t\) over which the data are averaged when the simulation reaches a quasi-equilibrium state, in our case \(\Delta t^{\mbox{\tiny MC}}=10^8\) per each stress-temperature state. In all simulations the temperature step is \(\Delta T = 0.1\) K. The simulation is started from a pure austenite state at \(T=310\) K for a given external stress \(\boldsymbol\sigma\). According to the Metropolis algorithm a random site is flipped either from austenite to martensite or vice versa. If a random number \(r\) uniformly distributed in \([0;1]\) is smaller than the probability resulting from the change of the Hamiltonian \(r < \exp\left(-\beta\Delta\,\hat{\!\mathcal{H}} (T)\right)\), then the change of the state is accepted, otherwise not. The change of the Hamiltonian is given by \(\Delta\,\hat{\!\mathcal{H}} (T) = \,\hat{\!\mathcal{H}} (T,\mathbf{\boldsymbol{\mu}}^1)-\,\hat{\!\mathcal{H}} (T,\mathbf{\boldsymbol{\mu}}^0)\), where \(\mathbf{\boldsymbol{\mu}}^1\) and \(\mathbf{\boldsymbol{\mu}}^0\) are configurations after and before the toggling, respectively. The procedure is repeated \(10^8\) times, and the average values over time are computed. Then the temperature is decreased by \(0.1\) K and the procedure is repeated again. Moreover, the change of the average martensite fraction is computed with respect to the change of the temperature (see Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}). The increments are repeated up to \(T=190\) K. # Results ## Width of the transformation zone The simulation of thermally induced transformations shows a considerable influence of the interface energy on the transformation kinetics. This result holds for all stress states in the elastic regime. The higher the interface energy the faster the transformation proceeds, see Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}. This narrowing of the transformation zone implies an increased phase stability. However, the temperature of parity of phases \(f_{_A}(T^\ast) = f_{_M}(T^\ast)\) is a function of the internal energies and the external stress and does not depend on the interface energy. To characterize the width of the transformation zone we introduce a parameter \(w\;[1/K]\), defined as a mean squared derivative of the martensitic fraction with respect to the temperature \[w = \int\limits_0^\infty \left[\frac{\partial \langle f_{_M} \rangle}{\partial T}-\left\langle\frac{\partial \langle f_{_M} \rangle}{\partial T}\right\rangle\right]^2\,dT, \label{eq:4}\] where \(\left\langle \partial \langle f_{_M} \rangle /\partial T \right\rangle\) is the mean slope of the martensitic phase fraction. Fig. [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"} illustrates a smooth decrease of the width parameter \(w\) with the interface energy. Note that the width is higher for smaller stresses if the interface energy is below \(45-50\) J/m\(^2\), at this value the curves for different stresses collapse. ## Effective compliance To demonstrate how the interface energy affects the damping properties of materials obeying martensitic transformations, we remark that the deformation wave speed \(v\) (in a one-dimensional case only longitudinal waves are present) is inversely proportional to the square root of the product of the density and the compliance \[1/v \sim \sqrt{\rho C(U_{_{AM}})},\] where \(C = \partial \varepsilon/\partial \sigma\). For shape memory alloys the density is almost independent from the phase, whereas the compliance increases significantly within the transformation region as shown in Fig. [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}. For a given temperature the peak compliance increases with increasing interface energy. Thus, a shock wave, which is capable of initiating a stress-induced transformation, slows down within the transformation zone. The higher the interface energy, the more pronounced is this deceleration process. A further important factor for damping properties is the hysteretic behavior of the material, therefore it is the future step in our study to introduce a physically meaningful hysteresis mechanism. # Conclusions and perspectives Using one-dimensional Monte-Carlo lattice simulations based on a statistical micromechanical model, we demonstrated for different external stresses a smooth increase in the martensite phase fraction with decreasing temperature. The width of the martensite transformation is shown to depend on the austenite-martensite interface energy: the higher this energy, the sharper the transformation. Note that to demonstrate better this trend and to partly compensate the one-dimensionality of our system, we deliberately exaggerated the interface energy. By analyzing the transformation in stress-temperature space, we show the influence of the interface energy on the effective compliance of the material subjected to a martensitic transformation. In agreement with the stress-temperature curve generally observed experimentally for the onset of the martensite transformation , we obtain an almost linear relation between stress and temperature for the parity curve (inset in Fig. [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"}). Our results suggest a possible direction in the improvement of shape-memory alloys aimed at increasing the interface energy between the martensite and the austenite phases. Use of materials with a sharper martensite transformation (owing to a high interface energy) will result in more sensitive and energy-efficient smart systems: switches, actuators, etc. For an investigation of the entire properties and quantitative comparisons with experiments it is inevitable to simulate at least a two-dimensional lattice system including multiple martensitic variants and internal stresses. Interface energies between martensitic variants \[as in \] and the choice of the variant under complex external loads and temperatures are also under consideration. Although it is easy to extend the Monte-Carlo simulations for an arbitrary dimension of the lattice, the incorporation of internal stresses in the formulation is not trivial and the work in this direction is in progress. Moreover, we are working on a thermodynamically coherent introducing of the latent heat and the internal friction in the system in order to reproduce a transformation hysteresis.

{'timestamp': '2014-01-23T02:01:19', 'yymm': '1401', 'arxiv_id': '1401.5510', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5510'}

# Introduction In this paper we consider the stochastic differential equation \[\label{eq:main} X_{t} = x + W_{t} + \int_{0}^{t}b(s, X_s)\,ds.\] In the paper A.M. Davie proved the following theorem: The proof of Davie is quite self-contained, but rather technically complicated. In particular it does not rely on the uniqueness of strong solutions. It turns out that in some cases the pathwise uniqueness can be proved with a slightly simpler approach. The main idea is to use the Hölder regularity of the flow generated by the strong solution proved in and a modification of the van Kampen uniqueness theorem for ordinary differential equations with a Lipschitz flow and continuous coefficients (see ). This approach also enables us to extend Davie's result to some other classes of irregular drifts. # Auxiliary results The proof of Davie uses the following estimate: An interesting discussion of this inequality and some similar problems can be found in. The original proof of Davie is quite long and relies on some explicit computations for the Gaussian kernel. Since our approach to Davie's theorem in the case of a Borel measurable drift also uses this estimate, below we present a proof which seems to be less technical than the reasoning in. The next proposition will play the crucial role in the proof of the main results. The existence of a flow possessing properties 1--3 is proved in (see Theorem 1.2). Instead of property 4 the authors of prove (see their Lemma 5.11) a slightly weaker assertion that almost surely for any fixed \(s, t \in [0, 1], \ s \leqslant t\) the mapping \(\varphi_{s, t}\) is Hölder continuous. For the sake of completeness, we present below a sketch of the proof of Proposition [\[pr:flow\]](#pr:flow){reference-type="ref" reference="pr:flow"} with necessary references to, and the key details of the proof of property 4. .1in **Step 1.** (See, Theorem 3.3, Lemma 3.4, and Lemma 3.5.) Let \[L_{p}^{q}(T) = L^{q}\bigl([0, T], L^{p}(\mathbb{R}^{d})\bigr)\] \[\mathbb{H}_{\alpha, p}^{q}(T) = L^{q}\bigl([0, T], W^{\alpha, p}(\mathbb{R}^{d})\bigr), \ \mathbb{H}_{p}^{\beta, q}(T) = W^{\beta, q}\bigl([0, T], L^{p}(\mathbb{R}^{d})\bigr)\] \[H_{\alpha, p}^{q}(T) = \mathbb{H}_{\alpha, p}^{q}(T) \cap \mathbb{H}_{p}^{1, q}(T)\] Let \(U:[0, T]\times\mathbb{R}^{d}\rightarrow\mathbb{R}^{d}\) be a solution to the equation \[\label{eq:pde} \left\{ \begin{aligned} &\frac{\partial U}{\partial t} + \frac{1}{2}\Delta U + b \cdot \nabla U = \lambda U-b\\ &U(T, x) = 0 \end{aligned} \right.\] for sufficiently large positive \(\lambda\) such that \[\|U\|_{H_{2, p}^{q}(T)} = \|D_{t}U\|_{L_{p}^{q}} + \|U\|_{H_{2, p}^{q}(T)} \leqslant C(d, T, p, q, \lambda)\|b\|_{L_{p}^{q}(T)},\] \[\sup\limits_{t \in [0, T]}\|\nabla U\|_{C_{b}(\mathbb{R}^{d})} \leqslant \frac{1}{2}.\] Then the family of mappings \(\psi_{t}:\mathbb{R}^{d}\rightarrow\mathbb{R}^{d}\) defined by the formula \[\psi_{t}(x) = x + U(t, x)\] possesses the following properties: 1. for each \(t \in [0, T]\) the mapping \(\psi_{t},\ \psi_{t}^{-1}\) is a \(C^{1}\)-diffeomorphism of \(\mathbb{R}^{d}\), 2. uniformly in \(t \in [0, T]\) the mappings \(\psi_{t},\ \psi_{t}^{-1}\), have globally bounded Hölder-continuous derivatives with respect to the space variable, 3. the mapping \((t, x) \mapsto \psi_{t}(x)\) belongs locally to the class \(H_{2, p}^{q}(T)\). **Step 2.** (See, Proposition 4.3.) The next step is transforming the original equation [\[eq:main\]](#eq:main){reference-type="ref" reference="eq:main"} (considered as a stochastic equation with the identity diffusion matrix and a Borel measurable drift) into an equation with more regular coefficients by means of the family of the homeomorphisms constructed at the previous step. Let us apply Itô's formula to the process \(X_t\) and the function \(U\) (see, p. 4): \[\begin{gathered} dU(t, X_t) = \frac{\partial U}{\partial t}(t, X_t)\,dt + \nabla U(t, X_t)\bigl(b(t, X_t)\,dt + dW_t\bigr) + \frac{1}{2}\Delta U(t, X_t)\,dt=\\ =\lambda U(t, X_t)-b(t, X_t)dt + \nabla U(t, X_t)\,dW_t\\ \end{gathered}\] Then the process \[Y_t := \psi_t(t, X_t) = X_t + U(t, X_t)\] has the stochastic differential \[\begin{gathered} dY_t = \lambda U(t, \psi_{t}^{-1}(Y_t))\,dt + \bigl[I + \nabla U(t, \psi_{t}^{-1}(Y_t))\bigr]\,dW_t= \widetilde{b}(t, Y_t)\,dt + \widetilde{\sigma}(t, Y_t)\,dW_t,\\ \widetilde{b}(t, y) = \lambda U(t, \psi_{t}^{-1}(y)), \ \widetilde{\sigma}(t, y) = I + \nabla U(t, \psi_{t}^{-1}(Y_t))\\ \end{gathered}\] **Step 3.** (See, Proposition 5.2,, p. 13--14.) Taking into account the aforementioned properties of the mappings \(\psi_t\) it is not difficult to see that it suffices to prove the existence of a uniformly Hölder-continuous flow for the transformed equation. Below we prove only the uniform Hölder-continuity of the desired flow since all other details (e.g., the proof of its existence) can be found in. We have \[\label{eq:transformed_main} dY_t = \widetilde{b}(t, Y_t)\,dt + \widetilde{\sigma}(t, Y_t)\,dW_t.\] Let us show that for each \(a \geqslant 2\) there exists a constant \(C(a, T)\) such that for any \(x, y \in \mathbb{R}^{d}\) the following estimate holds: \[\mathbb{E}\sup\limits_{t \in [0, T]}|Y_{t}^{x}-Y_{t}^{y}|^{a} \leqslant C(a, T)\bigl(|x-y|^{a} + |x-y|^{a-1}\bigr),\] In this case the existence of a uniformly Hölder-continuous flow will follow from the well-known Kolmogorov continuity theorem. Following, , let us define an auxiliary process \[A_t := \int_{0}^{t}\frac{\|\widetilde{\sigma}(s, Y_{s}^{y})-\widetilde{\sigma}(s, Y_{s}^{x})\|^{2}}{|Y_{s}^{y}-Y_{s}^{x}|^2}I_{\{Y_{s}^{y} \neq Y_{s}^{x}\}}\,ds\] Then (see, Lemma 4.5) for each \(k \in \mathbb{R}\) we have \[\label{eq:exp_estimate} \mathbb{E}\bigl[e^{k A_T}\bigr] < \infty\] (in the proof of this inequality the Sobolev regularity of \(\widetilde{\sigma}\) plays the crucial role). Let \[Z_t := Y_{t}^{y}-Y_{t}^{x}.\] Applying Itô's formula to the process \(Z_t\) and the function \(f: x \mapsto |x|^{a},\) where \(a \geqslant 2,\) we obtain \[\begin{gathered} \frac{1}{a}d|Z_{t}|^{a} = \Bigl<\bigl(\widetilde{b}(t, Y_{t}^{y})-\widetilde{b}(t, Y_{t}^{x})\bigr)\,dt, Z_{t}^{a-1}\Bigr> + \\ + \Bigl<\bigl(\widetilde{\sigma}(t, Y_{t}^{y})-\widetilde{\sigma}(t, Y_{t}^{x})\bigr)dW_t, Z_{t}^{a-1}\Bigr> + \\ + \frac{1}{2}Tr\Bigl([\sigma(t, Y_{t}^{y})-\sigma(t, Y_{t}^{x})][\sigma(t, Y_{t}^{y})-\sigma(t, Y_{t}^{x})]^{t} D^{2}f(Z_t)\Bigr)\,dt,\\ \bigl[D^{2}f(Z_t)\bigr]_{i, j} = \delta_{i,j}|Z_t|^{a-2} + (a-2)Z_{t}^{i}Z_{t}^{j}|Z_t|^{a-4}.\\ \end{gathered}\] Using the Lipschitz continuity of \(\widetilde{b}\) and the definition of the process \(A_t\) we obtain the inequality \[d|Z_{t}|^{a} \leqslant C|Z_t|^{a}\,dt + C|Z_t|^{a}\,dA_{t} + \Bigl<\bigl(\widetilde{\sigma}(t, Y_{t}^{y})-\widetilde{\sigma}(t, Y_{t}^{x})\bigr)dW_t, Z_{t}^{a-1}\Bigr>\] Let \[M_t := \int_{0}^{t}\Bigl<\bigl(\widetilde{\sigma}(t, Y_{t}^{y})-\widetilde{\sigma}(t, Y_{t}^{x})\bigr)dW_t, Z_{t}^{a-1}\Bigr>.\] Since the coefficient \(\widetilde{\sigma}\) is bounded and all moments of the random variable \(|Z_t|\) are finite (see, Proposition 2.7), the process \(M_t\) is a square-integrable continuous martingale. Then we have \[\begin{gathered} de^{-CA_t}|Z_t|^{a} =-Ce^{-CA_{t}}|Z_{t}|^{a}dA_t + e^{-CA_t}d|Z_t|^{a}\leqslant\\ \leqslant-Ce^{-CA_{t}}|Z_{t}|^{a}dA_t + e^{-CA_t}C|Z_t|^{a}\,dt + e^{-CA_t}C|Z_t|^{a}\,dA_{t} + e^{-CA_t}\,dM_t =\\ =Ce^{-CA_t}|Z_t|^{a}\,dt + e^{-CA_t}\,dM_t\\ \end{gathered}\] Consequently, the following estimate holds: \[\mathbb{E}e^{-CA_t}|Z_t|^{a} \leqslant |x-y|^{a} + C\int_{0}^{t}\mathbb{E}e^{-CA_t}|Z_t|^{a}\,dt.\] Applying Gronwall's inequality we obtain the estimate \[\mathbb{E}e^{-CA_t}|Z_t|^{a} \leqslant |x-y|^{a}e^{CT}.\] Taking into account Hölder's inequality and the estimate [\[eq:exp_estimate\]](#eq:exp_estimate){reference-type="ref" reference="eq:exp_estimate"} we have \[\mathbb{E}|Z_t|^{a} = \mathbb{E}e^{CA_t}e^{-CA_t}|Z_t|^a \leqslant \Bigl[\mathbb{E}e^{2CA_t}\Bigr]^{\frac{1}{2}}\Bigl[\mathbb{E}e^{-2CA_t}|Z_t|^{2a}\Bigr]^{\frac{1}{2}} \leqslant C(a, T)|x-y|^{a}\] The next chain of inequalities easily follows from Doob's martingale inequality and the boundedness of \(\widetilde{\sigma}\): \[\begin{gathered} \mathbb{E}\sup\limits_{t \in [0, T]}e^{-2CA_t}|Z_t|^{2a} \leqslant \\ \leqslant 4|Z_0|^{2a} + 4\mathbb{E}\sup\limits_{t \in [0, T]}\Bigl|\int_{0}^{t}Ce^{-CA_s}|Z_s|^{a}\,ds\Bigr|^{2} + 4\mathbb{E}\sup_{t \in [0, T]}\Bigl|\int_{0}^{t}e^{-CA_s}\,dM_s\Bigr|^2 \leqslant\\ \leqslant 4|x-y|^{2a} + 4C^{2}T\mathbb{E}\int_{0}^{T}e^{-2CA_t}|Z_t|^{2a}\,dt + \\ + 16\mathbb{E}\int_{0}^{T}e^{-2CA_t}\|\widetilde{\sigma}(t, Y_{t}^{y})-\widetilde{\sigma}(t, Y_{t}^{x})\|^{2}|Z_t|^{2a-2}\,dt \leqslant\\ \leqslant K|x-y|^{2a} + K\mathbb{E}\int_{0}^{t}e^{-2CA_s}|Z_s|^{2a}\,ds + K\mathbb{E}\int_{0}^{t}e^{-2CA_s}|Z_s|^{2a-2}\,ds \leqslant \\ \leqslant K(a, T)\bigl(|x-y|^{2a} + |x-y|^{2a-2}\bigr).\\ \end{gathered}\] Therefore, \[\begin{gathered} \mathbb{E}\sup\limits_{t \in [0, T]}|Z_t|^{a} \leqslant \mathbb{E}e^{CA_T}\sup\limits_{t \in [0, T]}e^{-CA_t}|Z_t|^{a}\leqslant\\ \leqslant \Bigl[\mathbb{E}e^{2CA_T}\Bigr]^{\frac{1}{2}} \Bigl[\mathbb{E}\sup\limits_{t \in [0, T]}e^{-2CA_t}|Z_t|^{2a}\Bigr]^{\frac{1}{2}} \leqslant K'(a, T)\bigl(|x-y|^{a} + |x-y|^{a-1}\bigr)\\ \end{gathered}\] It is now easy to complete the proof. # Main results To illustrate the main idea let us prove Davie's theorem for some (possibly unbounded) drift coefficients \(b\) possessing Hölder's continuity with respect to the space variable. It is worth noting that the reasoning from can not be directly applied in this case, since they essentially use the global boundedness of the drift. Now we show how to prove the original result of Davie (his Theorem [\[th:borel_case\]](#th:borel_case){reference-type="ref" reference="th:borel_case"}) in the case where \(b\) is just Borel measurable. Similarly to the proof of Theorem [\[th:holder_case\]](#th:holder_case){reference-type="ref" reference="th:holder_case"}, it is readily seen that without loss of generality we can assume that \(b(t, x) = b(t, x)I_{\{|x| < N\}}\) and \(\|b\|_{\infty} \leqslant 1\). In this case for each \(\alpha \in (0, 1)\) the equation [\[eq:main\]](#eq:main){reference-type="ref" reference="eq:main"} \(P\)-almost surely possesses a Hölder-continuous flow of solutions that will be denoted by \(X(s, t, x, W)\). The main aim of the reasoning below is to find a substitute for the Hölder condition on the coefficient \(b\) that would allow us to repeat the proof of Theorem [\[th:holder_case\]](#th:holder_case){reference-type="ref" reference="th:holder_case"} with minor changes. Below we will need the following set of functions: \[\begin{gathered} Lip_{N}\bigl([r, u], \mathbb{R}^{d}\bigr) := \\ := \bigl\{h \in C\bigl([r, u], \mathbb{R}^{d}\bigr) \ \mid \ |h(t)-h(s)| \leqslant |t-s| \ s,t \in [r, u], \ \max\limits_{s \in [r, u]}|h(s)| \leqslant N \bigr\} \end{gathered}\] with the uniform metric \(\varrho(h_1, h_2) = \|h_1-h_2\|_{\infty}\). Now let us temporarily fix \(N > 0\) and \(r, u \in [0, T]\) such that \(l = u-r \leqslant \frac{1}{2}\). Let \[\varphi(h, W) := \int_{r}^{u}b(s, W_s + h(s))\, ds.\] We can now proceed to the proof of Theorem [\[th:borel_case\]](#th:borel_case){reference-type="ref" reference="th:borel_case"}. ***Acknowledgment*** I would like to thank V.I. Bogachev for fruitful discussions and comments. A part of this work was done during a visit to the Mathematical Institute of Burgundy. I would like to thank Shizan Fang for his hospitality and useful remarks. This work has been supported by the RFBR project 12-01-33009.

{'timestamp': '2014-01-22T02:12:28', 'yymm': '1401', 'arxiv_id': '1401.5455', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5455'}

# Introduction Closed-loop flow control is of major academic and industrial interest. At the interface of control theory and fluid mechanics it is pertinent to many engineering domains, such as aeronautics and combustion. It can be used to reduce aerodynamic drag of an automobile or an airplane, increase combustion efficiency, or enhance mixing. Control of amplifier flows like boundary layers, mixing layers, jets or separated flows is particularly relevant and challenging. Indeed, amplifier flows are globally stable, however convective instabilities will amplify disturbances while being advected downstream (). Incoming perturbations are likely to be amplified to the point where they disrupt the entire flow. Nullifying these disturbances before they can be amplified by the flow is a great challenge for the flow control community (). Typically when considering a laminar amplifier flow, the control objective can be to inhibit the transition to turbulence. Examples of such flows abound, a much-studied amplifier flow being that over a backward-facing step (BFS) () which presents an unsteady region of convective instability. Another example is that of the flow over a cavity, used for studying the control of global instabilities ().\ Control of amplifier flows has been the subject of much research (). A control law can be computed using one of two ways. One possibility is to compute the model using beforehand knowledge of the physics of the flow (). When derived directly from the Navier-Stokes equations these models are of very high order and require reduction before they can be used in a realistic setting. Model reduction is still a rich and very active research field, see. In some cases a physical analysis of the flow can yield simple models leading to efficient control laws as shown in. The second option is system identification as suggested by. In this case, the flow is probed until a model can be derived from its responses. This approach is data based: it seeks to build an input-output model for the flow from empirical observations. Such an approach has been applied with success to the control of the recirculation bubble behind a BFS, see.\ \ The BFS is considered as a benchmark geometry for the study of amplifier flows: separation is imposed by a sharp edge creating a strong shear layer susceptible to Kelvin-Helmholtz instability. Upstream perturbations are amplified in the shear layer leading to significant downstream disturbances. This flow has been extensively studied both numerically and experimentally ().\ \ The principle of feed-forward control is to act on the flow upon detection of an event as opposed to the more common feed-back control where one reacts to an event. Feed-forward algorithms have been successfully used in flow control in numerical simulations (). Recently have shown the effectiveness of a feed-forward algorithm computed using an Auto-Regressive Moving-Average Exogenous model (ARMAX) to capture the relevant dynamics of the flow. The resulting control law leads to reduced energy levels and fluctuations. The aim of this work is to determine the feasibility and robustness of this approach in an experimental setting. # Experimental Setup ## Water tunnel Experiments were carried out in a hydrodynamic channel in which the flow is driven by gravity. The flow is stabilized by divergent and convergent sections separated by honeycombs. The quality of the main stream can be quantified in terms of flow uniformity and turbulence intensity. The standard deviation \(\sigma\) is computed for the highest free stream velocity featured in our experimental set-up. We obtain \(\sigma = 0.059\) cm.s\(^{-1}\) which corresponds to turbulence levels of \(\frac{\sigma}{U_{\infty}}=0.0023\). For the present experiment the flow velocity is \(U_{\infty} =\) 2.1 cm.s\(^{-1}\) giving a Reynolds number based on step height \(Re_h=\frac{U_{\infty}h}{\nu} = 430\). Following the assumptions of Reynolds number was chosen to ensure a sub-critical linear 2D flow. ## Backward-Facing Step geometry and upstream perturbation The BFS geometry and the main geometric parameters are shown in figure [\[fig:dimensions\]](#fig:dimensions){reference-type="ref" reference="fig:dimensions"}. BFS height is \(h=1.5\) cm. Channel height is \(H=7\) cm for a channel width \(w=15\) cm. The vertical expansion ratio is \(A_y = \frac{H}{h+H} = 0.82\) and the span-wise aspect ratio is \(A_z=\frac{w}{h+H}=1.76\). The injection slot is located \(d / h =2\) upstream of the step edge.\ The principle of the method described in is to devise an input-output model for the flow based on experimental data. This model is used to compute actuation aimed at negating incoming upstream noise, thereby preventing its amplification. Because our sensor is 2D in the symmetry plane and our actuator can only deliver span-wise homogeneous actuation, a 2D upstream perturbation is required for effective control. As shown in figure [\[fig:dimensions\]](#fig:dimensions){reference-type="ref" reference="fig:dimensions"} a 2D obstacle with a rounded leading edge of height \(O_h=0.8\) cm has been placed at \(d_h=15\) cm upstream from jet injection (12 h from the step edge). Because of the low Reynolds number the flow stays 2D. ## Sensor: 2D real-time velocity fields computations The sensors used as inputs for the closed-loop experiments are visual sensors, i.e. regions of the 2D PIV (Particle Image Velocimetry) velocity fields measured jn the symmetry plane as shown in figure [\[fig:window\]](#fig:window){reference-type="ref" reference="fig:window"}. The flow is seeded with 20 \(\mu\)m neutrally buoyant polyamid seeding particles. They are illuminated by a laser sheet created by a 2W continuous laser beam operating at \(\lambda\) = 532 nm. Images of the vertical symmetry plane are recorded using a Basler acA 2000-340km 8bit CMOS camera. Velocity field computations are run on a Gforce GTX 580 graphics card. The algorithm used to compute the velocity fields is based on a Lukas-Kanade optical flow algorithm called FOLKI developed by. Its offline and online accuracy has been demonstrated and detailed by. Furthermore this acquisition method was successfully used in. The size of the velocity fields is \(17.2\times4.6\) cm\(^2\). They are computed every \(\delta_t=20ms\), for a sampling frequency \(F_s=25Hz\). ## Uncontrolled flow The swirling strength criterium \(\lambda_{Ci}\) is an effective way of detecting vortices in 2D velocity fields introduced and improved by. For 2D data the swirling criterium is defined as \(\lambda_{Ci}=\frac{1}{2}\sqrt{4 \det(\nabla \bf{u})-\tr(\nabla \bf{u})^2}\) (when this quantity is real).\ Figures [\[fig:rolls_ups\]](#fig:rolls_ups){reference-type="ref" reference="fig:rolls_ups"} a) and b) show the mean velocity amplitude fields for the uncontrolled flow with and without obstacle. Figures [\[fig:rolls_ups\]](#fig:rolls_ups){reference-type="ref" reference="fig:rolls_ups"} c) and d) show \(\lambda_{Ci}\) snapshots of the uncontrolled flow with and without the obstacle, highlighting the perturbations caused by the upstream obstacle. Figure [\[fig:rolls_ups\]](#fig:rolls_ups){reference-type="ref" reference="fig:rolls_ups"} d) shows the steady stream of vortices created by the obstacle interacting with the recirculation. Quantitatively, \(\lambda_{Ci}\) is an order of magnitude higher than for the flow without obstacle.\ The boundary layer thickness at the step edge for the flow with and without obstacles are \(\delta =\) 1.34h and \(\delta =\) 1.73h respectively. The turbulent kinetic energy (TKE) is defined as \(\epsilon(x,y,t)=\frac{1}{2}(u'(x,y,t)^2+v'(x,y,t)^2)\), where u', v' are longitudinal and vertical velocity fluctuations. The figure [\[fig:mean_TKE_un\]](#fig:mean_TKE_un){reference-type="ref" reference="fig:mean_TKE_un"} shows the time-averaged TKE field \(<\epsilon(x,y)>_t\) downstream of the step for the case with the upstream obstacle. The field exhibits two regions of high TKE. The lower region corresponds to the recirculation bubble. The upper region corresponds to residual perturbations induced by vortices shed by the upstream obstacle. ## Actuation actuation is a gaussian flow sink/source placed above the step, which is not experimentally feasible. In our case, actuation is provided by a flush slot jet, 0.1 cm long and 9 cm wide. This actuation has been chose to obtain a perturbation as homogeneous along the span-wise direction as possible. The jet angle to the wall is 45\(^o\). The slot is located 3 cm (2h) upstream the step edge (figure [\[fig:dimensions\]](#fig:dimensions){reference-type="ref" reference="fig:dimensions"}). Jet flow is induced using water from a pressurized tank. It enters a plenum and goes through a volume of glass beads designed to homogenize the incoming flow. Jet amplitude is controlled by changing tank pressure. Because channel pressure is higher than atmospheric pressure this allows us to provide both blowing and suction. The convection time from jet injection to measurement area is 2 s (\(<0.5Hz\)). The maximum actuation frequency \(f_a\) is about 1Hz which is sufficient for these experiments.\ The control law output velocities, these are converted into pressure commands using the transfer function described in figure [\[fig:pressure_to_amplitude\]](#fig:pressure_to_amplitude){reference-type="ref" reference="fig:pressure_to_amplitude"}. # ARMAX model ## Introduction An ARMAX model is used because it can be derived from experimental data,. Furthermore it has been shown by that it is particularly well adapted at modeling the BFS flow when in the linear regime. Two exogenous inputs \(s(t),u(t)\) and one output \(m(t)\) are used. The first exogenous input \(s(t)\) measures fluctuations of spatially averaged \(\lambda_{Ci}\) (small grey area on figure [\[fig:window\]](#fig:window){reference-type="ref" reference="fig:window"}). Such a sensor is well suited to the detection of upstream vortices created by the obstacle. The second exogenous input is jet flow velocity \(u(t)\).\ Output \(m(t)\) is a measure of TKE fluctuations in the recirculation region. The control objective is to negate the incoming perturbations created by the obstacle in order to reduce overall downstream TKE fluctuations. TKE is averaged over the whole downstream velocity field (large grey area on figure [\[fig:window\]](#fig:window){reference-type="ref" reference="fig:window"}): \[\begin{gathered} \label{eq:m} m(t) = \frac{\int \epsilon(x,y,t) dx dy}{\int dx dy} \end{gathered}\] Following the equation for the model is defined in eq [\[eq:ARMAX\]](#eq:ARMAX){reference-type="ref" reference="eq:ARMAX"}: \[\begin{gathered} \label{eq:ARMAX} \underbrace{m(t)+\sum_{k=1}^{n_a} a_k m(t-k)}_{auto-regressive}=\underbrace{\sum_{k=n_{du}}^{n_{du}+n_{bu}} b_k^u u(t-k)}_{exogenous\:1}+\underbrace{\sum_{k=n_{ds}}^{n_{ds}+n_{bs}} b_k^s s(t-k)}_{exogenous\:2}+E(t)\\ E(t)=\underbrace{\sum_{k=n_{1}}^{n_{c}} c_k e(t-k)}_{moving\:average}+e(t)\nonumber \end{gathered}\] To achieve feed-forward control, the effects of upstream sensing \(s(t)\) and actuation \(u(t)\) on the output \(m(t)\) must be quantified. For a pure feed-forward control, upstream estimation should be independent of actuation, see. During control, \(u(t)\) is a function of \(s(t)\). For our experimental setup we found that interference between actuation and the upstream sensor causes the control algorithm to saturate actuation. To avoid this effect, an inclined jet has been used instead of a wall normal jet. Moreover since \(s(t)\) only measures the presence of vortices it is weakly affected by downstream actuation compared to vertical velocity for example. Special care must be given to lower actuation amplitude as much as possible so that it does not affect the upstream sensor. Figure [\[fig:corr_u\_s\]](#fig:corr_u_s){reference-type="ref" reference="fig:corr_u_s"} shows the cross correlation function between \(s(t)\) and \(u(t)\) for two cases: the calibration case and a case where there is interference (jet amplitude is too high) between the upstream sensor and the actuator. Interference results in high correlation between \(s\) and \(u\) whereas in our calibration case correlation is negligible. Coefficients \((a_k, b_k^u, b_k^s)\) are computed to minimize error \(e(t)\) at all times. To calibrate the model the user must provide time series for both inputs and outputs, the longer the better. Values for \(n_a,n_{du}, n_{bu}, n_{ds}, n_{bs}\) are tied to the physics of the flow and are determined by the user. These coefficients are linked to time delays in the flow system. The flow time history required for the model to work properly is given by \(n_a.\delta_t\) (auto-regressive part). \(n_{du}.\delta_t\) and \(n_{ds}.\delta_t\) are the times required for the respective inputs to affect the output; they are linked to flow convective velocity. \(n_{bu}.\delta_t\) and \(n_{bs}.\delta_t\) represent input time scales. They correspond to the time during which upstream effects impact the output signal. Finally \(n_c\) is used to model noise and ensures robustness (). This value is chosen iteratively, once all other coefficients have been fixed, to get the best possible fit between experimental data and model output. ## Model Computation Figure [\[fig:calibration_series\]](#fig:calibration_series){reference-type="ref" reference="fig:calibration_series"} shows a small segment of the calibration time series. The forcing law \(u(t)\) used in these series is one of pseudo random pulses. Pulses are made to occur at random intervals, long enough for the effects of the previous pulse to have subsided before the next pulse. During these intervals the only input to the system is \(s(t)\). This allows the effects of actuation and upstream perturbations to be computed using a single time series. Impulse amplitude for actuation \(u(t)\) should be chosen such that it is high enough to affect the output \(m(t)\) but low enough to avoid perturbations of the upstream output \(s(t)\). Calibration data were acquired over 25 minutes. Figure [\[fig:auto_m\]](#fig:auto_m){reference-type="ref" reference="fig:auto_m"} shows the auto correlation function for \(m(t)\). A quasi-oscillatory behavior can be observed. It can be used to choose \(n_a\) which is such that \(n_a.\delta_t\) equals half the oscillatory period, as recommended by. Figure [\[fig:exp_impulse_response\]](#fig:exp_impulse_response){reference-type="ref" reference="fig:exp_impulse_response"} shows the response to an impulse, which can be used to evaluate the coefficients \(n_{du},n_{bu},n_{ds},n_{bs}\). The time delay \(t_{du}\)=2.5s between the beginning of the actuation and the response gives \(t_{du}=n_{du}.\delta_t\). The upstream sensor is located 3.5 cm upstream the jet injection. Assuming perturbations travel at channel velocity, this implies a time delay of \(t_{ds} \approx\) 1.7 s for an upstream disturbance to affect the output, thus \(n_{ds}.\delta_t= t_{du} + t_{ds}\).\ Let \(t_{bu}\) be the time during which an impulse in \(u\) affects the output, as shown in figure [\[fig:exp_impulse_response\]](#fig:exp_impulse_response){reference-type="ref" reference="fig:exp_impulse_response"}, then \(n_{bu}.\delta_t=t_{bu}\). Because the response to an impulse in \(s\) is more difficult to distinguish we assume \(n_{bs}=n_{bu}\). Finally \(n_c\) is chosen after the other coefficients have been fixed in order to get the best possible agreement between model and real outputs. Table [1](#tab:coeff){reference-type="ref" reference="tab:coeff"} summarizes the final coefficients used in the computation of the ARMAX model using the Matlab *armax* function (), it also shows the corresponding time delays and averages in seconds. Figure [\[fig:armax_output_validation\]](#fig:armax_output_validation){reference-type="ref" reference="fig:armax_output_validation"}a compares ARMAX output to the source signal for the calibration series. Agreement is good at 96 %. Figure [\[fig:armax_output_validation\]](#fig:armax_output_validation){reference-type="ref" reference="fig:armax_output_validation"}b compares ARMAX output to the source signal for the validation series; agreement is slightly lower at 94 %. ## Linearity A major underlying assumption of this approach is the linearity of the system. In our setup this was checked by imposing periodic pulsed forcing, with varying amplitudes. Figure [\[fig:linearity\]](#fig:linearity){reference-type="ref" reference="fig:linearity"} shows the phase averaged, spatially averaged TKE evolution in response to an actuation impulse. Impulse amplitude ratio is also given for comparison. A change in impulse amplitude leads to a proportional change in response amplitudes, confirming the linear behavior of the flow. Linearity was also checked when varying the size of the window where TKE is computed. Averaging over smaller windows, closer to the step, where non-linearities are weaker, did not improve the system linearity. # Results ## Control law Figures [\[fig:impulse\]](#fig:impulse){reference-type="ref" reference="fig:impulse"}a and [\[fig:impulse\]](#fig:impulse){reference-type="ref" reference="fig:impulse"}b show the impulse response for both exogenous inputs. These figures show impulse responses are qualitatively similar, however they differ in amplitude.\ Impulses responses can help determine if a model is \"controllable\" and whether or not the objective (negating TKE fluctuations) is *a priori* attainable, which makes them an invaluable diagnostics tool. To achieve fluctuation suppression the control law suggested by was computed.\ Only perturbations detected in \(s(t)\) can potentially be canceled out. Other sources of disturbance are not modeled and are ignored by the control law. Equation [\[eq:transfer\]](#eq:transfer){reference-type="ref" reference="eq:transfer"} illustrates how the output signal can be written as a combination of the input signals. \[m(t)=\sum_{k=0}^{\infty} h_k^s s(t-k)+ h_k^u u(t-k) \label{eq:transfer}\] The coefficients \(h_k^s,h_k^u\) are obtained by computing the impulse response of the ARMAX model as described in equation 4.2 and [\[eq:h\]](#eq:h){reference-type="ref" reference="eq:h"} for an impulse response \(s(t=0)=1,u(t=0)=1\). \[\begin{gathered} \forall k \: m_{impulse \: s}(t=k)=h_k^s\: s(0)\\ \forall k \: m_{impulse \: u}(t=k)=h_k^u\:u(0) \label{eq:h} \end{gathered}\] These coefficients can be used to express \(m(t)\) as a function of \(s(t), u(t)\) as shown in equation [\[eq:discrete\]](#eq:discrete){reference-type="ref" reference="eq:discrete"}. Previously \(s(t)\) was used to compute the model, here it is used as an input which allows us to compute \(u(t)\). This is done over 2000 time steps (T = 80 s). \[M_T=H_u U^f+G_u U^P+G_s S^P \label{eq:discrete}\] with \[\begin{gathered} M_T=\begin{pmatrix} m_{t} \\ m_{t+1} \\ \vdots \\ m_{t+T} \end{pmatrix}, U^f=\begin{pmatrix} u_{t} \\ u_{t+1} \\ \vdots \\ u_{t+T} \end{pmatrix}, U^P=\begin{pmatrix} u_{t-1} \\ u_{t-2} \\ \vdots \\ u_{t-T} \end{pmatrix}, S^P=\begin{pmatrix} s_{t} \\ s_{t-1} \\ \vdots \\ s_{t-T} \end{pmatrix}\\ H_u=\begin{pmatrix} h_0^u & & & \\ h_1^u& h_0^u & & \\ \cdots & \cdots & \ddots & \\ h_T^u & \cdots & \cdots & h_0^u \end{pmatrix}, G_u= \begin{pmatrix} h_1^u & \cdots &\cdots & h_T^u \\ h_2^u& \cdots & h_T^2 & 0 \\ \cdots & \reflectbox{\(\ddots\)} & & 0 \\ h_u^T & \cdots & \cdots & 0 \\ 0 & 0 & 0 & 0 \end{pmatrix}, G_s= \begin{pmatrix} h_0^s & \cdots &\cdots & h_T^s \\ h_1^u& \cdots & h_T^s & 0 \\ \cdots & \reflectbox{\(\ddots\)} & & 0 \\ h_T^s & \cdots & \cdots & 0 \\ \end{pmatrix} \end{gathered}\] Our goal is to find \(U^f\) such that \(M^T=0\) thus \(U^f=(-H_u^+G_u)U^P+(-H_u^+)S^P\). Because our interest is in actuation at time \(t\), we have \(u(t)=U^f(1)\). This is computed at every time step. One should note the similarities with model predictive control (MPC), where the model is iteratively updated in conjunction with a cost minimizing control law at each time step, see. \(H_u^+\) denotes the pseudo-inverse (). A simple inverse amplifies high frequencies, yielding an impractical control law. Using a pseudo-inverse with non zero tolerance dampens high frequencies giving a smoother and hardware viable control law. In practice the tolerance level must be chosen such that actuation can follow the control law. Since actuator cannot achieve changes faster than 1 Hz, the tolerance level was chosen such that the impulse response control law did not exhibit fluctuations above 1 Hz, leading to a value of 2.5. Figure [\[fig:control_law\]](#fig:control_law){reference-type="ref" reference="fig:control_law"}a compares the controlled and uncontrolled response of m(t) to an impulse in \(s(t)\). Figure [\[fig:control_law\]](#fig:control_law){reference-type="ref" reference="fig:control_law"}b shows the corresponding non dimensional control law \(a_0 (t) = u_j (t) / U_{\infty}\). These figures show that while complete fluctuation negation is impossible, fluctuation damping is achievable. Such a control law will negate a portion of upstream disturbances. Furthermore since part of the perturbation will not have the chance to be further amplified in the shear layer this should result in noteworthy reduction in downstream TKE fluctuations. found a far greater reduction for the impulse responses. One of the reasons for this is the location of the actuator, at the wall in our experiment, instead of in the bulk above the wall in the numerical simulation. The vortices created by the obstacle travel too far from the wall (approximately one step height) to be as successfully suppressed. ## Control results Figure [\[fig:control_results\]](#fig:control_results){reference-type="ref" reference="fig:control_results"} shows a comparison between outputs for the controlled and uncontrolled flow. Comparison was done over 14 minutes (21000 iterations). The results clearly show a reduction in fluctuations for the controlled flow (-35 %). Moreover a reduction in mean value is also observed (-15 %). The mean value reduction is an added benefit of fluctuation reduction. Better performances could be expected when considering the impulse responses. Additional noise sources not accounted for by the upstream sensor are likely to be present in an experimental flow, contributing to degraded performance.\ Figure [\[fig:TKE_2D\]](#fig:TKE_2D){reference-type="ref" reference="fig:TKE_2D"} shows the mean TKE field for the controlled and uncontrolled flows in the region of interest. The reduction in mean TKE is clear, as is a slight augmentation in recirculation size. Furthermore the effects of control are heterogeneous: while the TKE in the recirculation is mainly unaffected, the region of high TKE induced by the obstacle is successfully suppressed.\ Figure [\[fig:freq_u\]](#fig:freq_u){reference-type="ref" reference="fig:freq_u"}a shows the non-dimensional control output sampled over one minute. One can see that the control signal is one of periodic suction. Figures [\[fig:freq_u\]](#fig:freq_u){reference-type="ref" reference="fig:freq_u"}b and [\[fig:freq_u\]](#fig:freq_u){reference-type="ref" reference="fig:freq_u"}c show the frequency spectra for \(s(t)\) and \(u(t)\). A double peak is present in both spectra for the same frequency. This explains the physical processes involved during control. An incoming vortex is detected as a spike in \(s(t)\). The response is a sharp aspiration as shown by figure [\[fig:control_law\]](#fig:control_law){reference-type="ref" reference="fig:control_law"}. Thus, the control is operating in opposition. # Conclusion For the first time, an experimental implementation of a feed-forward control algorithm based on a ARMAX model was conducted on a backward-facing step flow. Results show the validity of such an approach. Nevertheless, to ensure successful implementation special care should be given to actuation, in particular to prevent contamination of the upstream sensor. Moreover, this approach is limited to the linear regime of the flow.\ Analyzing impulse responses gives valuable insight into the flows controllability as well as the potential for success of the method. While these responses tell us full negation of upstream disturbances is impossible, the computed model was able to reliably predict flow responses and yield a control law able to reduce energy levels and fluctuations. Future work should involve span-wise sensors and actuators thus allowing span-wise heterogeneous disturbances to be controlled as proposed and evaluated numerically by.

{'timestamp': '2014-03-25T01:11:05', 'yymm': '1401', 'arxiv_id': '1401.5635', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5635'}

# Introduction {#sintro} In the standard cosmological scenario, non-baryonic Dark Matter (DM) plays a decisive role, representing approximately 27% of the mass-energy content of the Universe. The existence and properties of weakly interacting massive DM particles are inferred from mass measurements of galaxies and galaxy clusters derived from kinematics, gravitational lensing or other probes. Additional indirect evidences come from the influence DM has on the formation and evolution of the Large Scale Structure of the Universe, and from its effects on the temperature anisotropies observed in the Cosmic Microwave Background. A widely adopted assumption in cosmology is that DM is a pressureless fluid. Although such hypothesis has not been directly proven yet, in a Universe where the DM fluid has a large pressure, there would not be enough time for cosmological structures to grow between the recombination epoch and today. Moreover, the observed properties of the Large Scale Structure are consistent with the current pressureless DM scenario. Despite considerable efforts for a direct detection of DM with underground experiments (e.g., DAMA[^1], XENON[^2], PICASSO[^3], XMASS[^4], CDMS[^5]), no significant signal has been found to date and the nature of DM remains unknown. The current pressureless scenario sets the DM Equation of State parameter to be zero, by definition. The aim of this Letter is to directly test this assumption by constraining the EoS parameter of the galaxy cluster fluid, using new accurate measurements of a cluster mass profile from lensing and kinematics analyses. first proposed to combine the analyses of the rotational curves and of the lensing signal in spiral galaxies to constrain the amount of DM pressure. generalized the approach without assuming any model for both the DM EoS and the rotational curves. Finally, extended the method to the case of galaxy clusters. In this Letter, we apply this method to the \(z=0.44\) cluster MACS 1206.2-0847 (hereafter MACS 1206), which has been studied in detail as part of the Cluster Lensing And Supernova survey with Hubble. have used new high quality imaging of this cluster obtained with Subaru and the Hubble Space Telescope (HST), to derive its mass density profile from weak-lensing distortion, magnification, and strong-lensing analyses. (hereafter B13) have used an unprecedented data-set of about 600 measured redshifts of cluster members, obtained as part of a VLT/VIMOS large programme (ID 186.A-0798), to determine the cluster mass profile over the radial range of 0.05-2.5 virial radii by applying the Jeans equation and the caustic technique of. By using the complementary information provided by the kinematics and lensing mass profile determinations, in this Letter, we constrain the DM EoS parameter which encapsulates information on the amount of fluid pressure. # Theoretical Framework {#stheo} A static, spherically symmetric spacetime is described by the metric: \[\label{eqmetric} \mathrm{d} s^{2}=-e^{2\Phi(r)/c^2}\, \mathrm{d} t^{2} + e^{2\lambda(r)} \mathrm{d} r^{2} +r^{2}\, \mathrm{d}\theta^2 + r^{2}\,\sin^2\!\theta\mathrm{d}\varphi^2 \,,\] where \(\Phi(r)\) and \(\lambda(r)\) are two arbitrary generic functions of the metric. One can apply to such metric the Einstein field equations, \(G_{\alpha\beta} = \left( 8\pi G/c^4\right) T_{\alpha\beta}\), by using the stress-energy tensor of an ideal gas with radial, \(T_{rr}\), and tangential, \(T_{\theta \theta}\), pressure components as \(T_{\alpha\beta}= \left(\rho+\frac{p}{c^2}\right) u_{\alpha} u_{\beta} + p g_{\alpha\beta}\). In these equations, \(G_{\alpha\beta}\) is the Einstein tensor, \(G\) is Newton's gravitational constant, \(c\) is the speed of light in vacuum, \(u_{\alpha}\) is the fluid's four velocity, and \(g_{\alpha \beta}\) is the metric tensor. It is then possible to obtain the density profile and the pressure profiles in the radial and transverse directions: \[\begin{aligned} \label{eqden_prof} \rho(r)\, \!&=&\! \, \frac{1}{4\pi} \frac{m^{\prime}(r)}{r^2} \,, \\ \label{eqpr_prof} p_r(r) \!&=&\!-\frac{c^4}{8\pi G} \frac{2}{r^2}\, \left[ \frac{G m(r)}{c^2 r}-r\,\frac{\Phi^{\prime}(r)}{c^2}\left( 1-\frac{2\, G m(r)}{c^2 r} \right) \right], \\ \label{eqpt_prof} \nonumber p_t(r) \, \!&=&\! \, \frac{c^4}{8\pi G} \left\lbrace \left( 1-\frac{2\,G m(r)}{c^2 r} \right) \frac{1}{c^2} \left[ \frac{\Phi'(r)}{r} + \frac{\Phi'(r)^2}{c^2} +\Phi''(r) \right] \right. \\ &~&-\left. \frac{G}{c^2} \left(\frac{m(r)}{r}\right)^\prime \left(\frac{1}{r} +\frac{\Phi^\prime(r)}{c^2}\right) \right\rbrace \,, \end{aligned}\] where prime denotes the partial derivative with respect to the r-coordinate, \(m(r)\) is the mass within a sphere of radius \(r\), and the metric function \(e^{2\lambda(r)}\) is defined as \[\label{eqm_fun} e^{-2\lambda(r)} \stackrel{\Delta}{=} 1-\frac{2\, G\, m (r) }{c^2\,r}\.\] For a medium with isotropic pressure, the metric Eq. [\[eqmetric\]](#eqmetric){reference-type="ref" reference="eqmetric"} (describing the geometry of the spacetime inside the cluster) matches the Schwarzschild metric (describing the geometry of the space time outside the cluster) at \(r=\tilde{r}\), where \(\tilde{r}\) is the size of the cluster, only if \(p_r=p_t=0\). Thus, to check the validity of the pressureless assumption, throughout this Letter, we consider a fluid with an anisotropic pressure and the most general definition of EoS: \[\label{eqeos} w(r) \,=\, \frac{p_r(r) + 2\, p_t(r)}{3\, c^2 \rho(r)} \,.\] The functions \(\Phi(r)\) and \(m(r)\) fully characterize the metric, and the density and pressure profiles, and, consequently, the EoS. We calculate the two metric functions by using the mass profiles of the cluster derived from the velocity distribution of cluster galaxies and from the gravitational lensing measurements. In general relativity, all the mass-energy content of the cluster shapes the gravitational potential well. However, the two probe particles used in the aforementioned analyses, the galaxies and the photons, perceive the gravitational potential in different ways, due to their distinct properties. To calculate the trajectory of the galaxy test particles, with velocity \(v\ll c\) (B13 measured the velocity dispersion of the MACS 1206 cluster \(\sigma_{los} = 1087^{+53}_{-55} \; \mathrm{km\, s}^{-1}\) ), we relate Einstein's field equations to the Poisson equation in the weak field approximation \((2\Phi \ll c^2 \;\mathrm{and}\; 2mG/r\ll c^2 )\). Thus, the (0,0) component of Einstein's field equations reads \[\label{eqpoiss} R_{00} \approx \nabla^2 \Phi = \frac{4 \pi G}{c^2} \left( c^2 \rho + p_r + 2 p_t \right)\] for the cluster fluid, where the metric potential \(\Phi\) is different from \(\Phi_N\), the Newtonian potential. In the Newtonian limit, \(\rho\gg p/c^2\) and we recover the usual Poisson equation \(\nabla^2 \Phi_N = 4 \pi G \, \rho\). From Eq. [\[eqpoiss\]](#eqpoiss){reference-type="ref" reference="eqpoiss"}, we can see that the mass profile derived from the kinematics analysis, \(m_k (r)\), depends only on the (0,0) part of the metric. Therefore the relation between the Poisson equation (Eq. [\[eqpoiss\]](#eqpoiss){reference-type="ref" reference="eqpoiss"}) and this mass profile is \[\label{eqmkin} m_k (r) = \frac {r^2}{G} \nabla{\Phi}(r)\.\] Since, in the lensing analysis, the probe particles travel at the speed of light, the full treatment of the geodesics is needed, even in the case of weak field approximation. [see also @faber06] derived the expression of the lensing potential by applying Fermat's principle to the geodesics of the photons moving through the cluster potential described by the metric Eq.[\[eqmetric\]](#eqmetric){reference-type="ref" reference="eqmetric"}. In this context, the light-ray trajectory is fully described by the relativistic analogue of the refractive index. For the metric Eq.[\[eqmetric\]](#eqmetric){reference-type="ref" reference="eqmetric"}, the effective refractive index in the weak field approximation is \[\begin{aligned} \label{eqref_index} \nonumber n(r) \, \!&=&\! \, 1-\frac{\Phi (r)}{c^2}-\frac{G}{c^2} \int \frac{m(r)}{r^2} dr +\\ &~& + \,O \left[ \left( \frac{2Gm}{c^2r} \right)^2, \frac{2Gm}{r} \frac{\Phi}{c^2}, \frac{\Phi^2}{c^4} \right] \end{aligned}\] and it is possible to define the lensing potential as \[\label{eqphilen} 2 \Phi_l (r) = \Phi (r) + G \int \frac{m(r)}{r^2} dr \.\] The trajectory of a light particle is determined entirely by the effective refractive index \(n(r)\), thus the bending of light is not only due to \(\Phi (r)\) (see Eq. [\[eqpoiss\]](#eqpoiss){reference-type="ref" reference="eqpoiss"}), but also to an extra term due to the relativistic-pressure of the fluid. The mass profile derived from the lensing analysis can be related to the lensing potential through the Poisson equation \[\label{eqrholen} \rho_l (r) = \frac{1}{4 \pi G} \nabla^2 \Phi_l (r)\] \[\label{eqmlen} m_l(r) = \frac{r^2}{G} \Phi^{\prime}_l (r) = \frac {r^2 }{2 G} \Phi^{\prime}(r) + \frac{m(r)}{2} = \frac{m_k(r)}{2} + \frac{m(r)}{2} \.\] Note that, while according to Eq.[\[eqpoiss\]](#eqpoiss){reference-type="ref" reference="eqpoiss"} the observed galaxy kinematics depends on the metric component \(g_{00}\) alone, the observed gravitational lensing potential reflects contributions from both \(g_{00}\) and \(g_{rr}\). Using the kinematic and the lensing mass profiles, we can finally determine the two metric functions \(\Phi (r)\) (from Eq.[\[eqmkin\]](#eqmkin){reference-type="ref" reference="eqmkin"}) and \(m(r)\) (from Eq.[\[eqmlen\]](#eqmlen){reference-type="ref" reference="eqmlen"}) and calculate the EoS of the cluster fluid (Eq.[\[eqeos\]](#eqeos){reference-type="ref" reference="eqeos"}). # The MACS 1206 cluster Mass Profiles {#sdata} The X-ray selected MACS 1206 cluster, at redshift 0.44, has been observed in the course of the CLASH survey. HST observations were completed in 2011. A detailed strong lensing model, based on the identification of 50 multiple lensed images of 13 background galaxies, was presented by. The combination of the inner mass density profile from this model with weak lensing shear and magnification measurements from *Subaru* multi-band images led to a reliable determination of the mass density profile of MACS 1206 out to \(\sim 2\) Mpc[^6]. The spectroscopic observations with the VLT, which led to a total of 2749 objects with reliable redshift measurements in the cluster field and the kinematic analysis, are described in B13. Using the projected phase-space distribution of these objects and several techniques for the rejection of interlopers, 592 cluster members were identified. This large spectroscopic sample was used to determine the kinematic mass profile out to the virial radius (\(\sim2\) Mpc) by solving the Jeans equation with the MAMPOSSt technique, further extended to 5 Mpc with the caustic method of. The kinematic determination of the cluster mass profile is in very good agreement both with the lensing determination, and with that based on X-ray *Chandra* observations, under the assumption of hydrostatic equilibrium. The fact that different probes of the cluster mass profile converge to similar results[^7] suggests that systematic effects in the mass determination are negligible and that the cluster is dynamically relaxed. Moreover, the analysis of did not find significant level of substructure within this cluster, when using the most conservative membership selection. The concentric distribution of different mass components (dark matter, stellar light and gas) also underscores an equilibrium configuration. All these properties make MACS 1206 an ideal candidate for testing the EoS of the cluster fluid. parametrized the lensing mass profile of MACS 1206 with the NFW model. The same model was found by B13 to provide the highest likelihood fit to the kinematic data. We therefore use the NFW model parametrization of the kinematic and lensing mass profiles in our analyses. To check the sensitivity of our results on the kinematic mass profile used, we also consider the and profiles. These models are quite different from NFW and yet were found to provide acceptable fits to the kinematic data (B13). Unlike the lensing determination of the cluster mass profile, the kinematic determination also requires modeling the velocity anisotropy profile of the tracers of the gravitational potential, \(\beta(r)\). B13 considered three possible ansatz models, named 'O', \[\beta_{\mathrm{O}}(r) \,=\, \beta_{\infty} \frac{r-r_{-2}}{r+r_{-2}} \,, \label{eq:betao}\] 'T', \[\beta_{\mathrm{T}}(r) \,=\, \beta_{\infty} \frac{r}{r+r_{-2}} \,, \label{eq:betat}\] and 'C' with a constant anisotropy with radius. In Eqs. [\[eq:betao\]](#eq:betao){reference-type="ref" reference="eq:betao"} and [\[eq:betat\]](#eq:betat){reference-type="ref" reference="eq:betat"}, \(r_{-2}\) is the scale radius at which the logarithmic derivative of the mass density profile equals \(-2\), \(d \,\ln \rho /d\, \ln r =-2\) and \(\beta_{\infty}\) is the anisotropy value at large radii. The O model gives the smallest product of relative errors in the two free parameters of the mass profile (\(r_{200}\)[^8] and \(r_{-2}\)), namely it maximizes the ratio \((r_{200} \, r_{-2})/(\delta r_{200} \, \delta r_{-2})\), where \(\delta r_{200}\) and \(\delta r_{-2}\) are the (symmetrized) errors on \(r_{200}\) and \(r_{-2}\), respectively. Therefore the NFW+O model is adopted as our reference model in the following analysis. # Results In this section, we present and discuss our results on the EoS of the DM fluid as obtained from the analysis of the lensing and kinematic mass profiles of MACS 1206 described above. Derivatives of the potential and the mass profiles in Eqs.([\[eqden_prof\]](#eqden_prof){reference-type="ref" reference="eqden_prof"}-[\[eqpt_prof\]](#eqpt_prof){reference-type="ref" reference="eqpt_prof"}) are computed directly from the models of and B13. For details see Section 2 and 3. In Fig.[\[figeos_profiles\]](#figeos_profiles){reference-type="ref" reference="figeos_profiles"}, we show the resulting EoS parameter \(w(r)\) (Eq.[\[eqeos\]](#eqeos){reference-type="ref" reference="eqeos"}) as a function of the cluster centric radius. The statistical errors have been calculated via \(10^4\) MonteCarlo resamplings by propagating the uncertainties on the parameters of the lensing and kinematic mass profile models, \(r_{200}\) and \(r_{-2}\). As for the parameters derived from kinematics analysis, B13 shows that \(r_{200}\) and \(r_{-2}\) have uncorrelated errors thus we use the probability distributions shown in Fig. 9 of B13 to explore the parameter space in the MonteCarlo sampling. shows that the joint weak and strong lensing contours of the mass profile model are elliptical, so we can assume that they are Gaussian distributed but with covariance between the model parameters. We perform a MonteCarlo Markov Chain fitting to the weak and strong lensing radial profiles to generate posterior probability distribution functions in \(r_{200}\) and \(r_{-2}\) to consider the covariance between these two parameters in the calculation of the errors on the EoS parameter profile. In Fig.[\[figeos_profiles\]](#figeos_profiles){reference-type="ref" reference="figeos_profiles"}, the shaded area indicates the error computed from the \(16^{th}\) and \(84^{th}\) percentiles of the probability distribution at varying radii, for the reference NFW+O model. In Table [1](#t:w){reference-type="ref" reference="t:w"}, we list the mean values of \(w\), calculated over the radial range \(0.5\, \mathrm{Mpc} \leq r\leq r_{200}\), where \(w(r)\) is approximately constant. Since the errors on the EoS parameter at different radii, \(w(r)\), are highly correlated, we list \(w\) errors in Table [1](#t:w){reference-type="ref" reference="t:w"} obtained by computing the average of the \(w\) uncertainties over the considered radial range. We limit the radial range to radii \(\leq r_{200}\) because at larger radii both the kinematics and the lensing determinations of the cluster mass profile might be affected by systematics (deviation from dynamical relaxation for the kinematics, and contamination by a large-scale structure filament for the lensing). Moreover, the DM, whose EoS we want to constrain, dominates the mass budget in this radial range. Both in the Figure and in the Table, we show how our results are modified when we use different models for the anisotropy and kinematic mass profiles. The results on \(w(r)\) are sensitive to the adopted models, but still within the statistical uncertainties of the reference models. In Table [1](#t:w){reference-type="ref" reference="t:w"}, all the mean values are consistent with zero, thus with the usually adopted pressureless assumption. Note that the mean values cannot be directly inferred from the curves of Fig.[\[figeos_profiles\]](#figeos_profiles){reference-type="ref" reference="figeos_profiles"} because the errors are asymmetric. In particular, for the reference NFW+O model, we find a radially averaged value \[w = 0.00\pm 0.15 \rm{(stat)} \pm 0.08 \rm{(syst)} \,,\] where the statistical error is listed in Tab. [1](#t:w){reference-type="ref" reference="t:w"} and the systematic error reflects the peak to peak variation in the mean values of \(w\) obtained using different mass and anisotropy models. In our estimate, we cannot disentangle the mass profile of the baryons from that of the DM, however baryons contribute at most \(15\%\) of the cluster total mass at all radii. Moreover, we estimate that the baryon contribution to \(w\) is \(\sim 10^{-5}\). This is found by using the EoS of an ideal gas for the hot intra-cluster medium. We can therefore assume that the EoS parameter found in our study describes the behavior of the DM fluid in the cluster. A previous analysis of the DM EoS has found tentative evidence for a negative value of \(w(r)\). This result was obtained by using the lensing and kinematic mass determinations of two clusters: Coma and CL0024+1654, which however are known to contain major substructures, and their presence could affect the kinematic mass profile determinations. In particular, CL0024+1654 is possibly in a post-merger state after the core passage of two clusters occurring along the line of sight. In any case, our constraints on \(w\) are significantly more precise (by a factor of \(\gtrsim 3\) than those obtained by thanks to the significantly better quality of both our lensing and kinematic data. [\[t:w\]]{#t:w label="t:w"} # Summary and Conclusions In this Letter, we discussed how the pressureless assumption for the DM fluid can be quantitatively verified and we obtained the most stringent constraint on the DM EoS available to date by using high quality kinematic and lensing mass analyses of the relaxed CLASH cluster MACS 1206. We confirmed the pressureless assumption, namely \(w = 0.00\pm 0.15 \rm{(stat)} \pm 0.08 \rm{(syst)}\). We find no radial dependence of \(w\) outside the central (0.5 Mpc) region. The CLASH ­ VLT final sample will provide accurate mass profiles for 12 clusters, allowing us to place stronger constraints on the DM EoS parameter by stacking the information from all clusters. This will reduce our statistical errors, and, most importantly, possible systematic effects in the mass profile determinations such as departure from dynamical equilibrium, and contamination by large scale structure along the line of sight. If a departure from sphericity of the cluster potential well is detected, it is still possible, in principle, to apply the method used in this Letter to calculate the DM EoS parameter. showed how, given a mass distribution, it is always possible to recover the corresponding density and pressure distributions in absence of any particular potential symmetry and thus calculate the DM EoS parameter. From an observational point of view, cluster orientation and asphericity can systematically affect the mass profile determinations and consequently the EoS parameter. It is possible to reduce the impact of these errors by stacking results derived from a large sample of clusters. [^1]: http//people.roma2.infn.it/ dama/web/home.html [^2]: http//xenon.astro.columbia.edu/ [^3]: http//www.picassoexperiment.ca/ [^4]: http//www-sk.icrr.u-tokyo.ac.jp/xmass/index-e.html [^5]: http://cdms.berkeley.edu/ [^6]: and B13 adopted \(\Omega_m =0.3, \Omega_{\Lambda} =0.7, h=0.7\). [^7]: Note that the similarity of the kinematic and lensing mass profiles does not imply a pressureless fluid \((w(r) =0)\), since, in first approximation, \(w(r)\) does not depend on the profiles, but on their derivatives \[w(r)\approx \frac{2}{3}\frac {m_k^{\prime}(r)-m_l^{\prime}(r)}{2 m_l^{\prime}(r)-m_k^{\prime}(r)}\]. [^8]: The radius \(r_{200}\) is the radius of a sphere with mass overdensity \(\Delta\)=200 times the critical density of the Universe at the cluster redshift. We use \(r_{200} \simeq 2\) Mpc from and B13.

{'timestamp': '2014-01-24T02:00:35', 'yymm': '1401', 'arxiv_id': '1401.5800', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5800'}

# Introduction Between 10\(^{8}\) and 10\(^{9}\) years after the Big Bang, the intergalactic medium (IGM) changed from being cold and neutral to being warm and ionized. An early generation of massive stars with very low metal content producing copious UV radiation is the most likely cause of reionization. X-rays from accretion onto compact objects, either stellar mass or supermassive, are thought to have heated the IGM and may have made a \(\sim 10\) per cent contribution to reionization. X-rays from active galactic nuclei (AGN) dominate the X-ray background at \(z < 5\), but recent work places strong constraints on X-ray emission from AGN at high \(z\). Recent work on the formation of the first stars suggest that they had typical masses of a few tens of \(M_{\odot}\) and formed in small multiple systems often dominated by a binary. The formation of supermassive objects, \(\ge 10^{5} M_{\odot}\), was rare. Thus, accretion onto stellar-mass compact objects likely dominated the X-ray emission in the early universe with a significant fraction occurring in binary systems. It is not possible to detect HMXBs in the early universe using current instruments to obtain direct observational constraints. A number of authors have considered the effects of HMXBS on the early universe with inferences about their properties drawn from theory or observations of Milky Way black hole binaries and have assumed hard power-law spectra with photon indices of 1.3--2.0 sometimes with a soft thermal component providing a minority of the total energy output. Here, we attempt to infer the properties of early HMXBs based on local analogues in nearby galaxies. The early universe was highly metal deficient. Metallicity appears to have a strong impact on the formation and properties of X-ray binaries. We discuss the properties of X-ray binaries found in low metallicity environments in the local universe in Section [2](#sec:binaries){reference-type="ref" reference="sec:binaries"} and some of the implications for the study of binaries in the early universe in Section [3](#sec:implications){reference-type="ref" reference="sec:implications"}. Section [4](#sec:summary){reference-type="ref" reference="sec:summary"} is a brief summary. # X-ray binaries at low metallicity {#sec:binaries} The bulk of X-ray emission from galaxies not hosting an active nucleus arises from compact objects formed at the endpoint of the stellar evolution of massive stars (typically \(> 8 M_{\odot}\)) in binary systems. The X-ray binaries that form within 100 Myr after birth of their parent stars are dominated by high mass X-ray binaries (HMXBs) in which the companion star has a high mass and therefore evolves rapidly. HMXBs also dominated the X-ray binary population of the early universe due to their rapid evolution. The number of HMXBs and the total X-ray luminosity they produce is well correlated with the host galaxy star formation rate. Furthermore, the occurrence of very bright HMXBs or ultraluminous X-ray sources (ULXs), defined here as having luminosities above \(10^{39} \rm \, erg \, s^{-1}\), is correlated with recent starburst activity. The number of ULXs in and the total X-ray luminosity of a galaxy, normalised by its star formation rate (SFR), increases with decreasing metallicity. These trends are enhanced at very low metalliticies, \(Z/Z_{\odot} < 0.1\), where the number of ULXs increases by a factor of \(7 \pm 3\) (normalized to SFR) relative to near-solar-metallicity galaxies and the total X-ray luminosity increases by a factor of \(9.7 \pm 3.2\). Population synthesis simulations of the formation and evolution of high mass X-ray binaries in low metallicity environments also suggest enhanced X-ray emission. found that the number of HMXBs (per unit SFR) increased about a factor of 3 between solar and Small Magellanic Cloud (SMC) metallicity \(Z/Z_{\odot} = 0.2\). found that the number of HMXBs at \(Z/Z_{\odot} = 0.02\) increased by a factor of \(\sim 3.5\) relative to solar metallicity and the number of ULXs increased by a factor of \(\sim 5\) due to an increase in the number of binary systems evolving through the pathways that form Roche-lobe overflow systems and supergiant wind accretors. Also, the maximum black hole mass increases with decreasing metallicity, reaching as high as \(80 M_{\odot}\) for \(Z/Z_{\odot} = 0.01\) compared with \(\sim 20 M_{\odot}\) for \(Z/Z_{\odot} \sim 1\). Thus, it is possible to form more luminous X-ray binaries at low metallicity without violating the Eddington limit. ## X-ray spectra {#sec:spectra} It is clear that the number of HMXBs and their total X-ray luminosity is enhanced at low metallicity. A key question for understanding the contribution of X-ray binaries to heating and reionization is the shape of their emission spectrum. The HMXB X-ray luminosity function (XLF) is well described by a power-law distribution with an exponent near-1.6 and a cut-off near \(10^{41} \rm \, erg \, s^{-1}\). The XLF in low metallicity galaxies has the same form and exponent. Figure [\[xlf\]](#xlf){reference-type="ref" reference="xlf"} shows the integrated luminosity from sources with luminosities above a given threshold. The shallow slope of the HMXB XLF implies that the total luminosity is dominated by bright sources. Half of the flux from active sources (with \(L_X > 10^{36} \rm \, erg \, s^{-1}\)) is produced by sources with \(L_X > 4 \times 10^{39} \rm \, erg \, s^{-1}\), 71 percent by sources with \(L_X > 10^{39} \rm \, erg \, s^{-1}\), and 90 per cent by sources with \(L_X > 10^{38} \rm \, erg \, s^{-1}\). The total luminosity and, hence, the spectral shape of the emission from a population of HMXBs is dominated by ULXs and HMXBs in high luminosity states. Black hole X-ray binaries in the Milky Way exhibit hard spectra, following a power-law form with photon index near 1.7, at low luminosities, and then transition to states with softer spectra at higher luminosities. These softer spectra are a combination of a thermal component with temperatures up to a few keV and steep power-laws with photon indices near 2.5. These sources reach luminosities up to \(\sim 10^{39} \rm \, erg \, s^{-1}\). ULXs produce a large fraction of the total X-ray luminosity and exhibit spectra that show distinct curvature in the 2--8 keV band. This was first demonstrated with high statistical quality XMM-Newton spectra. The curvature in ULX spectra has recently been confirmed with Suzaku and joint NuSTAR/XMM-Newton observations. Suzaku observations of the starburst galaxy M82 show that its X-ray spectrum, seen in Chandra imaging to be dominated by a ULX, has a cut-off energy of \(5.7 \pm 0.7\) keV when fitted with an exponentially cut-off model. XMM-Newton/NuSTAR spectra of the ULX NGC 1313 X-2 show cut-off energies near 2.5 keV while those of NGC 1313 X-1 show cut-off energies in the range 5--9 keV. The HMXB in the very low metallicity galaxy I Zw 18 (\(Z/Z_{\odot} = 0.019\)), shows a transition from a hard power-law spectum at low luminosity to a curved spectrum at high luminosity. When fitted with an exponentially cut-off model, the high luminosity spectrum has a cut-off energy of \(2.1 \pm 0.6\) keV and a photon index of \(0.80 \pm 0.26\). Some recent papers, e.g.  have assumed a hard power-law with photon index of 1.7 for the HMXB spectrum based on measurements of HMXBs in nearby galaxies made by using Chandra. These results are influenced by the low numbers of counts in the Chandra spectra. find that a power law is unacceptable for 39 per cent of the spectra with more than 1000 counts but only 8 per cent of the spectra with fewer than 200 counts. The average photon index was derived using only those spectra for which a power law provided a statistically acceptable fit, systematically excluding those spectra exhibiting evidence for curvature and indirectly excluding a higher fraction of spectra with larger numbers of counts. Other work has considered the transition to softer spectra at the luminosities seen from Milky Way X-ray binaries, but not the curved spectra seen from ULXs. The higher quality XMM-Newton, NuSTAR, and Suzaku spectra provide a better measurement of the true spectral shape of the very bright HMXBs that dominate the total HMXB luminosity and consistently show curvature at high energies. It is possible to get some constraints on the spectral shape of the X-ray emission from HMXBs at moderate redshifts from deep X-ray surveys. performed a stacking analysis of star-forming galaxies in the Chandra Deep Field South and calculated the counts in two energy bands (0.5--2 keV and 2--8 keV) for sets of galaxies in different redshift intervals. The galaxies are significantly detected in the harder band only for the redshift bands \(z=0-1\) and \(z = 1-2\). The hardness ratios are \(0.28 \pm 0.08\) and \(0.36 \pm 0.11\), respectively. Using the known response of Chandra it is possible to calculate the expected counts ratio for an exponentially cut-off power-law model with a given cut-off energy and photon index--note that it is necessary to take into account the redshift in order to interpret the cut-off energy in the galaxy rest frame. Figure [\[cowie_cutpl\]](#cowie_cutpl){reference-type="ref" reference="cowie_cutpl"} shows the exponentially cut-off power law model parameters consistent with the measured counts ratios for the \(z=0-1\) and \(z = 1-2\) redshift bands. The allowed parameter range is consistent with the parameters measured for the ULXs mentioned above. # Implications {#sec:implications} Studies of galaxies with low metallicity suggest that the total luminosity of early X-ray binaries (normalised to SFR) is likely enhanced relative to near-solar metallicity systems and that the total spectrum is likely dominated by systems in high luminosity states with curved spectra. This spectral curvature may potentially affect: constraints on early X-rays from measurements of the soft X-ray background, constraints on the ratio of X-ray luminosity to SFR at high redshift obtained from deep X-ray surveys, the detectability of the first HMXBs, the morphology of X-ray heating, and the 21 cm radio signal from the epoch of reionization. Figure [\[zspec\]](#zspec){reference-type="ref" reference="zspec"} shows power law and exponentially cut-off power law spectra at redshifts of \(z=0\) and \(z=6\) in the top panel and the observed flux in the 0.5--2 keV band as a function of redshift for these models in the bottom panel--essentially the 'K-correction' for each spectral model. This band is often used in deep X-ray surveys. For a fixed observational energy band, an exponential cut-off in the spectrum has a strong effect on the observed flux as the red-shifted cut-off energy moves into the band. Because of the rather hard photon index measured for I Zw 18, this effect is somewhat negated at low redshifts because that spectrum rises with energy at energies below the cut-off. Spectral curvature affects observational constraints on X-rays from early binaries. With the assumption of a hard, power law spectrum for the HMXB emission, a significant HMXB contribution to reionization is excluded. However, even with a break energy as high as 5 keV, the cut-off energy of emission from \(z = 10\) would be redshifted to 0.5 keV, below the 0.5--2 keV band typically used in deep X-ray surveys. If the total HMXB emission spectrum has significant curvature as seen in the ULX spectra, then constraints from the soft X-ray background on the emission from early X-ray binaries become very weak. X-ray emission sufficient to fully power reionization is allowed. Deep X-ray surveys can also be used to constrain the emission from HMXBs and the ratio of X-ray luminosity to SFR (\(L_X\)/SFR) at high redshift. As noted above, performed a stacking analysis of star-forming galaxies in the Chandra Deep Field South. In addition to calculating the total X-ray emission in various redshift bands, they also estimated the total SFR of the target galaxies in each band. Correcting for the effect of redshift on the observed flux in the 0.5--2 keV band using a power-law model with a photon index of 2, they concluded there is no evidence for redshift evolution of \(L_X\)/SFR. In contrast, Basu-Zych et al. (2013a) found an increase in \(L_X\)/SFR at high redshifts. Figure [\[lxsfrz\]](#lxsfrz){reference-type="ref" reference="lxsfrz"} shows the effect of spectral curvature on the redshift dependent estimates of \(L_X\)/SFR of Cowie et al. (2012. The increase in the K-correction caused by the spectral curvature progressively increases the estimated values of \(L_X\)/SFR at increasing redshifts. The results then suggest a moderate increase in \(L_X\)/SFR at intermediate redshifts, \(z=3-5\), and are consistent with relatively large enhancements at high redshifts. For the exponentially cut-off power law spectrum with a cut-off energy of 5 keV, the upper limits (90 per cent confidence) on the enhancement in \(L_X\)/SFR are 4.2 and 6.2 in the two highest redshift bins, \(z=\)`<!-- -->`{=html}5--6 and 6--7. With a cut-off energy of 2.1 keV, the corresponding upper limits are 7.2 and 11.9. The latter constraints are consistent with the increase in \(L_X\)/SFR expected due to the decrease in metallicity at high redshifts while the former would suggest that the results from low-metallicity nearby galaxies, mainly blue compact dwarfs in the samples of and, may overpredict the X-ray emission from the larger galaxies detected at high redshifts. Spectral curvature combined with interstellar absorption within the Milky Way may limit the detectability of HMXBs at high redshifts. The lines of sight with the lowest column density within the Milky Way are within the Lockman hole and have absorption column densities near \(N_H = 6 \times 10^{19} \rm \, cm^{-2}\). Figure [\[lockman\]](#lockman){reference-type="ref" reference="lockman"} shows the X-ray absorption for this column density along with an exponentially cut-off power law spectrum with cut-off energy of 2.1 keV at redshifts of \(z=0\), 6, and 12. It is apparent that interstellar absorption has a severe effect, strongly reducing the X-ray flux and the band over which X-rays are detected. We note that current CCD detectors operate efficiently only above energies of about 0.2 keV. This may place an even more severe constraint on the detectability of high redshift HMXBs. Spectral curvature will also affect how far X-rays from early HMXBs penetrate into the IGM and therefore the morphology of X-ray heating. This affects the 21-cm radio emission from the epoch of reionization and it may be possible to place constraints on the total spectrum of early X-ray emission when adequate 21-cm maps become available. For truly pristine gas with zero metallicity, the highest ionisation edge is at 54.4 eV so the effect of spectral curvature above 2 keV is not so pronounced. However, metal enrichment local to the site of HXMB formation could have an effect. Spectral curvature will also have an effect on the temporal extent of X-ray heating since it affects how the observed/absorbed spectrum evolved with redshift. # Summary {#sec:summary} Study of nearby high mass X-ray binaries can provide insight into HMXBs formed early in the history of the universe. The dominant factor leading to change in the properties of early HMXBs compared with current HMXBs is the reduced metallicity in the early universe. Multiple observational and theoretical studies have shown that X-ray binary production is enhanced at low metallicity in terms of the number of HMXBs produced for a given amount of star formation, particularly for HMXBs with high luminosities, and the total X-ray luminosity produced. The total luminosity of current populations of HMXBs is dominated by sources at high luminosities. These sources exhibit X-ray spectra that show curvature above 2 keV. The same is likely true of HMXB populations at high redshifts. The spectral curvature changes the K-correction for X-rays from HMXBs in a manner that greatly weakens the constraints on X-ray emission of early HMXBs obtained from the soft X-ray background. Applied to deep X-ray surveys of star forming galaxies, the modified K-correction suggests a moderate increase in the ratio of X-ray luminosity to SFR (\(L_X\)/SFR) at intermediate redshifts, \(z=3-5\), and is not inconsistent with a large enhancement at high redshifts, \(z=6-7\).

{'timestamp': '2014-02-07T02:14:43', 'yymm': '1401', 'arxiv_id': '1401.5687', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5687'}

# Introduction and Background It is known, that a number of physical phenomena can be described mathematically by Mathieu equation. For instance, the equation of motion for the flux lattice in the supeconductor theory is the following \[\begin{split} \label{Eq1} m\frac{d^2y(t)}{dt^2} + \eta\frac{dy(t)}{dt} + (K_0 + k \cos\omega{t})y(t) = \frac{BJ_0}{c}\cos\Omega{t}, \end{split}\] where \(B\) is the magnetic induction in the sample, \(m\) is the total mass (per unit length) of the flux lattice, \(\eta\) is the viscosity coefficient, \(\omega\) is the frequency of the modulating magnetic field, \(c\) is the light velocity, \(J_0\) and \(\Omega\) are the amplitude and the frequency of the microwave current, respectively, \(K_0\) and \(k\) are amplitudes of constant component and an alternating component in the function \(K(t)\) in the relationship \[\begin{split} \label{Eq2} F(t) =-K(t)y(t) \end{split}\] between the force \(F(t)\) and small displacement \(y(t)\) of the flux lattice from its equilibrium position, that is \[\label{Eq3} K(t) = K_0 + k \cos\omega{t}.\] The equation ([\[Eq1\]](#Eq1){reference-type="ref" reference="Eq1"}) is the inhomogeneous damped Mathieu equation. Mathieu equation is well knowm in the theory of differential equations, see for example, , ,. At that,the so-called general Mathieu equation is the following equation \[\begin{split} \label{Eq4} \frac{d^2y(t)}{dt^2} + [h-2\theta \cos({2t})] y(t) = 0, \end{split}\] where \(h\) and \(\theta\) are real or complex constants. The known solution of the general Mathieu equation ([\[Eq4\]](#Eq4){reference-type="ref" reference="Eq4"}) is built in the form \[\begin{split} \label{Eq5} y(t) = \exp({\mu t}) P(t), \end{split}\] where \(P(t)\) is a periodical function with the period, equal to \(\pi\), \(\mu\) is so-called characteristic index, depending on the values of \(h\) and \(\theta\). The function \[\begin{split} \label{Eq6} y(t) = \exp({-\mu t}) P(-t), \end{split}\] represents itself the second solution. The solutions ([\[Eq5\]](#Eq5){reference-type="ref" reference="Eq5"}) and ([\[Eq6\]](#Eq6){reference-type="ref" reference="Eq6"}) are linearly independent, they produce the fundamental system of the solutions, excluding the case, when \(i\mu \in Z\), that is, to the set of whole numbers. Further, the solution ([\[Eq5\]](#Eq5){reference-type="ref" reference="Eq5"}) is written formally in the form of the following infinite series \[\begin{split} \label{Eq7} y(t) = \sum_{n =-\infty}^{\infty}c_n\exp{(\mu + 2ni) t}, \end{split}\] in which for the set of coefficients \(\{c_n\}\) the following recurrent relations \[\begin{split} \label{Eq8}-\theta c_{n-1} +[h + (\mu + 2ni)^2] c_{n}-\theta c_{n+1} = 0, n \in Z \end{split}\] were obtained. It is the only algorithm for the numerical solution. In fact, the analytical solution of the general Mathieu equation ([\[Eq4\]](#Eq4){reference-type="ref" reference="Eq4"}) was not found. The existence the only algorithn for the numerical solution is inconveniently for the practical usage of Mathieu equation, especially in the cases of physical applications when analytical dependences are required to understand the physical processes. For instance, the authors of the work have preferred instead of trying to solve the equation ([\[Eq1\]](#Eq1){reference-type="ref" reference="Eq1"}) numerically to use the linearized equation \[\begin{split} \label{Eq9} m\frac{d^2\delta y}{dt^2} + \eta{d\delta y}{dt} + K_0\delta y =-k \cos\omega{t}y_0, \end{split}\] in which \(y_0\) is the particular solution of ([\[Eq1\]](#Eq1){reference-type="ref" reference="Eq1"}) for the case of \(k = 0\). Let us remark, that an (in)homogeneous damped Mathieu equation ([\[Eq4\]](#Eq4){reference-type="ref" reference="Eq4"}) can be reduced to an (in)homogeneous general Mathieu equation by the following transformation of the function \(y(t)\): \(y(t) = w(t)\exp(-\frac {1}{2}\eta t)\). The authors of the work have found, to the first order in \(k\), \(\eta\), and \(\omega\), that the value of an electric field \(E(t)\) induced is \[\begin{split} \label{Eq10} E(t) = \frac{B^2J_0\Omega}{|K_0|c^2} [1-\epsilon \cos(\omega{t}-\phi)] \sin(\Omega{t}-\alpha), \end{split}\] where \(\epsilon = \frac{k}{K_0}\), \(\tan\phi = \frac{2\eta\omega}{K_0}\), and \(\tan\alpha = \frac{\eta\Omega}{K_0}\). So, the equation ([\[Eq10\]](#Eq10){reference-type="ref" reference="Eq10"}) indicates that the induced \(E(t)\) field is amplitude modulated. In other words, the concrete physical mechanism and its analytical description were established, alhough for rather restricted ranges of the parameters used. Really, the equation ([\[Eq10\]](#Eq10){reference-type="ref" reference="Eq10"}) is valid for \(k \ll |K_0|\), \(\omega \ll \Omega\), and \(|K_0|\gg m\Omega^2\). At the same time, the authors of remark, that when static magnetic field \(H\) becomes too large \[larger than 800 G\], the flux structure becomes probably too complex for the simple model proposed to remain valid. It is clear on the given example, that an analytical solution of Mathieu equation remains to be very actual for its applications in physical sciences and in engineering. On the other hand, an analytical solution of Mathieu equation has also the mathematical theoretical aspect. It is determined by the fact that the solution of a number of differential equations is reduced to the solution of Mathieu equation. They, for example, are 1) \[\begin{split} \label{Eq11} (1-t^2)\frac{d^2y(t)}{dt^2} + t\frac{dy(t)}{dt} + (2a t^2 + b)y(t) = 0, \end{split}\] The transformation of variable \(t = cos z\) leads to the Mathieu equation \[\begin{split} \label{Eq12} \frac{d^2y(z)}{dz^2} + (a + b + a cos2z)y(z) = 0, \end{split}\] 2) \[\begin{split} \label{Eq13} 2t(t-1)\frac{d^2y(t)}{dt^2} + (2t-1)\frac{dy(t)}{dt} + (a t + b) y(t) = 0, \end{split}\] The transformation of variable \(t = cos^2 z\) leads to the Mathieu equation \[\begin{split} \label{Eq14} \frac{d^2y(z)}{dz^2}-(a + 2b + a cos2z)y(z) = 0, \end{split}\] 3) \[\begin{split} \label{Eq15} \frac{d^2y(t)}{dt^2} + (a\sin \lambda t + b) y(t) = 0, \end{split}\] The transformation of variable \(\lambda t = 2z + \frac{\pi}{2}\) leads to the Mathieu equation \[\begin{split} \label{Eq16} \frac{d^2y(z)}{dz^2}-(\frac{4b}{\lambda^2} + \frac{4a}{\lambda^2}cos2z) y(z) = 0, \end{split}\] 4) The equations \[\begin{split} \label{Eq17} &\frac{d^2y(t)}{dt^2} + (a\sin^2t + b) y(t) = 0,\\ &\frac{d^2y(t)}{dt^2} + (a\cos^2t + b) y(t) = 0 \end{split}\] are transformed to Mathieu equations by using of trigonometric formulae \[\begin{split} \label{Eq18} &2\sin ^2 t = 1-\cos 2t,\\ &2\cos ^2 t = 1 + \cos 2t \end{split}\] correspondingly. The aim of the given work is to find the analytical solution of the general Mathieu equation. # Results **Theorem.** The general solution of the general Mathieu equation can be represented analytically to be the superposition of Bessel functions of the first and the second kinds. **Proof.** It is evident, that the equation ([\[Eq1\]](#Eq1){reference-type="ref" reference="Eq1"}) can be represented in the form \[\begin{split} \label{Eq19} &\frac{1}{2} m\frac{d^2y}{dt^2} + \frac{1}{2}\eta\frac{dy}{dt} + \frac{1}{2}K_0 y + \frac{1}{2}k \exp( {i\omega{t}}) y-\frac{BJ_0}{2c}\cos\Omega{t} +\\ &[\frac{1}{2} m\frac{d^2y}{dt^2} + \frac{1}{2}\eta\frac{dy}{dt} + \frac{1}{2}K_0 y + \frac{1}{2}k \exp(-{i\omega{t}}) y-\\ & \frac{BJ_0}{2c}\cos\Omega{t}] = (\hat{L}_1 +\hat{L}_2) y = 0, \end{split}\] where \(\hat{L}_1\) and \(\hat{L}_2\) are differential operators \[\begin{split} \label{Eq20} &\hat{L}_1 = \frac{1}{2} m\frac{d^2}{dt^2} + \frac{1}{2}\eta\frac{d}{dt} + \frac{1}{2}K_0 + \frac{1}{2}k \exp( {i\omega{t}})-\frac{BJ_0}{2c}\cos\Omega{t} \\ &\hat{L}_2 = [\frac{1}{2} m\frac{d^2}{dt^2} + \frac{1}{2}\eta\frac{d}{dt} + \frac{1}{2}K_0 + \frac{1}{2}k \exp(-{i\omega{t}}) -\\ & \frac{BJ_0}{2c}\cos\Omega{t}]. \end{split}\] It is also evident that the partial solution of the starting equation will correspond to the intersection of sets of the solutions satysfying simultaneously to the equations \[\begin{split} \label{Eq21} &\hat{L}_1 y = 0,\\ &\hat{L}_2 y = 0. \end{split}\] So, we have to solve the equations \[\begin{split} \label{Eq22} &\frac{1}{2} m\frac{d^2y}{dt^2} + \frac{1}{2}\eta\frac{dy}{dt} + \frac{1}{2}K_0 y + \frac{1}{2}k \exp( {i\omega{t}}) y = \\ &\frac{BJ_0}{2c}\cos\Omega{t},\\ &\frac{1}{2} m\frac{d^2y}{dt^2} + \frac{1}{2}\eta\frac{dy}{dt} + \frac{1}{2}K_0 y + \frac{1}{2}k \exp(-{i\omega{t}}) y = \\ &\frac{BJ_0}{2c}\cos\Omega{t} \end{split}\] or in the equivalent form \[\begin{split} \label{Eq23} &\frac{d^2y}{dt^2} + \frac{\eta}{m}\frac{dy}{dt} + \frac{K_0}{m} y + \frac{k}{m} \exp({i\omega{t}}) y =\\ &\frac{BJ_0}{mc}\cos\Omega{t},\\ &\frac{d^2y}{dt^2} + \frac{\eta}{m}\frac{dy}{dt} + \frac{K_0}{m} y + \frac{k}{m} \exp(-{i\omega{t}}) y =\\ &\frac{BJ_0}{mc}\cos\Omega{t},\\ \end{split}\] The case \[\begin{split} \label{Eq24} &\frac{1}{2} m\frac{d^2y}{dt^2} + \frac{1}{2}\eta\frac{dy}{dt} + \frac{1}{2}K_0 y + \frac{1}{2}k \exp({i\omega{t}}) y-\\ &\frac{BJ_0}{2c}\cos\Omega{t} =\\ &-[\frac{1}{2} m\frac{d^2y}{dt^2} + \frac{1}{2}\eta\frac{dy}{dt} + \frac{1}{2}K_0 y + \frac{1}{2}k \exp(-{i\omega{t}}) y-\\ &\frac{BJ_0}{2c}\cos\Omega{t}] \end{split}\] can be also studied. There are a number of variants to simplify the solution of ([\[Eq24\]](#Eq24){reference-type="ref" reference="Eq24"}) using the symmetry of \(y(t)\). In the case, if \(y(t)\) is even, \(\frac{dy}{dt}\) is uneven, \(\frac{d^2y}{dt^2}\) is even from ([\[Eq24\]](#Eq24){reference-type="ref" reference="Eq24"}) we obtain by \(t'\rightarrow-t\) \[\begin{split} \label{Eq25} & m\frac{d^2y}{dt'^2} + K_0 y + \frac{1}{2}k \exp({i\omega{t'}}) y + \frac{1}{2}k \exp({-i\omega{t'}}) y =\\ &\frac{BJ_0}{c}\cos\Omega{t'}. \end{split}\] Given case is equivalent to the task above formulated by the relations ([\[Eq19\]](#Eq19){reference-type="ref" reference="Eq19"})-([\[Eq21\]](#Eq21){reference-type="ref" reference="Eq21"}). Further, if \(y(t)\) is even, \(\frac{dy}{dt}\) is uneven, \(\frac{d^2y}{dt^2}\) is even from ([\[Eq24\]](#Eq24){reference-type="ref" reference="Eq24"}) we obtain by \(t'\rightarrow-t\) \[\begin{split} \label{Eq26} \eta\frac{dy}{dt}-\frac{BJ_0}{2c}\cos\Omega{t} = 0,\\ \end{split}\] that is the simple differential equation of the first order, the solution of which is evident. The homogeneous equations, corresponding to inhomogeneous equations ([\[Eq1\]](#Eq1){reference-type="ref" reference="Eq1"}) that is the equations \[\begin{split} \label{Eq27} &\frac{d^2y}{dt^2} + \frac{\eta}{m}\frac{dy}{dt} + \frac{K_0}{m} y + \frac{k}{m} \exp({i\omega{t}}) y = 0\\ &\frac{d^2y}{dt^2} + \frac{\eta}{m}\frac{dy}{dt} + \frac{K_0}{m} y + \frac{k}{m} \exp(-{i\omega{t}}) y = 0 \end{split}\] can be solved strictly. Let us designate \(\frac{\eta}{m} = a\) \(\frac{K_0}{m} = b\), \(\frac{k}{m} = c\), \(i\omega = \lambda\), \(-i\omega = \lambda '\). Then, according to the solutions are \[\begin{split} \label{Eq28} &y(t)^{[1]} = \exp(-\frac{\eta}{2m}t)[C^{[1]}_1 J_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})) + \\ &C^{[1]}_2 Y_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2}))] \end{split}\] for the first equation in ([\[Eq27\]](#Eq27){reference-type="ref" reference="Eq27"}) and \[\begin{split} \label{Eq29} &y(t)^{[2]} = \exp(-\frac{\eta}{2m}t)[C^{[2]}_1 J_{\nu{'}} (\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2})) + \\ &C^{[2]}_2 Y_{\nu{'}}(\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2}))], \end{split}\] for the second equation in ([\[Eq27\]](#Eq27){reference-type="ref" reference="Eq27"}). Here \[\begin{split} \label{Eq30} &J_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})), J_{\nu{'}} (\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2})),\\ &Y_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})), Y_{\nu{'}}(\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2})) \end{split}\] are Bessel functions of the first kind and of the second kind of index \[\begin{split} \label{Eq31} \nu = \frac{\sqrt{({\frac{\eta}{m}})^2-4 \frac{k}{m}}}{i\omega} \end{split}\] for the equation ([\[Eq30\]](#Eq30){reference-type="ref" reference="Eq30"}) and of index \[\begin{split} \label{Eq32} \nu{'} = \frac{\sqrt{({\frac{\eta}{m}})^2-4 \frac{k}{m}}}{-i\omega} \end{split}\] for the equation ([\[Eq31\]](#Eq31){reference-type="ref" reference="Eq31"}). Then the set of functions, satisfying to the relation \[\begin{split} \label{Eq33} y(t)^{[1]}\cap y(t)^{[2]} \end{split}\] will be the solution of the homogeneous equation, corresponding to the starting inhomogeneous equation ([\[Eq1\]](#Eq1){reference-type="ref" reference="Eq1"}). Given conclusion is correct, if \(y(t)^{[1]}\cap y(t)^{[2]} \neq \varnothing\). In particular, it is satisfied for the case \(C^{[1]}_1 = 0\), \(C^{[2]}_1 = 0\) and by \[\begin{split} \label{Eq34} &C^{[1]}_2 Y_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})) =\\ &C^{[2]}_2 Y_{-\nu{}}(\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2})), \end{split}\] that is, in the case \[\begin{split} \label{Eq35} &C^{[1]}_2 \frac {J_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})) \cos\pi\nu-J_{-\nu}(\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2}))}{\sin\pi\nu} = \\ &C^{[2]}_2 \frac {-J_{-\nu}(\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2})) \cos\pi\nu + J_{\nu}(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2}))}{\sin\pi\nu}, \end{split}\] Hence we obtain the conditions, by which the relation \[\begin{split} \label{Eq36} C^{[1]}_2\exp(-\frac{\eta}{2m}t) Y_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})) \end{split}\] will be the solution of the homogeneous damped Mathieu equation. They are \[\begin{split} \label{Eq37} &\frac{C^{[1]}_2 \cos\pi\nu-C^{[2]}_2}{\sin\pi\nu } = 0,\\ &\frac{C^{[1]}_2\cos\pi\nu-C^{[1]}_2}{\sin\pi\nu} = 0 \end{split}\] that is, by \[\begin{split} \label{Eq38} &\frac{\cos^2\pi\nu-1}{\sin\pi\nu} = 0,\\ &C^{[1]}_2 \neq 0. \end{split}\] Consequently, \(\nu \in Z\), that is \(\nu\) has to belong to the set of whole numbers. Therefore, the relation ([\[Eq36\]](#Eq36){reference-type="ref" reference="Eq36"}) is the solution of homogeneous damped Mathieu equation by \(\forall C^{[1]}_2 \neq 0 \in C\) and by \(\nu \in Z\). It means, that there are restrictions on possible values of parameters in the starting homogeneous damped Mathieu equation. They are the following \[\begin{split} \label{Eq39} \nu = (\frac{\sqrt{({\frac{\eta}{m}})^2-4 \frac{k}{m}}}{i\omega}) \in Z, \end{split}\] that is \[\begin{split} \label{Eq40} \nu = (\frac{\sqrt{{4 \frac{k}{m}-(\frac{\eta}{m}})^2}}{\omega}) \in Z \end{split}\] It is evident that by \(\eta = 0\) we obtain the solution of the homogeneous undamped Mathieu equation, that is the solution of the general Mathieu equation. To obtain the fundamental system of the solutions of the homogeneous damped Mathieu equation, we have to find the second linearly independent solution. It can be easily done, if to take into account, that \[\begin{split} \label{Eq41} &J_{\nu{'}} (\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2})) = J_{-\nu}(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})) = \\ &(-1)^{\nu} J_{\nu{}} (\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})) \end{split}\] for \(\forall\) \(\nu \in Z\). Since the replacement \(-i\omega\rightarrow i\omega\) is equivalent to \[\begin{split} \label{Eq42} J_{\nu{'}} (\frac{2\sqrt{\frac{K_0}{m}}}{-i\omega}\exp(\frac{-i\omega t}{2}))\rightarrow J_{-\nu}(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})), \end{split}\] we obtain, that by \(C^{[1]}_2 = 0\), \(C^{[2]}_2 = 0\) the expression ([\[Eq33\]](#Eq33){reference-type="ref" reference="Eq33"}) that is \[\begin{split} \label{Eq43} C^{[1]}_1 \exp(-\frac{\eta}{2m}t)J_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})) \end{split}\] will be the second linearly independent solution of the homogeneous damped Mathieu equation \(\forall C^{[1]}_{1} \neq 0 \in C\) and by \(\forall\nu \in Z\) at \(C^{[1]}_{1} = C^{[2]}_{1}\) for even \(\nu \in Z\) and at \(C^{[1]}_{1} =-C^{[2]}_{1}\) for uneven \(\nu \in Z\), that is at \(C^{[1]}_1 = C^{[2]}_1 sgn(-1)^{\nu}\). Its linear independence from the first solution follows from linear independence of Bessel functions on the first and the second kinds. At fixed \(C^{[1]}_1\) and \(C^{[1]}_2\) the relations ([\[Eq36\]](#Eq36){reference-type="ref" reference="Eq36"}) and ([\[Eq43\]](#Eq43){reference-type="ref" reference="Eq43"}) represent themselves the fundamental system of the solution of homogeneous damped Mathieu equation. So, the general solution of homogeneous damped Mathieu equation is \[\begin{split} \label{Eq44} &\exp(-\frac{\eta}{2m}t)[C^{[1]}_1 J_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2})) +\\ &C^{[1]}_2 Y_\nu(\frac{2\sqrt{\frac{K_0}{m}}}{i\omega}\exp(\frac{i\omega t}{2}))] \end{split}\] The theorem is proved. Thus, we have found the general solution of the homogeneous damped Mathieu equation in the essentially more simple form, allowing its practical using in many applications without numerical calculations, since Bessel functions are well known.

{'timestamp': '2014-01-22T02:10:12', 'yymm': '1401', 'arxiv_id': '1401.5348', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5348'}

# Introduction ## Non-associates Let \(\bbK\) be an algebraic number field with \(r=[\bbK:\bbQ]\). Let \(\al_{1}\), ..., \(\al_{n}\) lie in \(\bbK\) and put \(L(\bX) = \al_{1}X_{1} + \cdots + \al_{n}X_{n}\). The set \(\cM = \left\{ L(\bx): \bx \in \bbZ^{n} \right\}\) is a \(\bbZ\)-module contained in \(\bbK\). For every subfield \(\bbL\) of \(\bbK\), let \(\cM^{\bbL}\) consist of the elements \(\be\) of \(\cM\) such that for every \(\al \in \bbL\) there is a non-zero rational integer \(z\) with \(z \al \be \in \cM\). For our purposes here, we single out a particular subgroup of the group of units in \(\bbL\): let \(\cU_{\cM}^{\bbL}\) be the group of elements in \(\cO_{\cM}^{\bbL}\) of norm \(1\). Note that if \(\cM\) is a full module, then \(\cM^{\bbK}=\cM\). It is known that there are only finitely many pairwise non-associate elements with given norm in a full module \(\cM\) (see ). While some upper bounds for the number of such non-associates are known (see the result from cited below), it is not known how well these bounds reflect the actual behaviour of these numbers. It would not be unreasonable to suspect that this number depends on the field \(\bbK\). In fact, given a result of Schmidt on norm-form equations to be cited below, one might even believe that this number depends only on \(r\). However, we show here that this is not correct. In particular, we have the following result. ## Norm form equations For \(i=1, \ldots, r\), we let \(\sigma_{i}\) denote the isomorphic embeddings of \(\bbK\) into \(\bbC\) and write \(\al^{(i)}=\sigma_{i}(\al)\) for any \(\al \in \bbK\). With \(L(\bX) = \al_{1}X_{1} + \cdots + \al_{n}X_{n}\), as above, and \(L^{(i)}(\bX) = \al_{1}^{(i)}X_{1} + \cdots + \al_{n}^{(i)}X_{n}\) for \(i=1,\ldots, r\). We can write \(F(\bX)\) in the form \[\label{eq:altform} F(\bX) = a L^{(1)}(\bX) \cdots L^{(r)}(\bX).\] This definition was formulated by Schmidt in the early 1970's and he showed that there are non-zero rational numbers \(m\) such that \(F(\bX)=m\) has infinitely many solutions in \(\bbZ^{n}\) if and only if \(F\) is degenerate. Moreover, he was also able to show that even if \(F\) is degenerate, then there is a notion of a family of solutions such that there are only finitely many families of solutions of \(F(\bX)=m\). That this is a natural notion of a family of solutions is probably best seen by means of examples and so we invite the reader to consult those presented in and. By the end of the 1980's, Schmidt had proven his quantitative subspace theorem and used it to establish upper bounds that depend only on \(m\), \(n\) and \(r\) for the number of solutions of the norm form equation \(F(\bX)=m\) when \(F\) is non-degenerate. At that time, Schmidt posed to this author the question of what sort of bounds one could obtain for the number of families of solutions of degenerate norm forms. The author obtained a bound depending only on \(m,n\) and \(r\) for the number of full submodules in subfields of \(\bbK\) such that any solution of \(F(\bX)=m\) must lie in the union of these submodules. Győry has independently established this same result. The most recent results in this area are due to Evertse and Győry. Their results are much more general than the following, but their Theorem 1 implies that the solutions of \(F(\bX)=1\) lie in the union of at most \[\left( 2^{33} r^{2} \right)^{n(n+1)(2n+1)/3-2}\] full submodules of subfields of \(\bbK\). As was mentioned in the previous paragraph, Győry's work in this area has been much more general. He has generalised the concept of a family of solutions for the norm form setting to that of decomposable form equations. Decomposable forms include not only norm forms, but also discriminant forms, index forms, resultant forms, reducible binary forms and other kinds of forms as well. Moreover, Győry considers these decomposable form equations over number fields and, more generally, over finitely generated fields. Similar to the author's results to be cited in the next paragraph, in, Győry obtained explicit upper bounds for the number of families which depend on certain "indices" associated with the module. For full modules, the author could only establish bounds for the number of families of solutions which depended more closely on the given module. In particular, Lemma V.2 of states that if \(\cO_{\cM}\) is the ring of coefficients of \(\cM\) and \([\cU_{\bbK}: \cU_{\cM}]\) is the index of the unit group of \(\cO_{\cM}\) in the unit group of \(\bbK\), then the solutions of the norm form equation \(F(\bX) = m\) lie in the union of at most \[\label{eq:unit-based-bound} [\cU_{\bbK}: \cU_{\cM}] { \tau \left( \left| m \right| \right) }^{r}\] families, where \(\tau\) is the function which counts the number of positive divisors of a rational integer. In Theorem V.2 of, an upper bound in terms of the coefficients of \(\cM\) was obtained. For a full module, \(\cM\), the number of families of solutions of the associated norm-form equation \(F(\bX)=m\) is equal to the number of non-associates in \(\cM\) of norm \(m/a\). Hence, from Theorem [\[thm:countereg\]](#thm:countereg){reference-type="ref" reference="thm:countereg"}, we obtain the following Corollary. Despite Corollary [\[cor:countereg\]](#cor:countereg){reference-type="ref" reference="cor:countereg"}, for norm forms in three variables which do not arise from full modules, we are able to get a bound of the desired form. In fact, this is only possible because Corollary [\[cor:countereg\]](#cor:countereg){reference-type="ref" reference="cor:countereg"} is not true when \(\bbK\) is a quadratic field. The method of proof of Theorem [\[thm:family-bound\]](#thm:family-bound){reference-type="ref" reference="thm:family-bound"} fails when \(F(\bX)\) is a norm form in four (or more) variables satisfying analogous conditions. By Schmidt's Subspace Theorem, all the solutions of \(F(\bX)=1\) correspond to elements of certain three-dimensional subspaces of \(\bbQ^{4}\). If one of these subspaces gives rise to a full module of rank \(3\), then, from Corollary [\[cor:countereg\]](#cor:countereg){reference-type="ref" reference="cor:countereg"}, it is possible for there to be an arbitrarily large number of families of solutions. Several questions and directions for further investigation come to mind. It would be of considerable diophantine interest to determine the nature of the dependence of the number of families of solutions on \(F(\bX)\) or \(\cM\). From the proof of Theorem [\[thm:countereg\]](#thm:countereg){reference-type="ref" reference="thm:countereg"}, it is clear that some dependence on \([\cU_{\bbK}: \cU_{\cM}]\) as in [\[eq:unit-based-bound\]](#eq:unit-based-bound){reference-type="eqref" reference="eq:unit-based-bound"} is necessary. Under what circumstances is the number of families independent of the module \(\cM\)? Bombieri and Schmidt have shown that \(O(r)\) is the correct order of growth for the number of solutions of Thue equations. What is the correct order of growth in Theorem [\[thm:family-bound\]](#thm:family-bound){reference-type="ref" reference="thm:family-bound"}? # Preliminary Lemmas to the Proof of Theorem \(\ref{thm:countereg}\) The following is Exercise 5 on page 93 of. We include a proof for completeness. The next lemma is the key result in establishing Theorem [\[thm:countereg\]](#thm:countereg){reference-type="ref" reference="thm:countereg"}. Now we provide some results about the sorts of full modules that we will use to construct our examples. We start with our definition and notation for them. # Proof of Theorem \(\ref{thm:countereg}\) {#sect:3} Let \(\bbK\) be a number field with \(r=[\bbK:\bbQ] \geq 3\) and let \(\eta\) be a unit in \(\bbK\) of norm \(1\) which is not a root of unity. Put \(\epsilon=\eta^{n_{r}}\), where \(n_{r}\) is as in Lemma [\[lem:seq\]](#lem:seq){reference-type="ref" reference="lem:seq"}. Let \(\alpha\) be a primitive element of the extension \(\bbK/\bbQ(\epsilon)\) whose minimal polynomial over \(\bbZ[\epsilon]\) is monic (with \(\alpha=1\) if \(\bbK=\bbQ(\epsilon)\)). For any positive integer \(N\), let \(k_{1}=p_{r,1}, \ldots, k_{N}=p_{r,N}\) be the first \(N\) elements of a sequence of primes satisfying the conditions in Lemmas [\[lem:seq\]](#lem:seq){reference-type="ref" reference="lem:seq"} and [\[lem:order\]](#lem:order){reference-type="ref" reference="lem:order"} (with \(\cO=\cM_{1}\left( \epsilon, \alpha \right)\)--note that this is an order in \(\bbK\) by our conditions on \(\alpha\)). Put \(K_{N}=k_{1} \cdots k_{N}\). We are now ready to apply Lemma [\[lem:more-units\]](#lem:more-units){reference-type="ref" reference="lem:more-units"} to prove Theorem [\[thm:countereg\]](#thm:countereg){reference-type="ref" reference="thm:countereg"}. We let \(\cM^{(i)}=\cM_{k_{i}}(\epsilon,\alpha)\), so that \(\cO^{(i)}=\cO_{k_{i}}(\epsilon,\alpha)\) from Lemma [\[lem:coeff-ring\]](#lem:coeff-ring){reference-type="ref" reference="lem:coeff-ring"}(i). Notice that \(\cO^{(i)}\) is a proper subset of \(\cM^{(i)}\), so condition (a) of Lemma [\[lem:more-units\]](#lem:more-units){reference-type="ref" reference="lem:more-units"} holds. Also \(\epsilon \in \cM^{(i)}\) and \(\epsilon \not\in \cO^{(i)}\), so condition (b) of Lemma [\[lem:more-units\]](#lem:more-units){reference-type="ref" reference="lem:more-units"} holds. Recall from the statement of Lemma [\[lem:more-units\]](#lem:more-units){reference-type="ref" reference="lem:more-units"} that \(\ell_{i}\) is the number of distinct multiplicative cosets of the form \(\epsilon^{v}\cU^{(i)}\), where \(\cU^{(i)}\) is the group of units of norm \(1\) in \(\cO^{(i)}\). From Lemmas [\[lem:seq\]](#lem:seq){reference-type="ref" reference="lem:seq"}(ii) and [\[lem:order\]](#lem:order){reference-type="ref" reference="lem:order"}, we know that \(\ell_{i} | \left( g_{r} \left( k_{i} \right)/n_{r} \right)\), which are all pairwise relatively prime and that \(\ell_{i}>1\). Therefore, condition (c) of Lemma [\[lem:more-units\]](#lem:more-units){reference-type="ref" reference="lem:more-units"} holds. Finally, from Lemma [\[lem:coeff-ring\]](#lem:coeff-ring){reference-type="ref" reference="lem:coeff-ring"}(ii), condition (d) of Lemma [\[lem:more-units\]](#lem:more-units){reference-type="ref" reference="lem:more-units"} holds. Since all the conditions in Lemma [\[lem:more-units\]](#lem:more-units){reference-type="ref" reference="lem:more-units"} are satisfied, \(\cM_{K_{N}}(\epsilon,\alpha) = {\displaystyle \bigcap_{i=1}^{N}} \cM_{k_{i}}(\epsilon,\alpha)\) (equality holding by Lemma [\[lem:coeff-ring\]](#lem:coeff-ring){reference-type="ref" reference="lem:coeff-ring"}(ii)) has at least \(2^{N}\) units that are non-associates. Hence Theorem [\[thm:countereg\]](#thm:countereg){reference-type="ref" reference="thm:countereg"} holds. # Preliminary Lemmas to the Proof of Theorem \(\ref{thm:family-bound}\) In the case of our theorem, i.e., \(n=3\) and \(r = [\bbK:\bbQ] > 3\), Lemma [\[lem:1\]](#lem:1){reference-type="ref" reference="lem:1"} tells us that if \(\cM^{\bbL} = \cM\) then \(\bbL = \bbQ\). This information turns out to be crucial in what follows. We will use the heights \(H(\cdot)\), \(H^{*}(\cdot)\) and \(\cH(\cdot)\) defined on pages 201 and 204 of. Let \(L=\sum_{j=1}^{n} \al_{j}X_{j}\) be a linear form with coefficients in an algebraic number field \(\bbK\) of degree \(r\) over \(\bbQ\) and let \(a \in \bbQ^{*}\) be such that the norm form \[F(\bX) = a \cN_{\bbK/\bbQ} \left( L (\bX) \right) = a \prod_{i=1}^{r} \left( \sum_{j=1}^{n} \al_{j}^{(i)} X_{j} \right)\] has its coefficients in \(\bbZ\). Then the height of F, \(H^{*}(F)\), is defined by \[H^{*}(F) = |a| \prod_{i=1}^{r} \left( \sum_{j=1}^{n} \left| \al_{j}^{(i)} \right|^{2} \right)^{1/2}.\] According to Lemma 1 of, for the absolute height \(H(L)\) of the linear form \(L\), we have \[H^{*}(F) = {\rm cont}(F) H(L)^{r},\] where \({\rm cont}(F)\) denotes the greatest common divisor of the coefficients of \(F\). Two norm forms \(F\), \(G\) are called *equivalent*, which we denote by \(F \sim G\), if \(G(\bX)=F(B\bX)\) for some matrix \(B \in {\rm SL}(n,\bbZ)\). Now \(H^{*}(\cdot)\) is not an invariant under this equivalence, so we define an invariant height of a norm form \(F\), \(\cH(F)\), by \(\cH(F) := \min_{G \sim F} H^{*}(G)\), where the minimum is taken over all norm forms \(G\) equivalent to \(F\). To proceed, we now divide the solutions of ([\[eq:nf\]](#eq:nf){reference-type="ref" reference="eq:nf"}) into large and small solutions and "jack up the height" of the norm form \(F(\bX)\). The point of this last process, which will be explained shortly, is to replace \(F(\bX)\) by a finite number of other norm forms \(F_{j}(\bX)\) which are of sufficiently large height so that we can apply known diophantine techniques to obtain an upper bound on the number of solutions or, when it works for degenerate norm form equations, families of solutions of \(F_{j}(\bX)=1\). We create these new forms in such a way, via linear maps, that the number of solutions (or families of solutions) to the norm form equation \(F(\bX)=1\) is at most the sum of the number of solutions (or families of solutions) of each of the norm form equations \(F_{j}(\bX)=1\). Since we know the number of such norm forms, we can bound the number of solutions of \(F(\bX)=1\). For a prime \(p\), let \[A_{0} = \left( \begin{array}{cc} p & 0 \\ 0 & 1 \end{array} \right), \hspace{5mm} A_{j} = \left( \begin{array}{cc} 0 &-1 \\ p &-j \end{array} \right) \mbox{ for \(j=1, \ldots, p\).}\] For any \(n \geq 2\), we let \(E\) be the \((n-2) \times (n-2)\) identity matrix and consider the \(n \times n\) matrices \[B_{j} = \left( \begin{array}{cc} A_{j} & 0 \\ 0 & E \end{array} \right) \mbox{ for \(j=0, \ldots, p\).}\] We can use the linear maps induced by these matrices to express \[\bbZ^{n} = \bigcup_{j=0}^{p} B_{j} \bbZ^{n}.\] For \(j=0, \ldots, p\), we put \[F_{j}(\bX) = F \left( B_{j} \bX \right)\] and notice that we can express \(F_{j}(\bX)\) in the form \[F_{j}(\bX) = a_{j} \cN_{\bbK/\bbQ} \left( L_{j} (\bX) \right),\] where \(a_{j}\) is a non-zero rational number and \(L_{j}(\bX) = L \left( B_{j} \bX \right)\). Moreover, we shall assume that these \(F_{j}(\bX)\)'s are *reduced*, that is, \(\cH \left( F_{j} \right) = H^{*} \left( F_{j} \right)\). This idea comes from where it is noted that the number of solutions is unaffected by such an assumption. If we let \(p=125000^{3}+21\) (which is prime), then \[\label{eq:hbnd} H \left( L_{j} \right) = \cH { \left( F_{j} \right) }^{1/r} \geq p^{1/3} > 125000,\] for each \(j=0,\ldots,p\), by equations (5.3) and (5.5) of. In what follows, we shall drop the subscripts on the \(F_{j}\)'s and \(L_{j}\)'s in order to simplify our notation. It is also at this point where we introduce our definition of small and large solutions. To be able to estimate the number of small solutions we need the following lemma. Let us now turn to the large solutions. We first normalise the \(L^{(i)}(\bX)\)'s: put \(M_{i}(\bX) = { \left| L^{(i)} \right| }^{-1} L^{(i)}(\bX)\) for \(i=1,\ldots,r\). To count the families of large solutions, Schmidt used his quantitative subspace theorem. Here we shall use Evertse's refinement of this result. # Proof of Theorem \(\ref{thm:family-bound}\) Combining Lemmas [\[lem:small\]](#lem:small){reference-type="ref" reference="lem:small"} and [\[lem:big\]](#lem:big){reference-type="ref" reference="lem:big"}, along with our discussion of the relation between solutions of the \(F_{j}(\bX)=1\) and \(F(\bX)=1\), we see that the solutions of ([\[eq:nf\]](#eq:nf){reference-type="ref" reference="eq:nf"}) lie in the union of at most \[\left( 125 \, 000^{3} + 22 \right) \left( 2^{95} 3^{108} r^{9} + 2^{1570} 3^{1007} r^{3} \log^{2} r \right) < 1.1 \cdot 10^{965} r^{9}\] proper linear subspaces of \(\bbQ^{3}\), since \(r \geq 4\). The integer points in any proper linear subspace of \(\bbQ^{3}\) can be parametrised as \(\bx = T \by\) where \(T\) is a linear map from \(\bbQ^{2}\) into the subspace which sets up a 1-1 correspondence between \(\bbZ^{2}\) and the integer points in the subspace. Thus, restricting our attention to the integer points of the subspaces which arise, our norm form \(F(\bX)\) becomes \(F(T(\bY))\), which is a norm form in two variables with integer coefficients. We may also write them as \[F_{1}(\bY) = \cN_{\bbK/bQ} \left( \be_{1} Y_{1} + \be_{2} Y_{2} \right) = 1.\] We must now consider the fields \(K_{1} = \bbQ \left( \be_{1}/\be_{2} \right)\). If \(\left[ \bbK_{1}: \bbQ \right] \geq 3\), then the \(\bbZ\)-module generated by \(\be_{1}\) and \(\be_{2}\) is not a full module and so \(F_{1}(\bY)\) is a binary form of degree \(r\) which is not a power of a binary quadratic form. Thus, as a consequence of Theorem 1 of (take \(\epsilon\) sufficiently large), \(F_{1}(\bY)=\pm 1\) has at most \(5600r\) integer solutions. We need to consider \(F_{1}(\bY) = \pm 1\), since \(F_{1}(\bY)\) might be the power of a binary form of lower degree. So it remains to consider the case when \(\bbK_{1}\) is a quadratic field. But by Lemma [\[lem:quad\]](#lem:quad){reference-type="ref" reference="lem:quad"}, there are at most two families in this case. Thus for each subspace there are at most \(5600r\) families of solutions. Therefore, by our estimate above for the number of proper linear subspaces of \(\bbQ^{3}\) into which the solutions must fall, there are at most \(1.1 \cdot 10^{965} \cdot 5600r^{10} < 10^{969} r^{10}\) families of solutions to ([\[eq:nf\]](#eq:nf){reference-type="ref" reference="eq:nf"}).

{'timestamp': '2014-01-22T02:12:15', 'yymm': '1401', 'arxiv_id': '1401.5447', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5447'}

# Introduction While it is now abundantly clear that much of the mass in the extended outer stellar envelopes of galaxies (stellar halos hereafter) is stripped from dwarf galaxies as they tidally interact with the central galaxy ([@majewski-etal03]; [@bj05], hereafter ; @purcell-etal07; [@bell-etal08; @mcconnachie-etal09]; [@cooper-etal10], hereafter ; [@xue11; @ibata13]), a number of questions remain. Is *all* halo mass accreted (; [@rashkov12]), or are substantial fractions kicked up from the stellar disk (called *in situ*; [@kazantzidis08; @zolotov09; @font-etal11]) or formed within satellites after they have been accreted ([@tissera13]; Valluri et al., in prep.)? How much variation is expected from halo to halo (, [@bell-etal08; @font-etal11])? Many studies focus on these issues by comparing observations of stellar halos with models of stellar halo formation in a cosmological context. In most cases, cognizant that much of the stellar halo is accreted, comparisons are made to models in which the stellar halo is constituted *only* of accreted material: this simplifies interpretation, and any major discrepancies between the observations and such models could signal that this assumption is incorrect, giving insight into other possible modes of halo formation. A generic practical problem that one encounters when creating a theoretical numerical model of stellar halos is resolution. Although stellar halos are potentially rich with signposts of the hierarchical galaxy formation process, they contain a very small fraction of the total stellar mass in a galaxy. For example, the stellar halos of the Milky Way and similar-mass galaxies account for only \(\approx 1\)--\(10\%\) of their stellar content (; [@bailin11; @ibata13]). In order to resolve the tidal streams that constitute the halo and provide observational tests of the hierarchical merging paradigm, hundreds of thousands of particles must be used within the halo itself. Simulating the entire galaxy self-consistently at this resolution would then require hundreds of millions of particles, a task that would require tens of millions of CPU hours per galaxy with current algorithms and hardware. The problem is exacerbated by the stochastic nature of merger histories, which result in factors of several galaxy-to-galaxy variation in stellar halo properties even at a given galaxy mass and requires performing a significant number of these simulations in order to make robust predictions. The common solution to this problem is to not simulate the entire galaxy self-consistently. Pure \(N\)-body simulations are much faster than full hydrodynamic simulations at a given resolution, at the expense of not including any non-gravitational processes (e.g. ; [@libeskind-etal11], hereafter ). Furthermore, if the main body of the galaxy is replaced by an analytic potential, the number of particles gets dramatically reduced and it becomes feasible to simulate the halo at high resolution in a reasonable length of time . In these pure \(N\)-body methods (), dark matter (DM) particles must be labelled ("painted") to represent the stellar component in lieu of having a self-consistent method of generating stars. Other tactics that have been taken are to use lower resolution self-consistent simulations where the internal structure of the halo is poorly resolved (in low resolution simulations, unresolved physical processes may also have a significant impact on the derived halo properties, although in the case of [@font-etal11] the authors have tested that their conclusions are robust to a factor of \(2\) change in spatial resolution), or to use semi-analytic prescriptions that have explicitly no internal structure but only predict the total amount of halo material. Yet, there are important differences between the expectations of these different approaches. The difference that inspired this work was the factor-of-two difference in the degree of substructure predicted by and . As described by and, these models predict an amount of substructure (measured using the RMS of the model around a smooth halo profile, divided by the total number of stars) different by a factor of two or more from each other, in the sense that the models have considerably more substructure than those of (@schlaufman12 find a similar difference between the and @rashkov12 models). The models also have more substructure than the observations; interpreted this as a model--data discrepancy signaling the need for halo stars to form *in situ*; we interpret this as a *model--model* discrepancy signaling the need to better understand why two seemingly reasonable models should disagree so significantly. One potential point of distinction between the models was that, while had an analytic potential (and therefore would have been perhaps more likely to have a more structured halo), it included the potential from a *disk*. As discussed earlier, the inclusion of baryons in simulations of galaxy formation adds a degree of complexity and computational cost that both reduces the general applicability of the simulations and prohibits the construction of samples of stellar halos that adequately span the range of possible assembly histories. In this work, we use a hydrodynamical simulation from the McMaster Unbiased Galaxy Simulations project, which is of sufficiently high resolution that the satellites whose accretion we wish to follow are well resolved, to explore two crucial aspects of the relationship of baryons to dark matter relevant to stellar halo formation. Firstly, the dissipative formation of a disk changes the potential of the galaxy, enhancing the strength of the tidal field and affecting the orbits of halo stars. Secondly, we wish to explore the importance of the practice of "painting" stars onto dark matter particles: inasmuch as dark matter particles have not suffered dissipation, even the most bound dark matter particles have orbits that are very likely to be different from realistic stellar orbits, and this would affect the properties of the resulting stellar halos. The plan of this paper is as follows. In Section 2, we give an overview of the assumptions that have been used in previous work and the models that we will use to test their effects. In Section 3, we provide full details of how the simulations and models are generated. Section 4 contains the results from the different models, and in Section 5 we discuss what these results imply about the influence of the standard assumptions on halo models. Finally, our conclusions are presented in Section 6. # Overview There are two key assumptions that previous halo models have often adopted in order to make the problem tractable, which we label "painting" and "dark matter dynamics." *Painting* : [\[overview-painting\]]{#overview-painting label="overview-painting"} All high resolution stellar halo models consist of pure \(N\)-body simulations that contain only DM particles. In order to predict the properties of the luminous stellar halo, the authors "paint" stellar mass onto certain DM particles, and then measure the properties and structure of these painted particles. The methods used to paint vary: resolve each contributing subhalo into \(10^5\) DM particles, and then paint the most bound particles such that the luminosity follows a King profile; has equal-mass DM particles that are painted equally if they are sufficiently deep in the potential well; use a sophisticated semi-analytic galaxy formation model to determine the expected amount and distribution of star formation within each contributing subhalo and paint stellar masses onto the DM particles so as to contain the appropriate star formation history; and paint the most-bound DM particles within each subhalo equally, but with a stellar mass that varies between subhalos. *Dark Matter Dynamics* : When galaxies form, baryons cool and collapse into a centrifugally-rotating disk, whose morphology can then be altered by further accreted material and interactions with other galaxies; these processes do not occur in pure DM simulations, which do not have the capability to radiatively cool. The gravitational potential in which the stars that constitute the stellar halo orbit is therefore different in the real universe than in a pure DM simulation: it is more concentrated, and is flattened in the inner regions due to the disk. Different groups have taken different approaches to account for this effect: and use pure \(N\)-body cosmological simulations and neglect any changes in the potential due to baryonic physics; grow an analytic disk potential inside an analytic growing halo potential; and finally, a particularly interesting approach is that of , who compare a full hydrodynamic cosmological simulation to the identical DM-only simulation using the same initial conditions. They find that, when a gravitational-potential-based painting scheme is adopted (see above), the radial distribution of the stellar halo in both simulations is identical, and therefore argue that the effect of the baryonic physics can be taken entirely into account by the appropriate painting scheme. These two types of assumptions have remained largely untested, and their effect on the final properties of the predicted stellar halos are therefore unknown. Our goal is to use a set of control simulations in which we vary either the method by which we determine where "stars" lie in the simulation volume or the potential in which the particles orbit. To do this, we compare a full Smoothed Particle Hydrodynamics (SPH) simulation of galaxy formation from the McMaster Unbiased Galaxy Simulations with a DM-only simulation of the same initial conditions. We analyze 4 different models for the stellar halo that is formed: 1. SPH-STARS: the stars that form self-consistently in the SPH simulation that are accreted from satellites; 2. SPH-PAINTED: DM particles in the SPH simulation that are painted to match the mean stellar mass-DM mass relation of satellites in the simulation; 3. SPH-EXACT: DM particles in the SPH simulation that are painted to match the stellar mass of each individual satellite that contributes to the halo; and 4. DM-PAINTED: painted DM particles in the DM-only simulation. Full details of how each of these models is constructed is given in Section [3.3](#sec:modelhalos){reference-type="ref" reference="sec:modelhalos"}. Our painting schemes are calibrated using the luminous satellites within the SPH simulation. This allows us to directly compare the painted stars with those that form self-consistently in the simulations, since the same objects should have the same stellar content. The comparison between SPH-STARS and SPH-PAINTED halos isolates the effect of using painted DM particles instead of stars (and the SPH-EXACT halo can be used to determine what aspect of the painting scheme is responsible for any differences), while the comparison between the SPH-PAINTED and DM-PAINTED halos isolates the effect of the baryonic contribution to the gravitational potential. We emphasize here that although we will use the SPH-STARS halo as the reference model, this is not because we think it is a good approximation to stellar halos in the real universe. Galaxy formation simulations at the resolution of MUGS generically form too many stars by a factor of \(\sim 2\) , contain too large a fraction of their stellar mass in their spheroid (for example, MUGS galaxies have a mean bulge fraction of \(73\%\) compared to an observed value of \(\sim 40\%\) for comparable-luminosity observed galaxies; [@tasca11]), and cannot resolve the majority of streams that constitute the halo. However, the relative comparison between the models is valid: if we start with the same amount of stellar material in the same satellites, then the stellar halos should be similar if the assumptions we are testing are appropriate. # Simulations ## MUGS The simulations we analyze are (1) g15784 from the McMaster Unbiased Galaxy Simulations , and (2) a simulation with the same initial conditions but evolved purely using collisionless \(N\)-body dynamics, i.e. only with DM. These two simulations provide everything we need to cleanly measure the importance of the assumptions we are testing. The total mass of the galaxy within the virial radius (i.e. including subhalos) is \(1.4 \times 10^{12}~\Msun\) at \(z=0\), of which \(1.1 \times 10^{11}~\Msun\) is in the form of stars and \(1.0 \times 10^{11}~\Msun\) is in the form of gas. The simulation uses a \(\Lambda\)CDM cosmology with \(H_0=73~\mathrm{km~s^{-1}~Mpc^{-1}}\), \(\Omega_m=0.24\), \(\Omega_{\Lambda}=0.76\), \(\Omega_b=0.04\), and \(\sigma_8=0.76\). The DM particle mass is \(1.1 \times 10^{6}~\Msun\) in the SPH simulation and \(1.3\times 10^6~\Msun\) in the DM-only simulation [^1] the initial gas particle mass is \(2.2 \times 10^5~\Msun\), and the initial star particle mass is \(6.3 \times 10^4~\Msun\). The gravitational softening is \(312.5\) pc. The visible galaxy that forms at \(z=0\) has a prominent disk that can be traced to \(10\) kpc in gas and \(20\) kpc in stars, with a scale length of \(1.7\) kpc and has a bulge-to-total ratio of \(0.48\) within \(25\) kpc. It is therefore a good analog for an early-type disk galaxy. 87 snapshots of the SPH simulation were saved, while 62 snapshots of the DM simulation were saved. DM snapshots were spaced equally in time with \(\approx215\) Myr between snapshots; SPH snapshots exist at all of these times plus some additional times corresponding to convenient redshifts. ## Merger Trees Halos were found in the simulations using Amiga's Halo Finder , which generates a spatially-adaptive mesh on which the density field is measured, and then structures are found in the density field corresponding to a virialized spherical overdensity criterion. Because of the adaptive mesh, structures can be found on different scales, and so AHF generates a hierarchy of halos, subhalos, subsubhalos, etc. Finally, energetically unbound particles are removed from the particle lists corresponding to each halo. The output contains the list of particles associated with each halo in each snapshot. To follow the evolution of individual halos, we must associate dark matter halos in successive snapshots. To do this, we compare the list of particles IDs between each pair of halos, and assign halo \(h_i\) in snapshot \(s_i\) to be the progenitor of halo \(h_{i+1}\) in snapshot \(s_{i+1}\) if \(h_i\) contributes more particles to \(h_{i+1}\) than any other halo in \(s_i\) does. We also must be careful with substructure: AHF can assign particles to multiple halos, and in particular, the particles of a subhalo are usually also members of the parent halo. Therefore, if two subhalos merge within a parent halo, the parent halo may appear to be the progenitor of the merger product, because it contributes all of the particles in both subhalos. We therefore assign the progenitor to be the subhalo whose contribution to \(h_{i+1}\) most closely matches the actual number of particles in \(h_{i+1}\) in these cases. Linking together each halo with its successor results in a track, which we consider to be the evolution of an individual object. The merger tree "trunk" is the track that results in the parent halo at \(z=0\). A list of the maximum mass each track achieves and the snapshot at which that mass is reached is recorded and used for the particle painting (Section [3.3.2](#sec:paintingmethod){reference-type="ref" reference="sec:paintingmethod"}). ## Model Halos {#sec:modelhalos} ### Accreted Stars The SPH-STARS halo model consists of all accreted stars in the full SPH simulation. We define accreted stars as those that are not within the AHF particle list (i.e., outside of the virial radius) of the parent halo trunk at the first simulation output in which they appear, but which appear within the particle list of the parent halo at \(z=0\). Note that this neglects stars that form in satellites after they have been accreted into the parent halo, but which are later stripped ([@tissera13]; Valluri et al., in prep.). Although it is theoretically possible for a star particle to form in a satellite outside of the trunk but fall into the parent halo before the next output, outputs are placed closely enough together in time (at most \(215\) Myr) that the number of such particles should be small. Each star particle is born with mass \(6.3 \times 10^4~\Msun\), and then loses mass over time due to stellar evolution. The final stellar halo in the SPH-STARS model contains \(3.49 \times 10^{10}~\Msun\) in \(799248\) particles within \(1.1\) times the virial radius. Because some observational analyses specifically avoid regions with known satellites, we also construct a model halo where we exclude all particles contained in bound substructures that are found by AHF. This "NOSUBS" halo model contains \(1.78 \times 10^{10}~\Msun\) in \(409645\) particles. ### Painting Methods {#sec:paintingmethod} All halo models aside from SPH-STARS consist of dark matter particles (in either the DM-only or SPH simulation) that have been assigned a stellar mass. This stellar mass is assigned to merger tree tracks at their maximum-mass snapshot. #### SPH-PAINTED The SPH-PAINTED model is constructed in the full SPH simulation from DM particles that have been painted. In the SPH simulation, we can measure the actual stellar mass fraction for DM halos at the snapshot where they reach their maximum mass (Figure [\[fig:mstarcalib\]](#fig:mstarcalib){reference-type="ref" reference="fig:mstarcalib"}). The following relationship provides a good fit to the majority of halos: \[\label{eq:mpaint} M_* = 4.5 \times 10^6~\Msun \left(\frac{M_{\mathrm{DM}}}{10^9~\Msun}\right)^{1.7}.\] It is possible in the real universe to use the observed satellite galaxy luminosity function and the simulated dark matter subhalo mass function to estimate the stellar mass fraction satellites must have at infall (which usually corresponds to their time of maximum mass). Studies that do this require there to be a much steeper relationship with lower normalization than what is seen in equation ([\[eq:mpaint\]](#eq:mpaint){reference-type="ref" reference="eq:mpaint"}), with power law indices ranging from \(2.5\) to \(3\) and normalizations at \(10^9~\Msun\) ranging from \(2 \times 10^3\) to \(2 \times 10^4~\Msun\). The SPH simulations, however, produce many more stars in the small objects as a consequence of their well documented tendency to overcool. In order to compare the SPH-STARS and SPH-PAINTED model halos, and therefore determine the effect that particle tagging has on the predicted properties of halos, we must make the progenitors in these models as similar as possible. We therefore match the stellar mass of the progenitor to what forms in the hydrodnamic simulation rather than the stellar content of real galaxies. If the SPH simulations could perfectly reproduce satellite galaxies, then these calibrations would be identical. Equation ([\[eq:mpaint\]](#eq:mpaint){reference-type="ref" reference="eq:mpaint"}) is therefore the correct stellar mass fraction to adopt for this purpose. There are a minority of AHF groups (\(23\%\)) with an unusually high stellar mass content, with stellar masses greater than or equal to their dark masses. These objects are usually subgroups of larger galaxies, and turn out to be stellar concentrations (e.g., star clusters and spiral arms) that are not independent contributors to the merger history of the main galaxy and should not be painted separately. However, they are assigned a stellar mass that is commensurate with their relatively small dark mass, and are therefore also small and do not contribute significantly to the model halo. We paint the most bound[^2] \(1\%\) of the particles in the halos at the time of maximum mass, and divide the stellar mass evenly amongst these particles. We tested painting different fractions of most bound particles, and found little qualitative difference in the properties of the resulting stellar halos for values less than \(\sim 10\%\). Our approach is similar to what adopted to generate the halo analyzed by. The final SPH-PAINTED halo model contains \(2.25 \times 10^{10}~\Msun\) in \(5179\) particles, while the NOSUBS version contains \(1.58 \times 10^{10}~\Msun\) in \(4043\) particles. #### DM-PAINTED The DM-PAINTED model is constructed in the DM-only simulation from painted particles. In order to compare the DM-PAINTED and SPH-PAINTED model halos, and therefore determine the effect that non-gravitational physics has on the predicted properties of stellar halos, we must use the identical painting scheme. We therefore also paint the most bound \(1\%\) of the particles in the DM halos at the time of maximum mass with a total stellar mass from equation ([\[eq:mpaint\]](#eq:mpaint){reference-type="ref" reference="eq:mpaint"}), evenly divided amongst the particles. The AHF groups that are bound by their stellar mass, rather than their DM, are absent in the DM-only simulation, and therefore do not contribute at all to the model halo. The DM-PAINTED halo model contains \(1.45 \times 10^{10}~\Msun\) in \(4405\) particles, while the NOSUBS version contains \(1.22 \times 10^{10}~\Msun\) in \(3862\) particles. #### SPH-EXACT The SPH-EXACT model, like the SPH-PAINTED model, is constructed in the SPH simulation from painted DM particles. However, rather than using equation ([\[eq:mpaint\]](#eq:mpaint){reference-type="ref" reference="eq:mpaint"}) to determine the total stellar mass of each progenitor, the exact stellar mass for that progenitor in the SPH simulation is used, as shown in Figure [\[fig:mstarcalib\]](#fig:mstarcalib){reference-type="ref" reference="fig:mstarcalib"}. This allows us to ascertain whether any differences between SPH-STARS and SPH-PAINTED are due to the particular method of assigning stellar masses to DM halos, or whether they are generic to the enterprise of painting DM particles. The SPH-EXACT halo model contains \(6.42\times 10^{10}~\Msun\) in \(5179\) particles, while the NOSUBS version contains \(4.49 \times 10^{10}~\Msun\) in \(4043\) particles. The choice of painted particles are identical to the SPH-PAINTED model, but they are assigned different stellar masses. The total mass is higher than in SPH-PAINTED because the subhalos that do not fall on the best fit relation of Figure [\[fig:mstarcalib\]](#fig:mstarcalib){reference-type="ref" reference="fig:mstarcalib"} scatter systematically high. # Halo Structure ## Overview Two-dimensional projected maps of the density of the stellar halo models are shown in Figure [\[fig:densmap\]](#fig:densmap){reference-type="ref" reference="fig:densmap"}, which includes all accreted or painted particles, and Figure [\[fig:densmap_noss\]](#fig:densmap_noss){reference-type="ref" reference="fig:densmap_noss"}, which excludes those contained in bound substructure. The pixel size is adaptively expanded from a minimum of \(5\) kpc per side until there are at least 5 particles per pixel, so the signal-to-noise per pixel is approximately equal in the low density regions. We first note that the models unambiguously trace the evolution of the same galaxy: the massive satellites are recognizable in each model at similar locations. It is therefore valid to directly compare the quantitative structure measurements in the different models and be confident that the differences are due to differences in the assumptions of the models, not due to different evolutionary histories. Secondly, we note that the models show systematic differences. In particular, it is clear that the global concentration and shape of the stellar halos are different. There also appears to be more structure in the SPH-STARS model than in the SPH-PAINTED model, and yet more structure to the DM-PAINTED halo. We will quantify these differences below, which can be interpreted as an estimate of the magnitude of the systematic errors introduced by the assumptions going into the halo models. ## Concentration {#sec:concentration} To measure the different concentrations of the stellar halo models, we plot their cumulative stellar mass profiles in Figure [\[fig:masscprof\]](#fig:masscprof){reference-type="ref" reference="fig:masscprof"}. In the top panel, we include all star particles, while we focus on the bottom panel, where the bound subhalos have been removed. To assess the effect of painting, we compare the SPH-STARS and SPH-PAINTED models. The model containing accreted stars is significantly more centrally concentrated than the painted DM particles in the same simulation---for example, the half-mass radius is more than \(6\) times smaller. Because the SPH-STARS model has orders of magnitude more particles, resolution could conceivably be an issue. To assess the impact of the particle number, we have randomly sampled \(5179\) particles out of the SPH-STARS model to form a new "SPH-STARS (LOWSAMP)" model that has the same number of particles as the SPH-PAINTED model; this is shown in green, and has an identical radial distribution as the full SPH-STARS model. We therefore conclude that the assumption that painted DM particles follow the same distribution as self-consistently formed stars introduces a large systematic error in the overall concentration of the halo. In the bottom panel, the cumulative mass profiles of power law density distributions \(\rho \propto r^{\alpha}\) have been overlaid for various values of \(\alpha\). Steep slopes of \(\alpha <-3\) appear convex in this plot, while shallow slopes of \(\alpha >-3\) appear concave. The SPH-STARS halo is well described by a power law with a slope of \(\alpha \sim-3.5\), while the SPH-PAINTED halo transitions from a relatively shallow \(\alpha \sim-2\) in the inner regions to a much steeper \(\alpha \sim-4\) slope in the outer region. This general behavior agrees very well with what has been found in the literature. find a halo density profile that transitions from a shallow \(\alpha \sim-1\) at small radius to \(\alpha \sim-3.5\) at large radius, similar to what is seen in the SPH-PAINTED model, which has very similar assumptions. The accreted stars in have a power law slope of \(\sim-3\) at most radii, steepening to \(-3.5\) at large radii, which is not dissimilar to what we find in the SPH-STARS model, and while never quantified the density profile of the self-consistently-formed accreted stars in their hydrodynamic simulation, the cumulative mass profile of such stars in their figure 1 is an excellent match to the analogous SPH-STARS model in our Figure [\[fig:masscprof\]](#fig:masscprof){reference-type="ref" reference="fig:masscprof"}. If this difference is due to assigning an incorrect stellar mass to the accreted halos, then the SPH-EXACT halo should mirror SPH-STARS; if it is due to the choice of DM particles instead of stars, then it should mirror SPH-PAINTED. Figure [\[fig:masscprof\]](#fig:masscprof){reference-type="ref" reference="fig:masscprof"} shows that it is much more similar to SPH-PAINTED; it has the same overall profile shape, and a much more similar half-mass radius. However, it is undoubtedly more concentrated than SPH-PAINTED (the half-mass radius is a factor of \(2\) smaller). This indicates that halos that scatter high in Figure [\[fig:mstarcalib\]](#fig:mstarcalib){reference-type="ref" reference="fig:mstarcalib"} end up at systematically smaller radii, but that the more dominant effect is that painted DM particles are less concentrated than accreted stars. To assess the impact of the potential, we compare the SPH-PAINTED and DM-PAINTED models. We find that they have similar functional forms, but that the SPH-PAINTED model is significantly more concentrated; for example, the half-mass radius is \(1.7\) times smaller. Therefore, the baryonic contribution to the potential, which is itself more centrally concentrated than the DM, leads to a more centrally concentrated stellar halo. ## Shape {#sec:results-shape} Another difference between the models is their global sphericity. Figure [\[fig:axratplot\]](#fig:axratplot){reference-type="ref" reference="fig:axratplot"} shows the shape of the stellar distribution, which has been calculated using the 2nd moment tensor of the stellar mass in an iteratively-defined ellipsoidal shell of width \(25\%\) of the quoted radius. Both the SPH-STARS and SPH-PAINTED halos are somewhat oblate, with \(b/a \sim 0.8\)--\(1\) and a total flattening rising from \(c/a \sim 0.5\) in the inner regions up to \(0.8\) at the virial radius. On the other hand, model DM-PAINTED, which contains no disk, is very strongly prolate, with \(b/a \approx c/a \sim 0.4\)--\(0.7\) depending where it is measured. This is not surprising, since the dark matter halos of simulations with disks are strongly modified by the presence of the disk, becoming less flattened and more oblate, relative to the more flattened prolate dark matter halos that predominate in pure DM cosmological simulations. ## Substructure A key prediction of stellar halo models is the degree of substructure. A rough measurement of this is the variation in the stellar mass density within a shell of a given radius; this is similar to the "sigma/total" measurement used by. Formally, we divide the virial region of the halo into (initially) spherical shells, and then subdivide each shell into angular sectors of equal volume. The divisions between these sectors are spaced equally in azimuthal angle \(\phi\) and in the cosine of the polar angle \(\theta\). We use \(N_{\phi}=4\) azimuthal divisions and \(N_{\theta}=4\) polar divisions. This probes different physical scales at different radii, and therefore one should not compare the quantitative measurements between radial bins, but rather compare different models at the same radius. We compute the mean stellar mass density of each model within the entire shell , and the root mean square (rms) of the sector-to-sector variation, \(\sigma_{\rho}\). There is some contribution due purely to shot noise from the finite number of particles, \(\sigma_{\mathrm{shot}}\), the magnitude of which can be determined by noting that the total mass \(M\) within a sector is equal to the number of particles \(N\) times their mean mass : \[M = N \meanm\] \[\sigma^2_{\mathrm{shot}} = \left(\frac{\df M}{\df N}\right)^2 \sigma^2_N + \left(\frac{\df M}{\df\meanm}\right)^2 \sigma^2_{\meanm}\] \[= \meanm^2 N + N^2 \left(\frac{\sigma^2_m}{N}\right)\] \[= \meanm^2 N + N \left( \left<m^2\right>-\meanm^2 \right)\] \[= N \left<m^2\right>.\] Technically, this derivation assumes that the mass per sector within each radial bin is independent, while in reality there is an additional constraint that the sum of the masses of the sectors must equal the mass in the shell. However, with \(16\) sectors, the reduction of one degree of freedom only changes the shot noise by \(\approx 3\%\). We have verified the accuracy of this analytic expression using Monte Carlo experiments. The intrinsic sector-to-sector dispersion \(\sigma_{\rho}\) is then the measured rms minus the shot noise \(\sigma_{\mathrm{shot}}\), in quadrature. An additional complication is that the shapes of the halos are different. As noted by, densities at a given radius can vary by \(10\)--\(50\%\) due to the ellipticity of the density distribution, which could dominate the sector-to-sector dispersion if not taken into account. We therefore calculate the shape of the density distribution in a shell of geometric mean radius \(30\) kpc and width \(20\) kpc, using the method described in Section [4.3](#sec:results-shape){reference-type="ref" reference="sec:results-shape"}, and use these principal axes to define ellipsoidal shells in which to determine \(\sigma_{\rho}\). These shells have the same width and geometric mean radius as the corresponding spherical shells, and therefore the same volume. We also scale the particle coordinates in the principal axis frame by the lengths of the principal axes before we determine which angular sector it belongs to; this ensures that the sectors all have equal volumes regardless of the shape of the ellipsoid. The amount of structure seen in each halo model is shown in Figure [\[fig:sigma\]](#fig:sigma){reference-type="ref" reference="fig:sigma"}. The lefthand plots include all stellar mass, while the bound substructures have been removed in the righthand plots. The top plots show the dispersion within spherical shells, while the bottom plots use the ellipsoidal shells discussed above. The lefthand plots make it apparent that the different models sometimes predict dramatically different amounts of substructure at the same radius; however, there is no clear systematic pattern to the differences. Much of the substructure in these plots is due to the distinct satellite galaxies in the simulation rather than the diffuse halo, which are usually excluded from observational studies. We therefore focus on the righthand plots, in which bound substructures have been removed, and particularly on the bottom-right panel, where the ellipsoidal shape of the halo has been taken into account. The number of particles could impact the amount of structure, and certainly the error bars, complicating the comparison between the SPH-STARS and SPH-PAINTED halos. We therefore first compare the SPH-STARS (black) and SPH-STARS (LOWSAMP) (green) halos and find that the error bars are indeed significantly larger at lower resolution, but that the results always agree to within the error bars, giving us confidence that the error bars accurately portray the uncertainty in the measurement. Although it appears that the subsampled halo is systematically less structured, this is a coincidence of the randomly-sampled particles that constitute the LOWSAMP halo; with different random seeds, the LOWSAMP halo scatters both positive and negative around the full resolution halo, with a dispersion comparable to the quoted error bar. We therefore conclude that differences between halo models that are larger than the error bars are real discrepancies between the predictions. We assess the impact of using painted dark matter particles instead of stars by comparing SPH-STARS (black) to SPH-PAINTED (red). SPH-PAINTED is systematically less structured at all radii when using our preferred ellipsoidal bins, typically by a factor of \(2\). There is little difference between the dispersion measured in spherical vs. ellipsoidal shells, which is consistent with the similar global shapes of these halos (Figure [\[fig:axratplot\]](#fig:axratplot){reference-type="ref" reference="fig:axratplot"}). The SPH-EXACT halo (magenta) is indistinguishable from the SPH-PAINTED halo, indicating that the difference in the amount of substructure is entirely due to the use of painted DM particles. The impact of the different potential in the baryonic simulations can be seen by comparing SPH-PAINTED (red) to DM-PAINTED (blue). The dark matter-only simulation is systematically more structured at all radii, in this case by factors of typically \(3\), although as high as \(7\) at intermediate radii. Although the global halo shapes are different enough that they could introduce a significant dispersion in a structureless halo, the same discrepancy is seen when using spherical shells at almost all radii, allowing us to be confident that this is not an artifact of the more flattened halo in the dark matter-only simulation. # Discussion We have tested the effects of two common assumptions used in stellar halo models: (1) that stellar mass can be painted onto DM particles, and (2) that the baryonic changes in the potential can be ignored. We discuss below the effects of each assumption. ## Painting When we compare the SPH-STARS and SPH-PAINTED halo models, we find that SPH-PAINTED is less concentrated and less structured than SPH-STARS. Both models put, on average, the same stellar mass into the same progenitor objects and evolve them in the same potential, so it is surprising that there is such a large difference. This difference is entirely due to (a) assuming that the scatter in the stellar-DM mass relation does not correlate with the fate of the accreted halo, and (b) assuming that DM particles deep in the potential well of a subhalo (note that we only paint the most bound \(1\%\) of the DM particles) evolve similarly to the star particles, which are also found deep in the potential well. The SPH-EXACT halo, where the scatter is not an issue, is slightly more concentrated than SPH-PAINTED, indicating there is a slight tendency for halos with high stellar masses to wind up in the inner parts of the halo, but that this effect does not dominate the overall radial profile. Moreover, SPH-EXACT and SPH-PAINTED have identical substructure, indicating that the main reason that the SPH-STARS and SPH-PAINTED models differ is because DM particles deep in the potential well are different from stars deep in the potential well. Although these particles are co-located, the kinematics of the populations are not identical: the DM particles are dynamically hotter, having orbits that take them further from the center of the subhalo (a consequence of the more extended nature of the DM component). We postulate that this difference causes the DM particles to be stripped earlier than stars during the subhalo's accretion and orbit around the parent galaxy, and results in a less concentrated halo. The higher velocity dispersion of the DM also means the stripped debris is less coherent, resulting in less substructure. In both cases the debris orbits in the same potential, and it is therefore not surprising that painting does not affect the shape of the predicted stellar halo. It is important to note that we have formally only tested one particular painting scheme, and this scheme is less sophisticated than many that are used. Therefore, we must be careful about what lessons are generalizable. Firstly, we note that it is a generic property of galaxy formation physics that baryons are more concentrated than DM within subhalos, and as a direct consequence the DM particles at the same radii as star particles have different kinematic properties such as velocity dispersions and angular momenta. We therefore expect these results to generalize to any painting scheme that does not explicitly guarantee that the painted DM particles share not only the same spatial distribution as the stars within satellite galaxies, but also the same kinematic distribution. Secondly, it appears that the systematic difference that painting induces in the halo concentration can completely explain discrepancies in the literature between the radial density profiles in the models of , which used painting, and the accreted stellar halo in the hydro simulations of , which did not. advocate one particular painting scheme where DM particles are chosen to lie within a given depth of the subhalo potential well, and demonstrate that the resulting painted stellar halo model has the same concentration as the accreted stars in their hydro simulation. This painting scheme requires higher resolution than present in our simulations, so we cannot directly test it, but we find the match between the concentration of the stars and painted DM particles very encouraging; testing to see if the painted DM halo has the same degree of substructure as the star particles is another critical test that we would strongly advocate. However, there is one important aspect of this scheme that may compromise its use for stellar halo models: the scheme was calibrated so that the radial distribution of the diffuse halo, i.e. the stripped satellites, matched the SPH stars[^3] without regards to the properties of the satellites themselves. There is therefore no guarantee that the stellar masses (and consequently metallicities; [@tremonti-etal04]) or radial profiles of the satellites are correct, and in fact, scaling arguments suggest that the stellar masses, in particular, may not scale correctly with total DM halo mass. In other words, it may be that are building a realistic-looking halo out of the wrong pieces; if this is true, then properties like the metallicity structure, which is one of the key observables one would like to extract from stellar halo models, will be incorrect. Reproducing the properties of the satellites themselves is a critical test that any painting scheme must pass, and it is not immediately obvious whether it is possible for *any* method of painting DM particles to satisfy all of the necessary constraints. The most important and general point is that painting can easily introduce systematics at the factor-of-several level. We strongly urge modelers to perform tests like these to estimate the magnitude of the systematic error when introducing new painting schemes. ## Baryonic Potential {#sec:discussion-potential} When comparing the SPH-PAINTED and DM-PAINTED halo models, we find that DM-PAINTED is less concentrated, its shape is more prolate and less flattened, and it has more substructure. Both models use the same painting scheme on the same progenitors, but evolve them in a different gravitational potential. The critical differences between the potentials are that the SPH-PAINTED model has a parent galaxy that is more concentrated and has a disk, and also has satellites with deeper potential wells. The global shape of the potential clearly has an impact on the global shape of the stellar halo: when the potential is more concentrated, so is the stellar halo; when the potential is more spherical, so is the stellar halo; when the potential is more prolate, so is the stellar halo. The baryons also make it more difficult to strip particles out of the satellites, so the satellites must get closer to the center to get stripped and therefore deposit their tidal debris closer to the center. The reason for the different amounts of substructure is less clear. There are two physical mechanisms that could decrease the amount of substructure in SPH-PAINTED: differential precession of streams in the oblate potential of the disk, and changes in individual orbits due to the central concentration of the potential (below we will refer to this as "scattering" for simplicity, although for an extended central concentration, like a disk, this is primarily due to individual orbital types changing their shape rather than true scattering onto chaotic orbits that is seen for point-like central concentrations; [@debattista08; @valluri10]). We do not have the ability to independently assess each effect with this simulation, but we note that the global potential in the DM-only simulation is more prolate-triaxial than in the SPH simulation, and it is only in the disk region that the SPH potential has a significantly smaller \(c/a\) axis ratio. We may therefore expect differential precession to disrupt streams in the outer parts of the halo faster in the DM-PAINTED model but streams in the inner parts of the halo faster in the SPH-PAINTED model. Instead, the SPH-PAINTED model has less structure at all radii, but the effect is indeed strongest within \(40\) kpc. This suggests that both differential precession and orbital scattering are playing a role. The overproduction of substructure in pure DM models must be taken into account when comparing these models to observations. For example, determine that the Milky Way halo contains less structure than predicted by the pure DM models, and conclude that the Milky Way halo must also contain a smooth in situ component that reduces the total substructure. However, the factor by which find that the model overpredicts the structure is \(2\)--\(3\), which is of the same magnitude as the level of systematic overprediction of substructure we find for pure DM models. We therefore urge caution against overinterpreting differences between these models and observations that are smaller than the scale of the systematics that we find. ## Simulation Caveats Because the galaxy in the SPH simulation is not a perfect representation of a real galaxy, it is worth discussing how those differences might affect our conclusions. The primary differences between the simulated galaxy and a typical disk galaxy of the same mass are that (1) its potential is too concentrated and (2) its disk fraction is too small. We therefore might expect that effects we see that are due to the concentration of the potential might be overestimated, while those due to the diskiness of the potential might be underestimated. We argue that the systematic error due to painting primarily arises because the kinematics of DM particles deep in the potential of a satellite are different from those of stars deep in the potential. In a more realistic less concentrated potential, the kinematic differences between different particle types might be expected to be less, suggesting an overestimate of the level of this systematic effect. On the other hand, if the stellar distribution were diskier, the kinematics of the stars in that rotating disk would differ even more from those of the non-disky DM particles, suggesting we are underestimating the level of this systematic. We do not know a priori which of these effects would dominate, so the quantitative degree of the effect is uncertain, but the overall sign and approximate magnitude of the systematic are likely to be faithfully indicated by this work. The SPH-PAINTED stellar halo is more concentrated than the DM-PAINTED stellar halo due to the concentration of the baryonic potential. Because the baryonic potential is too concentrated in the SPH simulation, the degree to which the baryons concentrate the stellar halo could also be overestimated. The amount of substructure in a more realistic potential would be expected both to increase because of the less concentrated potential and decrease because of the diskier potential---we believe both phenomena are important in determining the amount of substructure, as discussed in section [5.2](#sec:discussion-potential){reference-type="ref" reference="sec:discussion-potential"}. Again, this may affect the quantitative measurement but is unlikely to affect the sign or overall magnitude of the systematic effect, which are our main conclusions. One further caveat is that the dynamical mass of star particles in the SPH-STARS model is significantly smaller than the mass of the dark matter particles, so the central region of the SPH simulation, where the stars dominate, is effectively simulated at higher resolution than in the DM-only simulation. One could therefore imagine that some of the differences we see are related to resolution rather than to baryonic effects. However, it is unlikely that this would drive our results. Increasing the mass resolution of a DM-only simulation, as we would need to do to match it to the mass resolution of the SPH stars, does not result in different structure at the scales of interest: higher-resolution DM halos are comparably triaxial, have comparable radial profiles, and have comparable numbers of higher-mass (i.e. resolved in the lower-resolution simulation) subhalos. A related issue, however, is that hydrodynamic simulations are much more sensitive to resolution, and so the detailed structure of our SPH galaxy and therefore of the stellar halo models built from it could be different than that of a galaxy simulated at higher resolution. However, the main reasons that we see a difference are because the baryons form a concentrated distribution, and because that distribution is disk-shaped; both of these facts will be true of any hydrodynamic simulation that forms a disk galaxy. Therefore, while the detailed structure of the halo model may differ at different resolutions, the comparison between the halo models should still yield a good estimate for the magnitude of the systematic effects that are introduced by the modeling assumptions. Overall, we therefore caution that the differences between the simulated galaxy and real galaxies may have a quantitative effect on our results. However, the sign and approximate magnitude of the systematic differences reported here are expected to be robust, with the possible exception of the large differences in concentrations between the SPH halo and the dark matter only case, which is expected to persist qualitatively but quantitatively may be substantially overestimated in a simulation such as ours with an overly concentrated baryonic component. # Conclusions We have examined two critical assumptions that are part of most models of the structure of stellar halos: using painted DM particles to represent stars, and the omission of baryonic contributions to the gravitational potential. We have used a controlled set of models where we can independently test their effects. In one test, we compare stars formed in a cosmological hydro simulation to painted DM particles in the same simulation to test the effect of painting, while in the other test we compare DM particles painted in a hydro simulation to DM particles painted in the identical way in a pure N-body realization of the same initial conditions to test the effect of the different potentials. We find that both sets of assumptions cause significant differences in the properties of the predicted stellar halos. Painting results in a *less concentrated* halo, with a half-mass radius \(\sim 6\) times larger, and a *less structured* halo, by a factor of \(\sim 2\). Some of the concentration difference is due to a systematic tendency for progenitors with high stellar mass to wind up at small radius, but most of the concentration difference and all of the structural difference is due to the different kinematics of DM particles and stars at the same location within a satellite. The omission of the baryonic contribution to the potential results in a halo that is *less concentrated*, by a factor of \(~1.7\) in half-mass radius, *more structured*, by a factor of \(2\)--\(7\), and *more prolate*, with \(b/a \sim c/a \sim 0.6\). The mechanisms that drive these changes are likely a combination of orbit scattering from the central density enhancement, differential precession when orbits are near the disk, and the overall prolateness of the dark matter halo. This is the first attempt we are aware of to estimate the magnitude of the systematic effects present in stellar halo models based on dark matter simulations, and the results are somewhat discouraging. The factor-of-a-few level of systematic uncertainty is similar to the factor by which some of the models are discrepant from observations of the Milky Way halo, meaning we cannot presently conclude anything about the origin of the halo from that discrepancy. The hydrodynamic simulations differ from observed galaxies in some important ways that could affect our results; most importantly, the baryons in the simulation are more concentrated, and the baryonic distribution is less disky. These could affect the quantitative measurements that we make, but because these two differences act in opposite directions, the overall sign and approximate magnitude of the systematic error we measure is unlikely to be affected (with the possible exception of the halo concentration, which could potentially be substantially overestimated). We also caution that parts of the simulations we are comparing operate at different resolution due to the presence of low-mass star particles; however, we have argued that this is unlikely to undermine our conclusions. Are there potentially methods that can be used to create high resolution stellar halo models that circumvent these difficulties? One possible way forward, which we are pursuing, is to couple a very sophisticated painting technique to an evolving halo potential that both self-consistently solves the DM dynamics and includes an analytic disk (which should itself be consistent with the properties of a galaxy forming within that DM halo, for example via a semi-analytic model). We stress that it is important for any painting technique to be tested, using methods such as the one we have used in this paper, to ensure that it reproduces the spatial distribution and substructure expected from stars that form with the same efficiency, and that it reproduces the properties (such as stellar mass fraction) of the satellite galaxies. Such a hybrid approach is likely to provide the best hope for modeling stellar halos in the near term. Looking forward, the highest resolution hydro simulations today are beginning to be able to produce interesting predictions, and future hydro simulations with hundreds of millions of particles within the virial radius are likely to provide the best models once they are feasible. We thank the referee, Joop Schaye, for a helpful report. This work has made use of the Shared Hierarchical Academic Research Computing Network (SHARCNET) Dedicated Resource Project: 'MUGS: The McMaster Unbiased Galaxy Simulations Project' (DR316, DR401 and DR437). This work was supported in part by National Science Foundation grant AST 1008324. MV was supported by National Science Foundation grant AST-0908346. VPD is supported by STFC Consolidated grant \# ST/J001341/1. GS received funding from the European Research Council under the European Union's Seventh Framework Programme (FP 7) ERC Grant Agreement n. \[321035\]. HMPC and JW acknowledge financial support from the National Science and Engineering Research Council. [^1]: The difference is because both simulations contain the same number of DM particles, but in the SPH simulation they account for only \(83\%\) of the total mass instead of the entire mass in the DM simulation. [^2]: In determining the "most bound" particles, we use the total potential plus kinetic energies. [^3]: Note that although the scheme was calibrated to recover the correct radial distribution, the fact that it was able to do so via tuning one parameter is not a trivial result: as we have demonstrated, many other painting schemes are not able to do this for *any* choice of parameters.

{'timestamp': '2014-01-23T02:00:47', 'yymm': '1401', 'arxiv_id': '1401.5489', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5489'}

# Introduction There are several different definitions of coherent states based on group-theoretical properties, which all play important roles in diverse areas of theoretical physics. Explicit examples of wave functions of coherent states can show aspects of quantum evolution and semiclassical phenomena in a clean way, but it is not always possible to find such explicit realizations in general-enough terms. However, physical properties can still be extracted if one works with moments of a state instead of wave functions, a description which is the basis also of canonical effective theory. Dynamical equations for the moments follow from generalizations of Ehrenfest's equations, while coherence is implemented by the condition that moments saturate uncertainty relations. The main link to group coherent states, studied in this article, is the imposition of an additional Casimir condition that restricts moments to states that belong to an irreducible representation of a Lie group describing the quantum system (or the quantization of a co-adjoint orbit in the dual Lie algebra of the group). Imposing the Casimir to be constant can be interpreted as a constraint on the original (non-symplectic) phase space with a Poisson structure given by the dual Lie algebra, suggesting that methods for effective constraints can be used. Because a Casimir operator \(\hat{C}\) commutes with all other operators, additional simplifications compared with general first-class constraints arise: (i) expectation values \(\langle\hat{O}\hat{C}\rangle\), which feature prominently in effective constraints, equal the symmetric version \(\frac{1}{2}\langle\hat{O}\hat{C}+\hat{C}\hat{O}\rangle\) and are guaranteed to be real for self-adjoint \(\hat{O}\) and \(\hat{C}\), and (ii) no gauge flow \(\delta\langle\hat{O}\rangle/\delta \epsilon= \langle[\hat{O},\hat{C}]\rangle/i\hbar\) generated by a Casimir constraint need be considered. (On a non-symplectic phase space, first-class constraints do not necessarily generate gauge flows.) These simplifications allow us to confirm, to all orders in a semiclassical or moment expansion, that the correct number of quantum degrees of freedom is left after imposing the Casimir condition. In a second step, we then combine the Casimir condition with the requirement that uncertainty relations be saturated, restricting moments to those of a group coherent state. Higher orders of moments are more difficult to manage at this level, but we will be able to demonstrate several interesting relationships between the different conditions imposed. Quantum cosmology presents an example in which standard group coherent states are not always available in general terms, that is with full squeezing, while the effective methods elaborated here do apply. We end our article with a detailed discussion, clarifying the contentious issue of cosmic forgetfulness which posits that certain pre-big bang models in loop quantum cosmology suffer from a severe lack of control on the pre-big bang state. This issue has lost its urgency with the recent discovery that the same models, when embedded consistently in a setting with perturbative inhomogeneity, lead to signature change at high density, so that no state can be evolved deterministically through the big bang. Nevertheless, the issue of cosmic forgetfulness may still be of interest from a mathematical perpective. # Effective Casimir constraints {#s:EffCas} For simplicity, we consider the case of a single Casimir condition \(C\), imposed on a non-symplectic phase space so that the submanifold \(C={\rm const}\) is symplectic. The phase-space function \(C\) can be seen as one coordinate of a Poisson manifold, such that \(C\) itself is a Casimir function in the Poisson sense, and submanifolds \(C={\rm const}\) are the symplectic leaves. Since \(C\), by definition, has a vanishing Poisson bracket with any other function on the Poisson manifold, it does not generate a Hamiltonian gauge flow when viewed as a constraint. Nevertheless, it can be identified as a first-class constraint owing to the non-symplectic nature of the phase space. (Notions equivalent for symplectic geometry and usually associated with first-class constraints, such as the properties of non-trivial gauge flows and symplectic properties of constraint sets, may no longer be equivalent for Poisson manifolds. The generalization of standard definitions therefore requires some care.) As an example, we may look at a 3-dimensional manifold equipped with coordinates (or basic functions) \(V,J_+,J_-\) and Poisson brackets \(\{V,J_+\}=J_-\), \(\{V,J_-\}=-J_+\) and \(\{J_+,J_-\}=-V\), for instance interpreted as the (dual) Lie algebra of \({\rm sl}(2,{\mathbb R})\). The Casimir function reads \(C=J_+^2+J_-^2-V^2\). Another way to interpret the same system is to use partially complex variables \(V\) with \(J=J_++iJ_-\) and \(J^*= J_+-iJ_-\). One can then realize the Poisson brackets \[\label{PBComplex} \{V,J\}=-iJ\quad,\quad \{V,J^*\}=iJ^*\quad,\quad \{J,J^*\}=2iV\] by functions \(V\) and \(J:=V\exp(-iP)\) of canonical variables \(V\) and \(P\) with \(\{V,P\}=1\). Setting the Casimir function to zero, \(C=JJ^*-V^2=0\), then amounts to a reality condition for \(P\). In this form, the Casimir condition plays a role in some cosmological models. Upon passing on to the quantum treatment of the system, the role of \(C\) as a first-class constraint (though one on a non-symplectic phase space) allows us to employ the effective-constraint methods developed in. To this end, we view the corresponding quantum system algebraically, based on the commutators of basic operators quantizing ([\[PBComplex\]](#PBComplex){reference-type="ref" reference="PBComplex"}). The state space of the algebra can be formulated geometrically by making use of the expectation-value functional \(\langle\cdot\rangle\), applied to all polynomials in basic operators. Expectation values of products of operators are identified as moments \[\label{Moments} \Delta(V^aJ_+^bJ_-^c):=\langle (\hat{V}-\langle\hat{V}\rangle)^a (\hat{J}_+-\langle\hat{J}_+\rangle)^b (\hat{J}_--\langle\hat{J}_-\rangle)^c\rangle_{\rm Weyl-ordered}\] of the basic operators in a state considered. Any product that is not Weyl-ordered can be rearranged as a sum of Weyl-ordered terms, some of which with explicit \(\hbar\)-factors. The set ([\[Moments\]](#Moments){reference-type="ref" reference="Moments"}) therefore prescribes the expectation values for all polynomials in the basic variables. A Poisson geometry of the quantum phase space, given by expectation values and moments of basic operators, is provided by the commutator, defining a Poisson bracket on expectation-value functionals by \[\label{Poisson} \{\langle\hat{A}\rangle,\langle\hat{B}\rangle\}= \frac{\langle[\hat{A},\hat{B}]\rangle}{i\hbar}\,,\] extended to all functions by the Leibniz rule. One element of this algebra is the Casimir or constraint operator \(\hat{C}\) to be imposed as a constraint, \(\hat{C}|\psi\rangle=0\). (For a non-zero Casimir, we can simply redefine the constraint operator as \(\hat{C}-{\rm const}\hat{\mathbf{1}}\) without changing commutators.) Evaluated in any state annihilated by \(\hat{C}\), the expectation values \(\langle\widehat{\rm pol}\hat{C}\rangle\) must all vanish for arbitrary polynomials \(\widehat{\rm pol}\) in the basic operators. For one quantum constraint operator we obtain an infinite number of constraint functions on the quantum phase space. For multiple ones, if the quantum constraints are first class, the system of effective constraints is first-class in the phase-space sense. In the special case of a Casimir constraint, all effective constraints are Casimir functions on the Poisson manifold defined by ([\[Poisson\]](#Poisson){reference-type="ref" reference="Poisson"}): it is straightforward to see that a constraint of the form \(\langle\widehat{\rm pol}\hat{C}\rangle=0\) weakly Poisson commutes with all quantum phase-space functions if \(\hat{C}\) commutes with all operators. For first-class constraints on symplectic phase spaces, the viability of these methods has already been demonstrated, addressing also the problem of time. In general, the ordering of effective constraints in the specific form \[\label{Cpol} C_{\rm pol}:=\langle\widehat{\rm pol}\hat{C}\rangle\] is important for the system of constraints in order to remain first class and to vanish in physical states. We may assume symmetric polynomials without loss of generality, because re-ordering terms would just contribute quantities proportional to lower-order constraints. However, even if the basic operators and \(\hat{C}\) are assumed to be self-adjoint with respect to some \(*\)-relation on the basic algebra, which we will do in what follows, the ordering in ([\[Cpol\]](#Cpol){reference-type="ref" reference="Cpol"}) is in general not symmetric, leading to the possibility of complex-valued effective constraints. For Casimir conditions, on the other hand, we have \([\widehat{\rm pol},\hat{C}]=0\) by definition, so that we can substitute the symmetric ordering \(\frac{1}{2} \langle \widehat{\rm pol}\hat{C}+ \hat{C}\widehat{\rm pol}\rangle\) without changing the expectation value. We are therefore dealing with real-valued effective constraints for all polynomial or moment orders. Moreover, Casimir conditions are easier to implement because no gauge flows need be considered: we have a vanishing \[\{\langle\hat{O}\rangle,\langle\widehat{\rm pol}\hat{C}\rangle\} = \frac{\langle [\hat{O},\widehat{\rm pol}]\hat{C}\rangle}{i\hbar}\approx 0\] on the solution space of the effective constraints, again using the commutation property of a Casimir. This simplification is the main reason that allows us to directly test effective methods at higher orders in moments. ## Removing degrees of freedom In classical systems, Casimir constraints remove phase-space degrees of freedom so that, in the absence of a gauge flow, the constraint surface is symplectic. The dimension of a symplectic manifold is restricted to be even, and its quantization corresponds to a fixed pattern of expectation values and moments: for every canonical pair \((q,p)\) there are two independent expectation values and, starting with \(n=2\) reaching ad infinitum, a tower of \(n+1\) moments for every integer \(n\), defined as in ([\[Moments\]](#Moments){reference-type="ref" reference="Moments"}) for a single canonical pair of basic operators: \[\label{MomentsCan} \Delta(q^ap^b):=\langle (\hat{q}-\langle\hat{q}\rangle)^a (\hat{p}-\langle\hat{p}\rangle)^b\rangle_{\rm Weyl-ordered}\.\] Semiclassically, a moment of order \(n\) behaves like \(O(\hbar^{n/2})\), giving rise to a general semiclassical expansion with a finite number of degrees of freedom per canonical pair at any fixed order \(n\). If this pattern is violated, one cannot interpret the degrees of freedom in the usual way of quantum mechanics, indicating either extra constraints if there are not enough free moments or spurious degrees of freedom if too many variables remain unrestricted. In particular, for a single Casimir constraint we must, by imposing \(C_{\rm pol}=0\), eliminate sufficiently many variables to leave a certain number of canonical pairs with their characteristic moments. For the example of one quadratic Casimir constraint on a 3-dimensional Poisson manifold such as the one of \((V,J_+,J_-)\), one can heuristically see that the reduction is correct, working locally and assuming that all effective constraints are independent. Taking an expectation value \(\langle\hat{C}\rangle=0\) restricts the basic expectation values to two independent ones (for fixed second-order moments). The next order of constraints is obtained for \(\langle\hat{A}\hat{C}\rangle=0\) with \(\hat{A}\) one of the three basic operators (or a linear combination of them). Fixing third-order moments, we obtain three constraints for the six second-order moments of three independent basic operators, just enough to restrict the second-order moments to three independent ones: two fluctuations and one correlation. To see the correct reduction in numbers for all orders, we note that one can, at least locally, view the Casimir function as a coordinate on the Poisson manifold transversal to symplectic leaves. Instead of using basic phase-space functions such as \(V\), \(J_+\) and \(J_-\), we can locally transform to Casimir--Darboux coordinates of a canonical pair \((q,p)\) on the symplectic leaf and \(C\) transversally. This decomposition is relevant also for the counting of degrees of freedom of the corresponding quantum system. As for constraints, in addition to the expectation value of \(\hat{C}\), we must remove all moments of \(\hat{C}\) with itself (such as the fluctuation \(\Delta C\)) as well as all cross moments with \(\hat{q}\) and \(\hat{p}\). These cross moments are nothing but the effective constraints \(\langle\widehat{\rm pol}\hat{C}\rangle=0\), with \(\widehat{\rm pol}\) a polynomial of degree \(n\) to restrict \(C\)-moments of order \(n+1\). Since there are as many cross-moments as constraints of this form, we obtain the correct number of degrees of freedom provided the constraints are all independent (and the local argument remains valid in quantum theory). In most cases, and always if the coordinate change to Casimir--Darboux coordinates is not global, the transformation from \((V,J_+,J_-)\) to \((q,p,C)\) is non-linear and relationships between \((V,J_+,J_-)\)-moments and \((q,p,C)\) are non-trivial. For instance, a moment of low order in one system may involve moments of all orders in the other. For a global statement about parameter counting, more-detailed considerations must be performed. ## Counting truncated constraint conditions For more generality, we work with a general finite-dimensional algebra of generators \(\hat{x}_i\), \(i=1, 2, \ldots M\), denoting their expectation values (or, occasionally, the corresponding classical values) by \(x_i\), \(i=1, 2, \ldots M\). To define the moments, we introduce \[\widehat{\Delta x}_i := \hat{x}_i-x_i \hat{\mathbf{1}}\] and their Weyl-ordered products \[\hat{e}_{\vec{i}} = \hat{e}_{(i_1, i_2, \ldots i_M)} := \left(\widehat{\Delta x}_1^{i_1} \widehat{\Delta x}_2^{i_2} \ldots \widehat{\Delta x}_M^{i_M} \right)_{\rm Weyl-ordered}\,,\] which form a linear basis for the (extended) algebra. We use a compact notation in which \(\vec{i}\) is an \(M\)-tuple of non-negative integers. Expectation values \[\label{Momentsx} \Delta(\vec{x}^{\vec{i}}):=\Delta(x_1^{i_1}\cdots x_M^{i_M}) = \langle \hat{e}_{\vec{i}} \rangle\.\] of the basis elements are the moments. For later use we define the degree \(| \vec{i} | := \sum_{n=1}^M i_n\) and a partial ordering \(\vec{i} \geq \vec{j}\) if \(i_n \geq j_n\) for all \(n\), so that \(\vec{i} > \vec{j}\) if \(\vec{i} \geq \vec{j}\) and \(\vec{i} \neq \vec{j}\). We will use \(\vec{i}!\) to denote \((i_1!i_2! \ldots i_M!)\) and, as already defined in ([\[Momentsx\]](#Momentsx){reference-type="ref" reference="Momentsx"}), \(\vec{x}^{\vec{i}}=x_1^{i_1}\cdots x_M^{i_M}\). By construction, \(\Delta(\vec{x}^{\vec{i}})=0\) for all \(\vec{i}\) with \(| \vec{i}| = 1\). Interpreting the \(\hat{x}_i\) as basic operators of a quantum system, we assume that their commutator algebra is linear (and follows from a direct quantization of Poisson brackets of the corresponding classical functions): \[\label{Algebra} [\hat{x}_i, \hat{x}_j] = i\hbar \epsilon_{ij}^{\ \ k} \hat{x}_k\] where \(\epsilon_{ij}^{\ \ k}\) are structure constants, identical to the structure constants of the classical Poisson algebra. For a semisimple Lie algebra, which we will assume in an example later in this article, \(\epsilon_{ijk}\) with the third index pulled down by contraction with the Killing metric, are totally antisymmetric. The case of the Weyl algebra \([\hat{q}, \hat{p}]=i\hbar \hat{\mathbf{1}}\) for a single canonical pair, as another important example, does not seem to fall within the current setting. However, we could treat \(\hat{q}\), \(\hat{p}\) and \(\hat{\mathbf{1}}\) as three 'generators' or, alternatively, start from the Heisenberg algebra \([\hat{q},\hat{p} ]= i\hbar \hat{z}\) and enforce a constraint \(\hat{C} = \hat{z}-\hat{\mathbf{1}}\) in the way described below. It follows that \[\label{DeltaComm} \left[ \widehat{\Delta x_i}, \widehat{\Delta x_j} \right] = i\hbar \sum_k \epsilon_{ij}^{\ \ k} \hat{x}_k = i \hbar \sum_k \epsilon_{ij}^{\ \ k} \left( \widehat{\Delta x_k} + x_k \hat{\mathbf{1}} \right)\] with the two right-hand-side terms respectively of polynomial degree \(1\) and \(0\) in the operators \(\widehat{\Delta x_i}\). (This relation is not formulated for linear operators, owing to the presence of expectation values that depend on a state. Nevertheless, the usual rules can be applied if one treats \(x_k\), during a calculation, as some real number and identifies it with the expectation value of \(\hat{x}_k\) only in the final results.) In this form, the relation plays an important role in considerations of orderings: it allows one to apply commutators to symmetrically order any product of \(\widehat{\Delta x}_i\)-s by adding terms of lower polynomial degree and proportional to powers of \(\hbar\), \[\begin{aligned} \hat{e}_{\vec{i}} \hat{e}_{\vec{j}} &=& \hat{e}_{(\vec{i}+\vec{j})} + \hbar \sum_{\tiny{\begin{array}{c} \vec{k} < \vec{i}+\vec{j} \\ |\vec{k}| \leq |\vec{i}+\vec{j}|-1 \end{array}}} {}^{(1)}\!\beta_{\vec{i}, \vec{j}}{}^{\vec{k}} \:\hat{e}_{\vec{k}} + \hbar^2 \sum_{\tiny{\begin{array}{c} \vec{k} < \vec{i}+\vec{j} \\ |\vec{k}| \leq |\vec{i}+\vec{j}|-2 \end{array}}} {}^{(2)}\!\beta_{\vec{i}, \vec{j}}{}^{\vec{k}}\:\hat{e}_{\vec{k}} + \cdots \end{aligned}\] where \({}^{(n)}\!\beta_{\vec{i}, \vec{j}}{}^{\vec{k}}\) are polynomials in the expectation values \(x_i\). ### Constraints and truncation {#sec:truncation} Following the general procedure of effective constraints, the Casimir condition of an operator \(\hat{C}\) is imposed by demanding \(\langle f(\hat{x}_1, \hat{x}_2, \ldots, \hat{x}_M) \hat{C} \rangle = 0\) for all polynomial functions \(f\). These are infinitely many conditions for infinitely many moments, but working order by order in the moments (or in a semiclassical expansion) one can truncate these conditions systematically using the basis \(\{\hat{e}_{\vec{i}} \}\). We introduce \[C_{\vec{i}} := \langle \hat{e}_{\vec{i}\ } \hat{C} \rangle = 0, \ \ \forall \ \vec{i} \in \mathbb{Z}_+^M\] with the convention \(\hat{e}_0=\hat{\mathbf{1}}\), so that \(C_0 := \langle \hat{C} \rangle\). In order for the truncation to be consistent, we must suitably combine the truncation of constraint functions according to the degree \(|\vec{i}|\) with a truncation of variables that feature in the system. Assume that we truncate at some order \(N \geq 2\) of the semiclassical expansion: we drop moments of degree greater than \(N\), that is all \(\Delta(\vec{x}^{\vec{i}})\) with \(| \vec{i} | > N\). The truncation of the system of constraints is more subtle. All \(C_{\vec{i}}\) are linear functions of the moments \(\Delta(\vec{x}^{\vec{j}})\), but they contain terms of three different types. It is useful to assign orders to these terms according to their type, as follows: - A term of the form \(f(x_1, x_2 \ldots x_M)\Delta(\vec{x}^{\vec{i}})\), where \(f\) is a polynomial in the expectation values not proportional to the classical constraint, is assigned semiclassical order equal to \(| \vec{i} |\). - A term of the form \(C(x_1, x_2 \ldots x_M)\Delta(\vec{x}^{\vec{i}})\), where \(C\) is the classical polynomial expression for the constraint, requires a special treatment. As an exception to the previous point, it is assigned semiclassical order \(| \vec{i} |+ 2\). This exception can be understood from the fact that \(C(x_1, x_2 \ldots x_M)\), although it may not vanish exactly when quantum corrections are included in \(C_0=C+O(\hbar)\), is of the order \(\hbar\) when the quantum constraint \(C_0=0\) is imposed. See also . - Terms such as \(\hbar^n f(x_1, x_2 \ldots x_M)\Delta(\vec{x}^{\vec{i}})\), which may arise upon reordering algebra elements, are assigned semiclassical order equal to \(| \vec{i} |+ 2n\). All these orders are consistent with the correspondence of order \(n\) to terms \(O(\hbar^{n/2})\) as it applies to the moments. To truncate, terms in any expression for a constraint function are dropped when they are of semiclassical order higher than \(N\). ### Counting degrees of freedom We begin by counting the degrees of freedom of an unconstrained system generated by polynomials in \(M\) basic variables (for a phase space of dimension \(M\)). We have \(M\) expectation values and, at each semiclassical order \(N \geq 2\), the degrees of freedom are represented by the independent functions \(\Delta(\vec{x}^{\vec{i}})\) with \(| \vec{i} | = N\). The task of counting degrees of freedom is simplified by rephrasing the problem: The number of such variables is given by the number of \(M\)-tuples of non-negative integers with \(|\vec{i}|=N\), a quantity that we will call \(\mathcal{N}_M(N)\). Each such \(M\)-tuple is produced by considering a row of \(N+M-1\) identical objects, marking \(M-1\) of them to serve as partitions. The value of \(i_n\) is then the number of unmarked objects between partition \((n-1)\) and partition \(n\). (Partitions \(0\) and \(M\) are assumed to be at the ends.) With this rephrasing we directly obtain \[\mathcal{N}_M(N) = {N+M-1 \choose M-1}\.\] Each constraint, in the absence of gauge flows, removes a single classical degree of freedom, and we expect the tower of effective constraint conditions to remove the corresponding moments. After the constraints are imposed, there should be as many degrees of freedom as for a system with \(M-1\) generators, with \(\mathcal{N}_{M-1}(N) = {N+M-2 \choose M-2}\) free variables at each semiclassical order. Using the identity \({a\choose b}-{a-1 \choose b} = {a-1\choose b-1}\), we conclude that the required number of independent conditions at each order should be \({N+M-2 \choose M-1}\). In order to show consistency of the effective procedure, we now proceed to counting the number of conditions independently. This question is more difficult to answer because constraint conditions \(C_{\vec{i}}\) generally mix terms of different orders. As a first step, we show that, truncated at a given order, the system of constraints is finite. The number of constraints at each order is then the number of additional non-trivial constraint conditions that arise when we raise the truncation order by one. A general constraint function has the form \[\begin{aligned} \label{eq:constraints} C_{\vec{i}} &=& \sum_{\vec{j} \geq \vec{i}} \frac{1}{(\vec{j}-\vec{i})!} \frac{\partial^{|\vec{j}-\vec{i}|} C}{\partial x_1^{j_1-i_1} \ldots \partial x_{M}^{j_M-i_M}} \Delta(\vec{x}^{\vec{j}}) + \hbar \sum_{|\vec{j}| \geq |\vec{i}|-1} {}^{(1)}\!\alpha _{C_{\vec{i}}}{}^{\vec{j}}\: \Delta(\vec{x}^{\vec{j}}) \nonumber \\ & & + \hbar^2 \sum_{|\vec{j}| \geq |\vec{i}|-2} {}^{(2)}\!\alpha _{C_{\vec{i}}}{}^{\vec{j}} \:\Delta(\vec{x}^{\vec{j}}) + \ldots + \hbar^{|\vec{i}|} \sum_{\vec{j}} {}^{(|\vec{i}|)}\!\alpha _{C_{\vec{i}}}{}^{\vec{j}}\: \Delta(\vec{x}^{\vec{j}}) \end{aligned}\] Here \({}^{(n)}\!\alpha _{C_{\vec{i}}}{}^{\vec{j}}\) are coefficients of semiclassical order zero, polynomial in the expectation values \(x_i\). The first sum comes from the Weyl-symmetric part of the element \(\hat{e}_{\vec{i}\ } \hat{C}\). Subsequent sums arise from its components that are antisymmetric in one, two and more adjacent pairs of moment-generating elements \(\widehat{\Delta x}_i\). Each antisymmetric pair can be reduced by using the commutation relations, producing the powers of \(\hbar\). For a Casimir constraint, \(\hat{C}_{\vec{i}}\) is guaranteed to be real; therefore, there must be an even number of commutators applied in each re-ordering step and we have only even powers of \(\hbar\), or \({}^{(n)}\!\alpha _{C_{\vec{i}}}{}^{\vec{j}}=0\) for odd \(n\). The important feature of the above expansion is that the lowest semiclassical order terms are \(C\Delta(\vec{x}^{\vec{i}})\) and \[\label{Cbar} \bar{C}_{\vec{i}}:= \sum_k \frac{\partial C}{\partial x_k} \Delta(\vec{x}^{(i_1,\ldots,i_{k-1},i_k+1,i_{k+1},\ldots,i_M)})\.\] The latter term has the lowest order, \(|\vec{i}|+1\), keeping in mind that for the purposes of truncating the constraints, \(C\) is of order \(2\). After truncation at order \(N\), constraints \(C_{\vec{i}}=0\) are satisfied identically for all \(|\vec{i}| >N-1\). This observation allows us again to rephrase the counting problem: The number of non-trivial conditions up to order \(N\) is the same as the number of non-negative integer \(M\)-tuples of degree \(N-1\) and less. As we go from truncation at order \(N-1\) to truncation at order \(N\), this number changes by the number of \(M\)-tuples of degree \(N-1\), which is exactly the quantity \(\mathcal{N}_M(N-1)={N+M-2 \choose M-1}\) required to be eliminated by the counting of degrees of freedom. Thus, provided the non-trivial constraint conditions remaining after truncation are functionally independent, they remove precisely one combinatorial degree of freedom in the quantum mechanical sense. In order to show independence, which is done in detail in, one considers the gradients \(\{{\rm d}_{\Delta}\bar{C}_{\vec{i}} \}_{1\leq |\vec{i}|\leq N-1}\) on the space of expectation values and moments. While the operator \(d_{\Delta}\) takes the gradient with respect to the quantum moments only, the functional independence conditions derived depend predominantly on the classical form of the constraint function. For a semiclassical state, applying careful truncations as before, one can conclude: (i) The truncated constraints are independent as long as \({\rm d}C\) (the gradient taken with respect to the expectation values only) is not comparable to \(\hbar\) or the moments in at least one coordinate direction, assuming that expectation values satisfy the classical constraint. This condition can be interpreted as the semiclassical analog of regularity of the constraints. (ii) For expectation values off the classical constraint surface, the constraint functions are functionally independent as long as, for some \(k\), neither \(\partial C/\partial x_k\) nor \(\partial^{N-2}(C^{-1})/\partial x_k^{N-2}\) are comparable to \(\hbar\) or the moments. The conditions may clearly be violated somewhere on the quantum phase space, but near the classical constraint surface terms such as \(1/C\) and its derivatives diverge and appear in the gradients. The gradients, unlike the constraints, can thus be considered "large." These conditions are sufficient, but not necessary, so that in some cases the constraints may be independent even if the conditions do not hold. To summarize, so long as the classical constraint is sufficiently regular, the truncated set of gradients \(\{ {\rm d}_{\Delta}\bar{C}_{\vec{i}} \}_{1\leq |\vec{i}|\leq N-1}\) is linearly independent for expectation values \(x_i\) lying in some neighborhood of the classical constraint surface \(C=0\), leading to functional independence of the truncated set of constraint functions in that region. # Uncertainty relations For moments of a group coherent state, in a quantization of a co-adjoint orbit on the dual of the Lie algebra corresponding to ([\[Algebra\]](#Algebra){reference-type="ref" reference="Algebra"}), we require that uncertainty relations be saturated. Together with the Casimir condition, several equations are then to be solved. We first derive uncertainty relations, beginning the usual procedure familiar from textbooks on quantum mechanics. We pick a pair \((\hat{x}_i,\hat{x}_j)\) of basic operators, assumed self-adjoint as noted before. Starting with a generic state \(\psi\) in a Hilbert-space representation of the basic algebra, we introduce three new states \(v_{x_i}:=\widehat{\Delta x_i}\psi\), \(w_{x_j}:=\widehat{\Delta x_j}\psi\) and \(u_{x_i,x_j}:= w_{x_j}-(v_{x_i}\cdot w_{x_j}/||v_{x_i}||^2) v_{x_i}\). The uncertainty relation for our pair of operators then follows from \[0\leq ||u_{x_i,x_j}||^2= ||w_{x_j}||^2-\frac{|v_{x_i}\cdot w_{x_j}|^2}{||v_{x_i}||^2}\,,\] with saturation if and only if \(u_{x_i,x_j}=0\). Inserting our specific expressions for \(u_{x_i}\) and \(w_{x_j}\) in terms of \(\psi\), we can express each term in the Schwarz inequality \(||v_{x_i}||^2||w_{x_j}||^2\geq |v_{x_i}\cdot w_{x_j}|^2\) in terms of moments. We easily obtain \(||v_{x_i}||^2=\Delta(x_i^2)\) and \(||w_{x_j}||^2=\Delta(x_j^2)\) and, with a little more re-ordering work, \[|v_{x_i}\cdot w_{x_j}|^2= |\langle (\hat{x}_i-x_i)(\hat{x}_j-x_j)\rangle|^2= |\Delta(x_ix_j)+ {\textstyle\frac{1}{2}} i\hbar \epsilon_{ij}{}^kx_k|^2= \Delta(x_ix_j)^2+ \frac{\hbar^2}{4}(\epsilon_{ij}{}^kx_k)^2\.\] Uncertainty relations \[\label{Uncertx} \Delta (x_i^2)\Delta (x_j^2)-\Delta(x_ix_j)^2\geq \frac{\hbar^2}{4} (\epsilon_{ij}{}^kx_k)^2\] in standard form then follow. ## Higher orders Higher-order moments are restricted by higher-order uncertainty relations. We can derive them by using non-linear polynomials in the \(\widehat{\Delta x_i}\) to define states \(v_{{\rm pol}_1}:=\widehat{{\rm pol}_1}\psi\) and \(w_{{\rm pol}_2}:=\widehat{{\rm pol}_2}\psi\) and proceeding as before. Without loss of generality, we require \(\widehat{{\rm pol}_{1/2}}\) to be Weyl-ordered. Unlike ([\[Uncertx\]](#Uncertx){reference-type="ref" reference="Uncertx"}), these higher-order relations mix moments of different orders. (For recent work on higher-order uncertainty relations with canonical basic operators, see.) The first ones beyond ([\[Uncertx\]](#Uncertx){reference-type="ref" reference="Uncertx"}), for instance, involve moments of second, third and fourth order, obtained when \({\rm pol}_1\) is linear and \({\rm pol}_2\) quadratic (or vice versa). Using the relations in the appendix, we compute \[\begin{aligned} ||w_{x_jx_k}||^2 &=& \Delta(x_j^2x_k^2)\\ &&-\frac{1}{6} \hbar^2 \left( \epsilon_{jk}{}^l \epsilon_{jl}{}^m \Delta(x_kx_m)-\epsilon_{jk}{}^l \epsilon_{kl}{}^m\Delta(x_jx_m)\right)-\frac{3}{4}\hbar^2 \epsilon_{jk}{}^l \epsilon_{jk}{}^m \Delta(x_lx_m) \nonumber \end{aligned}\] (no sum over repeated lower indices) and \[\begin{aligned} |\langle v_{x_i},w_{x_jx_k}\rangle|^2 &=& \left(\Delta(x_ix_jx_k)- \frac{1}{12}\hbar^2\left(\epsilon_{ik}{}^l \epsilon_{lj}{}^m+ \epsilon_{ij}{}^l \epsilon_{lk}{}^m\right)x_m\right)^2\\ && + \frac{1}{4}\hbar^2\left( \epsilon_{ij}{}^l\Delta(x_kx_l)+ \epsilon_{ik}{}^l\Delta(x_jx_l)\right)^2\. \nonumber \end{aligned}\] If we only consider terms of lowest order (six), this third-order uncertainty relation becomes \[\label{UncertThird} \Delta(x_j^2x_k^2)\Delta(x_i^2)-\Delta(x_ix_jx_k)^2\geq 0.\] Unless there are third-order correlations between the basic variables, the only implication is that fourth-order moments of the form \(\Delta(x_j^2x_k^2)\) must be positive, which already follows from their definition. The next order of uncertainty relations is more interesting and bounds fourth-order moments by a positive number. Defining \(\vec{\hat{x}}{}^{\vec{i}} = \left(\hat{x}_1^{i_1} \hat{x}_2^{i_2} \ldots \hat{x}_M^{i_M}\right)_{\rm Weyl-ordered}\), in general, the leading-order contribution to the uncertainty relation implied by \(\widehat{{\rm pol}_1}=\vec{\hat{x}}{}^{\vec{i}}\) and \(\widehat{{\rm pol}_2}=\vec{\hat{x}}{}^{\vec{j}}\) is of the form \[\label{UncertGen} \Delta(\vec{x}^{2\vec{i}})\Delta(\vec{x}^{2\vec{j}})- \Delta(\vec{x}^{2\vec{i}+2\vec{j}})\geq U\] where \(U\) follows from the squared imaginary part of \(\langle\vec{\hat{x}}{}^{\vec{i}}\vec{\hat{x}}{}^{\vec{j}}\rangle-\Delta(\vec{x}^{\vec{i}+\vec{j}})\). The leading order is obtained if exactly one commutator is applied in the re-ordering required to bring \(\langle\vec{\hat{x}}{}^{\vec{i}}\vec{\hat{x}}{}^{\vec{j}}\rangle\) into \(\Delta(\vec{x}^{\vec{i}+\vec{j}})\). It has two contributions, one of degree \(2(|\vec{i}|+|\vec{j}|+1)\) (after taking the square) from the \(\widehat{\Delta x_k}\)-term in ([\[DeltaComm\]](#DeltaComm){reference-type="ref" reference="DeltaComm"}), and one of order \(2(|\vec{i}|+|\vec{j}|)\) from the \(x_k\)-term. The latter is always of the same order as the leading contribution on the left-hand side of ([\[UncertGen\]](#UncertGen){reference-type="ref" reference="UncertGen"}), except when it happens to vanish. It always vanishes for third-order uncertainty relations ([\[UncertThird\]](#UncertThird){reference-type="ref" reference="UncertThird"}) because it contains only the vanishing \(\Delta(\vec{x}^{\vec{i}})\) with \(|\vec{i}|=1\). For higher orders, however, \(U\) in ([\[UncertGen\]](#UncertGen){reference-type="ref" reference="UncertGen"}) is non-zero to leading order, so that the familiar form of uncertainty relations is obtained, with the right-hand side non-zero and proportional to \(\hbar^2\). ## Relations for group coherent states Group coherent states obey Casimir conditions and saturate uncertainty relations. We will now explore the interplay of these different conditions, focusing on the case of a 3-dimensional semisimple Lie group of rank one. In particular, the counting problem, treated in the previous section for a Casimir condition irrespective of uncertainty relation, is again of relevance. On the unrestriced state space, ignoring the Casimir condition, we obtain an independent uncertainty relation ([\[Uncertx\]](#Uncertx){reference-type="ref" reference="Uncertx"}) for every pair of generators \(\hat{x}_i\). When the Casimir condition is imposed, some degrees of freedom are removed, and we should expect a smaller number of independent uncertainty relations. For instance, for a 3-dimensional Lie group of rank one, we start with three second-order uncertainty relations, one for every pair among three basic operators, but should have only one independent one for the 2-dimensional phase space left after the Casimir condition is imposed. This counting problem is difficult to analyze at arbitrary orders, but it is instructive to have a look at the leading orders of second-order moments that feature in ([\[Uncertx\]](#Uncertx){reference-type="ref" reference="Uncertx"}). Under the present assumptions, we have a quadratic Casimir \(C=k^{ij}x_ix_j-c=0\) with a constant \(c\), using the Killing metric \(k^{ij}\). To lowest order, such a condition implies that \[\label{CbarKill} \bar{C}_{\vec{i}}= 2x_ik^{ij}\Delta(\vec{x}^{\vec{i}} x_j)=0\,,\] using ([\[Cbar\]](#Cbar){reference-type="ref" reference="Cbar"}). For higher-order contributions, we must perform more-explicit calculations keeping track of all necessary re-orderings. It is then useful to write the corresponding Casimir operator in terms of \(\widehat{\Delta x_i}\): \[\begin{aligned} \hat{C}&=&k^{ij}\hat{x}_i\hat{x}_j-c\hat{\mathbf{1}}= k^{ij}(\widehat{\Delta x_i}+x_i) (\widehat{\Delta x_j}+x_j)-c\hat{\mathbf{1}}\\ &=& k^{ij} \widehat{\Delta x_i}\widehat{\Delta x_j}+ 2k^{ij} x_i \widehat{\Delta x_j}+ k^{ij} x_ix_j-c\hat{\mathbf{1}}\. \end{aligned}\] The effective Casimir condition then reads \[\label{CasEff} \langle\hat{C}\rangle= k^{ij} \Delta(x_ix_j)+ k^{ij}x_ix_j-c=0\,,\] and can be used to simplify \(\hat{C}\) "on-shell," that is when the basic Casimir condition holds: \[\label{CApp} \hat{C}\approx k^{ij} (\widehat{\Delta x_i}\widehat{\Delta x_j}-\Delta(x_ix_j))+ 2k^{ij} x_i \widehat{\Delta x_j}\.\] In this form, the condition is useful for higher-order effective constraints because the lower-order ones will already be taken care of. Moreover, ([\[CasEff\]](#CasEff){reference-type="ref" reference="CasEff"}) does not explicitly depend on \(c\), so that representation-independent relations between expectation values and moments of different orders will be obtained. (The actual values of moments depend on \(c\) and the representation once specific solutions to ([\[CasEff\]](#CasEff){reference-type="ref" reference="CasEff"}) are used.) The next-order conditions are obtained from \(\frac{1}{2}\langle \widehat{\Delta x_k}\hat{C}+ \hat{C}\widehat{\Delta x_k}\rangle\), in which we directly write the symmetric ordering (and use \(\widehat{\Delta x_k}\) because the \(x_k\)-terms subtracted from \(\hat{x}_k\) would just multiply lower-order constraints). However, not all terms in \(k^{ij}(\hat{x}_k\hat{x}_i\hat{x}_j+\hat{x}_i\hat{x}_j\hat{x}_k)\) are ordered totally symmetric and give rise to moments as defined in ([\[Momentsx\]](#Momentsx){reference-type="ref" reference="Momentsx"}). To change the ordering in this cubic case, we make use of the identity \[\frac{1}{2}(\hat{A}\hat{B}\hat{D}+\hat{B}\hat{D}\hat{A})= (\hat{A}\hat{B}\hat{D})_{\rm Weyl-ordered}+ \frac{1}{12}([[\hat{A},\hat{B}],\hat{D}]+ [[\hat{A},\hat{D}],\hat{B}])+ \frac{1}{4}\{\hat{A},[\hat{B},\hat{D}]\}\] and write \[\begin{aligned} \frac{1}{2}\left( \widehat{\Delta x_k}\hat{C}+ \hat{C}\widehat{\Delta x_k}\right)&=& k^{ij}(\widehat{\Delta x_i}\widehat{\Delta x_j}\widehat{\Delta x_k})_{\rm Weyl-ordered}\\ &&+ k^{ij} \left(\frac{1}{6}[[\widehat{\Delta x_k},\widehat{\Delta x_i}], \widehat{\Delta x_j}]-\Delta(x_ix_j)\widehat{\Delta x_k}\right)+2 k^{ij}x_i(\widehat{\Delta x_j}\widehat{\Delta x_k})_{\rm Weyl-ordered}\. \end{aligned}\] Its expectation value gives the third-order Casimir condition \[\frac{1}{2}\langle \widehat{\Delta x_k}\hat{C}+ \hat{C}\widehat{\Delta x_k}\rangle= k^{ij}\Delta(x_ix_jx_k)-\frac{\hbar^2}{6} k^{ij}\epsilon_{ki}{}^l\epsilon_{lj}{}^m x_m + 2k^{ij}x_i \Delta(x_jx_k)\approx0\.\] Even though the condition is derived from third-order operators, its leading term (the last one in the equation) is of second order because the Casimir operator does not depend just on \(\widehat{\Delta x_i}\). If moments (and explicit factors of \(\hbar\)) of higher than second order are ignored, the third-order Casimir condition restricts second-order moments by \[\label{ThirdCasimir} x^j\Delta(x_jx_k)\approx 0\,,\] where we raise the index using the Killing metric. This result is clearly a special case of ([\[CbarKill\]](#CbarKill){reference-type="ref" reference="CbarKill"}). To the next order, we obtain from \(\langle(\widehat{\Delta x_k}\widehat{\Delta x_l})_{\rm Weyl-ordered}\hat{C}\rangle=0\) the relation \[\begin{aligned} \label{FourthCasimir} &&2x^i\Delta(x_ix_kx_l)+ k^{ij} \left(\Delta(x_ix_jx_kx_l)- \Delta(x_ix_x)\Delta(x_kx_l)\right)\\ &&+\frac{1}{6}\hbar^2 k^{ij} \left(3\epsilon_{i(k}{}^m \epsilon_{l)j}{}^n \Delta(x_mx_n)+ 2\epsilon_{i(k}{}^m \epsilon_{l)m}{}^n \Delta(x_jx_n)+2 \epsilon_{mj}{}^n \epsilon_{i(k}{}^m \Delta(x_{l)}x_n)\right)\nonumber\\ &&-\frac{1}{6}\hbar^2 k^{ij} \epsilon_{ki}{}^m \epsilon_{lj}{}^n x_mx_n \approx0\. \nonumber \end{aligned}\] For totally antisymmetric \(\epsilon_{ijk}\), the second line simplifies to \(-\frac{5}{6} \hbar^2 \Delta(x_lx_k)-\frac{1}{6} \hbar^2\delta_{lk} \Delta(x^ix_i)\), and the last line to \(-\frac{1}{6}\hbar^2(\delta_{lk}x^ix_i-x_lx_k)\). ## Interplay of Casimir conditions and uncertainty relations With the more-explicit form of effective Casimir conditions found in the preceding subsection, we can see how the number of uncertainty relations is reduced. As before, we assume a 3-dimensional Lie algebra with three uncertainty relations ([\[Uncertx\]](#Uncertx){reference-type="ref" reference="Uncertx"}). We start with the relation obtained for the pair \((\hat{x}_1,\hat{x}_2)\) and use ([\[ThirdCasimir\]](#ThirdCasimir){reference-type="ref" reference="ThirdCasimir"}) to bring it to the form of one of the others. To this end, we multiply the left-hand side of ([\[Uncertx\]](#Uncertx){reference-type="ref" reference="Uncertx"}) with \((x^1)^2\) and rewrite using \[\begin{aligned} x^1\Delta(x_1^2) &=&-x^2\Delta(x_2x_1)-x^3\Delta(x_3x_1)\\ x^1\Delta(x_1x_2) &=&-x^2\Delta(x_2^2)-x^3\Delta(x_3x_2)\\ x^1\Delta(x_1x_3) &=&-x^2\Delta(x_2x_3)-x^3\Delta(x_3^2) \end{aligned}\] several times. We obtain \[\begin{aligned} (x^1)^2 (\Delta(x_1^2)\Delta(x_2^2)-\Delta(x_1x_2)^2) &=& ((x^2)^2\Delta(x_2^2)+2x^2x^3\Delta(x_2x_3)+(x^3)^2\Delta(x_3^2)) \Delta(x_2^2)\\ &&-(x^2\Delta(x_2^2)+x^3\Delta(x_2x_3))^2\\ &=& (x^3)^2(\Delta(x_2^2)\Delta(x_3^2)-\Delta(x_2x_3)^2)\,, \end{aligned}\] just as needed for the left-hand side of the uncertainty relation belonging to the pair \((\hat{x}_2,\hat{x}_3)\). On the right-hand side, we have \(x^1\epsilon_{12}{}^kx_k= x^1\epsilon_{123}x^3=-x^3\epsilon_{32k}x^k\), using the behavior of structure constants for rank-1 semisimple Lie algebras. Also the right-hand sides will then match, showing that the uncertainty relations for \((\hat{x}_1,\hat{x}_2)\) and \((\hat{x}_2,\hat{x}_3)\) are not independent. Similarly, one can show that the one for \((\hat{x}_1,\hat{x}_3)\) is not independent either, demonstrating that only one independent uncertainty relation is left after the Casimir condition is imposed. Including higher orders in the Casimir constraints, we have \[\begin{aligned} x^1\Delta(x_1^2) &=&-x^2\Delta(x_2x_1)-x^3\Delta(x_3x_1)-\frac{1}{2}\Delta(x^ix_ix_1)+ \frac{1}{6}\hbar^2 x_1\\ x^1\Delta(x_1x_2) &=&-x^2\Delta(x_2^2)-x^3\Delta(x_3x_2)-\frac{1}{2}\Delta(x^ix_ix_2)+ \frac{1}{6}\hbar^2 x_2\\ x^1\Delta(x_1x_3) &=&-x^2\Delta(x_2x_3)-x^3\Delta(x_3^2)-\frac{1}{2}\Delta(x^ix_ix_3)+ \frac{1}{6}\hbar^2 x_3 \end{aligned}\] and \[\begin{aligned} (x^1)^2 (\Delta(x_1^2)\Delta(x_2^2)-\Delta(x_1x_2)^2) &=& (x^3)^2(\Delta(x_2^2)\Delta(x_3^2)-\Delta(x_2x_3)^2)\\ &&-\frac{1}{2}x^j\Delta(x^ix_ix_j) \Delta(x_2^2)\\ &&+ x^3\left( \Delta(x^ix_ix_3)\Delta(x_2^2)-\Delta(x^ix_ix_2)\Delta(x_2x_3)\right)\\ && +\frac{1}{6}\hbar^2 \left((x^1x_1+x^2x_2-x^3x_3) \Delta(x_2^2)+ 2x^3x_2\Delta(x_2x_3)\right) \\ &&-\frac{1}{4}\Delta(x^ix_ix_2)^2+ \frac{1}{6}\hbar^2 x_2\Delta(x^ix_ix_2)-\frac{1}{36}\hbar^4 x_2^2 \. \end{aligned}\] In the second line one can use ([\[FourthCasimir\]](#FourthCasimir){reference-type="ref" reference="FourthCasimir"}), which brings in fourth-order moments. This shows that the interplay of different uncertainty relations subject to Casimir conditions is more complicated and mixes the orders. A more systematic description of higher-order uncertainty relations would be useful. # Example: Reality conditions of harmonic loop quantum cosmology As a detailed application, we consider, as introduced in Sec. [2](#s:EffCas){reference-type="ref" reference="s:EffCas"}, basic operators satisfying the \({\rm sl}(2,{\mathbb R})\) algebra \[\label{comm} [\hat{V},\hat{J}]=\hbar\hat{J}\quad,\quad {} [\hat{V},\hat{J}^{\dagger}]=-\hbar\hat{J}^{\dagger}\quad,\quad {} [\hat{J},\hat{J}^{\dagger}]=-2\hbar(\hat{V}+\hbar/2)\] subject to the condition \(\hat{J}\hat{J}^{\dagger}=\hat{V}^2\). (This example is not exactly of the form discussed previously, owing to the central extension by \(\hbar/2\). However, as this central extension is trivial, the same results hold true. One may absorb \(\hbar/2\) in the definition of \(\hat{V}\), but we will keep this term explicitly since it contributes to some quantum corrections.) For expectation values, the operator identity \(\hat{C}=\hat{J}\hat{J}^{\dagger}-\hat{V}^2\) implies the equation \[\label{reality} |J|^2-(V+\hbar/2)^2= (\Delta V)^2-\Delta(JJ^*)+\frac{1}{4}\hbar^2\] relating expectation values to fluctuations of \(\hat{V}\), \(\hat{J}\) and \(\hat{J}^{\dagger}\). Alternatively, we may write the condition as \[J_+^2+J_-^2-V^2= (\Delta V)^2-(\Delta J_+)^2-(\Delta J_-)^2+\frac{1}{2} \hbar(2V+\hbar)\] for moments of self-adjoint operators \(\hat{V}\), \(\hat{J}_+= \frac{1}{2}(\hat{J}+\hat{J}^{\dagger})\) and \(\hat{J}_-=-\frac{1}{2}i(\hat{J}-\hat{J}^{\dagger})\). Equation ([\[reality\]](#reality){reference-type="ref" reference="reality"}) reduces the number of degrees of freedom contained in the expectation values back to the usual canonical value of two. In this counting, the fluctuations on the right of ([\[reality\]](#reality){reference-type="ref" reference="reality"}) are considered fixed. Independent conditions for fluctuations and higher moments arise at higher order by considering operators \[\label{realityOp} \left((\Delta \hat{V})^i (\Delta \hat{J}_+)^j (\Delta\hat{J}_-)^k\right)_{\rm Weyl} (\hat{J}_+^2+\hat{J}_-^2-\hat{V}^2)\quad \mbox{with}\quad i+j+k>0\,,\] whose expectation values vanish. In terms of \(\Delta\)-operators, the reality condition reads \[\begin{aligned} \hat{C}&=& \widehat{\Delta J_+}^2+\widehat{\Delta J_-}^2-\widehat{\Delta V}^2+ 2J_+\widehat{\Delta J_+}+2 J_-\widehat{\Delta J_-}-(2V+\hbar)\widehat{\Delta V}\\ &&+ J_+^2+J_-^2-(V+\hbar/2)^2-\hbar^2/4\\ &\approx& \widehat{\Delta J_+}^2-\Delta(J_+^2) +\widehat{\Delta J_-}^2-\Delta(J_-^2)-\widehat{\Delta V}^2+\Delta(V^2)\\ &&+ 2J_+\widehat{\Delta J_+}+2 J_-\widehat{\Delta J_-}-(2V+\hbar)\widehat{\Delta V} \end{aligned}\] analogous to ([\[CApp\]](#CApp){reference-type="ref" reference="CApp"}). For symmetric reorderings of \(\Delta\)-operators, we then use the relations \[\begin{aligned} &=& i\hbar(\widehat{\Delta J_-}+J_-)\\{} [\widehat{\Delta V},\widehat{\Delta J_-}]&=&-i\hbar(\widehat{\Delta J_+}+J_+)\\{} [\widehat{\Delta J_+},\widehat{\Delta J_-}]&=& -i\hbar(\widehat{\Delta V}+V+\hbar/2)\. \end{aligned}\] For instance, the third-order moments \[\begin{aligned} \Delta(VJ_{\pm}^2) &\equiv& \left\langle ( \hat{V}-V ) ( \hat{J}_{\pm}-J_{\pm})^2 \right\rangle_{\rm Weyl-ordered} \\ \Delta(V^3) &\equiv&\left\langle ( \hat{V}-V)^3 \right\rangle \end{aligned}\] appear in the reality condition \[\begin{aligned} 0&=&\Delta(VJ_+^2)+\Delta(VJ_-^2)-\Delta(V^3)\\ &&+2J_+\Delta(VJ_+)+ 2J_-\Delta(VJ_-)-(2V+\hbar)\Delta(V^2)\\ &&-\frac{1}{6}\hbar^2(2V+\hbar) \end{aligned}\] following from the vanishing expectation value of \(\Delta \hat{V}(\hat{J}_+^2+\hat{J}_-^2-\hat{V}^2)\). Similarly, \[\begin{aligned} 0&=& \Delta(J_+^3)+\Delta(J_+J_-^2)-\Delta(J_+V^2)\\ &&+2J_+\Delta(J_+^2)+2J_-\Delta(J_+J_-)-2V\Delta(VJ_+)-\hbar \Delta(VJ_+) -\frac{\hbar^2}{6}J_+\,,\\ 0&=& \Delta(J_-^3)+\Delta(J_+^2J_-)-\Delta(J_-V^2)\\ &&+2J_-\Delta(J_-^2)+2J_+\Delta(J_+J_-)-2V\Delta(VJ_-)-\hbar \Delta(VJ_-)-\frac{\hbar^2}{6}J_-\. \end{aligned}\] Third-order reality conditions restrict semiclassical second-order moments. To leading order in \(\hbar\) (keeping only the "central charge" \(\hbar/2\) as a higher-order contribution to \(V\)), \[\begin{aligned} (V+\hbar/2)\Delta(V^2) &=& J_+\Delta(VJ_+)+J_-\Delta(VJ_-)\\ (V+\hbar/2)\Delta(VJ_+) &=& J_+\Delta(J_+^2)+J_-\Delta(J_+J_-)\\ (V+\hbar/2)\Delta(VJ_-) &=& J_-\Delta(J_-^2)+J_+\Delta(J_+J_-)\. \end{aligned}\] In terms of moments of complex variables, as derived in and used crucially to restrict initial values of moments for numerical solutions of equations of motion, this reads \[\label{realitySecond1} (V+\hbar/2)(\Delta V)^2={\rm Re}(J^*\Delta(VJ))= {\rm Re} J {\rm Re} \Delta(VJ) + {\rm Im} J {\rm Im} \Delta(VJ)\.\] The remaining third-order relations imply \[\begin{aligned} \label{realitySecond2} (V+\hbar/2) {\rm Re} \Delta(VJ) &=& \frac{1}{2}\left( {\rm Re} J {\rm Re} (\Delta J)^2 + {\rm Im} J {\rm Im} (\Delta J)^2 + {\rm Re} J \Delta(JJ^*) \right)\,, \nonumber \\ (V+\hbar/2){\rm Im} \Delta(VJ) &=& \frac{1}{2}\left( {\rm Re} J {\rm Im} (\Delta J)^2-{\rm Im} J {\rm Re} (\Delta J)^2 + {\rm Im} J \Delta(JJ^*) \right)\. \end{aligned}\] By our general considerations, none of the higher-order conditions restrict second-order moments further. (The lowest-order term given by ([\[CbarKill\]](#CbarKill){reference-type="ref" reference="CbarKill"}) is of order three or higher for Casimir conditions of order four or higher.) Initially, we have six second-order moments. By higher-order reality conditions, they are subject to three further conditions, leaving three degrees of freedom as expected for two fluctuations and one correlation. ## Uncertainty relations and existence of coherent states For the pairs \((\hat{V},\hat{J}_+)\), \((\hat{V},\hat{J}_-)\) and \((\hat{J}_+,\hat{J}_-)\) of self-adjoint operators, we obtain from ([\[Uncertx\]](#Uncertx){reference-type="ref" reference="Uncertx"}): \[\begin{aligned} (\Delta V)^2 (\Delta J_+)^2-\Delta(VJ_+)^2 &\geq& \frac{1}{4} \hbar^2J_-^2\label{uncert1}\\ (\Delta V)^2(\Delta J_-)^2-\Delta(VJ_-)^2 &\geq& \frac{1}{4}\hbar^2J_+^2\label{uncert2}\\ (\Delta J_+)^2(\Delta J_-)^2- \Delta(J_+J_-)^2 &\geq& \frac{1}{4}\hbar^2(V+\hbar/2)^2\. \label{uncert3} \end{aligned}\] Our previous counting argument has shown that one canonical pair and its moments are left after the Casimir condition is imposed. As in the general case of Casimir conditions, the reality conditions ensure that two of the uncertainty relations are indeed equivalent to the third one. A state saturating the uncertainty relation and obeying the Casimir condition is a coherent state for the group that provides the phase space. If we solve the saturated uncertainty relation (or its higher-order analogs) that remains after the Casimir condition has been used, we obtain moments that could belong to a coherent state, but it is not obvious that there is an actual (normalizable) wave function for them. In fact, constructing explicit wave functions for coherent states can be a complicated procedure. Fortunately, in a concrete example, one can show, without constructing the actual wave function, that there is always a wave function that produces "coherent" moments obtained by solving saturated uncertainty relations. To show this, we begin in the standard way used in quantum mechanics, where a Gaussian is obtained as the unique wave function saturating the canonical uncertainty relation. For our two basic operators \(\hat{V}\) and \(\hat{J}_+\) and some state \(\psi\) we introduce, as before, two states \(v:=(\hat{V}-V)\psi\) and \(w:=(\hat{J}_+-J_+)\psi\), with expectation values taken in the same state \(\psi\). From standard arguments it then follows that the \((V,J_+)\)-uncertainty relation is saturated if and only if \(u:=w-(v\cdot w/||v||^2)v\) vanishes. In our example, this equation reads \[\label{psiSat} \hat{J}_+\psi- \frac{\Delta(VJ_+)+\frac{1}{2}\langle[\hat{V},\hat{J}_+]\rangle}{(\Delta V)^2} \hat{V}\psi+ \left(\frac{\Delta(VJ_+)+\frac{1}{2}\langle[\hat{V},\hat{J}_+]\rangle}{(\Delta V)^2} V-J_+\right) \psi=0\.\] We may represent states as wave functions \(\psi=\sum_n \psi_n|n\rangle\) in terms of \(\hat{V}\)-eigenstates \(|n\rangle\) with \(\hat{V}|n\rangle=n|n\rangle\) (assuming the discrete series of representations, which is relevant for the quantum-cosmological application). On these eigenstates, using the realization \(\hat{J}=\exp(-i\hat{P})\) in terms of canonical operators \((\hat{V},\hat{P})\), \(\hat{J}|n\rangle=(n+1)|n+1\rangle\) acts like the product of \(\hat{V}\) with a raising operator, and the Hermitian combination as \[\hat{J}_+|n\rangle= \frac{1}{2}\left((n+1)|n+1\rangle+ n|n-1\rangle\right)\.\] The saturation equation ([\[psiSat\]](#psiSat){reference-type="ref" reference="psiSat"}) for a wave function \(\psi\) is therefore equivalent to a difference equation \[\label{Diff} \frac{1}{2}\left(n\psi_{n-1}+(n+1)\psi_{n+1}\right) -\alpha n\psi_n+\beta\psi_n=0\] where we denoted the coefficients in ([\[psiSat\]](#psiSat){reference-type="ref" reference="psiSat"}) as \[\label{alpha} \alpha=\frac{\Delta(VJ_+)+ \frac{1}{2}\langle[\hat{V},\hat{J}_+]\rangle}{(\Delta V)^2} =\frac{\Delta(VJ_+)+\frac{1}{2}i\hbar J_-}{(\Delta V)^2}\] and \[\beta=\alpha V-J_+\.\] Since expectation values and moments in \(\alpha\) and \(\beta\) themselves depend on \(\psi\), and therefore on all \(\psi_n\), the system of coupled equations defined by ([\[Diff\]](#Diff){reference-type="ref" reference="Diff"}) is non-linear and difficult to deal with exactly. For our purposes, however, it is sufficient to analyze the asymptotic behavior for \(n\to\pm\infty\), which is feasible. The range of \(n\) includes all integers, and therefore suitable fall-off conditions must be satisfied at both ends for a normalizable state in \(\ell^2\). However, it is sufficient to show that ([\[Diff\]](#Diff){reference-type="ref" reference="Diff"}) has at least one normalizable solution for large positive and negative values of \(n\), respectively, because one can always patch together a normalizable solution on \((-\infty,-1]\) with one on \((1,\infty]\) by choosing \(\psi_0\) so that ([\[Diff\]](#Diff){reference-type="ref" reference="Diff"}) holds for \(n=0\). For the existence of normalizable solutions at large \(n\), we consider the difference equation \(\psi_{n+1}-2\alpha\psi_n+ \psi_{n-1}=0\) with constant coefficients. Its solutions are \(\psi^{\pm}_n=k_{\pm}^n\) with \(k_{\pm}=\alpha\pm\sqrt{\alpha^2-1}\). Solutions \(\psi_{\pm}\) are normalizable if \(|k_{\pm}|<1\). These two numbers satisfy the relation \(k_+k_-=1\), and therefore there can be at most one normalizable solution. There is no normalizable solution if and only if both \(k_{\pm}\) lie on the unit circle, in which case \(k_+=k_-^*\). The latter condition can be fulfilled only if \(\alpha\) is real with \(|\alpha|\leq 1\). In all other cases, a normalizable solution exists and we are guaranteed to have a state saturating the uncertainty relation. Here, ([\[alpha\]](#alpha){reference-type="ref" reference="alpha"}) is not real unless \(J_-=0\), which is generic enough to conclude the existence of coherent states. Solving effective equations for moments produces their quantum parameters, even if an explicit wave function is unknown. In the present example, we can use these results to extend, where wave functions for \({\rm sl}(2,{\mathbb R})\)-coherent states were obtained for small correlations. The moments of group coherent states found here then allow us to address cosmic forgetfulness, for which potentially large correlations and squeezing are important. ## Moments of dynamical coherent states For simple Hamiltonian operators, one can find dynamical coherent states that saturate the uncertainty relation at all times. Harmonic loop quantum cosmology, based on the non-canonical algebra ([\[comm\]](#comm){reference-type="ref" reference="comm"}) with the Hamiltonian \(\hat{H}=\hat{J}_-\), provides such an example. Exact solutions for expectation values and moments in arbitrary states can first be found, and then restricted to those that saturate the uncertainty relation. The dynamical equations are slightly more compact if we use complex variables \((V,J,J^*)\) and the corresponding moments. Evolution in some time parameter \(\lambda\) is generated by \(\hat{H}\). As per Ehrenfest's equations, expectation values obey \[\begin{aligned} \label{Vdot} \frac{{\rm d}}{{\rm d} \lambda} V&=& \frac{1}{i\hbar}\langle[\hat{V},\hat{H}]\rangle= -\frac{1}{2}(J+J^*)\\ \frac{{\rm d}}{{\rm d} \lambda} J&=& \frac{1}{i\hbar}\langle[\hat{J},\hat{H}]\rangle= -(V+\hbar/2)=\frac{{\rm d}}{{\rm d} \lambda}J^*\,, \label{Jdot} \end{aligned}\] solved by \[\label{VJ} V(\lambda)+\hbar/2=A \cosh(\lambda-\lambda_0)\quad,\quad J(\lambda)=-A\sinh(\lambda-\lambda_0)+iH\] since \({\rm Im}J=J_-=\langle\hat{H}\rangle=:H\), and with integration constants \(A\) and \(\lambda_0\). Second-order moments satisfy the equations of motion \[\begin{aligned} \frac{{\rm d}}{{\rm d}\lambda}(\Delta V)^2 &=&-\Delta(VJ)-\Delta(VJ^*) \label{GVVdot}\\ \frac{{\rm d}}{{\rm d}\lambda}(\Delta J)^2 &=&-2\Delta(VJ)\quad,\quad \frac{{\rm d}}{{\rm d}\lambda}(\Delta J^*)^2 =-2\Delta(VJ^*)\\ \frac{{\rm d}}{{\rm d}\lambda}\Delta(VJ) &=& -\frac{1}{2} (\Delta J)^2-\frac{1}{2}\Delta(JJ^*)-(\Delta V)^2\\ \frac{{\rm d}}{{\rm d}\lambda}\Delta(VJ^*) &=&-\frac{1}{2} (\Delta J^*)^2- \frac{1}{2}\Delta(JJ^*)-(\Delta V)^2\\ \frac{{\rm d}}{{\rm d}\lambda}\Delta(JJ^*) &=& -\Delta(VJ)-\Delta(VJ^*) \label{GJbarJdot} \end{aligned}\] solved by \[\begin{aligned} (\Delta V)^2(\lambda) &=& \frac{1}{2}(c_3e^{-2\lambda}+c_4e^{2\lambda})- \frac{1}{4}(c_1+c_2)\\ (\Delta J)^2(\lambda) &=& \frac{1}{2}(c_3e^{-2\lambda}+c_4e^{2\lambda})+ \frac{1}{4}(3c_2-c_1)- i(c_5e^{\lambda}-c_6e^{-\lambda})\\ (\Delta J^*)^2(\lambda) &=& \frac{1}{2}(c_3e^{-2\lambda}+c_4e^{2\lambda})+ \frac{1}{4}(3c_2-c_1)+ i(c_5e^{\lambda}-c_6e^{-\lambda})\\ \Delta(VJ)(\lambda) &=& \frac{1}{2}(c_3e^{-2\lambda}-c_4e^{2\lambda})+ \frac{i}{2}(c_5e^{\lambda}+c_6e^{-\lambda})\\ \Delta(VJ^*)(\lambda) &=& \frac{1}{2}(c_3e^{-2\lambda}-c_4e^{2\lambda})- \frac{i}{2}(c_5e^{\lambda}+c_6e^{-\lambda})\\ \Delta(JJ^*)(\lambda) &=& \frac{1}{2}(c_3e^{-2\lambda}+c_4e^{2\lambda})+ \frac{1}{4}(3c_1-c_2)\. \end{aligned}\] These equations and some of the following derivations can be found in. We list them here because we will be able to generalize them in the next subsection, for which we have to refer to the older results. Here, \(c_1=-(\Delta V)^2+\Delta(JJ^*)\) which by the reality condition ([\[reality\]](#reality){reference-type="ref" reference="reality"}) equals \[\label{c1} c_1=A^2-H^2\] in terms of \(H=\langle\hat{H}\rangle\) and the integration constant \(A\) in the solutions \(V(\lambda)\) and \(J(\lambda)\) in ([\[VJ\]](#VJ){reference-type="ref" reference="VJ"}). Using \(\hat{H}=-\frac{1}{2}i(\hat{J}-\hat{J}^{\dagger})\) we obtain \[(\Delta H)^2 =-\frac{1}{4}\left((\Delta J)^2-2\Delta(JJ^*)+ (\Delta J^*)^2\right) = \frac{1}{2}(c_1-c_2) \label{c1c2}\.\] The remaining constants are subject to further conditions from reality and uncertainty bounds. From ([\[uncert1\]](#uncert1){reference-type="ref" reference="uncert1"}), requiring saturation, \[\label{Uncert} 4c_3c_4 = H^2\hbar^2+\frac{1}{4}(c_1+c_2)^2 \.\] Additional conditions follow from the other uncertainty relations, but they are equivalent once reality conditions are imposed. The second-order reality conditions ([\[realitySecond1\]](#realitySecond1){reference-type="ref" reference="realitySecond1"}) and ([\[realitySecond2\]](#realitySecond2){reference-type="ref" reference="realitySecond2"}), evaluated for the explicit solutions, are \(\lambda\)-dependent. Comparing coefficients of different powers of \(e^{\lambda}\), we obtain three independent equations \[\begin{aligned} c_5 = \frac{A}{2H}\left(2c_4-{\textstyle\frac{1}{2}}(c_1+c_2)\right)\quad &,&\quad c_6 = \frac{A}{2H}\left(2c_3-{\textstyle\frac{1}{2}}(c_1+c_2)\right) \label{c5c6}\\ c_5+c_6 &=& \frac{H}{A}(c_1-c_2)\. \label{c5plusc6} \end{aligned}\] We use ([\[c5plusc6\]](#c5plusc6){reference-type="ref" reference="c5plusc6"}) to eliminate \(c_5\) and \(c_6\) from the sum of the two equations in ([\[c5c6\]](#c5c6){reference-type="ref" reference="c5c6"}), \[\label{c3plusc4} c_3+c_4= \frac{H}{A}(c_5+c_6)+\frac{1}{2}(c_1+c_2)= \frac{H^2}{A^2} (c_1-c_2)+\frac{1}{2}(c_1+c_2)\,,\] and relate the remaining constants on the right-hand side to state parameters via \[c_1-c_2 = 2(\Delta H)^2 \quad,\quad c_1+c_2= 2c_1-(c_1-c_2)= 2(A^2-H^2-(\Delta H)^2)\] using ([\[c1c2\]](#c1c2){reference-type="ref" reference="c1c2"}) and ([\[c1\]](#c1){reference-type="ref" reference="c1"}). Of particular interest is the asymmetry of volume fluctuations at \(\lambda\to\infty\) compared with \(\lambda\to-\infty\). In the cosmological model based on the present example, this asymmetry corresponds to the relation between fluctuations long before and long after the "bounce" when \(V(\lambda)\) in ([\[VJ\]](#VJ){reference-type="ref" reference="VJ"}) is minimal at \(\lambda=\lambda_0\). It is given by the absolute value of the difference of \(c_3\) and \(c_4\), which we can obtain from the sum ([\[c3plusc4\]](#c3plusc4){reference-type="ref" reference="c3plusc4"}) using ([\[Uncert\]](#Uncert){reference-type="ref" reference="Uncert"}): \[\begin{aligned} \label{c3c4} (c_3-c_4)^2= (c_3+c_4)^2-4c_3c_4&=& \frac{H^4}{A^4}(c_1-c_2)^2+ \frac{H^2}{A^2}(c_1^2-c_2^2)-H^2\hbar^2\\ &=& 4H^2\left(\left(1-\frac{H^2}{A^2}\right)(\Delta H)^2- \frac{\hbar^2}{4}+ \left(\frac{H^2}{A^2}-1\right)\frac{(\Delta H)^4}{A^2}\right)\.\nonumber \end{aligned}\] The solutions of this subsection tell us how moments evolve, starting from some initial values that belong to an initial state. If the initial state saturates the uncertainty relation, it will always do so: the left-hand side of the relation is a Casimir function on the space of moments. If we combine our solutions with the results of the previous subsection, we are guaranteed the existence of dynamical coherent states, for we know that there is a coherent initial state whose moments must evolve as derived here. The question, addressed in, is then in how far the coherence of the state can restrict the asymmetry of fluctuations. ## Beyond coherent states, and cosmic forgetfulness At this stage, we have reproduced the asymmetry derived in from all three uncertainty relations in a different way, using reality conditions and only one of the uncertainty relations. Since we already know that second-order reality conditions reduce the number of uncertainty relations to just one, this result is not surprising. However, the rederivation allows a powerful generalization of the asymmetry formula to all semiclassical states, not just dynamically coherent ones. Our reality conditions are valid provided only that moments of order higher than second are subdominant, which is the most general definition of semiclassical states. The preceding derivation remains intact if we change the equality in ([\[Uncert\]](#Uncert){reference-type="ref" reference="Uncert"}) to an inequality once we depart from dynamical coherent states. In this way, the last formula, ([\[c3c4\]](#c3c4){reference-type="ref" reference="c3c4"}), changes to the inequality \[(c_3-c_4)^2\leq 4H^2\left(\left(1-\frac{H^2}{A^2}\right)(\Delta H)^2- \frac{\hbar^2}{4}+ \left(\frac{H^2}{A^2}-1\right)\frac{(\Delta H)^4}{A^2}\right)\] valid for all semiclassical states (or, more generally, whenever moments of order three and higher can be ignored, even if second-order moments are large compared to expectation values). In particular, the change of volume fluctuations is bounded by the quantum fluctuation \(\Delta H\) of the Hamiltonian. (In the cosmological model, \(H\) corresponds to the momentum of a free, massless scalar used to parameterize time.) The inequality derived here generalizes the identity found in for dynamical coherent states saturating the uncertainty relation. It is also consistent with the bound derived in for a larger class of states. The question of asymmetric volume fluctuations has been raised in and led to a lively debate in the literature on loop quantum cosmology. Even though the asymmetry of volume fluctuations is bounded by energy fluctuations, ensuring that an initial semiclassical state does not develop too-large volume fluctuations, the change of fluctuations may nevertheless be significant. Moreover, it depends sensitively on the initial values. If the moments of a state had to be known for long-term state evolution, the sensitivity and potential significant changes near the turning point of \(\langle\hat{V}\rangle\) render a reliable analysis of the state at very early times practically impossible, implying cosmic forgetfulness. (As noted, the issue is not as relevant in recent versions of the scenario, in which one is required to eliminate deterministic evolution through the "bounce" due to signature change at high density.) The possibility of significant changes in volume fluctuations becomes clear when one looks at the relative change of relative volume fluctuations, for dynamical coherent states given by \[\left|1-\frac{\lim_{\lambda\to-\infty}(\Delta V)^2(\lambda)}{\lim_{\lambda\to\infty} (\Delta V)^2(\lambda)}\right|= \frac{|c_3-c_4|}{c_4}\sim \frac{(\Delta H)^2}{(\Delta V)^2}\.\] If this value is near zero, volume fluctuations are nearly symmetric. However, the ratio on the right-hand side depends more sensitively on the precise state and cannot be restricted to be small without further assumptions. Moreover, in cosmology it is usually a safe assumption that matter behaves more quantum than geometry; thus, the energy fluctuation should be expected to be significantly larger than volume fluctuations and the right-hand side is much larger than one. Large changes of volume fluctuations are not ruled out, implying, together with the large sensitivity to initial values, that the early state cannot be reconstructed precisely even if evolution were deterministic. # Conclusions We have presented several constructions and results for moments of states in quantum systems corresponding to irreducible representations of groups. The Casimir conditions that select a representation can be dealt with using methods for effective first-class constrained systems. In contrast to the general case of first-class constraints, simplifications occur that allowed us to draw conclusions for arbitrary orders in a semiclassical or moment expansion. Our effective methods are then particularly useful to derive representation-independent relationships between moments of group coherent states. Some of our results on uncertainty relations can be applied more generally, even when no Casimir condition is imposed. These methods are useful whenever it is difficult to construct explicit wave functions for group (or other) coherent states. We have shown this in our cosmological example, in which the question of cosmic forgetfulness requires good control on all possible coherent states, not just on those of small squeezing which are easier to construct as wave functions. When no assumption on the amount of squeezing is made, even a coherent state does not allow good control on quantum parameters such as fluctuations compared over long time intervals. As also shown in this example, it is possible to show that coherent states exist even without knowing their wave functions. Effective equations for moments then allow one to compute the corresponding quantum parameters. Results obtained by these effective methods are no less reliable than those of more cumbersome calculations using explicit wave functions.

{'timestamp': '2014-01-22T02:10:19', 'yymm': '1401', 'arxiv_id': '1401.5352', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5352'}

# Introduction Recently the increased interest in the properties of the \(\eta\) and \(\eta'\) meson can be observed due to extensive experimental search of the \(\eta\) and \(\eta'\) bound states performed e.g. at COSY , ELSA , GSI , JINR , JPARC , LPI , and MAMI  as well as intensive theoretical investigations e.g. [@wilkin2; @bass; @eta-prime-mesic-Nagahiro; @Mesic-Bass; @EtaMesic-Hirenzaki; @eta-prime-mesic-Hirenzaki; @eta-prime-mesic-Nagahiro-Oset; @ETA-Friedman-Gal; @ETA-Gal-Cieply; @Wycech-Acta; @WYCECHGREE-GW-zGAla; @Mesic-Kelkar; @Bass10Acta]. Properties of \(\eta'\) in nuclear medium are related with the effects of \(U_A(1)\) anomaly at finite density , which is reflected in the large mass of the \(\eta'\) meson compared to the masses of the other members of the pseudoscalar meson nonet , and with the \(\eta\)-\(\eta'\) mixing . COSY--11 experiment  has provided already an important data for these studies , with the most precise direct measurement of the total width of the \(\eta'\) meson \(\Gamma_{\eta'}~\), and the first rough estimation of the \(\eta'\)-N interaction from the excitation function of the cross section for the \(pp\to pp\eta'\) reaction . Here we describe an analysis of the data used earlier for \(\Gamma_{\eta'}\) determination in view of the extraction of the production cross-section for the \(\eta'\) meson \(\sigma_{pp\to pp\eta'}\) in proton-proton collisions and an update of the \(\sigma_{pp\to pp\eta'}\) values presented previously . # Experiment In the reported measurement the \(\eta'\) meson was produced in proton-proton collisions reaction and its mass was reconstructed based on the momentum vectors of protons taking part in the \(pp~\to~pp\eta^{\prime}\) reaction which was measured at five different beam momenta using the COSY--11 detector setup  installed at the cooler synchrotron COSY  in Research Centre Jülich. The schematic view of the COSY--11 detector setup is presented in Fig. [\[fig:c11setup\]](#fig:c11setup){reference-type="ref" reference="fig:c11setup"}. The collision of a proton from the beam with a proton cluster target may cause an \(\eta'\) meson creation. In that case all outgoing nucleons have been registered by the COSY--11 detectors, whereas for the \(\eta^{\prime}\) meson identification the missing mass technique was applied. The COSY beam momentum and the dedicated zero degree COSY--11 facility enabled the measurement at an excess energy down the fraction of an MeV above the kinematic threshold for the \(\eta^{\prime}\) meson production. Modification of the COSY--11 target system allowed to decrease effective beam momentum spread and therefore enabled precise determination of the access energy \(Q\) with the precision of 0.10 MeV. Good control of the systematic uncertainties was possible due to measurement performed at five different values of \(Q\) and monitoring of the beam and target properties . On the other hand the achieved missing mass resolution in the order of the total width of the \(\eta'\) meson itself  improved significantly the \(\eta'\) production cross section measurement. The number of registered \(\eta'\) mesons was obtained from the missing mass spectra for each \(Q\) value and corrected for the detector geometrical acceptance and registration efficiency. The luminosity value was determined using comparison of the cross section of \(pp\to pp\) reaction determined by EDDA Collaboration  and the number of registered elastically scattered protons. # Results Since \(\sigma_{pp\to pp\eta'}\) measured at COSY--11 was obtained with the luminosity determination based on the EDDA data available that time , we updated these numbers accordingly to superseded data . COSY--11 measurement at \(Q=16.4\) MeV  was already reported with new EDDA data , whereas SPESIII  and DISTO  used different techniques for luminosity determination. The experimental data presented at Fig. [\[fig:cross\]](#fig:cross){reference-type="ref" reference="fig:cross"} are compared to the analytical parametrization derived by Fäldt and Wilkin  which takes into account final state interaction of the protons: \[\sigma_{pp\to pp\eta'}\left(Q\right)=C \frac{Q^2}{m_p p_{LAB}}\frac{1}{\left(1 + \sqrt{ 1 + \frac{Q}{\epsilon}}\right)^2}, \label{eq:colin}\] where \(Q\) denotes the excess energy, \(p_{LAB}\) beam momentum, \(m_p\) proton mass. The parameters \(\epsilon=0.75^{+0.20}_{-0.15}\) MeV and \(C=45^{+10}_{-9}\) mb denote the Coulomb distortion and constant factor, respectively, and have been determined by fitting this formula to the experimental data. Values of \(pp\to pp\eta'\) cross sections determined at COSY--11 are gathered in the Table [1](#tab:updated){reference-type="ref" reference="tab:updated"} apart of the new measurement reported here, which is still in the final stage of the analysis. :::

{'timestamp': '2014-04-15T02:15:22', 'yymm': '1401', 'arxiv_id': '1401.5924', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5924'}

# Introduction Inspired by the hierarchical structure of the visual cortex, recent studies on probabilistic models used deep hierarchical architectures to learn high order statistics of the data . One widely used architecture is a deep believe network (DBN), which is usually trained as stacked restricted Boltzmann machines (RBMs) [@HintonSalakhutdinov-2006a; @BengioLamblinEtAl-2006a; @ErhanBengioEtAl-2010a]. Since the original formulation of RBMs assumes binary input values, the model needs to be modified in order to handle continuous input values. One common way is to replace the binary input units by linear units with independent Gaussian noise, which is known as Gaussian-binary restricted Boltzmann machines (GRBMs) or Gaussian-Bernoulli restricted Boltzmann machines   first proposed by. The training of GRBMs is known to be difficult, so that several modifications have been proposed to improve the training. used a sparse penalty during training, which allowed them to learn meaningful features from natural image patches. trained GRBMs on natural images and concluded that the difficulties are mainly due to the existence of high-frequency noise in the images, which further prevents the model from learning the important structures. illustrated that in terms of likelihood estimation GRBMs are already outperformed by simple mixture models. Other researchers focused on improving the model in the view of generative models [@RanzatoKrizhevskyEtAl-2010a; @RanzatoHinton-2010a; @CourvilleBergstraEtAl-2011a; @LeHeessEtAl-2011a].   suggested that the failure of GRBMs is due to the training algorithm and proposed some modifications to overcome the difficulties encountered in training GRBMs. The studies above have shown the failures of GRBMs empirically, but to our knowledge there is no analysis of GRBMs apart from our preliminary work , which accounts the reasons behind these failures. In this paper, we extend our work in which we consider GRBMs from the perspective of density models, i.e. how well the model learns the distribution of the data. We show that a GRBM can be regarded as a mixture of Gaussians, which has already been mentioned briefly in previous studies [@Bengio-2009a; @TheisGerwinnEtAl-2011a; @CourvilleBergstraEtAl-2011a] but has gone unheeded. This formulation makes clear that GRBMs are quite limited in the way they can represent data. However we argue that this fact does not necessarily prevent the model from learning the statistical structure in the data. We present successful training of GRBMs both on a two-dimensional blind source separation problem and natural image patches, and that the results are comparable to that of independent component analysis (ICA). Based on our analysis we propose several training recipes, which allowed successful and fast training in our experiments. Finally, we discuss the relationship between GRBMs and above mentioned modifications of the model. # Gaussian-binary restricted Boltzmann machines (GRBMs) ## The model A Boltzmann Machine (BM) is a Markov Random Field with stochastic *visible* and *hidden* units, which are denoted as \(\mathbf{X}:=\left( X_1,\ldots,X_M\right)^T\) and \(\mathbf{H}:=\left( H_1, \ldots, H_N\right)^T\), respectively. In general, we use bold letters denote vectors and matrices. The joint probability distribution is defined as \[\begin{aligned} P\left( \mathbf{X},\mathbf{H} \right) &:=& \frac{1}{Z} \, \mathrm{e}^{-\frac{1}{T_0}E(\mathbf{X},\mathbf{H})}, \label{eqn:jointprobOfXH}\\ Z &:=& \int \int{\mathrm{e}^{-\frac{1}{T_0}E\left( \mathbf{x},\mathbf{h} \right)}} \mathrm{d} \mathbf{x}\,\mathrm{d} \mathbf{h} \label{eqn:partitionFunc} \end{aligned}\] where \(E\left( \mathbf{X},\mathbf{H} \right)\) denotes an *energy function* as known from statistical physics, which defines the dependence between \(\mathbf{X}\) and \(\mathbf{H}\). The temperature parameter \(T_0\) is usually ignored by setting its value to one, but it can play an important role in inference of BMs. The *partition function* \(Z\) normalizes the probability distribution by integrating over all possible values of \(\mathbf{X}\) and \(\mathbf{H}\), which is intractable in most cases. So that in training BMs using gradient descent the partition function is usually estimated using sampling methods. However, even sampling in BMs remains difficult due to the dependencies between all variables. An RBM is a special case of a BM where the energy function contains no terms combining two different hidden or two different visible units. Viewed as a graphical model, there are no lateral connections within the visible or hidden layer, which results in a bipartite graph. This implies that the hidden units are conditionally independent given the visibles and vice versa, which allows efficient sampling. The values of the visible and hidden units are usually assumed to be binary, i.e. \(X_m, H_n\in\left\{ 0, 1 \right\}\). The most common way to extend an RBM to continuous data is a GRBM, which assumes continuous values for the visible units and binary values for the hidden units. Its energy function is defined as \[\begin{aligned} E\left( \mathbf{X}, \mathbf{H} \right) : &=& \sum_i^M \frac{\left( X_i-b_i \right)^2}{2\sigma^2} -\sum_j^N c_j H_j -\sum_{i,j}^{M,N} \frac{ X_i w_{ij} H_j }{\sigma^2} \\ \label{eqn:energyFuncGauss1} &=& \frac{|| \mathbf{X}-\mathbf{b} ||^2}{2\sigma^2} -\mathbf{c}^T \mathbf{H} -\frac{ \mathbf{X}^T \mathbf{W} \mathbf{H} }{\sigma^2} , \label{eqn:energyFuncGauss2} \end{aligned}\] where \(||\mathbf{u}||\) denotes the Euclidean norm of \(\mathbf{u}\). In GRBMs the visible units given the hidden values are Gaussian distributed with standard deviation \(\sigma\). Notice that some authors use an independent standard deviation for each visible unit, which comes into account if the data is not whitened.\ The conditional probability distribution of the visible given the hidden units is given by \[\begin{aligned} P\left( \mathbf{X}|\mathbf{h} \right) &=& \frac{P\left( \mathbf{X}, \mathbf{h} \right)} {\int\limits{P\left( \mathbf{x}, \mathbf{h} \right) \mathrm{d} \mathbf{x} }} \\ & \stackrel{(\ref{eqn:jointprobOfXH}, \ref{eqn:energyFuncGauss2})}{=} & \frac{ \mathrm{e}^{\mathbf{c}^T \mathbf{h}} \prod\limits_i^M{ \mathrm{e}^{ \frac{X_i \mathbf{w}_{i *}^T \mathbf{h} }{\sigma^2} -\frac{|| X_i-b_i ||^2}{2 \sigma^2} } } } { \int\limits{\mathrm{e}^{\mathbf{c}^T \mathbf{h}} \prod\limits_i^M{ \mathrm{e}^{ \frac{x_i \mathbf{w}_{i *}^T \mathbf{h} }{\sigma^2} -\frac{|| x_i-b_i ||^2}{2 \sigma^2} } \mathrm{d} \mathbf{x} } } } \\ &=& \prod_i^M{ \frac{ \mathrm{e}^{ \frac{X_i \mathbf{w}_{i *}^T \mathbf{h} }{\sigma^2} -\frac{|| X_i-b_i ||^2}{2 \sigma^2} } }{ \int\limits \mathrm{e}^{ \frac{x_i \mathbf{w}_{i *}^T \mathbf{h} }{\sigma^2} -\frac{|| x_i-b_i ||^2}{2 \sigma^2} }\mathrm{d}x_i } } \label{eqn:probOfXH1} \\ &\stackrel{(\ref{eqn:quadraticExp})}{=}& \prod_i^M{ \frac{ \mathrm{e}^{ -\frac{|| X_i-b_i-\mathbf{w}_{i *}^T \mathbf{h} ||^2}{2 \sigma^2} } }{ \int\limits \mathrm{e}^{ -\frac{|| x_i-b_i-\mathbf{w}_{i *}^T \mathbf{h} ||^2}{2 \sigma^2} }\mathrm{d}x_i } } \label{eqn:probOfXH2} \\ &=& \prod_i^M{ \underbrace{ \mathcal{N}\left( X_i;b_i + \mathbf{w}_{i *}^T \mathbf{h}, \sigma^2 \right) }_{=\,P(X_i|\mathbf{h})} } \label{eqn:probOfXHGauss} \\ &=& \mathcal{N}\left( \mathbf{X};\mathbf{b} + \mathbf{W} \mathbf{h}, \sigma^2 \right), \label{eqn:probOfXHGaussVector} \end{aligned}\] where \(\mathbf{w}_{i*}\) and \(\mathbf{w}_{* j}\) denote the \(i\)th row and the \(j\)th column of the weight matrix, respectively. \(\mathcal{N}\left( x; \mathbf{\mu}, \sigma^2 \right)\) denotes a Gaussian distribution with mean \(\mu\) and variance \(\sigma^2\). And \(\mathcal{N}\left( \mathbf{X};\boldsymbol\mu, \sigma^2 \right)\) denotes an isotropic multivariate Gaussian distribution centered at vector \(\boldsymbol\mu\) with variance \(\sigma^2\) in all directions. From ([\[eqn:probOfXH1\]](#eqn:probOfXH1){reference-type="ref" reference="eqn:probOfXH1"}) to ([\[eqn:probOfXH2\]](#eqn:probOfXH2){reference-type="ref" reference="eqn:probOfXH2"}) we used the relation \[\begin{aligned} \frac{ax}{\sigma^2}-\frac{\left( x-b \right)^2}{2\sigma^2} &=& \frac{-x^2+2bx+2ax-b^2}{2\sigma^2} \nonumber\\ &=& \frac{-x^2+2bx+2ax-b^2+a^2-a^2+2ab-2ab}{2\sigma^2} \nonumber\\ &=& \frac{-(x-a-b)^2+a^2+2ab}{2\sigma^2}. \label{eqn:quadraticExp} \end{aligned}\] The conditional probability distribution of the hidden units given the visibles can be derived as follows \[\begin{aligned} P(\mathbf{H}|\mathbf{x}) &=& \frac{P(\mathbf{x},\mathbf{H})}{\sum\limits_{\mathbf{h}} P(\mathbf{x}, \mathbf{h})} \\ &\stackrel{(\ref{eqn:jointprobOfXH}, \ref{eqn:energyFuncGauss2})}{=}& \frac{ \mathrm{e}^{-\frac{||\mathbf{x}-\mathbf{b}||^2}{2 \sigma^2}} \prod\limits_j^N{ \mathrm{e}^{ \left( c_j + \frac{\mathbf{x}^T \mathbf{w}_{* j}}{\sigma^2} \right)H_j } } }{\sum\limits_{\mathbf{h}}{ \mathrm{e}^{-\frac{||\mathbf{x}-\mathbf{b}||^2}{2 \sigma^2}} \prod\limits_j^N{ \mathrm{e}^{ \left( c_j + \frac{\mathbf{x}^T \mathbf{w}_{* j}}{\sigma^2} \right)h_j } } } } \\ &=& \prod_j^N{ \underbrace{ \frac{\mathrm{e}^{ \left( c_j + \frac{\mathbf{x}^T \mathbf{w}_{* j}}{\sigma^2} \right)H_j } }{\sum\limits_{h_j}{ \mathrm{e}^{ \left( c_j + \frac{\mathbf{x}^T \mathbf{w}_{* j}}{\sigma^2} \right)h_j } } } }_{=\,P(H_j|\mathbf{x})} }. \\ \Longrightarrow \quad P(H_j=1|\mathbf{x}) &=& \frac{1}{1+\mathrm{e}^{-\left(c_j + \frac{\mathbf{x}^T \mathbf{w}_{* j}}{\sigma^2}\right) }} \label{eqn:probOfHXGauss} \end{aligned}\] \(P(\mathbf{H}|\mathbf{x})\) turns out to be a product of independent sigmoid functions, which is a frequently used non-linear activation function in artificial neural networks. ## Maximium likelihood estimation {#sec:trainingAlgorithm} Maximum likelihood estimation (MLE) is a frequently used technique for training probabilistic models like BMs. In MLE we have a data set \(\mathcal{\tilde{X}}=\left\{ \tilde{\mathbf{x}}_1, \ldots, \tilde{\mathbf{x}}_L \right\}\) where the observations \(\tilde{\mathbf{x}}_l\) are assumed to be independent and identically distributed (i.i.d.). The goal is to find the optimal parameters \(\tilde{\boldsymbol\Theta}\) that maximize the likelihood of the data, i.e. maximize the probability that the data is generated by the model . For practical reasons one often considers the logarithm of the likelihood, which has the same maximum as the likelihood since it is a monotonic function. The log-likelihood is defined as \[\ln P( \mathcal{\tilde{X}}; \boldsymbol\Theta ) = \ln{\prod_{l=1}^L{P\left( \tilde{\mathbf{x}}_l; \boldsymbol\Theta \right)}} = \sum_{l=1}^L{\ln{P\left( \tilde{\mathbf{x}}_l; \boldsymbol\Theta \right)}}. \label{}\] We use the average log-likelihood per training case denoted by \(\hat\ell\). For RBMs it is defined as \[\hat\ell := \left< \ln P( \mathcal{\tilde{X}}; \boldsymbol\Theta ) \right>_{\tilde{\mathbf{x}}} = \left< \ln \left( \sum_{\mathbf{h}} { \mathrm{e}^{-E \left( \tilde{\mathbf{x}},\mathbf{h} \right) } } \right) \right>_{\tilde{\mathbf{x}}}-\ln Z, \label{eqn:loglikelihood}\] where \(\tilde{\mathbf{x}} \in \mathcal{\tilde{X}}\). And \(\left< f(u) \right>_u\) denotes the expectation of the function \(f(u)\) with respect to variable \(u\). The gradient of the \(\hat\ell\) turns out to be the difference between the expectations of the energies gradient under the data and model distribution, which is given by \[\begin{aligned} \frac{ \partial \hat\ell}{ \partial \theta } &\stackrel{(\ref{eqn:loglikelihood}, \ref{eqn:partitionFunc})}{=}& \left< \sum\limits_{\mathbf{h}}{ \frac{ \frac{\mathrm{e}^{-E\left( \tilde{\mathbf{x}},\mathbf{h} \right)}}{Z} }{ \sum\limits_{\mathbf{h'}}{ \frac{\mathrm{e}^{-E \left( \tilde{\mathbf{x}},\mathbf{h'} \right)}}{Z} } } \left(-\frac{ \partial E \left( \tilde{\mathbf{x}}, \mathbf{h} \right) }{ \partial \theta } \right) } \right>_{\tilde{\mathbf{x}}}- \frac{1}{Z} \sum\limits_{\mathbf{h}} \sum\limits_{\mathbf{x}} { \mathrm{e}^{-E \left( \mathbf{x}, \mathbf{h} \right) }\left( -\frac{ \partial E \left( \mathbf{x},\mathbf{h} \right) }{ \partial \theta } \right) } \\ &\stackrel{(\ref{eqn:jointprobOfXH})}{=}& -\left< \sum_{\mathbf{h}}{ P\left( \mathbf{h}|\tilde{\mathbf{x}} \right) \frac{ \partial E \left( \tilde{\mathbf{x}},\mathbf{h} \right) }{ \partial \theta } } \right>_{\tilde{\mathbf{x}}} + \left< \sum_{\mathbf{h}}{ P\left( \mathbf{h}|\mathbf{x} \right) \frac{ \partial E \left( \mathbf{x},\mathbf{h} \right) }{ \partial \theta } } \right>_{\mathbf{x}}. \label{eqn:loglikelihooddiff} \end{aligned}\] In practice, a finite set of i.i.d. samples can be used to approximate the expectations in ([\[eqn:loglikelihooddiff\]](#eqn:loglikelihooddiff){reference-type="ref" reference="eqn:loglikelihooddiff"}). While we can use the training data to estimate the first term, we do not have any i.i.d. samples from the unknown model distribution to estimate the second term. Since we are able to compute the conditional probabilities in RBMs efficiently, Gibbs sampling can be used to generate those samples. But Gibbs-sampling only guarantees to generate samples from the model distribution if we run it infinite long. As this is impossible, a finite number of \(k\) sampling steps are used instead. This procedure is known as Contrastive Divergence - \(k\) (CD-\(k\)) algorithm, in which even \(k=1\) shows good results . The CD-gradient approximation is given by \[\frac{ \partial \hat\ell }{ \partial \theta } \approx -\left< \sum_{\mathbf{h}}{ P\left( \mathbf{h}|\tilde{\mathbf{x}} \right) \frac{ \partial E \left( \tilde{\mathbf{x}} ,\mathbf{h} \right) }{ \partial \theta } } \right>_{ \tilde{\mathbf{x}} } + \left< \sum_{\mathbf{h}}{ P( \mathbf{h}|\mathbf{x}^k ) \frac{ \partial E \left( \mathbf{x}^{(k)},\mathbf{h} \right) }{ \partial \theta } } \right>_{\mathbf{x}^{(k)}}, \\ \label{eqn:loglikelihooddiffapprox1}\] where \(\mathbf{x}^{(k)}\) denotes the samples after \(k\) steps of Gibbs sampling. The derivatives of the GRBM's energy function with respect to the parameters are given by \[\begin{aligned} \frac{ \partial E\left( \mathbf{X}, \mathbf{H} \right) }{ \partial \mathbf{b} } &=& -\frac{\mathbf{X}-\mathbf{b}}{\sigma^2}, \\ \label{eqn:energydiffvisbiasGauss} \frac{ \partial E\left( \mathbf{X}, \mathbf{H} \right) }{ \partial \mathbf{c} } &=& -\mathbf{H}, \\ \label{eqn:energydiffhidbiasGauss} \frac{ \partial E\left( \mathbf{X}, \mathbf{H} \right) }{ \partial \mathbf{W} } &=& -\frac{ \mathbf{X} \mathbf{H}^T }{\sigma^2}, \label{eqn:energydiffweightGauss}\\ \frac{ \partial E \left( \mathbf{X},\mathbf{H} \right) }{ \partial\,\mathbf{\sigma} } &=& -\frac{ || \mathbf{X}-\mathbf{b} ||^2 }{ \sigma^3 } + \frac{ 2 \, \mathbf{X}^T\mathbf{W}\mathbf{H} }{ \sigma^3 }, \label{eqn:energydiffvarianceGauss} \end{aligned}\] and the corresponding gradient approximations ([\[eqn:loglikelihooddiffapprox1\]](#eqn:loglikelihooddiffapprox1){reference-type="ref" reference="eqn:loglikelihooddiffapprox1"}) become \[\begin{aligned} \frac{ \partial \hat\ell }{ \partial \mathbf{b} } & \approx & \left<\frac{ \tilde{\mathbf{x}}-\mathbf{b} }{\sigma^2}\right>_{\tilde{\mathbf{x}}} -\left<\frac{ \mathbf{x}^{(k)}-\mathbf{b} }{\sigma^2}\right>_{\mathbf{x}^{(k)}} , \label{eqn:loglikelihooddiffvisbiasGauss}\\ \frac{ \partial \hat\ell }{ \partial \mathbf{c} } & \approx & \left< P\left( \mathbf{h} = \mathbf{1}|\tilde{\mathbf{x}} \right) \right>_{\tilde{\mathbf{x}}} -\left< P\left( \mathbf{h} = \mathbf{1}|\mathbf{x}^{(k)} \right) \right>_{\mathbf{x}^{(k)}} , \label{eqn:loglikelihooddiffhidbiasGauss}\\ \frac{ \partial \hat\ell }{ \partial \mathbf{w} } & \approx & \left< \frac{\tilde{\mathbf{x}}P\left( \mathbf{h} = \mathbf{1}|\tilde{\mathbf{x}} \right)^T}{\sigma^2} \right>_{\tilde{\mathbf{x}}}- \left< \frac{\mathbf{x}^{(k)}P\left( \mathbf{h} = \mathbf{1}|\mathbf{x}^{(k)} \right)^T}{\sigma^2} \right>_{\mathbf{x}^{(k)}} , \label{eqn:loglikelihooddiffweightGauss}\\ \frac{ \partial \hat\ell }{ \partial\,\mathbf{\sigma} } & \approx & \left< \frac{ || \tilde{\mathbf{x}}-\mathbf{b} ||^2 - 2 \, \tilde{\mathbf{x}}^T\mathbf{W}\, P\left( \mathbf{h} = \mathbf{1}|\tilde{\mathbf{x}} \right) }{\sigma^3} \right>_{ \tilde{\mathbf{x}} } \\ \nonumber &&-\left< \frac{ || \mathbf{x}^{(k)}-\mathbf{b} ||^2 - 2 \, {\mathbf{x}^{(k)}}^T\mathbf{W}\, P\left( \mathbf{h} = \mathbf{1}|\mathbf{x}^{(k)} \right) }{\sigma^3} \right>_{ \mathbf{x}^{(k)} }, \label{eqn:loglikelihooddiffvarianceGauss} \end{aligned}\] where \(P\left( \mathbf{h} = 1|\mathbf{x} \right) := (P\left( h_1 = 1 | \mathbf{x} \right), \cdots, P\left( h_N = 1 | \mathbf{x} \right))^T\), i.e. \(P\left( \mathbf{h} = 1|\mathbf{x} \right)\) denotes a vector of probabilities. ## The marginal probability distribution of the visible units From the perspective of density estimation, the performance of the model can be assessed by examining how well the model estimates the data distribution. We therefore take a look at the model's marginal probability distribution of the visible units, which can be formalized as a product of experts (PoE) or as a mixture of Gaussians (MoG))[^1]. ### In the Form of Product of Experts We derive the marginal probability distribution of the visible units \(P(\mathbf{X})\) by factorizing the joint probability distribution over the hidden units. \[\begin{aligned} P\left( \mathbf{X} \right) &=& \sum\limits_{ \mathbf{h} }{ P\left( \mathbf{X},\mathbf{h} \right) } \\ & \stackrel{(\ref{eqn:jointprobOfXH},\ref{eqn:energyFuncGauss1})}{=} & \frac{1}{Z} \,\mathrm{e}^{ -\frac{|| \mathbf{X}-\mathbf{b} ||^2}{2\sigma^2} } \prod_j^N{\sum_{h_j}{ \mathrm{e}^{ c_j+\frac{\mathbf{X}^T\mathbf{w}_{* j} }{\sigma^2} h_j } } } \\ &\stackrel{h_j\in \left\{ 0,1 \right\}}{=}& \frac{1}{Z} \prod_j^N{\left( \,\mathrm{e}^{-\frac{|| \mathbf{X}-\mathbf{b} ||^2}{2N\sigma^2}} + \mathrm{e}^{ c_j+\frac{\mathbf{X}^T\mathbf{w}_{* j} }{\sigma^2} -\frac{|| \mathbf{X}-\mathbf{b} ||^2}{2N\sigma^2}} \right)} \label{eqn:probOfXGauss1} \\ &\stackrel{(\ref{eqn:quadraticExp})}{=}& \frac{1}{Z} \prod_j^N \left( \,\mathrm{e}^{-\frac{|| \mathbf{X}-\mathbf{b} ||^2}{2N\sigma^2}} + \mathrm{e}^{\frac{||\mathbf{b}+N \mathbf{w}_{* j}||^2 -||\mathbf{b}||^2}{2N\sigma^2} + c_j-\frac{|| \mathbf{X}-\mathbf{b}-N \mathbf{w}_{* j} ||^2}{2N\sigma^2}}\right) \\ &=& \frac{1}{Z} \prod_j^N{\left( \sqrt{2\pi N \sigma^2} \right)^{M}} \Big[ \mathcal{N}\left( \mathbf{X};\mathbf{b}, N\sigma^2 \right) \nonumber \\ && +\,\mathrm{e}^{\frac{||\mathbf{b}+N \mathbf{w}_{* j}||^2 -||\mathbf{b}||^2}{2N\sigma^2} + c_j} \mathcal{N}\left( \mathbf{X};\mathbf{b}+N\mathbf{w}_{* j}, N\sigma^2 \right) \Big] \\ &=:&\frac{1}{Z} \prod_j^N{p_j\left( \mathbf{X} \right)} \label{eqn:probOfXGauss2}. \end{aligned}\] Equation ([\[eqn:probOfXGauss2\]](#eqn:probOfXGauss2){reference-type="ref" reference="eqn:probOfXGauss2"}) illustrates that \(P(\mathbf{X})\) can be written as a product of \(N\) factors, referred to as a product of experts . Each expert \(p_j(\mathbf{X})\) consists of two isotropic Gaussians with the same variance \(N\sigma^2\). The first Gaussian is placed at the visible bias \(\mathbf{b}\). The second Gaussian is shifted relative to the first one by \(N\) times the weight vector \(\mathbf{w}_{* j}\) and scaled by a factor that depends on \(\mathbf{w}_{* j}\) and \(\mathbf{b}\). Every hidden unit leads to one expert, each mode of which corresponds to one state of the corresponding hidden unit. Figure [\[fig:GRBMPdf\]](#fig:GRBMPdf){reference-type="ref" reference="fig:GRBMPdf"} (a) and (b) illustrate \(P(\mathbf{X})\) of a GRBM-2-2 viewed as a PoE, where GRBM-\(M\)-\(N\) denotes a GRBM with \(M\) visible and \(N\) hidden units. ### In the Form of Mixture of Gaussians {#sec:lowerOrderComp} Using Bayes'theorem, the marginal probaility of \(\mathbf{X}\) can also be formalized as: \[\begin{aligned} P\left( \mathbf{X} \right) &=& \sum\limits_{ \mathbf{h} }{ P\left( \mathbf{X}|\mathbf{h} \right) P(\mathbf{h}) } \\ \label{eqn:probOfXComponent} & \stackrel{}{=} & \sum\limits_{\mathbf{h}}{\mathcal{N}\left( \mathbf{X};\mathbf{b}+\mathbf{W}\mathbf{h}, \sigma^2 \right) \frac{\left( \sqrt{2\pi\sigma^2} \right)^M}{Z}\, \mathrm{e}^{\mathbf{c}^T \mathbf{h} + \frac{|| \mathbf{b} + \mathbf{W} \mathbf{h}||^2-||\mathbf{b}||^2}{2\sigma^2}} } \label{eqn:probOfXComponent1}\\ & \stackrel{}{=} & \underbrace{ \frac{\left( \sqrt{2\pi\sigma^2} \right)^{M}}{Z} }_{P(\mathbf{h}: \mathbf{h} \in \mathcal{H}_0)}\, \mathcal{N}\left( \mathbf{X};\mathbf{b},\sigma^2 \right) \nonumber \\ && + \sum_{j=1}^N{ \underbrace{ \frac{\left( \sqrt{2\pi\sigma^2} \right)^{M}}{Z} \, \mathrm{e}^{\frac{||\mathbf{b}+\mathbf{w}_{* j}||^2-||\mathbf{b}||^2}{2\sigma^2} + c_j} }_{P(\mathbf{h}_j: \mathbf{h}_j \in \mathcal{H}_1)} \mathcal{N}\left( \mathbf{X};\mathbf{b}+\mathbf{w}_{* j}, \sigma^2 \right)} \nonumber \\ && + \sum_{j=1}^{N-1}{ \sum_{k>j}^N{ \underbrace{ \frac{\left( \sqrt{2\pi\sigma^2} \right)^{M}}{Z} \, \mathrm{e}^{\frac{||\mathbf{b}+\mathbf{w}_{* j}+\mathbf{w}_{* k}||^2-||\mathbf{b}||^2}{2\sigma^2} + c_j + c_k} }_{P(\mathbf{h}_{jk}: \mathbf{h}_{jk} \in \mathcal{H}_2)} \mathcal{N}\left( \mathbf{X};\mathbf{b}+\mathbf{w}_{* j}+\mathbf{w}_{* k}, \sigma^2 \right) } } \nonumber \\ && + \ldots, \label{eqn:probOfXExpandGauss} \end{aligned}\] where \(\mathcal{H}_k\) denotes the set of all possible binary vectors with exactly \(k\) ones and \(M-k\) zeros respectively. As an example, \(\sum_{j=1}^{N-1}\sum_{k>j}^{N}P(\mathbf{h}_{jk}:\mathbf{h}_{jk} \in \mathcal{H}_{2}) = \sum_{\mathbf{h}\in\mathcal{H}_2}{P}(\mathbf{h})\) sums over the probabilities of all binary vectors having exactly two entries set to one. \(P(\mathbf{H})\) in ([\[eqn:probOfXComponent1\]](#eqn:probOfXComponent1){reference-type="ref" reference="eqn:probOfXComponent1"}) is derived as follows \[\begin{aligned} P(\mathbf{H}) &=& \int\limits{P\left( \mathbf{x}, \mathbf{H} \right) \mathrm{d} \mathbf{x} } \\ & \stackrel{(\ref{eqn:jointprobOfXH},\ref{eqn:energyFuncGauss2})}{=} & \frac{1}{Z} \int\limits{\mathrm{e}^{\mathbf{c}^T \mathbf{H}} \prod\limits_i^M{ \mathrm{e}^{ \frac{x_i \mathbf{w}_{i *}^T \mathbf{H} }{\sigma^2} -\frac{|| x_i-b_i ||^2}{2 \sigma^2} } \mathrm{d} \mathbf{x} } } \\ &=& \frac{\mathrm{e}^{\mathbf{c}^T \mathbf{H}}}{Z} \prod\limits_i^M{ \int\limits{ \mathrm{e}^{ \frac{x_i \mathbf{w}_{i *}^T \mathbf{H} }{\sigma^2} -\frac{|| x_i-b_i ||^2}{2 \sigma^2} } \mathrm{d} x_i } } \\ &\stackrel{(\ref{eqn:quadraticExp})}{=}& \frac{\mathrm{e}^{\mathbf{c}^T \mathbf{H}}}{Z} \prod\limits_i^M{\left( \mathrm{e}^{\frac{ (b_i+w_{i *}^T\mathbf{H})^2-b_i^2}{2\sigma^2}} \int\limits{ \mathrm{e}^{ \frac{|| x_i-b_i-\mathbf{w}_{i *}^T \mathbf{H} ||^2}{2\sigma^2} } \mathrm{d} x_i } \right)} \\ &=& \frac{\mathrm{e}^{\mathbf{c}^T \mathbf{H}}}{Z} \left( \sqrt{2\pi\sigma^2} \right)^M \mathrm{e}^{\sum\limits_i^M{\frac{( b_i + \mathbf{w}_{i *}^T \mathbf{H})^2-b_i^2}{2\sigma^2}} } \\ &=& \frac{\left( \sqrt{2\pi\sigma^2} \right)^M}{Z}\, \mathrm{e}^{\mathbf{c}^T \mathbf{H} + \frac{|| \mathbf{b} + \mathbf{W} \mathbf{H}||^2-||\mathbf{b}||^2}{2\sigma^2}} \label{eqn:probOfHGauss} \end{aligned}\] Since the form in ([\[eqn:probOfXExpandGauss\]](#eqn:probOfXExpandGauss){reference-type="ref" reference="eqn:probOfXExpandGauss"}) is similar to a mixture of isotropic Gaussians, we follow its naming convention. Each Gaussian distribution is called a *component* of the model distribution, which is exactly the conditional probability of the visible units given a particular state of the hidden units. As well as in MoGs, each component has a *mixing coefficient*, which is the marginal probability of the corresponding state and can also be viewed as the prior probability of picking the corresponding component. The total number of components in a GRBM is \(2^N\), which is exponential in the number of hidden units, see Figure [\[fig:GRBMPdf\]](#fig:GRBMPdf){reference-type="ref" reference="fig:GRBMPdf"} (c) for an example. The locations of the components in a GRBM are not independent of each other as it is the case in MoGs. They are centered at \(\mathbf{b}+\mathbf{W}\mathbf{h}\), which is the vector sum of the visible bias and selected weight vectors. The selection is done by the corresponding entries in \(\mathbf{h}\) taking the value one. This implies that only the \(M+1\) components that sum over exactly one or zero weights can be placed and scaled independently. We name them first order components and the anchor component respectively. All \(2^N-M-1\) higher order components are then determined by the choice of the anchor and first order components. This indicates that GRBMs are constrained MoGs with isotropic components. # Experiments ## Two-dimensional blind source separation The general presumption in the analysis of natural images is that they can be considered as a mixture of independent super-Gaussian sources, but see for an analysis of remaining dependencies. In order to be able to visualize how GRBMs model natural image statistics, we use a mixture of two independent Laplacian distributions as a toy example. The independent sources \(\mathbf{s}=\left( s_1, s_2 \right)^T\) are mixed by a random mixing matrix \(\mathbf{A}\) yielding \[\tilde{\mathbf{x}}' = \mathbf{A} \mathbf{s}, \label{eqn:MixData}\] where \(p\left( s_i \right) = \frac{\mathrm{e}^{-\sqrt{2}|s_i|}}{\sqrt{2}}\). It is common to whiten the data (see Section [4.1](#sec:preprocessing){reference-type="ref" reference="sec:preprocessing"}), resulting in \[\tilde{\mathbf{x}} = \mathbf{V} \tilde{\mathbf{x}}' = \mathbf{V} \mathbf{A} \mathbf{s}, \label{eqn:MixDataWht}\] where \(\mathbf{V}={\left< \tilde{\mathbf{x'}}\tilde{\mathbf{x'}}^T\right> }^ {-\frac{1}{2}}\) is the whitening matrix calculated with principle component analysis (PCA). Through all this paper, we used the whitened data. In order to assess the performance of GRBMs in modeling the data distribution, we ran the experiments for \(200\) times and calculated the \(\hat\ell\) for test data analytically. For comparision, we also calculated the \(\hat\ell\) over the test data for ICA[^2], an isotropic two-dimensional Gaussian distribution and the true data distribution[^3]. The results are presented in Table [1](#tab:BlindSrcSept){reference-type="ref" reference="tab:BlindSrcSept"}, which confirm the conclusion of  that GRBMs are not as good as ICA in terms of \(\hat\ell\). # Discussion The difficulties of using GRBMs for modeling natural images have been reported by several authors  and various modifications have been proposed to address this problem. analyzed the problem from the view of generative models and argued that the failure of GRBMs is due to the model's focus on predicting the mean intensity of each pixel rather than the dependence between pixels. To model the covariance matrices at the same time, they proposed the mean-covariance RBM (mcRBM). In addition to the conventional hidden units \(\mathbf{h}^m\), there is a group of hidden units \(\mathbf{h}^c\) dedicated to model the covariance between the visible units. From the view of density models, mcRBMs can be regarded as improved GRBMs such that the additional hidden units are used to depict the covariances. The conditional probabilities of mcRBM are given by \[P(\mathbf{X}|\mathbf{h}^m, \mathbf{h}^c) = \mathcal{N}\left( \mathbf{X}; \boldsymbol\Sigma \,\mathbf{W}\,\mathbf{h}^m , \boldsymbol\Sigma\right), \label{eqn:ProbmcRBM}\] where \(\boldsymbol\Sigma = \left( \mathbf{C}\,diag(\mathbf{P}\mathbf{h}^c)\,\mathbf{C}^T \right)^{-1}\) . By comparing ([\[eqn:ProbmcRBM\]](#eqn:ProbmcRBM){reference-type="ref" reference="eqn:ProbmcRBM"}) and ([\[eqn:probOfXHGauss\]](#eqn:probOfXHGauss){reference-type="ref" reference="eqn:probOfXHGauss"}), it can be seen that the components of mcRBM can have a covariance matrix that is not restricted to be diagonal as it is the case for GRBMs. From the view of generative models another explanation for the failure of GRBMs is provided by . Although they agree with the poor ability of GRBMs in modeling covariances,   argue that the deficiency is due to the binary nature of the hidden units. In order to overcome this limitation, they developed the spike-and-slab RBM (ssRBM), which splits each binary hidden unit into a binary spike variable \(h_j\) and a real valued slab variable \(s_j\). The conditional probability of visible units is given by \[P(\mathbf{X}|\mathbf{s},\mathbf{h},||\mathbf{X}||^2<R) = \frac{1}{B} \mathcal{N}\left(\mathbf{X}; \boldsymbol\Lambda^{-1} \sum_{j=1}^N{\mathbf{w}_{* j}s_jh_j}, \boldsymbol\Lambda^{-1} \right), \label{eqn:ssRBM}\] where \(\boldsymbol\Lambda\) is a diagonal matrix and \(B\) is determined by integrating the Gaussian \(\mathcal{N}(\mathbf{X}; \boldsymbol\Lambda^{-1} \sum_{j=1}^N{\mathbf{w}_{* j}s_jh_j}, \boldsymbol\Lambda^{-1} )\) over the ball \(||\mathbf{X}||^2<R\)  . In contrast to the conditional probability of GRBMs ([\[eqn:probOfXHGauss\]](#eqn:probOfXHGauss){reference-type="ref" reference="eqn:probOfXHGauss"}), \(\mathbf{w}_{* j}\) in ([\[eqn:ssRBM\]](#eqn:ssRBM){reference-type="ref" reference="eqn:ssRBM"}) is scaled by the continuous variable \(s_j\), which implies that the components can be shifted along their weight vectors. We have shown that GRBMs are capable of modeling natural image patches and that the reported failures are due to the training procedure. showed also that GRBMs could learn meaningful filters by using a sparse penalty. But this penalty changes the objective function and introduced a new hyper-parameter. addressed these training difficulties, by proposing a modification of PT and an adaptive learning rate. However, we claim that the reported difficulties of training GRBMs with PT are due to the mentioned gradient divergence problem. With gradient restriction we were able to overcome the problem and train GRBMs with normal PT successfully. # Conclusion In this paper, we provide a theoretical analysis of GRBM and showed that its product of experts formulation can be rewritten as a constrained mixture of Gaussians. This representation gives a much better insight into the capabilities and limitations of the model. We use two-dimensional blind source separation task as a toy problem to demonstrate how GRBMs model the data distribution. In our experiments, GRBMs were capable of learning meaningful features both in the toy problem and in modeling natural images. In both cases, the results are comparable to that of ICA. But in contrast to ICA the features are not restricted to be orthogonal and can form an over-complete representation. However, the success of training GRBMs highly depends on the training setup, for which we proposed several recipes based on the theoretical analysis. Some of them can be further generalized to other datasets or directly applied like the gradient restriction. In our experience, maximizing the \(\hat\ell\) does not imply good features and vice versa. Prior knowledge about the data distribution will be beneficial in the modeling process. For instance, our recipes are based on the prior knowledge of the natural image statistics, which is center peaked and has heavy tails. It will be an interesting topic to integrate prior knowledge of the data distribution into the model rather than starting modeling from scratch. Considering the simplicity and easiness of training with our proposed recipe, we believe that GRBMs provide a possible way for modeling natural images. Since GRBMs are usually used as first layer in deep belief networks, the successful training of GRBMs should therefore improve the performance of the whole network. [^1]: Some part of this analysis has been previously reported by . Thanks to the anonymous reviewer for pointing out this coincidence. [^2]: [\[ICA_algorithm\]]{#ICA_algorithm label="ICA_algorithm"}For the fast ICA algorithm  we used for training, the \(\hat\ell\) for super Gaussian sources can also be assessed analytically by \(\hat\ell =- \Big< \sum\limits_{j=1}^{N}\ln{2\cosh^2{\mathbf{w}_{* j}^T \tilde{\mathbf{x}}_l}} \Big>_{\tilde{\mathbf{x}}_l} + \ln{|\det \mathbf{W}|}\). [^3]: [\[ft:trueLL\]]{#ft:trueLL label="ft:trueLL"}As we know the true data distribution, the exact \(\hat\ell\) can be calculated by \({\hat\ell = -\sqrt{2}\Big< |\mathbf{u}_{1 *}\tilde{\mathbf{x}}_l| + |\mathbf{u}_{2 *}\tilde{\mathbf{x}}_l| \Big>_{\tilde{\mathbf{x}}_l} -\ln{2}+ \ln{|\det \mathbf{U}|}}\), where \(\mathbf{U} = (\mathbf{V}\mathbf{A})^{-1}\). [^4]: The Amari error  between two matrices \(\mathbf{A}\) and \(\mathbf{B}\) is defined as: \(\frac{1}{2N}\left( \sum\limits_{i=1}^{N} \sum\limits_{j=1}^{N} { \frac{|(\mathbf{A}\mathbf{B}^{-1})_{ij}|}{\max_k{|(\mathbf{A}\mathbf{B}^{-1})_{ik}|}} + \frac{|(\mathbf{A}\mathbf{B}^{-1})_{ij}|}{\max_k{|(\mathbf{A}\mathbf{B}^{-1})_{kj}|}} }\right)-1\).

{'timestamp': '2014-01-24T02:05:29', 'yymm': '1401', 'arxiv_id': '1401.5900', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5900'}

# Introduction On August 3rd, 1975 the low-mass X-ray binary black hole transient A0620-00, exhibited its most powerful outburst detected by the Sky Survey Experiment on board the Ariel V satellite in X-rays. On August 8th, this micro-quasar was also followed by the SAS-3 X-ray observatory. Subsequently it was also seen in different wavelengths, from radio to ultraviolet. At that time, A0620-00 became the most powerful X-ray source in the sky for almost two months. Five days after the discovery of A0620-00 intense variations on time scales of days, which reached a maximum value \(\sim 50\) times that of the Crab Nebula in the energy interval of \(1.5-6\) KeV, suggested that the source was an an excellent candidate for a stellar mass black hole with a stellar companion. This idea was further corroborated by the direct observations made by which resolved the binary components of the source. The estimated distance to A0620-00 is \(\sim \! 1\textrm{Kpc}\), being one of the nearest X-ray transients objects, hosting a black hole with a mass function \(f(m) = 3.18 \pm 0.16 \, M_{\odot}\). Using dynamical and stellar numerical models, the inclination of the accretion disc with respect to the orbit spanned by the black hole and the stellar companion yields \(i = 51^{\circ} \pm 0.9\), implying a black hole mass \(6.6\pm 0.25 \, M_{\odot}\), and an estimated distance to the source \(1.06 \pm 0.12 \, \textrm{Kpc}\) . The radio emission of A0620-00 was detected in 1975 [@Davis75; @Owen76], with no jet resolved. Since many X-ray transient systems containing a black hole have radio emission that follows their X-ray outburst with clear detections of relativistic outflows or jets, it was clear that a jet should have been produced in the X-ray outburst of A0620-00. infered the existance of that jet by compiling different radio observations, concluding that the speed of the jet \(\sim 0.9 c\), where \(c\) is the velocity of light. In this letter, we assume that the mechanism producing the observed light curve of A0620-00 is caused by variations in the injected flow at the base of the jet, which leads to the formation of shock waves that propagate along the jet. The hydrodynamical jet model presented in describes the motion of working surfaces inside a relativistic jet, which are able to fit the observed light curves of long gamma-rays bursts (lGRBs) as well as the light curve of the blazar PKS 1510-089. The shape of the X-ray light curve of the micro-quasar A0620-00 is similar to the one observed in lGRBs, showing an exponential rapid increase with a slow decay. With all these, the the physical ingredients of the phenomena that produces the light curve of A0620-00 can be considered similar to those ones ocurring in lGRB and on PKS 1510-089, but with different physical scales of energy, sizes, masses, accretion power rates, etc.. The letter is organised as follows. In Section [2](#data){reference-type="ref" reference="data"} we present the X-ray data of the light curve of A0620-00. In Section [3](#model){reference-type="ref" reference="model"} we briefly describe the main features of the hydrodynamic model by M09, and using that model we fit the observational data in Section [4](#sec:Fit){reference-type="ref" reference="sec:Fit"}. Finally, the results of our fits and the discussion of the main physical parameters inferred by the modelling are presented in Section [5](#discussion){reference-type="ref" reference="discussion"}. # Observational data {#data} The observational 1975-1976 X-ray light curve of the micro-quasar A0620-00 is shown in Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} and was kindly provided by Jeffrey McClintock. It consist of a composition of two independent lightcurves obtained by and, with instruments on board Ariel V and SAS-3 respectively. Both data count-rates have been converted to flux Crab units, according to the instruments specifications. With this it is possible to get a complete light curve of the 1975-1976 outburst, including a bump in the decaying outburst. Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} shows the plotted data on a linear scale, with the advantage of revealing the impressive outburst of 1975 and a clear second bump a few days after the maximum. To convert from Crabs to mJy, we use the conversion given by and for the Ariel V data (in the energy range \(1-13\) KeV) and the one in <http://heasarc.gsfc.nasa.gov/docs/sas3> for the SAS-3 satellite (in the energy range \(2-10\)KeV). This conversion is coherent with the results obtained by, for which \(1 \text{Crab} \approx 2.4 \times 10^{-11} \mathrm{W} \mathrm{m}^{-2}\) in the energy range \(1-13 \mathrm{KeV}\). In order to calculate the Luminosity \(L\) we multiply the obtained Flux \(F\) by \(4\pi D^2 \delta^{-p}\), where \(\delta:=1 / \Gamma(v_0)(1-(v_0/c)\cos\theta)\). For this particular case, since the inclination angle \(i \approx 51^\circ\), then the angle \(\theta\) between the jet and our line of sight is \(\approx 39^\circ\), with a distance to the source \(D = 1 \,\textrm{Kpc}\). The beaming index \(p\) for synchrotron radiation is \(3\) and we have chosen such value in accordance to the calculations of blazars and lGRBs, having in mind a unified radiative model for the flow inside all relativistic astrophysical jets. With this luminosity, and with the average jet bulk speed of \(v_{0}=0.9 c\), we are able to fit the observational data with the hydrodynamical model of M09. Attempts to model the light curve of A0620-00 were made by who noticed that this behaviour might well be understood modelling many "*synchrotron bubble*" ejections. Since micro-quasars are thought to be short scaled versions of quasars and are thus logical scaled counterparts of lGRB , it is quite natural to model their behaviour using the model by M09 to model their light curve. We thus assume that velocity and mass variations at the base of the jet of the micro-quasar A0620-00 produce internal shock waves that travel inside the expanding relativistic jet and that these shock waves in turn are able to reproduce its observed light curve. # The hydrodynamical model. {#model} Many relativistic jets show internal shock waves, which are due to the interaction of the jet with inhomogeneities of the surrounding medium, the bending of jets and time fluctuations in the velocity and mass of the ejected material. In particular the semi-analytical model of M09 is a hydrodynamical description of time fluctuations at the base of the jet that develop shock waves inside an expanding relativistic jet. The model of M09 produce internal shock waves by periodic oscillations of speed and mass discharge at the base of the jet. This mechanism injects fast fluid that overtakes slow one, producing an initial discontinuity which eventually forms a working surface expanding along the jet. The extra kinetic energy inside the working surface is thus radiated away. The efficiency converting factor between kinetic energy and observed radiation is assummed to be \(\sim 1\). This value was used by M09 and for lGRBs and the Blazar PKS1510-089. We have made such a choice, since a micro-quasar can be considered as a scaled version of a quasar. Furthermore, A0620-00 has been the most energetic X-ray micro-quasar and in this sense shares the same behaviour as PKS1510-089 which presented extreme \(\gamma\)-ray energy detections. As explained in Section [1](#introduction){reference-type="ref" reference="introduction"}, the micro-quasar A0620-00 behaves as an scaled typical lGRB and as such, the hypothesis used by M09 can be extended to this particular object. As we will discuss in section [5](#discussion){reference-type="ref" reference="discussion"}, this assumption yields physical parameters which are coherent with the expectations of typical micro-quasars. Following M09, we assume that the flow is injected at the base of the jet with a periodic velocity \(v\) given by: \[v(\tau)= v_0 + c \eta^2 \sin \omega \tau, \label{eq01}\] and a periodic mass injection rate: \[\dot{m}(\tau)= \dot{m}_0 + \epsilon^2 \sin \Omega \tau, \label{eq02}\] where \(\tau\) is the time measured in the proper frame of the source, the velocity \(v_0\) is the "average" velocity of the flow inside the jet, and \(\omega\) is the oscillation frequency of the flow. The positive constant parameters \(\eta^2\) and \(\epsilon^2\) are obtained by fitting the observational data, with the particular feature that \(\eta^2\) has to be sufficiently small so that the bulk velocity \(v(\tau)\) does not exceed the velocity of light \(c\). The mass injection rate \(\dot{m}_0\) is the "average" discharge of the flow at the base of the jet, and \(\Omega\) is its oscillation frequency. # Modelling the X-ray light curve {#sec:Fit} As previously discussed, the first outburst resembles the light curve of a typical lGRB. As such, we model that burst by assuming \(\dot{m} = \textrm{const.}\), in complete accordance to the calculations by M09. The bump in the decay of the first burst is modelled in two ways. The first is by assuming a new ejection with constant discharge added up to the first outburst. The second way is by assuming an oscillating mass discharge \(\dot{m}\) produced at a particular time while the first outburst decays. In the first burst, where \(\dot{m} = \textrm{const.}\), the semi-analytical model presented by M09, requires to know the values of \(v_0\), \(\eta^2\), \(\omega\) and \(\dot{m}\). The "mean" velocity value \(v_0\) can be taken from observational data. For this particular case, we choose the inferred value from a wide variety of radio observations modelled through ejection mechanisms by which yields a Lorentz factor \(\Gamma(v_0)=2.3\). Since the value of \(\eta^2\) has to be small due to the expected variations inside the jet, we start with a small value of \(\eta^2\) such that the bulk velocity of the flow \(v(\tau = 0) \sim 0.1 \times v_{0}\). The velocity variations \(v(\tau)\) are thus allowed to vary from this value up to the extreme upper limit \(\Gamma(v(\tau)) \sim 10\). As pointed by M09, the mass ejection rate is related to the observed luminosity \(L = \dot{m}c^{2}\) and is obtained directly from the fits of the light curve. The second burst can be described by two different mechanisms: (a) The mass discharge \(\dot{m}\) is kept constant and the velocity is the sum of the velocity as in equation [\[eq01\]](#eq01){reference-type="eqref" reference="eq01"} with an extra oscillating term \(\eta'^2 \sin \omega' \tau\), with the same \(v_0\), \(\eta^2\) and \(\omega\) used for the callibration of the first outburst. (b) The velocity is the same as the one used for the callibration of the first burst, and the mass discharge \(\dot{m}\) is allowed to oscillate as in equation [\[eq02\]](#eq02){reference-type="eqref" reference="eq02"}, with \(\dot{m}_0\) given by the results obtained with the callibration of the first outburst. Following, we set a dimensionless system of units to perform the required fitting. To do so, the luminosity is measured in units of the peak luminosity and the time in units of the FWHM of each particular outburst. This system of units is such that for the first outburst \(\omega = 1\) and \(\dot{m} = 1\), with the only unknown \(\eta^2\) obtained by a linear regression analysis to within \(10\)% of accuracy. For the case of the second outburst: (a) The only unknown is \(\eta'^2\) obtained with a further linear regression analysis. (b) The unknown quantity is \(\epsilon^2\) which can be obtained by another regression analysis. To return to the physical system of units one can recall at any particular step that the luminosity \(L = \dot{m} c^2\) (for the first outburst) and that the time \(t = \omega^{-1} \tau\) (case (a) of the second outburst), \(t = \omega'^{-1} \tau\) and \(t = \Omega^{-1} \tau\) (case (b) of the second outburst). # Discussion The results of the fits to the X-ray data presented in Section [2](#data){reference-type="ref" reference="data"} using the model by M09 are shown in Figures [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} and [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}. The obtained values for the physical parameters of the model are presented in Table [3](#table1){reference-type="ref" reference="table1"}. We have also included the maximum and minimum Lorentz factors, obtained for the bulk velocity of the flow. Direct inspection on the results of the Table show that \(\dot{m} \sim 10^{-9}-10^{-10} \textrm{M}_\odot \textrm{yr}^{-1}\), \(\omega^{-1} \sim 0.01-2 \, \text{days}\) with a maximum Lorentz factor \(2.3-3.6\). A00620-00 resulted to be an ideal target to test the model by since it closely resembles a lGRB in this outstanding outburst in x-rays. Future tests of the model have to be done with a wide variety of Light Curves from a large collection of micro-quasars.

{'timestamp': '2014-01-22T02:11:22', 'yymm': '1401', 'arxiv_id': '1401.5395', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5395'}

# Background {#background .unnumbered} State-of-the-art nanofabrication technologies have allowed for a drastic reduction in the size, and increase in quality, of nanomechanical systems, which have been the driving force behind radically increasing the sensitivity of numerous devices. Examples include accelerometers, mass sensors, electrometers, temperature sensors, force transducers and biosensors. The mass of a sensor and its ability to precisely measure physical quantities are intimately related, with smaller devices having superior sensitivity. However, as we continue to reduce device volume, it is difficult to find transduction methods that scale appropriately. Furthermore, it is generally the case that the target quantity is measured through the nanomechanical device's motion, hence as we move to more sensitive devices, we require a detection method with comparable precision. A solution to these issues has been found in the field of cavity optomechanics, which allows for quantum-limited, sub-am/\(\sqrt{\rm Hz}\) displacement sensitivity and device masses down to the pico/femto-gram range. Optomechanics describes the coupling of the mechanical motion of a device to an optical field, often to manipulate or detect its motion. It is advantageous to use an optical cavity, such as a whispering gallery mode (WGM) resonator, to provide this field, as the light in the optical cavity is able to sample the mechanics many times due to its long photon lifetime and leads to resonantly enhanced optomechanical coupling. In such a system, the motion of the mechanical device shifts the resonance frequency and phase of the optical cavity. By detecting this signal, it is possible to infer the motion of the device. At the same time, photons inside the cavity apply a radiation pressure to the mechanical resonator, which can be used to optomechanically dampen or amplify its motion, lending to a large number of interesting phenomena. Cavity optomechanical systems have been realized in a number of different geometries, including photonic crystal cavities, Fabry-Pérot etalons, WGM resonators and electronic microwave cavities. It is advantageous to fabricate cavity optomechanical devices on-chip, as it is therefore possible engineer both the optical and mechanical resonators to certain desired specifications. Modern nanofabrication technologies make it possible to produce devices with extremely accurate predefined dimensions, allowing feature sizes as small as 100 nm for foundry-based deep ultraviolet (DUV) optical lithography and 2 nm with electron beam lithography. This enables precise tailoring of important device parameters, such as the gap between mechanical and optical resonators, which controls the optomechanical coupling in our devices. Furthermore, devices fabricated using top-down lithography can be integrated into electronic on-chip devices and, in the case of optical lithography, can easily be mass produced. However, difficulties arise when trying to couple light into these on-chip devices, as a high on-chip density and planar geometry require a precise optical probe which can couple exclusively to a particular device. This problem has been solved by using dimpled tapered fibers, providing efficient and maneuverable coupling to on-chip optomechanical devices. In this article, we describe a process by which such a coupling system is produced, outlining the necessary steps while assuming no special knowledge *a priori*. We begin by investigating the fundamentals of tapered optical fibers, as well as describing an apparatus which can be used to fabricate and dimple them. Following this is a discussion of the procedure by which dimpled tapered fibers are produced. Custom-built optical access vacuum chambers, which are used for optomechanical coupling to on-chip devices, are also detailed. Finally, we explain our method of coupling to optomechanical devices with tapered fibers. Using these systems, we have been able to demonstrate the first ever on-chip optomechanical torsional sensors, as well as multidimensional detection of high frequency microcantilevers suitable for force sensing applications. # Discussion {#discussion .unnumbered} ## Single Mode Tapered Optical Fibers {#single-mode-tapered-optical-fibers .unnumbered} A crucial element in any optomechanical device is the method by which the optical field is injected, and subsequently collected, from the optical resonator in the system. While a number of different options exist, including free space optical coupling, grating couplers and fiber-to-waveguide coupling, we have chosen to use direct coupling from tapered optical fibers. Tapered fibers are more efficient, and require less on-chip space, than grating couplers, while free-space coupling is inconsistent with on-chip devices. It may prove that fiber-to-waveguide coupling is more efficient and stable than tapered fibers, but the versatiliy and maneuverability of tapered fibers remains a significant advantage. A tapered fiber is a standard optical fiber (silica core surrounded by a higher index cladding) that has had its initial diameter adiabatically reduced over a small length known as the *tapered region*. This can be performed either through hydrofluoric acid etching of an optical fiber, or by the heat-and-pull method. In this latter method, a small region of an optical fiber, known as the *hot-zone*, is heated to the point of melting and subsequently stretched to reduce its diameter. The final tapered fiber will then consist of three regions, the initial unstreched fiber, the taper transition, and the taper waist, all of which are detailed in Fig. [\[fibdia\]](#fibdia){reference-type="ref" reference="fibdia"}. From a conservation of mass argument, it can be shown that for a constant hot-zone of length \(L\), which is produced in the case of a stationary flame, the taper transition is exponential. Using a constant pull speed \(v\), this results in a taper waist diameter \(d\) that decreases with pull time \(t\) according to \[d = d_0 e^{-v t / L}, \label{expd}\] where \(d_0\) is the diameter of the initial untapered fiber. Following the heat-and-pull process, a new air-clad core exists in the taper waist, comprised of a composite material with an effective index determined by the indices and relative sizes of the initial core and cladding. This region can be modeled as a long, dielectric cylinder, for which Maxwell's equations can be solved analytically to determine the electromagnetic modes of the core (cladding) in terms of Bessel (modified Bessel) functions of the first (second) kind, as described in. In general, such a structure will support many modes, lending to the description of a multimode fiber. However, once the fiber's diameter drops below a critical value, known as the *single mode cut-off diameter*, only a single guided mode remains in the fiber, labeled the hybrid \(HE_{11}\) mode, as all other spatial modes decay evanescently. We look to determine this critical diameter for light with a free space wavelength, \(\lambda\), traveling in a fiber with a core of index of refraction, \(n_{co}\), surrounded by a cladding with index, \(n_{cl}\). This is done by matching the electromagnetic fields within the core and cladding according to the boundary conditions given by Maxwell's equations, resulting in the following expression for the single remaining mode \[\left[ \frac{J_1(x)}{x J_0(x)} + \frac{K_1(y)}{y K_0(y)} \right]\left[ \frac{n_{co}^2}{n_{cl}^2} \frac{J_1(x)}{x J_0(x)} + \frac{K_1(y)}{y K_0(y)} \right] = 0. \label{xyrel}\] In the above equation, \(J_{\nu}(x)\) is the Bessel function of the first kind and \(K_{\nu}(x)\) is the modified Bessel function of the second kind. As well, \(x = \frac{d }{2}\sqrt{k_{co}^2-\beta^2}\) and \(y = \frac{d}{2} \sqrt{\beta^2-k_{cl}^2}\), where \(k_{co}=2 \pi n_{co} / \lambda\) and \(k_{cl} = 2 \pi n_{cl} / \lambda\) are the magnitudes of the wavevector in the core and cladding, respectively, and \(\beta\) is the fiber's propagation constant. From these definitions, we can immediately derive the expression \[x^2 + y^2 = \frac{\pi^2 d^2 (n_{co}^2-n_{cl}^2)}{\lambda^2}. \label{xysq}\] Using this relationship between \(x\) and \(y\), we are able to determine the fiber diameter \(d_c\) such that Eq. [\[xyrel\]](#xyrel){reference-type="ref" reference="xyrel"} has only one solution for \(d <d_c\), indicating the point at which the penultimate mode ceases to exist. This value is the single mode cut-off diameter and is determined by numerically calculating the solutions to Eq. [\[xyrel\]](#xyrel){reference-type="ref" reference="xyrel"} while iteratively increasing \(d\) until a second solution emerges. It is also possible to determine an analytic expression for \(d_c\) in the weakly-guiding approximation (WGA). In this case, we take \(n_{co} \approx n_{cl}\), so that Eq. [\[xyrel\]](#xyrel){reference-type="ref" reference="xyrel"} becomes \[\frac{x J_0(x)}{J_1(x)} =-\frac{y K_0(y)}{K_1(y)}. \label{xyrelapprox}\] For the single mode cut-off, \(y=0\) (*i.e.* \(\beta = \pm k_{cl}\)), which is physically interpreted as the mode evanescently decaying into the cladding. Using \(\displaystyle \lim_{y \to 0} \frac{y K_0(y)}{K_1(y)} = 0\), we see that Eq. [\[xyrelapprox\]](#xyrelapprox){reference-type="ref" reference="xyrelapprox"} has solutions when \(xJ_0(x) = 0\). One solution will always exist for \(x=0\), corresponding to the single remaining mode below cut-off. The penultimate mode comes into existence when \(J_0(x)=0\) for the first time, which occurs at \(x = 2.4048\). Therefore, we can find \(d_c\) by rearranging Eq. [\[xysq\]](#xysq){reference-type="ref" reference="xysq"} to get \[d_c = \frac{ 2.4048 \lambda}{\pi \sqrt{n_{co}^2-n_{cl}^2}}. \label{cutoffd}\] The validity of the WGA is confirmed by comparing the results of Eq. [\[cutoffd\]](#cutoffd){reference-type="ref" reference="cutoffd"} to numerically calculated cutoff diameters, which are summarized for a number of situations in Table [\[cutofftab\]](#cutofftab){reference-type="ref" reference="cutofftab"}. At these diameters there exists a significant evanescent field surrounding the waist region of the tapered fiber. This allows for substantial overlap between an optical resonator's modes and the fiber's guided light when it is approached to an optical cavity. Likewise, light trapped inside the cavity will couple back into the fiber, which will be carried away as optomechanical signal. While this type of straight tapered fiber is useful for coupling to a single off-chip device, such as a microsphere, it is difficult to use as probe of on-chip devices, although it can be done if the device is cleaved to hang over the edge of the chip or isolated using a mesa. Instead, it is useful to introduce a small dimpled region to the fiber, producing a portion of the taper waist that can be used as a probe of a single on-chip optomechanical device. When combined with a precise positioning system, this allows for sampling of numerous devices with the localized coupling region at the tip of the dimple of the tapered fiber. ## Tapered Fiber Puller {#tapered-fiber-puller .unnumbered} To produce tapered fibers, we use a heat-and-pull method in which a flame from a hydrogen torch is used to soften or melt an optical fiber while simultaneously stretching it at a constant speed. In our system, we produce this flame using a custom-built mountable hydrogen torch, as seen in Fig. [\[tappull\]](#tappull){reference-type="ref" reference="tappull"}a, which is threaded using a 7/16\"-24 die (McMaster-Carr, Part No. 26005A128) producing standard threads that allow for interchangeability of torch tips. The tips we use are the HT and OX series purchased from National Torch (see Fig. [\[tappull\]](#tappull){reference-type="ref" reference="tappull"}d), which provide a wide variety of flame sizes useful for producing different sizes of tapered fibers. The hydrogen torch is fed by a needle valve-controlled line, allowing for a very small and stable flame using the OX-00 torch tip, with a single 0.51 mm diameter hole. This tip is chosen because it produces compact tapers (less than 1 cm in total length) which are ideal for our fiber holders, while maintaining a relatively high transmission efficiency (up to \(\sim\)`<!-- -->`{=html}80%). The hydrogen torch is mounted on a three-axis positioning system, consisting of automated \(xy\)-translation in the plane of the optical table on which the apparatus is mounted, along with perpendicularly oriented manual \(z\)-adjustment. The \(xy\)-translation system is based on a Zaber T-G-LSM200A200A two-axis gantry system. Each orthogonal axis is driven by a Zaber T-LSM200A linear motorized stage, allowing for a total travel range of 200 mm in either dimension with a minimum step size of 50 nm. Manual \(z\)-adjustment is provided by a New Focus 9063-COM gothic-arch translation stage mounted using a New Focus 9063-A angle bracket. The stage is manipulated by a Mitutoyo No. 906912 micrometer, providing a 25 mm travel range with 10 \(\mu\)m resolution. This system is used for precise and reproducible placement of the hydrogen torch flame as it heats the fiber, which is an important element required to consistently produce high quality tapered fibers. The fiber itself is held using two Newport 466A-710 dual arm V-groove fiber holders, each of which is connected to an adjustable optical post mounted on a Zaber T-LSM100A linear motorized stage. Each stage has a travel range of 100 mm with a resolution of 50 nm and can pull the melted fiber at speeds up to 7 mm/s. All of the Zaber stages are automated in software, allowing for precise, reproducible \(xy\)-positioning of the torch gantry, as well as the ability to set a consistent pull speed. The adjustable optical posts help to ensure that the fiber is level, as proper alignment is crucial for producing a low-loss taper. This entire setup is surrounded by a protective box, built from optical rails and acrylic sheets, which helps reduce flame instability due to air currents, as well as preventing contaminants from entering the system. Another method by which a fiber can be tapered is using a CO\(_2\) laser, which produces radiation with a wavelength ranging from 10.2-10.8 \(\mu\)m. Absorption of these photons by an optical fiber causes it to heat in proportion to the intensity of the beam and the cross section of the fiber being irradiated. Therefore, the power of a CO\(_2\) laser must be carefully controlled while pulling a fiber in order to ensure even heating. CO\(_2\) lasers have also been used as a heat source for other processes, namely in the production of high-\(Q\) silica WGM resonators, such as microspheres and bottles. We integrate a CO\(_2\) laser into our fiber pulling system by replacing our hydrogen torch with a cage mount system (see Fig. [\[tappull\]](#tappull){reference-type="ref" reference="tappull"}b) containing a 45 degree cube-mounted silvered mirror (Thorlabs-Product No. CM1-P01) and a plano-convex ZnSe lens (Thorlabs-Product No. LA7542-F) with a focal length of 25.4 mm, which can be used to focus the intense, infrared radiation from the laser onto the fiber. Attaching our lens to the torch positioning system, we gain full control of its position. This allows for defocusing of the CO\(_2\) laser beam, effectively controlling both the size and temperature of the hotspot on the fiber. Furthermore, the focused beam can be scanned along the fiber, allowing for a movable hotspot, which is required for bottle fabrication. Finally, manual height adjustment of the cage mount focusing system allows us to ensure that the CO\(_2\) beam will hit the center of the lens, reducing aberration. It is also possible to attach a microscope imaging system directly to our torch positioning gantry, as shown in Fig. [\[tappull\]](#tappull){reference-type="ref" reference="tappull"}c. The microscope is comprised of a 10X M Plan Apo long working distance infinity-corrected objective (Edmund Optics-Stock No. \#59-877) attached to an Optem Zoom 70XL lens system, allowing for 70\(\times\) magnification of the setup. This image is recorded using an Edmund Optics EO-5012C color USB webcam, providing a video feed to a nearby computer. To ensure proper lighting and image quality, light from an external Metaphase MP-LED-150 microscope LED illuminator is coupled into the lens system's coaxial illumination port using a fiber optic waveguide (Edmund Optics-Stock No. \#39-368). This system is very useful, as it allows for real time imaging of our completed tapered fibers (and other fabricated optical components) when dimpling or attaching it to its holder, with full three-axis control. As tapered fibers are quite fragile, it is difficult to move them without breaking. For this reason, we first attach the tapered fiber to a holder, creating a more robust system which can easily be relocated. To this end, we have also included a manually adjusted Newport Compact Dovetail DS40-XYZ three-dimensional linear positioning stage in our system, which allows for 1 \(\mu\)m sensitivity over a travel range of 14 mm in each of \(x\) and \(y\) and 5 mm in \(z\). This stage allows us to properly position and align the fiber holder, as well as gradually approach it to the fiber for gluing. In addition, it is used to position the fiber mold used in the dimpling process, which must be approached and raised precisely at the thinnest point of the tapered fiber. Another important aspect of the tapered fiber puller is the fiber transmission monitoring system, which allows us to determine the point at which the taper becomes single mode, as well as assess fiber losses due to tapering. To do this, we measure the transmitted power of light from a New Focus Velocity 6330 tunable diode laser through the fiber during the pulling process. To control the amount of injected power, laser light is first passed through a Thorlabs VOA50-APC variable optical attenuator (VOA) before it is coupled into the fiber using a mechanical splicer (Fiber Instrument Sales elastomeric lab splice-Part No. FIS114012), in which two straight cleaved fiber ends are butt coupled to each other with the aid of index matching gel (Fiber Instrument Sales matching gel-Part No. F10001V). Likewise, the fiber is mechanically spliced on its opposite end to a patch cable connected to a New Focus Model 1811 IR DC-125 MHz low noise photoreceiver. The DC signal from this photodiode is split off and recorded using an NI USB-6259 BNC DAQ card for the duration of a fiber pull, providing a record of transmission vs pull time, as seen in Fig. [\[fibtrans\]](#fibtrans){reference-type="ref" reference="fibtrans"}. ## Fiber Tapering Procedure {#fiber-tapering-procedure .unnumbered} To create tapered fibers, we begin with a Corning SMF-28e optical fiber that has a silica core and cladding diameter of 8.2 \(\mu\)m and 125 \(\mu\)m, respectvely, all of which is protected by an acrylate coating which extends out to a diameter of 245 \(\mu\)m. The indices of refraction and dimensions of the core and cladding are chosen such that this original fiber is single mode for wavelengths exceeding 1260 nm, which includes both the dispersionless and minimum loss wavelengths in silica of 1310 nm and 1550 nm, respectively. To begin the tapering process, the acrylate coating is removed using a Micro-Strip\(^{\textregistered}\) stripping tool over a region approximately 3 cm long in the center of an SMF-28e fiber around one meter in total length. This section of stripped fiber is subsequently cleaned using a solvent to remove any remaining acrylate. The tapering occurs in this stripped region, where the flammable acrylate has been removed. In addition, the two ends of the fiber are stripped of acrylate and cleaved flat using an Ericsson EFC11 fiber cleaver. Utilizing the mechanical splicers and index matching gel described above, these cleaved ends are spliced to two ends of a severed FC/APC patch cable, one of which leads to the photodiode, the other to the diode laser. This method of fiber splicing is ideal for this application, as it is quick and easy, allowing for a convenient input and removal of the tapering fiber to and from the optical circuit. Losses vary depending on fiber alignment for this splicing method, but we are only concerned with providing enough power to observe variations in fiber transmission. Once we have ensured that the splices provide sufficient power to the photodiode, the fiber is placed in the V-groove fiber holders, with the region prepared for tapering centered between them. At this point, the hydrogen torch is lit using a butane lighter and gas flow is adjusted to ensure a steady flame about 1 cm high. This flame is then approached towards the fiber until a small (a few mm) section begins to glow, indicating that the fiber is in a molten state. Once this point has been reached, the two pulling stages move in opposite directions, each at a constant speed generally chosen to be 40 \(\mu\)m/s. During each pull, the transmission through the fiber vs pull time is monitored, an example of which is presented in Fig. [\[fibtrans\]](#fibtrans){reference-type="ref" reference="fibtrans"} for the OX-00 torch tip. By monitoring fiber transmission, it is possible to determine the point at which the fiber waist has become single mode. This will be indicated as a stabilization of the fiber transmission (which is evident in Fig. [\[fibtrans\]](#fibtrans){reference-type="ref" reference="fibtrans"}) due to the fact that the lossy, higher order modes of the fiber have died out, leaving behind the single fundamental mode of the fiber. Using images from a scanning electron microscope (SEM-inset of Fig. [\[fibdiamvstime\]](#fibdiamvstime){reference-type="ref" reference="fibdiamvstime"}), we experimentally measured the diameters of our fibers at the single mode transition to be \(\sim\)`<!-- -->`{=html}1.1 \(\mu\)m, consistent with the theoretically predicted diameter for an air-clad fiber with an index of 1.47 (we expect our fibers to have an index of 1.4677) at 1550 nm (see Table [\[cutofftab\]](#cutofftab){reference-type="ref" reference="cutofftab"}). By measuring the time required to reach this transition from a single pull, it is possible to determine a value for the hot-zone length \(L\) by inverting Eq. [\[expd\]](#expd){reference-type="ref" reference="expd"}, provided that the pull speed and initial fiber diameter are known *a priori*. Using this parameter, we are able to predict the fiber waist diameter for a given pull time. Note that in order for this prediction to be accurate, care must be taken to ensure that all subsequent pulls have conditions matching the orginal one in order to ensure a consistent hot-zone length. This is readily accomplished using our system. A plot of fiber waist diameter vs pull time using the apparatus described here is presented in Fig. [\[fibdiamvstime\]](#fibdiamvstime){reference-type="ref" reference="fibdiamvstime"}, indicating excellent agreement between the hot-zone length of 1.30 mm determined using the single mode cutoff point and the fit value of 1.29 mm. This ability to predict the fiber waist diameter is useful, as it allows for fabrication of fibers whose diameters support a propagating mode that is phase matched with the resonance we are interested in, enhancing coupling of light from the tapered fiber to the optical resonator. At the point of single mode transition, the fiber waist diameter is small enough to produce the desired evanescent field required for coupling to an optical cavity, which can be seen in the inset of Fig. [\[fibtrans\]](#fibtrans){reference-type="ref" reference="fibtrans"}. However, it is often advantageous to continue pulling fibers to smaller diameters, further increasing the extent of the evanescent field outside the fiber geometry, allowing for a larger range of coupling before the fiber contacts the optical resonator. It is possible to create these sub-\(\mu\)m diameter fibers by continuing to pull for a small amount of time (\(\sim\)`<!-- -->`{=html}10 s) after the single mode transition has been reached. Using the OX-00 torch tip, diameters as small as 850 nm can be achieved before the fiber breaks due to the pressure of the flowing hydrogen gas from the torch. By using the HT-3, our largest torch tip, the flame size increases, nearly doubling the hot-zone to 2.4 mm, allowing for the fabrication of tapered fibers with diameters down to 500 nm and 98% transmission. This provides fibers with diameters small enough that they can be used as a probe of nitrogen vacancy center photoluminescence, as well as allow single mode guiding of 780 nm light (Table [\[cutofftab\]](#cutofftab){reference-type="ref" reference="cutofftab"}), which is used in aqueous biosensing applications. By monitoring transmission before and after the pull, it is also possible to determine the losses induced in the fiber due to the tapering process. This is important for determining the amount of power injected into the optical resonator, allowing for calculation of the number of photons confined in the optical resonator. For the OX-00 tip, a tapered fiber transmission efficiency of up to \(\sim\)`<!-- -->`{=html}80% is achieved. By using the HT-3 tip, with its larger hot-zone, a more adiabatic taper transition region is created allowing us to produce fiber tapers with transmission effiencies exceeding 99%, which is on par with state-of-the-art, ultralow loss fiber pullers. ## Fiber Dimpling Procedure {#fiber-dimpling-procedure .unnumbered} Once a tapered fiber has been pulled, it is possible to proceed with the dimpling procedure. We begin by taping a stripped Corning SMF-28e optical fiber to the \(xyz\)-positioning stage located opposite the hydrogen torch, mounting it perpendicular to the tapered fiber so that it can be used as a mold in the dimpling process (see Fig. [\[dimple\]](#dimple){reference-type="ref" reference="dimple"}a). The fiber mold is prepared by stripping off its acrylic coating and cleaning it with a solvent, producing a mold of 125 \(\mu\)m in diameter. In addition, graphite powder (SLIP Plate\(^{\textregistered}\) Tube-O-Lube\(^{\textregistered}\)) is applied to the fiber mold to prevent it from sticking to the tapered fiber. This graphite generally burns away when introduced to the hydrogen flame during the dimple annealing process, however, using too much graphite should be avoided as it can contaminate the dimple, inducing losses. To prevent this from happening, a fiber wipe or compressed air can be used to gently remove excess graphite. To continue, the torch is replaced by the microscope imaging system on the torch positioning gantry so that dimpling can be observed in real time. While watching with the microscope, the tapered fiber is detensioned by approximately 10 \(\mu\)m to reveal its thinnest point, which appears as a small bend upwards in the fiber (see Fig. [\[dimple\]](#dimple){reference-type="ref" reference="dimple"}a). The stripped fiber mold is centered on this point and manually raised to touch the tapered fiber using the \(z\)-positioning stage. The mold fiber is then raised approximately 5 mm, while simultaneously detensioning the tapered fiber, allowing the fiber to wrap itself around the mold producing the desired dimpled shape, as shown in Fig. [\[dimple\]](#dimple){reference-type="ref" reference="dimple"}b. During this process, the tapered fiber should remain tensioned tightly around the mold at all times to prevent it from twisting. At this point, a hydrogen flame produced by the tapering torch is introduced to anneal the fiber into a dimpled shape. For this process, one of the HT series torch tips is used, producing a wide flame allowing for the increase in heat distribution required for annealing. This flame is approached to the dimple by hand, touching the mold and tapered fiber lightly (for about one second) until it glows red (see Fig. [\[dimple\]](#dimple){reference-type="ref" reference="dimple"}c). The mold fiber is then slowly lowered in the same manner it was raised, this time tensioning the tapered fiber, until the mold is returned to its initial position. The dimple is then removed from the mold by using the unlit torch to flow hydrogen from below, applying a gentle pressure which releases the dimpled fiber. Typically, this process returns a dimple with minimal losses (\(\sim\)`<!-- -->`{=html}8%, see Fig. [\[dimple\]](#dimple){reference-type="ref" reference="dimple"}f). A microscope image of a dimpled fiber produced using this procedure is shown in Fig. [\[dimple\]](#dimple){reference-type="ref" reference="dimple"}e. ## Gluing Procedure {#gluing-procedure .unnumbered} Once a dimpled tapered fiber (or other optical component created by the fiber heating system) is produced, it must be carefully attached to its holder using the gluing apparatus. To begin this process, Devcon 5 Minute\(^{\textregistered}\) epoxy gel (No. 14240) is applied to both sides of the fiber holder, which can be seen in Fig. [\[gluefig\]](#gluefig){reference-type="ref" reference="gluefig"}a. Care is taken to ensure that both droplets of epoxy are approximately the same height, ensuring that they will contact both sides of the tapered region at the same time. Once the epoxy is applied to the fiber holder, it is placed on its holding plate located on the gluing apparatus. The fiber holder is then carefully aligned beneath the fiber, ensuring that the fiber will be glued in the appropriate place. Next, the fiber holder is slowly raised using the \(z\)-axis of the positioning stages until the fiber has been enveloped in epoxy on both sides of the taper. This initial epoxy is then left to dry (for about 30 minutes) allowing the fiber to be rigidly held on the fiber holder, drastically increasing its durability. Once the initial epoxy dries, a second round of gluing is typically applied to the fiber, which increases the strength of the fiber's attachment to the holder. This entire gluing process is monitored in real time using the microscope imaging system mounted on the positioning gantry, which is very helpful as we are able to definitively determine the point at which the fiber has been glued. As well, by imaging the tapered region, along with monitoring transmission down the fiber, we can determine whether or not the tapered fiber has survived the gluing process. Once the fiber has been properly glued in place, it can be transferred directly to the coupling chamber where it is fusion spliced to an existing optical circuit, allowing for injection of light into optomechanical devices. ## Optomechanical Coupling Chambers {#optomechanical-coupling-chambers .unnumbered} Our coupling chambers, which can be seen in Fig. [\[optcham\]](#optcham){reference-type="ref" reference="optcham"}, contain two separate positioning systems, with similar principles but different translation stages. In one such setup, the sample chip is placed on top of a stack of Attocube linear nanopositioning stages consisting of one ANPz101 stage mounted on top of two perpendicularly oriented ANPx101 stages. The chip is attached rigidly to a custom-machined adapter, which is fastened to the top \(z\)-positioning stage. This arrangement provides positioning with sub-nm precision over a total range of 5 mm. A picture of this setup can be seen in Fig. [\[optcham\]](#optcham){reference-type="ref" reference="optcham"}d. The other positioning system is built using Newport Agilis\(^{\rm TM}\) AG-LS25V6 vacuum compatible, piezo driven linear stages. Two of these stages are stacked on top of each other, resulting in orthogonal \(xy\)-positioning, with a third mounted at a 90 degree angle for \(z\)-translation using an EQ3 Series angle bracket purchased from Newport. As above, the chip containing our optomechanical devices is mounted on a custom-built platform, which is attached to the vertical translation stage providing full three-axis control. These stages provide 50 nm stepping resolution over their entire travel range of 12 mm. Each of these systems have different strengths, with the Attocube stack providing extremely precise positioning over a relatively large range, while the Agilis stages provide a more durable, inexpensive alternative with a larger range of motion. These positioning systems are used to approach the optomechanical devices found on the sample chip to a stationary dimpled tapered fiber, which is glued to a custom-machined fiber holder. The fiber was chosen to remain fixed as it is far less stable than the devices on the chip, so its mechanical noise is reduced by anchoring it to an immobile fiber holder. To allow interchangability between our two chambers, this positioning system is fastened to a custom-machined circular plate, containing 1/4\"-20 tapped holes in a square pattern with a spacing of 3/4\". This plate is then screwed onto a homemade aluminum base with 6 ports, each of which is sealed with an O-ring and provide electrical and optical input/output for the setup, as well as allowing for pressure control inside the chamber. The optical input/output port consists of fiber feedthroughs, each of which allows for both an input and output fiber, channelling light to and from the tapered fiber. Each fiber is glued in place using Varian Torr Seal high vacuum epoxy, which provides the appropriate seal required for vacuum. For the Attocube setup, the electrical port houses three hermetically sealed BNC feedthroughs. The other type of electrical port, which provides input/output for the Agilis stages, is comprised of a vacuum compatible 15-pin D-type connector housed in a KF50 feedthrough flange (Accu-Glass Products-Model No. 15D-K50). The vacuum environment provided by our coupling chambers removes airborne contaminants which can reduce the quality of the tapered fiber and optomechanical devices over time. It is also possible to remove such contaminants using a nitrogen purged environment, however, performing optomechanics in vacuum has the added advantage of increasing the mechanical quality factors of devices by drastically reducing viscous damping. The vacuum pump port is comprised of a KF25 adapter connected to a turbo pump backed by a dry scroll pump. By using a completely dry pumping system, we ensure that no oil is ever backstreamed into our system. This connection is made using vibration isolating bellows, which are passed through a cement block to further prevent vibrations from the pump reaching the optical table where the chamber is located. Using this pumping system, we can achieve chamber pressures as low as 10\(^{-6}\) torr. There also exists a release port, consisting of a Nupro B-4HK brass bellows-sealed valve, which allows for surrounding air to enter the system, re-establishing ambient pressure inside the chamber. All unused ports on the chamber base are covered with a blank port. This entire system is leak-checked using an Adixen ASM380 dry leak detector, ensuring it is properly sealed. On top of the chamber base is an aluminum cylinder approximately 10 cm long and 17 cm in diameter which provides the housing for the positioning stages and tapered fiber mount. An L-shaped boot gasket (Duniway Stockroom-Part No. VBJG7) is placed on each side of the cylinder providing a leak tight seal between it and both the base and its custom-machined lid. Optical access through the lid is provided by a 75 mm diameter optical flat glass window (Edmund Optics 1/4-Wave N-BK7-Stock No. \#62-606), which lays flush against an O-ring located in a recessed portion of the lid when the chamber is under vacuum. This window is located directly above the tapered fiber holder and positioning stages, which allows for real time monitoring of the optomechanical system (see Fig. [\[optcham\]](#optcham){reference-type="ref" reference="optcham"}c). It is therefore possible to view the tapered fiber and on-chip devices while attempting to couple between them, which is important for this process. The fiber and chip are imaged using the exact same imaging system described above for the tapered fiber puller. This is made possible by the fact that this microscope is oriented using a three dimensional arrangement of manually positioned New Focus 9063-COM gothic-arch translation stages mounted on a two legged custom-built stand with identical mounting plate to that used for manual \(z\)-positioning of the hydrogen torch in the tapered fiber puller, allowing interchangibility of the micrscope between the two systems. The Mitutoyo No. 906912 micrometers used to manipulate this positioning system, provide a 25 mm travel range with 10 \(\mu\)m resolution. This resolution is more than enough to view our devices in the \(xy\)-plane of the chip, as well as provide excellent focusing for our imaging setup. ## On-Chip Optomechanical Devices {#on-chip-optomechanical-devices .unnumbered} Due to the small feature sizes required for our optomechanical devices, we have chosen to use foundry-based nanofabrication, which provides high throughput of devices with a minimum feature size of 100 nm using top-down DUV photolithography. Each optomechanical device consists of an optical microdisk side-coupled to a mechanical nano/micro-resonator, such as a torsion paddle or a cantilever, as can be seen in Fig. [\[devfig\]](#devfig){reference-type="ref" reference="devfig"}. Each device is centered in a large etched area (approximately 100 \(\mu\)m \(\times\) 50 \(\mu\)m), which provides ample room for coupling using our dimpled tapered fiber method. The mask for these devices is designed using custom-programmed Python scripts utilizing the gdspy module, which generates a GDSII file containing our chip layout. This allows us to iterate through a number of device specifications, such as coupling gap, disk radius and mechanical resonator dimensions, providing a large parameter space in which we can explore different optomechanical regimes. These design files are then submitted through Canadian Microelectronic Corporation (CMC) Microsystems to the Interuniversity Microelectronics Center (IMEC) located in Leuven, Belgium. It is here our devices are fabricated on an 8 inch silicon-on-insulator (SOI) wafer, which consists of a 220 nm thick layer of single crystal silicon supported by a 2 \(\mu\)m layer of silicon dioxide. The single crystal silicon device layer is ideal for optomechanical devices, as it has negligible absorption in the telecom band around 1550 nm and a high index of refraction (\(n \approx 3.42\)), enhancing the mechanical resonator's perturbation of the optical cavity's evanescent field. Our devices are patterned onto this wafer using an excimer laser (193 nm-248 nm) and high-definition photomasks derived from our design files. These patterned wafers are then etched using either a standard or high dose recipe, producing optomechanical devices in the silicon layer which are held rigidly in place by the oxide buffer. This, along with a protective resist coating the entire wafer, help to prevent damage to the devices during transit. After we receive these wafers, a number of post-processing procedures must be performed in order to prepare our devices for measuring. We begin by dicing the 8 inch wafer wafer into 1 cm \(\times\) 1 cm chips using a diamond saw. Each chip is then ultrasonically cleaned with acetone and rinsed with isopropyl alcohol to remove the protective coating. Once the wafer has been diced and cleaned, a buffered oxide etch (BOE) is used to selectively remove the sacrificial oxide layer beneath our mechanical devices, which releases them, allowing them to oscillate freely. It is important to note that since BOE is a wet etch, we must ensure that our devices are dried using either a critical point drier or ultralight solvents, such as n-Pentane (C\(_5\)H\(_{12}\)), to prevent stiction. Once a chip has been etched and dried, it is ready to be placed in the chamber for measuring. ## Coupling Procedure {#coupling-procedure .unnumbered} Coupling to our optomechanical devices begins by locating the dimple of the tapered fiber using the imaging apparatus. This is done by searching for the portion of the fiber that is in focus at the lowest point (due to the fact that the dimple protrudes away from the rest of the fiber). Once the dimple is found, the nanopositioning stages are used to align the desired optomechanical device such that the lowest point of the dimple is able to couple light into the optical resonator. The precision of our nanopositioning stack allows for two methods by which we can couple light into the modes of our optical resonators. We can either bring the fiber close enough to the device such that the cavity's optical modes are excited by the fiber's evanescent field or we can simply touch the fiber to the optical resonator. Hovering has the advantage that the excited optical modes are less perturbed by the fiber's presence, resulting in reduced losses. However, by touching the fiber to the resonator, the mechanical instability of the fiber is further reduced. As well, by using this coupling method, it is possible to excite a larger number of optical modes, some of which have higher \(Q\)s and larger optomechanical coupling to the mechanical resonator. ## Data Acquisition: Side-of-Fringe and Homodyne Detection {#data-acquisition-side-of-fringe-and-homodyne-detection .unnumbered} Once we have coupled light into our optical resonators, we begin measuring our devices' mechanical motion using amplitude sensitive measurements in the "tuned-to-the-slope" regime, which is illustrated in Fig. [\[coupfig\]](#coupfig){reference-type="ref" reference="coupfig"}a. In this detection scheme, the mechanical device's motion causes the cavity resonance to shift, which is transduced by the slope of this lineshape into AC transmission fluctuations in the fiber, occuring at the mechanical resonance frequency. Therefore, by tuning our laser wavelength to the maximal slope of our optical cavity resonance, we provide optimal optomechanical transduction efficiency, as can be seen in Fig. [\[coupfig\]](#coupfig){reference-type="ref" reference="coupfig"}b. In general, this resonant enhancement scales with the system's optical quality factor which increases the slope of the resonance, however, this is convolved with other effects such as an optical mode's volume and overlap with mechanical motion. Using this method, we have probed devices with an angular resolution of 4 nrad/\(\sqrt{\rm Hz}\) corresponding to a torque transduction on the level of 4\(\times\)`<!-- -->`{=html}10\(^{-20}\) N\(\cdot\)m/\(\sqrt{\rm Hz}\), as well as displacement noise floors of 2 fm/\(\sqrt{\rm Hz}\) and force sensitivity of 132 aN/\(\sqrt{\rm Hz}\). It is also possible to perform phase-sensitive measurements on our devices in the "tuned-to-the-peak" regime using a balanced optical homodyne detection system, which can be used for quadrature and entanglement measurents. This method of detection also has a number of advantages, including cancellation of laser noise and the ability to lock the laser to the bottom of the optical resonance. Furthermore, since the laser's detuning from the cavity resonance is zero, a maximum number of photons are coupled into the cavity, which enhances the system's optomechancal coupling. After setting up one of these detection schemes, the AC transmission signal through the fiber is sent to a spectrum analyzer (SA), which outputs its frequency power spectrum. For an optomechanical signal, this includes the power spectral density (PSD) of the mechanical motion. Alternatively, a time series measurement of the voltage signal taken with an analog-to-digital converter (ADC) can be digitally analyzed to determine its spectral components. In our system, this is performed using a digital lock-in amplifier (Zurich Instruments HF2LI), which is a specialized ADC that first mixes an input voltage signal with a reference frequency, \(\omega_{\rm ref}\), shifting the frequencies of the input signal by \(\pm \omega_{\rm ref}\). Therefore the frequency information around the reference frequency of the input signal is now the low frequency component of the mixed signal. This has the advantage of requiring a data collection rate proportional to the bandwidth of the signal measurement, as opposed to a data collection rate proportional to the maximum frequency component. For example, when bandwidth of only 100 kHz centered at 10 MHz contains important spectral information, a data collection rate of \(\sim 10^5\) samples per second (SPS) can be used as opposed to a rate of \(\sim 10^7\) SPS, a reduction of about 100 times the data needed to acquire the signal of interest. After mixing, the lock-in applies a low-pass filter to reduce noise contributions from unwanted frequencies outside the measurement bandwidth, thus the time series output of a lock-in amplifier is the result of a convolution of the lock-in amplifier's filter response with the demodulated input signal, *i.e.* \[Z(t) = X(t) + i Y(t) = \{H(t) * e^{i \omega_{\rm ref} t} V(t)\} (t),\] where \(X(t)\) and \(Y(t)\) are the two outputs of a dual-phase lock-in amplifier, \(H(t)\) is the impulse response of the lock-in amplifier's filter, and \(V(t)\) is the input signal to the lock-in. Fourier transforming the output elucidates the convolution, giving \[Z(\omega) = H(\omega) V(\omega-\omega_{\rm ref}).\] Thus the spectrum of the lock-in amplifier's output is the spectrum of the input voltage translated in frequency by the reference frequency and enveloped by the lock-in amplifier's filter. The power spectrum can then be estimated by taking \(S_{\rm ZZ} = |Z(\omega)|^2\), or done in practise by using a PSD estimation algorithm such as Bartlett's method, giving \[S_{\rm ZZ} (\omega) = |H(\omega)|^2 S_{\rm VV} (\omega-\omega_{\rm ref}).\] This method of data acquisition allows for real time optimization of optomechanical transduction in our devices, which facilitates sensitive probing of their mechanical motion, allowing for precise measurements of physical quantities, such as forces and torques. # Conclusion {#conclusion .unnumbered} This article presents a method by which high efficiency optical coupling is achieved between a dimpled tapered fiber and nanofabricated on-chip optomechanical devices. By using a custom-built automated heat-and-pull fiber puller, it is possible to consistently produce tapered fibers of a predetermined diameter, which is often chosen to be less than the single mode cutoff diameter, providing ample evanescent field for coupling. Dimpling this tapered fiber, using a well-defined procedure, allows for production of an excellent localized probe of planar on-chip devices. Attaching this fiber to a robust holder permits it to be tranferred to special coupling chambers. In these chambers, optomechanical coupling is performed in an optical access vacuum environment and is mediated by high precision, nanopositioning stages. By using the amplitude sensitive "tuned-to-the-slope" detection scheme, angular resolution of 4 nrad/\(\sqrt{\rm Hz}\) and displacement transduction of 2 fm/\(\sqrt{\rm Hz}\) have already been demonstrated. It is anticipated that using this technique for optomechanical coupling, we will be able to continue to measure increasingly sensitive devices, approaching the measurement limits imposed by quantum mechanics. # Author's contributions {#authors-contributions .unnumbered} BDH, PHK and JPD conceived and designed the experiment. BDH, PHK and AJRM constructed the experiment. PHK performed nanofabrication post-processing for the on-chip devices. BDH, PHK, CD, AJRM and HR collected and analyzed data. BDH and AJRM performed theoretical calculations regarding the single mode cutoff diameter for tapered fibers. CD and PHK developed and optimized the procedure for producing dimpled tapered fibers. CD and HR created software for data taking and system manipulation. BDH, PHK, CD and JPD drafted the manuscript. All authors have read, approved and provided critical revisions for the final manuscript.

{'timestamp': '2014-01-23T02:00:38', 'yymm': '1401', 'arxiv_id': '1401.5482', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5482'}

# Introduction Accretion onto compact objects is almost always highly variable, but in black hole binaries the variability can be particularly dramatic. Disks in these systems exhibit transitions between a luminous state in which the emission contains a substantial thermal component, and a quiescent state dominated by non-thermal emission. The luminous state, which can be accompanied by a non-thermal "coronal\" component and disk wind, is identified with a geometrically thin disk extending close to the innermost stable orbit, while the quiescent state is consistent with the presence of a geometrically thick hot accretion flow and, often, a persistent jet. Hot accretion requires luminosities \(L \lesssim \alpha^2 L_{\rm Edd}\) [where \(L_{\rm Edd}\) is the Eddington luminosity and \(\alpha\) the usual viscosity parameter: @esin97], so the existence of a transition from a thin to a thick disk as the accretion rate wanes is not surprising. More puzzling is the behavior at the start of a new outburst. Rather than backtracking in a diagram of spectral hardness versus luminosity, the system remains in a non-thermal state up to a much higher luminosity---in some cases, approaching the Eddington limit. Transient radio emission from jets is seen as the disk enters the luminous thin disk state. The origin of this hysteretic cycle, and much else besides, is not understood. Candidate explanations include ideas as diverse as the conjectured dependence of \(\alpha\) on magnetic Prandtl number, and instabilities associated with disk warps. In this Letter, we describe a mechanism for hysteresis based on established properties of accretion flows in which angular momentum is transported by the magnetorotational instability (MRI). In common with previous authors, we argue that dynamically significant net fields thread the inner disk while it is hot and geometrically thick. Net fields increase the strength of the MRI, leading to \(\alpha \sim 1\) for net fields whose ratio of gas to magnetic pressure \(\beta \lesssim 10^2\). With such efficient transport the density is low, and this precludes transition to a thin disk until \(L \sim L_{\rm Edd}\). Once a thin disk has formed and the accretion rate has dropped significantly below Eddington, the net flux is able to diffuse out radially through the disk, and \(\alpha\) drops. The hot inner flow is then re-established at a much lower \(L \ll L_{\rm Edd}\). To complete the cycle, net flux is regenerated in the inner thick disk at the interface with the adjacent thin disk. The MRI acting near the interface generates loops of vertical field, some fraction of which accumulate stochastically within the hot flow. # Hysteresis from cyclic flux accumulation Early numerical simulations of the magnetorotational instability demonstrated that although the MRI can act as a dynamo--sustaining turbulence in the absence of mean fields--the strength of MRI turbulent transport increases when a net field is present., analyzing a set of unstratified local simulations that included net vertical or toroidal field, found that the saturated magnetic pressure \(\langle B^2/8\pi \rangle\) and viscosity parameter \(\alpha\) both scaled linearly with the net field \(B_0\). Recent work confirms this trend. Without net field, the MRI generates an \(\alpha \simeq 0.01\), but this is stimulated by any net vertical field (characterized by a mid-plane ratio of gas to magnetic pressure \(\beta\)) with \[\beta = \frac{\rho c_s^2}{B_z^2 / (8 \pi)} \lesssim 10^5.\] find that \(\alpha \simeq 0.08\) for \(\beta = 10^4\), and that \(\alpha \gtrsim 1\) for \(\beta = 10^2\). Disk winds are also produced as the net field increases, though their strength and structure remain somewhat uncertain. Hot accretion flows cannot exist for \(L \gtrsim \alpha^2 L_{\rm Edd}\), because radiative cooling is too efficient. The hysteresis observed in black hole disk state transitions is thus consistent with the hot state possessing a substantial net flux, leading to a high \(\alpha \sim 1\) and a transition luminosity approaching the Eddington limit, while the thin disk state is weakly magnetized with a lower \(\alpha\) and a small transition luminosity. (The hot flow could be "magnetically dominated\" with \(\beta \sim 1\), as in the model of, but need not be provided that \(\beta\) is low enough that \(\alpha \sim 1\).) The simulations suggest that the gap in transition luminosities could be as large as a factor of \(10^3\) to \(10^4\), though less extreme values are possible if flux accumulation is inefficient or the thin disk retains some net flux. Thin disks are unable to advect vertical field inward, so it is plausible to expect that an initially magnetized thin disk of scale \(r_{\rm out}\) will expel its net field on the diffusion time \(t_\eta \sim r_{\rm out}^2 / \eta\) associated with the turbulent diffusivity \(\eta\). If the thin disk expels net flux, how is it regenerated? We argue that flux accumulation in the vicinity of the black hole occurs due to the action of an MRI disk dynamo at the interface between an inner hot flow and an outer thin disk. Vertical loops of magnetic field, with radial scales ranging up to \(\sim h\) and zero net flux, are created at random in both the thick and thin disk regions. Loops formed deep within the inner hot flow are advected inward and do not contribute to flux accumulation, whereas loops entirely within the thin disk diffuse away. But close to the interface, field loops can be created where one footpoint is trapped in the inner hot zone while the other escapes into the thin disk. As the loop opens up and the footpoints lose causal contact, the inner hot flow is left with an element of net magnetic flux, of random sign, that is uncompensated. Inward advection of these uncorrelated elements of magnetic flux can lead to the stochastic buildup of a large net magnetic flux close to the black hole. Because the scale height of the thin disk is so small (\(h/r \sim 0.01\)), flux accumulation due to the MRI in the thin disk is probably too slow to be of interest. We therefore focus on the hot flow near the interface at radius \(r\), assuming \(h/r \sim 1\), a characteristic random vertical field strength corresponding to \(\beta_0\), and a time scale for random changes in flux \(t_{\rm dyn}\). The mid-plane gas pressure \(P \sim \alpha^{-1}(\Omega / r) \dot{M}\), where \(\dot{M}\) is the accretion rate and \(\Omega\) the angular velocity. During each dynamo cycle \(t_{\rm dyn} = k \Omega^{-1}\), \(N\) patches appear carrying positive or negative flux of magnitude \[|\Delta \Phi|_N \sim {\pi r^2 \over 2 N} \left( \frac{8\pi P}{\beta_0} \right)^{1/2}\] each. If a fraction \(f\) of these patches escapes to the thin disk and diffuses away, a net increment of flux \(\sim (f N)^{1/2} |\Delta \Phi|_N\) is acquired by the hot flow, where it can be advected inward. If these flux parcels accumulate as a random walk, the hot flow will have enough net field to attain \(\beta \sim 1\) after a time \[t \sim {4 N \over f} \beta_0 k \left( \frac{r}{r_g} \right)^{3/2} \frac{GM}{c^3}.\] Here \(r_g = GM/c^2\). MRI turbulence exhibits moderately well-defined dynamo cycles, with \(k \sim 10^2\) , and it is reasonable to assume flux cells of scale up to \(\sim r\) in a thick disk. It is not clear, however, how weakly magnetized the thick disk is when it first forms, or how readily flux is separated at the interface with the thin disk. For \(M = 10 \ M_\odot\) and \(\beta_0 = 10^3\) (a weak enough field that \(\alpha\) would be less than unity) we have \(t \sim 6 \times 10^5 (N/f)(r / 10^3 r_g)^{3/2} \ {\rm s}\). Provided that \(N/f\) is of order \(\sim 100\) or less this suggests that a dynamically important field can be regenerated during the quiescent phase of X-ray binaries, in far less time than the interval between outbursts. We also note that the magnetic field of the donor star will thread (at least) the outer disk, and that even if this ambient flux cannot be advected it may act as a boundary condition affecting the build up of flux within the hot flow. Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} illustrates how these elements may contribute to hysteresis in state transitions. In common with other authors, we assume that the action in the inner disk is sourced by large-scale variations in the accretion rate originating from further out. These variations are consistent with being due to thermal instability associated with the ionization of hydrogen . For typical parameters, thermal instability originates at radii \(r \sim 10^5 r_g\) and can be considered independent of the physics of state transitions. The quiescent state, with a hot inner accretion flow, corresponds to the thermally unstable outer disk also being quiesent, with a low accretion rate. The extent to which magnetic flux can accumulate in the inner flow will depend on the interval between outbursts, and on the details of the flux accumulation process. Moderately strong constraints can be deduced from the observation that steady jets are present during quiescence, if one assumes that these are powered by the black hole spin. In this case a simple estimate for the power is \(P_{BZ} = (\kappa/4 \pi c) \Phi^2 \Omega_H^2\), with \(\Phi\) the flux threading the horizon, \(\Omega_H\) the horizon angular velocity, and \(\kappa \simeq 0.05\) a constant . The efficiency of the Blandford-Znajek process in a thick disk is then \(\epsilon \propto \alpha^{-1} \beta^{-1}\), with a prefactor typically substantially below unity even for a rapidly spinning hole. Hence, even though we only require \(\beta \lesssim 10^2\) to reach \(\alpha \sim 1\), stronger fields would be needed to produce jets. A magnetically dominated accretion flow is also required in models where Quasi-Periodic Oscillations (QPOs) have frequencies linked to the magnetospheric radius. These arguments are not watertight (jets, for example, could also be produced from disk winds), but they motivate consideration of models where the hot state resembles a magnetically dominated accretion flow. For \(\alpha \sim 1\) the transition to a thermal state occurs at \(L \sim L_{\rm Edd}\). How this transition occurs is unclear, but if it takes place on the rapid thermal time scale the immediate consequence will be to reduce the disk pressure, so that the already strong magnetic fields become even more dominant. The high accretion rate, moreover, ensures that the nominally "thin" disk will in fact have a substantial \((h/r)\) [see e.g. the slim disk solutions of @abramowicz88]. These properties suggest that the newly formed inner disk will be able to advect flux inward even more efficiently than the precursor hot flow, with the angular momentum transport probably being dominated by wind loss rather than by the MRI . At the same time strong toroidal field will escape the disk vertically on a short time scale. The transient jets observed to form during the thick to thin transition could be a consequence of this flux expulsion, with an independent physical origin to the steady jets seen during quiescence. Even after a thin disk has formed and a quasi-steady accretion state has been attained, the disk close to the black hole will still be strongly magnetized. It is likely that a corona will be present, and inflow through the magnetically dominated atmosphere of the disk will in turn affect (and slow) flux expulsion. Over time flux will escape but if initially \(\beta \sim 1\) there will be a period during which \(\beta\) remains \(\le 10^2\), \(\alpha\) is of the order of unity, and the disk is close to the boundary between thin and hot accretion flows. Fluctuations in the accretion rate, which could originate at large scales in the disk [e.g. due to irradiation coupling between the inner disk and the thermally unstable zone: @king98] or from instabilities in flows where radiation pressure is significant, will therefore be able to trigger the observed reversions of the inner disk to a hot state. Only after \(\beta \ge 10^2\) will \(\alpha\) start to drop significantly. The inner disk will then remain stably in a thin disk state until a large decline of the accretion rate has occurred. Finally the accretion rate and luminosity drop to \(L \simeq \alpha^2 L_{\rm Edd}\), with \(\alpha = 0.02-0.1\), closer to the value expected for a zero net-field disk. This is the level needed to re-establish a hot and initially weakly magnetized inner flow, completing the cycle. # Discussion We have argued that the hysteresis observed in the cycle of black hole X-ray binary state transitions has an elementary explanation; the hot accretion flow present in the low / hard state is more strongly magnetized than the geometrically thin disk that forms subsequently. The strength of MRI disk turbulence increases with field strength and this, in turn, results in the thick-to-thin transition occurring at a higher accretion rate than the thin-to-thick transition. The MRI can be boosted even if the disk is threaded by vertical fields that reverse sign with radius, but we have emphasized the simplest possibility that the hot flow has an organized net field. A net field can build up during quiescence, and leak away during outburst, because only thick disks (or perhaps thin disks with strong coronae) are able to advect net fields radially. The details of this flux cycle are the most uncertain aspect of our model, and numerical simulations of the evolution of magnetic fields near the interface between thick and thin disks would be valuable. Observations indicate that the luminosity at the upper transition fluctuates considerably from source to source, and from cycle to cycle in the same source. These fluctuations might reflect the stochastic nature of flux accumulation, which could lead to a distribution of net magnetic field strengths and corresponding fluctuations in the value of \(\alpha\). The luminosity at the lower transition, typically \(\sim 10^{-2} L_{\rm Edd}\), shows much less scatter, suggesting that on this branch the net field is too weak to affect \(\alpha\). How and where thermal state transitions originate in black-hole binary accretion flows is uncertain, hence we are unable to propose a quantitative timeline for the cycle. However, a clear prediction of our model is that the ratio of luminosities between the cold-to-hot and hot-to-cold transitions is sensitive to the ratio of \(\alpha\)-parameters between the highly magnetized and weakly magnetized states. Timescales that depend on \(\alpha\) should reflect this dependence. and have already proposed that the hot state of black hole X-ray binaries is strongly magnetized, and the phenomenology of jets and variability in these systems appears to be consistent with this hypothesis. further suggested that QPO frequencies might be linked to the magnetospheric radius of such a flow. In our model, the dynamical importance of the net field varies cyclically as the inner disk transitions from a hot flow to a thin disk and back again. This suggests observational implications beyond those described in. Although coronal activity is possible whenever the MRI operates in a disk, its signatures may depend on the value of \(\beta\). The disk winds that are observed in absorption during the thin disk state may owe their prominence to being launched mainly by coronal activity (i.e., the buoyancy of turbulent fields generated by the MRI), in contrast to faster, lower-density winds launched magneto-centrifugally by the net fields expected to thread the disk prior to and immediately following the thick-to-thin transition. An intriguing possibility is that the collapse of the highly magnetized hot flow to a thin disk may trap an extremely strong, even suprathermal toroidal field generated by the MRI. The presence of such a field would thicken the disk, decreasing the viscous inflow time, and may also lead to a large color correction in the thermal component of the spectrum, for which observational evidence may already exist. We thank Jeff McClintock and Greg Salvesen for valuable discussions and comments on the manuscript. We acknowledge support from NASA's Astrophysics Theory Program under grant NNX11AE12G.

{'timestamp': '2014-01-23T02:00:31', 'yymm': '1401', 'arxiv_id': '1401.5475', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5475'}

# Methods Section The dynamics in our system are well described by an \(XY\) Hamiltonian of \(N\) coupled spins, \[\label{eq:HXY} H_{XY}=\sum_{i<j}J_{ij}(\sigma_i^+\sigma_j^-+\sigma_i^-\sigma_j^+)\,,\] where \(\sigma_i^{\pm}=(\sigma_i^x\pm i \sigma_i^y)/2\), with \(\sigma_i^{x,y}\) the usual Pauli matrices. We can use the Holstein--Primakoff transformation to exactly map the spins to hard-core-interacting bosons, \(\sigma_i^+\to a_i^\dagger\), \(\sigma_i^-\to a_i\), where \(a_i\) (\(a_i^\dagger\)) annihilates (creates) a boson on site \(i\). The resulting model, \(H_{\rm bos.}=\sum_{i<j}J_{ij}(a_i^\dagger a_j+h.c.)\), conserves the total number of particles (corresponding to conserved magnetisation in the original spin model, allowing us to treat the transverse-field term \(B \sum_i \sigma_i^z\) as a constant that we can neglect). In the single-particle subspace, diagonalising \(H_{\rm bos.}\) is equivalent to diagonalising the \(N\times N\) matrix with entries \(J_{ij}\). The result can be written in the form \[H_{\rm bos.}=\sum_{k}\hbar \omega_k a_k^\dagger a_k\,,\] where \(a_k^\dagger=\sum_i c_{i,k} a_i^\dagger\) creates an excitation in eigenmode \(k\). The mode functions \(c_{i,k}\) are the normalised eigenvectors of \(J_{ij}\). The eigenmode spectrum \(\omega_k\), i.e., the dispersion relation, depends on the boundary conditions and interaction range. Once \(J_{ij}\) is known, it can be determined unambiguously. A single-particle wave-packet constructed from these eigenmodes is what we call a magnon quasiparticle. A local perturbation of the fully-polarized state, i.e. a local quench, can be understood as the creation of a single magnon. After a spin flip at site \(\ell\), e.g. the system state evolves according to \[\label{eq:localQuenchEvolution} \ket{\psi(t)} = {\rm e}^{-i\, H_{\rm bos.} t / \hbar} a_\ell^\dagger \ket{0} = \sum_k c_{\ell,k}^*\,\, {\rm e}^{-i\, \omega_k\, t} a_k^\dagger \ket{0}\,,\] where \(\ket{0}\) is the vacuum. When the number of excitations exceeds one, the picture of non-interacting magnons is only an approximation, and one has to account for the presence of hard-core interactions. **Calculation of group velocity.** From Eq. [\[eq:localQuenchEvolution\]](#eq:localQuenchEvolution){reference-type="eqref" reference="eq:localQuenchEvolution"}, it becomes clear that the time evolution after a creation of a single excitation is determined by the magnon dispersion relation, and consequently by the associated group velocities. For translationally invariant systems, the modes with energies \(\omega_k\) are plane waves with well-defined wave-vector \(k\). In this case, the magnon group velocities are given by the well-known relation \(v_g=\partial \omega/\partial k\). In contrast, for a finite system with open boundary conditions and nearest-neighbour interactions, the mode functions are standing waves of the form \(c_i^{(k)}\sim\sin(k\cdot i)\), with \(k=n\pi/(N+1)\), \(n=1\dots N\). In the presence of finite-range interactions, the modes get distorted, but the number of nodes remains well defined. We can then still associate standing waves with the magnon modes and extract a maximal group velocity as \(v_g^{\max}\equiv \max_k (\omega_{k+\pi/(N+1)}-\omega_{k}) (N+1)/\pi\). **Encoding a spin-1/2 in an optical transition of a trapped ion.** To experimentally realise the spin Hamiltonian \(H_{XY}\), we identify the Zeeman states \(|S_{1/2},m=1/2\rangle\) and \(|D_{5/2},m^\prime=5/2\rangle\) of \(^{40}\)Ca\(^+\) with the \(|\!\downarrow\rangle\) and \(|\!\uparrow\rangle\) states of a spin-1/2 particle. The metastable \(D_{5/2}\) state has a lifetime of 1.16(2) s and is connected to the \(S_{1/2}\) ground state by an electric quadrupole transition at a wavelength of \(\lambda=729\) nm. The degeneracy of the ion's Zeeman states is lifted by a weak magnetic field of \(\sim\) 4 Gauss which allows us to initialise the \(\vert S_{1/2},m=1/2\rangle\) state using optical pumping techniques with a probability of about 99.9%. Choosing the \(|D_{5/2},m^\prime=5/2\rangle\) state for encoding \(|\!\uparrow\rangle\) has the advantage that spontaneous decay of the metastable state does not give rise to population loss from the computational state space. The static electric fields of the linear trap used to confine the ions axially induce electric quadrupole shifts which shift the energy of the \(|\!\uparrow\rangle\) state. Hence, the transition frequency \(\omega_0\) between the spin states is slightly inhomogeneous across the ion string. However, for our experimental parameters, these inhomogeneities are below \(20\) Hz and thus considerably smaller than the spin-spin coupling strength. **Realisation of variable-range spin-spin couplings in \(^{40}\)Ca\(^+\).** Variable-range spin-spin interaction of Ising type are realized by globally addressing the ions with a laser beam whose direction is orthogonal to the ion string axis. The laser off-resonantly couples the ions' electronic states representing \(|\!\downarrow\rangle\) and \(|\!\uparrow\rangle\) to the ion strings' collective modes of motion in the directions perpendicular to the string. The laser carries two frequencies \(\omega_\pm=\omega_0\pm\Delta\) which induces a Mølmer-Sørensen type interaction by coupling to all first-order sidebands of the transverse collective modes of motion (\(\hbar\omega_0\) is the energy difference between \(|\!\downarrow\rangle\) and \(|\!\uparrow\rangle\)). In the limit of weak coupling, the induced effective interaction between the spins is described by the Hamiltonian \[H = \hbar\sum_{i<j} J_{ij}\sigma^x_i\sigma^x_j \label{eq:methods:H}\] with spin-spin coupling constants \[J_{ij}=\Omega_i\Omega_j\frac{\hbar k^2}{2m}\sum_n\frac{b_{i,n}b_{j,n}}{\Delta^2-\nu_n^2}. \label{eq:methods:Jij}\] Here, \(\Omega_i\) denotes the Rabi frequency of each component of the bichromatic beam on ion \(i=1\dots N\), \(k=2\pi/\lambda\) and \(m\) the ion mass. The summation runs over all \(2N\) transverse modes, where \(\nu_m\) is the mode's oscillation frequency and \(b_{i,n}\) is the Lamb-Dicke factor which is proportional to the displacement of the \(i^{th}\) ion in the \(n^{th}\) collective mode. When the laser detuning \(\Delta\) is set to a value higher than the frequency of the highest transverse mode, the coupling becomes anti-ferromagnetic with a range which is described approximately by a power-law dependence \(J_{ij}\propto |i-j|^{-\alpha}\). The exponent \(\alpha\) can be varied between \(0\) and \(3\). The more similar the denominators in eq. ([\[eq:methods:Jij\]](#eq:methods:Jij){reference-type="ref" reference="eq:methods:Jij"}) become, the shorter the range of the interaction gets. This can be achieved by either increasing the laser detuning or by bunching up the transverse modes in frequency space by trapping the ions in a strongly anisotropic potential. In contrast to experiments engineering spin-spin interactions in trapped ions using Raman transitions connecting hyperfine states, our experiment uses a single-photon transition. An additional transverse field \(\hbar B\sum_i\sigma_i^z\) can be added to eq. ([\[eq:methods:H\]](#eq:methods:H){reference-type="ref" reference="eq:methods:H"}) by shifting both components of the bichromatic beam by an additional amount \(\delta=2B\). If \(B\gg J_{ij}\), joint spin flips coupling \(|\!\downarrow\rangle\spindown\leftrightarrow|\!\uparrow\rangle\spinup\) are suppressed. All local quench experiments presented in the paper were carried out in this regime where the number of excited spins is a conserved quantity. As dephasing due to magnetic field and laser noise is suppressed in subspaces with fixed numbers of excitations, the spin-spin dynamics can be followed over time scales of tens of milliseconds. In all experiments presented here, with the exception of the data shown in Fig. 4a, we trap ions in a harmonic potential with an axial frequency of \(0.219\) MHz and transverse frequencies of \(2.655\) and \(2.628\) MHz. The degeneracy of the transverse frequencies is slightly lifted to achieve efficient Doppler cooling. The detuning \(\Delta\) of the laser from the highest transverse mode is in the range of \(15\) to \(120\) kHz. To achieve \(\alpha=1.41\) in a 15-ion string (Fig. 4a), we lowered the axial confinement to 150 kHz. To reduce off-resonant excitation of the vibrational modes, frequency resolved sideband cooling of all radial vibrational modes to the ground state is employed at the beginning of each experiment. Rabi frequencies of about \(\Omega\approx(2\pi)\,125\) kHz are achieved by focussing about \(20\) mW of light to an elliptical Gaussian beam focus with beam waists \(w_{||}=380\,\mu\)m and \(w_\perp=33\,\mu\)m. In the seven-ion experiments, the intensity on the outer ions is about 8% lower than on the central ion. In the fifteen ion experiment, this number increases to about 20%. **Compensation of ac-Stark shifts.** Excitation of the \(S_{1/2}\leftrightarrow D_{5/2}\) quadrupole transition by the laser inducing the spin-spin interactions gives rise to ac-Stark shifts of the coupled levels. These shifts are caused by off-resonant excitation of dipole transitions coupling the \(S_{1/2}\) and \(P_{1/2}\) states to other excited states. For our experimental parameters, the light shifts are on the order of 2-3 kHz. Moreover, they vary from one ion to the other, reflecting the intensity inhomogeneities of the laser beam used to the interactions. To compensate the ac-Stark shifts, a third laser frequency is added to the bichromatic beam causing a light shift of the same strength but opposite sign. In order to keep the power of the compensating light field low, we chose to add a frequency component red-detuned by about 1 MHz from the \(|\!\downarrow\rangle\leftrightarrow|\!\uparrow\rangle\) transition. For this method to work, care has to be taken that there are no polarization or k-vector gradients across the ion string that might introduce intensity-independent coupling strength variations among different ions. The intensity of this frequency component is set to the right value by analyzing at which detuning of the bichromatic beam correlated spin flips are observed. **Single-ion addressing, state tomography and read-out.** Addressing of single ions is achieved by a strongly focused beam inducing a light shift on one of the ions in the string. The beam position can be switched within \(12~\mu\)s from one ion to any other ion using an acousto-optic deflector. Arbitrary single ion rotations can be built up from operations combining single-ion ac-Stark shifts with global interactions resonantly coupling the states \(|\!\downarrow\rangle\) and \(|\!\uparrow\rangle\). Combining these arbitrary single-ion rotations with spatially resolved detection of the ions' fluorescence on an EMCCD camera enables us to measure any observable that can be written as a tensor product of Pauli spin operators. For the measurement of the concurrence of spin-pairs to the left and to the right of the center spin in a chain of seven spins, we carry out quantum state tomography of the respective two-level systems. All required expectation values of two-spin observables \(\sigma_i^{\beta_i}\sigma_j^{\beta_j}\) (\(\beta=x, y, z\)) can be obtained from measurements in nine different measurement bases where ions \(1-3\) are projected onto the same set of basis states and ions \(4-7\) onto the states of a different measurement basis. In general, it is possible to measure the quantum state of any subset of spins of our system at any point in the dynamics. Error bars for quantities derived from the reconstructed density matrix are obtained from Monte-Carlo bootstrapping techniques. **Measurement of spin-spin coupling matrix elements.** For the measurement of the spin-spin coupling matrix \(J_{ij}\), the ions are initially prepared in \(|\!\downarrow\rangle\) by optical pumping. Next, all ions except \(i\) and \(j\) are transferred into an auxiliary Zeeman \(D_{5/2}\) state, which does not couple to the bichromatic beam inducing the spin-spin coupling, and the state of one of the ions still remaining in \(|\!\downarrow\rangle\) is flipped to \(|\!\uparrow\rangle\). Finally, we switch on the bichromatic beam coupling \({|\!\downarrow\rangle}_i{|\!\uparrow\rangle}_j\) to \({|\!\uparrow\rangle}_i{|\!\downarrow\rangle}_j\) and measure the frequency of oscillation at which the two ions exchange the shared excitation. **Estimation of \(\alpha\) from dispersion relation.** The spatial behaviour of the spin-spin couplings \(J_{ij}\) does not follow a perfect power law, making it difficult to extract an unambiguous exponent \(\alpha\) from a direct fit in real space. However, the magnon dispersion relation allows us to estimate an effective value for \(\alpha\). To do this, we compare the dispersion relation of a system with power-law interactions to the dispersion relation of a realistic system obeying the experimental parameters. The power-law exponent \(\alpha\) yielding the best fit provides an estimate for the interaction range. As seen in Fig. 1c, there is a close agreement between the dispersion relation from power-law interactions, the one simulated using experimental parameters, and the one using the measured coupling strengths. The corresponding \(\alpha\) allow us to classify the system behaviour, since the dispersion relation uniquely determines the dynamics in the single-magnon subspace (see Eq. [\[eq:localQuenchEvolution\]](#eq:localQuenchEvolution){reference-type="ref" reference="eq:localQuenchEvolution"}). Indeed, as demonstrated in the Extended Data Fig. 1, simulations using the estimated \(\alpha\) reproduce the main features of the measured magnetisation dynamics well. **Numerical simulations of the spin dynamics.** For numerical simulations, we use the measured trap frequencies and the intensity distribution of the ions across the string to calculate the spin-spin coupling matrix \(J_{ij}\). The coupling matrix and the measured laser-ion detuning are then used in a numerical integration of the equation of motion within the \(2^N\)-dimensional Hilbert space describing the spin system. For the simulation of the fifteen-ion experiments, we disregarded processes not conserving the spin excitation number in order to carry out the numerical integration within the one-excitation subspace. The only free parameter in the simulations is the overall intensity of the bichromatic laser field which could not be perfectly calibrated (deviations from the measured value were always below 5%). **Approximate light cones and Lieb--Robinson bound.** If interactions in a quantum many-body system are of short range (e.g. exponentially decreasing with distance), information cannot propagate arbitrarily fast. Therefore, when measuring an observable (such as the magnetisation \(\sigma_i^z\)) at a distance \(d>0\) from a local perturbation, the observed expectation value can change only after a certain time. A mathematically rigorous formulation of this concept, which we will sketch now, has first been given by Lieb and Robinson and later generalised by various authors (see, e.g., Refs. ). We denote the unperturbed initial state by \(\ket{\psi_0}\) and by \(\ket{\psi(t)}\) the state that has evolved during time \(t\) following the perturbation. Although the concept is more general, for consistency with Fig. 4 of the main text, we consider a local perturbation with \(\sigma_\ell^x\), \(\ell=(N+1)/2\), i.e., \(\ket{\psi(t=0)}=\sigma_\ell^x \ket{\psi_0}\). The change of any observable \(\mathcal {O}\) can then be bounded by \[|\bra{\psi(t)}\mathcal{O}\ket{\psi(t)}-\bra{\psi_0}\mathcal{O}\ket{\psi_0}| \leq \left|\left|\left[\mathcal{O}(t),\sigma_\ell^x(0)\right]\right|\right|\,,\] where \(\mathcal{O}(t)\) is the observable \(\mathcal{O}\) evolved in the Heisenberg picture of the unperturbed Hamiltonian, and where \(||\mathcal O||\) is the operator norm of \(\mathcal O\), i.e. its largest absolute value of its eigenvalues. The commutator on the right hand side quantifies how much it matters in which temporal order the operators \(\mathcal{O}\) and \(\sigma_\ell^x\) are applied. If interactions decrease exponentially with distance, one can bound this commutator by \[\label{eq:LRbound} \left|\left|\left[\mathcal{O}(t),\sigma_\ell^x(0)\right]\right|\right| \leq \, \left|\left|\mathcal{O}\right|\right|\, F(d,t)\,,\quad F(d,t)=C\, \mathrm{e}^{\,\mu\left(v|t|-d\right)}\,,\] where \(C\) and \(\mu\) are positive constants that depend on the interactions and lattice structure, and \(v\) is the so called Lieb--Robinson velocity. In its essence, the function \(F(d,t)\) provides an approximate light cone--information propagating faster than \(v\) is exponentially suppressed. The appearing constants are not unique , but as a reasonable choice one may identify the Lieb--Robinson velocity with the maximal group velocity \(v_g^{\rm max}\). To study the break down of the light-cone picture in systems with long-range interactions, we include in Fig. 4(a-c) of the main text the line \(t=d/v_g^{\rm max}\) that would bound the approximate light cone if the system had only nearest-neighbours interactions with strength \(\bar J\equiv\frac{1}{N-1}\sum_{i=1}^{N-1} J_{i,i+1}\). The initial excitation clearly propagates faster than this nearest-neighbour light cone. For the case of nearest-neighbour interactions, one can find a compact formulation of the Lieb--Robinson bound [\[eq:LRbound\]](#eq:LRbound){reference-type="eqref" reference="eq:LRbound"}, which we use in Fig. 4d to quantify its violation in long-range interacting systems. Assume a nearest-neighbor Hamiltonian \(H=\sum_{ij}h_{ij}\). Then, following Ref. , one can write \(F(d,t)=\sum_{m=d}^\infty \mathcal{N}(m) {(2 g |t|/\hbar)^m}/{m!}\,\), with \(g={\max}_{ij} ||h_{ij}||\). Here, \(\mathcal{N}(m)\) is the number of paths with length \(m\) that connect the quenched site \(\ell\) to the observable \(\mathcal{O}\) at distance \(d\). In one dimension, simple counting gives \(\mathcal{N}(m)=\left(\begin{array}{c} m \\ \frac{m-d}{2} \end{array}\right)\) if either \(m\) and \(d\) are both even or both odd, and \(\mathcal{N}(m)=0\) otherwise. Using this property, one can analytically evaluate the sum in \(F(d,t)\) which takes the particularly elegant form \[\label{eq:LRbound} \left|\left|\left[\mathcal{O}(t),\sigma_\ell^x(0)\right]\right|\right| \leq \, \left|\left|\mathcal{O}\right|\right|\, F(d,t)\,,\quad F(d,t)=2 I_d(4 g |t|/\hbar)\,,\] where \(I_d(x)\) is the modified Bessel function of the first kind. In Fig. 4d of the main text, we study the corresponding bound for \(\mathcal{O}=\sigma_i^z\) and \(g=\max_i J_{i,i+1}\), showing that the signal in our system propagates faster than allowed by the nearest-neighbour Lieb--Robinson bound.\

{'timestamp': '2014-01-27T02:07:29', 'yymm': '1401', 'arxiv_id': '1401.5387', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5387'}

# Introduction {#sec:Introduction} The optical Hall effect (OHE) is a physical phenomenon, which describes transverse and longitudinal magnetic field-induced birefringence, caused by the non-reciprocal magneto-optic response of electric charge carriers. The term OHE is used in analogy to the classic, electrical Hall effect, since the electrical Hall effect and certain cases of OHE observation can be explained by extensions of the classic Drude model for the transport of electrons in matter (metals). For the OHE, Drude's classic model is extended by a magnetic field and frequency dependency, describing the electron's momentum under the influence of the Lorentz force. As a result an antisymmetric contribution is added to the dielectric tensor \(\eps(\omega)\), whose sign depends on the type of the free charge carrier (electron, hole). The non-vanishing off-diagonal elements of the dielectric tensor reflect the magneto-optic birefringence, which lead to conversion of *p*-polarized into *s*-polarized electromagnetic waves, and vice versa. The OHE can be quantified in terms of the Mueller matrix, which characterizes the transformation of an electromagnetic wave's polarization state. Experimentally the Mueller matrix is measured by generalized ellipsometry (GE). During a GE measurement different polarization states of the incident light are prepared and their change upon reflection from or transmission through a sample is determined. An OHE instrument conducts GE measurements on samples in high, quasi-static magnetic fields, and detects the magnetic field induced changes of the Mueller matrix. Though several instruments with partial OHE instrument characteristics were described in the literature, most instruments did not fulfill all criteria for an OHE instrument. Nederpel and Martens developed in 1985 a single wavelength (444 nm) magneto-optical ellipsometer for the visible spectral range, but the instrument provided only low magnetic fields (\(B\leq 50\) mT). In 2003 Černe *et al.* presented a magneto-polarimetry instrument (\(B\leq 8\) T) for the mid-infrared spectral range (spectral lines of CO\(_2\) laser), and in 2004 Padilla *et al.* developed a terahertz-visible (6 to 20000 cm\(^{-1}\)) magneto-reflectance and-transmittance instrument (\(B\leq 9\) T). While both instruments provide high magnetic fields, and contain polarizers and photo-elastic-modulators, these instruments were not designed to record Mueller matrix data (GE). The full \(4\times 4\) Mueller matrix in the terahertz-mid-infrared spectral range (20 to 4000 cm\(^{-1}\)) can be measured by an instrument described in 2013 by Stanislavchuk *et al.*, but here the instrument is not designed for experiments with the sample in magnetic fields. The first full OHE instrument developed in 2006 by Hofmann *et al.* for the far-infrared (FIR) spectral range (30 to 650 cm\(^{-1}\)) provided magnetic fields up to \(6\) T and allowed sample temperatures between \(4.2\) K and room temperature. This first OHE instrument has since been successfully used to determine free charge carrier properties, including effective mass parameters for a variety of material systems. Later, OHE experiments were conducted in the terahertz (THz) spectral range, but were limited to room temperature and low magnetic fields (\(B\leq 1.8\) T). Since the magnitude of the OHE depends on the magnetic field strength, higher magnetic fields facilitate the detection of the OHE. Furthermore, the sensitivity to the OHE is greatly enhanced by phonon mode coupling, surface guided waves and Fabry-Pérot interferences. Since these effects appear from the THz to the mid-infrared (MIR) spectral range, depending on the structure and material of the sample, it is necessary to extend the spectral range covered by OHE instrumentation. An OHE instrument for the MIR, for example, can detect the magneto-optic response of free charge carriers enhanced by phonon modes present in the spectral range above 600 cm\(^{-1}\), which applies to many substrate materials, *e.g.*, SiC, Al\(_2\)O\(_3\), or GaN, as well as to many materials used for thin films, *e.g.*, III-V nitride semiconductors Al\(_{1-x}\)Ga\(_{x}\)N, Al\(_{1-x}\)In\(_{x}\)N or In\(_{1-x}\)Ga\(_{x}\)N. In addition, inter-Landau-level transitions can be studied in the MIR spectral range with a MIR OHE instrument. The extension to the THz spectral range enables the detection of the OHE in samples with low carrier concentrations. Furthermore, the strongest magneto-optic response can be observed at the cyclotron resonance frequency, which typically lies in the microwave/THz spectral range for moderate magnetic fields (few Tesla) and effective mass values comparable to the free electron mass. In this article, we present an OHE instrument, covering an ultra wide spectral range from 3 cm\(^{-1}\) to 7000 cm\(^{-1}\) (0.1--210 THz or 0.4--870 meV), which combines MIR, FIR and THz magneto-optic generalized ellipsometry in a single instrument. This integrated MIR, FIR and THz OHE instrument incorporates a commercially available, closed cycle refrigerated, superconducting 8 Tesla magnet-cryostat sub-system, with four optical ports, providing sample temperatures between \(T=1.4\) K and room temperature. The ellipsometer sub-systems were built in-house and operate in the rotating-analyzer configuration, capable of determining the normalized upper \(3 \times 3\) block of the sample Mueller matrix. The operation of the integrated MIR, FIR and THz OHE instrument is demonstrated by three sample systems. Combined experimental data from the MIR, FIR and THz spectral range of a single epitaxial graphene sample, grown on a 6*H*-SiC substrate by thermal decomposition, are shown. The MIR OHE data of the same epitaxial graphene sample is used to demonstrate the operation of the MIR ellipsometer sub-system of the integrated MIR, FIR and THz OHE instrument, over the full available magnetic field range of the instrument. The magneto-optic response of free charge carriers and quantum mechanical inter-Landau-level transitions is observed, and their polarization selection rules obtained therefrom are briefly discussed. A Te-doped, n-type GaAs substrate serves as a model system for the FIR spectral range of the FIR/THz ellipsometer sub-system. The OHE signal originating from conduction band electrons in a bulk material is discussed, and the concentration, mobility, and effective mass of the conduction band electrons is determined. Finally, OHE data from an AlGaN/GaN high electron mobility transistor structure (HEMT) from the THz spectral range of the FIR/THz ellipsometer sub-system is presented and analyzed. The data was recorded at different temperatures between \(T=1.5\) K and room temperature, representing the full sample temperature range of the instrument. The manuscript is organized as follows, in section [2](#sec:Theory){reference-type="ref" reference="sec:Theory"} dielectric and magneto-optic dielectric tensors are introduced, a brief theoretical overview on Mueller matrices and GE data-acquisition is given, and general GE data analysis procedures are introduced. Section [3](#sec:Experimental setup){reference-type="ref" reference="sec:Experimental setup"} gives a detailed description of the experimental setup, while in section [4](#sec:ACQ-ANAL){reference-type="ref" reference="sec:ACQ-ANAL"} data acquisition and data analysis procedures for OHE data are discussed. Examples of experimental results, demonstrating the operation of the integrated MIR, FIR and THz OHE instrument, are presented and discussed in section [5](#sec:samplesystems){reference-type="ref" reference="sec:samplesystems"}, which is followed by a short summary in section [6](#sec:Summary){reference-type="ref" reference="sec:Summary"}. # Theory {#sec:Theory} The evaluation of physical relevant parameters from the OHE requires the experimental observation and quantification of the OHE, and a physical model to analyze OHE data. Experimentally, the OHE is quantified in terms of the Mueller matrix \(\mathbf{M}_{\text{\tiny{OHE}}}\) by employing generalized ellipsometry (GE). The physical model which is used to analyze the observed transverse and longitudinal magneto-optic birefringence of the OHE is based on the magneto-optic dielectric tensor \(\eps_{\text{\tiny{OHE}}}(\mathbf{B})\), which is a function of the slowly varying external magnetic field \(\mathbf{B}\). If, among other parameters, the magneto-optic dielectric tensor of a sample is known, experimental Mueller matrices \(\mathbf{M}_{\text{\tiny{OHE}}}\) can be modeled from \(\eps_{\text{\tiny{OHE}}}(\mathbf{B})\) using the relation \[\mathbf{M}_{\text{\tiny{OHE}}}(\eps_{\text{\tiny{OHE}}}(\mathbf{B}))\;. \label{eqn:MM-DF}\] This relation is in general not invertible analytically, but can be used to determine the magneto-optic dielectric tensor from experimental Mueller matrix data through non-linear model regression analysis. Dielectric tensors, Mueller matrix calculus, generalized ellipsometry including data acquisition, as well as data analysis will be addressed in this section. ## Magneto-optical dielectric tensors {#sec:DFs} The optical response of a sample is here described by the dielectric tensor \(\eps\). If the dielectric tensor of the sample without a magnetic field is given by \(\eps_{_{\hspace{-1pt}\mathbf{B}=0}}\) and the change of the dielectric tensor induced by a magnetic field \(\mathbf{B}\) by \(\eps_{_{\hspace{-1pt}\mathbf{B}}}\), the magneto-optic dielectric tensor describing the OHE, can be expressed as \[\eps_{\text{\tiny{OHE}}}(\mathbf{B}) = \eps_{_{\hspace{-1pt}\mathbf{B}=0}} +\; \eps_{_{\hspace{-1pt}\mathbf{B}}}\;.\] The magneto-optic response of the sample described by \(\eps_{_{\hspace{-1pt}\mathbf{B}}}\) usually originates from bound and unbound charge carriers subjected to the magnetic field and is caused by the action of the Lorentz force. The magneto-optic response is anisotropic, and non-reciprocal in time. Thus, the corresponding magneto-optic contributions \(\chi_{\mathrm{+}}\) and \(\chi_{\mathrm{-}}\) to the permittivity tensor \(\bm{\chi}=\eps-\mathbf{I}\), where \(\mathbf{I}\) is the \(3\times 3\) identity matrix, originate from the interaction of right-and left-handed circularly polarized light with the sample, respectively. Without loss of generality, if the magnetic field \(\mathbf{B}\) is pointing in the *z*-direction, the polarization vector \(\mathbf{P}=\varepsilon_0\bm{\chi}\mathbf{E}\) can be described by arranging the electric fields in their circularly polarized eigensystem \(\mathbf{E}_e = (E_x + \mathrm{i}E_y, E_x-\mathrm{i}E_y, E_z) = (E_{+}, E_{-}, E_z)\) by \(\mathbf{P}_e = \varepsilon_0\bm{\chi}_e \mathbf{E}_e = \varepsilon_0(\chi_{\mathrm{+}}E_{+}, \chi_{\mathrm{-}}E_{-}, 0)\), where \(\text{i}=\sqrt{-1}\) is the imaginary unit. Transforming \(\mathbf{P}_e\) back into the laboratory system the change of the dielectric tensor induced by the magnetic field takes the form: [^1] \[\label{eq:LLTeps} \eps_{_{\hspace{-1pt}\mathbf{B}}} = \frac{1}{2} \begin{pmatrix} \hspace{11pt}(\chi_{\textbf{\tiny{+}}}+\chi_{\textbf{\tiny{--}}}) & \text{i}(\chi_{\textbf{\tiny{+}}}-\chi_{\textbf{\tiny{--}}}) & 0 \\ -\text{i}(\chi_{\textbf{\tiny{+}}}-\chi_{\textbf{\tiny{--}}}) & \hspace{3pt}(\chi_{\textbf{\tiny{+}}}+\chi_{\textbf{\tiny{--}}}) & 0 \\ 0 & 0 & 0 \\ \end{pmatrix}\;.\] Note, under field inversion \(\mathbf{B}\rightarrow-\mathbf{B}\), the polarizabilities for left-and right-handed circularly polarized light interchange. \(\eps_{_{\hspace{-1pt}\mathbf{B}}}\) is only diagonal if \(\chi_{\mathrm{+}}=\chi_{\mathrm{-}}\), and otherwise is non-diagonal with anti-symmetric off diagonal elements. ### Classic dielectric tensors (Lorentz-Drude model) Charged carriers, subject to a slowly varying magnetic field obey the classical Newtonian equation of motion (Lorentz-Drude model) \[\mathbf{m}\mathbf{\ddot{x}} + \mathbf{m}\GMA\mathbf{\dot{x}} + \mathbf{m} \omega_0^2 \mathbf{x} = q \mathbf{E} + q (\mathbf{\dot{x}} \times \mathbf{B})\;, \label{eqn:fcc-newton}\] where \(\mathbf{m}\), \(q\), \(\boldsymbol{\mu}=q\mathbf{m}^{-1}\GMA^{-1}\), \(\mathbf{x}\) and \(\omega_0\) represent the effective mass tensor, the electric charge, the mobility tensor, the spatial coordinate of the charged carrier and the eigenfrequency of the undamped system without external excitation and magnetic field, respectively. For a time harmonic electromagnetic plane wave with an electric field \(\mathbf{E}\rightarrow\mathbf{E}\exp(\text{i}\omega t)\) with angular frequency \(\omega\), the time derivative of the spatial coordinate of the charge carrier is \(\mathbf{\dot{x}}=\mathbf{v}\exp(\text{i}\omega t)\), where \(\mathbf{v}\) is the velocity of the charge carrier. With the current density \(\mathbf{j}=n q \mathbf{v}\) Eq. ([\[eqn:fcc-newton\]](#eqn:fcc-newton){reference-type="ref" reference="eqn:fcc-newton"}) reads \[\mathbf{E} = \frac{1}{nq} \left[ \text{i} \frac{\mathbf{m}}{q\omega} \left( \omega_0^2\mathbf{I}-\omega^2\mathbf{I}-\text{i}\omega\GMA \right) \mathbf{j} + (\mathbf{B} \times \mathbf{j}) \right]\;,\] where \(n\) is the charge carrier density. With the Levi-Cevita-Symbol \(\epsilon_{ijk}\)[^2], the conductivity tensor \(\mathbf{\sigma}\), the dielectric constant \(\varepsilon_0\), and using \(\mathbf{E}= \boldsymbol{\sigma}^{-1} \mathbf{j}\) and \(\eps=\frac{1}{\text{i} \varepsilon_0 \omega}\boldsymbol{\sigma}\) the dielectric tensor for charge carriers subject to the external magnetic field \(\mathbf{B}\) can be expressed as \[\label{eqn:fulldrude} \varepsilon_{ik} = \frac{nq^2}{\varepsilon_0} \left[ m_{ik} (\omega_0^2-\omega^2-\text{i}\omega\gamma_{ik}) - \text{i}\omega\epsilon_{ijk} q B_j \right]^{-1}\;.\] #### Polar lattice vibrations (Lorentz oscillator) {#sec:PhononDF} For isotopic effective mass tensors the cyclotron frequency \(\omega_\text{c}=\frac{q |B|}{m}\) can be defined. For the mass of the vibrating atoms of polar lattice vibrations, the cyclotron frequency is several orders of magnitude smaller than for effective electron masses, and can be neglected for the magnetic fields and spectral ranges discussed in this paper. Therefore, the dielectric tensor of polar lattice vibrations \(\eps^{\text{\tiny{L}}}\) can be approximated using Eq. ([\[eqn:fulldrude\]](#eqn:fulldrude){reference-type="ref" reference="eqn:fulldrude"}) with \(\mathbf{B}=0\). When assuming isotropic effective mass and mobility tensors, the result is a simple harmonic oscillator function with Lorentzian-type broadening. For materials with orthorhombic symmetry and multiple optical excitable lattice vibrations, the dielectric tensor can be diagonalized to \[\eps^{\text{\tiny{L}}} = \begin{pmatrix} \varepsilon_{x}^{\text{\tiny{L}}} & 0 & 0 \\ 0 & \varepsilon_{y}^{\text{\tiny{L}}} & 0 \\ 0 & 0 & \varepsilon_{z}^{\text{\tiny{L}}} \end{pmatrix}\;, \label{Substrate}\] where \(\varepsilon_{\text{\tiny{\textit{k}}}}^{\text{\tiny{L}}}\) (\(k=\{x,y,z\}\)) is given by \[\varepsilon_{\text{\tiny{\textit{k}}}}^{\text{\tiny{L}}}=\varepsilon_{\infty,\text{\tiny{\textit{k}}}}\prod^{l}_{j=1}\frac{\omega^2+\text{i}\omega\gamma_{\stext{LO},\text{\tiny{\textit{k,j}}}}-\omega^2_{\stext{LO},\text{\tiny{\textit{k,j}}}}}{\omega^2+\text{i}\omega\gamma_{\stext{TO},\text{\tiny{\textit{k,j}}}}-\omega^2_{\stext{TO},\text{\tiny{\textit{k,j}}}}}\;, \label{eqn:phonon2}\] where \(\omega_{\text{LO,\text{\tiny{\textit{k,j}}}}}\), \(\gamma_{\text{LO,\text{\tiny{\textit{k,j}}}}}\), \(\omega_{\text{TO,\text{\tiny{\textit{k,j}}}}}\), and \(\gamma_{\text{TO,\text{\tiny{\textit{k,j}}}}}\) denote the \(k=\{x,y,z\}\) component of the frequency and the broadening values of the \(j^{\text{th}}\) longitudinal optical (LO) and transverse optical (TO) phonon modes, respectively, while the index \(j\) runs over \(l\) modes. Further details can be found in Refs. , and a detailed discussion of the requirements to the broadening parameters, such as Im\(\left\{\varepsilon_{\text{\tiny{\textit{k}}}}^{\text{\tiny{L}}}\right\}\geq 0\), in Ref. . #### Free charged carriers (extended Drude model) {#sec:DrudeDF} For free charged carriers no restoring force is present and the eigenfrequency of the system is \(\omega_0=0\). For isotropic effective mass and conductivity tensors, and magnetic fields aligned along the *z*-axis Eq. ([\[eqn:fulldrude\]](#eqn:fulldrude){reference-type="ref" reference="eqn:fulldrude"}) can be written in the form \(\eps^{\text{\tiny{D}}}_{\text{\tiny{OHE}}}(\mathbf{B}) = \eps_{_{\hspace{-1pt}\mathbf{B}=0}}^{\text{\tiny{D}}} + \eps_{_{\hspace{-1pt}\mathbf{B}}}^{\text{\tiny{D}}}\), with the Drude dielectric tensor for \(B=0\) \[\eps_{_{\hspace{-1pt}\mathbf{B}=0}}^{\text{\tiny{D}}} = -\frac{\omega_{\text{p}}^2}{\omega (\omega+\text{i}\gamma)} \mathbf{I} = \varepsilon^{\text{\tiny{D}} \mathbf{I}\;,\] where \(\omega_{\text{p}}=\sqrt{\frac{n q^2}{m\varepsilon_0}}\) is the plasma frequency, and \(\varepsilon^{\text{\tiny{D}}}\) is the isotropic Drude dielectric function. The magneto-optic contribution to the dielectric tensor \(\eps_{_{\hspace{-1pt}\mathbf{B}}}^{\text{\tiny{D}}}\) for isotropic effective masses and conductivities can be expressed, using Eq. ([\[eq:LLTeps\]](#eq:LLTeps){reference-type="ref" reference="eq:LLTeps"}), through polarizability functions for right-and left-handed circularly polarized light \[\chi_{\mathrm{\pm}} = -\frac{\varepsilon^{\text{\tiny{D}}}}{1\mp \frac{\omega+\text{i}\gamma}{\omega_{\text{c}}}}\;,\] where \(\omega_{\text{c}}=\frac{q |B|}{m}\) is the isotropic cyclotron frequency. ### Non-classic dielectric tensors (Inter-Landau-level transitions) {#sec:LandauDF} The dielectric tensor \(\eps_{_{\hspace{-1pt}\mathbf{B}}}^{\text{\tiny{LL}}}\) describing a series of inter-Landau-level transitions can be approximated by a sum of Lorentz oscillators. The quantities \(\chi_{\mathrm{\pm}}\) in Eq. ([\[eq:LLTeps\]](#eq:LLTeps){reference-type="ref" reference="eq:LLTeps"}) are then expressed by \[\chi_{\pm} = e^{\pm\text{i}\phi}\sum_k \frac{A_{k}}{\omega^2-\omega^2_{0,k}-\text{i} \gamma_{k} \omega}\;, \label{eqn:landau}\] where \(A_{k}\), \(\omega_{0,k}\), and \(\gamma_{k}\) are amplitude, transition energy, and broadening parameter of the \(k^{\text{th}}\) inter-Landau-level transition, respectively, which in general depend on the magnetic field. The phase factor \(\phi\) was introduced empirically here to describe the experimentally observed line shapes of all Mueller matrix elements. For inter-Landau-level transitions in graphite or bi-layer graphene we find \(\phi=\pi/4\), and for inter-Landau-level transitions in single layer graphene \(\phi=0\). Note that for \(\phi=0\), the polarizabilities for left and right handed circularly polarized light are equal \((\chi_{\textbf{\tiny{+}}}=\chi_{\textbf{\tiny{--}}})\), and \(\eps_{_{\hspace{-1pt}\mathbf{B}}}^{\text{\tiny{LL}}}\) is diagonal. ## Mueller matrix calculus, GE and data acquisition {#sec:OHE} ### Stokes vector/Mueller matrix calculus {#sec:MM-calculus} The real-valued Stokes vector \(\mathbf{S}\) has four components[^3], carries the dimension of an intensity, and can quantify any polarization state of plane electromagnetic waves. If expressed in terms of the *p*-and *s*-coordinate system[^4], its individual components can be defined by \(S_1=I_p+I_s\), \(S_2=I_p-I_s\), \(S_3=I_{45}-I_{-45}\), and \(S_4=I_{\sigma+}-I_{\mathit{\sigma}-}\), with \(I_p\), \(I_s\), \(I_{45}\), \(I_{-45}\), \(I_{\sigma+}\), and \(I_{\sigma-}\) being the intensities for the *p*-, *s*-, +45\(^\circ\),-45\(^\circ\), right-and left-handed circularly polarized light components, respectively. The real-valued \(4\times 4\) Mueller matrix \(\mathbf{M}\) describes the change of electromagnetic plane wave properties (intensity, polarization state), expressed by a Stokes vector \(\mathbf{S}\), upon change of the coordinate system or the interaction with a sample, optical element, or any other matter \[S^{(\text{out})}_j=\sum^{3}_{i=1}M_{ij}S^{(\text{in})}_i,\;\;(j=1\ldots 4)\;, \label{eqn:multi_MM}\] where \(\mathbf{S}^{(\text{out})}\) and \(\mathbf{S}^{(\text{in})}\) denote the Stokes vectors of the electromagnetic plane wave before and after the change of the coordinate system, or an interaction with a sample, respectively. Note that all Mueller matrix elements of the GE data discussed in this paper, are normalized by the element \(M_{11}\), therefore \(|M_{ij}|\leq 1\) and \(M_{11}\overset{!}{\equiv}1\). ### Mueller matrix and OHE data {#sec:MM-OHE-data} The Mueller matrix can be decomposed in 4 sub-matrices, where the matrix elements of the two off-diagonal-blocks \(only deviate from zero if \textit{p}-\textit{s}-polarization mode conversion appears, while the matrix elements in the two on-diagonal-blocks \text{ \scriptsize \) \begin{bmatrix} M_{11}&\hspace{-3pt}M_{12}\\ M_{21}&\hspace{-3pt}M_{22}\\ \end{bmatrix} \( \normalsize and \scriptsize \) \begin{bmatrix} M_{33}&\hspace{-3pt}M_{34}\\ M_{43}&\hspace{-3pt}M_{44}\\ \end{bmatrix} \( }\) mainly contain information about *p*-*s*-polarization mode conserving processes. *p*-*s*-polarization mode conversion is defined as the transfer of energy from the *p*-polarized channel of an electromagnetic plane wave to the *s*-polarized channel, or vice versa. Polarization mode conversion can appear when the *p*-*s*-coordinate system is different for \(\mathbf{S}^{(\text{in})}\) and \(\mathbf{S}^{(\text{out})}\)[^5], or when a sample shows birefringence, for example. In particular, polarization mode conversion appears if the dielectric tensor of a sample possesses non-vanishing off-diagonal elements. Therefore, in Mueller matrix data from optically isotropic samples, ideally all off-diagonal-block elements vanish, while, for example, magneto-optic birefringence can cause non-zero off-diagonal-block elements in the Mueller matrix. Here, we define OHE data as Mueller matrix data from an OHE experiment \[Eq. ([\[eqn:MM-DF\]](#eqn:MM-DF){reference-type="ref" reference="eqn:MM-DF"})\] with magnetic field \(\pm \mathbf{B}\) \[\label{OHE-MM-approx} \mathbf{M}^\pm_{\text{\tiny{OHE}}} = \mathbf{M}(\eps_{_{\hspace{-1pt} \mathbf{B}=0}}\hspace{-4pt}+\hspace{-2pt}\eps_{_{\hspace{-2pt}\pm \mathbf{B}}})\;.\] Furthermore, we define the derived OHE datasets \(\delta\mathbf{M}^{\pm}\) as difference data between the Mueller matrix datasets, measured at the magnetic field \(\pm \mathbf{B}\) and the corresponding zero field dataset \[\label{delta-MM-pm} \begin{split} \delta\mathbf{M}^{\pm} &= \mathbf{M}^\pm_{\text{\tiny{OHE}}} \hspace{-2pt}-\hspace{-2pt} \mathbf{M}_0\\ &= \Delta\mathbf{M}(\eps_{_{\hspace{-1pt} \mathbf{B}=0}},\eps_{_{\hspace{-2pt}\pm \mathbf{B}}})\;, \end{split}\] where \(\mathbf{M}_0=\mathbf{M}(\eps_{_{\hspace{-1pt}\mathbf{B}=0}})\) is the Mueller matrix of the zero field experiment and \(\Delta\mathbf{M}(\eps_{_{\hspace{-1pt} \mathbf{B}=0}},\eps_{_{\hspace{-2pt}\pm \mathbf{B}}})\) is the magnetic field induced change of the Mueller matrix. This form of presentation is in particular advantageous in case the magnetic field causes only small changes in the Mueller matrix, and provides improved sensitivity to magnetic field dependent model parameters during data analysis. Another form of presentation for derived OHE data is \[\label{delta-MM-p-m} \begin{split} \delta\mathbf{M}^{+}\pm\delta\mathbf{M}^{-} \hspace{4pt} = \hspace{10pt} &\Delta\mathbf{M}(\eps_{_{\hspace{-1pt} \mathbf{B}=0}},\eps_{_{\hspace{-2pt}+\mathbf{B}}})\\ \pm &\Delta\mathbf{M}(\eps_{_{\hspace{-1pt} \mathbf{B}=0}},\eps_{_{\hspace{-2pt}-\mathbf{B}}})\;, \end{split}\] which can be used to inspect symmetry properties of magneto-optic Mueller matrix data, and can help to improve the sensitivity to magnetic field dependent model parameters during data analysis. ### Mueller matrix data acquisition (GE) {#sec:data acquisition} Rotating element ellipsometers can be classified into two categories: (i) rotating analyzer ellipsometers (RAE) in polarizer-sample-rotating-analyzer (\(PSA_R\)) or rotating-polarizer-sample-analyzer (\(P_RSA\)) configuration, capable to measure the upper left \(3\times 3\) block of the Mueller matrix; (ii) rotating compensators ellipsometers (RCE) in polarizer-sample-rotating-compensator-analyzer (\(PSC_RA\)) or polarizer-rotating-compensator-sample-analyzer (\(PC_RSA\)) configuration, capable to measure the upper left \(3\times 4\) or \(4\times 3\) block of the Mueller matrix, respectively. The Mueller matrices of a polarizer \(\mathbf{P}\), analyzer \(\mathbf{A}\), compensator \(\mathbf{C}(\delta)\) with phase shift \(\delta\), coordinate rotation along beam path \(\mathbf{R}\left(\theta\right)\) by an angle \(\theta\), and of the sample \(\mathbf{M}\) are given by \[\begin{aligned} &\mathbf{P}=\mathbf{A}\hspace{-10pt}&=\frac{1}{2} &\begin{bmatrix} \hspace{8pt}1&\hspace{17pt}1&\hspace{17pt}0&\hspace{18pt}0&\hspace{5pt}{}\\ \hspace{8pt}1&\hspace{17pt}1&\hspace{17pt}0&\hspace{18pt}0&\hspace{5pt}{}\\ \hspace{8pt}0&\hspace{17pt}0&\hspace{17pt}0&\hspace{18pt}0&\hspace{5pt}{}\\ \hspace{8pt}0&\hspace{17pt}0&\hspace{17pt}0&\hspace{18pt}0&\hspace{5pt}{} \end{bmatrix}\;,\\ &\mathbf{C}\left(\delta\right) &=\ \ \ &\begin{bmatrix} \hspace{8pt}1&\hspace{17pt}0&\hspace{10pt}0&\hspace{0pt}0\\ \hspace{8pt}0&\hspace{17pt}1&\hspace{10pt}0&\hspace{0pt}0\\ \hspace{8pt}0&\hspace{17pt}0&\hspace{8pt}\cos \delta&\hspace{2pt}\scalebox{0.6}[1.0]{\(-\)}\sin \delta\\ \hspace{8pt}0&\hspace{17pt}0&\hspace{10pt}\sin \delta&\hspace{7pt}\cos \delta \end{bmatrix}\;,\\ &\mathbf{R}\left(\theta\right) &=\ \ \ &\begin{bmatrix} \hspace{8pt}1&\hspace{6pt}0 &\hspace{-12pt}0 &\hspace{1pt}0&\hspace{5pt}{} \\ \hspace{8pt}0&\hspace{4pt}\cos 2\theta_j &\hspace{-1pt}\sin 2\theta_j &\hspace{1pt}0&\hspace{5pt}{} \\ \hspace{8pt}0&\hspace{-2pt}\scalebox{0.6}[1.0]{\(-\)}\hspace{1pt}\sin 2\theta_j &\hspace{-1pt}\cos 2\theta_ j &\hspace{1pt}0&\hspace{5pt}{} \\ \hspace{8pt}0&\hspace{7pt}0 &\hspace{-12pt}0 &\hspace{1pt}1&\hspace{5pt}{} \end{bmatrix}\;,\\ &\mathbf{M} &=\ \ \ &\begin{bmatrix} \hspace{3pt}M_{11}&\hspace{4pt}M_{12}&\hspace{4pt}M_{13}&\hspace{4pt}M_{14}&\hspace{-2pt}{} \\ \hspace{3pt}M_{21}&\hspace{4pt}M_{22}&\hspace{4pt}M_{23}&\hspace{4pt}M_{24}&\hspace{-2pt}{} \\ \hspace{3pt}M_{31}&\hspace{4pt}M_{22}&\hspace{4pt}M_{33}&\hspace{4pt}M_{34}&\hspace{-2pt}{} \\ \hspace{3pt}M_{41}&\hspace{4pt}M_{42}&\hspace{4pt}M_{43}&\hspace{4pt}M_{44}&\hspace{-2pt}{} \end{bmatrix}\;, \end{aligned}\] respectively. Execution of the matrix multiplication characteristic for the corresponding ellipsometer type shows that, due to the rotation of optical elements, the measured intensity at the detector is typically sinusoidal. Fourier analysis of the detector signal provides Fourier coefficients, which are used to determine the Mueller matrix of the sample (see sec. [4.1](#sec:OHE-acq){reference-type="ref" reference="sec:OHE-acq"}). ## Data analysis {#sec:data analysis} Ellipsometry is an indirect experimental technique. Therefore, in general, ellipsometric data analysis invokes model calculations to determine physical parameters in dielectric tensors or the thickness of layers, for instance. Sequences of homogeneous layers with smooth and parallel interfaces are assumed in order to calculate the propagation of light through a layered sample, by the 4\(\times\)`<!-- -->`{=html}4 matrix formalism. To best match the generated data with experimental results, parameters with significance are varied and Mueller matrix data is calculated for all spectral data points, angles of incidence and magnetic fields. During the mean square error (MSE) regression, the generated Mueller matrix data \(M_{ij,k}^{\text{G}}\) is compared with the experimental Mueller matrix data \(M_{ij,k}^{\text{E}}\) and their match is quantified by the MSE \[\text{MSE}=\sqrt{\frac{1}{9S-K}\sum^{4}_{i=1}\sum^{4}_{j=1}\sum^{S}_{k=1}\left(\frac{M_{ij,k}^{\text{E}}-M_{ij,k}^{\text{G}}}{\sigma_{M_{ij,k}^{\text{E}}}}\right)^2}\;,\] where \(S\), \(K\) and \(\sigma_{M_{ij,k}^{\text{E}}}\) denotes the total number of data points, the total number of parameters varied during the non-linear regression process, and the standard deviation of \(M_{ij,k}^{\text{E}}\), obtained during the experiment, respectively. For fast convergence of the MSE regression, the Levenberg-Marquardt fitting algorithm is used. The MSE regression is interrupted when the decrease in the MSE is smaller than a set threshold and the determined parameters are considered as best model parameters. The sensitivity and possible correlation of the varied parameters is checked and, if necessary, the model is changed and the process is repeated. # Integrated MIR, FIR and THz OHE instrument {#sec:Experimental setup} Figure [\[fig:CAD-overview\]](#fig:CAD-overview){reference-type="ref" reference="fig:CAD-overview"} shows (side view) the integrated MIR, FIR and THz OHE instrument with its four sub-systems: (A) the MIR ellipsometer sub-system, (B) the FIR/THz ellipsometer sub-system, (C) the magneto-cryostat sub-system, and (D) the magneto-cryostat transfer sub-system. In order to utilize the magneto-cryostat sub-system with the MIR or the FIR/THz ellipsometer sub-system the magneto-cryostat transfer sub-system was installed. The integrated MIR, FIR and THz OHE instrument contains multiple light sources and detectors, and covers a spectral range from 3 cm\(^{-1}\) to 7000 cm\(^{-1}\) (0.1--210 THz or 0.4--870 meV). Both ellipsometer sub-systems can be operated without the magneto-cryostat sub-system, in a variable angle of incidence ellipsometry mode[^6] (\(\Phi_a=30^{\circ}\dots 90^{\circ}\)). Figure [\[fig:2dsetup\]](#fig:2dsetup){reference-type="ref" reference="fig:2dsetup"} shows a schematic overview (top view) of all major components in the integrated MIR, FIR and THz OHE instrument. ## MIR ellipsometer sub-system {#sec:MIR} The upper part of Fig. [\[fig:2dsetup\]](#fig:2dsetup){reference-type="ref" reference="fig:2dsetup"} shows a schematic drawing (top view) of the optical configuration of the MIR sub-system of the integrated MIR, FIR and THz OHE instrument. The MIR ellipsometer sub-system is composed of (i) the MIR source unit, (ii) the polarization state preparation unit, (iii) the MIR goniometer unit, and (iv) the polarization state detection unit. To minimize absorption due to water vapor, the complete beam path of the MIR ellipsometer sub-system is purged with dried air. Due to the high magnetic stray-fields (see Fig. [\[fig:CAD\]](#fig:CAD){reference-type="ref" reference="fig:CAD"}), all opto-mechanical components in the polarization state preparation and detection units were designed and manufactured without ferromagnetic materials (with exception of the stepper motors). The MIR source unit of the MIR ellipsometer sub-system is a Bruker Vertex 70 Fourier-transform-infrared spectrometer (Fig. [\[fig:2dsetup\]](#fig:2dsetup){reference-type="ref" reference="fig:2dsetup"}: MIR-FTIR) with a silicon carbide globar light source (spectral range 580--7000 cm\(^{-1}\)). After being collimated, the light beam passes the interferometer (potassium bromide (KBr) beam splitter), is reflected by a plane mirror, and exits the MIR source unit. The beam enters the polarization state preparation unit. Inside the polarization state preparation unit the beam passes a beam steering plane mirror assembly (m\(_1\)), a source selection and beam focusing assembly (R\(_1\)) and a rotation stage assembly (P\(_1\)-b\(_1\)-St\(_1\)), respectively. The beam steering plane mirror assembly (m\(_1\)) is composed of an opto-mechanic mount and a plane first surface gold mirror. The source selection and beam focusing assembly (R\(_1\)) comprises two sub-assemblies (detailed drawing: Fig. [\[fig:beamlift\]](#fig:beamlift){reference-type="ref" reference="fig:beamlift"} a), the rotatable plane mirror sub-assembly[^7] (stepper motor, axis extension, mechanic mount for axis extension, opto-mechanic mount for mirror, plane first surface gold mirror) and the beam focusing off-axis paraboloid stage sub-assembly (opto-mechanic mount for paraboloid, gold surface \(90^{\circ}\) off-axis paraboloid with an effective focal length of \(f_\text{e}=350\) mm[^8]), which focuses the beam onto the sample position. The focused beam then reaches the rotation stage assembly (P\(_1\)-b\(_1\)-St\(_1\)) which has a nominal angular resolution of 0.045\(^\circ\). The rotation stage assembly (detailed drawing: Fig. [\[fig:beamlift\]](#fig:beamlift){reference-type="ref" reference="fig:beamlift"} b) contains a KRS-5 substrate based wire grid polarizer (P\(_1\)), which is mounted in a hollow tube, which is fitted into two polymer bearings with glass balls, which are held by an aluminum block. A 48-tooth polymer gear is mounted to the hollow tube, and is connected via a Kevlar timing belt (b\(_1\)) to a 12-tooth polymer gear on a stainless steel shaft (gear ratio: 1:4), leading to the stepper motor (St\(_1\)). After passing the rotation stage assembly, the polarized and focused beam leaves the polarization state preparation unit. The beam is then reflected by, or transmitted through the sample (S\(_1\)). The sample can be mounted on a sample holder, attached to the MIR goniometer unit (G\(_1\)) (commercially available 2-circle goniometer 415, Huber Diffraktionstechnik), or inside the magneto-cryostat sub-system (M1). If the magneto-cryostat sub-system is used, reflection type measurements can only be conducted at a \(\Phi_a=45^{\circ}\) angle of incidence. A detailed description of the magneto-cryostat sub-system, its sample mount, and the optical window configuration is given in section ([3.3](#sec:Magnet){reference-type="ref" reference="sec:Magnet"}). The beam then enters the polarization state detection unit, which is mounted to the rotatable arm of the MIR goniometer unit (G\(_1\)). The polarization state detection unit contains a rotation stage assembly (A\(_1\)-b\(_2\)-St\(_2\)), a beam collimation and detector selection assembly (R\(_2\)), and three beam focusing/detection assemblies (o\(_1\)-MCT, o\(_2\)-DTGS\(_1\) and o\(_3\)-B\(_1\)). The rotation stage assembly (A\(_1\)-b\(_2\)-St\(_2\)) is equivalent to the one in the polarization state preparation unit (P\(_1\)-b\(_1\)-St\(_1\), Fig. [\[fig:beamlift\]](#fig:beamlift){reference-type="ref" reference="fig:beamlift"} b), but in the polarization state detection unit the KRS-5 substrate based wire grid polarizer serves as the analyzer (A\(_1\)) of the MIR ellipsometer sub-system. The beam collimation and detector selection assembly (R\(_2\)) is composed of two sub-assemblies: the rotatable plane mirror sub-assembly and the collimating off-axis paraboloid stage sub-assembly (gold surface \(90^{\circ}\) off-axis paraboloid, \(f_\text{e}=350\) mm). Both sub-assemblies are equivalent to the source selection and beam focusing assembly in the polarization state preparation unit (Fig. [\[fig:beamlift\]](#fig:beamlift){reference-type="ref" reference="fig:beamlift"} a), but are used in reverse order to first collimate and then redirect the beam to one of the three beam focusing/detection assemblies using the rotatable mirror. Each beam focusing/detection assembly contains a beam focusing off-axis paraboloid stage sub-assembly and a detector sub-assembly. All detector sub-assemblies contain an opto-mechanic mount and a detector. The beam focusing off-axis paraboloid stage sub-assemblies (o\(_\text{1-3}\)) are composed of an opto-mechanic mount and a 90\(^{\circ}\) off-axis paraboloid with an uncoated gold surface. The focal lengths of the off-axis paraboloids are matched to the corresponding detector. The off-axis paraboloids (o\(_\text{1}\), o\(_\text{2}\)) for the liquid nitrogen cooled HgCdTe-detector sub-assembly (MCT) and the pyroelectric, solid state deuterated triglycine sulfate detector sub-assembly (DTGS\(_\text{1}\)), each have an effective focal length of 38 mm (1.5 in). The off-axis paraboloid o\(_\text{3}\) for the liquid helium cooled bolometer detector sub-assembly (B\(_\text{1}\)) has an effective focal length of 190.5 mm (7.5 in). The signal of the used detector is fed back into the MIR-FTIR-spectrometer to record interferograms. For more information on the data acquisition and processing see section [2.3](#sec:data analysis){reference-type="ref" reference="sec:data analysis"}. ## FIR/THz ellipsometer sub-system {#sec:FIR-THz} The lower part of Fig. [\[fig:2dsetup\]](#fig:2dsetup){reference-type="ref" reference="fig:2dsetup"} shows a schematic drawing (top view) of the optical configuration of the FIR/THz ellipsometer sub-system of the integrated MIR, FIR and THz OHE instrument. The FIR/THz ellipsometer sub-system can be divided in five units: (i) the FIR source unit, (ii) the THz source unit, (iii) the polarization state preparation unit, (iv) the FIR/THz goniometer unit, and (v) the polarization state detection unit. For measurements in the FIR spectral range, the FIR/THz ellipsometer sub-system is operated in analyzer-step mode, while for measurements in the THz spectral range the FIR/THz ellipsometer sub-system is operated in continuously rotating analyzer mode. To minimize absorption due to water vapor, the complete beam path of the FIR/THz ellipsometer sub-system can be purged with dried air. Due to the high magnetic stray-fields (see Fig. [\[fig:CAD\]](#fig:CAD){reference-type="ref" reference="fig:CAD"}), all opto-mechanical components in the THz source unit, polarization state preparation unit and polarization state detection unit were designed and manufactured without ferromagnetic materials (with exception of the stepper motors and the THz source). The FIR source unit of the FIR/THz ellipsometer sub-system is a Bruker Vertex V-70 FTIR-spectrometer (Fig. [\[fig:2dsetup\]](#fig:2dsetup){reference-type="ref" reference="fig:2dsetup"}: FIR-FTIR). The spectrometer is equipped with a silicon beam splitter but otherwise identical to the spectrometer used as MIR source unit in the MIR ellipsometer sub-system.\ The THz source unit comprises five assemblies: the THz source and THz-beam collimation assembly (Fig. [\[fig:2dsetup\]](#fig:2dsetup){reference-type="ref" reference="fig:2dsetup"}: BWO-o\(_\text{4}\)), the optical chopper assembly (Ch), the beam steering plane mirror assembly (m\(_\text{3}\)), polarization state rotator assembly (PR-b\(_\text{3}\)-St\(_\text{3}\)), and the beam steering plane mirror assembly (m\(_\text{4}\)). The THz source and THz-beam collimation assembly contains a THz source sub-assembly and a THz-beam collimating off-axis paraboloid stage sub-assembly. The THz source sub-assembly is composed of an opto-mechanic mount and the backward wave oscillator (BWO) THz source (Microtech). The BWO source emits THz radiation with a high brilliance (bandwidth \(\sim\)`<!-- -->`{=html}2 MHz) and a high output power (\(\sim\)`<!-- -->`{=html}0.1--0.01 W). The THz radiation is almost perfectly linearly polarized and the orientation of the polarization is fixed in space. The base frequency range of the BWO is 107--177 GHz, which can be converted to higher frequency bands using GaAs Schottky diode frequency multipliers. The spectral range accessible by the BWO can be expanded to 220--350 GHz (\(\times 2\) multiplier), 330--525 GHz (\(\times 3\) multiplier), 650--1040 GHz (\(\times 2\) and \(\times 3\) multiplier) and 980--1580 GHz (double \(\times 3\) multiplier). For further details on the BWO based THz source and THz ellipsometry are given in Ref.  and references therein. The THz radiation from the BWO is collimated by the THz-beam collimating off-axis paraboloid stage sub-assembly, composed of an opto-mechanic mount and a 90\(^{\circ}\) off-axis paraboloid (o\(_\text{4}\)) with an uncoated gold surface and an effective focal length \(f_e=60\) mm. The THz-beam then reaches the optical chopper assembly (Ch), which contains an opto-mechanic mount and a 3 bladed optical chopper, driven by a linear motor. The 3 bladed optical chopper is rotated with a frequency of \(f_c=3.8\) Hz, resulting in a optical chopping frequency of \(f_o=11.4\) Hz, which is close to the optimal frequency response of the Golay cell detector (\(f_\text{opt}\sim\) 12--15 Hz). After interaction with THz-beam steering plane mirror assembly (m\(_\text{3}\)) (opto-mechanic mount and plane first surface gold mirror), the THz-beam is redirected to the polarization state rotator assembly (PR). The polarization state rotator assembly (Fig. [\[fig:PR\]](#fig:PR){reference-type="ref" reference="fig:PR"}) is an odd-bounce image rotation system. The polarization state rotator is designed to rotate the polarization state of an incoming electromagnetic beam azimuthally in a non-deviating, non-displacing fashion (with respect to the incoming electromagnetic beam direction), and is used to pre-align the polarization direction of the THz-beam with the polarizing axis of the wire-grid polarizer in the polarization state preparation unit.[^9] The polarization state rotator assembly is composed of a stepper motor (St\(_3\)) with a 12-tooth polymer gear, which is connected to a 48-tooth polymer gear (gear ratio: 1:4) via a Kevlar timing belt (b\(_3\)), rotating a PEEK cage (PR) which contains three opto-mechanic mounts with plane first surface gold mirrors (rotation axis parallel to the incoming and outgoing THz-beam). After reflection on a THz-beam steering plane mirror assembly (m\(_\text{4}\)) (opto-mechanic mount and plane first surface gold mirror), the THz-beam leaves the THz source unit.\ The polarization state preparation unit contains four assemblies: a FIR-beam steering plane mirror assembly (m\(_\text{6}\)), a THz-beam steering plane mirror assembly (m\(_\text{5}\)), a source selection and beam focusing assembly (R\(_\text{3}\)), and a rotation stage assembly (P\(_\text{2}\)-b\(_\text{4}\)-St\(_\text{4}\)). The FIR-and THz-beam steering plane mirror assemblies (m\(_\text{6,5}\)) are identical and both are composed of an opto-mechanic mount and a plane first surface gold mirror. The plane first surface gold mirrors redirect, depending on the spectral range the FIR/THz ellipsometer sub-system is operated in, the FIR-or THz-beam to the source selection and beam focusing assembly. The source selection and beam focusing assembly (R\(_\text{3}\)) comprises two sub-assemblies, the rotatable plane mirror sub-assembly and the beam focusing off-axis paraboloid stage sub-assembly, which are equivalent to those in the source selection and beam focusing assembly in the polarization state preparation unit of the MIR ellipsometer sub-system (Fig. [\[fig:beamlift\]](#fig:beamlift){reference-type="ref" reference="fig:beamlift"} a). Depending on the orientation of the plane first surface gold mirror in the rotatable plane mirror sub-assembly, either the FIR-or THz beam is directed to beam focusing off-axis paraboloid stage sub-assembly (gold surface \(90^{\circ}\) off-axis paraboloid, \(f_\text{e}=350\) mm). The focused beam is then routed through the rotation stage assembly, which contains two polyethylene substrate based wire-grid polarizers (P\(_\text{2}\)), but is otherwise identical to the rotation stage assembly in the polarization state preparation unit of the MIR ellipsometer sub-system (Fig. [\[fig:beamlift\]](#fig:beamlift){reference-type="ref" reference="fig:beamlift"} b). The beam then leaves the polarization state preparation unit. The beam is then reflected by, or transmitted through the sample (S\(_\text{2}\)). The sample can be mounted on a sample holder, attached to the FIR/THz goniometer unit (G\(_\text{2}\)) (commercially available, 2-circle goniometer 415, Huber Diffraktionstechnik), or inside the magneto-cryostat sub-system (M2). If the magneto-cryostat sub-system is used, reflection type measurements can only be conducted at \(\Phi_a=45^{\circ}\) angle of incidence. A detailed description of the magneto-cryostat sub-system, its sample mount, and the optical window configuration is given in section [3.3](#sec:Magnet){reference-type="ref" reference="sec:Magnet"}. The beam then enters the polarization state detection unit, which comprises a rotation stage assembly (A\(_\text{2}\)-b\(_\text{5}\)-St\(_\text{5}\)), a beam collimation and detector selection assembly (R\(_\text{4}\)), and three beam focusing/detection assemblies (m\(_\text{7}\)-o\(_\text{5}\)-g, o\(_\text{6}\)-DTGS\(_\text{2}\) and o\(_\text{7}\)-B\(_\text{2}\)). The beam is routed through the rotation stage assembly for the analyzer of the FIR/THz ellipsometer sub-system, which contains two polyethylene substrate based wire-grid polarizers (A\(_\text{2}\)), but is otherwise identical to the rotation stage assembly in the polarization state detection unit of the MIR ellipsometer sub-system (Fig. [\[fig:beamlift\]](#fig:beamlift){reference-type="ref" reference="fig:beamlift"} b). The rotation stage assembly for the analyzer of the FIR/THz ellipsometer sub-system can be operated in step mode for FIR measurements or in continuous rotation mode for THz measurements. The beam is then collimated (gold surface \(90^{\circ}\) off-axis paraboloid, \(f_\text{e}=350\) mm) and redirected to the selected detector by the beam collimation and detector selection assembly (R\(_\text{4}\)), identical to the beam collimation and detector selection assembly in the polarization state detection unit of the MIR ellipsometer sub-system (Fig. [\[fig:beamlift\]](#fig:beamlift){reference-type="ref" reference="fig:beamlift"} a). The beam focusing and Golay-cell-detector assembly (m\(_\text{7}\)-o\(_\text{5}\)-g) contains a beam steering plane mirror sub-assembly (m\(_\text{7}\)) (opto-mechanic mount, plane first surface gold mirror), a beam focusing off-axis paraboloid stage sub-assembly (o\(_\text{5}\)) (opto-mechanic mount, gold surface \(90^{\circ}\) off-axis paraboloid, \(f_\text{e}=60\) mm), and a Golay-cell detector sub-assembly (g) (opto-mechanic mount, Golay-cell detector). The beam focusing and DTGS detector assembly (o\(_\text{6}\)-DTGS\(_\text{2}\)) comprises a beam focusing off-axis paraboloid stage sub-assembly (o\(_\text{6}\)) (opto-mechanic mount, gold surface \(90^{\circ}\) off-axis paraboloid, \(f_\text{e}=38\) mm), and a solid state deuterated triglycine sulfate detector sub-assembly (DTGS\(_\text{2}\)) (opto-mechanic mount, Bruker Vertex V-70 DTGS detector). Alternatively, the beam focusing and bolometer detector assembly (o\(_\text{7}\)-B\(_\text{2}\)), composed of a beam focusing off-axis paraboloid stage sub-assembly (o\(_\text{7}\)) (opto-mechanic mount, gold surface \(90^{\circ}\) off-axis paraboloid, \(f_\text{e}=190.5\) mm) and the bolometer detector sub-assembly (B\(_\text{2}\)) (opto-mechanic mount; commercially available, liquid helium cooled bolometer detector, Infrared Laboratories Inc.), can be used. For THz measurements the bolometer or the golay cell detector can be chosen, while for FIR measurements only the bolometer or the DTGS detector provide frequency responses fast enough to record interferograms. ## Magneto-cryostat sub-system {#sec:Magnet} The central piece of equipment of the integrated MIR, FIR and THz OHE instrument is the commercially available, superconducting, closed cycle magneto-cryostat sub-system (7T-SpectromagPT, Oxford Instruments) with four optical ports (Fig. [\[fig:CAD\]](#fig:CAD){reference-type="ref" reference="fig:CAD"}). The design of the integrated MIR, FIR and THz OHE instrument allows the usage of the magneto-cryostat sub-system with the MIR and the FIR/THz ellipsometer sub-system by employing the magneto-cryostat transfer sub-system. The magneto-cryostat sub-system can be subdivided into the magnet head, the primary cooling cycle, the secondary cooling cycle, the sample holder and the optical windows. The magnet head contains two magnet coils, which are mounted around the sample position and are fabricated in a split-coil pair design. In a spherical volume of 10 mm diameter around the sample, magnetic fields up to \(B=8\) T with an inhomogeneity of less than 0.3 % can be achieved. The magnetic field can be reversed and points towards one of the optical windows. Therefore, for reflection type OHE measurements, the magnetic field lies within the plane of incidence and forms an angle of \(45^{\circ}\) with the sample normal. This leads to a magnetic field component \(B_c=|\mathbf{B}|/\sqrt{2}\) perpendicular to the sample surface The primary cooling cycle comprises a pulse tube cooler (SRP-082, SHI Cryogenics), high pressure helium lines and a helium compressor (F-70, Sumitomo Heavy Industries). Ultra-high-purity helium (UHP-He) gas at high pressure is provided by the helium compressor, and guided by a high pressure helium line to the pulse tube cooler. The pulse tube cooler is thermally coupled to the superconducting magnet coils and allows to cool the magnet coils to temperatures of \(T\approx 3.1\) K. The pulse tube cooler also pre-cools the UHP-He in the secondary cooling cycle. The UHP-He gas is then guided back to the helium compressor by a high pressure helium line, and is reused. The secondary cooling cycle uses UHP-He gas to cool the sample. The UHP-He gas in the closed cycle, is circulated by an oil-free, dry scroll pump (XDS-10, Edwards). When the sample is at base temperature, the UHP-He gas leaves the outlet of the scroll pump at a pressure of \(\sim\) 0.5 bar, and is pumped through a zeolite and a liquid nitrogen trap, in order to extract possible contaminants leaking into the closed cycle. The UHP-He gas then passes a distiller spiral, which is thermally coupled to the pulse tube cooler (primary cooling cycle), and condenses the UHP-He gas into liquid helium (LHe). The LHe flow is controlled by a needle valve, and the LHe is then injected into a heat exchanger. The heat exchanger is attached to the double-walled, hollow variable temperature inset (VTI) cryostat (Fig. [\[fig:sampleholder\]](#fig:sampleholder){reference-type="ref" reference="fig:sampleholder"}). Finally, the scroll pump reduces the gas pressure above the LHe to 1.5-3 mbar and thereby cools the VTI to a minimal temperature of \(T=1.4\) K. The sample is thermally coupled to the VTI by a static exchange gas (UHP-He). A resistive heater allows to warm the sample up to room temperature, without bringing the temperature of the superconducting magnet coils above the critical temperature. The sample holder of the magneto-cryostat sub-system (Fig. [\[fig:sampleholder\]](#fig:sampleholder){reference-type="ref" reference="fig:sampleholder"}) can hold up to two samples at a time. If two samples are mounted the optimal sample size is \(0.5\times 12\times 12\) mm\(^3\), while the maximum sample size is \(1\times 30\times 30\) mm\(^3\) (only one sample can be mounted). The sample position is adjustable in the vertical direction (\(\pm15\) mm) and rotationally around the vertical axis of the VTI by 360\(^{\circ}\) (angle of incidence alignment). All other degrees of freedom necessary for sample alignment (linear motion, rotation, tip/tilt) can only be accessed by moving the sample together with the magneto-cryostat sub-system. The magneto-cryostat sub-system can be moved parallel to the incoming beam, using the magneto-cryostat transfer sub-system. The alignment perpendicular to the incoming beam and the rotational alignment of the magneto-cryostat sub-system is achieved by sliding the magneto-cryostat sub-system in the magneto-cryostat holding frame of the magneto-cryostat transfer sub-system (see sec. [3.4](#sec:Magnet_transfer){reference-type="ref" reference="sec:Magnet_transfer"}). Four screws in the same magneto-cryostat holding frame of the magneto-cryostat transfer sub-system are used for tip/tilt alignment. After successful alignment, in order to minimize motion, the magneto-cryostat sub-system is clamped to the magneto-cryostat holding frame of the magneto-cryostat transfer sub-system. When the light beam is routed through the magnet head, it passes a set of exterior and interior optical windows. For measurements in the FIR/THz spectral range, the exterior optical windows are made of 0.27 mm thick homo-polypropylene films, while for the MIR spectral range potassium bromide (KBr) windows are used. All four exterior windows are purged on the exterior side with dried air, to prevent condensation of moisture from the ambient air. The exterior optical windows of the magnet can be exchanged and arranged for both transmission-and reflection-type measurements. The latter window configuration allows reflection-type measurements over the full spectral range of the OHE instrument, by simply moving the magneto-cryostat between the ellipsometer sub-systems without warming up the superconducting coils or the sample. The interior four optical windows on the VTI are made of polished diamond, grown by chemical vapor deposition (CVD), with a thickness of 0.35 mm, a diameter of 14 mm and an average surface roughness of \(Ra\leq 15\) nm (arithmetic average). To reduce Fabry-Pérot interferences in the interior optical windows the interior windows were wedged by an angle of \(\sim\)`<!-- -->`{=html}0.5\(^{\circ}\). The interior windows were glued into stainless steel window frames by a cryogenic two component epoxy, leaving a clear aperture of \(\sim\)`<!-- -->`{=html}12 mm (see cutout in Fig. [\[fig:sampleholder\]](#fig:sampleholder){reference-type="ref" reference="fig:sampleholder"}). In general, optical windows affect experimental ellipsometry data, especially if the optical window material is birefringent. Therefore all optical windows were characterized by transmission GE on a commercial MIR ellipsometer (J.A.Woollam Co., Inc.). No birefringence was observed in the MIR spectral range. Nevertheless, through mounting of the optical windows, and in particular through stress due to the vacuum in the magneto-cryostat sub-system, strain induced birefringence cannot be completely excluded and therefore window effects are included in the analysis of the OHE data (see sec. [4.2](#sec:OHE-analysis){reference-type="ref" reference="sec:OHE-analysis"}). ## Magneto-cryostat transfer sub-system {#sec:Magnet_transfer} The magneto-cryostat transfer sub-system (Fig. [\[fig:CAD-overview\]](#fig:CAD-overview){reference-type="ref" reference="fig:CAD-overview"}) contains the magneto-cryostat transfer frame, the MIR source unit frame and the FIR/THz source unit frame. The magneto-cryostat transfer frame comprises the magneto-cryostat transfer assembly, the MIR goniometer unit platform assembly and FIR/THz goniometer unit platform assembly. The magneto-cryostat transfer assembly is built from alumina extrusions. The grooves of the alumina extrusions are used as guides for a Teflon roll based rail sub-assembly. On top of the Teflon roll based rail sub-assembly the magneto-cryostat holding frame (black alumina extrusions in Fig. [\[fig:CAD-overview\]](#fig:CAD-overview){reference-type="ref" reference="fig:CAD-overview"} and [\[fig:CAD\]](#fig:CAD){reference-type="ref" reference="fig:CAD"}) is mounted, which is used for alignment purposes (see sec. [3.3](#sec:Magnet){reference-type="ref" reference="sec:Magnet"}). The MIR and FIR/THz goniometer unit platform assemblies are mounted into the magneto-cryostat transfer frame and are platforms for the MIR and FIR/THz goniometer unit of the ellipsometer sub-systems. The MIR and FIR/THz source unit frames are built from alumina extrusions, equipped with a 19-inch rack mounting system, and provide platforms for the source units of the MIR and FIR/THz ellipsometer sub-systems, respectively. # OHE data acquisition and analysis {#sec:ACQ-ANAL} ## OHE data acquisition {#sec:OHE-acq} All spectroscopic ellipsometer sub-systems of the integrated MIR, FIR and THz OHE instrument are operated in a \(PSA_R\) configuration, capable to measure the upper left \(3\times 3\) block of the Mueller matrix. Note that this does not affect the ability to determine certain sample properties related to anisotropy, which can be obtained from the off-diagonal-block elements \(M_{13}\), \(M_{23}\), \(M_{31}\) and \(M_{32}\). For rotating-analyzer based ellipsometers with lossless and ideal polarizing optical elements, the stokes vector of a beam of light at the detector \(\mathbf{I}^{\text{D}}\) can be described within the Mueller matrix formalism by \[\label{eqn:MatMulElli} \mathbf{I}^{\text{D}} \hspace{-2pt}=\hspace{-2pt} \mathbf{R}(-\theta_\text{A}) \hspace{2pt \mathbf{A} \hspace{2pt \mathbf{R}(\theta_\text{A}) \hspace{2pt \mathbf{M} \hspace{2pt \mathbf{R}(-\theta_\text{P}) \hspace{2pt \mathbf{P} \hspace{2pt \mathbf{R}(\theta_\text{P}) \hspace{2pt \mathbf{I}^{\text{S}}\;,\] where \(\mathbf{I}^{\text{S}}\) is the stokes vector of the light leaving the source, and \(\mathbf{P}\), \(\mathbf{A}\), \(\mathbf{R}\left(\theta_j\right)\) and \(\mathbf{M}\) are the Mueller matrices of a polarizer, analyzer, coordinate rotation by the angle \(\theta_j\) (polarizer: \(j=\text{P}\); analyzer: \(j=\text{A}\)) and sample, respectively If the analyzer is rotated with a constant angular frequency \(\omega_\text{A}\), and both, light source and detector exhibit no polarization dependency, *e.g.* the source emits unpolarized light with intensity \(I^{\text{S}}_0\) and the detector is only sensitive to the total intensity \(I^{\text{D}}_0\),[^10] the ratio of these quantities is \[\label{eqn:fouriercoef} \begin{split} \frac{I^{\text{D}}_0}{I^{\text{S}}_0}\hspace{-2pt} = \hspace{-2pt} \frac{1}{4} \hspace{-2pt} \left[ \lambda(\mathbf{M}\hspace{-1pt},\hspace{-1pt}\theta_{\hspace{-0pt}\text{P}}\hspace{-2pt})\hspace{-2pt} \right. & + \hspace{-2pt} \alpha(\mathbf{M}\hspace{-1pt},\hspace{-1pt}\theta_{\hspace{-0pt}\text{P}}\hspace{-2pt}) \cos (2 \omega_\text{A} t) \hspace{-2pt}\\ & + \left. \hspace{-2pt} \beta(\mathbf{M}\hspace{-1pt},\hspace{-1pt}\theta_{\hspace{-0pt}\text{P}}\hspace{-2pt}) \sin(2 \omega_\text{A} t) \right]\;, \end{split}\] with the time harmonic Fourier coefficients \[\label{eqn:NCSparameter} \begin{split} \lambda (\mathbf{M},\theta_\text{P}) =& M_{11} + M_{12} \cos(2 \theta_\text{P}) + M_{13} \sin(2 \theta_\text{P})\\ \alpha (\mathbf{M},\theta_\text{P}) =& M_{21} + M_{22} \cos(2 \theta_\text{P}) + M_{23} \sin(2 \theta_\text{P})\\ \beta (\mathbf{M},\theta_\text{P}) =& M_{31} + M_{32} \cos(2 \theta_\text{P}) + M_{33} \sin(2 \theta_\text{P})\;. \end{split}\] Individual Mueller matrix elements \(M_{ij}\) are determined from Fourier coefficients measured at different polarizer orientations \(\theta_\text{P}\) on the input side (see subroutine Mueller matrix regression in Fig. [\[fig:flow_aqu\]](#fig:flow_aqu){reference-type="ref" reference="fig:flow_aqu"}). The complete sequence of operations executed by the OHE instrument during the GE and OHE data acquisition is summarized in Fig. [\[fig:flow_aqu\]](#fig:flow_aqu){reference-type="ref" reference="fig:flow_aqu"}. Since the intensity of light at the detector and the Fourier coefficients (employed to determine the Mueller matrix elements) depend on the absolute angular positions of all rotating optical elements \[Eq. ([\[eqn:fouriercoef\]](#eqn:fouriercoef){reference-type="ref" reference="eqn:fouriercoef"}) and ([\[eqn:NCSparameter\]](#eqn:NCSparameter){reference-type="ref" reference="eqn:NCSparameter"})\], the knowledge of these absolute angular positions is crucial for the operation of every ellipsometer. Furthermore, all spectroscopic ellipsometers have to account for non-ideal polarization characteristics of sources and detectors, as well as commonly present non-idealities of MIR and FIR optical elements (such as polarizers), in their calibration routine. The absolute positions of the optical components used in the OHE instrument, with respect to the *p*-and *s*-coordinate system (the *s*-axis is parallel to the goniometer axis, the *p*-axis is perpendicular to the beam), as well as the non-idealities of the optical elements are carefully calibrated prior to the OHE data acquisition. All ellipsometers sub-systems are calibrated prior to the experiments ## OHE data analysis {#sec:OHE-analysis} During OHE data analysis, additional non-idealities, not included in the ellipsometer calibration, introduced by the magneto-cryostat sub-system have to be considered. In particular, three effects are considered: (i) in-and-out-of-plane anisotropy in the optical window, (ii) changes in the alignment of mirrors and/or off-axis paraboloids in the polarization state preparation and detection units due to magnetic forces on ferromagnetic components (ellipsometric coordinate system change), and (iii) imperfect sample alignment (ellipsometric coordinate system and angle-of-incidence change). In order to account for sample misalignment in the model based data analysis described in section [2.3](#sec:data analysis){reference-type="ref" reference="sec:data analysis"}, the angle of incidence and the sample tilt angle[^11] are included as model parameters. In order to model the combined non-idealities due to strain induced birefringence in the optical windows and minute changes in the alignment of optical elements on the input (output) side, an additional Mueller matrix \(\mathbf{M}_{\text{\tiny{in}}}\) (\(\mathbf{M}_{\text{\tiny{out}}}\)) was included in the model based data analysis (Sec. [2.3](#sec:data analysis){reference-type="ref" reference="sec:data analysis"}). The Mueller matrices \(\mathbf{M}_{\text{\tiny{in}}}\) and \(\mathbf{M}_{\text{\tiny{out}}}\) are assumed to have no dispersion (wavelength independent). The best-model Mueller matrix \(\mathbf{M}_{\text{\tiny{best}}}\), used for MSE regression as described in section [2.3](#sec:data analysis){reference-type="ref" reference="sec:data analysis"} reads \[\mathbf{M}_{\text{\tiny{best}}} = \mathbf{M}_{\text{\tiny{out}}} \mathbf{M}_{\text{\tiny{mod}}} \mathbf{M}_{\text{\tiny{in}}} \;,\] where \(\mathbf{M}_{\text{\tiny{mod}}}\) represents the sample Mueller matrix calculated by the \(4\times 4\) matrix formalism.[^12] For the analysis of OHE data, two strategies may be used exploiting either model-free or model-based approaches. Model free analysis provides semi-quantitative results, by studying trends in amplitudes or spectral positions of features in OHE or derived OHE data vs. the magnitude of the magnetic field \(|\mathbf{B}|\). The model free analysis can provide insight into the symmetry properties of magneto-optic dielectric tensors, and an example for the model free analysis of derived OHE data \(\delta\mathbf{M}^{+}\) vs. \(|\mathbf{B}|\) is given in section [5.1](#sec:Graphene){reference-type="ref" reference="sec:Graphene"} for the case of epitaxial graphene. The model-based data analysis approach provides more quantitative parameters than the model-free data analysis approach, and can be used to determine model parameters such as the free charge carrier concentration or the effective mass parameters. During the data analysis the GE, OHE and derived OHE datasets are analyzed simultaneously to determine physical model parameters of \(\eps_{_{\hspace{-1pt} \mathbf{B}=0}}\) and \(\eps_{_{\hspace{-2pt}\mathbf{B}}}\) by a single, consistent optical model (Fig. [\[fig:flow\]](#fig:flow){reference-type="ref" reference="fig:flow"}). - In the first step, \(\mathbf{M}_0=\mathbf{M}(\eps_{_{\hspace{-1pt}\mathbf{B}=0}})\) (GE data obtained at \(B=0\) T) is analyzed only. During the analysis, all model parameters *independent* of the magnetic field, such as those describing, for example, the polar lattice resonances or layer thickness are varied until the calculated data match the measured data as closely as possible. - In the second step, OHE data \(\mathbf{M}^\pm_{\text{\tiny{OHE}}}\) for \(B\neq 0\) T is included in the analysis, and all model parameters *dependent* on the magnetic field, such as, for example, effective mass or inter-Landau-level transition parameters are varied. In addition, derived OHE data \(\delta\mathbf{M}^{\pm}\) and/or \(\delta\mathbf{M}^{+}\pm\delta\mathbf{M}^{-}\) can be included into the simultaneously analyzed dataset. These derived OHE datasets do not increase the amount of information collected during the experiment, but help (*a*) to visualize magnetic field induced changes, and (*b*) improve the sensitivity to magnetic-field dependent model parameters during OHE data analysis.[^13] The parameters determined in the first analysis step are used as starting values for the second analysis step and are varied if necessary. Note that an optical model, describing the GE dataset (\(B=0\) T) sufficiently well, might not be capable to describe the complete dataset including OHE and/or derived OHE data correctly. If necessary the optical model has to be adjusted, and the data analysis procedure has to be repeated. Additional information on the data analysis strategies can be found in Ref. . # Results and Discussions {#sec:samplesystems} In this section, experimental data from the integrated MIR, FIR and THz OHE instrument is presented. First, a combined dataset of derived OHE data from the MIR, FIR and THz spectral range from 7.5 cm\(^{-1}\) to 4000\(^{-1}\) (0.22--120 THz or 0.9--500 meV), and a magnetic field of \(|\mathbf{B}|=\)`<!-- -->`{=html}4 T, for an epitaxial graphene sample grown on 6*H*-SiC is shown (Fig. [\[fig:full_range\]](#fig:full_range){reference-type="ref" reference="fig:full_range"}). In the spectral range below approximately 500-600 cm\(^{-1}\) the experimental data reveals an OHE signal enhanced by Fabry-Pérot interferences in the 6*H*-SiC. An OHE signal, enhanced by coupling with the 6*H*-SiC phonon mode, is observed near 1000 cm\(^{-1}\). Furthermore, between 500 and 4000 cm\(^{-1}\) inter-Landau-level transitions in single layer graphene are observed in the on-diagonal block elements of the Mueller matrix. Then derived OHE data from the individual MIR, FIR and THz spectral ranges of the integrated MIR, FIR and THz OHE instrument, and the corresponding best-model calculated data are shown exemplarily. We present results from OHE experiments on an epitaxial graphene sample grown on 6*H*-SiC, a Te doped n-type GaAs substrate and an AlGaN/GaN high electron mobility transistor structure (HEMT), representing the MIR, FIR, and THz spectral range of the integrated MIR, FIR and THz OHE instrument, respectively. The selected experimental datasets demonstrate the full spectral, magnetic field and temperature range of the integrated MIR, FIR and THz OHE instrument, as well as analysis strategies. Effects from free charge carriers in bulk and in two dimensional confinement as well as quantum mechanical effects (inter-Landau-level transitions) are observed and discussed. ## The MIR optical Hall effect---Graphene {#sec:Graphene} Exemplarily, epitaxial graphene on 6*H*-SiC was investigated to demonstrate the MIR ellipsometer sub-system of the OHE instrument (Sec. [3.1](#sec:MIR){reference-type="ref" reference="sec:MIR"}). The epitaxial graphene sample was grown in an argon atmosphere at 1400 \(^\circ\)C, by sublimating Si from the polar *c*-face (000\(\bar{1}\)) of a semi-insulating 6*H*-SiC substrate. From previous measurements on C-face 4*H*-SiC , the number of graphene layers is estimated to be 10-20. Further details on growth conditions are beyond the scope of this manuscript, and can be found in Ref. . The OHE experiment was conducted at an angle of incidence of \(\Phi_a\) = 45\(^\circ\), while the magnetic field was aligned along the reflected beam. Experimental data was recorded in the spectral range from 600 to 4000 cm\(^{-1}\) with a spectral resolution of 1 cm\(^{-1}\), using the HgCdTe detector, while the sample was kept at a temperature of \(T=1.5\) K. Figure [\[fig:Graphene1\]](#fig:Graphene1){reference-type="ref" reference="fig:Graphene1"} shows derived OHE data \(\delta\mathbf{M}^{+}\) (green, dotted line) for \(B=8\) T (effective field parallel to sample normal \(B_c=|\mathbf{B}|/\sqrt{2}\sim 5.66\) T) and results from best model calculations (red, solid line). The model free analysis approach provides valuable information. The OHE data shows several resonances, which can be divided in three groups: a peak near \(\nu=1000\) cm\(^{-1}\) labeled FCC (triangles, pink), a set of features labeled LL\(_{\stext{BLG}}\) (rectangles, blue) and a set of peaks labeled LL\(_{\stext{SLG}}\) (diamonds, orange). While the resonances labeled FCC and LL\(_{\stext{BLG}}\) are present in all Mueller matrix elements, the features labeled LL\(_{\stext{SLG}}\) are only present in the on-diagonal-block elements of the Mueller matrix. Since non-vanishing off-diagonal-block elements in the Mueller matrix are inherently tied to non-vanishing off-diagonal elements in the underlying dielectric tensor, the dielectric tensor has to be diagonal \((\varepsilon_{xy}=\varepsilon_{yx}=0)\). Furthermore, the magneto-optic contributions of LL\(_{\stext{SLG}}\) to the permittivity tensor in its representation for circularly polarized light \[Eq. ([\[eq:LLTeps\]](#eq:LLTeps){reference-type="ref" reference="eq:LLTeps"})\] must satisfy \(\chi_{\mathrm{+}}=\chi_{\mathrm{-}}\). In other words, the physical processes leading to the fingerprints labeled LL\(_{\stext{SLG}}\) are polarization conserving-while the processes leading to FCC and LL\(_{\stext{BLG}}\) are polarization mixing (polarization selection rules). The origin of the individual processes can be determined by field-dependent measurements. Figure [\[fig:Graphene2\]](#fig:Graphene2){reference-type="ref" reference="fig:Graphene2"} (c) displays the magnetic field dependence of a representative on-diagonal-block Mueller matrix element \(\delta M_{33}^{+}\) of the derived OHE dataset for \(B=1\dots 8\text{~T}\) in \(0.1\text{~T}\) increments (the corresponding representative off-diagonal-block Mueller matrix element is omitted here and the interested reader is referred to Ref. ). At a first glance it can be noted that the resonance labeled FCC increases in amplitude while its spectral position is not affected. This indicates that the resonance labeled FCC is the SiC phonon mode (best model parameter: \(l=1\), \(\varepsilon_{\infty,{\stext{x,y}}}=6.00\), \(\gamma_{\stext{LO},{\stext{x,y}}}=4.74\) cm\(^{-1}\), \(\omega_{\stext{LO},{\stext{x,y}}}=972.72\) cm\(^{-1}\), \(\gamma_{\stext{TO},{\stext{x,y}}}=1.18\) cm\(^{-1}\), \(\omega_{\stext{TO},{\stext{x,y}}}=799.31\) cm\(^{-1}\), \(\varepsilon_{\infty,{\stext{z}}}=5.84\), \(\gamma_{\stext{LO},{\stext{z}}}=2.64\) cm\(^{-1}\), \(\omega_{\stext{LO},{\stext{z}}}=966.99\) cm\(^{-1}\), \(\gamma_{\stext{TO},{\stext{z}}}=0.50\) cm\(^{-1}\), and \(\omega_{\stext{TO},{\stext{z}}}=798.73\) cm\(^{-1}\)) enhanced, magneto-optic fingerprint from free charge carriers. In contrast, the spectral positions of the resonances labeled LL\(_{\stext{SLG}}\) shift to higher energies and their amplitudes increase with increasing magnetic field strength. A detailed analysis of the peaks reveals a \(\sqrt{B}\)-dependence of their spectral positions, indicative for inter-Landau-level transitions in the Dirac-type band structure of single-layer graphene with the Fermi level close to the charge neutrality point . Landau level energies in single layered graphene are given by \(E_{\stext{SLG}}^{\stext{LL}}(n) =\mbox{sign}(n)E_0\sqrt{|n|}\) with \(E_0=\tilde{c}\sqrt{2\hbar e|B_c|}\), where \(\tilde{c}\) is the average velocity of Dirac fermions in graphene. Optical selection rules for transitions between levels \(n'\) and \(n\) demand \(|n'|=|n|\pm 1\). The Fermi velocity is determined as \(\tilde{c}=(1.01\pm0.01)\times 10^6\)m/s, in very good agreement with Refs. . The analysis of the magnetic field dependence of the resonances labeled LL\(_{\stext{BLG}}\) reveals their origin from the bi-layer branch of inter-Landau-level transitions in ABA-stacked (Bernal) multi-layered graphene. Inter-Landau-level transitions from bi-layer and tri-layer graphene were identified. For further details the interested reader is referred to Ref. . ## The FIR optical Hall effect---GaAs {#sec:GaAs} The *n*-type GaAs substrate investigated here, was moderately doped with tellurium (\(N\approx 10^{18}\) cm\(^{-3}\)) and is opaque to FIR radiation. All Measurements were conducted at an angle of incidence of \(\Phi_a=45^{\circ}\), while the magnetic field was parallel to the reflected beam. Figure [\[fig:GaAs_8T\]](#fig:GaAs_8T){reference-type="ref" reference="fig:GaAs_8T"} shows experimental data from the FIR spectral range (50--650 cm\(^{-1}\)) obtained at a sample temperature of \(T=300\) K. The sub-figures (a) and (b) display GE data (\(B=0\) T) and derived OHE data \(\delta\mathbf{M}^{+}-\delta\mathbf{M}^{-}\) for \(B=8\) T, respectively. Without magnetic field, GaAs is optically isotropic, i.e., for \(B=0\) T all off-diagonal-block elements (\(M_{13}\), \(M_{23}\), \(M_{31}\), and \(M_{32}\)) in (a) are zero. In contrast, due to the underlying magneto-optic dielectric tensor symmetry, all on-diagonal-block elements in (b) are zero, indicating the anti-symmetry of the off-diagonal elements of the dielectric tensor, as well as their sign change under field inversion. Note that for a bulk material and a field orientation that is not perpendicular to the sample surface, all off-diagonal elements of the dielectric tensor are different from zero. This also leads to the asymmetry between corresponding Mueller matrix elements with interchanged indices (\(ij\) and \(ji\)). The optical model for the GaAs substrate consists of a single, semi-infinite layer which contains two contributions, the optical response of polar lattice vibrations and free charge carriers described by the Drude model. Due to the crystal symmetry of GaAs, the dielectric tensor for the lattice vibration \[Eq. ([\[eqn:phonon2\]](#eqn:phonon2){reference-type="ref" reference="eqn:phonon2"}) and ([\[Substrate\]](#Substrate){reference-type="ref" reference="Substrate"})\] has identical diagonal elements, *e.g.* \(\varepsilon_{x}^{\text{\tiny{L}}}=\varepsilon_{y}^{\text{\tiny{L}}}=\varepsilon_{z}^{\text{\tiny{L}}}\) and possesses only one TO-LO resonance in the FIR spectral range. Best model parameter for the lattice vibrations were determined as \(\varepsilon_{\infty}=(9.269\pm 0.009)\), \(\omega_{\text{\tiny{TO}}}=(268.24\pm 0.06)\) cm\(^{-1}\), \(\omega_{\text{\tiny{LO}}}=(290.96\pm 0.07)\) cm\(^{-1}\) and \(\gamma_{\text{\tiny{TO}}}=\gamma_{\text{\tiny{LO}}}=(1.9\pm 0.2)\) cm\(^{-1}\). The model parameters for the free-charge-carrier contribution \(\log N=(17.926\pm 0.001)\) cm\(^{-3}\), \(\mu=(1789\pm 11)\) cm\(^{2}\)/Vs and \(m=(0.0738\pm 0.0001)\) m\(_{\text{e}}\) are in good agreement with results in previous publications. ## The THz optical Hall effect---HEMT {#sec:HEMT} The HEMT structure investigated here, was grown on a semi-insulating 4*H*-SiC substrate by metal-organic chemical vapor deposition at temperatures of 1050 \(^{\circ}\)C. First, a nominally 100 nm thick AlN nucleation layer was grown, followed by a 1800 nm thick GaN buffer and a 20 nm thick Al\(_{0.25}\)Ga\(_{0.75}\)N electron barrier layer. Figure [\[fig:HEMT\]](#fig:HEMT){reference-type="ref" reference="fig:HEMT"} displays the temperature dependence of non-vanishing off-diagonal-block Mueller matrix elements of derived OHE data \(\delta M_{31}^{+}-\delta M_{31}^{-}\) and \(\delta M_{32}^{+}-\delta M_{32}^{-}\) for \(B=3\) T in the spectral range from 0.22 to 0.32 THz. The experimental OHE data is depicted as dotted lines (green), while best-model calculated OHE data is plotted as solid lines (red) for different temperatures between 1.5 K and 300 K. The observed OHE is caused by free charge carriers in the high mobility channel of the HEMT structure, which is enhanced by the Fabry-Pérot interference in the \((356\pm 1)\) \(\mu\)m thick SiC substrate. The optical model is composed of a SiC substrate layer, a AlN nucleation layer, a GaN buffer layer, a layer for the GaN HEMT channel and a AlGaN layer. The thickness of the GaN HEMT layer was set to \(d=1\) nm and not varied during data analysis. The HEMT layer thickness was used to calculate the sheet charge density \(N_s=N d\), where \(N\) is the bulk charge density. Derived THz OHE data was analyzed simultaneously with GE data recorded with a commercial MIR ellipsometer. During model based OHE data analysis the parameters for the sheet charge density \(N_s\), the mobility \(\mu\) and the effective mass parameter \(m\) were determined. The sheet carrier density was found to be constant within the error limits at a value of \(N_s=N d=(5\pm 1) \times 10^{12}\) cm\(^{-2}\) for all temperatures. The mobility and the effective mass parameters were determined as \(m=(0.22\pm 0.01) m_0\) (\(m=(0.36\pm 0.03) m_0\)) and \(\mu=(7800\pm 410)\) cm\(^{2}\)/Vs (\(\mu=(1711\pm 150)\) cm\(^{2}\)/Vs) at \(T=1.5\) K (\(T=300\) K), where \(m_0\) is the mass of the free electron. Further details are omitted here and the interested reader is referred to an in depth discussion in Ref. . # Summary {#sec:Summary} In this article, we have given an overview over theoretical and experimental aspects of the OHE, and have successfully demonstrated the operation of an integrated MIR, FIR and THz optical Hall effect instrument. Two in-house built ellipsometers sub-systems, operating in the rotating-analyzer configuration, were employed to determine the upper \(3\times 3\) block of the normalized Mueller matrix in an ultra wide spectral range from 3 cm\(^{-1}\) to 7000 cm\(^{-1}\) (0.1--210 THz or 0.4--870 meV). Different aspects of the integrated MIR, FIR and THz optical Hall effect instrument, such as the cryogen-free, superconducting 8 Tesla magneto-cryostat sub-system, and the optical setup of the integrated MIR, FIR and THz optical Hall effect instrument, as well as the data acquisition and data analysis strategies were discussed in detail. For demonstration purposes of the operation of the integrated MIR, FIR and THz optical Hall effect instrument in the MIR, FIR and THz spectral range, three sample systems were studied: an epitaxial graphene sample (MIR), a GaAs substrate (FIR) and a AlGaN/GaN HEMT structure (THz). The presented data covered the full magnetic field range (\(B=0\dots 8\) T) and temperature range (\(T=1.5\dots 300\) K) of the system. We have demonstrated that the integrated MIR, FIR and THz optical Hall effect instrument can be used to determine the parameters of free charge carriers in bulk, and in two-dimensional confinement. Furthermore, it was shown that the integrated MIR, FIR and THz optical Hall effect instrument is capable to investigate the symmetry properties of magneto-optic, dielectric tensors of quantized systems, and was able to determine the polarization selection rules of inter-Landau-level transitions. The integrated MIR, FIR and THz optical Hall effect instrument has a wide range of applications, from determining free charge carrier properties contact-free, over the access to free charge carrier properties in buried layers, to magneto-optic quantum phenomena, polarization selection rules, to symmetry properties of magneto-optic dielectric tensors.

{'timestamp': '2014-01-22T02:10:37', 'yymm': '1401', 'arxiv_id': '1401.5372', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5372'}

# Test of standard theorem styles Ahlfors' Lemma gives the principal criterion for obtaining lower bounds on the Kobayashi metric. Then our main theorem: This is a test of a citing theorem to cite a theorem from some other source. # Test of number-swapping This is a repeat of the first section but with numbers in theorem heads swapped to the left. Ahlfors' Lemma gives the principal criterion for obtaining lower bounds on the Kobayashi metric. Then our main theorem: # An Algorithm for the Vector Domination Problem The algorithm is a divide and conquer algorithm, splitting the two sets \(A\) and \(B\) into sets \(A^+\),\(A^-\), \(B^+\) and \(B^-\) depending on if the first coordinate is larger or smaller than some value \(a\). For our purposes, we choose \(a\) as a weighted median that is computed across both sets \(A\) and \(B\). The weighted median of a collection of real numbers with associated real weights is a number such the both the total weight of all numbers smaller than the median and the total weight of all numbers larger than the median are at most half of the total weight. Note that the weighted median can be computed in linear time using minor modifications to the standard algorithms to compute the unweighted median. The algorithm can be extended for the case where \(|A| \neq |B|\). If \(|A| = O(|B|)\), above analysis immediately gives us a runtime of \((|A| |B|)^{1-s(c)/2}\). If \(|A| \gg |B|\), we partition \(A\) into \(\frac{|A|}{|B|}\) subsets of size \(|B|\) each and solve the Vector Domination problem for each subset individually. The whole algorithm then runs in time \(|B|^{2-s(c)} \frac{|A|}{|B|} = |A| |B|^{1-s(c)}\). Theorem [\[thmmain\]](#thmmain){reference-type="ref" reference="thmmain"} then follows directly from Lemmas [\[lemilp\]](#lemilp){reference-type="ref" reference="lemilp"} and [\[lemalgo\]](#lemalgo){reference-type="ref" reference="lemalgo"}. # Conclusions We give a first algorithm for \(\zeroone\) integer programs on \(n\) variables and \(cn\) constraints that improves over exhaustive search by a factor \(2^{\text{poly(1/c)}n}\). The result does generalize to ILP where the variables are constrained by any finite set of values, not just Boolean variables. Under the Strong Exponential Time Hypothesis, this is qualitatively optimal in the sense that we can only expect exponential improvement over exhaustive search if the number of constraints is linear. However, for the special case of formulas in conjunctive normal form the best algorithms achieve savings that are polylogarithmic in \(\frac1c\). It is open if we can get the same for \(\zeroone\) ILP. This result is not comparable to our earlier work. While \(\zeroone\) ILP is a special case of depth two threshold circuits, our result requires a linear number of constraints instead of a linear number of wires. It is still open if we can find an algorithm for general depth two threshold circuits that works for a linear number of gates (i.e. constraints). Lastly, there are countless problems that reduce to Integer Programming in a natural way. Some of these applications might benefit from this result. Alternatively, there might be problems that reduce directly to the Vector Domination problem and benefit from the corresponding subroutine. # ILP and the Vector Domination problem {#Domination} In this section we give a reduction from \(\zeroone\) Integer Linear Programming to the *Vector Domination* problem. In all definitions below, \(u \geq v\) for two vectors \(u\) and \(v\) of the same dimensions is used to denote an element-wise comparison. Both problems have a trivial exhaustive search algorithm. For \(\zeroone\) ILP, that algorithm runs in time \(O\left( 2^n \text{poly}(n,m)\right)\), while the trivial algorithm for the Vector Domination problem runs in time \(O(|A| |B| d)\). We can reduce \(\zeroone\) ILP to the Vector Domination problem, which shows that if the Vector Domination problem allows for an algorithm faster than exhaustive search, then so does \(\zeroone\) ILP. The reduction was introduced by Williams for the special case of Boolean vectors (he called the problem the Cooperative Subset Query problem). Note that the reduction can be trivially adapted to any ILP variant where the variables can take values from a constant size set. # Introduction A large number of real world problems from social sciences, economics, logistics and other areas are very naturally expressed as integer programs. Various variants of Integer Programming have been studied, such as bounds on the solution vector, pure or mixed integer programs, and linear, nonlinear or even nonconvex constraints, as well as a number of other restrictions on the constraints. Most forms of Integer Programming are \(\NP\)-hard, with some variants in \(\P\) (such as linear, totally unimodular constraints) and the variant with nonconvex constraints and unbounded solution vector is even undecidable. For a survey on the large body of work for solving various variants of Integer Programming exactly or approximately, we refer to the survey by Genova and Guliashki. In this paper we concentrate on the special case of \(\zeroone\) Integer Linear Programming (ILP). Given \(n\) Boolean variables and \(m\) linear constraints, the problem is to find an assignment of either \(0\) or \(1\) to the variables such that all constraints are satisfied. For this special case, we omit the objective function to be optimized and only consider the problem of deciding if a set of constraints is feasible. Since any objective value can be at most exponentially large in the input size, binary search can reduce the optimization problem to the feasibility problem with only polynomial overhead. The problem can trivially be solved in time \(O(2^n \text{poly}(n,m))\) using exhaustive search. On the other hand, an algorithm to solve the problem in time \(O\left(2^{(1-s)n}\right)\) for \(s >0\) when \(m\) is superlinear in \(n\) would contradict the Strong Exponential Time Hypothesis (\(\SETH\)), which says that for every \(s>0\) there is a \(k\) such that \(\ksat\) cannot be solved in time \(O\left(2^{(1-s)n}\right)\). \(\cnfsat\) on \(m\) clauses is a special case of \(\zeroone\) ILP on \(m\) constraints and the Sparsification Lemma allows us to reduce \(\ksat\) to \(\cnfsat\) on \(c(k)n\) many clauses. Short of proving or disproving \(\SETH\), we can then ask the following question: Given a \(\zeroone\) integer linear program on \(n\) variables and a linear number of constraints, is it possible to decide feasibility faster than exhaustive search? We answer this question affirmatively. In particular, we prove the following. For the special case of \(\cnf\) satisfiability, Schuler gives an algorithm that improves over exhaustive search by an exponential factor if the number of clauses is linear. With savings (the \(s(c)\) in above theorem) that are inverse polylogarithmic in \(c\), as opposed to polynomial, Schuler's algorithm runs considerably faster. Recently, Williams gave an algorithm for \(\zeroone\) ILP that improves over exhaustive search even for a polynomial number of constraints. The algorithm runs in time \(2^{(1-s)n}\). For \(s = \frac1{\text{polylog}(m)}\). Since \(s\) is subconstant, even for linear \(m\), Williams' result is not directly comparable to our result. Also note that a subconstant \(s\) does not contradict \(\SETH\). Our result is a follow-up to earlier work of a subset of the authors, where we considered the more general class of depth two threshold circuits. Depth two threshold circuits generalize the problem from a conjunction of linear constraints to a threshold function of linear constraints. However, the algorithm for depth two threshold circuits we gave only improves over exhaustive search by an exponential factor if the number of nonzero coefficients (across all constraints) is linear in the number of inputs. Furthermore, the savings of our algorithm for depth two threshold circuit was exponential in \(\frac1c\). The key idea of our algorithm is to reduce the problem to the *Vector Domination problem*, the problem of finding a pair of vectors such that one vector dominates the other in every coordinate. As such, the main idea stays the same as in our previous work. The main technical contribution of the current paper is an algorithm for the Vector Domination problem that improves over the trivial algorithm for vectors of dimension \(O(\log N)\) (where \(N\) is the number of vectors), whereas earlier algorithms, such as the one used for our previous result, only worked for dimensions up to \(\delta \log N\), where \(\delta\) is a sufficiently small number. Whereas algorithms for small dimensions have been discussed before, in particular Bentley and Chan, we believe that we give the first algorithm improving over exhaustive search for all dimensions possible without refuting \(\seth\).

{'timestamp': '2014-02-20T02:02:03', 'yymm': '1401', 'arxiv_id': '1401.5512', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5512'}

# Introduction Spatial multiplexing is a key feature of MIMO communication networks. The DoF region, which is the capacity region normalized by the logarithm of SNR in high SNR regimes, is a metric that captures the spatial multiplexing property. The DoF region of the MIMO BC with no CSIT was first shown in, for the two user case and later in for the general \(K\)-user BC. In this paper, we provide a new proof for the results obtained in the mentioned papers based on a simple lemma. The paper is organized as follows. Section [2](#system){reference-type="ref" reference="system"} introduces the system model and the characterization of the DoF region. Our new proof is provided in section [3](#proof){reference-type="ref" reference="proof"}. Based on this proof, the capacity region of a certain \(K\)-user MIMO BC with CDIT and an outer bound for the DoF region of the MIMO IC with no CSIT are provided in section [4](#capacity){reference-type="ref" reference="capacity"} and section [5](#IC){reference-type="ref" reference="IC"}, respectively. Section [6](#conc){reference-type="ref" reference="conc"} concludes the paper. Throughout the paper, \((.)^H\) and \(R_{\geq 0}\) denote the conjugate transpose and the set of non-negative real numbers, respectively. Also, \(f\sim o(\log P)\) is equivalent to \(\lim_{P\to \infty}\frac{f}{\log P}=0\). # System Model and Main Results {#system} We consider a MIMO BC, in which a transmitter with \(M\) antennas sends independent messages \(W_1,\ldots,W_K\) to \(K\) users (receivers), where each receiver is equipped with \(N_i\) receive antennas (\(i=1,2,\ldots,K\)). In a flat fading scenario, the discrete-time baseband received signal of user \(i\) at channel use \(t\) can be written as \[\label{e1} \textbf{{Y}}_i(t)=\textbf{{H}}_i^H(t)\textbf{{X}}(t)+\textbf{{Z}}_i(t)\ \ \ i=1,2,\ldots,K\] where \(\textbf{{Y}}_i(t)\in C^{N_i\times 1}\) is the received signal at receiver \(i\), \(\textbf{{X}}(t)\in C^{M\times 1}\) is the transmitted signal satisfying the power constraint \(E[\|\textbf{{X}}\|^2]\leq P\), \(\textbf{{H}}_i(t)\in C^{M\times N_i}\) is the channel matrix of user \(i\) and \(\textbf{{Z}}_i(t)\in C^{N_i\times 1}\) is the additive white Gaussian noise at receiver \(i\). The elements of \(\textbf{{H}}_i(t)\) and \(\textbf{{Z}}_i(t)\) are independent identically distributed circularly symmetric complex Gaussian random variables with unit variance. These elements are also assumed i.i.d. across the users. Let \(\textbf{H}_i^n=\{\textbf{H}_i(1),\textbf{H}_i(2),\ldots,\textbf{H}_i(n)\}\) be the set of channel matrices of user \(i\) up to channel use \(n\). We assume no channel state information at the transmitter and perfect channel state information at the receiver (CSIR) i.e., at channel use \(n\), user \(i\) has perfect knowledge of \(\textbf{H}_i^n\). The rate tuple \((R_1,R_2,\ldots,R_K)\) is achievable if the probability of error in decoding \(W_i\) at user \(i (i=1,\ldots,K)\) can be made arbitrarily small with sufficiently large coding length. Analysis of the capacity region \(C(P)\), which is the set of all the achievable rate tuples, is not always tractable. Instead, we consider the DoF region, which is a simpler metric independent of the transmit power, and is defined as \(\{(d_1,\ldots,d_K)|\exists (R_1,R_2,\ldots,R_K)\in C(P) \mbox{\ such that\ } d_i=\lim_{P\to \infty}\frac{R_i}{\log P} \forall i\}\). At very high SNRs, the effect of additive noise can be neglected and what remains is the interference caused by other users' signals. Therefore, the DoF region could also be interpreted as the region constructed by the number of interference-free private data streams that users receive simultaneously per channel use. **Theorem 1.** The DoF region of the \(K\)-user MIMO BC with no CSIT and perfect CSIR is given by \[\label{e2} D=\{(d_1,d_2,\ldots,d_K)\in R_{\geq 0}^K| \sum_{i=1}^K\frac{d_i}{r_i}\leq 1\}\] where \(r_i=\min\{M,N_i\}\). The region is achieved by orthogonal transmission schemes, such as time sharing across the users. # Proof of the theorem 1 {#proof} Unlike and, the proof is not based on the degradedness of the MIMO BC under no CSIT. Without loss of generality, we assume \(N_1\geq N_2\geq\ldots\geq N_K\) and we enhance the channel by giving the message of user \(i\) to users \(i+1, i+2,\ldots, K\). We also assume that each user not only knows its own channel, but also has perfect knowledge of the other users' channels. In other words, perfect global CSIR is assumed. It is obvious that this assumption does not reduce the outer bound which means that the bound with CSIR is inside the bound with global CSIR; however, the achievability is based on only CSIR not global CSIR. According to the Fano's inequality \[\label{e3} nR_i \leq I(W_i;\tilde{\textbf{Y}}_i^n|\Omega^n,W_{i+1},\ldots,W_K)+\epsilon_n\ \ \ \ i=1,2,\ldots,K\] where \(W_{K+1}=\emptyset\), \(\tilde{\textbf{Y}}_i^n=\{\textbf{Y}_i(1),\textbf{Y}_i(2),\ldots,\textbf{Y}_i(n)\}\) is the extension of the received signal at user \(i\) over \(n\) channel uses and \(\Omega^n=\{\textbf{H}_1^n,\textbf{H}_2^n,\ldots,\textbf{H}_K^n\}\) is the global channel state information up to channel use \(n\). We decompose the received observation of user \(i\) as \(\tilde{\textbf{Y}}_i^n=(\textbf{Y}_i^n,\hat{\textbf{Y}}_i^n)\) where \(\textbf{Y}_i^n\) is the set of \(r_i\) linearly independent observations and \(\hat{\textbf{Y}}_i^n\) can be reconstructed by linear combination of the elements in \(\textbf{Y}_i^n\) within noise level. From the chain rule of mutual information, \[\begin{aligned} \label{e4} nR_i &\leq I(W_i;\textbf{Y}_i^n|\Omega^n,W_{i+1},\ldots,W_K)\nonumber \\ &\ \ \ +I(W_i;\hat{\textbf{Y}}_i^n|\Omega^n,W_{i+1},\ldots,W_K,\textbf{Y}_i^n)+\epsilon_n \nonumber \\ &=I(W_i;\textbf{Y}_i^n|\Omega^n,W_{i+1},\ldots,W_K)\nonumber\\ &\ \ \ +\underbrace{h(\hat{\textbf{Y}}_i^n|\Omega^n,W_{i+1},\ldots,W_K,\textbf{Y}_i^n)}_{o(\log P)}\nonumber\\&\ \ \-\underbrace{h(\hat{\textbf{Y}}_i^n|\Omega^n,W_{i},\ldots,W_K,\textbf{Y}_i^n)}_{o(\log P)}+\epsilon_n\ \ i=1,2,\ldots,K. \end{aligned}\] For simplicity, we ignore \(\epsilon_n\) and the terms with \(o(\log P)\) and write \[\begin{aligned} \sum_{i=1}^K\frac{nR_i}{r_i} &\leq \sum_{i=1}^K\frac{I(W_i;\textbf{Y}_i^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_i} \nonumber \\ &\leq \underbrace{\frac{h(\textbf{Y}_K^n|\Omega^n)}{r_K}}_{\leq n\log P} +\sum_{i=1}^{K-1}\left[\frac{h(\textbf{Y}_i^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_i}-\right.\nonumber\\ &\left.\ \ \frac{h(\textbf{Y}_{i+1}^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_{i+1}}\right] \label{e5} \end{aligned}\] where we have used the fact that \(\frac{h(\textbf{Y}_1^n|\Omega^n,W_{1},\ldots,W_K)}{r_1}\sim o(\log P)\), since with the knowledge of \(\Omega^n,W_{1},\ldots,W_K\), the observation \(\textbf{Y}_1^n\) can be reconstructed within noise distortion. Before going further, the following lemma, which is an extension of lemma 1 in, is needed. **Lemma**. Let \(\Gamma_N=\{Y_1,Y_2,\ldots,Y_N\}\) be a set of \(N(\geq2)\) arbitrary random variables and \(\Psi_i^{j}(\Gamma_N)\) be a sliding window of size \(j\) over \(\Gamma_N\) (\(1\leq i,j \leq N\)) starting from \(Y_i\) i.e., \[\Psi_i^{j}(\Gamma_N) = Y_{(i-1)_N+1},Y_{(i)_N+1},\ldots,Y_{(i+j-2)_N+1}\] where \((.)_N\) defines the modulo \(N\) operation. Then, \[\begin{gathered} \label{e6} (N-m)h(Y_1,Y_2,\ldots,Y_N|A)\leq \sum_{i=1}^{N}h(\Psi_i^{N-m}(\Gamma_N)|A) \\ 1\leq m\leq N-1 \end{gathered}\] where \(A\) is an arbitrary condition. It is obvious that lemma 1 in is a special case of the above lemma for \(m=1.\) Each term in the summation of ([\[e5\]](#e5){reference-type="ref" reference="e5"}) can be written as \[\begin{aligned} &\frac{h(\textbf{Y}_i^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_i}-\frac{h(\textbf{Y}_{i+1}^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_{i+1}}=\nonumber\\ &\frac{r_{i+1}h(\textbf{Y}_i^n|\Omega^n,W_{i+1},\ldots,W_K)\!-\!r_ih(\textbf{Y}_{i+1}^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\label{e15.75} \end{aligned}\] \[\begin{aligned} &= \frac{r_{i+1}h(\textbf{Y}_{i,1}^n,\textbf{Y}_{i,2}^n,\ldots,\textbf{Y}_{i,r_i}^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\nonumber\\ &\ \ \-\frac{\!r_ih(\textbf{Y}_{i+1}^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\\ &\leq \frac{\sum_{p=1}^{r_i}h(\Psi_p^{r_{i+1}}(\Gamma_{r_i})|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\nonumber\\ &\ \ \-\frac{r_ih(\textbf{Y}_{i+1}^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}} \label{e16} \end{aligned}\] \[\begin{aligned} &= \sum_{p=1}^{r_i}\left[\frac{h(\Psi_p^{r_{i+1}}(\Gamma_{r_i})|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right.\nonumber\\&\ \ \left.-\frac{h(\textbf{Y}_{i+1}^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}} \right]\label{e17}\\ &= \sum_{p=1}^{r_i}\left[\frac{h(\textbf{A}_{p,i,n}\textbf{X}^n+\textbf{B}_{p,i,n}|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right.\nonumber\\&\ \ \left.-\frac{h(\textbf{C}_{i,n}\textbf{X}^n+\textbf{D}_{i,n}|\Omega^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}} \right]\label{e17.5}\\ &=0 \label{e18} \end{aligned}\] where in ([\[e16\]](#e16){reference-type="ref" reference="e16"}), since \(r_{i+1}\leq r_i\), the result of the previous lemma is applied in which \(\Gamma_{r_i}=\{\textbf{Y}_{i,1}^n,\textbf{Y}_{i,2}^n,\ldots,\textbf{Y}_{i,r_i}^n\}\) is the set of \(r_i\) linearly independent elements in \(\textbf{Y}_i^n\). In ([\[e17.5\]](#e17.5){reference-type="ref" reference="e17.5"}), we write \(\Psi_p^{r_{i+1}}(\Gamma_{r_i})\) and \(\textbf{Y}_{i+1}^n\) as large \(nr_{i+1}\) dimensional vectors as follows. \(\Psi_p^{r_{i+1}}(\Gamma_{r_i})=\textbf{A}_{p,i,n}\textbf{X}^n+\textbf{B}_{p,i,n}\) and \(\textbf{Y}_{i+1}^n=\textbf{C}_{i,n}\textbf{X}^n+\textbf{D}_{i,n}\) where \(\textbf{A}_{p,i,n}\) and \(\textbf{C}_{i,n}\) (\(\in C^{nr_{i+1}\times nM}\)) capture the channel coefficients over the \(n\) channel uses, \(\textbf{X}^n\) is the \(nM\) dimensional input vector and \(\textbf{B}_{p,i,n}\) and \(\textbf{D}_{i,n}\) capture the noise vectors over the \(n\) channel uses. Since \(\textbf{A}_{p,i,n}\) and \(\textbf{C}_{i,n}\) are identically distributed channel coefficients and \(\textbf{B}_{p,i,n}\) and \(\textbf{D}_{i,n}\) are identically distributed noise terms, the arguments of the differential entropies in ([\[e17.5\]](#e17.5){reference-type="ref" reference="e17.5"}) are statistically equivalent (i.e., have the same probability density function). Since the entropies are only a function of the distribution, we conclude that the two entropies in the difference are equal which results in ([\[e18\]](#e18){reference-type="ref" reference="e18"}). Therefore, ([\[e5\]](#e5){reference-type="ref" reference="e5"}) is simplified to \[\sum_{i=1}^K\frac{nR_i}{r_i}\leq n\log P.\] After dividing both sides by \(n\log P\) and taking the limit \(n,P \to \infty\), we get \[\sum_{i=1}^K\frac{d_i}{r_i}\leq 1.\] The above DoF region is achieved by a simple time sharing across the users where the global CSIR assumption is not necessary. **Remark 1**. The DoF region remains unchanged under the assumption of different noise distributions across the users. In this case, ([\[e18\]](#e18){reference-type="ref" reference="e18"}) does not hold anymore, since the terms in the differential entropies are no longer statistically equivalent due to different noise distributions. In this case, we further enhance the channel by giving all the noise vectors to all the users. Therefore, ([\[e4\]](#e4){reference-type="ref" reference="e4"}) is modified as \[\begin{aligned} nR_i &\leq I(W_i;\textbf{Y}_i^n,\Lambda^n|\Omega^n,W_{i+1},\ldots,W_K)\nonumber\\&=I(W_i;\textbf{Y}_i^n|\Omega^n,\Lambda^n,W_{i+1},\ldots,W_K)\nonumber\\&\ \ \ +\underbrace{I(W_i;\Lambda^n|\Omega^n,W_{i+1},\ldots,W_K)}_{=0} \end{aligned}\] where \(\Lambda^n\) denotes the set of all the noise vectors across the users (extended over \(n\) channel uses). Following the same approach, ([\[e17.5\]](#e17.5){reference-type="ref" reference="e17.5"}) is modified as \[\begin{aligned} \label{e22} &\sum_{p=1}^{r_i}\left[\frac{h(\textbf{A}_{p,i,n}\textbf{X}^n+\textbf{B}_{p,i,n}|\Omega^n,\Lambda^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right.\nonumber\\ &\ \ \left.-\frac{h(\textbf{C}_{i,n}\textbf{X}^n+\textbf{D}_{i,n}|\Omega^n,\Lambda^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}} \right]. \end{aligned}\] The matrices \(\textbf{A}_{p,i,n}\) and \(\textbf{C}_{i,n}\), which contain the channel coefficients, have the same distribution, however the vectors \(\textbf{B}_{p,i,n}\) and \(\textbf{D}_{i,n}\), which contain the noise terms, are no longer statistically equivalent. Hence, by taking the expectation over all the noise realizations, ([\[e22\]](#e22){reference-type="ref" reference="e22"}) becomes \[\begin{aligned} \label{e23} &\sum_{p=1}^{r_i}\left(E_{\Lambda^n}\left[\frac{h(\textbf{A}_{p,i,n}\textbf{X}^n+\textbf{B}_{p,i,n}|\Omega^n,\Lambda^n=\lambda^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right.\right.\nonumber\\&\left.\left.-\frac{h(\textbf{C}_{i,n}\textbf{X}^n+\textbf{D}_{i,n}|\Omega^n,\Lambda^n=\lambda^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right]\right) \end{aligned}\] where \(\lambda^n\) is a realization of \(\Lambda^n\). By applying the realization to the arguments of the differential entropies, ([\[e23\]](#e23){reference-type="ref" reference="e23"}) becomes \[\begin{aligned} \label{e24} &\sum_{p=1}^{r_i}\left(E_{\Lambda^n}\left[\frac{h(\textbf{A}_{p,i,n}\textbf{X}^n+B_{p,i,n}|\Omega^n,\Lambda^n=\lambda^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right.\right.\nonumber\\&\left.\left.-\frac{h(\textbf{C}_{i,n}\textbf{X}^n+D_{i,n}|\Omega^n,\Lambda^n=\lambda_n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right]\right) \end{aligned}\] \[\begin{aligned} &= \sum_{p=1}^{r_i}\left(E_{\Lambda^n}\left[\frac{h(\textbf{A}_{p,i,n}\textbf{X}^n|\Omega^n,\Lambda^n=\lambda^n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right.\right.\nonumber\\&\ \ \left.\left.-\frac{h(\textbf{C}_{i,n}\textbf{X}^n|\Omega^n,\Lambda^n=\lambda_n,W_{i+1},\ldots,W_K)}{r_ir_{i+1}}\right]\right)\label{e25} \\ &=0\label{e26} \end{aligned}\] where \(B_{p,i,n}\) and \(D_{i,n}\) are the realizations for \(\textbf{B}_{p,i,n}\) and \(\textbf{D}_{i,n}\), respectively. In ([\[e25\]](#e25){reference-type="ref" reference="e25"}), we have used the fact that constant addition does not change the differential entropies, and in ([\[e26\]](#e26){reference-type="ref" reference="e26"}), statistical equivalence between the arguments of the entropies is used. Therefore, the region in the theorem 1 is still an outer bound for the DoF region under the assumption of different noise distributions and since it is achievable, it is still the optimal DoF region in this case. The only difference is in the achievability i.e., since the noise can be non-Gaussian, the Gaussian distribution may no longer be optimal for the input and the optimal input distribution depends on the distribution of the noise in such a way that conditioned on the realization of the channel, the received signal becomes Gaussian. **Remark 2**. It is obvious that the assumptions of 1) Gaussian channel distribution and 2) independent channels across the users, were not used in the proof. It means that the proof can also be applied to other correlated channel distributions as long as the channel distributions are identical across the users. # Capacity region analysis {#capacity} In this section we consider the simplest assumptions in the beginning of section [2](#system){reference-type="ref" reference="system"} i.e., i.i.d. Gaussian channels and noise vectors. We also assume \(M\geq N_1\geq N_2\geq\ldots\geq N_K\) which results in \(r_i = N_i (i=1,\ldots,K)\). Since the SNR is not necessarily infinite (in contrast to the DoF analysis), all the \(o(\log P)\) terms should be replaced with their exact values. The first one is the term in ([\[e4\]](#e4){reference-type="ref" reference="e4"}) which is zero here, since \(M\geq N_1\geq N_2\geq\ldots\geq N_K\) and therefore, \(\tilde{\textbf{Y}}_i^n=\textbf{Y}_i^n\). From the Fano's inequality, \[\begin{aligned} \sum_{i=1}^K\frac{nR_i}{r_i} &\leq \sum_{i=1}^K\frac{I(W_i;\textbf{Y}_i^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_i} \nonumber \\ &\leq \frac{h(\textbf{Y}_K^n|\Omega^n)}{r_K}-\underbrace{\frac{h(\textbf{Y}_1^n|\Omega^n,W_{1},\ldots,W_K)}{r_1}}_{n\log (2\pi e)}+ \nonumber \end{aligned}\] \[\label{e29} \underbrace{\sum_{i=1}^{K-1}\!\left[\!\frac{h(\textbf{Y}_i^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_i}-\!\!\right. \left. \frac{h(\textbf{Y}_{i+1}^n|\Omega^n,W_{i+1},\ldots,W_K)}{r_{i+1}}\!\!\right]}_{\leq 0}\!\!.\] From the above results, we get an outer bound for the achievable rate region as \[\label{e27.5} \sum_{i=1}^K\frac{R_i}{r_i}\leq \frac{h(\textbf{Y}_K^n|\Omega^n)}{nr_K}.\] Therefore, an outer bound for the ergodic capacity region is \[\label{e27.75} \sum_{i=1}^K\frac{R_i}{r_i}\leq \frac{\max_{\mathbf{\Sigma}_{X}:\mbox{tr}(\mathbf{\Sigma}_{X})\leq P}E\left[\log \det(\textbf{I}_{r_K}+\textbf{H}_K^H\mathbf{\Sigma}_X\textbf{H}_K)\right]}{r_K}\] and since the channels have i.i.d. Gaussian elements, the optimal input covariance matrix is \(\frac{P}{M}\textbf{I}_M\). Hence, \[\begin{aligned} \label{e28} C^o(P)=\{&(R_1,R_2,\ldots,R_K)\in R_{\geq 0}^K| \nonumber \\&R_i\leq E\left[\log \det(\textbf{I}_{r_i}+\frac{P}{M}\textbf{H}_i^H\textbf{H}_i)\right]\ \forall i\nonumber\\&\sum_{i=1}^K\frac{R_i}{r_i}\leq \frac{E\left[\log \det(\textbf{I}_{r_K}+\frac{P}{M}\textbf{H}_K^H\textbf{H}_K)\right]}{r_K}\} \end{aligned}\] It is obvious that the outer bound is more affected by the capacity of the point-to-point link from the transmitter to the user with the lowest number of receive antennas. **Definition.** We define a class of channels (a set of matrices) \(\Theta(p,q,m)\) where each channel (matrix) in this class has its elements drawn from the distribution \(p\) in such a way that the optimal input covariance matrix for achieving the capacity of the point-to-point link from the transmitter to the virtual user defined by this channel is diagonal with equal entries. The details for this condition are given in. We also assume that for each channel in this class, all the singular values have the distribution \(q\). In other words, \[\begin{aligned} &\Theta(p,q,m)=\left\{H\in C^{m\times n}\ \forall n|\mbox{ Elements of }H\sim p,\nonumber\right.\\&\arg\max_{\mathbf{\Sigma}_{X}:\mbox{tr}(\mathbf{\Sigma}_{X})\leq P}E\left[\log \det(\textbf{I}_{n}+H^H\mathbf{\Sigma}_XH)\right]=\frac{P}{m}\textbf{I}_m, \nonumber\\&\mbox{and }\lambda_i(H^HH)\sim q, \forall i=1,\ldots,\mbox{rank}(H)\}. \end{aligned}\] **Theorem 2.** In a \(K\)-user Gaussian MIMO BC with \(M\geq N_1\geq N_2\geq\ldots\geq N_K\) and all the channels from the class of \(\Theta(p,q,M)\), the capacity region with CDIT is given by \[\begin{aligned} \label{e30} C(P)=\{&(R_1,R_2,\ldots,R_K)\in R_{\geq 0}^K|\nonumber\\&\sum_{i=1}^K\frac{R_i}{r_i}\leq E_q\left[\log (1+\frac{P}{M}\lambda)\right]\} \end{aligned}\] where \(E_q\left[\log (1+\frac{P}{M}\lambda)\right]=\int \log (1+\frac{P}{M}x)q(x)dx.\) A special case of theorem 2 was shown for the two user Gaussian MIMO BC in, in which all the eigenvalues of \(\textbf{H}_k^H\textbf{H}_k (k=1,2)\) are unity. # mimo interference channel with no csit {#IC} Consider a \(K\)-user MIMO IC with \(K\) transmitters and \(K\) receivers equipped with \(M_i\) and \(N_i\) antennas, respectively (\(i=1,2,\ldots,K\)). The input-output relationship at channel use \(t\) is given by \[\textbf{{Y}}_i(t)=\sum_{j=1}^K\textbf{{H}}_{i,j}^H(t)\textbf{{X}}_j(t)+\textbf{{Z}}_i(t)\ \ \ i=1,2,\ldots,K\] where \(\textbf{{Y}}_i(t)\) is the received signal at receiver \(i\), \(\textbf{{H}}_{i,j}\) is the channel matrix from the transmitter \(j\) to the receiver \(i\), \(\textbf{{X}}_j(t)\) is the transmitted vector by the transmitter \(j\) satisfying \(E[\|\textbf{X}_j\|^2]\leq P\) and \(\textbf{{Z}}_i(t)\) is the noise vector at the receiver \(i\). We assume that the channels are drawn from the same distribution, while the noise vectors could have different distributions. We also assume perfect CSIR (each receiver knows all the incoming channels to it from all the transmitters) and no CSIT. ## 2-user MIMO IC For the two user case, theorems 2 and 3 in are combined into theorem 5 in. Here, we provide an alternative proof for it. We assume \(N_1\leq N_2\) and \(r_i=\min(M_2,N_i)\). By giving the message of user 1 to user 2, we have \[\begin{aligned} \frac{nR_1}{r_1}+\frac{nR_2}{r_2} &\leq \frac{I(W_1;\tilde{\textbf{Y}}_1^n|\Omega^n)}{r_1}+\frac{I(W_2;\tilde{\textbf{Y}}_2^n|\Omega^n,W_{1})}{r_2}\label{e40} \\ &= \frac{h(\tilde{\textbf{Y}}_1^n|\Omega^n)}{r_1}-\overbrace{\frac{h(\tilde{\textbf{Y}}_2^n|\Omega^n,W_{1},W_2)}{r_2}}^{o(\log P)} \nonumber \\ &\ \ \ +\frac{h(\tilde{\textbf{Y}}_2^n|\Omega^n,W_{1})}{r_2}-\frac{h(\tilde{\textbf{Y}}_1^n|\Omega^n,W_{1})}{r_1}\\ &\leq \frac{n\min(N_1,M_1+M_2)}{r_1}\log P \nonumber \\ &\ \ \ +\underbrace{\frac{r_1h(\textbf{Y}_2^n|\Omega^n,W_{1})-r_2h(\textbf{Y}_1^n|\Omega^n,W_{1})}{r_1r_2}}_{\leq 0}\label{e42}\\ &\leq \frac{n\min(N_1,M_1+M_2)}{r_1}\log P \end{aligned}\] where in ([\[e40\]](#e40){reference-type="ref" reference="e40"}), \(\tilde{\textbf{Y}}_1^n\) and \(\tilde{\textbf{Y}}_2^n\) are the same as those in ([\[e3\]](#e3){reference-type="ref" reference="e3"}) and we have neglected all the terms with \(o(\log P)\) henceforth. In ([\[e42\]](#e42){reference-type="ref" reference="e42"}), \(h(\tilde{\textbf{Y}}_1^n|\Omega^n)\) is maximized when \(\tilde{\textbf{Y}}_1^n\) is Gaussian received from a transmitter with \(M_1+M_2\) antennas. Also, in the term \([\frac{h(\tilde{\textbf{Y}}_2^n|\Omega^n,W_{1})}{r_2}-\frac{h(\tilde{\textbf{Y}}_1^n|\Omega^n,W_{1})}{r_1}]\), since the entropies are conditioned on \(W_1\), \(\textbf{X}_1(1),\textbf{X}_1(2),\ldots,\textbf{X}_1(n)\) are known. Therefore, the extensions of \(\textbf{H}_{11}^H(t)\textbf{X}_1(t)\) and \(\textbf{H}_{21}^H(t)\textbf{X}_1(t)\) over \(n\) channel uses can be removed from \(\tilde{\textbf{Y}}_1^n\) and \(\tilde{\textbf{Y}}_2^n\), respectively. What remains is a broadcast channel with a transmitter having \(M_2\) transmit antennas. With a difference of \(o(\log P)\), we can replace \(\tilde{\textbf{Y}}_1^n\) and \(\tilde{\textbf{Y}}_2^n\) with their linearly independent elements \(\textbf{Y}_1^n\) and \(\textbf{Y}_2^n\), respectively as in ([\[e4\]](#e4){reference-type="ref" reference="e4"}). Since \(r_1\leq r_2\), following the same approach as in the formulae ([\[e15.75\]](#e15.75){reference-type="ref" reference="e15.75"}) to ([\[e18\]](#e18){reference-type="ref" reference="e18"}), we get the non-positive term in ([\[e42\]](#e42){reference-type="ref" reference="e42"}). Therefore, the outer bound is \[\begin{aligned} D^o=\{&(d_1,d_2)\in R_{\geq 0}^2|\ d_i\leq \min(M_i,N_i)\ i=1,2\mbox{ and }\nonumber\\ &\frac{d_1}{r_1}+\frac{d_2}{r_2}\leq\frac{\min(N_1,M_1+M_2)}{r_1}\}. \end{aligned}\] ## \(K\)-user MIMO IC It is obvious that an outer bound for the DoF region of the MIMO IC can be obtained if the transmitters cooperate to make a broadcast channel with \(M_T=\sum_{i=1}^KM_i\) antennas at the base station. Following the same proof in this paper for the broadcast channel, we get \[\begin{aligned} D^o=\{&(d_1,d_2,\ldots,d_K)\in R_{\geq 0}^K|\nonumber\\&d_i\leq \min(M_i,N_i)\ \forall i\mbox{ and } \sum_{i=1}^K\frac{d_i}{\min(M_T,N_i)}\leq 1\}. \end{aligned}\] According to theorem 9 in, the above outer bound is tight provided that either \(N_i\leq M_i\ \forall i\) or \(N_i=N \geq M=M_i \ \forall i\) where in the former time sharing across the users and in the latter receive zero-forcing and time sharing are the achievable schemes, respectively. # Conclusion {#conc} In this paper, a novel proof for the DoF region of the \(K\)-user MIMO BC with no CSIT was provided. Motivated by the proof, the capacity region of a specific class of the \(K\)-user Gaussian MIMO BC with CDIT is derived. Also, an outer bound for the DoF region of the MIMO IC with no CSIT is provided.

{'timestamp': '2014-01-27T02:06:05', 'yymm': '1401', 'arxiv_id': '1401.5676', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5676'}

# INTRODUCTION The development of lasers of high-intensity and high-frequency made possible the experimental investigations of many atomic systems and revealed new phenomena. For example, the strong electromagnetic coupling of two autoionizing states and their coherent interaction, in some cases, could lead to partial stabilization and population trapping in the ground state . The interaction of atomic autoionizing states with an external electromagnetic field has been considered in a number of papers . An experimental and theoretical investigation of the effects of the strong electromagnetic coupling of two autoionizing states on the photoionization properties of Mg demonstrated the coherent interaction between the autoionizing states . Many experiments involving autoionizing states, have been performed over the last years and the observed dynamics present very interesting features. In this study, we consider the case of helium atom in which the ground state is coupled to the 2s2p \({^1}P^o\) autoionizing state, through a laser with frequency \(\omega_1\). A second laser, with frequency \(\omega_2\), couples the 2s2p \({^1}P^o\) state to the \(2p^2~{^1}S\) autoionizing state. The latter has not been observed in any photoexcitation experiment and the results of the present study can be used for a new experimental investigation. A similar system involving double autoionization resonance, has been studied theoretically in detail, in connection with laser-induced transitions between triply excited hollow states . Here we present only the essential formulae of our approach, since the the complete theoretical treatment has been published elsewhere . The atomic parameters of the states involved are calculated by an *ab-initio* approach. In Sec. 2 we give a short description of our theoretical methodology and the basic dynamic equations describing the problem. In Sec. 3 we solve the equations and present and analyze our results. Finally, we conclude in Sec. 4 with suggestions for prospective experimental investigations. # Theoretical approach We consider Helium atom subject to two linear polarized laser fields with frequencies \(\omega_1\) and \(\omega_2\), respectively. The relative phase between them is ignored and the total laser field has the form: \[E(t) = E_1 (t)\exp (i\omega _1 t) + E_2 (t)\exp (i\omega _2 t) + c.c..\] The frequencies \(\omega_1\) and \(\omega_2\) are chosen so as to be tunable around the selected resonance transitions, namely \(1s^2~{^1}S \rightarrow 2s2p~{^1}P^o\) and \(2s2p~{^1}P^o \rightarrow 2p^2~{^1}S\). For the time-dependent field amplitudes \(E_i(t),~i=1,2\) we choose a convenient form for the pulse envelope, namely a \(\sin^2\), avoiding the long tails of a Gaussian which make the numerics more difficult, without significantly affecting the results. The explicit form is: \[E_i(t)= {\cal E}_i^{(0)} \sin^{2}(\frac{\pi t}{\tau _i} ), \quad \textrm{with}~ 0 \leq t \leq {\tau _i},\] with \({\cal E}_i^{(0)}\) the the maximum field strength and \({\tau}_i/2\) is the Full Width at Half Maximum (FWHM). We assume that \({\tau}_i\) is few picoseconds and the simultaneous action of \(E_i(t), i=1,2\), i.e. \({\tau}_1\)=\({\tau}_2\). The next step is to solve the time-dependent Schrödinger equation. In the following sections we use the notation \(\left| g \right\rangle\) for the ground state \(1s^2~{^1}S\) of He, and \(\left| a \right\rangle\), \(\left| {E_a } \right\rangle\) and \(\left| b \right\rangle\), \(\left| {E_b } \right\rangle\) for the discrete and continua parts belonging to the \(2s2p~{^1}P^o\) and \(2p^2~{^1}S\) doubly excited states, respectively. Note that these autoionizing states are single channel Feshbach resonances. The wave function for this standard model system shown in fig.1, can be expressed as: \[\begin{aligned} \left| {\Psi (t)} \right\rangle &=& C_g(t)\left| g \right\rangle + C_a(t)\left| a \right\rangle + C_b(t)\left| b \right\rangle \nonumber \\ &+& \int {dE_a C_{E_a } (t)\left| {E_a } \right\rangle } + \int {dE_b C_{E_b } (t)\left| {E_b } \right\rangle }. \end{aligned}\] The Hamiltonian operator of the system is written as: \(H=H_0 + V + V_d\), with \(H_0\left| {\mu} \right\rangle = E_{\mu}\left| {\mu} \right\rangle, ~\mu=a,b,g\), and \(V\) being the configuration interaction coupling the discrete parts of the doubly excited states to the continua and \(V_d = V_d(t)\) is the field-atom interaction. Projection of the individual states in the expansion of \(\left| {\Psi (t)} \right\rangle\) leads to a set of coupled differential equations containing amplitudes for the discrete parts as well as for the continua. The introduction of (i) the slowly varying amplitudes \(c_{i} (t)\) which are defined by \(c_{i} (t)\) = \(C_{i} (t)e^{i(E_i/\hbar+\Delta \omega)t}\) and \(\Delta \omega\) is the sum of the frequencies of the absorbed photons, (ii) the application of the rotating wave approximation(RWA), and (iii) the adiabatic elimination of the continua lead to the following set of equations for the discrete-state amplitudes: \[{i \hbar} \frac{{\partial }}{{\partial t}} {\bf c}(t) = {\bf H}(t) {\bf c}(t),\] where \[{\bf H}(t) = \left[ \begin{array}{ccc} {S_g-\frac{i}{2}\gamma _g } & \tilde \Omega _{ga} & {S_{gb}-\frac{i}{2}\gamma _{gb} } \\ \tilde \Omega_{ag} & -\delta _1-\frac{i}{2}(\Gamma _a + \gamma _a ) & \tilde \Omega _{ab} \\ {S_{bg}-\frac{i}{2}\gamma _{bg} } & \tilde \Omega _{ba} &-{\delta _1-\delta _2-\frac{i}{2}(\Gamma _b + \gamma _b )} \end{array} \right], \label{eq:H}\] and \({\bf c}(t) = \left[ c_{g}(t), c_{a}(t), c_{b}(t) \right]^T\). In equation (5) \(\delta _1 = \omega _1-(E_a^{(0)} + S_a-E{}_g-S_g)\), \(\delta _2 = \omega _2-(E_b^{(0)} + S_b-E_a^{(0)}-S_a )\) are the detunings, \(E_a^{(0)}\), \(E_b^{(0)}\) and \(\Gamma _a\), \(\Gamma _b\) the resonance energy and width of the doubly excited states, \(S_g\), \(S_a\), \(S_b\), \(S_{gb}\) and \(\gamma _g\), \(\gamma _a\), \(\gamma _b\), \(\gamma_{gb}\) the laser induced shifts and widths. The generalized complex Rabi frequencies \(\tilde \Omega _{ga}\) and \(\tilde \Omega _{ab}\) are defined as: \[\tilde \Omega _{ga} = \Omega _{ga} \left( {1-\frac{i}{{q_a }}} \right) = \frac{1}{2}E_1 (t)D^{(z)}_{ga} \left( {1-\frac{i}{{q_a }}} \right)\] and \[\tilde \Omega _{ab} = \Omega _{ab} \left( {1-\frac{i}{{q_{ab} }}} \right) = \frac{1}{2}E_2 (t)D^{(z)}_{ab}\left( {1-\frac{i}{{q_{ab} }}} \right)\] with \(D^{(z)}\) being the dipole along the polarization direction \(z\), and \(q_a\) and \(q_{ab}\) the Fano line shape parameter for the transition \(\left| g \right\rangle \rightarrow \left| a \right\rangle\) and its generalization for the transition \(\left| a \right\rangle \rightarrow \left| b \right\rangle\). Obviously the Hamiltonian of equation ([\[eq:H\]](#eq:H){reference-type="ref" reference="eq:H"}) is *Non-Hermitian*. The coefficients \(c_{i}(t)\) are slowly varying in the sense that the transformation \(c_{i} (t)\) = \(C_{i} (t)e^{i(E_i/\hbar+\Delta \omega)t}\) has removed their rapid variation. For the derivation of the above equations the laser induced continuum-continuum couplings have been neglected. It has been shown that the line shapes are not affected by continuum-continuum transitions and the same is true and for the total photoionization rate . Also, preliminary calculations showed that we can ignore the laser-induced couplings between the discrete parts \(\left| a \right\rangle\) and \(\left| b \right\rangle\) of the resonance states and the non-resonant part of the continua \(\left| E_b \right\rangle\) and \(\left| E_a \right\rangle\) respectively, and the second order effect of the laser induced coupling of the ground state \(\left| g \right\rangle\) to \(\left| b \right\rangle\) via the non-resonant part of the continua \(\left| E_a \right\rangle\). By solving the above equations the ionization yield into each channel and the total ionization probability can be calculated. The total ionization probability is: \[P{\rm{ (}}t{\rm{) }} = {\rm{ 1 }}-\left| {{\rm{c}}_{\rm{g}} (t)} \right|^2-\left| {{\rm{c}}_{\rm{a}} (t)} \right|^2-\left| {{\rm{c}}_{\rm{b}} (t)} \right|^2\]    \(q_a\)       \(q_{ab}\)        \(E_a^{(0)}\)          \(\Gamma _a\)     \(~~~~E_b^{(0)}\)         \(\Gamma _b\)         \(\Omega _{ga}\)         \(\Omega _{ab}\)         \(\gamma _g\)     ------------- ---------------- --------------------- ------------------------- --------------------- ------------------------ ---------------------------- --------------------------- --------------------------- -2.79 -714 -0.6928 1.37 \(\times 10{^{-3}}\) -0.6214 2.15\(\times 10{^{-4}}\) \(0.038 \frac{E_{1}(t)}{2}\) \(2.14 \frac{E_{2}(t)}{2}\) \(0.47 \frac{I_{1}(t)}{4}\) : Coupling of the states involved in the 3\(\times\)`<!-- -->`{=html}3 model examined here. All values are given in atomic units. \(E_a^{(0)}\), \(E_b^{(0)}\) and \(\Gamma _a\), \(\Gamma _b\) are the energies and the widths of the 2s2p\({^1}P^o\) and \(2p^2~{^1}S\) states, respectively. *Atomic structure parameters*. We have calculated all of the parameters pertaining to the atomic levels coupled by the process described above, through an *ab initio* approach. In order to take into account the electron correlation, very accurately, a large number of configurations were selected. The MCHF method  has been used to perform the present calculations. The MCHF wave-function expansion for the ground state of He was over a set of 15 configuration states coupled to form a \(^{1}S\) term. The radial wave functions for the different orbitals were obtained by the MCHF procedure, varying all the orbitals simultaneously. Minimization of the total energy yielded an energy of-2.9033 a.u. to be compared to the accurate value of-2.903724 a.u. from extensive variational calculations. The autoionizing states 2s2p \({^1}P^o\) and \(2p^2\) \({^1}S\) of He are calculated using a partition of the function space within the framework of the Feshbach formalism . Using appropriate \(\it Q\) and \(\it P\) projection operators, we can represent these resonances as quasibound states embedded in a continuum. The localized wavefunctions and of these states are also calculated by a MCHF approach. For 2s2p \({^1}P^o\) we used 27 configuration states and the energy obtained for this state is:-0.69256 a.u. For the \(2p^2\) \({^1}S\) state we used 30 two-electron configuration states and the energy obtained for this state is:-0.62218 a.u. The correlation effects are important for an accurate description of these states. This can be seen from their configuration expansion, which for the 2s2p \({^1}P^o\) is of the form: \[\psi(2s2p~ {^1}P^o) = 0.953 (2s2p)-0.291 (2p3d)-0.076 (3s3p) +...\] while for the \(2p^2\) \({^1}S\) state we have: \[\psi(2p^{2}~{^1}S) = 0.787 (2p^2)-0.554 (2s^2) + 0.173 (3s^2)-0.139 (3p^2) + 0.179 (3d^2) +...\] The autoionization widths of these states are calculated by the well known complex-coordinate method . We choose the open channel component of the resonant wavefunctions to be: \[u(1s\varepsilon \ell~ ^{1}L^{\pi}) = \hat{ \cal A} \left( \phi_ {1s}(r_1) \sum_{i} c_i \chi_i ( \rho_2 ^{*}) Y_{\ell m}(\Omega_2) \right)\] with \(\chi_i ( \rho ^{*}) = (\rho ^{*})^{k_i} e^{-a_i \rho ^{*}}\). The radial function \(\phi_ {1s}(r)\) was kept fixed to the hydrogen-like orbital of He\(^+\) whereas, in the variant of the complex-coordinate approximation followed here, for \(\chi_i ( \rho ^{*})\) the radial coordinate takes the form \(\rho_i ^{*}=r_i e^{-i\theta}\). The non-linear parameters \(a_i\) and the expansion coefficients \(c_i\) are subject to a variational optimization for the calculation of the complex energy eigenvalues pertaining to the autoionizing resonant states. Our results for the energy position, including the energy shift, and the width, of the autoionizing states presented here, can be compared to other more elaborate calculations . The wave-functions described above were also used for the calculation of the dipole moments for the various transitions involved. All the parameters that enter in the calculation are shown in table 1. We note here that the use of the complex-coordinate method provides a powerful tool for the calculation of the complex Rabi frequency \(\tilde \Omega _{ab}\) for a transition between autoionizing states. # Results and discussion The solution of the system of differential equations given by Eq. (4) can provide us the information about the temporal evolution of the system under consideration. We have chosen to study the response of the system under the simultaneous action of the electric field \(E_1\) with frequency \(\omega_1\) and a pulse duration \({\tau _1}\)=5 ps and of the electric field \(E_2\) with frequency \(\omega_2\) and of the same duration. Figures 2 and 3 show the photoionization yield of He\(^+\) as a function of the detunings of the laser sources for a series of intensities and detunings. As it can be seen from Fig. 2 the line shape changes significantly with the intensity of the laser coupling the 2s2p \({^1}P^o\) and \(2p^2\) \({^1}S\) doubly excited states. At low intensities we have a line shape for the lowest \({^1}P^o\) Feshbach resonance of He which is a typical Beutler-Fano profile . As the intensity is increased a doublet appears due to the ac Stark splitting as a result of the laser induced oscillation between 2s2p \({^1}P^o\) and \(2p^2\) \({^1}S\). This structure is known as an Autler-Townes doublet and the separation between the two peaks carries information about the dipole matrix element coupling the doubly excited states. In Fig. 3 the frequency \(\omega_1\) is on resonance (\({\delta}_1\)=0) while the laser frequency \(\omega_2\) is varied (\({\delta}_2 \ne 0\)). In this case the coupling of the doubly excited states reveals a window resonance on the photoionization cross section. The number of the electrons emitted in the energy region of each autoionizing state is proportional to the ionization signal given by the formula: \[S_i = \int\limits_{-\infty }^{ + \infty } {\left| {c_i (t)} \right|^2 \Gamma _i dt}\] with \(i=a, b\). The magnitude of \(S_i\) depends on the time evolution of the coefficients \(c_i\) and, in a more crucial way, on the values of the autoionizing widths of the resonant states. We have calculated \(S_a\) and \(S_b\) for various values of \(I_2\) ranging from \(10^8 W/cm^2\) to \(2 \times 10^{11} W/cm^2\) and for (\({\delta}_1 \ne 0\), \({\delta}_2=0\)) and our results are shown in fig. 4a and 4b. # CONCLUSIONS To our knowledge the autoionizing state \(2p^2\) \({^1}S\) of He has not been observed in any multiphoton process. There has not been any detection of this state by two-photon processes and the energy position of this resonance is measured only in scattering experiments. In this study we showed that, the coupling of this state via a laser field of moderate intensity to another autoionizing state, has significant and detectable results in the photoionization cross section of He. The prospective pump-probe two color experiments utilizing sources as, for example, the Deutsches Elektronen-Synchrotron Free Electron Laser (DESY FEL) provide an apparatus for the detection of the resonant coupling of autoionizing states as these described here . The laser intensities used in our study are attainable and proposed experiments with Free-Electron Laser (FEL) sources are readily available .

{'timestamp': '2014-01-24T02:11:54', 'yymm': '1401', 'arxiv_id': '1401.5503', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5503'}

# Introduction {#sec:intro} Dust grains play an essential role in some fundamental physical processes in the interstellar medium (ISM). First, they dominate the radiative transfer of stellar light in the ISM. In particular, the extinction curve, that is, the wavelength dependence of optical depth for dust absorption plus scattering is known to reflect the dust materials and grain size distribution (e.g., @mathis77, hereafter MRN; @draine03). Second, the dust surface is the main site for the formation of some molecular species, especially H\(_2\). The rate of dust surface reaction is proportional to the total surface area of dust grains. Since the extinction curve and the total grain surface area both depend strongly on the grain size distribution, clarifying the regulating mechanism of grain size distribution is of particular importance in understanding those important roles of dust in the ISM. MRN showed that a mixture of silicate and graphite with a grain size distribution (number of grains per grain radius) proportional to \(a^{-3.5}\), where \(a\) is the grain radius (\(a=0.001\)--\(0.25~\mu\)m), reproduces the Milky Way extinction curve. This size distribution is referred to as the MRN size distribution. and made more detailed models of the Milky Way extinction curve. In both of these models, the abundance of grains whose radii are beyond the maximum in the MRN size distribution (0.25 \(\mu\)m) is so low that the contribution of such large grains to the total dust mass is negligible. The existence of micrometer (\(\mu\)m)-sized grains is suggested in dense molecular cloud cores (called dense cores in this paper). The so-called "coreshine" refers to emission in the mid-infrared (especially the 3.6-\(\mu\)m Spitzer Infrared Array Camera (IRAC) band) from deep inside dense cores of molecular clouds. It is detected in about half of the cores studied by . The emission is interpreted as light scattered by dust grains with typical sizes of \(\sim 1~\mu\)m, which is much larger than the maximum grain radius in the diffuse interstellar medium (\(\sim 0.25~\mu\)m; MRN). We refer to \(\mu\)m-sized grains as "very large grains (VLGs)" in this paper. Formation of VLGs by coagulation in dense cores has been theoretically investigated by . Based on the timescale on which grains grow up to \(\mu\)m sizes by coagulation, they argued that dense cores are sustained over several free-fall times. Since their main aim was to constrain the lifetime of dense cores, the impact of VLG formation on the grain size distribution in the entire ISM was beyond their scope. A certain fraction of VLGs formed in dense cores may be injected into the diffuse ISM when the dense cores disperse. Efficient formation of VLGs would contradict the MRN grain size distribution in which the maximum grain radius is \(\sim 0.25~\mu\)m. Thus, based on an estimation of the total abundance of VLGs in the Milky Way, we examine the consistency between the formation of VLGs suggested by coreshine and the Galactic extinction curve (or the MRN grain size distribution). In this paper, the abundance of VLGs stands for the ratio of the total VLG mass to the total dust mass (including VLGs), and is denoted as \(\phi_\mathrm{VLG}\) (see Section [2](#sec:estimate){reference-type="ref" reference="sec:estimate"}). By definition, \(0\leq\phi_\mathrm{VLG}\leq 1\). There are some indications that VLGs exist in the ISM. One of the indications of interstellar VLGs is provided by meteorites. Large interstellar grains (\(> 1~\mu\)m) are known to exist in chondritic meteorites. Such grains were identified based on their extremely anomalous (way off from the average solar system isotope ratios) isotopic compositions. This implies that interstellar dust grains must have resided and survived in a dense core that ended up forming the solar system. Another indication of interstellar VLGs comes from direct detection of interstellar grains in the heliosphere by *Ulysses* and *Galileo*. These experiments have shown that the volume mass density of VLGs is comparable to the total dust volume mass density derived from the typical dust-to-gas ratio in the diffuse ISM in the Galaxy. This seems contradictory to the above grain size distribution derived by MRN, who found that most of the dust grains have radii less than 0.25 \(\mu\)m. Thus, it has been argued that the density of VLGs is enhanced in the solar neighborhood. Nevertheless, it is still interesting to compare the VLG abundance expected from the formation in dense cores with the measurements, in order to quantify what fraction of the observed VLGs can be explained by the formation in dense cores. We may also need to consider stellar sources of dust grains, especially asymptotic giant branch (AGB) stars and supernovae (SNe) for the production of VLGs. Indeed, the size distribution of grains produced by AGB stars is suggested to be biased toward large (\(\gtrsim 0.1~\mu\)m) sizes from the observations of spectral energy distributions , although showed that the grains are not single-sized. Theoretical studies have also shown that the dust grains formed in the winds of AGB stars have typical sizes \(\gtrsim 0.1~\mu\)m. SNe (Type II SNe) are also considered to produce relatively large (\(>0.01~\mu\)m) grains because small grains are destroyed by reverse shocks before they are ejected into the ISM. However, the timescale of dust supply from stars is longer than the shattering timescale by an order of magnitude. Therefore, even if VLGs are supplied from stars, they probably fail to survive in the ISM. In this paper, we do not treat the stellar production of VLGs because of the difficulty in their survival, but focus on their formation in dense cores, motivated by the new evidence of VLGs--coreshine. In this paper, we estimate the abundance of VLGs in the Galaxy, assuming that dense cores are the main sites for the formation of VLGs. This paper is organized as follows. In Section [2](#sec:estimate){reference-type="ref" reference="sec:estimate"}, we formulate and estimate the abundance of VLGs in the Galaxy. In Section [3](#sec:obs){reference-type="ref" reference="sec:obs"}, we compare our estimates with some observations. In Section [4](#sec:discussion){reference-type="ref" reference="sec:discussion"}, we discuss our results and implications for the dust evolution in galaxies. In Section [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}, we give our conclusions. # Estimate of the total VLG mass {#sec:estimate} ## Formation rate of VLGs in the Galaxy We estimate the supply rate of VLGs (\(\mu\)m-sized grains) in the Galaxy. Motivated by coreshine as evidence of VLGs in dense cores, we examine the hypothesis that dense cores are the main sites for the formation of VLGs in the Galaxy (see also Introduction). We assume that all dense molecular cloud cores (dense cores) eventually convert a significant fraction of dust grains into VLGs by coagulation after their lifetimes. The formation rate of VLGs (the total mass of VLGs is denoted as \(M_\mathrm{VLG}\)) in dense cores in the Galaxy, \([dM_\mathrm{VLG}/dt]_\mathrm{form}\), is estimated as \[\begin{aligned} \left[\frac{dM_\mathrm{VLG}}{dt}\right]_\mathrm{form}\equiv \frac{X_\mathrm{core}M_\mathrm{dust}(1-\phi_\mathrm{VLG})f_\mathrm{VLG}(1-\epsilon )} {\tau_\mathrm{core}}, \label{eq:dmdt} \end{aligned}\] where \(X_\mathrm{core}\) is the mass fraction of dense cores to the total gas mass, \(M_\mathrm{dust}\) is the total dust mass in the Galaxy (\(X_\mathrm{core}M_\mathrm{dust}\) is the total dust mass contained in the dense cores), \(\phi_\mathrm{VLG}\equiv M_\mathrm{VLG}/M_\mathrm{dust}\) is the ratio of the VLG mass to the total dust mass (the factor \(1-\phi_\mathrm{VLG}\) means that we need to subtract the dust that has already become VLGs), \(f_\mathrm{VLG}\) is the fraction of dust that is eventually coagulated to \(\mu\)m sizes in the dense cores, \(\epsilon\) is the star formation efficiency in the dense cores (the factor \(1-\epsilon\) means that the gas that is not included in stars is assumed to be dispersed into the ISM), and \(\tau_\mathrm{core}\) is the lifetime of dense core (i.e., the timescale of VLG formation). Note that Eq. ([\[eq:dmdt\]](#eq:dmdt){reference-type="ref" reference="eq:dmdt"}) *should not* be regarded as an ordinary differential equation, but just gives an estimate for the VLG formation rate. Since dense cores are also the sites of star formation, the star formation rate of the Galaxy is estimated by dividing the total gas mass contained in the dense cores with their lifetime (i.e., the timescale of star formation): \[\begin{aligned} \psi =\frac{\epsilon X_\mathrm{core}M_\mathrm{gas}}{\tau_\mathrm{core}}, \label{eq:sfr} \end{aligned}\] where \(M_\mathrm{gas}\) is the total gas mass in the Galaxy (\(X_\mathrm{core}M_\mathrm{gas}\) is the total gas mass in dense cores). This equation converts the core formation rate (\(X_\mathrm{core}M_\mathrm{gas}/\tau_\mathrm{core}\)) to the star formation rate, and serves to eliminate the core formation rate, which is unknown observationally compared with the star formation rate. By introducing the dust-to-gas ratio, \(\mathcal{D}\equiv M_\mathrm{dust}/M_\mathrm{gas}\) and using Eq. ([\[eq:sfr\]](#eq:sfr){reference-type="ref" reference="eq:sfr"}), we obtain \[\begin{aligned} \frac{X_\mathrm{core}}{\tau_\mathrm{core}}= \frac{\mathcal{D}\psi}{\epsilon M_\mathrm{dust}}.\label{eq:core} \end{aligned}\] Inserting Eq. ([\[eq:core\]](#eq:core){reference-type="ref" reference="eq:core"}) into Eq. ([\[eq:dmdt\]](#eq:dmdt){reference-type="ref" reference="eq:dmdt"}), we finally get the following estimate for the VLG formation rate: \[\begin{aligned} \left[\frac{dM_\mathrm{VLG}}{dt}\right]_\mathrm{form}= \frac{1-\epsilon}{\epsilon}\, \mathcal{D}(1-\phi_\mathrm{VLG})\psi f_\mathrm{VLG}. \label{eq:dmdt2} \end{aligned}\] This VLG formation rate could be implemented in a larger framework of dust enrichment, which is capable of calculating the evolution of the total dust mass in the Galaxy, to calculate the evolution of \(M_\mathrm{VLG}\) in a consistent way with \(M_\mathrm{dust}\) or \(\mathcal{D}\). However, this is not necessary for the purpose of estimating the total VLG mass in the Galaxy. The timescale of dust enrichment (i.e., the timescale of the variation of \(M_\mathrm{dust}\) or \(\mathcal{D}\)) in the Galaxy is roughly the metal-enrichment timescale (\(\sim\) several Gyr), which is much longer than the lifetime of VLGs (typically determined by the shattering timescale \(\sim 10^8\) yr; @hirashita10). Therefore, we can assume that \(M_\mathrm{dust}\) and \(\mathcal{D}\) are constant within the lifetime of VLGs. In such a case, the total mass of VLGs can be approximately estimated as follows. It is shown that \(\mu\)m-sized grains are shattered in the diffuse ISM by grain--grain collisions under the grain motion induced by turbulence. Shattering also occurs in supernova shocks. Thus, we assume that the lifetime of VLGs is determined by the shattering timescale, \(t_\mathrm{shat}\) (\(\sim 10^8\) yr; @hirashita10). The destruction rate of VLGs can thus be approximately estimated as \(M_\mathrm{VLG}/t_\mathrm{shat}\), and the equilibrium between the VLG formation and destruction is achieved on a timescale of \(\sim t_\mathrm{shat}\). Since, as mentioned above, \(t_\mathrm{shat}\) is much shorter than the evolution timescale of the dust mass, the variation of \(\mathcal{D}\) in \(t_\mathrm{shat}\) can be neglected. Therefore, the total mass of VLGs in the Galaxy is estimated by the equilibrium between the formation and destruction under a fixed \(\mathcal{D}\): \[\begin{aligned} M_\mathrm{VLG} & \sim & \left[\frac{dM_\mathrm{VLG}}{dt}\right]_\mathrm{form} \cdot t_\mathrm{shat}\nonumber\\ & = & \frac{1-\epsilon}{\epsilon}\, \mathcal{D}(1-\phi_\mathrm{VLG})\psi f_\mathrm{VLG}t_\mathrm{shat}. \label{eq:M_VLG_main} \end{aligned}\] ## Estimates of various quantities {#subsec:estimate} We adopt dust-to-gas ratio \(\mathcal{D}=0.01\), the same value as adopted in our previous calculations of coagulation in dense cores. The star formation rate in the Galaxy can be estimated from the total luminosity of OB stars. obtained a star formation rate of 1.3 M\(_\odot~\mathrm{yr}^{-1}\) based on the total OB star luminosity derived observationally by. According to, almost all the dust mass is in VLGs after coagulation in dense cores. Here we conservatively assume that half of the dust mass is converted to VLGs after coagulation in dense cores (i.e., \(f_\mathrm{VLG}=0.5\)). For \(t_\mathrm{shat}\), we adopt \(10^8\) yr according to, who considered shattering under the grain motion driven by interstellar turbulence. As calculated by, shattering can also take place in supernova shocks. Even if the supernova shocks are the main site of shattering, a similar timescale is obtained for shattering (Section [4.3](#subsec:variation){reference-type="ref" reference="subsec:variation"}). Using the above values, we obtain \[\begin{aligned} M_\mathrm{VLG} & \sim & 5\times 10^5\,(1-\phi_\mathrm{VLG})\, \frac{1-\epsilon}{\epsilon} \left(\frac{\mathcal{D}}{0.01}\right)\nonumber\\ & & \times\left(\frac{\mathcal{\psi}}{1~\mathrm{M}_\odot~\mathrm{yr}^{-1}}\right) \left(\frac{f_\mathrm{VLG}}{0.5}\right) \left(\frac{t_\mathrm{shat}}{10^8~\mathrm{yr}}\right)~\mathrm{M}_\odot. \label{eq:M_VLG} \end{aligned}\] For comparison, we also estimate the total dust mass. The total gas mass in the Milky Way is \(M_\mathrm{gas}\sim 5\times 10^9\) M\(_\odot\). By multiplying the dust-to-gas ratio, the dust mass is estimated as \[\begin{aligned} M_\mathrm{dust}=M_\mathrm{gas}\mathcal{D}\sim 5\times 10^7 \left(\frac{\mathcal{D}}{0.01}\right) \left(\frac{M_\mathrm{gas}}{5\times 10^9~\mathrm{M}_\odot}\right) ~\mathrm{M}_\odot.\label{eq:M_dust} \end{aligned}\] Dividing Eq. ([\[eq:M_VLG\]](#eq:M_VLG){reference-type="ref" reference="eq:M_VLG"}) with Eq. ([\[eq:M_dust\]](#eq:M_dust){reference-type="ref" reference="eq:M_dust"}), we obtain the following expression by recalling that \(\phi_\mathrm{VLG}=M_\mathrm{VLG}/M_\mathrm{dust}\): \[\begin{aligned} \frac{\phi_\mathrm{VLG}}{1-\phi_\mathrm{VLG}}= 0.01\,\frac{1-\epsilon}{\epsilon}\, \left(\frac{\tau_\mathrm{SF}}{5\times 10^9~\mathrm{yr}}\right)^{-1} \left(\frac{f_\mathrm{VLG}}{0.5}\right) \left(\frac{t_\mathrm{shat}}{10^8~\mathrm{yr}}\right), \label{eq:phi} \end{aligned}\] where \(\tau_\mathrm{SF}\equiv M_\mathrm{gas}/\psi\) is the star formation (gas consumption) time. The star formation efficiency \(\epsilon\) in molecular cloud cores is around 0.1--0.3. Thus, we can assume that \(\phi_\mathrm{VLG}\ll 1\) in the Galaxy. In this case, we simply replace the left-hand side of Eq. ([\[eq:phi\]](#eq:phi){reference-type="ref" reference="eq:phi"}) with \(\phi_\mathrm{VLG}\). # Comparison with observational data {#sec:obs} ## Direct detection Interstellar dust grains with radii \(a\gtrsim 0.1~\mu\)m can be detected directly in space. In Fig. [\[fig:distribution\]](#fig:distribution){reference-type="ref" reference="fig:distribution"}, we show the grain mass distribution for interstellar grains in the heliosphere observed by the *Ulysses* and *Galileo* spacecraft. Small grains with typically \(a<0.1~\mu\)m are excluded from heliospheric plasma because of large charge-to-mass ratios. Thus, we are only interested in the data at \(a> 0.1~\mu\)m. In Fig. [\[fig:distribution\]](#fig:distribution){reference-type="ref" reference="fig:distribution"}, we also plot the grain mass distribution calculated by but scaled so that the total mass density of VLGs is \(\phi_\mathrm{VLG}\mathcal{D}\mu m_\mathrm{H}n_\mathrm{H}\), where \(\mu\) is the gas mass per hydrogen atom (1.4), \(m_\mathrm{H}\) is the mass of hydrogen atom, \(n_\mathrm{H}=0.1~\mathrm{cm}^{-3}\) is the number density of hydrogen nuclei in the local ISM. adopted \(n_\mathrm{H}=0.22~\mathrm{cm}^{-3}\), but the difference by a factor of 2 does not affect the conclusions. The functional form of the distribution is taken from the maximal coagulation model with a number density of hydrogen nuclei of \(10^5\) cm\(^{-3}\) at \(t=5.5t_\mathrm{ff}\) (\(t_\mathrm{ff}\) is the free-fall time), when the peak is located at \(a=1~\mu\)m. The peak roughly reflects the mass density of VLGs, and is not sensitive to the choice of \(t\) in any case. From Eq. ([\[eq:phi\]](#eq:phi){reference-type="ref" reference="eq:phi"}), we adopt VLG-to-total-dust mass ratio \(\phi_\mathrm{VLG}=0.01(1-\epsilon )/\epsilon\) with star formation efficiency \(\epsilon =0.1\)--0.3 (Section [2.2](#subsec:estimate){reference-type="ref" reference="subsec:estimate"}). Note that the grain mass distribution is expressed as \(m^2n(m)=(4\pi\rho_\mathrm{gr}/9)n'(a)a^4\), where \(\rho_\mathrm{gr}\) is the grain material density, and \(n'(a)\) is the grain size distribution,[^1] which is related to \(n(m)\) by \(n(m)\, dm=n'(a)\,da\). We adopt the same grain material density, \(\rho_\mathrm{gr}=3.3\) g cm\(^{-3}\), as in. As we observe in Fig. [\[fig:distribution\]](#fig:distribution){reference-type="ref" reference="fig:distribution"}, the mass density of VLGs expected with \(\phi_\mathrm{VLG}\sim 0.02\)--0.09 (for \(\epsilon =0.3\)--0.1) is much lower than observed in space. This means that the large mass density observed in the heliosphere cannot be explained by the abundance of VLGs formed in dense cores. From the analysis of their velocities, the VLGs detected in space are not likely to be the remains of the large grains that were produced by coagulation in the dense core inside which the Solar System formed. If dense cores are the site of VLG formation, we need some mechanism of enhancing the VLG abundance locally to explain the spacecraft data. The physical mechanism of such an enhancement is still unknown. For comparison, we also show the MRN grain size distribution (\(n'(a)\propto a^{-3.5}\), with the upper and lower limits of grain radius, 0.001 \(\mu\)m and 0.25 \(\mu\)m, respectively). The normalization of the MRN grain size distribution is determined so that the total dust mass density is \(\mathcal{D}\mu m_\mathrm{H}n_\mathrm{H}\) with dust-to-gas ratio \(\mathcal{D}=0.01\). This MRN size distribution is representative of the grain size distribution in the diffuse ISM of the Galaxy. The abundance of VLGs expected from star formation efficiency \(\epsilon =0.1\)--0.3 is well below the extrapolation of the MRN grain size distribution to \(a\sim 1~\mu\)m. This is consistent with a small VLG-to-total-dust mass ratio, \(\phi_\mathrm{VLG}\ll 1\); that is, VLGs make only a small contribution to the total dust mass in the Galaxy. ## Extinction curves {#subsec:extinc} The wavelength dependence of interstellar extinction, the so-called extinction curve, is a viable tool to derive the grain size distribution. The MRN size distribution was derived by fitting the averaged interstellar extinction curve in the Milky Way. For the MRN size distribution, the upper limit of the grain radius is 0.25 \(\mu\)m. Here we examine if the inclusion of VLGs is consistent with the Galactic extinction curve or not. A possible caveat of extinction curve fitting is that the solution is not unique. For example, MRN adopted graphite for the carbonaceous material because of the strong 2175 Å bump, while proposed hydrogenated amorphous carbon. used the latter species for the carbonaceous dust component and reproduced both the extinction and emission of dust in the Galaxy. The relative contribution between silicate and carbonaceous dust also depends on the material properties adopted. Therefore, strictly speaking, adopting a specific model reproducing the extinction curve means that the following result only serves as an example of the effects of VLGs. However, our results below are valid at least qualitatively as long as large grains tend to show flat extinction curves in the optical and the 2175 Å bump is produced by small grains. We assume that grains are composed of two species: silicate and graphite. Extinction curves are calculated by using the same optical properties of silicates and graphite as in. That is, we adopt the optical constants from and calculate extinction cross-sections by the Mie theory. The cross-sections are weighted by the grain size distributions, and summed up for silicate and graphite. The fraction of silicate to the total dust mass is assumed to be 0.54, and the rest is assumed to be graphite . The contributions from the MRN size distribution and from the VLGs to the total extinction are proportional to \((1-\phi_\mathrm{VLG})\) and \(\phi_\mathrm{VLG}\), respectively. In Fig. [\[fig:extinc\]](#fig:extinc){reference-type="ref" reference="fig:extinc"}, we show the results. The extinction is normalized to the \(V\) band (0.55 \(\mu\)m), and the normalized extinction at wavelength \(\lambda\) is denoted as \(A_\lambda /A_V\). We find that, even for star formation efficiency \(\epsilon =0.1\) (equivalent to VLG-to-total-dust mass ratio \(\phi_\mathrm{VLG}=0.09\)), there is only a slight difference from the extinction curve for the MRN model[^2] with a little enhancement in infrared extinction, and slightly lower carbon bump and ultraviolet extinction. All these differences can be explained by the contribution from VLGs, as is clear from the extinction curve of the VLG component (dashed line in Fig. [\[fig:extinc\]](#fig:extinc){reference-type="ref" reference="fig:extinc"}). However, the difference is negligible compared with the typical variation in the Milky Way. \(R_V\equiv A_V/(A_V-A_B)\), which is an indicator of the flatness of extinction curve, is 3.6 and 3.7 for MRN and \(\epsilon =0.1\), respectively. For \(\epsilon =0.3\), the difference is even smaller. Therefore, our estimate of the VLG abundance is within the acceptable range as far as the variation of extinction curve is concerned. We also show an extreme case where \(\phi_\mathrm{VLG}=0.5\), considering that the above *Ulysses* and *Galileo* data show a mass density of VLGs comparable to the total dust mass density in the diffuse ISM (i.e., the MRN component in Fig. [\[fig:distribution\]](#fig:distribution){reference-type="ref" reference="fig:distribution"}). The deviation from the mean Galactic extinction curve is clear in this case. For this extinction curve, we obtain \(R_V=4.6\), which is still in the range of the variation in the Milky Way extinction curve, but is not typical of the diffuse ISM. This again supports the view that the high VLG abundance is due to local enhancement. also calculated the extinction curve based on 's model modified for the *Ulysses* and *Galileo* measurements. They obtained \(R_V=5.8\), the difference from our models being due to their higher abundance of VLGs. They also conclude that the large excess of \(\mu\)m-sized interstellar grains in the heliosphere is not representing the typical diffuse medium in the Milky Way but is probably due to local enhancement of VLG abundance. # Discussion {#sec:discussion} ## In the context of galaxy evolution The above estimate of \(\phi_\mathrm{VLG}\) (the VLG-to-total-dust mass ratio) is applicable to the Milky Way. Galaxies in general are expected to have a variety in \(\phi_\mathrm{VLG}\). The central hypothesis in this paper is that VLGs form in dense cores. This is probably true in solar-metallicity environments where dust-to-gas ratio is large enough for coagulation to occur efficiently in dense cores. In contrast, in a galaxy whose dust-to-gas ration is lower, coagulation in dense cores may not be efficient enough to produce VLGs; in other words, \(f_\mathrm{VLG}\) is smaller if dust-to-gas ratio (or metallicity) is lower. Since dust-to-gas ratio has a positive correlation with metallicity, we expect that \(\phi_\mathrm{VLG}\) is smaller in lower-metallicity galaxies. showed that the major source of grains with \(a>0.1~\mu\)m is stars (SNe and AGB stars) in low-metallicity (or low-dust-to-gas-ratio) environments. Therefore, if coagulation is not efficient, we expect that most of the VLGs (if they exist) are of stellar origin. As mentioned in Introduction, the size distributions of grains produced by AGB stars and SNe are suggested to be biased toward large (\(\gtrsim 0.1~\mu\)m) sizes. However, there is still an uncertainty in the size distribution of dust grains formed in stellar ejecta. For more quantitative estimates of relative importance between stellar VLGs and dense-core VLGs, we need to construct a framework that takes into account all dust formation and destruction mechanisms as done by. Since they did not include VLG formation in dense cores, it is necessary to implement it into their framework in the future. The stellar origin of VLGs is worth considering further. As mentioned in Introduction, large interstellar grains (\(>1~\mu\)m) exist in chondritic meteorites. Given their large isotope anomalies that are reflective of stellar nucleosynthesis processes, the grains may have formed in stellar ejecta. However, it is unlikely that grains formed in stellar ejecta have such a high abundance as observed by the spacecraft, since the timescale of dust supply from stellar sources is likely to be longer than the destruction timescale by shattering and sputtering . Recently, have suggested that the lifetime of silicate dust is comparable to the destruction timescale, which means that the possibility of VLG formation in stellar ejecta is still worth investigating. The existence of VLGs can have a significant influence on the extinction curve if there is local enhancement of VLG abundance as indicated by the spacecraft missions (see the case of \(\phi_\mathrm{VLG}=0.5\) in Fig. [\[fig:extinc\]](#fig:extinc){reference-type="ref" reference="fig:extinc"}). Thus, if we are concerned with the local variation of extinction curves, the contribution from VLGs should be taken into account. On global scales of a galaxy, in contrast, the abundance of VLGs is not high enough to affect the extinction curve, as long as the values of the parameters in Eq. ([\[eq:phi\]](#eq:phi){reference-type="ref" reference="eq:phi"}) are similar to the Galactic values. The far-infrared spectral energy distribution is also affected by the existence of VLGs. The equilibrium temperature (\(T_\mathrm{eq}\)) of a dust grain depends on the grain radius as \(T_\mathrm{eq}\propto a^{1/(\beta +4)}\), where \(\beta (\sim 2)\) is the power-law index of the mass absorption coefficient \(\kappa_\nu\) as a function of frequency \(\nu\), \(\kappa_\nu\propto\nu^\beta\). Because of this dependence, VLGs tend to be colder than "normal" grains with \(a\lesssim 0.1~\mu\)m. If the abundance of VLGs is high, we should consider a significant contribution from such a "cold" dust component. ## Star formation efficiency in dense cores The star formation efficiency in dense cores, \(\epsilon\), is a key factor for the VLG-to-total-dust mass ratio \(\phi_\mathrm{VLG}\). It enters our formulation in two ways. The first is through Eq. ([\[eq:dmdt\]](#eq:dmdt){reference-type="ref" reference="eq:dmdt"}), in which the factor \((1-\epsilon )\) expresses the fraction of dust that is not included in stars formed in dense cores. The second is through Eq. ([\[eq:sfr\]](#eq:sfr){reference-type="ref" reference="eq:sfr"}), which connects the core formation rate to the star formation rate. As a result, \(\phi_\mathrm{VLG}\propto (1-\epsilon )/\epsilon\) (as long as \(\phi_\mathrm{VLG}\ll 1\)). If \(\epsilon\) is nearly unity, \(\phi_\mathrm{VLG}\sim 0\), which means that all VLGs once formed in dense cores are eventually included in stars. In contrast, if \(\epsilon\lesssim 0.01\), \(\phi_\mathrm{VLG}\sim 1\). This corresponds to a case in which the total mass of dense cores is large. The volume density of VLGs measured by *Ulysses* and *Galileo* is comparable to that of the entire dust population expected from the dust-to-gas ratio in the ISM, \(\mathcal{D}=0.01\) (MRN in Fig. [\[fig:distribution\]](#fig:distribution){reference-type="ref" reference="fig:distribution"}). For example, \(\phi =0.5\) corresponds to \(\epsilon =0.01\), indicating an extremely low star formation efficiency. Considering that star formation efficiencies observationally estimated for dense cores are \(\sim 0.1\)--0.3 (Section [2.2](#subsec:estimate){reference-type="ref" reference="subsec:estimate"}), the high VLG abundance is probably due to local enhancement (see also Section [3.2](#subsec:extinc){reference-type="ref" reference="subsec:extinc"}). ## Variation of other parameters {#subsec:variation} In galaxies, \(\tau_\mathrm{SF}\) and \(t_\mathrm{shat}\) may also change. The gas consumption time \(\tau_\mathrm{SF}\) ranges from \(\sim 10^8\) yr to \(\sim 10^{11}\) yr for nearby galaxies. In a starburst galaxy with \(\tau_\mathrm{SF}\sim 10^8\) yr, if we adopt \(\epsilon =0.1\), \(f_\mathrm{VLG}=0.5\) and \(t_\mathrm{shat}=10^8\) yr, we obtain \(\phi_\mathrm{VLG}\simeq 0.8\). This large value is due to the enhanced formation rate of dense cores. Therefore, the abundance of VLGs can be enhanced in starburst galaxies. If the typical dust size in starburst galaxies is really large and VLGs are continuously supplied there, it may have important consequence for the dust destruction efficiency by sputtering. Since the timescale of dust destruction by thermal sputtering is proportional to the grain size, the VLGs survive longer than smaller grains. In particular, extreme starbursts at high redshift seems to have more dust than can be produced by stars with a dust lifetime expected for the "normal-sized" grains. The discrepancy is explained by grain growth through accretion of gas-phase metals in the ISM. However, if the VLGs dominate the total dust mass, their longer survival would make the discrepancy smaller, requiring less grain growth than is thought in previous studies. This is also true for nearby starburst galaxies. However, in starburst galaxies, shattering may also be efficient. Grains acquire larger velocities in the ionized medium than in the neutral medium because less dissipative magnetohydrodynamic nature of the ionized medium leads to more efficient gyroresonance (resonance between the magnetohyrodynamical waves with the gyromotion of a grain) acceleration of grains. As a result, \(t_\mathrm{shat}\) is shorter in the ionized medium. Therefore, if the ISM is highly ionized as a result of high star formation activities in starburst galaxies, \(t_\mathrm{shat}\) may decrease, compensating the decrease of \(\tau_\mathrm{SF}\) in Eq. ([\[eq:phi\]](#eq:phi){reference-type="ref" reference="eq:phi"}). In this paper, we have considered turbulence as the source of grain motion for shattering. Shattering also occurs in supernova shocks. showed that shattering in supernova shocks is as efficient as shattering in turbulence. They derived the timescale of shattering in supernova shocks \(\sim 0.01\tau_\mathrm{SF}\), which gives the same order of magnitude as \(t_\mathrm{shat}\) above. If we adopt this expression, the dependence of \(\phi_\mathrm{VLG}\) on \(\tau_\mathrm{SF}\) cancels out (Eq. [\[eq:phi\]](#eq:phi){reference-type="ref" reference="eq:phi"}). Therefore, if supernova shocks are the main site of shattering, the VLG abundance is insensitive to the star formation activity. In summary, it is not clear if the abundance of VLGs increases or decreases in starburst galaxies. Nevertheless, the expression in Eq. ([\[eq:phi\]](#eq:phi){reference-type="ref" reference="eq:phi"}) is useful to understand the dependence of the VLG abundance on star formation activity and shattering timescale. # Conclusion {#sec:conclusion} We have estimated the abundance of very large grains (VLGs), whose radii are typically \(\gtrsim 1~\mu\)m, in the Galaxy. Coreshine in dense molecular cloud cores (dense cores) is taken as evidence for such VLGs. Assuming that VLGs are formed in dense cores, we have estimated the abundance of VLGs in the Galaxy. The VLG abundance relative to the total dust mass is estimated as \(\phi_\mathrm{VLG}\sim 0.01(1-\epsilon )/\epsilon (\tau_\mathrm{SF}/5\times 10^9~\mathrm{yr})^{-1} (f_\mathrm{VLG}/0.5)(t_\mathrm{shat}/10^8~\mathrm{yr})\), where \(\epsilon\) is the star formation efficiency in dense cores, \(\tau_\mathrm{SF}\) is the timescale of gas consumption by star formation, \(f_\mathrm{VLG}\) is the fraction of dust mass eventually coagulated into VLGs in dense clouds, and \(t_\mathrm{shat}\) is the lifetime of VLGs (determined by shattering). Adopting typical star formation efficiencies \(\epsilon\sim 0.1\)--0.3, we obtain \(\phi_\mathrm{VLG}\sim 0.02\)--0.09 for the Galaxy. This abundance is well below the value detected by *Ulysses* and *Galileo*. Thus, if the VLGs originate from dense cores, local enhancement of VLG abundance in the solar neighborhood is necessary. We have also examined the effect of VLGs on the extinction curve, finding that the effect is negligible even with the upper value of the above range, \(\phi_\mathrm{VLG}\sim 0.09\). With \(\phi_\mathrm{VLG}\sim 0.5\), which is near the value of the above spacecraft data, the extinction curve is still in the range of the variation in Galactic extinction curves, but is not typical of the diffuse ISM. This again supports the idea that the high VLG abundance in the heliosphere is due to local enhancement. Finally, it is worth noting that the explicit dependence of \(\phi_\mathrm{VLG}\) on \(\tau_\mathrm{SF}\), \(f_\mathrm{VLG}\), and \(t_\mathrm{shat}\), can be used to estimate the VLG abundance in galaxies as well as in the Milky Way.

{'timestamp': '2014-01-23T02:05:05', 'yymm': '1401', 'arxiv_id': '1401.5570', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5570'}

# Introduction {#sec:intro} There have been many efforts aiming at understanding the formation process of star clusters, the birthplace of the majority of stars . Recent theoretical studies suggest that stellar feedback such as protostellar outflows and stellar radiation is a key to understanding the process of star formation in clustered environments (*outflow*: Li & Nakamura 2006; Matzner 2007; Nakamura & Li 2007; Cunninghan et al. 2009; Carroll et al. 2009, *radiation*: Fall et al. 2010; Peters et al. 2010; Dib 2011; Colin et al. 2013, *both*: Murray et al. 2010; Hansen et al. 2012). However, the exact role of stellar feedback in clustered star formation remains controversial. Two main scenarios have been proposed for the role of stellar feedback in clustered star formation. In the first scenario, stellar feedback is envisioned to destroy the dense cluster-forming clump as a whole, which terminates further star formation. In this case, star formation should be rapid and brief. Because supersonic turbulence decays very quickly, star formation needs to be terminated within a couple of turbulent crossing times, to achieve low star formation efficiencies (SFEs) that are often observed in nearby cluster-forming regions. The magnetic field is also considered to play only a minor or negligible role in this scenario, where the global gravitational collapse leads to rapid star formation. Hereafter, we refer to this scenario as the rapid star formation, where the primary role of stellar feedback, particularly radiation feedback from massive stars, is to terminate star formation quickly. In contrast, in the second scenario, the stellar feedback is envisioned to play the role of maintaining the internal turbulent motions of the clumps. Here, the star formation should be slow and can last for several free-fall times or longer, as the dissipated turbulence is replenished by stellar feedback . The magnetic field is also considered to play an important role in slowing down the global collapse and further star formation. The parent clump is kept close to a quasi-virial equilibrium by the combination of stellar feedback and magnetic field. Hereafter, we refer to this scenario as the slow star formation. Since the stellar feedback plays a very different role for these two scenarios, clarifying its role in clustered star formation is crucial to constrain how clustered star formation proceeds. In the present paper, we focus on the role of protostellar outflow feedback among various stellar feedback mechanisms because the outflow feedback is likely to be a leading mechanism for regulating star formation in nearby cluster-forming regions where no UV light-emitting massive stars are formed. Even for high-mass-star-forming regions, the outflow feedback is expected to play a key role in regulating star formation at least in the early and intermediate stages before massive stars form. We refer to this slow star formation scenario as the outflow-regulated cluster formation. A number of previous studies have attempted to address how the kinetic energies of the observed outflows influence the ambient gas for individual cluster-forming clumps [e.g., @hatchell07; @stanke07; @swift08; @maury09; @arce10; @curtis10; @nakamura11a; @nakamura11b; @narayanan12]. The main conclusion of these studies is that the energy injection rate by molecular outflows is generally larger than the turbulence enegy dissipation rate, and thus the outflow feedback has enough energy to sustain the turbulent motions. However, the outflow feedback is a momentum-driven feedback because radiative energy loss is efficient in the clumps. Here, we compile the outflow data of several cluster-forming clumps, and verify the role of outflow feedback in cluster-forming clumps by using the momentum dissipation and injection rates, in addition to the energy dissipation and injection rates. To facilitate comparison with observations, we will first construct an analytic version of the outflow-regulated cluster formation model and explore its observational consequences in Section [2](#sec:model){reference-type="ref" reference="sec:model"}. We compare the model with the results of molecular outflow surveys toward nearby cluster-forming regions in Section [3](#sec:obs){reference-type="ref" reference="sec:obs"}. Finally, we summarize the main conclusion in Section [4](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}. # Analytic Model of Outflow-Regulated Cluster Formation {#sec:model} Recent numerical simulations have demonstrated that protostellar outflows can indeed inject turbulent motions into cluster-forming clumps . A common drawback of this type of simulations is the use of periodic boundary conditions, which prevent the outflows from leaving the simulation box, leading to an overestimate of the efficiency of outflow feedback. However, reduced the speed of the outflows before they leave the computation box and reached the same conclusion. Thus, the periodic boundary condition does not change the main conclusion. analytically estimated the star formation rate per free-fall time of a parent clump on the basis of the outflow-regulated cluster formation scenario. They assumed (1) that the turbulence momentum dissipation rate balances the outflow momentum injection and (2) that the clump is kept close to a virial equilibrium with the internal turbulent speed equal to the virial speed. The numerical simulations of cluster formation have shown that the above two conditions are reasonably achieved. Here, assuming the above two conditions, we derive several physical quantities that can be compared directly with the observations. ## Turbulence Momentum Dissipation Rate Consider a clump with mass \(M_{\rm cl}\) and radius \(R_{\rm cl}\). The mean density and column density can be calculated, respectively, as \[\rho = {M_{\rm cl} \over 4\pi R_{\rm cl}^3/3} = 6.5 \times 10^{-20} \ {\rm g \ cm}^{-3} \left({M_{\rm cl} \over 500 \ M_\odot}\right) \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right)^{-3} \ ,\] and \[\Sigma={M_{\rm cl} \over \pi R_{\rm cl}^2} = 0.13 \ {\rm g \ cm}^{-2} \left({M_{\rm cl} \over 500 \ M_\odot}\right) \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right)^{-2} \.\] In the outflow-regulated cluster formation model, the outflow feedback replenishes the internal supersonic turbulence and the clump is kept close to a quasi-virial equilibrium . In that case, the momentum dissipation rate of the internal turbulent motion in the clump, \(dP_{\rm turb}/dt\), should balance the outflow momentum injection rate, \(dP_{\rm out}/dt\), as \[{dP_{\rm turb} \over dt} + {dP_{\rm out} \over dt} = 0 \ , \label{eq:turbulence}\] where the momentum dissipation rate of the internal turbulent motion is defined as \[\begin{aligned} {dP_{\rm turb} \over dt} &=&-0.21 {M_{\rm cl} \sigma_{\rm 3D} \over t_{\rm diss}} \, \nonumber \\ &\simeq& -6.4 \times 10^{-4} M_\odot \ {\rm km \ s}^{-1} \ {\rm yr}^{-1} \left({M_{\rm cl} \over 500 \ M_\odot}\right) \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right)^{-1} \left({\sigma_{\rm 1D} \over {\rm km \ s}^{-1}}\right)^2 \ , \label{eq:turbdiss} \end{aligned}\] where \(\sigma_{\rm 3D}\) is the three-dimensional velocity dispersion and \(\sigma_{\rm 1D}\) (= \(3^{-1/2} \sigma_{\rm 3D}\)) is one-dimensional velocity dispersion, which is smaller than the Full-Width-at-Half-Maximum (FWHM) line width by a factor of \(\sqrt{8 \ln 2}\). The momentum dissipation time \(t_{\rm diss}\) is given by \[\begin{aligned} t_{\rm diss} &=& {R_{\rm cl} \over \sigma_{\rm 3D}} \nonumber \\ &\simeq& 0.28 {\rm Myr} \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right) \left({\sigma_{\rm 1D} \over {\rm km \ s}^{-1}}\right)^{-1} \ . \end{aligned}\] Here, Equation ([\[eq:turbdiss\]](#eq:turbdiss){reference-type="ref" reference="eq:turbdiss"}) is derived from Equation (8) of , by assuming that the clump mass is constant and that the turbulence driving length scale is equal to the clump diameter of \(2R_{\rm cl}\). In the outflow-regulated cluster formation model, the internal three-dimensional velocity dispersion should be equal to the virial speed as \[\begin{aligned} \sigma_{\rm 3D} &=& \sqrt{\frac{3}{5}af_{\rm B}\frac{GM_{\rm cl}}{R_{\rm cl}}} \nonumber \, \\ &\simeq& 1.5 \ {\rm km \ s}^{-1} \left({a \over 5/3}\right)^{1/2} \left({f_{\rm B} \over 0.5}\right)^{1/2} \left({M_{\rm cl} \over 500 \ M_\odot}\right)^{1/2} \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right)^{-1/2} \label{eq:virial velocity} \end{aligned}\] where the order-of-unity dimensionless parameter, \(a\), accounts for the effects of density distribution in the clump gravitational energy. For a uniform sphere and a centrally-condensed sphere with \(\rho \propto r^{-2}\), \(a\) is equal to 1 and 5/3, respectively. Here, we adopt \(a=5/3\) because the cluster-forming clumps tend to be centrally-condensed. We take into account the magnetic support by multiplying the virial speed by a factor \(f_{\rm B}\), where \(0 \lesssim f_{\rm B} \lesssim 1\). ## Momentum Injection Rate due to Protostellar Outflow Feedback The outflow momentum injection rate is defined as \[\frac{dP_{\rm out}}{dt} = f_{\rm out} \epsilon_{\rm SFR} f_W P_* \ ,\] where \(\epsilon _{\rm SFR}\) is the star formation rate, and \(P_*\) the momentum per one solar mass of star formed, \(f_{\rm w}\) the fraction of the outflow that contributes to the generation of molecular outflows, and \(f_{\rm out}\) the fraction of the molecular outflow momentum that is converted into the clump internal turbulence momentum. The value \(f_{\rm out}\) is uncertain, but is expected to be in the range of 0.1 to 1, following the numerical simulations. According to Figure 2 of, the clump with about \(10^3 M_\odot\) has reached a quasi-virial equilibrium at the velocity dispersion of about 5 \(c_s\) in a few clump free-fall times \(t_{\rm ff}\), where \(c_s\) is the isothermal sound speed. The total amount of the specific momentum injected into the clump almost linearly increases with time and reaches about 10 \(c_s\) in \(3 t_{\rm ff}\), where the free-fall time is estimated by taking into account the fact that the clump mean density has increased by a factor of a few in a quasi-virial equilibrium state. The outflow momentum injection rate is estimated to be \(f_{\rm out}M_{\rm cl} \times 10 c_s / 3 t_{\rm ff}\), where we made the approximation that the clump mass is constant because the total mass of stars formed is only a few percent the clump mass. On the other hand, the turbulence dissipation rate is roughly estimated to be \(0.21 M_{\rm cl} \times 5 c_s / t_{\rm diss}\) and the dissipation time \(t_{\rm diss}\) is comparable to \(t_{\rm ff}\). Balancing the momentum injection rate agaist the turbulence dissipation rate yields \(f_{\rm out} \sim 0.3\). The actual value of \(f_{\rm out}\) may depend on the mass and size of the clump. For less massive, small clumps, \(f_{\rm out}\) may be smaller because the outflow lobes are easier to break out of the clump. The typical outflow speed \(V_W\) is about \(10^2\) km s\(^{-1}\) , which is assumed to be constant, independent of the stellar mass. Replacing \(\epsilon_{\rm SFR}\) by the star formation rate per free-fall time SFR\(_{\rm ff}\), Equation ([\[eq:outflow\]](#eq:outflow){reference-type="ref" reference="eq:outflow"}) is rewritten as \[\begin{aligned} \frac{dP_{\rm out}}{dt} &=& f_{\rm out} {\rm SFR}_{\rm ff} {M_{\rm cl} \over t_{\rm ff}} f_{\rm W} V_{\rm W} \nonumber \\ &\simeq& 2.3 \times 10^{-4} M_\odot {\rm km \ s}^{-1} \ {\rm yr}^{-1} \left(f_{\rm out} \over 0.3 \right) \left({{\rm SFR}_{\rm ff} \over 0.01}\right) \left({f_{\rm W} \over 0.4}\right) \left({V_{\rm W} \over 100 \ {\rm km \ s}^{-1}}\right) \left({M_{\rm cl} \over 500 \ M_\odot}\right)^{3/2} \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right)^{-3/2} \ , \label{eq:outflow} \end{aligned}\] where the free-fall time is defined as \[\begin{aligned} t_{\rm ff} &=& \sqrt{3\pi \over 32G\rho} \nonumber \\ &\simeq& 0.26 \ {\rm Myr} \left({M_{\rm cl} \over 500 \ M_\odot}\right)^{-1/2} \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right)^{3/2} \. \end{aligned}\] ## Star Formation Rate Per Free-Fall Time Using Equations ([\[eq:turbulence\]](#eq:turbulence){reference-type="ref" reference="eq:turbulence"}), ([\[eq:turbdiss\]](#eq:turbdiss){reference-type="ref" reference="eq:turbdiss"}), ([\[eq:virial velocity\]](#eq:virial velocity){reference-type="ref" reference="eq:virial velocity"}), and ([\[eq:outflow\]](#eq:outflow){reference-type="ref" reference="eq:outflow"}), the star formation rate per free-fall time expected from the outflow-regulated cluster formation model is given by \[\begin{aligned} {\rm SFR}_{\rm ff} &\simeq& 0.13 a f_{\rm B} f_{\rm out} f_{\rm w}^{-1}V_{\rm w}^{-1} \frac{GM_{\rm cl}}{R_{\rm cl}^2} t_{\rm ff} \nonumber \\ &=& 0.02 \left(\frac{f_{\rm B}}{0.5}\right) \left(f_{\rm out} \over 0.3 \right)^{-1} \left(\frac{f_{\rm w}}{0.4}\right)^{-1} \left(\frac{V_{\rm w}}{10^2 \ {\rm km \ s^{-1}}}\right)^{-1} \left(\frac{M_{\rm cl}}{500 \ M_\odot}\right)^{1/2} \left(\frac{R_{\rm cl}}{0.5 \ {\rm pc}}\right)^{-1/2} \. \label{eq:sfr_ff} \end{aligned}\] For the fiducial numbers of the physical quantities of nearby cluster-forming clumps, the star formation rate per free-fall time is estimated to be a few percent, implying that the outflow-regulated cluster formation model predicts slow star formation. If the star formation rate per free-fall time can be estimated from the observations, then the number of protostars formed in a free-fall time, \(N\), is given by \[\begin{aligned} N &\simeq& M_{\rm cl} {\rm SFR}_{\rm ff}^{\rm obs} / M_{*} \nonumber \\ &=& 10 \left({{\rm SFR}_{\rm ff}^{\rm obs} \over 0.01}\right) \left({M_{\rm cl} \over 500 \ M_\odot}\right) \left({M_* \over 0.5 \ M_\odot}\right)^{-1} \ , \label{eq:number of stars} \end{aligned}\] where \(M_*\) is the mean mass of a protostar assuming the standard stellar IMF and here we adopt \(M_{*} \simeq 0.5 M_\odot\), following. Then, the average number of protostars observed at any given time can be estimated as \[N_{\rm obs} \simeq N {t_{\rm life} \over t_{\rm ff}} \ ,\] where \(t_{\rm life}\) is the typical lifetime of protostars. Assuming that the protostars are the Class 0/I objects, \(t_{\rm life}\) is estimated to be \(t_{\rm life} \sim 0.4\) Myr. From the observed number of Class 0/I objects, we can derive the star formation rate per free-fall time from \[{\rm SFR}_{\rm ff}^{\rm obs} \simeq 0.01 \left({N_{\rm obs} \over 20}\right) \left({M_{\rm cl} \over 500 \ M_\odot}\right)^{-3/2} \left({M_* \over 0.5 \ M_\odot}\right) \left({t_{\rm life} \over 0.4 \ {\rm Myr}}\right)^{-1} \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right)^{3/2} \. \label{eq:sfr_obs}\] ## Dynamical Impact of Protostellar Outflow Feedback To assess how the outflow feedback influences the clump dynamics, we apply the virial analysis to the cluster-forming clump. The virial equation of a spherical clump is written as \[{1 \over 2}\frac{\partial ^2 I}{\partial t^2} = 2 U + W,\] where \(I\) is the moment of inertia, \(U\) is the clump kinetic energy, and \(W\) is the clump gravitational energy. The term \(U\) and \(W\) are given, respectively, by \[\begin{aligned} U &=& {3M_{\rm cl} \sigma_{\rm 1D}^2 \over 2} \\ &=& 750 M_\odot \ {\rm km^2 \ s}^{-2} \left(M_{\rm cl} \over 500 \ M_\odot \right) \left(\sigma_{\rm 1D} \over {\rm km \ s}^{-1}\right)^2 \ , \end{aligned}\] and \[\begin{aligned} W &=&-{3 \over 5}af_B{GM^2 \over R} \\ &\simeq&-1075 M_\odot {\rm km^2 \ s}^{-2} \left(a \over 5/3 \right)\left(f_B \over 0.5\right) \left(M_{\rm cl} \over 500 \ M_\odot\right)^2 \left(R_{\rm cl} \over 0.5 \ {\rm pc} \right)^{-1} \. \end{aligned}\] In the outflow-regulated cluster formation model, the clump is in quasi-virial equilibrium, i.e., \(\alpha_{\rm vir} \sim 1\), where \(\alpha_{\rm vir}\) is the virial parameter defined as \[\begin{aligned} \alpha_{\rm vir} &\equiv&-{2U \over W} = {5 \sigma_{\rm 1D}^2 R_{\rm cl} \over a f_B GM_{\rm cl}} \\ &\simeq& 1.4 \left({a \over 5/3}\right)^{-1} \left(f_B \over 0.5\right)^{-1} \left({\sigma_{\rm 1D} \over {\rm km \ s}^{-1}}\right)^2 \left({R_{\rm cl} \over 0.5 \ {\rm pc}}\right) \left({M_{\rm cl} \over 500 \ M_\odot}\right)^{-1} \. \end{aligned}\] If the total outflow kinetic energy, \(E_{\rm out}\), is significantly smaller than the clump kinetic and gravitational energies, \(U\) and \(-\)W, then the outflow energy injection contributes little to the clump dynamical state. To evaluate the impact of outflow feedback on the clump dynamics, we introduce the following non-dimensional quantity, \[\eta_{\rm out} \equiv-{2E_{\rm out} \over W} \.\] If \(\eta_{\rm out}\) is as small as \(\sim\) 0.1, then the outflow feedback is expected to play only a minor role in the clump dynamics. On the other hand, if \(\eta_{\rm out}\) is larger than \(\sim 1\), the role of outflow feedback would be much more significant: the clump material is expected to be dispersed away from the parent clump, and subsequent star formation is suppressed. Here we consider only protostellar outflow feedback as the potential clump disruption mechanism. This limits the clump mass to less than about 3000 M\(_\odot\), above which UV radiation from O stars is likely to dominate the clump destruction assuming the standard stellar IMF and star formation efficiencies. # Confronting Model with Molecular Outflow Surveys {#sec:obs} Recently, several extensive molecular outflow surveys using the \(^{12}\)CO lines have been carried out toward nearby cluster-forming regions. Here, we compare some characteristics of the outflow-regulated cluster formation model with the survey results toward 8 nearby cluster-forming clumps, (1) B59, (2) L1551, (3) L1641N, (4) Serpens Main Cloud, (5) Serpens South, (6) \(\rho\) Oph, (7) IC 348, and (8) NGC 1333. The masses of the clumps range from a few \(\times 10 M_\odot\) to \(10^3 M_\odot\). All the cluster-forming clumps have distances smaller than about 400 pc, so that the molecular outflows can be identified in reasonable spatial resolutions even with single-dish telescopes. Since all the clumps contain no massive stars that would emit strong UV radiation, the outflow feedback is expected to be the leading stellar feedback mechanism in these regions. It is worth noting that carried out detailed analysis of large-scale \(^{13}\)CO (\(J=1-0\)) mapping data toward the Perseus molecular cloud and found a number of parsec-scale expanding bubbles that are presumably driven by stellar winds from intermediate-mass protostars. These bubbles are also expected to contribute to turbulence driving in the clouds. Similar bubbles are also found in other regions like the Orion A molecular cloud We present a brief summary of the molecular outflow surveys toward these 8 clumps in Table [\[tab:obssummary\]](#tab:obssummary){reference-type="ref" reference="tab:obssummary"}, where the target name, observed \(^{12}\)CO transition, telescope, receiver, observed period, velocity resolution (\(\Delta V\)), effective angular resolution (\(\theta_{\rm eff}\)), and rms noise level (\(\Delta T_{\rm mb}\)) are listed. Physical parameters of these clumps derived from the observations are also summarized in Table [\[tab:target\]](#tab:target){reference-type="ref" reference="tab:target"}, where the distance assumed, mass, radius, mean surface density, 1D velocity dispersion, clump kinetic energy, gravitational energy, virial parameter, molecular line used, and reference are presented. The masses and radii are rescaled from the values presented in the literature by taking into account the different distances assumed. The virial parameters obtained from the observations are shown as a function of clump mass in Figure [\[fig:virial\]](#fig:virial){reference-type="ref" reference="fig:virial"}. For B59, we only consider the central round clump and do not take into account the north-east and U-shape ridges (see Figure 1 of ). We note that all clumps except Serpens Main and Serpens South were observed with the same molecular line tracer \(^{13}\)CO, whereas Serpens Main and Serpens South were observed with different molecular line tracers, C\(^{18}\)O and N\(_2\)H\(^+\), respectively. In general, the velocity dispersions derived from these high density tracers are smaller than those derived with \(^{13}\)CO. But, its effect is expected to be minor for the estimation of the virial parameters because the clumps are centrally-condensed, and thus the virial parameters are not so sensitive to the density. In fact, for a clump with \(\rho \propto r^{-2}\) and constant velocity dispersion, the virial parameter is a constant independent of radius. The other quantities such as \(dP_{\rm turb}/dt\), virial velocity, \(dP_{\rm out}/dt\) predicted by the model are also constant for all radii when the density distribution follows \(\rho \propto r^{-2}\). It is worth noting that for the Serpens South clump, CO appears to be highly depleted (Nishitani et al. 2014, in prep.) and \(^{12}\)CO and \(^{13}\)CO are significantly self-absorbed. Even the C\(^{18}\)O emission does not follow the dense clump well. So, we adopt the mass estimated from the Spectral Energy Distribution (SED) fitting of the *Herschel* data, for which the column density is summed up in the area enclosed by a contour line of \(5\times 10^{22}\) cm\(^{-2}\). For all clumps, the virial parameters are also estimated by omitting the effect of magnetic field, i.e., \(f_B = 1\) is assumed. Table [\[tab:target\]](#tab:target){reference-type="ref" reference="tab:target"} shows that the virial parameters of the clumps are close to unity for all the clumps except for the Serpens South and \(\rho\) Oph clumps. This suggests that almost all the clumps are not far from the virial equilibrium. For Serpens South and \(\rho\) Oph, the virial parameters are found to be very small, \(\alpha_{\rm vir} \sim 0.2\). For the Serpens South clump, found that the infall motions are too slow compared to the free-fall velocity, implying that the magnetic support may play a role in the clump support and therefore we speculate that the "effective" virial parameter including the effect of magnetic support is close to unity. The \(\rho\) Oph clump also has a relatively small velocity dispersion, which is much smaller than the free-fall velocity. However, the magnetic field appears not to be spatially well-ordered, suggesting that the magnetic field may not be as important for the clump support as in Serpens South. It remains unclear why \(\rho\) Oph has a small virial parameter and appears to be relatively quiescent. In fact, only one Class 0 object, VLA1623, is found in the clump. The \(\rho\) Oph clump has very high visual extinction, which may suggest the presence of foreground molecular gas, or the \(\rho\) Oph clump may be elongated along the line-of-sight, which would lead to an under-estimate of the virial parameter and an over-estimate of the angle dispersion of the near IR polarization vectors. The deficiency of Class 0 objects may be due to high extinction or, the star formation activity may be temporarily inactive. Either way, the \(\rho\) Oph clump is somewhat different from the other cluster-forming clumps in our sample. In the following, we will use the observational data to address two specific questions that lie at the heart of the outflow-regulated cluster formation model: (1) Does the outflow feedback has enough momentum to supply the dissipated turbulent motions? (2) Is the star formation in the surveyed region fast or slow? In addition, we will try to determine whether the outflow feedback has enough kinetic energy to unbind the parent clump, which is required in the competing model of rapid cluster formation. ## Outflow-Generated Turbulence In the outflow-regulated cluster formation model, the turbulence momentum dissipation rate should balance the outflow momentum injection rate. In Table [\[tab:outflow lobe\]](#tab:outflow lobe){reference-type="ref" reference="tab:outflow lobe"}, we present the turbulence momentum dissipation rate \(dP_{\rm turb}/dt\), the outflow momentum injection rate \(dP_{\rm out}/dt\), their ratio \((dP_{\rm out}/dt)/(dP_{\rm turb}/dt)\), the outflow momentum, the outflow kinetic energy \(E_{\rm out}\), and the ratio \(\eta_{\rm out}\) for all 8 cluster-forming clumps. The outflow parameters are derived from the quantities presented in the references shown in Table [\[tab:obssummary\]](#tab:obssummary){reference-type="ref" reference="tab:obssummary"} after applying corrections to account for different distances adopted and different assumptions. To estimate these quantities, we assume the following: (1) the inclination angles of the outflow axes are around \(57.3 ^\circ\), (2) the outflow material is optically-thin, and (3) the typical dynamical time of the outflows is \(3\times 10^4\) year. The second assumption leads to an under-estimation of the outflow momentum injection rates and the outflow kinetic energies by a factor of a few or more. Also, the outflow momenta and energies are underestimated at least by a factor of a few because the low-velocity components are omitted due to the difficulty in separating such outflow components from the ambient clump material. For example, for \(\rho\) Oph, the emission whose LSR velocity is in the range from 1 km s\(^{-1}\) to 6.5 km s\(^{-1}\) is not taken into account in estimating the physical parameters of the outflows, although this velocity width of 5.5 km s\(^{-1}\) is much larger than the FWHM velocity width of the clump (1.5 km s\(^{-1}\)). For the Serpens Main Cloud and Serpens South, the emissions with 6 km s\(^{-1}\) \(\le V_{\rm LSR} \le 10\) km s\(^{-1}\) and 4 km s\(^{-1}\) \(\le V_{\rm LSR} \le 11\) km s\(^{-1}\) are omitted, respectively, and the velocity widths of 4 km s\(^{-1}\) and 7 km s\(^{-1}\) are about twice and 5 times larger than the FWHM velocity widths of 2 km s\(^{-1}\) and 1.2 km s\(^{-1}\), respectively. In total, the outflow masses, momenta, and kinetic energies derived from the observations are likely to be underestimated by an order of magnitude. This underestimation of the outflow physical quantities may be compensated by the factor \(f_{\rm out}\), which is the fraction of the molecular outflow momentum that is converted into the clump turbulent momentum, and is expected to be around a few \(\times 10^{-1}\). Therefore, we assume that the momentum injection rates derived from the observations (presented in Table [\[tab:outflow lobe\]](#tab:outflow lobe){reference-type="ref" reference="tab:outflow lobe"}) are comparable to the outflow momentum injection rate. We present the ratio \((dP_{\rm out}/dt)/(dP_{\rm turb}/dt)\) as a function of clump mass in Figure [\[fig:dp/dt\]](#fig:dp/dt){reference-type="ref" reference="fig:dp/dt"}. According to Table [\[tab:outflow lobe\]](#tab:outflow lobe){reference-type="ref" reference="tab:outflow lobe"} and Figure [\[fig:dp/dt\]](#fig:dp/dt){reference-type="ref" reference="fig:dp/dt"}, for all clumps except \(\rho\) Oph, the outflow momentum injection rate is comparable to or larger than the turbulence dissipation rate. Therefore, we conclude that the outflows can maintain supersonic turbulence in the cluster-forming clumps. For the three least massive clumps, B59, L1551, and L1641N, the ratios between \(dP_{\rm out}/dt\) and \(dP_{\rm turb}/dt\) tend to be larger. This is presumably due to the fact that these clumps are nearest and the numbers of outflow lobes and protostars are smaller, and thus it is easier to distinguish between the outflow components and ambient clump gas. We note that in previous studies, the energy dissipation and injection rates are compared to assess whether the outflow feedback can maintain the turbulent motions in the cluster-forming clumps. The main conclusion is that the energy injection rate due to the outflow feedback is generally larger than the energy dissipation rate, and thus the outflow feedback has enough energy to maintain the turbulent motions. However, the outflow feedback is a momentum-driven feedback because radiative energy loss is efficient in the clouds and clumps. Thus, our approach of using the momentum dissipation and injection rates is likely to be more appropriate. ## Dynamical Impact of Outflow Feedback As shown in Table [\[tab:target\]](#tab:target){reference-type="ref" reference="tab:target"}, almost all clumps appear close to virial equilibrium. The outflow feedback should provide additional force in the clump material. If the outflows have enough energies to disperse the surrounding gas, then the outflow feedback can quench further star formation. Here, we measure the dynamical effect of the outflow feedback in the clump destruction, using the non-dimensional parameter \(\eta_{\rm out}\), the ratio between 2\(E_{\rm out}\) and \(-W\). In the last column of Table [\[tab:outflow lobe\]](#tab:outflow lobe){reference-type="ref" reference="tab:outflow lobe"} and Figure [\[fig:eta\]](#fig:eta){reference-type="ref" reference="fig:eta"}, we present the values \(\eta_{\rm out}\) derived from the observations. We note that presumably only a fraction of \(E_{\rm out}\) contributes to the dispersal of the clump material because a significant fraction of the outflow kinetic energy escapes out of the clump once the outflow breaks out. This fraction is expected to be larger for less massive, smaller clumps. However, this effect may be compensated by the fact that the outflow kinetic energies derived from the observations are likely to be underestimated by an order of magnitude. Therefore, we use the values of \(\eta_{\rm out}\) presented in Table [\[tab:outflow lobe\]](#tab:outflow lobe){reference-type="ref" reference="tab:outflow lobe"} to assess the dynamical impact of the outflow feedback. For the three least massive clumps, B59, L1551, and L1641N, the values of \(\eta_{\rm out}\) are large, indicating that the outflow feedback has potential to impact the clump structure and dynamics significantly. For L1641N, there is evidence that the stellar feedback may have dispersed the clump material significantly. For the intermediate mass clumps such as the Serpens Main Cloud and Serpens South, the outflow kinetic energy may partly influence the clump dynamical evolution. In contrast, for massive clumps, the outflow feedback appears to play a minor role in the global clump dynamics. In other words, it is likely difficult to destroy the whole clumps only by the current outflow activity. This suggests that whether the outflow feedback can destroy the cluster-forming clumps or not may depend on the clump mass. For massive clumps, the outflow feedback appears unable to disperse the clump material significantly and thus the star formation may proceed a relatively long time. ## Star Formation Rate Per Free-Fall Time Table [\[tab:sfr\]](#tab:sfr){reference-type="ref" reference="tab:sfr"} summarizes the number of protostars (Class 0 and I objects) observed in the individual clumps, and the star formation rates per free-fall time derived from the observations and Equation ([\[eq:sfr_ff\]](#eq:sfr_ff){reference-type="ref" reference="eq:sfr_ff"}). The star formation rates per free-fall time derived from the observations and predicted by the model are presented in Figure [\[fig:sfr\]](#fig:sfr){reference-type="ref" reference="fig:sfr"}. For B59, L1551, and L1641N, we adopt the results of ,, and, respectively. For the Serpens Main Cloud and \(\rho\) Oph, the numbers of protostars are calculated using the results of the *Spitzer* observations. For Serpens South, we count the protostars located within the circle indicated in Figure 1 of . We also add the Class 0 sources identified within the circle by. For IC 348 and NGC 1333, we adopt the numbers shown in. Here, we adopt the typical lifetime for Class I objects of 0.4 Myr on the basis of the results of the *Spitzer* Gould Belt Survey, although the lifetime may depend on the interstellar environments somewhat. For all clumps, the observed SFR\(_{\rm ff}^{\rm obs}\) and SFR\(_{\rm ff}\) predicted by the outflow-regulated cluster formation model stays as low as a few percent. Taking into account that the quantities have uncertainty at least by a factor of a few, we conclude that they are consistent with each other, supporting the slow cluster formation model. If SFR\(_{\rm ff} = 10 \sim\) a few \(\times 10\) % as suggested by the rapid cluster formation model, the lifetime of the Class I objects should be of order of \(10^4\) yr, which is too short. # Conclusion {#sec:conclusion} In the present paper, we constructed an analytic model of the outflow-regulated cluster formation scenario and confronted some of the model predictions with recent outflow surveys toward 8 nearby cluster-forming clumps: B59, L1551, L1641N, Serpens Main Cloud, Serpens South, IC 348, and NGC 1333. We found that the observational results support the outflow-regulated cluster formation model in general. The main conclusions are summarized below. - We constructed an analytic model of the outflow-regulated cluster formation, in which we assumed that the turbulence dissipation rate is balanced by the outflow momentum injection rate in a cluster-forming clump that is in virial equilibrium. In this model, the star formation rate per free-fall time is predicted to be a few percent. - Most of the surveyed cluster-forming clumps have virial parameters close to unity, indicating that the internal turbulent motions play an important role in the clump support, and that the clumps are close to virial equilibrium in general. The exceptions are Serpens South and \(\rho\) Oph, where the virial parameters are estimated to be as small as \(\sim\) 0.2. In Serpens South, revealed the existence of globally-ordered magnetic field that appears to be roughly perpendicular to the main filament, indicating that the magnetic support is important. In contrast, for \(\rho\) Oph, no globally-ordered magnetic field has been observed. However, the clump does not appear to be globally collapsing at the free-fall rate despite its slow internal turbulent motions. It remains unclear why \(\rho\) Oph appears relatively quiescent. It might be elongated along the line-of-sight, so that the virial parameter is underestimated. - For most of the clumps, the outflow momentum injection rate is comparable to or larger than the turbulence momentum dissipation rate. We note that the outflow momenta are underestimated in this paper because the outflow gas is assumed to be optically-thin and the low-velocity outflow components are ignored. The actual outflow momentum injection rates should be larger by a factor of a few or more. Thus, we conclude that the outflow feedback can maintain supersonic turbulence in the surveyed nearby cluster-forming regions. - However, the outflow kinetic energy is only a fraction of the clump gravitational energy except for the three least massive clumps, B59, L1551 and L1641N. Therefore, we conclude that the outflow feedback is not enough to disperse the whole clump at least for the intermediate-mass and massive clumps. - Using the numbers of Class 0/I objects, the star formation rates per free-fall time are estimated to be a few percent for all 8 clumps, which is consistent with the outflow-regulated scenario of slow cluster formation. FN is supported in part by a Grant-in-Aid for Scientific Research of Japan (A, 24244017), and ZYL by NASA NNH10AH30G and NNX14AB38G, and NSF AST-1313083.

{'timestamp': '2014-01-22T02:11:41', 'yymm': '1401', 'arxiv_id': '1401.5417', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5417'}

# Introduction {#sec:intro} The fields of quantum optics, plasmonics and metamaterials all share the broad purpose of advancing our understanding of light and exploring new ways in which it can be efficiently controlled. The specific focus of each of these fields, however, is markedly different. While quantum optics is the study of light-matter interactions at the quantum level, typically between single atoms and photons , the physical size of the supporting media is of the order of the wavelength of the light or larger, *e.g.* in waveguides and cavities. Plasmonics and metamaterials, on the other hand, consider subwavelength media where more compact guiding and confinement structures become possible . In particular, metamaterials are composed of periodic lattices of identical subwavelength unit cell scatterers, each of which governs completely the electromagnetic properties of the entire bulk material . The advancement of metamaterials has led to the development of 'left-handed' negative refractive index materials  and opened up many exciting applications, such as the super lens , transformation optics  and electromagnetic cloaking . Recently the fields of quantum optics, plasmonics and metamaterials have increasingly started to overlap with each other. One example is the effort to bring metamaterials from the microwave  to the optical regime , which requires the unit cells to be scaled down to the nanoscale to ensure they are subwavelength. Plasmonic structures such as metal nanoparticles (MNPs) are natural electric resonators at optical frequencies  and so the metamaterial community has increasingly looked towards plasmonics to design optical metamaterials. One of the big challenges, however, has been developing plasmonic magnetic resonators that can be combined with the electric resonators in order to achieve an optical negative refractive index . A possible route to address this challenge is through the use of unit cells based on quantum plasmonics . Quantum plasmonics combines the fields of quantum optics and plasmonics in order to study the quantum features of surface plasmons, as well as investigating the intense interaction of quantum emitters with localized plasmonic fields at nanostructures. The extra functionality provided by quantum emitter-plasmonic structures in the form of their interaction, tunability and associated interference effects may provide advantages over previous metamaterial designs. In this work, we explore such a scenario, combining quantum optics, plasmonics and metamaterials by investigating the integration of two-level semiconductor quantum dots (QDs) into a negative permeability metamaterial design made up of plasmonic magnetic resonators. The motivation of this work is to bring together phenomena from metamaterials and quantum plasmonics in order to address the problem of achieving a negative refractive index. This work links in with recent studies of transmission array quantum metamaterials , where theoretical investigations have looked into the possibility of engineering the bulk optical properties of mature quantum systems, such as dielectric-based cavity-atom arrays  and superconducting Josephson-qubit lines . Recent experimental work related to these studies has also investigated probing metamaterials in the quantum regime . In our work we utilize two different types of phenomenon in constructing a quantum plasmonic metamaterial. The first is the formation of an effective optical magnetic dipole from a unit cell of a metamaterial consisting of a coplanar ring of MNPs supporting localized surface plasmon (LSP) modes , as shown in Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}. The second phenomenon we use is the Fano interference , observed when light scatters from an interacting MNP-QD 'metamolecule' system . By replacing each MNP in the coplanar ring of the unit cell with an MNP-QD metamolecule, we are able to transfer the Fano interference into the magnetic resonance of the ring. When the MNP-QD nanorings are then used as the unit cell of a metamaterial, the Fano interference manifests itself in the material's bulk permeability. We exploit this interference to introduce control over the metamaterial's optical properties, providing tunable and nonlinear responses. The paper is organized as follows. In Section II, we introduce the physics of the interacting MNP-QD metamolecule and show that a Fano interference profile is present in the linear polarizability of the joint system. In Section III, we discuss the 'bare' MNP nanoring magnetic resonator (without QDs) and calculate the bulk permeability of a metamaterial composed of such inclusions. In Section IV, we incorporate the MNP-QD design from Section II into the nanoring and calculate the permeability of our proposed quantum plasmonic metamaterial. We show how the Fano interference effect can be used to tune the metamaterial's properties as well as introduce a nonlinear macroscopic response. We conclude our study in Section V by summarizing our work and outlining future directions of research, including how one might achieve a negative refractive index based on this approach. # The MNP-QD metamolecule {#metamo} When light of a particular frequency is incident upon a metal nanoparticle of subwavelength size a localized surface plasmon mode (LSP) is excited . The LSP is a non-propagating excitation of the conduction electrons in the MNP and is associated with a large enhancement of the electric field within and in the near field of the MNP. Placing quantum emitters such as atoms, quantum dots and nitrogen vacancy centers within the MNPs strong near field enhances the interaction between them . Typically, light interacts weakly with atoms and other quantum emitters due to the large size mismatch between the spatial extent of the field and the emitter. The MNP provides an interface between the incident light and the emitter, acting much like an antenna. As a result, the coupling frequency between the MNP field and the emitter can become very large. When this coupling frequency dominates all damping rates in the system, the strong-coupling regime is achieved . Here, the MNP and emitter coherently exchange energy so rapidly that the two can no longer be considered separately and must be viewed as a joint system. However, due to the large Ohmic and radiative damping associated with the MNP field mode, it is rarely the case that a MNP-emitter system is able to reach this regime, despite the large coupling frequency. Typically the coupling frequency, while lower than the MNP field's damping rate, is much larger than the emission rate of the emitter. This is a regime where the spontaneous emission of the emitter can be enhanced . Furthermore, a Fano interference can occur between the incident field and the excited field in the MNP-emitter system, leading to a characteristic Fano profile in the frequency of the scattered field . This interference is ubiquitous in wave mechanics and occurs when a discrete system interacts with a continuum . In the present case being considered, the former is the emitter and the latter is the MNP. This particular MNP-emitter type of system has been studied in depth using both a semi-classical  and a fully quantum mechanical model . The semi-classical model is perfectly suitable to examine the Fano interference in the weak-driving field limit. However, in the strong-field limit the semi-classical model breaks down and some of the nonlinear behavior predicted is invalidated by quantum noise . In order to study nonlinear effects in our metamaterial design our model must be able to operate in the strong-field limit. As such, our model is set up from the beginning within a quantum framework. In this work we focus on producing Fano interferences in a MNP-QD system and exploiting them in our metamaterial design. Thus, we are interested in the explicit optical response of the MNP-QD system. The response of this system to incident light is characterized by a frequency dependent polarizability, \(\alpha(\omega)\) . For a given incident field amplitude \(E_{0}\) it is defined as \[\alpha(\omega) = \frac{p_{MNP-QD}(\omega)}{E_0},\] where \(p_{MNP-QD}(\omega)\) is the amplitude of the metamolecule's dipole moment. The polarizability is a complex function where the real part describes the dispersion of the material and the imaginary part its absorption. To derive the MNP-QD system's polarizability, we first define its Hamiltonian, which is given by \[\hat{H} = \hat{H}_{0} + \hat{H}_{int} + \hat{H}_{drive}, \label{Hdrive0}\] where the individual terms are \[\begin{aligned} &\hat{H}_0 = \hbar\omega_{0}\hat{a}^{\dagger}\hat{a} + \hbar\omega_{x} \hat{\sigma}^{\dagger}\hat{\sigma}, \\ &\hat{H}_{int} = i\hbar g(\hat{\sigma}\hat{a}^{\dagger}-\hat{\sigma}^{\dagger}\hat{a} ),\\ &\hat{H}_{drive} =-E_{0}\mu(\hat{\sigma}e^{-i\omega t} + \hat{\sigma}^{\dagger}e^{i\omega t})\\ \nonumber & \quad \quad \quad-E_{0}(\chi^*\hat{a}e^{-i\omega t} + \chi\hat{a}^{\dagger}e^{i\omega t}). \end{aligned}\] Here, \(\omega_{0}\) and \(\omega_{x}\) are the resonance frequencies of the MNP plasmonic field mode and the QD respectively, and \(\omega\) is the external driving field frequency. The MNP resonant frequency \(\omega_0\) can be derived using the Frohlich condition  and by modeling the permittivity of silver using the Drude model, leading to the relation \(\omega_0 = \frac{\omega_p}{\sqrt{\epsilon_{\infty} + 2\epsilon_b}}\). The plasma frequency of silver is taken as \(\omega_p = 2\pi \times 4.35\) THz  and \(\epsilon_{\infty}\) is the ultra-violet permittivity of silver, which is set to \(\epsilon_{\infty} = 5\) . Finally, \(\epsilon_b\) is the permittivity of the background material in which the MNP-QD system is embedded. In Eq. ([\[Hdrive0\]](#Hdrive0){reference-type="ref" reference="Hdrive0"}), the term \(\hat{H}_0\) is the free energy Hamiltonian of the MNP and QD, where \(\hat{a}^\dag\) (\(\hat{a}\)) is the creation (annihilation) operator for the MNP plasmonic mode and \(\hat{\sigma}^\dag\) (\(\hat{\sigma}\)) is the raising (lowering) operator for the QD. The term \(\hat{H}_{int}\) describes the near-field interaction between the QD and the MNP plasmonic mode, while \(\hat{H}_{drive}\) accounts for the driving of the system by an external electric field \(E_0\). The coupling of the MNP plasmonic mode to the QD and the driving field are characterized by \(g\) and \(\chi\) respectively, and \(\mu\) is the dipole moment of the QD. The above Hamiltonians do not account for any losses the system may incur due to interactions with an external environment. The system can lose energy both radiatively to the electromagnetic vacuum, as well as due to Ohmic losses in the metal. These environmental couplings can be modeled as an interaction of the system with a bath of quantized harmonic oscillators . Treating these interactions with Born-Markov approximations enables the use of a master equation in Lindblad form, \(\dot{\hat{\rho}} = \mathcal{\hat{L}}(\hat{\rho})\), which gives a complete description of the system dynamics . Here, the Lindblad operator acts as follows \[\mathcal{\hat{L}}(\hat{\rho}) = \frac{i}{\hbar}[\hat{\rho},\hat{H}] + \hat{L}_0 + \hat{L}_x,\] where \[\hat{L}_j = \frac{\gamma_j}{2}(2\hat{c}_j\hat{\rho}\hat{c}_j^{\dagger}-[\hat{c}_j^{\dagger}\hat{c}_j,\hat{\rho}]_+),\] \(j = \{0, x\}\) and \(\hat{c}_{0}\) (\(\hat{c}_x\)) represents \(\hat{a}\) (\(\hat{\sigma}\)). For the decay rates, \(\gamma_x\) is the spontaneous emission rate of the QD and \(\gamma_0\) accounts for both Ohmic, \(\gamma_{nr}\), and radiative damping, \(\gamma_r\), of the LSP, where \(\gamma_0 = \gamma_{nr} + \gamma_r\). The spontaneous emission rate of the QD is taken as \(\gamma_x = 80\times10^{9}\) rad s\(^{-1}\) . On the other hand, the Ohmic damping of the MNP is \(\gamma_{nr} = \gamma + \frac{\gamma^3(2\epsilon_b + \epsilon_{\infty})}{\omega_{p}}\), where \(\gamma\) is the damping frequency of silver which we take as \(\gamma = 2.7 \times 10 ^{13}\) rad s\(^{-1}\) . The radiative emission is calculated from a dipole scattering formula, \(\gamma_{r} = \frac{2k^3\omega_0r^3}{\epsilon_{\infty}+2\epsilon_b}\) , where \(k\) is the wavenumber of the light and \(\epsilon_0\) is the free space permittivity. Radiative scattering dominates for larger MNPs which are more efficient antennas, while for small MNPs the Ohmic damping dominates as the mean free path of the conduction band electrons decreases . In order to find the dipole moment of the MNP-QD system and hence its polarizability we need to find the expectation values of the system operators. By working in the Heisenberg picture we can calculate the equations of motion for the expectation values, *i.e.* the Maxwell-Bloch (MB) equations, which we express here in a frame rotating with the driving field frequency \(\omega\), \[\begin{aligned} &\langle\dot{\hat{a}}\rangle =-(i\Delta_0 + \frac{\gamma_0}{2})\langle\hat{a}\rangle + g\langle\hat{\sigma}\rangle + \frac{i\chi E_0}{\hbar}, \label{MB1} \\ &\langle\dot{\hat{\sigma}}\rangle =-(i\Delta_x + \frac{\gamma_x}{2})\langle\hat{\sigma}\rangle-g\langle\hat{a}\rangle +2g\langle\hat{a}\hat{\sigma}^{\dagger}\hat{\sigma}\rangle \label{MB2} \\ & \qquad ~~ + \frac{i\mu E_0}{\hbar}(1-2\langle\hat{\sigma}^{\dagger}\hat{\sigma}\rangle), \nonumber \end{aligned}\] where \(\Delta_{0(x)} = (\omega_{0(x)}-\omega)\). In the general case, the above equations are difficult to solve as they are not in a closed form and thus form an infinite hierarchy of equations . However, we can make approximations that transform the equations into more amenable semi-classical equations. This can be done by making the assumption that the QD and the MNP field are separate systems and factoring the term \(\langle\hat{a}\hat{\sigma}^{\dagger}\hat{\sigma}\rangle\) into its light and matter components. This is a reasonable assumption when considering that the large damping of the MNP field inhibits coherent interactions . The MB equations can then be simplified further by assuming a weak driving field. In this case, the excited state population of the QD, \(\langle\hat{\sigma}^{\dagger}\hat{\sigma}\rangle\), is taken to be negligible . The MNP-QD system described above is a driven-dissipative one and therefore we are interested in calculating the polarizability when the system reaches a non-equilibrium steady state (NESS), *i.e.* when the system operators \(\dot{\hat{O}} = 0\). Using the above simplifications the NESS value of the MNP plasmonic field annihilation operator can be found to be \[\langle\hat{a}\rangle = \frac{g\langle\hat{\sigma}\rangle}{i\Delta_0+ \frac{\gamma_0}{2}} + \frac{i\chi E_0}{\hbar(i\Delta_x + \frac{\gamma_0}{2})}. \label{expmnp}\] To ensure the above quantum framework describes the physics of the system we compare its results to those predicted by classical theory. In this way we can calculate the parameters \(g\) and \(\chi\) , \[\begin{aligned} &g = \frac{S\mu}{d^3}\sqrt{\frac{3\eta r^3}{4\pi\epsilon_0\hbar}}, \label{g}, \\ &\chi =-i\epsilon_b\sqrt{12\eta\epsilon_0\pi\hbar r^3}. \label{chi2} \end{aligned}\] This derivation is expanded upon in Appendix. [\[appA\]](#appA){reference-type="ref" reference="appA"}. Here, the distance between the MNP and the QD is given by \(d\) and the MNP radius is \(r\). The dipole moment of the QD is \(\mu = er_0\), where the dipole moment radius is \(r_0 = 0.9\) nm (corresponding to 43.22 Debye) . The background permittivity is \(\epsilon_b = 2.2\), the parameter \(S = 2~(-1)\) is set for the external driving field being parallel (perpendicular) to the MNP-QD separation vector and \(\eta = \frac{(\gamma^2(2\epsilon_b + \epsilon_{\infty}) + \omega_p^2)^2}{2(2\epsilon_b+\epsilon_{\infty})^{\frac{3}{2}}\omega^3_p}\). In Appendix [\[appA\]](#appA){reference-type="ref" reference="appA"}, the dipole moment of the MNP field is shown to be \[p_{_{MNP}} = \chi^{*}\langle\hat{a}\rangle \label{mnpdipole}\] From this, an expression for the polarizability of the joint MNP-QD metamolecule system can be derived as \[\begin{aligned} \alpha(\omega) = \frac{\chi^*\langle\hat{a}\rangle + \mu\langle\hat{\sigma}\rangle}{E_0}. \end{aligned}\] Then, by solving the MB equations, Eqs. ([\[MB1\]](#MB1){reference-type="ref" reference="MB1"}) and ([\[MB2\]](#MB2){reference-type="ref" reference="MB2"}), in the steady state, an analytic expression for the system's polarizability is found to be \[\begin{aligned} \alpha(\omega) &=& \frac{i\mu^{2}}{\hbar(i\Delta_x + \frac{\gamma_x}{2} + \frac{g^2}{i\Delta_0 + \frac{\gamma_0}{2}})} + \frac{i|\chi|^{2}}{\hbar(i\Delta_0 + \frac{\gamma_0}{2} + \frac{g^2}{i\Delta_x + \frac{\gamma_x}{2}})} \nonumber \\ & &+ \frac{i\mu g(\chi^{*}-\chi)}{\hbar((i\Delta_x +\frac{\gamma_x}{2})(i\Delta_0 +\frac{\gamma_0}{2}) + g^{2})}. \end{aligned}\] In Fig. [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"} (a) and (b) we show the imaginary and real parts of the metamolecule's polarizability for a range of driving field frequencies, with the resonant frequency of the MNP and QD set to be equal. One can clearly see the Fano interference profile due to the MNP-QD interaction. In addition, looking closer at the imaginary part of the polarizability in Fig. [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"} (d), one can see that the interference suppresses light absorption at the resonance frequency. The real part of the polarizability is used to calculate the dispersion of the MNP-QD molecule. The frequency regions either side of the resonance are governed by anomalous dispersion where the polarizability decreases with increasing frequency, whereas at resonance there is a sharp increase in polarizability with increasing frequency, *i.e.* normal dispersion. This effect is also seen in EIT systems where it is responsible for slow light propagation . The Fano interference effect can be amplified by increasing the MNP-QD coupling frequency. In the above example, this coupling frequency is quite strong due to the small separation distance chosen, \(g/\omega_0 =5\times10^{-4}\). However, if the QD is placed too close to the MNP, then higher order multipoles are excited in the MNP and the dipole approximation breaks down . This should be avoided if we wish to use this scatterer in a metamaterial design using dipole formulae. The QD must also be placed further than 1 nm from the MNP surface in order to avoid electron tunneling . We have placed the QD at a distance of \(2r\) which is sufficient for higher order multipoles to be negligible  as well as to avoid tunneling effects . # The MNP nanoring {#mnpring} We now consider a ring of MNPs in a specific configuration that has recently been studied for its application as a magnetic resonator in the visible regime . Our goal is to describe this 'bare' system quantum mechanically so that we can incorporate the metamolecule from the previous section, and thereby investigate the tunability and nonlinear response provided by the QD emitter. In order to calculate the permittivity or permeability of a metamaterial, it is common practice to isolate either its electric or magnetic response with a particular type of incident field. Despite the permittivity and permeability being calculated using a special type of excitation method, these characteristic functions of the metamaterial approximate well the response of the metamaterial to an arbitrary form of incident field . Thus in our work we will concentrate on isolating the magnetic response of the ring. To achieve this, we direct a high frequency magnetic field along the ring's normal axis , as shown in Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}. The MNPs thus feel the following electric field \[\bm{E_{0}} = \frac{i\omega\mu_{0}RH_{0}}{2}\bm{\hat{\phi}},\] induced by the time varying incident magnetic field \(\bm{H} = H_{0}e^{-i\omega t}\bm{\hat{z}}\). Here, \(R\) is the radius of the ring, \(\omega\) is the frequency of the driving field, and \(\bm{\hat{\phi}}\), and \(\bm{\hat{z}}\) are unit vectors in the cylindrical coordinate system \((\bm{\hat{R}},\bm{\hat{\phi}},\bm{\hat{z}})\). The displacement field induced in each MNP is also directed along the azimuthal direction. Due to this symmetry, there is no net electrical response and we are therefore able to isolate the magnetic response of the system. A circular displacement field current is set up which acts as a magnetic dipole, whose magnitude is given by  \[m = \frac{-i\omega p_{MNP}NR}{2}, \label{magdipole}\] where \(N\) is the number of electric dipoles in the ring and \(p_{MNP}\) is the dipole moment of a single MNP. In Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} the ring configuration we consider is shown for a material with a magnetic response in the \(\bm{\hat{z}}\) direction only. Thus, the material is anisotropic. To make an isotropic material with the same response in all directions, a cubic lattice consisting of three orthogonal arrays of nanorings should be used. This can be achieved in a face-centered cubic lattice, where up to four different nanoring orientations can be included in a single unit cell . Furthermore, in our calculations we concentrate on the case of \(N\)=4 as it is the minimum number of MNPs in the ring such that the magnetic dipolar response dominates higher order multipoles . This is essential for the validity of characterizing the metamaterial's magnetic response with the permeability parameter, \(\mu_{eff}\), which we now derive . We start by calculating the dipole moment of one of the MNP inclusions using a quantum framework. Although strictly speaking this approach is not required as the process is essentially classical, we set up the quantum formalism now so that it can be used when we integrate MNP-QD metamolecules into the ring in the next section. Thus the equations of motion derived using this framework are also valid in the classical regime. The Hamiltonian of the bare system is as follows \[\begin{aligned} &\hat{H} = \hat{H}_{0} + \hat{H}_{int} + \hat{H}_{drive}, \end{aligned}\] where the individual terms are \[\begin{aligned} &\hat{H}_0 =\displaystyle\sum\limits_{n=0}^{N-1} \hbar\omega_{0}\hat{a}_n^{\dagger}\hat{a}_n, \\ &\hat{H}_{int} =\displaystyle\sum\limits_{n,m=0}^{N-1} \hbar J_{nm}(\hat{a}_n^{\dagger}\hat{a}_m +\hat{a}_m^{\dagger}\hat{a}_n)~~~n\not = m\\ &\hat{H}_{drive} =-E_{0}\displaystyle\sum\limits_{n=0}^{N-1}(\chi^*\hat{a}_ne^{-i\omega t} +\chi\hat{a}_n^\dag e^{i\omega t}). \end{aligned}\] Here, the inter-MNP coupling frequency is given by \(J_{nm}\), which for nearest neighbor coupling we denote as \(J_1\) and for next-nearest neighbor coupling as \(J_2\). The expressions for each are derived in Appendix [\[appB\]](#appB){reference-type="ref" reference="appB"} and given by \[\begin{aligned} J_{1(2)} =-12\pi\epsilon_0\epsilon_b^2r^3\eta Q_{1(2)}, \end{aligned}\] where \(Q_{1(2)}\) is the nearest (next-nearest) neighbor scalar interaction term between the MNPs which includes both near field and radiative interactions . Explicitly we have \[\begin{aligned} &Q_{1} = \frac{e^{i\sqrt{2}kR}}{16\sqrt{2}\pi\epsilon_{0}\epsilon_bR^5}(-2k^2R^4 + 3R^2(1-ik\sqrt{2}R)),\\ &Q_{2} = \frac{e^{i2kR}}{128\pi\epsilon_{0}\epsilon_bR^5}(-16k^2R^4 + 4R^2(1-i2kR)), \end{aligned}\] where \(k\) is the wave vector of the light, \(k = \omega\sqrt{\mu_0\mu_b\epsilon_0\epsilon_b}\), where \(\epsilon_0~(\mu_0)\) is the free space permittivity (permeability) and \(\epsilon_b~(\mu_b)\) is the relative permittivity (permeability) of the background medium. In the Heisenberg picture, using the full system Hamiltonian, the equation of motion for the expectation value of the annihilation operator of each MNP field mode, \(\langle\hat{a}_n\rangle\), can be found and for \(N=4\) they can be written as \[\begin{aligned} \langle\dot{\hat{a}}_n\rangle & =-(i\Delta_0 + \frac{\gamma_0}{2})\langle\hat{a}_n\rangle-iJ_1\langle\hat{a}_{n+1}\rangle-iJ_1\langle\hat{a}_{n-1}\rangle \label{aringexp} \\ & ~~~~-iJ_2\langle\hat{a}_{n+2}\rangle +\frac{i\chi E_0}{\hbar}, \nonumber \end{aligned}\] where the indices are written in modulo 4. This set of coupled equations can be solved in a straightforward manner in the steady state as the system's symmetry means that the expectation value of each dipole is the same. Using Eq. ([\[chi2\]](#chi2){reference-type="ref" reference="chi2"}) and Eq. ([\[mnpdipole\]](#mnpdipole){reference-type="ref" reference="mnpdipole"}) the dipole moment of a single MNP can be written as \[p_{_{MNP}} =\bigg(\frac{-|\chi|^2\omega\mu_{0}RH_{0}}{2\hbar(i\Delta_0 + \frac{\gamma_0}{2})}\bigg)\bigg(1 + \frac{i(2J_1 + J_2)}{i\Delta_0 + \frac{\gamma_0}{2}}\bigg)^{-1}. \label{MNPdip}\] Then, using Eq. ([\[magdipole\]](#magdipole){reference-type="ref" reference="magdipole"}) we can calculate the magnetic polarizability of a single nanoring, \(\alpha_m = \frac{m}{H_0}\). The effective permeability of the macroscopic composite system (metamaterial) can be calculated using the Maxwell-Garnett mixing formula , \[\mu_{eff} = 1 + \frac{1}{N_{d}^{-1}(\alpha_{m}^{-1} + i\frac{k^3}{6\pi})-\frac{1}{3}}, \label{mueff}\] where \(N_{d}\) is the volume concentration of nanorings in the composite system. The imaginary term in the denominator is only necessary when the rings are part of a regular three dimensional array. In this case, the radiative damping of the magnetic dipole is cancelled out . In Fig. [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"} we plot the effective permeability of a metamaterial with nanorings that have 2, 3 and 4 MNP inclusions. As mentioned earlier, only for \(N=4\) is the effective permeability physically meaningful, however, it is informative to plot \(N=2\) and 3, as they show the effect the inter-MNP coupling has on red-shifting the nanoring's magnetic resonance from the electric resonance of a single MNP (vertical dashed line). In both Fig. [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"} (a) and (b) one can see the material's resonance properties, including negative real values in panel (b). For the MNP nanoring we use the parameters of Ref. . Apart from a slightly smaller volume concentration of \(N_d\) = (96nm)\(^{-3}\), all other parameters are the same. # The MNP-QD nanoring {#mnpqdring} We now take the nanoring design from the previous section and replace each MNP with the MNP-QD metamolecule from Section [2](#metamo){reference-type="ref" reference="metamo"}, as shown in Fig. [\[fig4F\]](#fig4F){reference-type="ref" reference="fig4F"} (c). In this case there are two magnetic dipoles excited by the incident magnetic field; one set up by the ring of MNPs and the other by the QD ring. We use Eq. ([\[mueff\]](#mueff){reference-type="ref" reference="mueff"}) again to calculate the effective permeability of a metamaterial composed of these MNP-QD nanorings. However, in this case we must deal with two magnetic dipole excitations, as well as taking into account MNP-QD interactions. The Hamiltonian for the system is \[\begin{aligned} &\hat{H} = \hat{H}_{0} + \hat{H}_{int}+ \hat{H}_{drive}, \end{aligned}\] where the individual terms are \[\begin{aligned} &\hat{H}_0 =\displaystyle\sum\limits_{n=0}^{N-1} \hbar\omega_{0}\hat{a}_n^{\dagger}\hat{a}_n + \displaystyle\sum\limits_{n=0}^{N-1}\hbar\omega_{x} \hat{\sigma}_n^{\dagger}\hat{\sigma}_n, \\ &\hat{H}_{int} =\displaystyle\sum\limits_{n,m=0}^{N-1} \hbar J_{nm}(\hat{a}_n^{\dagger}\hat{a}_m + \hat{a}_m^{\dagger}\hat{a}_n)~~~n\not = m \\ &\quad \quad +\displaystyle\sum\limits_{n,m=0}^{N-1} \hbar I_{nm}(\hat{\sigma}_n^{\dagger}\hat{\sigma}_m + \hat{\sigma}_m^{\dagger}\hat{\sigma}_n)~~~n\not = m \nonumber \\ &\quad \quad + \displaystyle\sum\limits_{n,m=0}^{N-1} i\hbar g_{nm}(\hat{a}_n^{\dagger}\hat{\sigma}_{m} + \hat{a}_n\hat{\sigma}_{m}^{\dagger}),\nonumber \\ &\hat{H}_{drive} =-E_{0}\displaystyle\sum\limits_{n=0}^{N-1}(\chi^*\hat{a}_ne^{-i\omega t} + \chi\hat{a}_n^\dag e^{i\omega t})\\ \nonumber &\quad \quad \quad-E_{0}\mu\displaystyle\sum\limits_{n=0}^{N-1}(\hat{\sigma}_ne^{-i\omega t} + \hat{\sigma}_n^\dag e^{i\omega t}). \end{aligned}\] First we will calculate the effective permeability in the weak-field limit. In this case we can derive steady state MB matrix equations accounting for each site \[\begin{aligned} {\bf A}\bar{a} = {\bf B}\bar{\sigma} + \bar{c} \\ {\bf D}\bar{\sigma} =-{\bf B}\bar{a} + \bar{e} \end{aligned}\] Where \(\bar{a}\) and \(\bar{\sigma}\) are vectors which represent the expectation values for \(\hat{a}\) and \(\hat{\sigma}\) at each site in the ring, given as \[\bar{a} = \begin{pmatrix}\langle\hat{a}_1\rangle\\\langle\hat{a}_2\rangle\\\langle\hat{a}_3\rangle\\\langle\hat{a}_4\rangle\end{pmatrix}, \bar{\sigma} = \begin{pmatrix}\langle\hat{\sigma}_1\rangle\\\langle\hat{\sigma}_2\rangle\\\langle\hat{\sigma}_3\rangle\\\langle\hat{\sigma}_4\rangle\end{pmatrix}.\] The matrix \({\bf A}\) represents the MNP-MNP interactions in the nanoring, given as \[{\bf A} = \begin{pmatrix}i\Delta_0 + \frac{\gamma_0}{2}&iJ_1&iJ_2&iJ_1\\iJ_1&i\Delta_0 + \frac{\gamma_0}{2}&iJ_1&iJ_2\\iJ_2&iJ_1&i\Delta_0 + \frac{\gamma_0}{2}&iJ_1\\iJ_1&iJ_2&iJ_1&i\Delta_0 + \frac{\gamma_0}{2}\end{pmatrix},\] where the MNP-MNP coupling frequency \(J_n\) was defined in the previous section. The matrix \({\bf D}\) represents the QD-QD interactions in the nanoring, given as \[{\bf D} = \begin{pmatrix}i\Delta_x + \frac{\gamma_x}{2}&iI_1&iI_2&iI_1\\iI_1&i\Delta_x + \frac{\gamma_x}{2}&iI_1&iI_2\\iI_2&iI_1&i\Delta_x + \frac{\gamma_x}{2}&iI_1\\iI_1&iI_2&iI_1&i\Delta_x + \frac{\gamma_x}{2}\end{pmatrix},\] where \(I_1\) and \(I_2\) are the nearest neighbor and next-nearest neighbor coupling frequencies, given by \[I_{1(2)} = \frac{\mu^2}{\hbar}Q_{1(2)}.\] The matrix \({\bf B}\) is the MNP-QD coupling matrix, given as \[{\bf B} = \begin{pmatrix}g_1&0&-g_2&0\\0&g_1&0&-g_2\\-g_2&0&g_1&0\\0&-g_2&0&g_1\end{pmatrix}.\] Here, the coupling frequency \(g_1\) is for same-site MNP-QD interactions, while \(g_2\) is for an MNP coupling with its next-nearest QD neighbor. From Fig [\[fig4F\]](#fig4F){reference-type="ref" reference="fig4F"} (c) we can see that the same-site and next-nearest neighbor interactions are transverse (S =-1). Due to the azimuthal external electric field exciting the ring, the same-site QD and next-nearest neighbour QD relative to each MNP are driven in opposite directions. As such they are out of phase, this is represented by the minus signs in the matrix. We find that the next-nearest neighbour QD works to reduce the influence of the same-site QD on each MNP. Fortunately due to the stronger same-site interaction frequency the Fano interference still occurs. Similarly both nearest neighbour QDs relative to an MNP are also out of phase, however in this case their interaction frequency is the same and their effect on the MNP is cancelled out. Finally, the vectors \(\bar{c}\) and \(\bar{e}\) represent the external driving of the MNPs and the QDs by the induced electric field, \[\bar{c} = \frac{\chi\omega\mu_0R_1H_0}{2\hbar }\begin{pmatrix}1\\1\\1\\1\end{pmatrix}, \bar{e} = \frac{\mu\omega\mu_0R_2H_0}{2\hbar }\begin{pmatrix}1\\1\\1\\1\end{pmatrix},\] where we have taken into account the differing radii of the MNP (\(R_1\)) and QD (\(R_2\)) nanorings. We can solve these equations to calculate the dipole moment of each MNP and QD within the nanoring, \[\begin{aligned} P_{MNP} = \chi^*\bar{a}_1 = \chi^*({\bf A} + {\bf B} ({\bf D}^{-1}){\bf B})^{-1}({\bf B}({\bf D}^{-1})\bar{e} + \bar{c}),\\ P_{QD} = \mu\bar{\sigma}_1 = \mu({\bf D} + {\bf B} ({\bf A}^{-1}){\bf B})^{-1}({\bf B}({\bf A}^{-1})\bar{c} + \bar{e}). \end{aligned}\] Due to the symmetry of the system the dipole moment is the same on each site for the MNPs and also for the QDs. Following the procedure in Section III we calculate the magnetic dipole of both the MNP and QD rings. The total magnetic polarizability of the MNP-QD nanoring can be found using the relation \(\alpha_m = \frac{m_{_{MNP}} + m_{_{QD}}}{H_0}\) and Eq. ([\[mueff\]](#mueff){reference-type="ref" reference="mueff"}) can be used to calculate the effective permeability of a metamaterial made from the MNP-QD nanorings. The effective permeability (\(\mu_{eff}\)) of the metamaterial is shown in Fig. [\[fig4F\]](#fig4F){reference-type="ref" reference="fig4F"} (a) and (b). Due to the red-shift of the magnetic resonance of the MNP ring, as shown in Section [3](#mnpring){reference-type="ref" reference="mnpring"} (*c.f.* Fig. [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}), the QD resonances have been red-shifted in order to ensure Fano interference. The MNP ring has a radius of \(R_1 = 38\) nm, while the QD ring has a radius of \(R_2 = 6\) nm. Thus the same site MNP-QD separation is \(d = 32\) nm, as used for the MNP-QD molecule in Section I. However, we can see that the Fano interference present in the effective permeability is more prominent than that observed in the polarizability of an individual MNP-QD molecule (*c.f.* Fig. [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"}). This is due to the multiple QDs in the nanoring interacting with each MNP either directly or mediated through MNP-MNP interactions. In Fig. [\[fig4F\]](#fig4F){reference-type="ref" reference="fig4F"} (d) we show how the introduction of the QDs in the nanoring design changes the real part of the permeability from positive to negative through the Fano interference. While the magnitude of the Fano-affected Re(\(\mu_{eff}\)), when negative, is not very large, as the magnitude is dependent on the strength of the magnetic resonator, the strength of the resonance can be amplified by increasing the number of sites in the ring . From the above analysis, one can see that the integration of QDs in the MNP nanoring has transferred the Fano line-shape of the metamolecule to the effective permeability of the metamaterial. This provides an extra degree of control over the metamaterial's response. An example of this is the ability to tune the frequency at which certain phenomena occur. For example, from Fig. [\[fig4F\]](#fig4F){reference-type="ref" reference="fig4F"} (d) one can see that at the MNP-QD resonance point the Fano dip causes a negative real permeability. Thus, by dynamically shifting the detuning between the MNPs and the QDs, one can shift the frequency at which the metamaterial has negative permeability, as shown in Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}. However, the trade-off for this tunability is that as the QDs are detuned away from the magnetic resonance the bandwidth narrows. The bandwidth, \(\delta\), of the dip varies from \(0.05\) THz to \(0.01\) THz as we detune the QD away from the MNP nanoring's resonance. This still compares favorably to the bandwidth found in EIT experiments with cold rubidium atoms, where a dip bandwidth of \(\delta = 50\) MHz is observed . So far, all calculations have been confined to the weak driving field limit. However if we want to study the nonlinear properties of the nanoring metamaterial we need to consider a strong driving field. If the MNP-QD metamolecule is driven strongly by an intense light field then its optical scattering properties are modified dramatically. The two-level QD becomes saturated by the driving field and its interaction with the MNP field disappears. This nonlinear Fano effect cannot be predicted by classical or semiclassical theory, and has been studied previously in isolated MNP-QD systems and recently observed in quantum-well structures . In Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} the parameters used in our study are chosen in order to show this effect for a single MNP-QD system of the nanoring. The MNP and the QD are driven by the same external field, and one can see that as the intensity of the field increases (from the top row to the bottom row), the Fano dip in the metamolecule's polarizability is washed out by the saturation of the QDs population. These results have been computed by solving the full master equation numerically. The numerical approach involves solving the eigenproblem, \[\hat{\mathcal{L}}(\hat{\rho}_{SS}) = 0,\] where \(\hat{\rho}_{SS}\) is the NESS density matrix of the system. The difficulty here lies in the unbounded dimensions of the bosonic MNP field mode, whose Hilbert space is infinite. In order to capture the non-classical behavior of the system the truncation of the dimension of each of the MNP field's Hilbert space has a lower bound of \(d = 15\). This problem is well suited to the quantum optics toolbox developed by Tan . Our nanoring contains a minimum of four MNPs which makes its combined Hilbert space very large. Thus while the formalism we have developed and studied in this work is now in place, it is unfortunately too computationally intensive at present to study the saturation of the nanoring. However, logically if the QD is saturated in the MNP-QD metamolecule it will also be saturated when coupled to the MNPs in the nanoring. The addition of the QDs into the MNP nanoring cause the material's permeability to become negative at their resonance frequency, as seen in Fig. [\[fig4F\]](#fig4F){reference-type="ref" reference="fig4F"} and Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}. Thus, if the QD was saturated by a separate control field then the permeability could be controlled and varied with the light intensity between positive and negative values. In future work, techniques from many-body quantum systems  may be used together with our formalism in order to make the computation accessible. # Conclusion We have developed a quantum optical model to describe the dynamics of a negative permeability metamaterial integrated with two-level QD systems. Using this model we found that the Fano interference of a MNP-QD metamolecule can manifest itself in the macroscopic magnetic response of a metamaterial consisting of MNP-QD nanorings. We have shown that this effect can be used to tune the properties of the metamaterial. Our model is also useful to study nonlinear effects that arise when the metamaterial is driven strongly. We showed an example of this by studying how the non-linear Fano effect can affect our nanoring. Each MNP has its own electric response, however the frequencies at which the MNP nanoring metamaterial has a negative permeability and the frequencies at which it has a negative permittivity do not overlap. Even with inclusion of the QDs this problem remains. We have performed calculations which show that the Fano profile in the electric and magnetic scattered fields cannot be tuned independently in our scheme. Without this ability it is not possible to ensure a frequency overlap. Instead, in future work we intend to investigate how to incorporate the magnetic nanoring resonators with a broadband negative permittivity background , using various types of lattice configuration. In this way by dynamically tuning the magnetic response, we may also be able to control the metamaterial's refractive index. However, even for a material that has only a negative permeability, recent work has shown interesting quantum dynamics can be observed in the spontaneous emission interference of an emitter placed in close proximity . The ability to tune and saturate the magnetic response in this scenario may open up new additional features. Another direction of future work would be to exam how fluorescence quenching of the QD by the MNP would affect our system. In our parameter space we do not expect quenching to be a major factor , however it would be interesting to quantify at what point this approximation no longer holds for the system we have studied. Furthermore, in our calculations we have used similar parameters to previous studies . However, it has been noted recently that the Ohmic damping used for the MNP fields may well be an underestimation , resulting from the discrepancy between the Drude model used theoretically and experimental results at higher frequencies. Using more realistic damping it has been shown that rings with spherical MNPs no longer have a strong enough magnetic resonance to achieve \({\rm Re}(\mu) < 0\). However, this problem may potentially be resolved using MNP's with embedded gain material  or different type of nanostructures in place of the MNPs. Indeed, by considering more complex, strongly polarized plasmonic nanostructures within the ring  negative permeability has been shown to be possible, whilst ensuring damping is correctly accounted for. Here, Morits and Simovski have explored the use of dimers  and nanoprisms , with the latter providing negative permeability in the visible regime. In both cases as only dipole interactions are considered, the basic theoretical model we have developed in this work using MNPs can be transferred to these more complex nanostructures in a straightforward manner with the qualitative results of our analysis remaining valid. # Appendix {#appendix .unnumbered}

{'timestamp': '2014-01-22T02:11:30', 'yymm': '1401', 'arxiv_id': '1401.5402', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5402'}

# Introduction Recall that a subgroup \(H\) of \(G\) is said to be \(s\)-quasinormal (or \(s\)-permutable) in \(G\) if \(H\) is permutable with every Sylow subgroup \(P\) of \(G\) (that is, \(HP=PH\)). The \(s\)-permutableity of a subgroup of a finite group \(G\) often yields a wealth of information about the group \(G\) itself. In the past, it has been studied by many scholars (such as \[1-2\], \[7-9\], \[13\], \[17\]). Recently, Huang \[10\] introduced the following concept: Note that, for a class \(\mathfrak{F}\) of groups, a chief factor \(H/K\) of a group \(G\) is called \(\mathfrak{F}\)-central (see \[16\] or \[4, Definition 2.4.3\]) if \([H/K](G/C_G(H/K))\in \mathfrak{F}\). The symbol \(Z_\infty ^\frak{F} (G)\) denotes the \(\mathfrak{F}\)-hypercenter of a group \(G\), that is, the product of all such normal subgroups \(H\) of \(G\) whose \(G\)-chief factors are \(\mathfrak{F}\)-central. A subgroup \(H\) of \(G\) is said to be \(\mathfrak{F}\)-hypercenter in \(G\) if \(H\leq Z_\infty ^\frak{F}(G)\). By using this new concept, Huang \[10\] has given some conditions under which a finite group belongs to some formations. In this paper, we will go to further into the influence of \(\mathfrak{F_{\mathrm s}}\)-quasinormal subgroups on the structure of finite groups. New characterizations of some classes of finite groups are obtained. All groups considered in the paper are finite and \(G\) denotes a finite group. The notations and terminology in this paper are standard, as in \[4\] and \[14\]. # Preliminaries Let \(\frak{F}\) be a class of finite groups. Then \(\frak{F}\) is called a formation if it is closed under homomorphic image and every group \(G\) has a smallest normal subgroup (called \(\mathfrak{F}\)-residual and denoted by \(G^\mathfrak{F}\)) with quotient is in \(\mathfrak{F}\). \(\mathfrak{F}\) is said to be saturated if it contains every group \(G\) with \(G/\Phi(G) \in \mathfrak{F}\). \(\mathfrak{F}\) is said to be \(S\)-closed (\(S_n\)-closed) if it contains all subgroups (all normal subgroups, respectively) of all its groups. We use \(\mathfrak{N}\), \(\mathfrak{U}\), and \(\mathfrak{S}\) to denote the formations of all nilpotent groups, supersoluble groups and soluble groups, respectively. The following known results are useful in our proof. **Lemma 2.1** \[8, Lemma 2.2\]. Let \(G\) be a group and \(H\leq K\leq G\). (1) If \(H\) is \(s\)-permutable in \(G\), then \(H\) is \(s\)-permutable in \(K\); (2) Suppose that \(H\) is normal in \(G\). Then \(K/H\) is \(s\)-permutable in \(G/H\) if and only if \(K\) is \(s\)-permutable in \(G\); (3) If \(H\) is \(s\)-permutable in \(G\), then \(H\) is subnormal in \(G\); (4) If \(H\) and \(F\) are \(s\)-permutable in \(G\), the \(H\cap F\) is \(s\)-permutable in \(G\); (5) If \(H\) is \(s\)-permutable in \(G\) and \(M\leq G\), then \(H\cap M\) is \(s\)-permutable in \(M\). **Lemma 2.2** \[10, Lemma 2.3\]. Let \(G\) be a group and \(H\leq K\leq G\). (1) \(H\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(G\) if and only if there exists a normal subgroup \(T\) of \(G\) such that \(HT\) is \(s\)-permutable in \(G\), \(H_G\leq T\) and \(H/H_G\cap T/H_G\leq Z_\infty ^\frak{F} (G/H_G)\); (2) Suppose that \(H\) is normal in \(G\). Then \(K/H\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(G/H\) if and only if \(K\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(G\); (3) Suppose that \(H\) is normal in \(G\). Then, for every \(\mathfrak{F_{\mathrm s}}\)-quasinormal subgroup \(E\) of \(G\) satisfying (\(|H|\),\(|E|\))=1, \(HE/H\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(G/H\); (4) If \(H\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(G\) and \(\frak{F}\) is S-closed, then \(H\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(K\); (5) If \(H\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(G\), \(K\) is normal in \(G\) and \(\frak{F}\) is \(S_n\)-closed, then \(H\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(K\); (6) If \(G\in \frak{F}\), then every subgroup of \(G\) is \(\mathfrak{F_{\mathrm s}}\)-quasinormal in \(G\). **Lemma 2.3** \[6, Lemma 2.2\]. If \(H\) is a \(p\)-subgroup of \(G\) for some prime \(p\) and \(H\) is \(s\)-permutable in \(G\), then: (1) \(H\leq O_{p}(G)\); (2) \(O^{p}(G)\leq N_{G}(H)\). **Lemma 2.4** \[18\]. If \(A\) is a subnormal subgroup of a group \(G\) and \(A\) is a \(\pi\)-group, then \(A\leq O_\pi(G)\). **Lemma 2.5** \[15, II, Lemma 7.9\]. Let \(N\) be a nilpotent normal subgroup of \(G\). If \(N\neq 1\) and \(N\cap \Phi(G)=1\), then \(N\) is a direct product of some minimal normal subgroups of \(G\). **Lemma 2.6** \[5, Lemma 2.3\]. Let \(\mathfrak{F}\) be a saturated formation containing \(\mathfrak{U}\) and \(G\) a group with a normal subgroup \(E\) such that \(G/E\in \mathfrak{F}\). If \(E\) is cyclic, then \(G\in \mathfrak{F}\). Recall that a subgroup \(H\) of \(G\) is said to be \(\mathfrak{F}\)-supplemented in \(G\) if there exists a subgroup \(T\) of \(G\) such that \(G=HT\) and \(T\in \mathfrak{F}\), where \(\mathfrak{F}\) is some class of groups. The following Lemma is clear. **Lemma 2.7** Let \(\mathfrak{F}\) be a formation and \(H\) a subgroup of \(G\). If \(H\) has an \(\mathfrak{F}\)-supplement in \(G\), then: (1) If \(N\unlhd G\), then \(HN/N\) has an \(\mathfrak{F}\)-supplement in \(G/N\). (2) **If \(\mathfrak{F}\) is \(S\)-closed and \(H\leq K\leq G\), then \(H\) has an \(\mathfrak{F}\)-supplement in \(K\).** **Lemma 2.8** \[10, Theorem 3.1\]. Let \(\mathfrak{F}\) be an \(S\)-closed saturated formation containing \(\mathfrak{U}\) and \(G\) a group. Then \(G\in \mathfrak{F}\) if and only if \(G\) has a normal subgroup \(E\) such that \(G/E \in \mathfrak{F}\) and every maximal subgroup of every non-cyclic Sylow subgroup of \(E\) not having a supersoluble supplement in \(G\) is \(\mathfrak{U}_{\mathrm s}\)-quasinormal in \(G\). **Lemma 2.9** \[10, Theorem 3.2\]. Let \(\mathfrak{F}\) be a saturated formation containing \(\mathfrak{U}\) and \(G\) a group. Then \(G\in \mathfrak{F}\) if and only if \(G\) has a soluble normal subgroup \(E\) such that \(G/E \in \mathfrak{F}\) and every maximal subgroup of every non-cyclic Sylow subgroup of \(F(E)\) not having a supersoluble supplement in \(G\) is \(\mathfrak{U}_{\mathrm s}\)-quasinormal in \(G\). **Lemma 2.10** \[3, Main Theorem\]. Suppose \(G\) has a Hall \(\pi\)-subgroup and \(2\notin \pi\). Then all the Hall \(\pi\)-subgroups are conjugate in \(G\). **Lemma 2.11** \[6, Lemma 2.5\]. Let \(G\) be a group and \(p\) a prime such that \(p^{n+1}\nmid |G|\) for some integer \(n\geq 1\). If \((|G|,(p-1)(p^{2}-1)\cdot \cdot \cdot (p^{n}-1))=1\), then \(G\) is \(p\)-nilpotent. The generalized Fitting subgroup \(F^{*}(G)\) of a group \(G\) is the product of all normal quasinilpotent subgroups of \(G\). We also need in our proofs the following well-known facts about this subgroups (see \[12, Chapter X\]). **Lemma 2.12**. Let \(G\) be a group and \(N\) a subgroup of \(G\). (1) If \(N\) is normal in \(G\), then \(F^{*}(N)\leq F^{*}(G)\). (2) If \(N\) is normal in \(G\) and \(N\leq F^{*}(G)\), then \(F^{*}(G)/N\leq F^{*}(G/N)\). (3) \(F(G)\leq F^{*}(G)=F^{*}(F^{*}(G))\). If \(F^{*}(G)\) is soluble, then \(F^{*}(G)=F(G)\). (4) \(C_{G}(F^{*}(G))\leq F(G)\). (5) \(F^{*}(G)=F(G)E(G)\), \(F(G)\cap E(G)=Z(E(G))\) and \(E(G)/Z(E(G))\) is the direct product of simple non-abelian groups, where \(E(G)\) is the layer of \(G\). **Lemma 2.13** \[8, Lemma 2.15-2.16\]. (1) If \(H\) is a normal soluble subgroup of a group \(G\), then \(F^{*}(G/\Phi (H))=F^{*}(G)/\Phi (H)\). (2) If \(K\) is a normal \(p\)-subgroup of a group \(G\) contained in \(Z(G)\), then \(F^{*}(G/K)=F^{*}(G)/K\). # New Characterization of supersoluble groups **Proof.** The necessity is obvious since \(Z_\infty ^\frak{S}(G)=G\) whenever \(G\in \mathfrak{S}\). Hence we only need to prove the sufficiency. Suppose that the assertion is false and let \(G\) be a counterexample of minimal order. Then \(\textit{p}=2\) by the well known Feit-Thompson Theorem of groups of odd order. We proceed the proof via the following steps: *(1) \(O_{2}(G)=1\).* Assume that \(N=O_{2}(G)\neq 1\). Then \(P/N\) is a Sylow 2-subgroup of \(G/N\). Let \(M/N\) be a maximal subgroup of \(P/N\). Then \(M\) is a maximal subgroup of \(P\). By the hypothesis and Lemma 2.2(2), \(M/N\) is \(\mathfrak{S_{\mathrm s}}\)-quasinormal in \(G/N\). The minimal choice of \(G\) implies that \(G/N\) is soluble. It follows that \(G\) is soluble, a contradiction. Hence (1) holds. *(2) \(O_{2^{'}}(G)=1\).* Assume that \(D=O_{2^{'}}(G)\neq 1\). Then \(PD/D\) is a Sylow 2-subgroup of \(G/D\). Suppose that \(M/D\) is a maximal subgroup of \(PD/D\). Then there exists a maximal subgroup \(P_{1}\) of \(P\) such that \(M=P_{1}D\). By the hypothesis and Lemma 2.2(3), \(M/D=P_{1}D/D\) is \(\mathfrak{S_{\mathrm s}}\)-quasinormal in \(G/D\). Hence \(G/D\) is soluble by the choice of \(G\). It follows that \(G\) is soluble, a contradiction. *(3) Final contradiction.* Let \(P_{1}\) be a maximal subgroup of \(P\). By the hypothesis, there exists a normal subroup \(K\) of \(G\) such that \(P_{1}K\) is \(s\)-permutable in \(G\) and \((P_{1}\cap K)(P_{1})_{G}/(P_{1})_{G}\leq Z_\infty ^\frak{S}(G/(P_{1})_{G})\). Note that \(Z_\infty ^\frak{S}(G)\) is a soluble normal subgroup of \(G\). By (1) and (2), we have \((P_{1})_{G}=1\) and \(Z_\infty ^\frak{S}(G)=1\). This induces that \(P_{1}\cap K=1\). If \(K=1\), then \(P_{1}\) is \(s\)-permutable in \(G\) and so \(P_{1}=1\) by (1) (2) and Lemma 2.3(1). This means that \(|P|=2\). Then by \[14, (10.1.9)\], \(G\) is \(2\)-nilpotent and so \(G\) is soluble, a contradiction. We may, therefore, assume that \(K\neq 1\). If \(2\mid |K|\), then \(|K_{2}|=2\), where \(K_{2}\) is a Sylow 2-subgroup of \(K\). By \[14, (10.1.9)\] again, we see that \(K\) is 2-nilpotent, and so \(K\) has a normal 2-complement \(K_{2^{'}}\). Since \(K_{2^{'}}\) char \(K\unlhd G\), \(K_{2^{'}}\unlhd G\). Hence by (2), \(K_{2^{'}}=1\). Consequently \(|K|=2\), which contradicts (1). If \(2\nmid |K|\), then \(K\) is a \(2^{'}\)-group. Hence by (2), \(K\leq O_{2^{'}}(G)=1\), also a contradiction. This completes the proof. **Proof.** Suppose that the assertion is false and let \(G\) be a counterexample of minimal order. Then: *(1) Each proper subgroup of \(G\) containing \(A\) is supersoluble.* Let \(A\leq M< G\). Then \(M=M\cap AB=A(M\cap B)\). Obviously, \(M\cap B\) is a Hall subgroup of \(M\) and every Sylow subgroup of \(M\cap B\) is cyclic. By Lemma 2.2(4), every maximal subgroup of every non-cyclic Sylow subgroup of \(A\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(M\). The minimal choice of \(G\) implies that \(M\) is supersoluble. *(2) Let \(H\) be a non-trivial normal \(p\)-subgroup of \(G\) for some prime \(p\). If \(H\) contains some Sylow \(p\)-subgroup of \(A\) or a Sylow \(p\)-subgroup of \(A\) is cyclic or \(H\leq A\), then \(G/H\) is supersoluble.* If \(A\leq H\), then \(G/H=BH/H\cong B/(B\cap H)\) is supersoluble. Now we can assume that \(A\nleq H\). Clearly, \(G/H=(AH/H)(BH/H)\), where \(AH/H\) is subnormal in \(G/H\) and \(BH/H\) is supersoluble. Let \(Q/H\) be any non-cyclic Sylow \(q\)-subgroup of \(AH/H\) and \(Q_{1}/H\) a maximal subgroup of \(Q/H\). Then there exists a non-cyclic Sylow \(q\)-subgroup \(A_{q}\) of \(A\) such that \(Q=A_{q}H\) and a maximal subgroup \(A_{1}\) of \(A_{q}\) such that \(Q_{1}=A_{1}H\). If \(H\leq A\), then the assertion holds by the choice of \(G\) and Lemma 2.2(2). We may, therefore, assume that \(H\nleq A\). Let \(P\) be a Sylow \(p\)-subgroup of \(A\). Assume that \(P\) is cyclic or \(P\leq H\). Then \(p\neq q\). Clearly, \(Q_{1}\cap A_{q}=A_{1}\) is a maximal subgroup of \(A_{q}\). By the hypothesis, \(A_{1}\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). Therefore, \(Q_{1}/H=A_{1}H/H\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G/H\) by Lemma 2.2(3). This shows that the conditions of the theorem are true for \(G/H\) and so \(G/H\) is supersoluble by the minimal choice of \(G\). *(3) There exists at least one Sylow subgroup of \(A\) which is non-cyclic.* It follows from the well known fact that a group \(G\) is supersoluble if all its Sylow subgroups are cyclic. *(4) \(G\) is soluble.* If \(A\neq G\), then \(A\) is supersoluble by (1). Let \(p\) be the largest prime divisor of \(|A|\). Then \(A_{p}\unlhd A\). By Lemma 2.4, \(A_{p}\leq O_{p}(G)\). By (2), \(G/O_{p}(G)\) is supersoluble. It follows that \(G\) is soluble. We now only need to consider the case that \(A=G\). If \(G\) is not soluble and let \(p\) be the minimal prime divisor of \(|G|\). Then \(p=2\) by the well-known Feit-Thompson Theorem. Hence by Lemma 3.1, \(G\) is soluble. *(5) \(G\) has a unique minimal normal subgroup \(N\) such that \(N=O_p(G)=C_G(N)\) is a non-cyclic \(p\)-subgroup of \(G\) for some prime \(p\) and \(G=[N]M\), where \(M\) is a supersoluble maximal subgroup of \(G\).* Let \(N\) be an arbitrary minimal normal subgroup of \(G\). By (4), \(N\) is a \(p\)-group. If \(p\in \pi (B)\), then the Sylow \(p\)-subgroups of \(G\) are cyclic and so the Sylow \(p\)-subgroups of \(A\) are cyclic. If \(p\notin \pi (B)\), then clearly, \(N\subseteq A\). Hence by (2), \(G/N\) is supersoluble. If \(N\) is cyclic, then by Lemma 2.6, \(G\) is supersoluble, a contradiction. Since the class of all supersoluble groups is a saturated formation, \(N\) is the only minimal normal subgroup \(N\) of \(G\) and \(\Phi(G)=1\). This implies that (5) holds. *(6) \(N\) is not a Sylow subgroup of \(G\) and \(Z_\infty ^\mathfrak{U}(G)=1\).* By (5), clearly, \(Z_\infty ^\mathfrak{U}(G)=1\). Assume that \(N\) is a Sylow \(p\)-subgroup of \(G\). Let \(N_{1}\) be a maximal subgroup of \(N\). Then by hypothesis, \(N_{1}\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). Hence there exists a normal subgroup \(K\) of \(G\) such that \(N_{1}K\) is \(s\)-permutable in \(G\) and \(N_{1}\cap K\leq Z_\infty ^\frak{U}(G)=1\) since \((N_{1})_{G}=1\). It follows that \(N_{1}\leq N_{1}\cap N\leq N_{1}\cap K=1\). Hence \(|N|=p\). This contradiction shows that \(N\) is not a Sylow \(p\)-subgroup of \(G\). *(7) \(A\) is supersoluble.* If \(A\) is not supersoluble, then \(G=A\) by (1). Let \(q\) be the largest prime divisor of \(|G|\) and \(Q\) is a Sylow \(q\)-subgroup of \(G\). Then \(QN/N\) is a Sylow \(q\)-subgroup of \(G/N\). Since \(G/N\) is supersoluble, \(QN/N\unlhd G/N\). It follows that \(QN\unlhd G\). Let \(P\) be a non-cyclic Sylow \(p\)-subgroup of \(G=A\). If \(p=q\), then \(P=Q=QN\unlhd G\). Therefore \(N=O_p(G)=P\) is the Sylow \(p\)-subgroup of \(G\), a contradiction. Assume that \(q> p\). Then clearly \(QP=QNP\) is a subgroup of \(G\). Since \(N\nleq \Phi (G)\), \(N\nleq \Phi (P)\) by \[11, III, Lemma 3.3(a)\]. Let \(P_1\) be a maximal subgroup of \(P\) such that \(N\nleq P_{1}\). Then \((P_1)_G=1\). By the hypothesis, \(P_1\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). Hence, there exists a normal subgroup \(T\) of \(G\) such that \(P_1T\) is \(s\)-permutable in \(G\) and \(P_1\cap T\leq Z_\infty^\mathfrak{U}(G)=1\). Obviously, \(T\neq 1\) (In fact, if \(T=1\), then \(P_{1}\leq O_{p}(G)=N\) by Lemma 2.3(1). Hence \(P_{1}=N\) or \(P=N\). This is impossible). Thus \(N\leq T\), and so \(P_1\cap N\leq P_1\cap T=1\). This induces that \(|N|=|P:P_{1}|=p\), which contradicts (5). Thus (7) holds. *(8) The final contradiction.* Let \(q\) be the largest prime divisor of \(|A|\) and \(A_{q}\) a Sylow \(p\)-subgroup of \(A\). Since \(A\) is supersoluble by (7), \(A_{q}\unlhd A\). Hence \(A_{q}\leq O_{q}(G)\). If \(q\mid |B|\), then \(O_{q}(G)\leq G_{q}\), where \(G_{q}\) is a cyclic Sylow \(q\)-subgroup of \(B\) and so \(O_{q}(G)\) is cyclic. In view of (2), \(G/O_{q}(G)\) is supersoluble. It follows that \(G\) is supersoluble, a contradiction. Hence \(q\nmid |B|\). Then, \(A_{q}\) is a Sylow \(q\)-subgroup of \(G\) and so \(A_{q}=O_{q}(G)\neq 1\). This means that \(q=p\) and so \(N=A_{p}=G_{p}\), which contradicts (6). The final contradiction completes the proof. **Proof.** We first prove that the theorem is true if \(\mathfrak{F}=\mathfrak{U}\). Suppose that the assertion is false and consider a counterexample for which \(|G||H|\) is minimal. Then: *(1) \(H=G\) and \(F^{*}(G)=F(G)\).* By Lemma 2.8, \(F^{*}(H)\) is supersoluble. Hence \(F^{*}(H)=F(H)\) by Lemma 2.12(3). Since \((H,H)\) satisfies the hypothesis, the minimal choice of \((G,H)\) implies that \(H\) is supersoluble if \(H< G\). Then \(G\in \mathfrak{U}\) by Lemma 2.9, a contradiction. *(2) Every proper normal subgroup \(N\) of \(G\) containing \(F^{*}(G)\) is supersoluble.* Let \(N\) be a proper normal subgroup of \(G\) containing \(F^{*}(G)\). By Lemma 2.12, \(F^{*}(G)=F^{*}(F^{*}(G))\leq F^{*}(N)\leq F^{*}(G)\). Hence \(F^{*}(N)=F^{*}(G)\). Let \(M\) be a maximal subgroup of any non-cyclic Sylow subgroup of \(F^{*}(N)\). If there exists a supersoluble subgroup \(T\) such that \(G=MT\), then \(N=M(N\cap T)\) and \(N\cap T\in \mathfrak{U}\). This means that \(M\) has a supersoluble supplement in \(N\). Now assume that \(M\) has no supersoluble supplement in \(G\). Then by hypothesis and Lemma 2.2(4), \(M\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(N\). This shows that \((N,N)\) satisfies the hypothesis. Hence \(N\) is supersoluble by the minimal choice of \((G,H)\). *(3) If \(p\in \pi(F(G))\), then \(\Phi(O_{p}(G))=1\) and so \(O_{p}(G)\) is elementary abelian. In particular, \(F^{*}(G)=F(G)\) is abelian and \(C_{G}(F(G))=F(G)\).* Suppose that \(\Phi(O_{p}(G))\neq 1\) for some \(p\in \pi(F(G))\). By Lemma 2.13(1), we have \(F^{*}(G/\Phi (O_{p}(G)))=F^{*}(G)/\Phi (O_{p}(G))\). By using Lemma 2.2, we see that the pair \((G/\Phi (O_{p}\) \((G)),F^{*}(G)/\Phi (O_{p}(G)))\) satisfies the hypothesis. The minimal choice of \((G,H)\) implies \(G/\Phi (O_{p}(G))\) \(\in \mathfrak{U}\). Since \(\mathfrak{U}\) is a saturated formation, we obtain that \(G\in \mathfrak{U}\), a contradiction. This means that \(\Phi(O_{p}(G))=1\) and so \(O_{p}(G)\) is elementary abelian. Hence \(F^{*}(G)=F(G)\) is abelian and \(F(G)\leq C_{G}(F(G))\). Put \(N=C_{G}(F(G))\). Then, clearly, \(F(N)=F(G)\). If \(N=G\), then \(F(G)\leq Z(G)\). Let \(P_1\) be a maximal subgroup of some Sylow \(p\)-subgroup of \(F(G)\). Then \(F(G/P_1)=F(G)/P_1\) by Lemma 2.13(2). Hence \((G/P_1, F(G)/P_1)\) satisfies the hypothesis and so \(G/P_1\in \mathfrak{F}\). Then since \(P\leq Z(G)\), we obtain \(G\in \mathfrak{F}\). This contradiction shows that \(N<G\). Hence by (2), \(N\) is soluble and so \(C_{N}(F(N))\subseteq F(N)\). It follows that \(N=C_{G}(F(G))=F(G)\). *(4) \(G\) has no normal subgroup of prime order contained in \(F(G)\).* Suppose that \(L\) is a normal subgroup of \(G\) contained in \(F(G)\) and \(|L|=p\). Put \(C=C_{G}(L)\). Clearly, \(F(G)\leq C\unlhd G\). If \(C<G\), then \(C\) is soluble by (2). Since \(G/C\) is cyclic, \(G\) is soluble. Then by the hypothesis and Lemma 2.9, \(G\in \mathfrak{U}\), a contradiction. Hence \(C=G\) and so \(L\leq Z(G)\). By Lemma 2.13(2) \(F^{*}(G/L)=F^{*}(G)/L=F(G)/L\). Hence \(G/L\) satisfies the hypothesis by Lemma 2.2. The minimal choice of \((G,H)\) implies that \(G/L\in \mathfrak{U}\) and consequently \(G\) is supersoluble, a contradiction. *(5) For some \(p\in \pi(F(G))\), \(O_{p}(G)\) is a non-cyclic Sylow \(p\)-subgroup of \(F(G)\).* Clearly, \(F(G)=O_{p_{1}}(G)\times O_{p_{2}}(G)\times \cdot \cdot \cdot \times O_{p_r}(G)\) for some primes \(p_i\), \(i=1,2,\cdots, r.\) If all Sylow subgroups of \(F(G)\) are cyclic, then \(G/C_{G}(O_{p_{i}}(G))\) is abelian for any \(i\in \{1\cdot \cdot \cdot r\}\) and so \(G/\cap^{r}_{i=1}C_{G}(O_{p_{i}}(G))=G/C_{G}(F(G))=G/F(G)\) is abelian. Therefore \(G\) is soluble. It follows from Lemma 2.9 and the hypothesis that \(G\in \mathfrak{U}\), a contradiction. *(6) Every maximal subgroup of every non-cyclic Sylow subgroup of \(F(G)\) has no supersoluble supplement in \(G\).* Let \(P\) be a non-cyclic Sylow subgroup of \(F(G)\) and \(P_{1}\) a maximal subgroup of \(P\). Then \(P=O_{p}(G)\) for some \(p\in \pi(F(G))\). If \(P_{1}\) has a supersoluble supplement in \(G\), that is, there exists a supersoluble subgroup \(K\) of \(G\) such that \(G=P_{1}K=O_{p}(G)K\), then \(G/O_{p}(G)\simeq K/K\cap O_{p}(G)\) is supersoluble and so \(G\) is soluble. Hence as above, \(G\in \mathfrak{U}\), a contradiction. *(7) \(P\cap \Phi(G)\neq 1\), for some non-cyclic Sylow subgroup \(P\) of \(F(G)\).* Assume that \(P\cap \Phi(G)=1\). Then \(P=R_{1}\times R_{2}\times\) \(\cdot \cdot \cdot \times R_{m}\), where \(R_{i}(i\in \{1,\cdot \cdot \cdot m\})\) is a minimal normal subgroup of \(G\) by Lemma 2.5. We claim that \(R_{i}\) are of order \(\textit{p}\) for all \(i\in \{1,\cdot \cdot \cdot m\}\). Assume that \(|R_{i}|>\textit{p}\), for some \(\textit{i}\). Without loss of generality, we let \(|R_{1}|>\textit{p}\). Let \(R^{*}_{1}\) be a maximal subgroup of \(R_{1}\). Obviously, \(R^{*}_{1}\neq 1\). Then \(R^{*}_{1}\times R_{2}\times\) \(\cdot \cdot \cdot \times R_{m}=P_{1}\) is a maximal subgroup of \(P\). Put \(T=R_{2}\times\) \(\cdot \cdot \cdot \times R_{m}\), Clearly \((P_{1})_{G}=T\). By (6) and the hyperthesis, \(P_{1}\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). Hence by Lemma 2.2(1), there exists a normal subgroup \(N\) of \(G\) such that \((P_{1})_{G}\leq N\), \(P_{1}N\) is \(s\)-permutable in \(G\) and \(P_{1}/(P_{1})_{G} \cap N/(P_{1})_{G}\leq Z_\infty ^\frak{U} (G/(P_{1})_{G})\). Assume that \(P_{1}/(P_{1})_{G} \cap N/(P_{1})_{G}\neq 1\). Let \(Z_\infty ^\frak{U} (G/(P_{1})_{G})=V/(P_{1})_{G}=V/T\). Then \(P/T \cap V/T \unlhd G/T\). Since \(P\cap V\geq P_{1}\cap N\cap V\geq P_{1}\cap N> (P_{1})_{G}=T\), we have \(P/T \cap V/T\neq 1\). Because \(P/T\simeq R_{1}\) and \(R_{1}\) is a minimal normal subgroup of \(G\), \(P/T\subseteq V/T\). This implies that \(|R_{1}|=|P/T|=p\). This contradiction shows that \(P_{1}\cap N=(P_{1})_{G}=T\). Consequently \(P_{1}N=R^{*}_{1}TN=R^{*}_{1}N\) and \(R^{*}_{1}\cap N=1\). Since \(R_{1}\cap N\unlhd G\), \(R_{1}\cap N=1\) or \(R_{1}\cap N=R_{1}\). But since \(R^{*}_{1}\cap N=1\), we have that \(R_{1}\cap N=1\). Thus \(R^{*}_{1}=R^{*}_{1}(R_{1}\cap N)=R_{1}\cap R^{*}_{1}N\) is \(s\)-permutable in \(G\). It follows from Lemma 2.3(2) that \(O^{p}(G)\leq N_{G}(R_{1}^{*})\). Thus \(|G:N_{G}(R_{1}^{*})|\) is a power of \(p\) for every maximal subgroup \(R_{1}^{*}\) of \(R_{1}\). This induces that \(p\) divides the number of all maximal subgroups of \(R_{1}\). This contradicts \[11, III, Theorem 8.5(d)\]. Therefore \(|R_{i}|=p\), which contradicts (4). Thus (7) holds. *(8) \(F(G)=P\) is a \(p\)-group, \(P\) contains a unique minimal normal subgroup \(L\) of \(G\) and \(L\subseteq \Phi (G)\).* Suppose that \(1\neq Q\) is a Sylow \(q\)-subgroup of \(F(G)\) for some prime \(q\neq p\) and let \(L\) be a minimal normal subgroup of \(G\) contained in \(P\cap \Phi (G)\). By (3), \(Q\) is elementary abelian. By Lemma 2.12, \(F^{*}(G/L)=F(G/L)E(G/L)\) and \([F(G/L),E(G/L)]=1\), where \(E(G/L)\) is the layer of \(G/L\). Since \(L\leq \Phi (G)\), \(F(G/L)=F(G)/L\). Now let \(E/L=E(G/L)\). Since \(Q\) is normal in \(G\) and \([F(G)/L,E/L]=1\), we have \([Q,E]\leq Q\cap L=1\). It follows from (3) that \(F(G)E\leq C_{G}(Q)\unlhd G\). If \(C_{G}(Q)< G\), then \(C_{G}(Q)\) is supersoluble by (1) and (2). Thus \(E(G/L)=E/L\) is supersoluble and consequently \(F^{*}(G/L)=F(G)/L\) by Lemma 2.12(5). Now, by Lemma 2.2, we see that \((G/L,F(G)/L)\) satisfies the hypothesis. The minimal choice of \((G,H)\) implies that \(G/L\) is supersoluble and so is \(G\). This contradiction shows that \(C_{G}(Q)=G\), i.e. \(Q\leq Z(G)\), which contradicts (4). Thus \(F(G)=P\). Let \(X\) be a minimal normal subgroup of \(G\) contained in \(P\) with \(X\neq L\). Let \(E/L=E(G/L)\) is the layer of \(G/L\). As above, we see that \(F^{*}(G/L)=F(G/L)E(G/L)\) and \([F(G)/L,E/L]=1\). Hence \([X,E]\leq X\cap L=1\), i.e., \([X,E]=1\). It follows from (3) that \(F(G)E\leq C_{G}(X)\unlhd G\). If \(C_{G}(X)<G\), then \(C_{G}(X)\) is supersoluble by (1) and (2). Thus \(E(G/L)=E/L\) is supersoluble and consequently \(F^{*}(G/L)=F(G)/L\). Obviously, \(G/L\) satisfies the hypothesis. By the choice of \((G,H)\), we have that \(G/L\) is supersoluble and so is \(G\), a contradiction. Hence \(C_{G}(X)=G\), i.e. \(X\leq Z(G)\), which also contradicts (4). Thus \(L\) is the unique minimal normal subgroup of \(G\) contained in \(P\). Finally, \(L\subseteq \Phi (G)\) by (7). *(9) \(L<P\).* Suppose \(L=P\). Let \(P_{1}\) be a maximal subgroup of \(P\) such that \(P_1\) is normal in some Sylow subgroup of \(G\). Then \((P_{1})_{G}=1\). By the hypothesis and (8), \(P_{1}\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). Hence there exists a normal subgroup \(K\) of \(G\) such that \(P_{1}K\) is \(s\)-permutable in \(G\) and \(P_{1}\cap K\leq Z_\infty ^\mathfrak{U}(G)\). If \(P_{1}\cap K\neq 1\), then \(1<P_{1}\cap K\leq P\cap Z_\infty ^\mathfrak{U}(G)\), which implies that \(P=P\cap Z_\infty ^\mathfrak{U}(G)\) and \(|P|=p\) since \(P\) is a minimal normal subgroup of \(G\). This contradicts (4). So we may assume \(P_{1}\cap K=1\). Since \(P\) is a minimal normal subgroup of \(G\), \(P\cap K=P\) or \(1\). If \(P\cap K=P\), then \(P\subseteq K\), and so \(|P|=p\), which contradicts (4). If \(P\cap K=1\), then \(P\cap P_{1}K=P_{1}(P\cap K)=P_{1}\). Hence \(P_{1}\) is \(s\)-permutable in \(G\). Then by Lemma 2.3(2), \(O^{p}(G)\leq N_{G}(P_{1})\). This induces that \(P_{1}\unlhd G\). This means that \(P_{1}=(P_{1})_{G}=1\) and \(|P|=p\), also a contradiction. *(10) Final contradiction (for \(\mathfrak{F}=\mathfrak{U}\)).* By (3) and (8), \(P\) is an elementary abelian group, and so \(L\) has a complement in \(P\), \(T\) say. Let \(P_{1}=TL_{1}\), where \(L_{1}\) is a maximal subgroup of \(L\). Then \(1\neq P_{1}\) and clearly \(P_1\) is a maximal subgroup of \(P\) such that \(P_1\) is normal in some Sylow subgroup of \(G\). Hence by (6), \(P_{1}\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\) and \((P_{1})_{G}=1\) since \(L\) is the unique minimal normal subgroup of \(G\) contained in \(P\). Hence there exists a normal subgroup \(S\) of \(G\) such that \(P_{1}S\) is \(s\)-permutable in \(G\) and \(P_{1}\cap S\leq Z_\infty ^\mathfrak{U}(G)\). If \(P_{1}\cap S\neq 1\), then \(1< P_{1}\cap S\leq P\cap Z_\infty ^\mathfrak{U}(G)\) and so \(G\) has a minimal normal subgroup \(N\) of order \(p\) contained in \(P\), which is contrary to (4). Hence \(P_{1}\cap S=1\). If \(P\cap S\neq 1\), then \(L\leq P \cap S\) and so \(L_{1}\leq S\), which contradicts \(P_{1}\cap S= 1\). If \(P\cap S=1\), then \(P_{1}=P_{1}(P\cap S)=P\cap P_{1}S\) is \(s\)-permutable in \(G\). Hence \(O^{p}(G)\leq N_{G}(P_{1})\) by Lemma 2.3(2). It follows that \(P_{1}\unlhd G\), which contradicts \((P_{1})_{G}=1\). The final contradiction shows that the theorem holds when \(\mathfrak{F}=\mathfrak{U}\). Now we prove that the theorem holds for \(\mathfrak{F}\). Since \(H/H\in \mathfrak{U}\), by the assertion proved above and Lemma 2.2, we see that \(H\) is supersoluble. In particular, \(H\) is soluble and hence \(F^{*}(H)=F(H)\). Now by using Lemma 2.9, we obtain that \(G\in \mathfrak{F}\). This completes the proof of the theorem. # New Characterization of \(p\)-nilpotent groups **Proof.** Assume that \(p^{n+1}\mid |G|\). Let \(P_{n1}\) be an \(n\)-maximal subgroup of \(P\). By hypothesis, \(P_{n1}\) has a \(p\)-nilpotent supplement \(T_{1}\) in \(G\). Let \(K_{1}\) be a normal Hall \(p'\)-subgroup of \(T_{1}\). Obviously, \(K_{1}\) is a Hall \(p'\)-subgroup of \(G\). Hence \(G=P_{n1}T_{1}=P_{n1}N_{G}(K_{1})\). We claim that \(K_{1}\unlhd G\). Indeed, if \(K_{1}\ntrianglelefteq G\), then \(N_{P}(K_{1})=N_{G}(K_{1})\cap P\neq P\) since \(T_{1}\subseteq N_{G}(K_{1})\). Therefore, there exists a maximal subgroup \(P_{2}\) of \(P\) such that \(N_{P}(K_{1})\leq P_{2}\). Let \(P_{n2}\) be an \(n\)-maximal subgroup of \(P\) contained in \(P_{2}\). Since \(P=P\cap G=P\cap P_{n1}N_{G}(K_{1})=P_{n1}(P\cap N_{G}(K_{1}))=P_{n1}N_{P}(K_{1})\), we have \(P_{n1}\neq P_{n2}\). By hypothesis, \(P_{n2}\) has a \(p\)-nilpotent supplement in \(G\). With the same discussion as above, we can find a Hall \(p'\)-subgroup \(K_{2}\) of \(G\) such that \(G=P_{n2}N_{G}(K_{2})=P_{2}N_{G}(K_{2})\). If \(p=2\), then by Lemma 2.10, \(K_{1}\) conjugates with \(K_{2}\) in \(G\). If \(p>2\), then \(G\) is soluble by Feit-Thompson Theorem. Hence, \(K_{1}\) also conjugates with \(K_{2}\) in \(G\). This means that there exists an element \(g\in P_{2}\), such that \((K_{2})^{g}=K_{1}\). Then \(G=(P_{2}N_{G}(K_{2}))^{g}=P_{2}N_{G}(K_{1})\). Hence, \(P=P\cap G=P\cap P_{2}N_{G}(K_{1})=P_{2}(P\cap N_{G}(K_{1}))=P_{2}N_{P}(K_{1})=P_{2}\). This contradiction shows that \(p^{n+1}\nmid |G|\). Thus \(G\) is \(p\)-nilpotent by Lemma 2.11. **Proof.** Suppose the Lemma is false and let \(G\) be a counterexample of minimal order. By Lemma 2.11, we have \(p^{n+1}\mid |G|\). Hence \(P\) has a non-trivial \(n\)-maximal subgroup. We proceed via the following steps: *(1) \(O_{p'}(G)=1\).* If \(O_{p'}(G)\neq 1\). Then we may choose a minimal normal subgroup \(N\) of \(G\) such that \(N\leq O_{p'}(G)\). Clearly, \((|G/N|,(p-1)(p^{2}-1)\cdot \cdot \cdot (p^{n}-1))=1\) and \(PN/N\) is a Sylow \(p\)-subgroup of \(G/N\). Assume that \(L/N\) is an \(n\)-maximal subgroup of \(PN/N\). Then, clearly, \(L/N=M_{p}N/N\), where \(M_{p}\) is an \(n\)-maximal subgroup of \(P\). By hypothesis, \(M_{p}\) either has a \(p\)-nilpotent supplement or is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). By Lemma 2.7(1) and Lemma 2.2(3), we see that \(G/N\) (with respect to \(PN/N\)) satisfies the hypothesis. The minimal choice of \(G\) implies that \(G/N\) is \(p\)-nilpotent and consequently \(G\) is \(p\)-nilpotent, a contradiction. *(2) \(P\) has a maximal subgroup \(P_{1}\) such that \(P_{1}\) has no \(p\)-nilpotent supplement in \(G\)* (This follows from Lemma 4.1). *(3) \(G\) is soluble.* Suppose that \(G\) is not soluble. Then \(p=2\) by the well known Feit-Thompson Theorem. Assume that \(O_2(G)\neq 1\). By Lemma 2.7 and Lemma 2.2(2), \(G/O_2(G)\) satisfies the hypothesis. Hence \(G/O_2(G)\) is \(2\)-nilpotent. It follow that \(G\) is soluble, a contradiction. Now assume that \(O_2(G)=1\). Then \((P_n)_G=1\), where \(P_n\) is an \(n\)-maximal subgroup of \(P\). Since \(P_{n}\) has no \(p\)-nilpotent supplement in \(G\), \(P_n\) is \(\mathfrak{U}_{\mathrm s}\)-quasinormal in \(G\) by the hypothesis. Hence there exists \(K\unlhd G\) such that \(P_nK\) is \(s\)-permutable in \(G\) and \(P_n\cap K\leq Z^\mathfrak{U}_\infty (G)\). If \(K=1\), then \(P_n\leq O_2(G)=1\) by Lemma 2.3(1), a contradiction. Thus, \(K\neq 1\). If \(Z^\mathfrak{U}_\infty (G)\neq 1\), then there exists a minimal normal subgroup \(H\) of \(G\) contained in \(Z^\mathfrak{U}_\infty (G)\). Hence \(H\) is of prime power order. This is impossible since \(O_{2'}(G)=1\) and \(O_2(G)=1\). Hence \(P_n\cap K=1\) and so \(2^{n+1}\nmid |K|\). Then by Lemma 2.11, \(K\) has a normal Hall \(2'\)-subgroup \(T\). Since \(T\) char \(K\unlhd G\), \(T\unlhd G\). It follows from (1) that \(T=1\). Consequently, \(K\leq O_2(G)=1\), a contradiction again. Hence (3) holds. *(4) \(N=O_p(G)\) is the only minimal normal subgroup of \(G\) and \(G=[N]M\), where \(M\) is a maximal subgroup of \(G\) and \(M\) is \(p\)-nilpotent.* Let \(N\) be a minimal normal subgroup of \(G\). By (1) and (3), \(N\) is an elementary abelian \(p\)-group and \(N\leq O_p(G)\). By Lemma 2.7(1) and Lemma 2.2(2), \(G/N\) satisfies the hypothesis and so \(G/N\) is \(p\)-nilpotent. Since the class of all \(p\)-nilpotent groups is a saturated formation, \(N\) is the unique minimal normal subgroup of \(G\) and \(\Phi (G)=1\). Hence \(O_p(G)=N=C_{G}(N)\), and consequently \(G=[N]M\), where \(M\) is a \(p\)-nilpotent maximal subgroup of \(G\). Thus (4) holds. *(5) The final contradiction.* Let \(P_{n}\) be an \(n\)-maximal subgroup of \(P\) such that \(P_n\leq P_1\). Then \(P_{n}\) has also no \(p\)-nilpotent supplement in \(G\). Hence there exists a normal subgroup \(K\) of \(G\) such that \(P_{n}K\) is \(s\)-permutable in \(G\) and \((P_{n}\cap K)(P_{n})_{G}/(P_{n})_{G}\leq Z_\infty ^\frak{F}(G/(P_{n})_{G})\). We claim that \((P_{n})_{G}=1\). Indeed, if \((P_{n})_{G}\neq 1\), then by (2), \(O_{p}(G)=N=(P_{n})_{G}\). Hence \(G=NM=(P_{n})_{G}M=P_{n}M\), which contradicts (2). Therefore, \(P_{n}\cap K\leq Z_\infty ^\frak{U}(G)\). If \(K=1\), then \(P_{n}\) is \(s\)-permutable in \(G\), and so \(P_{n}\leq O_{p}(G)=N\) and \(O^{p}(G)\leq N_{G}(P_{n})\) by Lemma 2.3. Hence \(1\neq P_{n}\leq P_{n}^{G}=P_{n}^{O^{p}(G)P}=P_{n}^{P}=(P_{n}\cap N)^{P}\leq (P_{1}\cap N)^{P}=P_{1}\cap N \leq N\). On the other hand, obviously, \(N\leq P_{n}^{G}\). Thus \(N=P_{n}^{G}=P_{1}\cap N\). It follows that \(N\leq P_{1}\), and so \(G=NM=P_{1}M\). This means that \(P_{1}\) has a \(p\)-nilpotent supplement in \(G\). This contradiction shows that \(K\neq 1\). If \(P_{n}\cap K=1\), then \(p^{n+1}\nmid |K|\). By Lemma 2.11, \(K\) is \(p\)-nilpotent and \(K_{p'}\leq O_{p'}(G)=1\) by (1). Hence \(K=N=O_{p}(G)\). It follows from Lemma 2.3(1) that \(P_{n}K=K\) and so \(P_{n}\cap K\neq 1\), a contradiction. Hence \(P_{n}\cap K\neq 1\). This means that \(Z_\infty ^\frak{U}(G)\neq 1\) and so \(N\leq Z_\infty ^\frak{U}(G)\). Consequently, \(|N|=|O_p(G)|=p\). Therefore, \(G/N\cong G/C_{G}(N)\) is isomorphic with some subgroup of \(Aut(N)\) of order \(p-1\). Since \((|G|,(p-1)(p^{2}-1)\cdot \cdot \cdot (p^{n}-1))=1\), \(G/N=1\). Consequently, \(G=N\) is an elementary abelian \(p\)-group. The final contradiction completes the proof. **Proof.** The necessity is obvious. We only need to prove the sufficiency. Suppose it is false and let \(G\) be a counterexample of minimal order. By Lemma 2.7(2) and Lemma 2.2(4), every \(n\)-maximal subgroup of \(P\) either has a \(p\)-nilpotent supplement or is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(E\). Hence \(E\) is \(p\)-nilpotent by Lemma 4.2. Then, \(E\neq G\). Let \(T\) be a normal Hall \(p'\)-subgroup of \(E\). Clearly, \(T\unlhd G\). We proceed the proof via the following steps: *(1) \(T=1\), and so \(P=E\unlhd G\).* Suppose that \(T\neq 1\). Since \(T\) is a normal Hall \(p'\)-subgroup of \(E\) and \(E\unlhd G\), then \(T\unlhd G\). We show that \(G/T\) (with respect to \(E/T\)) satisfies the hypothesis. Indeed, \((G/T)/(E/T)\simeq G/E\) is \(p\)-nilpotent and \(E/T=PT/T\) is a \(p\)-group. Suppose that \(M_{n}/T\) is an \(n\)-maximal subgroup of \(PT/T\) and \(P_{n}=M_{n}\cap P\). Then \(P_{n}\) is an \(n\)-maximal subgroup of \(P\) and \(M_{n}=P_{n}T\). By the hypothesis, \(P_{n}\) either has a \(p\)-nilpotent supplement or is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). By Lemma 2.7(1) and Lemma 2.2(3), \(M_{n}/T=P_{n}T/T\) either has a \(p\)-nilpotent supplement or is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G/T\). The minimal choice of \(G\) implies that \(G/T\) is \(p\)-nilpotent. This implies that \(G\) is \(p\)-nilpotent. This contradiction shows \(T=1\). Hence \(P=E\unlhd G\). *(2) Let \(Q\) be a Sylow \(q\)-subgroup of \(G\), where \(q\) is a prime divisor of \(|G|\) with \(q\neq p\). Then \(PQ=P\times Q\).* By (1), \(P=E\unlhd G\), \(PQ\) is a subgroup of \(G\). By Lemma 2.7(2) and Lemma 2.2(4), every \(n\)-maximal subgroup of \(P\) either has a \(p\)-nilpotent supplement or is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(PQ\). By using Lemma 4.2, we have that \(PQ\) is \(p\)-nilpotent. Hence \(Q\unlhd PQ\) and thereby \(PQ=P\times Q\). *(3) The final contradiction.* From (2), we have \(O^p(G)\leq C_G(P)\). **This induces that \(E=P\leq Z_\infty(G)\). Therefore \(G\) is \(p\)-nilpotent.** The final contradiction completes the proof. **Proof.** The necessity is obvious. We only need to prove the sufficiency. Suppose that it is false and let \(G\) be a counterexample with \(|G||H|\) is minimal. Let \(P\) be an arbitrary given Sylow \(p\)-subgroup of \(F(H)\). Clearly, \(P\unlhd G\). We proceed the proof as follows. *(1) \(\Phi(G)\cap P=1\).* If not, then \(1\neq \Phi(G)\cap P \unlhd G\). Let \(R=\Phi(G)\cap P\). Clearly, \((G/R)/(H/R)\simeq G/H\in \mathfrak{F}\). By Gasch\(\ddot{u}\)tz theorem (see \[11, III, Theorem 3.5\]), we have that \(F(H/R)=F(H)/R\). Assume that \(P/R\) is a Sylow \(\textit{p}\)-subgroup of \(F(H/R)\) and \(P_{1}/R\) is a maximal subgroup of \(P/R\). Then \(P\) is a Sylow \(p\)-subgroup of \(F(G)\) and \(P_{1}\) is a maximal subgroup of \(P\). By Lemma 2.2(2) and the hypothesis, \(P_{1}/R\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G/R\). Now, let \(Q/R\) be a maximal subgroup of some Sylow \(\textit{q}\)-subgroup of \(F(H/R)=F(H)/R\), where \(q\neq p\). Then \(Q=Q_{1}R\), where \(Q_{1}\) is a maximal subgroup of the Sylow \(\textit{q}\)-subgroup of \(F(H)\). By hypothesis, \(Q_{1}\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). Hence \(Q/R=Q_{1}R/R\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G/R\) by Lemma 2.2(3). This shows that \((G/R, H/R)\) satisfies the hypothesis. The minimal choice of \((G,H)\) implies that \(G/R\) is \(p\)-nilpotent. It follows that \(G\) is \(p\)-nilpotent, a contradiction. Hence (1) holds. *(2) \(P=\langle x_{1}\rangle \times \langle x_{2}\rangle \times\) \(\cdot \cdot \cdot \times \langle x_{m}\rangle\), where every \(\langle x_{i}\rangle\) \((i\in \{1\cdot \cdot \cdot m\})\) is a normal subgroup of \(G\) with order \(p\).* By (1) and Lemma 2.5, \(P=R_{1}\times R_{2}\times\) \(\cdot \cdot \cdot \times R_{m}\), where \(R_{i} \ (i\in \{1\cdot \cdot \cdot m\})\) is a minimal normal subgroup of \(G\). We now prove that \(R_{i}\) is of order \(\textit{p}\), for \(i\in \{1\cdot \cdot \cdot m\}\). Assume that \(|R_{i}|>\textit{p}\), for some \(\textit{i}\). Without loss of generality, we let \(|R_{1}|>\textit{p}\) and \(R^{*}_{1}\) be a maximal subgroup of \(R_{1}\). Then, \(R^{*}_{1}\neq 1\) and \(R^{*}_{1}\times R_{2}\times\) \(\cdot \cdot \cdot \times R_{m}=P_{1}\) is a maximal subgroup of \(P\). Put \(T=R_{2}\times\) \(\cdot \cdot \cdot \times R_{m}\). Then, clearly, \((P_{1})_{G}=T\). By hypothesis, \(P_{1}\) is \(\mathfrak{U_{\mathrm s}}\)-quasinormal in \(G\). Hence by Lemma 2.2(1), there exists a normal subgroup \(N\) of \(G\) such that \((P_{1})_{G}\leq N\), \(P_{1}N\) is \(s\)-permutable in \(G\) and \(P_{1}/(P_{1})_{G} \cap N/(P_{1})_{G}\leq Z_\infty ^\frak{U} (G/(P_{1})_{G})\). Assume that \(P_{1}/(P_{1})_{G} \cap N/(P_{1})_{G}\neq 1\). Let \(Z_\infty ^\frak{U} (G/(P_{1})_{G})=V/(P_{1})_{G}=V/T\). Then \(P_{1}\cap N\leq V\) and \(P/T \cap V/T \unlhd G/T\). Since \(P\cap V\geq P_{1}\cap N\cap V\geq P_{1}\cap N> (P_{1})_{G}=T\), \(P/T \cap V/T\neq 1\). As \(P/T\simeq R_{1}\) and \(R_{1}\) is a minimal normal subgroup of \(G\), we have \(P/T\subseteq V/T\). This implies that \(|R_{1}|=|P/T|=p\). This contradiction shows that \(P_{1}\cap N=(P_{1})_{G}=T\). Consequently, \(P_{1}N=R^{*}_{1}TN=R^{*}_{1}N\) and \(R^{*}_{1}\cap N=1\). Since \(R_{1}\cap N\unlhd G\), \(R_{1}\cap N=1\) or \(R_{1}\cap N=R_{1}\). If \(R_1\cap N=R_1\), then \(R_1^*\subseteq R_1\subseteq N\), which contradicts \(R^{*}_{1}\cap N=1\). Hence \(R_{1}\cap N=1\). It follows that \(R^{*}_{1}=R^{*}_{1}(R_{1}\cap N)=R_{1}\cap R^{*}_{1}N\) is \(s\)-permutable in \(G\). Thus \(O^{p}(G)\leq N_{G}(R_{1}^{*})\) by Lemma 2.3(2). This induces that for every maximal subgroup \(R_{1}^{*}\) of \(R_{1}\), we have that \(|G:N_{G}(R_{1}^{*})|=p^{\alpha}\), where \(\alpha\) is an integer. Let \(\{ R_{1}^{*},R_{2}^{*},\cdot \cdot \cdot ,R_{t}^{*}\}\) be the set of all maximal subgroups of \(R_{1}\). Then \(p\) divides \(t\). This contradicts to \[11, III, Theorem 8.5(d)\]. Thus (2) holds. *(3) \(G/F(H)\) is \(p\)-nilpotent.* By (2), \(F(H)=\langle y_{1}\rangle \times \langle y_{2}\rangle \times\) \(\cdot \cdot \cdot \times \langle y_{n}\rangle\), where \(\langle y_{i}\rangle \ (i\in \{1\cdot \cdot \cdot n\})\) is a normal subgroup of \(G\) of order \(p\). Since \(G/C_{G}(\langle y_{i}\rangle)\) is isomorphic with some subgroup of \(Aut(\langle y_{i}\rangle)\), \(G/C_{G}(\langle y_{i}\rangle)\) is cyclic. Hence, \(G/C_{G}(\langle y_{i}\rangle)\) is \(p\)-nilpotent for every \(i\). It follows that \(G/\cap^{n}_{i=1}C_{G}(\langle y_{i}\rangle)\) is \(p\)-nilpotent. Obviously, \(C_{G}(F(G))=\cap^{n}_{i=1}C_{G}(\langle y_{i}\rangle)\). Hence \(G/C_{G}(F(G))\) is \(p\)-nilpotent. Consequently, \(G/(H\cap C_{G}(F(G)))=G/C_{H}(F(H))\) is \(p\)-nilpotent. Since \(F(H)\) is abelian, \(F(H)\leq C_{H}(F(H))\). On the other hand, \(C_{H}(F(H))\leq F(H)\) since \(H\) is soluble. Thus \(F(H)=C_{H}(F(H))\) and so \(G/F(H)\) is \(p\)-nilpotent. *(4) If \(K\) is a minimal normal subgroup of \(G\) contained in \(H\), then \(K\subseteq F(H)\) and \(G/K\) is \(p\)-nilpotent.* Let \(K\) be an arbitrary minimal normal subgroup of \(G\) contained in \(H\). Then \(K\) is an elementary abelian \(\textit{p}\)-group for some prime \(\textit{p}\) since \(H\) is soluble. Hence \(K\leq F(H)\). By Lemma 2.2(2) and (3), we see that \(G/K\) (with respect to \(H/K\)) satisfies the hypothesis. The minimal choice of \((G,H)\) implies that \(G/K\) is \(p\)-nilpotent. *(5) The final contradiction.* Since the class of all \(p\)-nilpotent groups is a saturated formation, by (2) and (4), we see that \(K=F(H)=\langle x\rangle\) is the unique minimal normal subgroup of \(G\) contained in \(H\), where \(\langle x\rangle\) is a cyclic group of order \(p\) for some prime \(p\). Since \(G/K\) is \(p\)-nilpotent, it has a normal \(p\)-complement \(L/K\). By Schur-Zassenhaus Theorem, \(L=G_{p'}K\), where \(G_{p'}\) is a Hall \(p'\)-subgroup of \(G\). Since \(p\) is the prime divisor of \(|G|\) with \((|G|,p-1)=1\) and \(N_{L}(K)/C_{L}(K)\simeq Aut(K)\) is a subgroup of a cyclic group of order \(p-1\), we see that \(N_{L}(K)=C_{L}(K)\). Then, by Burnside Theorem (see \[14, (10.1.8)\]), we have that \(L\) is \(p\)-nilpotent. Then \(G_{p'}\) \(\emph{char}\) \(L\unlhd G\), so \(G_{p'}\unlhd G\). Hence \(G\) is \(p\)-nilpotent. The final contradiction completes the proof.   The authors are very grateful to the helpful suggestions of the referee.

{'timestamp': '2014-01-23T02:04:32', 'yymm': '1401', 'arxiv_id': '1401.5560', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5560'}

# Introduction We study the boundedness of the vertical square function \[Vf(x) = \left( \int_{0}^{\infty} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \right)^{1/2}.\] Here the linear operators \(\theta_{t}\), \(t > 0\), have the form \[\theta_{t}f(x) = \int_{\mathbb{R}^{n}} s_{t}(x,y)f(y) \, d\mu(y). \label{Intro:theta}\] The appearing measure \(\mu\) is a Borel measure in \(\mathbb{R}^n\) which is only assumed to satisfy, for some \(m\), the upper bound \[\label{powerBound} \mu(B(x,r)) \lesssim r^{m}, \qquad x \in \mathbb{R}^{n}, \; r > 0.\] Moreover, for some \(\alpha > 0\), the kernels \(s_{t}\) satisfy the size and continuity conditions \[\label{eq:size} |s_t(x,y)| \lesssim \frac{t^{\alpha}}{(t+|x-y|)^{m+\alpha}}\] and \[\label{eq:yhol} |s_t(x,y)-s_t(x,z)| \lesssim \frac{|y-z|^{\alpha}}{(t+|x-y|)^{m+\alpha}}\] whenever \(|y-z| < t/2\). The following is our main theorem. Hofmann proved the \(L^2\) boundedness of the square function under these local \(L^q\) testing conditions in the case that \(\mu\) is the Lebesgue measure. In the non-homogeneous case Lacey and the first named author proved the \(L^2\) boundedness but only with local \(L^2\) testing conditions. Our main theorem is an extension of these two state of the art results. Indeed, we consider general measures and general exponents simultaneously. The aforementioned two references are the most obvious predecessors of our main theorem, but the whole story up to this point is rather long. One can consider \(Tb\) theorems at least for square functions and Calderón--Zygmund operators. Then they can be global or local. And if they are local, they can be with the easier \(L^{\infty}/\textup{BMO}/T^{2,\infty}\) type testing assumptions, or with the more general \(L^s\), \(s < \infty\), type assumptions. Moreover, in the latter case the range of the exponents (in the Calderón--Zygmund world more than one set of testing functions appear) one can use is a very significant problem. Lastly, the fact that whether one considers the homogeneous or non-homogeneous theory is a major factor. All of these big story arcs are relevant for the context of the current paper. We now try to give at least some of the key references of local \(Tb\) theorems. The first local \(Tb\) theorem, with \(L^{\infty}\) control of the test functions and their images, is by Christ. Nazarov, Treil and Volberg proved a non-homogeneous version of this theorem. The point compared to the global \(Tb\) theorems is as follows. The accretivity of a given test function \(b_Q\) is only assumed on its supporting cube \(Q\), i.e., \(|\int_Q b_Q\,d\mu| \gtrsim \mu(Q)\). While in a global \(Tb\) one needs a function which is simultaneously accretive on all scales. But the remaining conditions are still completely scale invariant: \(b_Q \in L^{\infty}(\mu)\) and \(Tb_Q \in L^{\infty}(\mu)\). This scale invariance of the testing conditions is the main thing one wants to get rid of. The non-scale-invariant \(L^s\) type testing conditions were introduced by Auscher, Hofmann, Muscalu, Tao and Thiele. Their theorem is for perfect dyadic singular integral operators and the assumptions are of the form \(\int_Q |b^1_Q|^p \lesssim |Q|\), \(\int_Q |b^2_Q|^q \lesssim |Q|\), \(\int_Q |Tb^1_Q|^{q'} \lesssim |Q|\) and \(\int_Q |T^*b^2_Q|^{p'} \lesssim |Q|\), \(1 < p, q \le \infty\). Extending the result to general Calderón--Zygmund operators is complicated (it is almost done by now--but not completely). Hofmann established the result for general operators but only assuming the existence of \(L^{2+\epsilon}\) test functions mapping to \(L^2\). Auscher and Yang removed the \(\epsilon\) by proving the theorem in the sub-dual case \(1/p + 1/q \le 1\). Auscher and Routin considered the general case under some additional assumptions. The full super-dual case \(1/p + 1/q > 1\) is by Hytönen and Nazarov, but even then with the additional buffer assumption \(\int_{2Q} |Tb^1_Q|^{q'} \lesssim |Q|\) and \(\int_{2Q} |T^*b^2_Q|^{p'} \lesssim |Q|\). This was the main story for the Calderón--Zygmund operators for doubling measures. For square functions the situation is a bit more clear with the need for only one exponent \(q\). The case \(q = 2\) is implicit in the Kato square root papers, , and explicitly stated and proved in and. The case \(q > 2\) is weaker than this. The hardest case \(q \in (1,2)\) is due to Hofmann as already mentioned. Some key applications really need the fact that one can push the integrability of the test functions to \(1+\epsilon\) (see again ). The non-homogeneous world is yet another story. The whole usage of these non-scale-invariant testing conditions is a huge source of problem in this context. One reason lies in the fact that even if we have performed a stopping time argument which gives us that a fixed test function \(b_F\) behaves nicely on a cube \(Q\), for example that \(\int_Q |b_F|^2\, d\mu \lesssim \mu(Q)\), we cannot say much what happens in the stopping children of \(Q\). That is, in a stopping child \(Q'\) of \(Q\) we cannot use the simple argument \[\int_{Q'} |b_F|^2 \,d\mu \le \int_{Q} |b_F|^2 \,d\mu \lesssim \mu(Q) \lesssim \mu(Q')\] which would only be available if \(\mu\) would be doubling. The non-homogeneous case \(q = 2\) for square functions is the very recent work of Lacey and the first named author. The case \(p=q=2\) for Calderón--Zygmund operators is by the same authors. For relevant dyadic techniques see also the Lacey--Vähäkangas papers and, and Hytönen--Martikainen. To recap the context, in this paper we consider non-homogeneous square functions and push \(q\) to the range \(q \in (1,2)\). We still mention that the study of the boundedness of non-homogeneous square functions was initiated by the recent authors in. This was a global \(Tb\). The key technique was the usage of good (in a probabilistic sense) Whitney regions. A scale invariant local \(Tb\) is by the current authors together with T. Orponen. In that paper we also study the end point theory, \(L^p\) theory, and various counter-examples (e.g. the failure of the change of aperture with general measures and the difference between conical and vertical square functions). We conclude the introduction by a remark and setting up some notation. ## Notation We write \(A \lesssim B\), if there is a constant \(C>0\) so that \(A \leq C B\). We may also write \(A \approx B\) if \(B \lesssim A \lesssim B\). For a number \(a\) we write \(a \sim 2^k\) if \(2^k \leq a <2^{k+1}\). We then set some dyadic notation. Consider a dyadic grid \(\D\) in \(\mathbb{R}^n\). For \(Q, R \in \D\) we use the following notation: - \(\ell(Q)\) is the side-length of \(Q\); - \(d(Q,R)\) denotes the distance between the cubes \(Q\) and \(R\); - \(D(Q,R):=d(Q,R)+\ell(Q)+\ell(R)\) is the long distance; - \(\widehat{Q}=Q\times (0, \ell(Q))\) is the Carleson box associated with \(Q\); - \(W_Q=Q\times (\ell(Q)/2, \ell(Q))\) is the Whitney region associated with \(Q\); - \(\text{ch}(Q)=\{Q' \in \D: Q' \subset Q, \ell(Q') = \ell(Q)/2\}\); - \(\text{gen}(Q)\) is determined by \(\ell(Q)=2^{\text{gen}(Q)}\); - \(Q^{(k)} \in\D\) is the unique cube for which \(\ell(Q^{(k)})=2^k \ell(Q)\) and \(Q \subset Q^{(k)}\); - \(\langle f \rangle_Q = \mu(Q)^{-1}\int_Q f\,d\mu\). # Structure of the proof and basic reductions ## Reduction to a priori bounded operators \(V\) {#reduction.apriori.bounded} In this subsection we say the following. Suppose we have proved the \(L^q(\mu)\) bound of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}, i.e., the quantitative bound \(\|V\|_{L^q(\mu) \to L^q(\mu)} \lesssim 1+V_{\textup{loc}, q}\), under the additional a priori finiteness assumption \(\|V\|_{L^q(\mu) \to L^q(\mu)} < \infty\). Then the \(L^q(\mu)\) bound of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} automatically follows without the a priori assumption. To this end, define \(s^i_t(x,y) = s_t(x,y)\) if \(1/i \le t \le i\), and \(s_t^i(x,y) = 0\) otherwise. These kernels are clearly in our original class--they satisfy [\[eq:size\]](#eq:size){reference-type="eqref" reference="eq:size"} and [\[eq:yhol\]](#eq:yhol){reference-type="eqref" reference="eq:yhol"} with kernel constants bounded by those of \(V\). Define \[V_if(x) := \left( \int_{1/i}^{i} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \right)^{1/2} = \left( \int_{0}^{\infty} |\theta_{t}^if(x)|^{2} \, \frac{dt}{t} \right)^{1/2},\] where \[\theta_{t}^if(x) = \int_{\mathbb{R}^{n}} s_{t}^i(x,y)f(y) \, d\mu(y).\] Let us note that the \(V_i\) are bounded operators on \(L^q(\mu)\). Let \[M_{\mu}f(x) = \sup_{r>0} \frac{1}{\mu(B(x,r))} \int_{B(x,r)} |f|\,d\mu(y).\] This centred maximal function is a bounded operator on \(L^p(\mu)\) for every \(p \in (1,\infty)\). Notice that \(|\theta_t f(x)| \lesssim M_{\mu} f(x)\) for every \(t > 0\) and \(x \in \mathbb{R}^n\). Using this we see that \(\|V_i\|_{L^q(\mu) \to L^q(\mu)} \le [2 \log i]^{1/2} \|M_{\mu}\|_{L^q(\mu) \to L^q(\mu)} < \infty\). By monotone convergence we have that \[\begin{aligned} \|Vf\|_{L^q(\mu)} &= \lim_{i \to \infty} \|V_if\|_{L^q(\mu)} \\ & \le \limsup_{i \to \infty} \|V_i\|_{L^q(\mu) \to L^q(\mu)} \|f\|_{L^q(\mu)} \\ &\lesssim \limsup_{i \to \infty} (1+V_{\textup{loc}, q}^i) \|f\|_{L^q(\mu)} \\ &\le (1 + V_{\textup{loc}, q}) \|f\|_{L^q(\mu)}. \end{aligned}\] So it suffices to prove Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} under the assumption \(\|V\|_{L^q(\mu) \to L^q(\mu)} < \infty\)--a piece of information that will be used purely in a qualitative way. ## Reduction to a q-Carleson estimate We begin by stating a \(T1\) in \(L^q(\mu)\) (the case \(q = 2\) is in ). The proof of this \(T1\) is indicated in Appendix [\[App:T1\]](#App:T1){reference-type="ref" reference="App:T1"}. Define, say for \(\lambda \geq 3\), \[\car_V(q, \lambda) := \mathop{\sup_{Q \subset \mathbb{R}^n}}_{\textup{cube}} \Big[ \frac{1}{\mu(\lambda Q)} \int_{Q} \Big( \int_0^{\ell(Q)} |\theta_t 1_{Q}(x)|^2 \frac{dt}{t} \Big)^{q/2} \,d\mu(x)\Big]^{1/q}\] and \[\mcar_V(q, \lambda) := \mathop{\sup_{Q \subset \mathbb{R}^n}}_{\textup{cube}} \Big[ \frac{1}{\mu(\lambda Q)} \int_{Q} \Big( \int_0^{\ell(Q)} |\theta_t 1(x)|^2 \frac{dt}{t} \Big)^{q/2} \,d\mu(x)\Big]^{1/q}.\] Then, for \(q \in (1,2]\), we have that there holds that \[\label{eq:Intro.qt1} \|V\|_{L^q(\mu) \to L^q(\mu)} \le C_1(1+ \mcar_V(q,9)) \le C_2(1 + \car_V(q,3)).\] Assuming the existence of the \(L^q\) test functions as in Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} we then prove that \[\label{Intro:Carleson} \car_V(q,3) \le C_3 (1 +V_{\textup{loc}, q}) + C_2^{-1} \|V\|_{L^q(\mu) \to L^q(\mu)}/2.\] We call this the key inequality. Combining [\[eq:Intro.qt1\]](#eq:Intro.qt1){reference-type="eqref" reference="eq:Intro.qt1"} and [\[Intro:Carleson\]](#Intro:Carleson){reference-type="eqref" reference="Intro:Carleson"} gives that \[\|V\|_{L^q(\mu) \to L^q(\mu)} \le C (1 +V_{\textup{loc}, q}) + \|V\|_{L^q(\mu) \to L^q(\mu)}/2\] ending the proof. We will now start the proof of the key inequality [\[Intro:Carleson\]](#Intro:Carleson){reference-type="eqref" reference="Intro:Carleson"}. This task is completed in Section [8](#section:nested){reference-type="ref" reference="section:nested"}. In Appendix [\[App:T1\]](#App:T1){reference-type="ref" reference="App:T1"} we indicate the proof of the \(T1\) theorem in \(L^q(\mu)\), i.e., the first estimate of [\[eq:Intro.qt1\]](#eq:Intro.qt1){reference-type="eqref" reference="eq:Intro.qt1"}. # Random and stopping cubes/ Martingale difference operators ## Random dyadic grids At this point we need to set up the basic notation for random dyadic grids (these facts are essentially presented in this way by Hytönen ). Let \(\D_0\) denote the standard dyadic grid, consisting of all the cubes of the form \(2^k(\ell + [0,1)^n)\), where \(k \in \mathbb{Z}\) and \(\ell \in \mathbb{Z}^n\). We also denote \(\D_{0, k}=\{ Q \in \D_0: \ell(Q)=2^k \}\). A generic dyadic grid, parametrized by \(w \in (\{0,1\}^n)^\mathbb{Z}\), is of the form \(\D(w)= \cup_{k \in \mathbb{Z}} \D_k(w)\), where \(\D_k(w)=\{Q_0+x_k(w): Q_0 \in \D_{0, k}\}\) and \(x_k(w)=\sum_{j<k} w_j 2^j\). The notation \(Q_0 + w := Q_0 + \sum_{j<k} w_j 2^j\), \(Q_0 \in \mathcal{D}_{0,k}\), is convenient. We get random dyadic grids by placing the natural product probability measure \(P_w\) on \((\{0,1\}^n)^\mathbb{Z}\) (thus the coordinate functions \(w_j\) are independent and \(P_w(w_j=\eta)=2^{-n}\) if \(\eta \in \{0,1\}^n\)). We fix the constant \(\gamma \in (0,1)\) to be so small that \[\gamma \leq \alpha/(2m+2\alpha) \quad \text{and} \quad m \gamma/(1-\gamma) \leq \alpha/4,\] where \(\alpha > 0\) appears in the kernel estimates and \(m\) appears in \(\mu(B(x,r)) \lesssim r^m\). A cube \(R \in \mathcal{D}\) is called \(\D\)-bad if there exists another cube \(Q \in \mathcal{D}\) so that \(\ell(Q) \ge 2^r \ell(R)\) and \(d(R, \partial Q) \le \ell(R)^{\gamma}\ell(Q)^{1-\gamma}\). Otherwise it is good. We denote the collections of good and bad cubes by \(\D_{\textrm{good}}\) and \(\D_{\textrm{bad}}\) respectively. The following properties are known (see e.g. ). - For a fixed \(Q_0 \in \mathcal{D}_0\) the set \(Q_0 + w\) depends on \(w_j\) with \(2^{j} < \ell(Q_0)\), while the goodness (or badness) of \(Q_0 + w\) depends on \(w_j\) with \(2^{j} \ge \ell(Q_0)\). In particular, these notions are independent (meaning that for any fixed \(Q_0 \in \mathcal{D}_0\) the random variable \(w \mapsto 1_{\textup{good}}(Q_0+w)\) and any random variable that depends only on the cube \(Q_0+w\) as a set, like \(w \mapsto \int_{Q_0+w} f\,d\mu\), are independent). - The probability \(\pi_{\textrm{good}} := P_{w}(Q _0+ w \textrm{ is good})\) is independent of \(Q_0 \in \mathcal{D}_0\). - \(\pi_{\textrm{bad}} :=1-\pi_{\textrm{good}} \lesssim 2^{-r \gamma}\), with the implicit constant independent of \(r\). The parameter \(r \lesssim 1\) is a fixed constant which is at least so large that \(2^{r(1-\gamma)} \ge 10\). The following lemma is stated without proof since the first part was proved on page 25 of and the second is lemma 2.10 of. ## Collections of stopping cubes Let \(\mathcal{D}\) be a dyadic grid in \(\mathbb{R}^n\) and let \(Q^* \in \mathcal{D}\) be a fixed dyadic cube with \(\ell(Q^*) = 2^s\). Set \(\mathcal{F}_{Q^*}^0 = \{Q^*\}\) and let \(\mathcal{F}_{Q^*}^1\) consist of the maximal cubes \(Q\in \D\), \(Q \subset Q^*\), for which at least one of the following two conditions holds: 1. \(| \langle b_{Q^*}\rangle_Q | < 1/2\); 2. \(\langle |b_{Q^*}|^q\rangle_Q > 2^{q'+1}A^{q'}\). Here \(A\) is a constant such that \(\|b_R\|_{L^q(\mu)}^q \le A\mu(R)\) for every cube \(R \subset \mathbb{R}^n\). Next, we repeat the previous procedure by replacing \(Q^*\) with a fixed \(Q \in \mathcal{F}_{Q^*}^1\). The combined collection of stopping cubes resulting from this is called \(\mathcal{F}_{Q^*}^2\). This is continued and we set \(\mathcal{F}_{Q^*}=\bigcup_{j=0}^\infty \mathcal{F}_{Q^*}^j\) Finally, for every \(Q \in \D\), \(Q \subset Q^*\), we let \(Q^\alpha \in \mathcal{F}_{Q^*}\) be the minimal cube \(R \in \mathcal{F}_{Q^*}\) for which \(Q \subset R\). The next lemma follows. We now state the classical Carleson embedding theorem. The next proposition is a Carleson embedding on \(L^p(\mu)\), where the Carleson condition itself depends on \(p\). This kind of Carleson is also well-known, of course, but we state and prove this general version here for the convenience of the reader. ## Twisted martingale difference operators and square function estimates If \(Q \in \D\), \(Q\subset Q^*\), and \(f \in L^1_{\textup{loc}}(\mu)\), we define the twisted martingale difference operators \[\Delta_Q f = \sum_{Q' \in \, \textrm{ch}(Q)} \Big[\frac{\langle f \rangle_{Q'}}{\langle b_{(Q')^a}\rangle_{Q'}}b_{(Q')^a}-\frac{\langle f \rangle_Q}{\langle b_{Q^a}\rangle_Q}b_{Q^a}\Big]1_{Q'}.\] Note that on the largest \(Q^*\) level we agree (by abuse of notation) that \(\Delta_{Q^*} = E^b_{Q^*} + \Delta_{Q^*}\), where \(E^b_{Q^*}f = \langle f \rangle_{Q^*} b_{Q^*}\). Therefore, we have that \(\int \Delta_Q f \,d\mu = 0\) if \(Q \subsetneq Q^*\). We also define \[\Delta_k f= \Delta^{Q^*}_k f := \sum_{Q \in \D_k: Q \subset Q^*} \Delta_Q f.\] Notice that if \(\ell(Q^*)=2^s\), then \(k \leq s\), that is, only cubes inside the fixed \(Q^*\) are considered. We now state some lemmata which contain the square function estimates we need in our proof. The first one was proved by Stein on page 103 of: The proof of the next lemma is quite hard. It was proved by Lacey and the first named author (but only stated in \(L^2(\mu)\)). But we will not need the full strength of this, since our function is bounded. Therefore, instead of using the next lemma, we will indicate a somewhat simpler proof in the \(|f| \le 1\) case, which is the only thing we will need. This is not that easy either but we include the key details for the convenience of the reader. But for us the following consequence is enough (and we will indicate the proof of this simpler statement): In the \(|f| \le 1\) case we can get rid of the assumption \(Q^a = F\) as follows: # Reductions towards the proof of the key inequality We will estimate the quantity \[\begin{aligned} \Big[ \int_{Q_0} \Big( \int_0^{\ell(Q_0)} |\theta_t f(x)|^2 \frac{dt}{t} \Big)^{q/2} \,d\mu(x)\Big]^{1/q} \end{aligned}\] for an arbitrary fixed cube \(Q_0 \subset \mathbb{R}^n\) and for an arbitrary fixed function \(f\) satisfying that \(|f| \le 1_{Q_0}\) (the choice \(f = 1_{Q_0}\) would suffice). Let \(s\) be defined by \(2^{s-1} \leq \ell(Q_0) < 2^{s}\). ## Reduction to a dyadic setting of good geometric data {#reduction.good.R} For a fixed \(w \in (\{0,1\}^n)^\mathbb{Z}\) and \(x \in Q_0\) we have that \[\begin{aligned} \int_{0}^{\ell(Q_0)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \leq \mathop{\sum_{R \in \mathcal{D}(w)}}_{\ell(R) \leq 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t}. \end{aligned}\] Recall the constants from [\[eq:Intro.qt1\]](#eq:Intro.qt1){reference-type="eqref" reference="eq:Intro.qt1"}. To prove [\[Intro:Carleson\]](#Intro:Carleson){reference-type="eqref" reference="Intro:Carleson"} we note that by above it is enough to prove that \[\label{eq:red.good1} E_w \Big[ \int_{\mathbb{R}^n} 1_{Q_0}(x)\Big( \mathop{\sum_{R \in \mathcal{D}(w)}}_{\ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{q/2}\,d\mu(x)\Big]^{1/q}\] can be bounded by \[\mu(3Q_0)^{1/q}.\] We can estimate the quantity in [\[eq:red.good1\]](#eq:red.good1){reference-type="eqref" reference="eq:red.good1"} by \[\begin{aligned} &E_w \Big[ \int_{\mathbb{R}^n} 1_{Q_0}(x)\Big( \mathop{\sum_{R \in \mathcal{D}(w)_{\textup{good}}}}_{\ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{q/2}\,d\mu(x)\Big]^{1/q} \\ &+ E_w \Big[ \int_{\mathbb{R}^n} 1_{Q_0}(x) \Big( \mathop{\sum_{R \in \mathcal{D}(w)_{\textup{bad}}}}_{\ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{q/2}\,d\mu(x)\Big]^{1/q}. \end{aligned}\] Using \(E g^{\alpha} \le (Eg)^{\alpha}\) for \(\alpha \in (0,1]\), we see (with \(\alpha = 1/q\) and \(\alpha = q/2\)) that \[\begin{aligned} &E_w \Big[ \int_{\mathbb{R}^n} 1_{Q_0}(x)\Big( \mathop{\sum_{R \in \mathcal{D}(w)_{\textup{bad}}}}_{\ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{q/2}\,d\mu(x)\Big]^{1/q} \\ &\le \Big[ \int_{\mathbb{R}^n} 1_{Q_0}(x) \Big( E_w \mathop{\sum_{R \in \mathcal{D}(w)_{\textup{bad}}}}_{\ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{q/2}\,d\mu(x)\Big]^{1/q}. \end{aligned}\] Using the fact that \(w \mapsto 1_{\textup{bad}}(R_0 + w)\) is independent of \(w \mapsto 1_{R_0 + w}(x)\) for every \(R_0 \in \mathcal{D}_0\), and that \(E_w 1_{\textup{bad}}(R_0 + w) \le c(r) \to 0\) when \(r \to \infty\), we have \[\begin{aligned} E_w \Big[ \int_{\mathbb{R}^n} 1_{Q_0}(x)& \Big( \mathop{\sum_{R \in \mathcal{D}(w)_{\textup{bad}}}}_{\ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{q/2}\,d\mu(x)\Big]^{1/q} \l \\ &\le c(r)^{1/2}\|Vf\|_{L^q(\mu)} \le (2C_2)^{-1} \|V\|_{L^q(\mu) \to L^q(\mu)} \mu(3Q_0)^{1/q} \end{aligned}\] fixing \(r \lesssim 1\) large enough (note that \(c(r) = C(n, \alpha, m)2^{-r\gamma}\)). We have reduced to showing that uniformly on \(w \in (\{0,1\}^n)^\mathbb{Z}\) the quantity \[\begin{aligned} \Big[ \int_{\mathbb{R}^n} 1_{Q_0}(x) \Big( \mathop{\sum_{R \in \mathcal{D}(w)_{\textup{good}}}}_{\ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{q/2}\,d\mu(x)\Big]^{1/q} \end{aligned}\] can be dominated by \(C_3(1 + V_{\textup{loc},q}) \mu(3Q_0)^{1/q}\). We fix one \(w\) and write \(\D = \D(w)\). ## Decomposition of \(f\) {#reduction.begin} Since \(f \in L^q(\mu)\) is supported in \(Q_0\) we may expand \[f = \mathop{\mathop{\sum_{Q^* \in \mathcal{D}}}_{\ell(Q^*) = 2^s}}_{Q_0 \cap Q^* \ne \emptyset} \mathop{\sum_{Q \in \mathcal{D}}}_{Q \subset Q^*} \Delta_Q f. \label{eq:md-decomp}\] Notice that there are only finitely many such \(Q^*\) and always \(Q^* \subset 3 Q_0\). Define \[A_\kappa f(x) := \Big(\mathop{\sum_{R \in \mathcal{D}_{\textup{good}}}}_{2^{-\kappa} < \ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{1/2}\] and \[A f(x) := \Big(\mathop{\sum_{R \in \mathcal{D}_{\textup{good}}}}_{\ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{1/2}.\] Notice that for \(x \in Q_0\) there holds that \[\begin{aligned} &\Big|Af(x)-A_\kappa\Big( \sum_{Q^*} \mathop{\sum_{Q \subset Q^*}}_{\ell(Q) > 2^{-\kappa}} \Delta_Q f\Big)(x)\Big| \\ &\le |Af(x)-A_{\kappa}f(x)| + \Big|A_{\kappa}f(x)-A_{\kappa}\Big( \sum_{Q^*} \mathop{\sum_{Q \subset Q^*}}_{\ell(Q) > 2^{-\kappa}} \Delta_Q f\Big)(x)\Big| \\ &\le \Big(\mathop{\sum_{R \in \mathcal{D}_{\textup{good}}}}_{\ell(R) \le 2^{-\kappa}} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{1/2} + A_{\kappa}\Big( f-\sum_{Q^*} \mathop{\sum_{Q \subset Q^*}}_{\ell(Q) > 2^{-\kappa}} \Delta_Q f\Big)(x) \\ &\le \Big( \int_0^{2^{-\kappa}} |\theta_{t}f(x)|^{2} \, \frac{dt}{t} \Big)^{1/2} + V\Big( f-\sum_{Q^*} \mathop{\sum_{Q \subset Q^*}}_{\ell(Q) > 2^{-\kappa}} \Delta_Q f\Big)(x). \end{aligned}\] It follows by dominated convergence and the fact that \(V\) is bounded on \(L^q(\mu)\) that \[\begin{aligned} \lim_{\kappa \to \infty} \Big\| 1_{Q_0}\Big( Af-A_{\kappa} \Big( \sum_{Q^*} \mathop{\sum_{Q \subset Q^*}}_{\ell(Q) > 2^{-\kappa}} \Delta_Q f\Big) \Big)\Big\|_{L^q(\mu)} = 0. \end{aligned}\] We have reduced to showing that \[\label{heart.estimate} \Big[ \int_{\mathbb{R}^n} 1_{Q_0}(x) \Big( \mathop{\sum_{R \in \mathcal{D}_{\textup{good}}}}_{2^{-\kappa} < \ell(R) \le 2^s} 1_R(x) \int_{\ell(R)/2}^{\ell(R)} \Big| \mathop{\mathop{\sum_{Q \in \mathcal{D}}}_{Q \subset Q^*}}_{\ell(Q) > 2^{-\kappa}} \theta_{t} \Delta_Q f(x)\Big|^{2} \, \frac{dt}{t}\Big)^{q/2}\,d\mu(x)\Big]^{1/q}\] can be dominated by \(C_3(1 + V_{\textup{loc},q}) \mu(Q^*)^{1/q}\) for every fixed \(\kappa\) and for every fixed \(Q^*\). We used the fact that \[\theta_t\Big( \sum_{Q^*} \mathop{\sum_{Q \subset Q^*}}_{\ell(Q) > 2^{-\kappa}} \Delta_Q f \Big) = \sum_{Q^*} \mathop{\sum_{Q \subset Q^*}}_{\ell(Q) > 2^{-\kappa}} \theta_t \Delta_Q f,\] since the sum is finite for every \(\kappa\). To fix only one \(Q^* \subset 3Q_0\) we used the fact that \(\#\{Q^* \in \mathcal{D}:\, \ell(Q^*) = 2^s \textup{ and } Q^* \cap Q_0 \ne \emptyset\} \lesssim 1\). ## Splitting the summation We will split the sum [\[heart.estimate\]](#heart.estimate){reference-type="eqref" reference="heart.estimate"} in to the following four pieces: - \(\ell(Q)< \ell(R)\); - \(\ell(Q)\geq \ell(R)\) and \(d(Q,R)> \ell(R)^\gamma \ell(Q)^{1-\gamma}\); - \(\ell(R) \leq \ell(Q) \leq 2^r \ell(R)\) and \(d(Q,R) \leq \ell(R)^\gamma \ell(Q)^{1-\gamma}\); - \(\ell(Q)>2^r\ell(R)\) and \(d(Q,R) \leq \ell(R)^\gamma \ell(Q)^{1-\gamma}\). We call the second sum the separated sum, the third sum the diagonal sum and the last sum the nested sum. Thus, [\[heart.estimate\]](#heart.estimate){reference-type="eqref" reference="heart.estimate"} is bounded by \[I_{\ell(Q)<\ell(R)} + I_{\textup{sep}} + I_{\textup{diag}} + I_{\textup{nested}}.\] We bound these four pieces in the four subsequent chapters. # The case \(\ell(Q)< \ell(R)\) {#section:Q<R} We start by proving the following lemma. Let \(Q \in \D\) and \(R \in \D_\textup{good}\) be such that \(\ell(R)/\ell(Q)=2^\ell\) and \(D(Q,R)/\ell(R) \sim 2^j\) for \(\ell \ge 1\) and \(j \ge 0\). Assume also that \((x,t) \in W_R\). Since \(\ell(Q) < \ell(R) \le 2^s\), we have \(\int \Delta_Q f\,d\mu= 0\). Using this we write \[|\theta_t \Delta_Q f (x)|=\Big| \int_Q [s_t(x,y)-s_t(x,c_Q)] \Delta_Q f(y)\, d\mu(y)\Big|.\] Using the estimate [\[kernel.est.Q\<R\]](#kernel.est.Q<R){reference-type="eqref" reference="kernel.est.Q<R"} we now see that \[|\theta_t \Delta_Q f (x)| \lesssim 2^{-\alpha \ell} 2^{-3\alpha j/4} \ell(S_0)^{-m} \int_Q |\Delta_Q f(y)| d\mu(y),\] where \(Q, R \subset S_0 :=Q^{(\ell+j+\theta(j))}\) (by (1) of Lemma [\[lemma.R\<S(Q)\]](#lemma.R<S(Q)){reference-type="ref" reference="lemma.R<S(Q)"}). We can now see that \(I_{\ell(Q) < \ell(R)}\) can be dominated by \[\sum_{j, \ell}2^{-\frac{\alpha}{2}(\ell+\frac{3}{4}j)} \Big\| \Big(\sum_{k \le s} \sum_{R \in \mathcal{D}_{k, \textup{good}}}1_{R} \Big( \mathop{\sum_{Q \in \mathcal{D}_{k-\ell}:\, Q \subset Q^*}}_{D(Q,R)/\ell(R)\sim 2^j } \ell(S_0)^{-m} \int |\Delta_Q f| d\mu \Big)^2\Big)^{1/2} \Big\|_{L^q(\mu)}.\] Let us fix \(j, \ell, k\). Set \(\tau_j(k):= j + \theta(j) + k = \textup{gen}(S_0)\). We have by disjointness considerations and the fact that \(Q, R \subset S_0\) that \[\begin{aligned} \sum_{R \in \mathcal{D}_{k, \textup{good}}} &1_{R} \Big( \mathop{\sum_{Q \in \mathcal{D}_{k-\ell}:\, Q \subset Q^*}}_{D(Q,R)/\ell(R)\sim 2^j } \ell(S_0)^{-m} \int |\Delta_Q f|\, d\mu \Big)^2\\ &=\Big( \sum_{R \in \mathcal{D}_{k, \textup{good}}}1_{R} \mathop{\sum_{Q \in \mathcal{D}_{k-\ell}:\, Q \subset Q^*}}_{D(Q,R)/\ell(R)\sim 2^j } 2^{-m \tau_j(k)} \int |\Delta_Q f|\, d\mu \Big)^2\\ &=\Big( \sum_{S \in \D_{\tau_j(k)}}\mathop{\sum_{R \in \mathcal{D}_{k, \textup{good}}}}_{R\subset S}1_{R} \mathop{\sum_{Q \in \mathcal{D}_{k-\ell}:\, Q \subset Q^*}}_{D(Q,R)/\ell(R)\sim 2^j } 2^{-m \tau_j(k)}\int |\Delta_Q f|\, d\mu \Big)^2 \\ &\lesssim \Big( \sum_{S \in \D_{\tau_j(k)}} \frac{1_S}{\mu(S)} \int_S |\Delta_{k-\ell} f| d\mu \Big)^2 \\ &= [E_{\tau_j(k)}(|\Delta_{k-\ell} f|)]^2. \end{aligned}\] Note that for fixed \(j, \ell\) there holds by Stein's inequality (Lemma [\[lemma-Stein\]](#lemma-Stein){reference-type="ref" reference="lemma-Stein"}) and estimate [\[eq:MD-square.function.estimate\]](#eq:MD-square.function.estimate){reference-type="eqref" reference="eq:MD-square.function.estimate"} that \[\begin{aligned} \Big\| \Big(\sum_{k \le s} [E_{\tau_j(k)}(|\Delta_{k-\ell} f|)]^2 \Big)^{1/2}\Big\|_{L^q(\mu)} &\lesssim \Big\| \Big(\sum_{k \le s} |\Delta_k f|^2 \Big)^{1/2}\Big\|_{L^q(\mu)} \lesssim \mu(Q^*)^{1/q}. \end{aligned}\] We may now conclude that \(I_{\ell(Q) < \ell(R)} \lesssim \mu(Q^*)^{1/q}\). # The separated sum {#section:separated} We first prove the following lemma. Let \(Q \in \D\), \(R \in \D_{\textup{good}}\) be such that \(d(Q,R) > \ell(R)^{\gamma}\ell(Q)^{1-\gamma}\), \(\ell(Q)/\ell(R)=2^\ell\) and \(D(Q,R)/\ell(Q) \sim 2^j\) for \(\ell, j \geq 0\). If \((x,t) \in W_R\) we have by [\[kernel.est.separ\]](#kernel.est.separ){reference-type="eqref" reference="kernel.est.separ"} that \[|\theta_t \Delta_Q f (x)|\lesssim 2^{-(\ell+j)\alpha/4}\ell(S_0)^{-m} \int |\Delta_Q f(y)|\,d\mu(y),\] where \(Q, R \subset S_0 :=Q^{(j+\theta(j+\ell))}\) (by (2) of Lemma [\[lemma.R\<S(Q)\]](#lemma.R<S(Q)){reference-type="ref" reference="lemma.R<S(Q)"}). We may deduce that \(I_{\textup{sep}}\) can be dominated by \[\sum_{j, \ell}2^{-\alpha(\ell+j)/4} \Big\| \Big(\sum_{k \le s} \sum_{R \in \mathcal{D}_{k, \textup{good}}} 1_{R} \Big( \mathop{\sum_{Q \in \mathcal{D}_{k+\ell}: Q \subset Q^*}}_{D(Q,R)\sim 2^j \ell(Q)} \ell(S_0)^{-m}\int |\Delta_Q f|\,d\mu \Big)^2\Big)^{1/2} \Big\|_{L^q(\mu)}.\] A completely analogous estimate to that of the previous section shows that \(I_{\textup{sep}} \lesssim \mu(Q^*)^{1/q}\). # The diagonal sum {#section:diagonal} Let \(Q \in \D\), \(R \in \D_{\textup{good}}\) be such that \(\ell(Q) / \ell(R) = 2^{\ell}\) and \(D(Q,R) / \ell(Q) \sim 2^j\). Since we are in the diagonal summation \(I_{\textup{diag}}\) we have that \(\ell, j \lesssim 1\). If \((x,t) \in W_R\) we have that \[|s_t(x,y)| \lesssim t^{-m} \approx \ell(R)^{-m} \approx \ell(S_0)^{-m},\] where \(Q, R \subset S_0 :=Q^{(j+\theta(j+\ell))}\) (by (2) of Lemma [\[lemma.R\<S(Q)\]](#lemma.R<S(Q)){reference-type="ref" reference="lemma.R<S(Q)"}). It is now clear by the previous arguments that \(I_{\textup{diag}} \lesssim \mu(Q^*)^{1/q}\). # The nested sum {#section:nested} In this case one uses the goodness of \(R\) to conclude that one must actually have that \(R \subset Q\). Therefore, things reduce to proving that \[\label{Inested} \Big\| 1_{Q_0} \Big( \mathop{\sum_{R \in \mathcal{D}_{\textup{good}}:\, R \subset Q^*}}_{2^{-\kappa} < \ell(R) < 2^{s-r}} 1_{R} \int_{\ell(R)/2}^{\ell(R)} \Big| \mathop{\sum_{\ell=r+1}^{s-\textup{gen(R)}}} \theta_t \Delta_{R^{(\ell)}} f \Big|^2\, \frac{dt}{t}\Big)^{1/2} \Big\|_{L^q(\mu)} \lesssim \mu(Q^*)^{1/q}.\] We bound the right hand side of [\[Inested\]](#Inested){reference-type="eqref" reference="Inested"} by \(I_{\textup{nested}, 1} + I_{\textup{nested}, 2}\), where \[\begin{aligned} I_{\textup{nested}, 1} = \Big\| 1_{Q_0} \Big( \mathop{\sum_{R \in \mathcal{D}_{\textup{good}}:\, R \subset Q^*}}_{2^{-\kappa} < \ell(R) < 2^{s-r}} 1_{R} \int_{\ell(R)/2}^{\ell(R)} \Big| \mathop{\sum_{\ell=r+1}^{s-\textup{gen(R)}}} \theta_t (1_{R^{(\ell)} \setminus R^{(\ell-1)}}\Delta_{R^{(\ell)}} f) \Big|^2\, \frac{dt}{t}\Big)^{1/2} \Big\|_{L^q(\mu)} \end{aligned}\] and \[\begin{aligned} I_{\textup{nested}, 2} = \Big\| 1_{Q_0} \Big( \mathop{\sum_{R \in \mathcal{D}_{\textup{good}}:\, R \subset Q^*}}_{2^{-\kappa} < \ell(R) < 2^{s-r}} 1_{R} \int_{\ell(R)/2}^{\ell(R)} \Big| \mathop{\sum_{\ell=r+1}^{s-\textup{gen(R)}}} \theta_t (1_{R^{(\ell-1)}}\Delta_{R^{(\ell)}} f) \Big|^2\, \frac{dt}{t}\Big)^{1/2} \Big\|_{L^q(\mu)}. \end{aligned}\] ## The sum \(I_{\textup{nested}, 1}\) The following lemma is the key to handling this sum. We now see using this lemma that \(I_{\textup{nested}, 1}\) can dominated by \[\begin{aligned} & \sum_{\ell\geq r+1} 2^{-\alpha \ell/2} \Big\| 1_{Q_0} \Big( \sum_{k\le s-\ell} \Big(\sum_{R \in \mathcal{D}_{k,\textup{good}}:\, R \subset Q^*} 1_{R} 2^{-(k+\ell)m} \int |\Delta_{R^{(\ell)}} f(y)| \,d\mu(y)\Big)^2\Big)^{1/2} \Big\|_{L^q(\mu)} \\ &\lesssim \sum_{\ell\geq r+1} 2^{-\alpha \ell/2} \Big\| 1_{Q_0} \Big( \sum_{k\le s-\ell} \Big(\sum_{S \in \mathcal{D}_{k+\ell}:\, S \subset Q^*} \frac{1_S}{\mu(S)} \int |\Delta_{S} f(y)| \,d\mu(y)\Big)^2\Big)^{1/2} \Big\|_{L^q(\mu)} \\ &= \sum_{\ell\geq r+1} 2^{-\alpha \ell/2} \Big\| 1_{Q_0} \Big( \sum_{k\le s-\ell} \Big(\sum_{S \in \mathcal{D}_{k+\ell}} \frac{1_S}{\mu(S)} \int_S |\Delta_{k+\ell} f(y)| \,d\mu(y)\Big)^2\Big)^{1/2} \Big\|_{L^q(\mu)} \\ &\lesssim \Big\| \Big( \sum_{k \le s} [ E_{k} (|\Delta_{k}f|)]^2\Big)^{1/2} \Big\|_{L^q(\mu)} \lesssim \mu(Q^*)^{1/q}. \end{aligned}\] The last inequality follows from Stein's inequality [\[Stein\]](#Stein){reference-type="eqref" reference="Stein"} and [\[eq:MD-square.function.estimate\]](#eq:MD-square.function.estimate){reference-type="eqref" reference="eq:MD-square.function.estimate"}. ## The sum \(I_{\textup{nested}, 2}\) We begin by recording the following bound: We now have to do a case study. ## The case \((R^{(\ell-1)})^a = (R^{(\ell)})^a\) In this case we may write \[\begin{aligned} \label{eq:split1} 1_{R^{(\ell-1)}} \Delta_{R^{(\ell)}} f =-1_{(R^{(\ell-1)})^c} B_{R^{(\ell-1)}} b_{(R^{(\ell)})^a} + B_{R^{(\ell-1)}}b_{(R^{(\ell)})^a}, \end{aligned}\] where \[B_{R^{(\ell-1)}} = \frac{\langle f \rangle_{R^{(\ell-1)}}}{\langle b_{(R^{(\ell-1)})^a}\rangle_{R^{(\ell-1)}}}-\frac{\langle f \rangle_{R^{(\ell)}} }{\langle b_{(R^{(\ell)})^a}\rangle_{R^{(\ell)} }}\] with the minus term missing if \(\ell(R^{(\ell)}) = 2^s\). Accretivity condition gives that \[\begin{aligned} |B_{R^{(\ell-1)}}| \mu(R^{(\ell-1)}) \lesssim \Big| \int_{R^{(\ell-1)}} B_{R^{(\ell-1)}} b_{(R^{(\ell)})^a}\,d\mu\Big| &= \Big| \int_{R^{(\ell-1)}} \Delta_{R^{(\ell)}} f\,d\mu\Big|. \end{aligned}\] Combining with Lemma [\[lem:kk\]](#lem:kk){reference-type="ref" reference="lem:kk"} we see that for \((x,t) \in W_R\) there holds that \[|\theta_t(1_{(R^{(\ell-1)})^c} B_{R^{(\ell-1)}} b_{(R^{(\ell)})^a})(x)| \lesssim 2^{-\alpha \ell/2} \frac{1}{\mu(R^{(\ell-1)})} \int_{R^{(\ell-1)}} |\Delta_{R^{(\ell)}} f|\,d\mu.\] So to control the sum with the first term of [\[eq:split1\]](#eq:split1){reference-type="eqref" reference="eq:split1"} it is enough to note that for a fixed \(\ell \ge r+1\) there holds that \[\begin{aligned} &\Big\| 1_{Q_0} \Big( \sum_{k \le s-\ell} \Big( \sum_{R \in \D_k:\, R \subset Q^*} \frac{1_R}{\mu(R^{(\ell-1)})} \int_{R^{(\ell-1)}} |\Delta_{R^{(\ell)}} f(y)|\,d\mu(y) \Big)^2\Big)^{1/2} \Big\|_{L^q(\mu)} \\ & \le \Big\| 1_{Q_0} \Big( \sum_{k \le s-\ell} \Big( \sum_{S \in \D_{k+\ell-1}} \frac{1_S}{\mu(S)} \int_{S} |\Delta_{k+\ell} f(y)|\,d\mu(y) \Big)^2\Big)^{1/2} \Big\|_{L^q(\mu)} \\ &= \Big\| \Big( \sum_{k \le s} [ E_{k-1} (|\Delta_{k}f|)]^2\Big)^{1/2} \Big\|_{L^q(\mu)} \lesssim \mu(Q^*)^{1/q}. \end{aligned}\] In the last step we again used Stein's inequality [\[Stein\]](#Stein){reference-type="eqref" reference="Stein"} and [\[eq:MD-square.function.estimate\]](#eq:MD-square.function.estimate){reference-type="eqref" reference="eq:MD-square.function.estimate"}. We will not touch the second term of [\[eq:split1\]](#eq:split1){reference-type="eqref" reference="eq:split1"} yet--it will become part of the paraproduct. ## The case \((R^{(\ell-1)})^a = R^{(\ell-1)}\) We decompose \[\begin{aligned} 1_{R^{(\ell-1)}} \Delta_{R^{(\ell)}}f = \Big(\frac{\langle f \rangle_{R^{(\ell-1)}}}{\langle b_{R^{(\ell-1)}}\rangle_{R^{(\ell-1)}}}&b_{R^{(\ell-1)}}-\frac{\langle f \rangle_{R^{(\ell)}} }{\langle b_{(R^{(\ell)})^a}\rangle_{R^{(\ell)} }}b_{(R^{(\ell)})^a} \Big) \\ &+ 1_{(R^{(\ell-1)})^c} \frac{\langle f \rangle_{R^{(\ell)}} }{\langle b_{(R^{(\ell)})^a}\rangle_{R^{(\ell)} }}b_{(R^{(\ell)})^a}. \end{aligned}\] The term in the parenthesis will become part of the paraproduct, and we do not touch it further in this subsection. For the second term, using the construction of the stopping time and Lemma [\[lem:kk\]](#lem:kk){reference-type="eqref" reference="lem:kk"}, we have for \((x,t) \in W_R\) that \[\Big|\theta_t \Big( 1_{(R^{(\ell-1)})^c} \frac{\langle f \rangle_{R^{(\ell)}} }{\langle b_{(R^{(\ell)})^a}\rangle_{R^{(\ell)} }}b_{(R^{(\ell)})^a} \Big)(x)\Big| \lesssim 2^{-\alpha \ell/2} |\langle f \rangle_{R^{(\ell)}}|.\] We say that \(R \in \mathcal{S}_\ell\), if \((R^{(\ell-1)})^a = R^{(\ell-1)}\). To control the corresponding sum we note that \[\begin{aligned} \Big\| 1_{Q_0} &\Big( \mathop{\sum_{R \in \mathcal{D}_{\textup{good}}:\, R \subset Q^*}}_{2^{-\kappa} <\ell(R)<2^{s-r}}1_{R}\int_{\ell(R)/2}^{\ell(R)} \Big| \mathop{\sum_{\ell=r+1}^{s-\textup{gen(R)}}}_{R \in \mathcal{S}_\ell} \frac{\langle f \rangle_{R^{(\ell)}} }{\langle b_{(R^{(\ell)})^a}\rangle_{R^{(\ell)} }} \theta_t( 1_{(R^{(\ell-1)})^c} b_{(R^{(\ell)})^a}) \Big|^2\frac{dt}{t}\Big)^{1/2} \Big\|_{L^q(\mu)}\\ &\lesssim \Big\| 1_{Q_0} \Big( \mathop{\sum_{R \in \mathcal{D}_{\textup{good}}:\, R \subset Q^*}}_{2^{-\kappa} <\ell(R)<2^{s-r}}1_{R} \mathop{\sum_{\ell=r+1}^{s-\textup{gen(R)}}}_{R \in \mathcal{S}_\ell}2^{-\alpha \ell/2} |\langle f \rangle_{R^{(\ell)}}|^2 \Big)^{1/2} \Big\|_{L^q(\mu)}\\ &\leq \Big\| 1_{Q_0} \Big( \sum_{\ell \ge r+1} 2^{-\alpha \ell /2} \sum_{S \in \D:\, S \subset Q^*} |\langle f \rangle_{S}|^2 A_S \Big)^{1/2} \Big\|_{L^q(\mu)} \\ &\lesssim \Big\| 1_{Q_0} \Big( \sum_{S \in \D:\, S \subset Q^*} |\langle f \rangle_{S}|^2 A_S \Big)^{1/2}\Big\|_{L^q(\mu)} \\ & \lesssim \mu(Q^*)^{1/q}, \end{aligned}\] where we denoted \[A_S(x) := \mathop{\sum_{S' \in \textup{ch}(S)}}_{(S')^a = S'} 1_{S'}(x).\] For the final estimate one can use the fact that \(|f| \le 1\) to throw away the averages, and then use Hölder with exponent \(p:= 2/q > 1\) together with Lemma [\[lem:mescar\]](#lem:mescar){reference-type="ref" reference="lem:mescar"}: \[\begin{aligned} \Big\|& \Big( \sum_{S \in \D:\, S \subset Q^*} A_S \Big)^{1/2}\Big\|_{L^q(\mu)} \\ & \le \Big[ \mu(Q^*)^{1-1/p} \Big( \sum_{F \in \mathcal{F}_{Q^*}} \mu(F) \Big)^{1/p} \Big]^{1/q} \\ &\lesssim ( \mu(Q^*)^{1-1/p} \mu(Q^*)^{1/p} )^{1/q}\\ & = \mu(Q^*)^{1/q}. \end{aligned}\] ## The Carleson estimate for the paraproduct Combining the above two cases and collapsing the remaining telescoping summation we are left with: \[\begin{aligned} \Big\| 1_{Q_0} \Big(& \mathop{\sum_{R \in \mathcal{D}_{\textup{good}}:\, R \subset Q^*}}_{2^{-\kappa} <\ell(R)<2^{s-r}}1_{R} \int_{\ell(R)/2}^{\ell(R)} \Big| \frac{\langle f \rangle_{R^{(r)}}}{\langle b_{(R^{(r)})^a} \rangle_{R^{(r)}}} \theta_t b_{(R^{(r)})^a}\Big|^2\,\frac{dt}{t}\Big)^{1/2} \Big\|_{L^q(\mu)} \\ &\lesssim \Big\| 1_{Q_0} \Big( \mathop{\sum_{S\in \D}}_{S\subset Q^*} \mathop{\sum_{R \in \D }}_{R^{(r)}=S} 1_{R} \int_{\ell(R)/2}^{\ell(R)} \big| \theta_t b_{S^a}\big|^2\,\frac{dt}{t}\Big)^{1/2} \Big\|_{L^q(\mu)}\\ & = \Big\| 1_{Q_0} \Big( \sum_{F \in \mathcal{F}_{Q^*}} \sum_{S: S^\alpha=F} \mathop{\sum_{R \in \D }}_{R^{(r)}=S} 1_{R} \int_{\ell(R)/2}^{\ell(R)} \big| \theta_t b_{F}\big|^2\,\frac{dt}{t}\Big)^{1/2} \Big\|_{L^q(\mu)}\\ & \leq \Big\| 1_{Q_0} \Big( \sum_{F \in \mathcal{F}_{Q^*}} \sum_{R: R \subset F} 1_{R} \int_{\ell(R)/2}^{\ell(R)} \big| \theta_t b_{F}\big|^2\,\frac{dt}{t}\Big)^{1/2} \Big\|_{L^q(\mu)}\\ & \leq \Big\| 1_{Q_0} \Big( \sum_{F \in \mathcal{F}_{Q^*}} 1_{F} \int_{0}^{\ell(F)} \big| \theta_t b_{F}\big|^2\,\frac{dt}{t} \Big)^{1/2} \Big\|_{L^q(\mu)}\\ & \leq \sum_{j\geq 0} \Big( \sum_{F \in \mathcal{F}^j_{Q^*}} \Big\| \Big(1_{F} \int_{0}^{\ell(F)} \big| \theta_t b_{F}\big|^2\,\frac{dt}{t}\Big)^{1/2} \Big\|^q_{L^q(\mu)} \Big)^{1/q}\\ & \le V_{\textup{loc}, q} \sum_{j\geq 0} \Big( \sum_{F \in \mathcal{F}^j_{Q^*}} \mu(F)\Big)^{1/q} \lesssim V_{\textup{loc}, q} \mu(Q^*)^{1/q}. \end{aligned}\] In the first inequality we used the stopping time conditions and the fact that \(|f|\leq 1\), while the penultimate inequality follows from assumption \((4)\) of theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}.

{'timestamp': '2014-01-22T02:12:31', 'yymm': '1401', 'arxiv_id': '1401.5457', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5457'}

# Introduction More than fifty years after Philip Anderson suggested the localization of de Broglie electron waves its analogous manifestation involving classical electromagnetic (EM) waves is still under discussion. In particular, several reported observations of localization of light in three dimension have been questioned. A recent theoretical study even suggests the absence of light localization in a random three-dimensional (3D) ensemble of point scatterers. The situation appears to be much clearer in constrained geometries having a finite number \(N\) of propagating EM modes at a given frequency \(f\). In particular, the physics of wave transport in disordered waveguides of length \(L\) much exceeding their width \(w\) is currently well established. The latter geometry is referred to as quasi-one-dimensional (quasi-1D) when \(w\) is smaller or of the order of the transport mean free path \(l^*\) due to disorder because transport equations for ensemble-averaged quantities (such as, e.g., the average intensity) turn out to involve only one spatial dimension. In contrast to the case of fully 3D systems, Anderson localization should always take place provided that the dimensionless conductance \(g_0\sim N l^{\ast}/L\) is less than one. A complete set of data on Anderson localization in quasi-1D disordered systems has been presented by the group of Azriel Genack about a decade ago. These experiments concerned microwaves propagating in hollow copper tubes filled with random assemblies of spheres having sizes of the order of the wavelength of the propagating radiation. More recent experiments on light propagation in silicon waveguides further support the localization scenario in quasi-1D waveguide geometries. In this article, we present a numerical study of EM wave transport in disordered quasi-1D waveguides at terahertz frequencies. The interest of studying Anderson localization of THz waves is twofold. First, the refractive index contrast between the scatterers and the background medium can be exceptionally high, of the order of 5 to 1 for common materials. Second, it has been shown that the electric field amplitude of THz waves can be imaged non-invasively, thus providing a unique tool to study the spatial distribution of the electric field in and close to the localized regime with high spatial resolution. We expect that coherent terahertz waves in ferroelectric crystals will provide a powerful system, outside of established fields such as optical-frequency EM waves, microwaves, acoustic waves and matter waves, to quantitatively study Anderson localization effects. As a first step towards this goal, we present a numerical study of terahertz wave transport in quasi-1D disordered waveguides for which exact and approximate theoretical predictions are available for the the average transport coefficients near the onset of Anderson localization, as well as for the full distribution function of transmission fluctuations. These quantities can be directly measured in future experiments. Moreover, we establish crucial parameters of the system, which maximize localization effects in structured LiNbO\(_{3}\) waveguides in order to guide future experimental and theoretical studies. We note that our approach explicitly models the experimental situation in all relevant details, and is therefore distinctive from many other numerical studies in which either point scatterers or surface roughness induce diffuse transport and localization. # Phonon-polaritons in ferroelectric crystals Coupled electromagnetic-lattice vibrational waves, referred to as phonon-polaritons, can be generated within ferroelectric crystals through impulsive stimulated Raman scattering, or alternatively, through optical rectification by irradiating a crystal with a pulsed femtosecond laser. For the ferroelectric crystal LiNbO\(_3\), at frequencies below 2 THz, phonon-polariton dispersion is light-like with propagation speeds of *c*/*n* where *c* is the speed of light in vacuum and the refractive index is *n* = 5.1. Recently developed pump-probe techniques allow for the generation and imaging of both the time-dependent amplitude and phase of phonon-polaritonic fields. In addition to powerful imaging methods, scattering sites can be readily fabricated in the form of micrometer-scale cylindrical air-filled holes in a LiNbO\(_{3}\) crystal using micro-machining methods such as laser ablation and optical lithography. The spatial resolution requirements for both machining structured waveguides and imaging propagating polaritonic waves in ferroelectric crystals are nearly two orders of magnitude less stringent than is the case for visible light. For instance, the wavelength of a phonon-polariton at a frequency of 1 THz in LiNbO\(_{3}\) is roughly 60 \(\mu\)m. The large refractive index contrast between LiNbO\(_{3}\) and air at terahertz frequencies induces strong scattering of phonon-polaritons from the holes, leading to a situation that is favorable for the observation of Anderson localization. # Simulations of terahertz wave transport Since dispersion of phonon-polariton waves in LiNbO\(_3\) is light-like, we simulate phonon-polaritonic transport through disordered quasi-1D waveguides by solving the Helmholtz equation for the electromagnetic field in the frequency domain using finite element methods in two dimensions. The waveguide is supposed to be rectangular, with the length \(L\) much exceeding the width \(w\). Perfectly reflecting boundary conditions are imposed at the outer edges of the waveguide. To eliminate back reflections at the ends of the waveguide, a 0.5 mm-long perfectly matched layer is used. It effectively absorbs all incident radiation over approximately one wavelength. To tune the scattering strength in the system, calculations are performed for two scatterer radii \(a = 25\) and 50 \(\mu\)m, and at three excitation frequencies \(f = 0.5\), 1 and 1.5 THz. These parameters are readily accessible in experiments using LiNbO\(_{3}\) crystals. We consider air-filled scatterers having refractive index \(n_{\mathrm{sc}} = 1\) and fix the width of the waveguide to \(w = 500\) \(\mu\)m. The presence of the scatterers lowers the effective refractive index of the waveguide \(n_{\mathrm{eff}}\). Using Bruggeman effective medium theory, we find that \(n_{\mathrm{eff}}\) decreases from 5.1 at areal filling fraction of the scatterers \(\phi = 0\) down to 4.93 and 4.09, at \(\phi = 0.0193\) and 0.196, respectively, for \(a = 50\) \(\mu\)m. Over a 0.5--1.5 THz frequency range, the number of transverse propagating modes \(N = kw/\pi\) is therefore \(N=7\) to 25 for an empty waveguide and \(N = 6\) to 24 for the areal filling fractions considered here. To calculate the average transmission, ensembles consisting of 100 arrangements of randomly positioned circular scatterers are generated at areal filling fractions up to \(\phi \sim 0.2\). To generate a given arrangement, scatterers are successively added to the system using a random number generator with the constraint that no two scatterer centers lie closer than \(2(a + 1)\) \(\mu\)m. The additional 1 \(\mu\)m is employed to avoid meshing problems during simulation runs. This method of placing circular scatterers results in a hard-sphere liquid-like pair correlation distribution, the multiple scattering properties of which are well studied both theoretically and experimentally. An adaptive, variable-size Delaunay triangular grid is created using commercially available finite element simulation software (COMSOL Multiphysics 4.3). The tradeoff between numerical accuracy and computation time dictates using an average grid point density of \(430\) \(\mu\)m\(^{-2}\) having maximum and minimum spacing of 10 and 4.4 \(\mu\)m, respectively. Results differ by less than 1% after increasing the total number of grid points by a factor of 30. The input of the waveguide is excited with a monochromatic plane wave polarized transverse to the long axis of the waveguide and solutions to the two-dimensional Helmholtz equation for the in-plane electromagnetic field are found. For illustration purposes, we show in Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"} grayscale plots of the wave intensity inside the waveguide at two different frequencies for randomly placed 50 \(\mu\)m radius scatterers at a filling fraction \(\phi = 0.196\). Transmission coefficients are determined by calculating the power transmitted downstream from the scatterers, and then dividing the result by the power transmitted through the same waveguide without scatterers. # Results Next we present the results or our a numerical study of terahertz wave transport in quasi-1D disordered waveguides and it's comparison to the exact and approximate theoretical predictions reported in the literature. ## Average transmission Representative results for average transmission \(\langle T \rangle\) as a function of waveguide length are shown in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"} for different areal densities of scatterers at the frequency \(f = 0.5\) THz. To characterize the influence of scatterer density on the crossover from diffusive to localized transport, we calculated the average transmission for waveguides having various areal filling fractions of air-filled scatterers ranging from \(\phi = 0.039\) to 0.196 for \(a = 50\) \(\mu\)m and from \(\phi = 0.049\) to 0.197 for \(a = 25\) \(\mu\)m. For brevity, we focus our discussion on results for *a* = 50 \(\mu\)m; similar results are obtained for all three frequencies and both scatterer sizes explored in this study. To keep the same level of accuracy throughout the theoretical discussion, we use simplified expressions for the dimensionless conductance \(g_0\) and the average diffuse transmission \(\langle T_0 \rangle \simeq g_0/N\) taking into account internal reflections of waves at the entry and exit of the waveguide through the so-called extrapolation length \(z_0 = (\pi/4) l^*\): \[g_0 = \left( \frac{L}{ \label{dimcond}\] where \[T \(and its probability distribution\)P(T/T )\((see below), cannot be evaluated to the same level of accuracy. Fits using the diffuse regime result---Eq.\ (\ref{dimcond}) and\)T = T_0 = g_0/N\(---deviate appreciably from the data for waveguide lengths\)L \]\langle T \rangle \simeq \frac{g_0}{N} \left[ 1-\frac{1}{3 g_0} + \frac{1}{45g_0^2} + \frac{2}{945 g_0^3} + \ldots \right]. \label{power series}\[ As can be seen in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}, fits to the numerical data using Eq. ([\[power series\]](#power series){reference-type="ref" reference="power series"}) with \(l^*\) as the only fit parameter are excellent. Moreover, we illustrate in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}(a) that the approximations made in the series expansion, Eq. ([\[power series\]](#power series){reference-type="ref" reference="power series"}) truncated after the fourth term, are small as long as \(L/ The analysis of a number of areal filling fractions for different frequencies allows us to study the dependence of transport mean free path on scatterer density. We find that the inverse transport mean free path\)`<!-- -->`{=html}1/l\^\*\(obtained from the fits increases linearly with\]for\[, whereas at larger\],\)`<!-- -->`{=html}1/l\^\*\(grows slightly faster. For arrangements of monodisperse scatterers having no spatial correlations, the transport mean free path can be expressed as\)l\^\* = l/(1-)\(, where\)l = 1/(Q\_)\(is the scattering mean free path and\[is the scattering anisotropy parameter. A linear fit having no intercept to the first four points in the inset of Fig.\ \ref{Fig3}(b), and using\)= 0.5\(, yields\)Q\_ = 32\(mm\)\^-1\(. This is not too far from the value obtained using exact Mie scattering theory for a long cylinder in an infinite plane. Using the latter theory we find that as a function of frequency,\)Q\_\(is peaked at\)f = 1\(THz with a maximum of\)Q\_ \(mm\)\^-1\(and that the anisotropy parameter\]is approximately equal to 0.5 over the frequency range from 0.5 to 1.5 THz. We note that we do not expect\)Q\_\(of a cylindrical scatterer in the waveguide and in the homogenous infinite space to be identical. It will somewhat depend on the finite number of transverse modes accessible for the scattered waves. Moreover, there might be some effect due to the finite scatterer size as it becomes comparable to the width of the waveguide. In addition, at higher values of\)\(,\)`<!-- -->`{=html}1/l\^\*\(deviates slightly from the linear scaling\)`<!-- -->`{=html}1/l\^\* \(possibly because of short-range positional correlations, which are known for dense colloidal dispersions to either decrease or increase\)l\^\*\(relative to the uncorrelated system, depending upon the size, number density and refractive index of the scatterers. However, the study of the complex interplay between these competing effects and their influence on\)l\^\*\(is beyond the scope of the present article. \subsection{Fluctuations of transmission} Complementary to the average transmission, transmission fluctuations provide a second measure for the onset of localization. Using larger ensembles consisting of 5000 scatterer configurations, we determine probability distributions\)P(s)\(of the normalized transmission\)s = T/T \(for\)= 0.079\(,\)N = 7\(and\)f = 0.5\(THz and waveguide lengths ranging from\)L = 2\(to 20 mm. In Fig.\ \ref{Fig4}, numerical results are compared with the theoretical prediction: \begin{eqnarray} P(s) &=& \int\limits_{-i \infty}^{i \infty} \frac{d x}{2 \pi i} \mathrm{exp} \left[x s-\Phi(x) \right], \label{ps1} \\ \Phi(x) &=& g \ln^2 \left( \sqrt{1+x/g} + \sqrt{x/g} \right). \label{ps2} \end{eqnarray} Theory predicts that the distribution is parameterized by a single parameter\)g = g \(that we adjust to fit the numerical data. As\)L\(increases to values greater than\) The large fluctuations of \(T\) can be quantified by the variance of \(s = T/\langle T \rangle\) that we show in Fig. [\[Fig5\]](#Fig5){reference-type="ref" reference="Fig5"}(b) as a function of waveguide length \(L\). Numerical data (symbols) are compared with the theoretical results \(\mathrm{var}(s) = 2/3 \langle g \rangle\) (dashed line) and \(\mathrm{var}(s) = (2/3 \langle g) \rangle (1+3l^*/2L)/(1+l^*/L)^3\) (solid line), where \(\langle g \rangle\) is obtained from Eq. (6.23) of Ref. \[see the solid line in Fig. [\[Fig5\]](#Fig5){reference-type="ref" reference="Fig5"}(a)\]. The equation derived in Ref. takes into account the finite-size effects and better agrees with the numerical results for short samples whereas the simpler equation \(\mathrm{var}(s) = 2/3 \langle g \rangle\) seems to do a better job at large \(L > 10\) mm. We would like to stress here that lines in Fig. [\[Fig5\]](#Fig5){reference-type="ref" reference="Fig5"} are not fits to numerical data but theoretical results obtained using \(l^* = 0.81\) mm extracted from fits of Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}. Good overall agreement between theory and numerics, together with the good quality of fits shown in Fig. [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"}, suggest that Eq. ([\[ps1\]](#ps1){reference-type="ref" reference="ps1"}) for the distribution of total transmission may be a good approximation even in the localized regime, provided that the parameter \(g\) is understood as \(\langle g \rangle\) and computed using the exact theory available for this quantity. This is a nontrivial result because Eq. ([\[ps1\]](#ps1){reference-type="ref" reference="ps1"}) was initially derived for weak disorder and can be rigourously justified only for \(g = g_0 \gg 1\). Indications of qualitative validity of Eq. ([\[ps1\]](#ps1){reference-type="ref" reference="ps1"}) in the localized regime were already contained in some of the previous experimental studies. In contrast to these studies in which the actual values of \(\langle g \rangle\) were not known, our results allow for a quantitative test of Eq. ([\[ps1\]](#ps1){reference-type="ref" reference="ps1"}) because we can compare the values of \(\langle g \rangle\) extracted from the fits with the values following from the exact theory \[see Fig. [\[Fig5\]](#Fig5){reference-type="ref" reference="Fig5"}(a)\]. Such a comparison shows that Eq. ([\[ps1\]](#ps1){reference-type="ref" reference="ps1"}) has quite a reasonable degree of precision that is likely to be sufficient for description of experimental data. # Conclusions Our numerical results clearly demonstrate the feasibility of observation of Anderson localization of THz waves in quasi-1D disordered waveguides under realistic experimental conditions. Due to the high contrast in refractive index between the scatterers and the matrix medium attainable in the THz frequency range, the localization length can be made quite short, typically of the order of several mm, already for moderate area filling fractions of scatterers \(\phi \sim 0.1\). The results of our numerical simulations are in overall good quantitative agreement with existing approximate and exact theories for the average transmission and its full statistical distribution provided that in the latter case, the parameter \(g\) of the distribution is understood as the average conductance \(\langle g \rangle\) and calculated using the exact theory available for it (see, e.g., Ref. ) and not as the bare conductance \(g_0\) defined through the geometrical parameters of the waveguide according to Eq. ([\[dimcond\]](#dimcond){reference-type="ref" reference="dimcond"}). Moreover, our study shows that all relevant length scales, such as the scatterer size \(a\), the waveguide width \(w\) and length \(L\), the transport mean free path \(l^{*}\) are rather well separated which allows to clearly distinguish between different regimes of propagation. Finally, structures with design parameters close or even exactly equal to those used in our simulations can be manufactured with high precision as has been previously demonstrated. This opens the door for future experimental studies of EM wave localization with THz waves that promise improved spatial resolution and better experimental access to the EM field amplitude rather than intensity. The present project has been financially supported by the the Swiss National Science Foundation (projects 132736, 142571, 140943 and 149867) and the Swiss State Secretariat for Education, Research and Innovation (SERI). S.E.S. acknowledges support from the Federal Program for Scientific and Scientific-Pedagogical Personnel of Innovative Russia for 2009--2013 (contract No. 14.B37.21.1938).

{'timestamp': '2014-01-23T02:08:55', 'yymm': '1401', 'arxiv_id': '1401.5660', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5660'}

# Introduction Top-quality type IIa diamond crystals suitable for advanced applications in X-ray crystal optics at synchrotrons and X-ray free-electron lasers (XFELs) have recently become available due to refinements in the high-pressure high-temperature (HPHT) crystal synthesis method. Diamond crystal plates with (111) surface orientation are of primary importance for front-end diffracting X-ray optics due to a greater intrinsic energy bandwidth of the 111 Bragg reflection and the resulting flux of the reflected X-rays compared with those of higher-order reflections. However, production of the (111) crystal plates with large defect-free regions is a more challenging task compared with production of plates of other orientations close to the \[001\] direction. This is due to (\(i\)) the (001) diamond HPHT growth sector has the best crystal quality and the lowest impurity concentration; and (\(ii\)) the {111} crystal faces are the most resistant to polishing since these planes have the highest atomic density. The primary application of diamond (111) crystal plates in X-ray optics are the high-heat-load double-crystal monochromators at synchrotron undulator beamlines (e.g., ). A scheme of a double-crystal monochromator in Bragg reflection geometry is shown in Fig. [\[fig:dcm\]](#fig:dcm){reference-type="ref" reference="fig:dcm"}. If the first crystal is made sufficiently thin, the beam transmitted through it can be used in a parallel experiment downstream, i.e., beam multiplexing. Utilization of X-ray monochromators with low absorption crystals for beam multiplexing was pioneered at the TROIKA beamline at the ESRF (European Synchrotron Radiation Facility, Grenoble, France) . This beam multiplexing approach gains even more importance for hard X-ray free-electron lasers, which recently redefined the frontiers of X-ray sciences. Using a single straight electron trajectory of an XFEL, multi-user operations cannot be achieved as is commonly accomplished at storage-ring-based synchrotron sources. However, multiplexing can be performed by means of X-ray optics, which enables simultaneous delivery of portions of the XFEL beam to several experiments. This yields an increase in the total number of performed experiments and thus reduces the high XFEL operating cost per experiment. Highly developed crystal fabrication and processing methods for silicon (Si) have led to a few attempts to use Si crystals for beam-multiplexing XFEL monochromators. Greater X-ray absorption in Si requires utilization of ultra-thin crystals (thicknesses \(\approx\) 5-10 \(\mu\)m) to achieve sufficient transmittance of the XFEL beam (30-80%) over a photon energy of 4-10 keV. High-quality ultra-thin Si crystals for XFEL beam-multiplexing monochromators can be manufactured using state-of-the-art processing methods; however, such ultra-thin crystals have been found unstable in the XFEL beam . Similar levels of transmittance can be achieved with crystal thicknesses of 50-100 \(\mu\)m in the case of diamond due to its lower X-ray absorption. The greater crystal thickness and Young's modulus for diamond ensure a substantially greater stiffness of the sample. Diamond crystals with such thicknesses are expected to be more stable in the XFEL beam if an appropriate crystal mounting scheme is provided. Because of the small divergence (\(\lesssim\) 1 \(\mu\)rad rms) of the XFEL beam its wavefront is particularly sensitive to imperfections of X-ray optics. Disturbances of the wavefront introduced by the optics should be much less or at least comparable to the beam divergence, which is a challenging requirement, especially for diamond crystal optics prone to crystal defects. Along with the presence of intrinsic defects, a mounting-induced crystal strain is another main factor that leads to deterioration of the crystal diffraction performance, which disturbs the radiation wavefront. The mounting solution previously developed by some of the co-authors for the diamond (001) self-seeding XFEL monochromator was found to show limitations in heat transfer between the crystal plate and the graphite holder where the plate was loosely mounted. In this work we report on the production of 100-\(\mu\)m-thick and 300-\(\mu\)m-thick type IIa HPHT diamond crystal plates with (111) orientation. Each crystal plate has a region with low concentration of defects (working region) of about 5 \(\times\) 2 mm\(^2\) sufficient for use in XFEL optics. More importantly, we present a solution for strain minimization as a special mounting of such crystals on a rigid substrate. All parts of the optical assembly were fabricated out of diamond, which improved heat transfer and radiation hardness of the device. The mounting-induced strain was evaluated using double-crystal X-ray topography in rocking curve diffraction imaging mode. The mounting induced crystal distortion (1.5 \(\mu\)rad rms variation in the rocking curve peak position across the working region) was substantially smaller than the full width at half maximum (FWHM) of the rocking curve width (25-30 \(\mu\)rad) and still comparable to the divergence of the XFEL beam. Two such assemblies were installed into the vacuum tank of the beam-multiplexing monochromator at the XPP instrument of the Linac Coherent Light Source and successfully tested. The monochromator was operated in Bragg reflection geometry, as shown in Fig. [\[fig:dcm\]](#fig:dcm){reference-type="ref" reference="fig:dcm"}. The capability of splitting the XFEL beam into a pink (transmitted) and a monochromatic (reflected) branch was demonstrated, which enables the use of the XFEL beam in two experiments simultaneously. # Diamond (111) crystal plates Type IIa diamond single crystals were grown at the Technological Institute for Superhard and Novel Carbon Materials (TISNCM) using the temperature-gradient method at high static pressure and high temperature (e.g., ). The temperature of crystallization was 1460 \(^{\circ}\)C at a pressure of 5.5 GPa. After the crystallization process, diamond crystals were cut by a laser along the \(\{111\}\) crystal plane with a \(\simeq\) 2\(^{\circ}\) angular offset from the plane. The angular offset was introduced to facilitate mechanical polishing of the crystal plates. The plates were polished to a micro-roughness of \(\approx\) 10 nm (rms). Preliminary selection of crystal plates with low density of crystal defects was made using Lang X-ray topography. For the final selection of diamond crystal plates, white-beam X-ray topography was performed at the MRCAT 10BM (bending magnet) beamline  of the Advanced Photon Source. White-beam X-ray topographs were obtained in Laue (transmission) geometry on the diamond plates enclosed in Kapton film holders. The source-to-sample distance was \(L\simeq\) 27 m, the effective source size at the beamline was \(s \approx 300\) \(\mu\)m, and the distance from the sample to the observation plane was \(d \simeq\) 0.1 m. Thus, the expected spatial resolution in the observation plane was \(\delta = sd/L \simeq\) 1 \(\mu\)m, which was comparable with the photographic resolution of the utilized X-ray film (AGFA STRUCTURIX D3-SC). We note that the image resolution was somewhat reduced due to crystal instability caused by heat load of the X-rays during the shortest available exposure time (\(\approx\) 2 s). Multiple images were taken to mitigate this problem. High-quality topographs were obtained from the \(3\bar{1}\bar{1}\) reflections. They are shown in Fig. [\[fig:wbt\]](#fig:wbt){reference-type="ref" reference="fig:wbt"}(a) and (b) for the selected 300-\(\mu\)m-thick and 100-\(\mu\)m-thick crystal plates, respectively. The dotted box in Fig. [\[fig:wbt\]](#fig:wbt){reference-type="ref" reference="fig:wbt"}(a) and (b) shows the 5 \(\times\) 2 mm\(^2\) working regions originated for the most part from the (001) growth sector with a low concentration of defects. A few defects of about 10 \(\mu\)m in size are noticeable in the working regions of the crystals. These can be attributed to minor dislocations and micro-inclusions. The diagonal lines represent stacking faults that likely originate at the growth sector boundaries. Nevertheless, the crystals are of remarkable quality in the 5 \(\times\) 2 mm\(^2\) region, especially if compared to commercially available type IIa diamond (111) crystals (e.g., the crystal studied earlier by some of the authors ). # Crystal mounting method The relative intrinsic energy bandwidth of X-ray Bragg reflection is a very small quantity (\(\Delta E/E \approx 10^{-5}-10^{-7}\)), which is related to the small intrinsic angular width \(\Delta \theta\) \(\approx\) \(1-100\) \(\mu\)rad (except Bragg-backscattering cases \(\theta \simeq\) 90\(^{\circ}\)): \[\frac{\Delta E}{E} = \frac{\Delta \theta}{\tan{\theta}}.\] Since the relative energy bandwidth is small, a similarly small relative crystal strain \(\delta a / a \sim \Delta E /E\) (\(a\) is the crystal lattice parameter) can distort the shape of the intrinsic curves and reduce performance of the crystal as a monochromator. Thin Bragg crystals are particularly affected since relative deformations of the crystal are larger for a given applied stress (force). A simple fact that a thin crystal has a fixed contact point with a holder (achieved via bonding using glue, tape, or even grease) can be distinguished in Bragg diffraction from a situation where the crystal has a nearly free boundary (i.e., motion of the crystal is constrained only by friction against a crystal holder). Such a nearly free boundary condition is appropriate for certain applications; however, in this case the crystal is very susceptible to vibrations. Also, thermal transfer between the crystal and the holder could be limited due to reduced surface contact. These factors may severely deteriorate diffraction performance of the crystal under an intense incident X-ray beam. In this work, an alternative approach to mounting a diffracting crystal is explored, which includes a gentle pressure (a spring force) applied to the crystal placed on a rigid substrate. This force is applied in a controlled manner, such that the resulting strain is evaluated and limited to desired specifications. The optical assembly is schematically shown in Fig. [\[fig:design\]](#fig:design){reference-type="ref" reference="fig:design"}. In order to minimize issues related to differential thermal expansion, and to improve radiation hardness and thermal transfer all parts of the assembly were manufactured out of diamond materials. To improve thermal transfer between the substrate and the crystal their surfaces in contact were polished. A high-quality type IIa HPHT diamond (111) crystal plate is mounted on a substrate fabricated out of polycrystalline chemical vapor deposited (CVD) diamond. The substrate is sufficiently thick (500 \(\mu\)m) to provide rigid support for the diamond plate. The surface of the substrate in contact with the diamond plate was polished to \(\approx\) 10 nm (rms) micro-roughness. The substrate has two small rectangular openings for insertion of restrainers made of low-quality type IIa HPHT diamond and one large rectangular window (5 \(\times\)`<!-- -->`{=html}2 mm\(^2\)) for passage of X-rays transmitted through the diamond plate. The diamond plate has two small rectangular cuts on the sides such that the restrainers prevent lateral displacement of the plate on the substrate. The restrainers have grooves for insertion of 15-\(\mu\)m-thick CVD diamond stripes (CVD springs) that provide gentle force on the diamond plate, thus preventing its vertical displacement. The force acting on the diamond is \[F \propto \frac{\alpha }{L^2}k, \label{eq:force}\] where \(\alpha\) is the bending angle of the CVD spring as shown in Fig. [\[fig:design\]](#fig:design){reference-type="ref" reference="fig:design"}(b), \(L\) is the length of the spring, and k is the flexural stiffness of the spring. The spring can be approximated by a cantilever beam with an area moment of inertia \(I\), such that \(k = EI\), where \(E\) is the Young's modulus. Thus, for a fixed angle \(\alpha\) (in our design \(\alpha \simeq 15^{\circ}\)) the force can be controlled by either variation of the length of the spring or variation of the spring stiffness. Increasing the length of the spring was found to be insufficient in our experiments to reach very small forces to minimize mounting strain in diamond crystal plates with \(\approx\) 100-\(\mu\)m thickness. An additional spring force reduction was achieved by changing the stiffness of the CVD springs. The best solution to substantially reduce the stiffness of the springs was to introduce periodic strain-relief cuts (perforation) in the spring as shown in Fig. [\[fig:spring\]](#fig:spring){reference-type="ref" reference="fig:spring"}. All diamond components were manufactured at TISNCM using HPHT and CVD diamond synthesis methods and a controlled precision laser cutting of diamond materials. Polycrystalline CVD diamond films were deposited onto 640-\(\mu\)m-thick Si single-crystal substrates. Prior to deposition the Si substrates were ultrasonically treated in diamond powder suspension in ethanol. Diamond deposition process was carried out in a bell jar-type Microwave Plasma-Assisted Chemical Vapor Deposition (MWPA CVD) reactor. The substrate temperature was maintained at 860-870 C\(^\circ\). The methane concentration was 6% in a total flow of 500 sccm. The microwave power was in the range 3-3.2 kW at a total pressure of about 200 mbar during deposition of thick films for production of the 500-\(\mu\)m-thick CVD diamond substrates, and 130 mbar during deposition of thin flexible films for production of the 15-\(\mu\)m-thick CVD springs. After the deposition the diamond films were separated from the silicon substrates by dissolution of the substrates in 44 wt.% solution of KOH in water at 50\(^\circ\). The devices were assembled and cleaned at the Advanced Photon Source. Microscope images of the all-diamond assemblies are shown in Fig. [\[fig:asmb\]](#fig:asmb){reference-type="ref" reference="fig:asmb"}. Figure [\[fig:asmb\]](#fig:asmb){reference-type="ref" reference="fig:asmb"}(a) shows an assembly with a 300-\(\mu\)m-thick diamond crystal plate mounted using non-perforated CVD diamond springs. Figure [\[fig:asmb\]](#fig:asmb){reference-type="ref" reference="fig:asmb"}(b) shows an assembly with a 100-\(\mu\)m-thick diamond crystal plate mounted using perforated CVD diamond springs. Figure [\[fig:asmb\]](#fig:asmb){reference-type="ref" reference="fig:asmb"}(c) shows an assembly with the same 100-\(\mu\)m-thick diamond crystal plate mounted using non-perforated CVD diamond springs. Figure [\[fig:asmb\]](#fig:asmb){reference-type="ref" reference="fig:asmb"}(d) shows an enlarged view of the perforated spring acting on the 100-\(\mu\)m-thick diamond crystal plate, and Fig. [\[fig:asmb\]](#fig:asmb){reference-type="ref" reference="fig:asmb"}(e) shows an enlarged side perspective view of a restrainer with an inserted CVD spring. # Characterization of mounting-induced strain ## Experimental Rocking curve measurements and rocking curve imaging  were the main diagnostic methods for evaluating the mounting-induced crystal strain in the all-diamond assemblies. Double-crystal topography using a Cu K\(\alpha\) rotating anode X-ray source was performed to map the rocking curve of diamond crystal plates mounted in the holder assemblies. A Si (220) first crystal with an asymmetry angle \(\eta_{\mathrm{Si}} \simeq\) 22.2\(^{\circ}\) was used to collimate an X-ray beam incident on the diamond crystal plate (C) mounted in the all-diamond assembly (Fig. [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"}). The diamond crystal was set for the 111 reflection. The Bragg angle of the diamond crystal was \(\theta_{\mathrm{C}} \simeq\) 22.0\(^{\circ}\), which was close to the Bragg angle of the collimator crystal \(\theta_{\mathrm{Si}} \simeq\) 23.7\(^{\circ}\). In order to align the diamond crystal for the 111 reflection and to obtain the rocking curve of the entire crystal (total rocking curve), the intensity reflected from a uniformly illuminated diamond crystal was measured using a scintillation detector (SD). A CCD camera (with a resolution \(60\times60\) \(\mu\)m\(^2\)) replacing SD was used to obtain a series of X-ray diffraction images at different angular positions of the diamond crystal through the rocking curve of the 111 reflection. The image data were sorted such that a local rocking curve was obtained for every pixel, thus making it possible to map the rocking curve parameters over the entire crystal (i.e., rocking curve imaging). The angular resolution in the rocking curve scan was \(\approx\) 4 \(\mu\)rad, which is small compared with the FWHM of the ideal rocking curve in the double-crystal geometry (\(25\) \(\mu\)rad). The latter was evaluated using the dynamical theory of X-ray diffraction assuming a uniform angular distribution of the radiation incident on the first crystal (over an angular range exceeding the angular acceptance region of the crystal) and a photon energy distribution of the Cu K\(\alpha\) characteristic lines. ## Analysis of total and local rocking curves Figure [\[fig:rc_abc\]](#fig:rc_abc){reference-type="ref" reference="fig:rc_abc"} shows measured total rocking curves (blue circles and lines), the theoretical rocking curve (solid black line), and local rocking curves (green circles and lines) from a selected pixel at the center of the working region on the diamond (111) crystal plate. Figure [\[fig:rc_abc\]](#fig:rc_abc){reference-type="ref" reference="fig:rc_abc"}(a) shows rocking curves for the 300-\(\mu\)m-thick crystal mounted using non-perforated CVD diamond springs (as shown in Fig. [\[fig:asmb\]](#fig:asmb){reference-type="ref" reference="fig:asmb"}(a)). The FWHM of the local rocking curve matches that of the theoretical curve (25 \(\mu\)rad) while the total curve exhibits an additional broadening. This observation suggests that the broadening of the total rocking curve is due to the presence of defects in the crystal (i.e., intrinsic strain). The defects are located outside the working region of the crystal as will be shown further. Figure [\[fig:rc_abc\]](#fig:rc_abc){reference-type="ref" reference="fig:rc_abc"}(b) similarly shows rocking curves for the 100-\(\mu\)m-thick crystal mounted using perforated CVD diamond springs (as shown in Fig. [\[fig:asmb\]](#fig:asmb){reference-type="ref" reference="fig:asmb"}(b)). In this case, both the total rocking curve and the local rocking curve exhibit an additional broadening compared to the FWHM of the theoretical rocking curve for the 100-\(\mu\)m-thick crystal (26 \(\mu\)rad). Figure [\[fig:rc_abc\]](#fig:rc_abc){reference-type="ref" reference="fig:rc_abc"}(c) shows rocking curves for the 100-\(\mu\)m-thick crystal mounted using non-perforated CVD diamond springs. The broadening of the local rocking curve becomes somewhat larger, which suggests that the use of perforated springs is preferable for minimization of strain. However, this observation needs an additional verification, since there could be a variation in the local rocking curve from one pixel to another. The values of FWHM of the measured total and local rocking curves and those of the theoretical rocking curves are given in Table [1](#tab:fwhm){reference-type="ref" reference="tab:fwhm"} for each case. The FWHM of the total rocking curve (36 \(\mu\)rad) is about the same for (b) and (c), which indicates that the broadening is dominated by crystal defects and that this parameter is not very sensitive to mounting strain. The broadening is smaller than the intrinsic angular width (Darwin width) of the reflection (\(\simeq\) 23.1 \(\mu\)rad). For certain applications of diamond in X-ray optics this validates the quality of the entire crystal and the mounting method using either springs. However, in our case the desired level of broadening should be less than or comparable with the angular divergence of the XFEL beam (1-2 \(\mu\)rad rms) to minimize disturbance of the radiation wavefront. Therefore, a more detailed characterization of the working crystal region is required. [\[tab:fwhm\]]{#tab:fwhm label="tab:fwhm"} ## Analysis of the rocking curve topographs The comparative analysis of the widths of the total and local rocking curves does not show a complete picture since it does not take into account a distribution of the rocking curve parameters over the crystal plate. A more sensitive indicator of the crystal strain is a local change in the lattice parameter or a tilt of the Bragg plane, which lead to a shift in the peak position of the local rocking curve. In our diffraction imaging analysis this shift is represented by the center of mass of the local rocking curve (COM). Figure [\[fig:maps1e\]](#fig:maps1e){reference-type="ref" reference="fig:maps1e"} shows maps (topographs) of the rocking curve COM and FWHM for the 300-\(\mu\)m-thick plate mounted using the non-perforated springs (a), the 100-\(\mu\)m-thick plate mounted using the perforated springs (b), and the 100-\(\mu\)m-thick plate mounted using the non-perforated springs (c). These maps resemble the white-beam topographs given above. In particular, higher crystal quality in the working region and the diagonal lines (note higher contrast in FWHM) representing growth sector boundaries are clearly observed. The COM topographs indicate that a substantial strain is present in the upper part of the crystals. In the COM topographs for (a) and (b) the contrast is predominantly due to the intrinsic strain (presence of defects). In the case (c) the COM gradient is stronger, which indicates the presence of mounting-induced strain. The topographs of the working region are shown separately in Fig. [\[fig:maps1e_wkr\]](#fig:maps1e_wkr){reference-type="ref" reference="fig:maps1e_wkr"}. The distribution of the FWHM is fairly uniform across the working region in each case. The COM topographs for (a) and (b) are again quite uniform, while in the case (c) a distinct gradient is observed. The topographs of the working region illustrate that for the 300-\(\mu\)m-thick plate mounting it is acceptable to use non-perforated springs (Fig. [\[fig:maps1e_wkr\]](#fig:maps1e_wkr){reference-type="ref" reference="fig:maps1e_wkr"}(a)). However, for the 100-\(\mu\)m-thick plate mounting use of the non-perforated springs results in a higher level of strain (Fig. [\[fig:maps1e_wkr\]](#fig:maps1e_wkr){reference-type="ref" reference="fig:maps1e_wkr"}(c)) compared to the case of perforated springs (Fig. [\[fig:maps1e_wkr\]](#fig:maps1e_wkr){reference-type="ref" reference="fig:maps1e_wkr"}(b)). A simple statistical analysis of the rocking curve parameters was performed across the working region in each case. The values of standard deviation of the COM, standard deviation of the FWHM, and the average value for the FWHM are given in Table [2](#tab:wkrstat){reference-type="ref" reference="tab:wkrstat"}. The largest variation of the local rocking curve peak position (COM) was for the 100-\(\mu\)m-thick plate mounted using non-perforated CVD springs (3.5 \(\mu\)rad). This value is still small compared to the average FWHM (31 \(\mu\)rad), which ensures a reasonable performance of the crystal in Bragg diffraction. The use of the perforated diamond CVD springs yields further minimization of the mounting-induced strain as reflected by the reduced standard deviation of the COM of the local rocking curve (1.5 \(\mu\)rad). [\[tab:wkrstat\]]{#tab:wkrstat label="tab:wkrstat"} # Conclusions In summary, all-diamond X-ray optical assemblies holding type IIa diamond (111) crystal plates were fabricated for the beam-multiplexing diamond monochromator at the LCLS. It is demonstrated how requirements on crystal quality and crystal mounting lead to an advanced application of diamond materials in XFEL optics. Two crystal plates of 100 \(\mu\)m and 300 \(\mu\)m thicknesses with low concentration of defects in a region of 5\(\times\)`<!-- -->`{=html}2 mm\(^2\) were selected using white-beam X-ray topography. A dedicated crystal mounting method was developed for minimization of mounting strain. The mounting strain was evaluated using double-crystal X-ray topography in the rocking curve imaging mode. The variation of the rocking curve peak position over the working region of the crystal plates was found to be substantially smaller than the angular width of the curve and comparable with the angular divergence of the XFEL beam. With these assemblies installed in the beam-multiplexing monochromator at the LCLS, the capability of splitting the XFEL beam into a pink and a monochromatic branch was demonstrated. It was found that the measured bandwidth and throughput of the monochromator closely match the theoretical values, and the resulting beam profiles are only minimally disturbed.

{'timestamp': '2014-01-24T02:04:24', 'yymm': '1401', 'arxiv_id': '1401.5879', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5879'}

# Introduction The physical effects of spatial confinement on the properties of ring polymers is important to the physical understanding of the entropy-driven segregation of the two bacterial daughter chromosomes upon cell division and the structure of eukaryotic metaphase chromosomes [@hancock-eu-crowding; @mirny1; @mirny2]. For rod-like bacteria cells such as *E. coli*, *Bacillus subtilis*, or *streptobacillus* a directed motion and segregation of duplicated chromosomes along the cell axis is detected after DNA replication, see, for instance, Ref. . In eukaryotes, upon decondensation of the chromosomes in a strongly limited space inside the nucleus, the existence of chromosomal territories indicates an ultra-slow polymer mixing dynamics. Knotting of DNA molecules in tight spaces inside viral capsids another example of external polymer confinement in biology. *In vitro*, the elongation and compaction of long DNA molecules confined in nano-channels upon increasing fraction of the crowding agent was indeed detected. Internal polymer confinement *in vivo* is due to macromolecular crowding, which enforces DNA condensation in bacterial cells where the volume fraction occupied by crowding macromolecules such as RNA, ribosomes, or other biomacromolecules reaches \(\phi=30\dots35\%\). The abundance of crowding agents effects a viscoelastic environment that severely alters the diffusional dynamics of endogeneous cytoplasmic granules and of submicron tracers inside living cells. Concurrently the internal dynamics of polymers and the macromolecular association kinetics inside biological cells are dramatically changed. The effect of various polymeric crowders on the opening-closing dynamics of DNA hairpins has recently been experimentally probed in Ref. . Crowding can also facilitate phase separation and compartmentalisation of the bacterial cytoplasm. In theoretical models, inert spherical obstacles are often used to mimic highly crowded interiors of bacterial and eukaryotic cells. Crowding particles cause effective interactions between the polymer segments of the same chain and between the two chains in confinement, as studied in the present paper. From the theoretical perspective, overlapped segments of long polymer chains experience entropic repulsion scaling with the number of overlapping polymer blobs. In a dense polymer melt the entanglements of the chains also slow down the polymer dynamics. In the presence of obstacles, the extension and dynamics of ring polymers on the lattice was analysed in Refs. . A number of simulations studies of polymers under external confinement in various geometries appeared in the literature in recent years [@ital-plane; @ital-sphere; @ital-review; @koreans12; @likos12; @reich13rings]. In particular, the size scaling of ring polymers in dense melts was analysed by computer simulations in Ref. . As the concentration of rings \(c\) grows and the effective volume available for their expansion decreases, the scaling exponent for the radius of gyration \(\left< R_g^2\right>\simeq n^{2\nu}\) decreases from \(\nu=3/5\) to \(\nu\approx0.3\), mirroring impeded polymer extension. Neighbouring rings in dense melts thus induce a spherical caging effect, and their dimension was shown to scale as \(\left< R_g^2 \right>\sim c^{-0.59}\) in terms of the ring concentration. The segregation of semi-flexible macromolecules in nano-channels was shown theoretically in Ref. . Ring polymers in confinement were successfully used to model the bacterial chromosome and to rationalise the implications of supercoiling for the contact maps of eukaryotic interphase chromosomes. More specific with respect to our present study, the entanglement propensity of ring and linear polymers under external cylindrical confinement with respect to the phenomenon of DNA homologous recombination were analysed recently in Ref. . It was shown that linear chains penetrate into one another significantly easier than ring polymers. Finally, the threading of ring polymers inside a polymer gel was recently studied by simulations. The statistical properties of linear and ring polymers in the presence of crowding effects were considered in a number of theoretical and simulations studies. In particular, the behaviour of single knotted polymer rings on a regular lattice of obstacles was simulated in Refs. . For random-loop and self-avoiding polymers in the presence of crowding the computer modelling in Refs.  was shown to give rise to a non-monotonic dependence of \(\left< R_g^2\right>\) on the volume fraction occupied by the crowders, \(\phi\), featuring a slight minimum in the chain dimensions at \(\phi\approx0.2\). Related to this, the rates of chemical diffusion-limited reactions in molecularly crowded media in confined environments was shown to reveal a maximum at \(\phi\approx0.2\). In a biophysical context, the effect of crowding on gene regulation was studied with respect to facilitated diffusion and target search on DNA by DNA-binding proteins in Refs. . The translocation of polymers between the two reservoirs with crowders of equal and non-equal sizes has also been rationalised by simulations. The implications of crowding environments inside nano-channels have also been recently examined by simulations. Finally, the effects of crowders on the looping probabilities of polymer chains in the presence of external confinement and molecular crowding is also important. Below, we take a step further and analyse the joint effect of external confinement and internal crowding for two unknotted ring polymers in a model rod-shaped bacteria. More concretely, we perform molecular dynamics simulations for two polymer rings confined in a cylindric volume in the presence of mobile crowders, which are subject to the same thermal bath, as illustrated in Fig. [\[fig-scheme-chains\]](#fig-scheme-chains){reference-type="ref" reference="fig-scheme-chains"}. We analyse how the entropic repulsion of these thermally agitated ring polymers becomes altered under these crowding conditions. The paper is organised as follows. In the next Section, we present the details of the simulations model. In Section [3](#sec-results){reference-type="ref" reference="sec-results"} we rationalise the effects of external cylindrical confinement and internal confinement by the crowding obstacles. The main results for the static properties of the mutual overlap of the polymer rings and their dynamic characteristics are presented. In Section [4](#sec-conclusions){reference-type="ref" reference="sec-conclusions"} we discuss the basic results and their implications to the biological system and with respect to polymer physics. # Model and implementation of the simulations ## Polymer chains The standard finitely extensible non-elastic (FENE) potential is used to model the interactions between the monomers in our polymer chains in the bead-spring coarse-grained model of the DNA molecule, namely, \[U_{\mathrm{FENE}}(r)=-\frac{k}{2}r_ {\mathrm{max}}^2\ln\left(1-\frac{r^2}{r_{ \mathrm{max}}^2}\right).\] Here \(k\) is the spring constant acting between nearest-neighbour beads and \(r_{ \mathrm{max}}\) is the maximum allowed separation between neighbouring monomers. The total number of monomers in the ring polymers varies in the simulations in the range \(n=60\dots350\). Excluded-volume interactions between the polymer segments are introduced by the truncated Lennard-Jones repulsion (Weeks-Chandler-Andersen potential), that is, \[U_{\mathrm{LJ}}(r)=\left\{\begin{array}{ll}4\epsilon[(\sigma/r)^{12}-(\sigma/r)^{6}] +\epsilon, & r<2^{1/6}\sigma\\[0.2cm] 0, & \mbox{ otherwise}\end{array}\right.\] We here introduced the monomer-monomer distance \(r\), \(\sigma\) is the monomer diameter, and \(\epsilon\) is the strength of the potential. We set \(k=30\), \(r_ {\mathrm{max}}=1.5\sigma\), and \(\epsilon=1\), where all energies are measured in units of the thermal energy \(k_BT\), and distances are measured in units of \(\sigma\). Analogous repulsive 6-12 Lennard-Jones potentials parameterise the chain-obstacle, chain-wall, and crowder-wall contact interactions. Along the axial \(z\) coordinate in our cylindrical geometry a harmonic potential is applied once a monomer attempts to move outside of the cylinder at \(z<0\) or \(z>L\). The dynamics of the position \(\mathbf{r}_i(t)\) of the \(i\)th monomer in a polymer chain is described by the Langevin equation \[\begin{aligned} \nonumber m\frac{d^2\mathbf{r}_{i}(t)}{dt^2}&=&-\sum_{j=1,j\neq i}^{n}\nabla[U_{\mathrm{LJ}} (\mathbf{r}_i-\mathbf{r}_j)+U_{\mathrm{FENE}}(\mathbf{r}_i\mathbf{-r}_j)]\\ &&-\sum_{j=1}^{N_{cr}}\nabla U_ {\mathrm{LJ}}(\mathbf{r}_i-\mathbf{r}_{cr,j})- \nabla U_ {\mathrm{LJ}}(\mathbf{r}_i-\mathbf{R}_{\mathrm{cyl}})-)+\mathbf{F}(t). \end{aligned}\] Here \(m\) is the monomer mass, \(monomer velocity, and\)F(t)\(represents Gaussian delta-correlated noise with\[-correlations\)\<F(t)F(t')\>=6 equations are used for the dynamics of the positions \(\mathbf{r}_{cr,j}\) of the crowding molecules in the presence of the confining cylinder at \(\mathbf{R}_{ \mathrm{cyl}}\). Similarly to the procedure described in Ref. , we implement the velocity Verlet algorithm with the characteristic integration time-step of \(\Delta t=0.01\). The monomer size is set to \(\sigma=4\) nm determining the chain thickness that stays constant for different ring lengths simulated below. This thickness represents the effective physical DNA diameter including hydration water shells and electrostatic effects. Our approach thus differs from that taken in Ref. , where the mixing of ring polymers of different lengths without a crowding agent was studied and the polymers were assumed to become thinner as they get longer. This assumption was used in Ref.  to keep a constant volume fraction \(\phi_p\) of the polymer chains \(V_p\) in the simulation box of volume \(V\), and it is estimated that \]\phi_p=\frac{\mbox{DNA volume}}{\mbox{\emph{E. Coli} volume}}=\frac{V_p}{V}\sim1 \dots5,\[ depending on the DNA thickness (the bare DNA diameter plus the electrostatic repulsive salt-dependent shell around it). We present the ring-ring contact number and overlap distance of rings for the fixed \(\phi_p\) in Fig. [\[fig-contacts-final\]](#fig-contacts-final){reference-type="ref" reference="fig-contacts-final"}. The equilibration time of the polymer rings in our simulations depends on their length, the cell cylinder radius \(R\), and the volume fraction \(\phi\) of crowding particles in the simulation box. For typical parameters of the ring length \]l=n\sigma\sim200\sigma\[ and \(R=4.5\sigma\) used in simulations below, the chains equilibrate after \(\sim4 \times10^6\) simulation steps in the absence of crowders. The ring equilibration time grows with the chain length, and longer measurement times are required in order to sample the conformations of the polymer chains. The equilibration time also grows with the volume fraction \(\phi\) of crowders due to the slow-down of the polymer dynamics, see below. ## Crowding and confinement We distinguish two types of volume confinement for the polymer chains: external confinement by the cylindrical cell walls and internal confinement by mobile crowding obstacles. The model cell in our simulations is represented by an impenetrable cylinder of length \(L=35\sigma\) and radius \(R=3.5\sigma\dots5.5 \sigma\). By the internal confinement we mimic the highly poly-disperse soup of various proteins, RNA, cytoskeletal elements, and organelles in the cell cytoplasm. The crowders in the bacterial cytoplasm have an average molecular weight of \(\mathrm{ MW}\approx40\dots67\) kDa and diameter of 4\(\dots\)`<!-- -->`{=html}8 nm. We neglect here the poly-dispersity in crowder sizes observed in real cells and for simplicity assign to the crowders the same size \(\sigma_{cr}=\sigma\) as for the chain monomers. The crowders are simulated as spherical particles of unit mass (similarly to the polymer bead), with systematically varying volume occupancy \(\phi\). Each polymer monomer therefore corresponds to \(\approx12\) base pairs of the double-stranded DNA and \(\mathrm{MW} \approx12\times0.66\) kDa. Every step in our simulations then corresponds to a real time of \(\tau_0=\sigma\sqrt{12\times660~\mathrm{Da}/(k_BT)}\approx0.23\) ns. Simulating the crowders as particles with the more realistic value of MW of 67 kDa will slow down the crowder and polymer dynamics, renormalising the elementary simulation time unit to \(\tau_0\approx0.66\) ns. Varying the volume fraction of crowders in the simulations in the range \(0<\phi \lesssim0.3\) we mimic the response of a cell to the changes in external osmolarity, exerting a pressure on the outer cell membrane causing dehydration (osmotic upshift). This volume fraction is computed per free solution volume, i.e., \]\phi=\frac{V_{cr}}{V-V_{p}}=\frac{N_{cr}v}{\pi R^2 L-nv},\[ where \(v=4\pi(\sigma/2)^3/3\) is the volume of one chain monomer or of one crowding particle and \(N_{cr}\) is the number of crowding particles in the box. For the chain length \(n=200\) and cell length \(L=35\sigma\) the volume fraction of the two polymer rings is \(\phi_p\approx0.155\), \(0.094\), and \(0.063\) (close to the DNA crowding in *E.coli* ) for the respective cylinder radii \(R=3.5\sigma\), \(4.5\sigma\), and \(5.5\sigma\). We consider only excluded volume interactions according to the above-mentioned interaction potentials and neglect other interactions within the ring polymers, including electrostatic interactions. The latter can be of importance for tightly bent and circular DNAs, particularly at low-salt, weak-screening conditions. Our model also neglects effects of hydrodynamic interactions (both for rings and crowders), which can alter short-time polymer dynamics, but should not affect the static overlap properties of the rings. Polymer relaxation under confinement with and without hydrodynamic interactions was studied by computer simulations. For the relaxation time \(\tau_R\) of a polymer ring consisting of \(n\) monomers in a long cylindrical pore of radius \(R\) the relation \(\tau_R\sim n^2R^{0.9}\) was predicted. ## Ring contacts and decay of correlations For each simulation step \(t\) we determine the number of contacts \(N_{AB}(t)\) between the two ring polymers as follows. Each monomer is surrounded by a sphere with contact radius \(r_c=1.25\dots2\sigma\) that defines the overlap volume. If the centre of mass of a monomer of another chain stays within this contact sphere for the contact time \(t_c\), the contact is recorded as established. The time \(t_c\) is a measure of the internal dynamics of two intermingled rings. Within this time scale, the change in distances between contacting monomers should be smaller than \(r_c\). This validates the choice of the temporal and spacial thresholds for counting the number \(N_{AB}\) of ring-ring contacts. The distance \(r_c\) represents the 'radius of action' within which the monomers are supposed to be involved in some physical interactions. For the DNA, this can be electrostatic or protein-mediated inter-molecular contacts. Clearly, the results of counting the number \(N_{AB}\) of contacts depends on the threshold distance \(r_c\) (for comparison, the choice of \(r_c=1.5 \sigma\) was used in Ref. ). We analyse the dependence of the contact number on the contact distance \(r_c\) in Fig. [\[fig-rc-all\]](#fig-rc-all){reference-type="ref" reference="fig-rc-all"} below. The average \(\left<N_{AB}\right>\) is computed via averaging over various polymer configurations after the system reached its equilibrium. The contact volume \(V_{AB}\) of the rings is estimated as \(\left<N_{AB}\right>\) multiplied by \(\sim 1/2\) of the volume of the contact sphere, \(\left<V_{AB}\right>\approx\left<N_{AB} \right>4\pi(r_c/2)^3/3\). In addition to the three-dimensional inter-chain contact probability, scaling with \(\left<N_{AB}\right>\), we compute the one-dimensional mutual overlap length of two rings along the \(z-\)axis of the confining cylinder, \(\left<l_{AB}\right>\). Note that we consider only torsionally relaxed rings, with no effects of super-coiling. The latter would result in more branched and topologically complex polymer structures, likely with more extensive contacts. Following Ref. , we define the auto-correlation function (ACF) of ring-ring contacts via the contact number as follows \]\label{eq-ACF} \mathrm{ACF}(\Delta)=\frac{\left<N_{AB}(t+\Delta)N_{AB}(t)\right>-\left<N_{AB}(t+ \Delta)\right>\left<N_{AB}(t)\right>}{\left<N_{AB}(t)^{2}\right>-\left<N_{AB}(t) \right>^2}.\[ The averaging \(\left<\dots\right>\) is performed over the times times \(t\) along the generated trace \(N_{AB}(t)\) with the corresponding lag time \(\Delta\). The ACF characterises the decay of correlations in the overlap number of rings. The equilibration time in all our simulations is at least 50 times longer than the correlation time of the corresponding ACF for the inter-chain contact number for the chosen parameters. An additional quantity characterising the ring-ring overlap is the relative position of their centres of mass, \]\Delta z_{CM}=z_{CM,A}-z_{CM,B}.\[ From the corresponding probability density \(p_{AB}(\Delta z_{CM})\) along the cylinder axis we compute the free energy of the overlap of the two rings in terms of \]F(z_{CM})=-k_BT\log[p_{AB}(z_{CM})]\[ in the Shannon sense. # Results {#sec-results} ## Dimensions and contacts of polymer rings We verified that the extension of an unconfined ring polymer scales with its length as \]\left<R_{g}^2(n)\right>\sim n^{6/5},\[ consistent with the results reported in Ref.  (not shown). Our polymer rings are very flexible, the effective persistence length \(l_p\) being of the order of the monomer size (data not shown). Because of this extreme chain flexibility, we cannot analyse the implications of confinement onto the chain persistence (as compared to Ref.  where ring polymers were shown to stiffen substantially in tight confinement). Under the cylindrical confinement, the ring size \(\left<R_g^2\right>\) naturally reaches a saturation for long chains. Once the chain dimensions overcome the size of the cylindrical cell, the polymer starts to folds on itself and its apparent scaling exponent \(\nu\) decreases. The initial ring configurations at \(t=0\) generated in the simulations are well-separated, positioned at the opposite sides of the confining cylinder. They exhibit a fast initial relaxation followed by a roughly exponential relaxation dynamics. At the later stages, when the polymers experience external confinement by the cylinder and the other ring, a non-exponential relaxation dynamics sets in. The spectrum of chain fluctuations in frequency space in the presence of external confinement and crowding becomes altered as well. The general trend is that the instantaneous number of ring-ring contacts \(N_{AB} (t)\) fluctuates strongly in the course of the simulations, compare Fig. [\[fig-NAB-crazy\]](#fig-NAB-crazy){reference-type="ref" reference="fig-NAB-crazy"}. This trend is the same as in recent simulations for a similar system presented in the Supplementary Material of Ref. . We observe that in the presence of crowders the ring-ring separation becomes more pronounced, and the probability density function \(\mathrm{PDF}(N_{AB})\) of their contact numbers exhibits a peak at \(\left<N_{AB}\right>=0\). The spread of \(N_{AB}\) is slightly more localised in the presence of crowders, but both at crowded and non-crowded conditions the distributions \(p(N_{AB})\) have long tails, as evidenced in Fig. [\[fig-n-pdf\]](#fig-n-pdf){reference-type="ref" reference="fig-n-pdf"} A. The relative centre-of-mass position of the two rings, \(\Delta z_{CM}\), shows a larger spread in the presence of crowders, see Fig. [\[fig-n-pdf\]](#fig-n-pdf){reference-type="ref" reference="fig-n-pdf"}B. ## Ring swapping For some choices of the volume and the aspect ratio of the confining cylinder as well as for shorter polymer lengths, the directed distance \(\Delta z_{CM}\) between the centres of mass of the two ring polymers exhibits clear alternations between two states while the rings are well separated near the ends of the confining cylinder, see Fig. [\[fig-NAB-XCM\]](#fig-NAB-XCM){reference-type="ref" reference="fig-NAB-XCM"}. For such conditions, the diffusion times of the rings along the cylinder are relatively short, so they can pass one another and swap positions. At time instances when the rings are well separated the number of ring-ring contacts is minimal, while at almost vanishing centre of mass separation, \(z_{CM}\sim0\), the overlap of the rings is maximal and thus typically the \(N_{AB}(t)\) traces are peaking at these instances, see the time series of \(\Delta z_{CM}(t)\) and \(N_{AB}(t)\) shown in Fig. [\[fig-NAB-XCM\]](#fig-NAB-XCM){reference-type="ref" reference="fig-NAB-XCM"}. Note that the centre of mass distance \(|\Delta z_{CM}|\leq L\), and the case of \(\Delta z_{CM}>0\) corresponds to the situation when ring A is located on the left half and ring B on the right half of the confining cylinder. This type of dynamics is reminiscent of the periodic tumbling of polymers in shear flows, characterised by configurations with large extensions alternating with states of strong chain contraction, see, e.g., the studies reported in Refs. . For these conditions of well separated rings, the distribution of residence times \(t_r\), that each ring is situated in one of the two well-separated states close to the cylinder ends, is shown in Fig. [\[fig-XCM-dist\]](#fig-XCM-dist){reference-type="ref" reference="fig-XCM-dist"} for different crowding fractions \(\phi\). The mean residence time \(\langle t_r\rangle\) extracted from these histogram is the characteristic time scale for the internal ring swapping dynamics. As shown in the inset of Fig. [\[fig-XCM-dist\]](#fig-XCM-dist){reference-type="ref" reference="fig-XCM-dist"}, \(\langle t_r\rangle\) mildly increases with increasing \(\phi\). The maximum of the residence time histograms in Fig. [\[fig-XCM-dist\]](#fig-XCM-dist){reference-type="ref" reference="fig-XCM-dist"} shifts at higher crowding fractions to larger values because of the associated slower polymer dynamics. As illustrated in Fig. [\[fig-FE-2wells\]](#fig-FE-2wells){reference-type="ref" reference="fig-FE-2wells"} the effective free energy for mixing the two rings has a double-well shape. The height of the barrier separating the two minimum states amounts to several \(k_BT\) for the parameters used in the simulations. As the residence times \(t_r\) in these separated ring states increases, the height of the free energy barrier between them decreases. This is due to a slower polymer dynamics at high crowding fractions, as discussed below. It also demonstrates that the free energy landscape is no true equilibrium measure, as known from the theory of polymer translocation. Longer rings squeezed into the same confining cylinder reveal a slower swapping dynamics and the residence times in well-separated states grow until no swapping at all can be observed during the simulation time. Likewise, the exchange of rings in the simulation box is prohibited for smaller cylinder radii \(R\) (not shown). ## Correlations of ring-ring contacts As demonstrated in Fig. [\[fig-NAB-crazy\]](#fig-NAB-crazy){reference-type="ref" reference="fig-NAB-crazy"}, the number of ring-ring contacts fluctuates strongly and irregularly. To find a typical time-scale for this variation, we compute the ACF ([\[eq-ACF\]](#eq-ACF){reference-type="ref" reference="eq-ACF"}) of the ring-ring contact number from the \(N_{AB}(t)\) time traces. We start with the crowding-free case \(\phi=0\). The resulting curves in Fig. [\[fig-ACF\]](#fig-ACF){reference-type="ref" reference="fig-ACF"} show a fast relaxation at short times and turn to a nearly exponential decay at intermediate lag times \(\Delta\). At long times, the ACF drops to zero, indicating a complete loss of correlations. Some fluctuations of the \(\mathrm{ACF}(\Delta)\) at \(\Delta\to\infty\) indicate insufficient statistics in the calculation of the time average ([\[eq-ACF\]](#eq-ACF){reference-type="ref" reference="eq-ACF"}). From Fig. [\[eq-ACF\]](#eq-ACF){reference-type="ref" reference="eq-ACF"}, we observe that the initial decay of the ACF is slower for smaller cylinder radii, as expected. This is due to a larger space fraction in the simulation box being filled by the polymer monomers so that their motions get restricted to a larger extent, with many chains' moves being prohibited. For longer rings confined in the same cylinder, the ACF decays slower with the lag time \(\Delta\), again due to a smaller space available for the chains (not shown). Note that the intermediate-time decay exhibits comparable slopes in the logarithmic plot of Fig. [\[eq-ACF\]](#eq-ACF){reference-type="ref" reference="eq-ACF"}. ## Contacts and overlap of polymer rings: crowding effects Let us now study the effects of the internal confinement due to crowding in more detail. We first consider a single ring polymer under the cylindrical confinement in the presence of crowding agents. The results for the mean squared gyration radius \(\langle R_g^2\rangle\) are shown in Fig. [\[fig-RG2-single\]](#fig-RG2-single){reference-type="ref" reference="fig-RG2-single"}. We observe that the component of the radius of gyration measured along the cylinder axis is a slowly decreasing function of \(\phi\). Crowding particles thus act as a depletant, that effects ring shrinkage. For a less severe external confinement (larger cylinder radius \(R\)), we observe that the ring is more confined along the cylinder axis, but simultaneously more extended in the cylinder cross-section (\(x-y\) plane), as shown in Fig. [\[fig-RG2-single\]](#fig-RG2-single){reference-type="ref" reference="fig-RG2-single"}. Here we do not elaborate on the variation of the Flory scaling exponent \(\nu\) of the gyration radius for a single ring as function of the external confinement and crowding (for such results see, e.g., the results reported in Ref. ). In the following we concentrate on the overlap properties of two polymer rings in the cylindrical simulations cell. At higher fractions \(\phi\) of crowders the correlation time of maintainig the established contacts between polymer rings increases due to the slower polymer dynamics, following a larger effective viscosity in a denser soup of crowders, i.e., the Rouse polymer dynamics becomes effectively slowed down by surrounding crowding particles. Concurrently, the same effect is responsible for a slower decay of the contact autocorrelations at higher values of \(\phi\), as shown in Fig. [\[fig-acf-crowd\]](#fig-acf-crowd){reference-type="ref" reference="fig-acf-crowd"}. The associated correlation time \(\tau_C\) extracted from an exponential fit of the \(\mathrm{ACF}(\Delta)\) curves exhibits the power-law dependence \]\tau_C(\phi)\sim\phi^{3/2}\[ on the fraction of crowders, see the inset of Fig. [\[fig-acf-crowd\]](#fig-acf-crowd){reference-type="ref" reference="fig-acf-crowd"}. The \(3/2\) exponent indicates that the changes due to crowding are indeed a volume effect. Note that the single-ring relaxation time \(\tau_R\) should not be confused with \(\tau_C\) for the ring-ring contacts. The dependence on the presence of mobile crowders on the three-dimensional contact characteristics and the effective one-dimensional overlap properties of polymer rings of fixed length are analysed in Fig. [\[fig-contacts-1D-crowding\]](#fig-contacts-1D-crowding){reference-type="ref" reference="fig-contacts-1D-crowding"}. We observe that the average number of ring-ring contacts \(\left<N_{AB}\right>\) increases significantly with the decrease of the cylinder radius \(R\), i.e., when the chains are forced into a stronger contact by the external confinement, see Fig. [\[fig-contacts-1D-crowding\]](#fig-contacts-1D-crowding){reference-type="ref" reference="fig-contacts-1D-crowding"}A. As function of the internal confinement due to crowding, in some situations the number of ring-ring contacts \(N_{AB}\) exhibits a weakly non-monotonic dependence, see, e.g., the blue symbols in Fig. [\[fig-contacts-1D-crowding\]](#fig-contacts-1D-crowding){reference-type="ref" reference="fig-contacts-1D-crowding"}B. For weaker external confinement (larger \(R\)) we observe a mildly increasing dependence while it is decreasing for the smaller cylinder radius. This behaviour indicates a tradeoff between crowding and external confinement. The effective overlap length of the rings along the cylinder axis is, in contrast, a very reproducible function with the functional relation \]\left< l_{AB}(\phi)\right>\approx\left< l_{AB}(0)\right>(1-\phi). \label{eq-ideal-mix}\(\) of the crowding fraction \(\phi\), compare Fig. [\[fig-contacts-1D-crowding\]](#fig-contacts-1D-crowding){reference-type="ref" reference="fig-contacts-1D-crowding"}A. This fact indicates a nearly ideal mixing of polymer monomers and crowding particles, as if the chain connectivity plays a minor role. The absolute values of \(l_{AB}\) for different cylinder radii vary only marginally. The decrease in Eq. ([\[eq-ideal-mix\]](#eq-ideal-mix){reference-type="ref" reference="eq-ideal-mix"}) can be understood from a shrinkage of individual ring polymers by the crowders, as rationalised for longitudinal ring dimensions in Fig. [\[fig-RG2-single\]](#fig-RG2-single){reference-type="ref" reference="fig-RG2-single"}B. We note a relatively small value of the overlap length at all crowding densities used in simulations. It is consistent with the results of Ref.  where, in the absence of crowders, a very limited inter-penetration and overlap of the two polymer rings was obtained. We also systematically examined the effect of the contact distance \(r_c\), defining the overlap of both polymer monomers and crowders, on the number of ring-ring contacts established in the simulations. We find that the average number of contacts naturally grows with \(r_c\), compare Fig. [\[fig-rc-all\]](#fig-rc-all){reference-type="ref" reference="fig-rc-all"}. We also note that the error bars somewhat increase with \(\phi\) and \(r_c\) but always stay smaller than the symbol size. Here and below, as proposed in Ref. , for the ring-ring contacts the error bars are computed with the blocking method introduced for correlated data sets in Ref. . The statistical effects of the polymer-crowder mixing are analysed in terms of their distributions in the simulation cell. We find that for weak and moderate crowding fractions there is an accumulation of crowders near the cylinder ends, as evidenced in Fig. [\[fig-pdf\]](#fig-pdf){reference-type="ref" reference="fig-pdf"}A. In turn, the polymer monomers are located preferentially off the middle of the cylinder, see Fig. [\[fig-pdf\]](#fig-pdf){reference-type="ref" reference="fig-pdf"}B. We also observe that at small \(\phi\) the crowding particles are effectively excluded from regions occupied by the polymers, thus facilitating ring-ring contacts. At stronger crowding, a peak of crowding particles in the middle of the cell (between the polymer rings) emerges, see the blue curve in Fig. [\[fig-pdf\]](#fig-pdf){reference-type="ref" reference="fig-pdf"}A. These mid-positioned crowding particles trigger an entropy-driven segregation of polymer rings, and their three-dimensional contact number \(N_{AB}\) decreases at larger \(\phi\) values (Fig. [\[fig-contacts-final\]](#fig-contacts-final){reference-type="ref" reference="fig-contacts-final"}B). The mixing properties of polymers and crowders can be probed by the cumulative probability distribution of their monomers shown Fig. [\[fig-pdf\]](#fig-pdf){reference-type="ref" reference="fig-pdf"}C. The decrease of \(N_{AB}(\phi)\) at high \(\phi\) is both due to a progressive emergence of crowders in between the polymers and a longitudinal shrinkage of each of the rings with \(\phi\). The ideal \((1-\phi)\) polymer-crowder mixing is realised at high \(\phi\), when the sum of the distributions of the polymer monomers and crowders is almost constant throughout the simulation cell, see Fig. [\[fig-pdf\]](#fig-pdf){reference-type="ref" reference="fig-pdf"}C. Note that in order to sample the polymer configurations equally well at varying crowding fractions, longer simulation times are usually required at high \(\phi\) values. ## Variation of the polymer length In the previous sections, most of the results presented were obtained for a constant polymer length of \(n=200\) monomers. To be able to scale up the our computations, we computed the ring-ring overlap length and contact number for longer polymers and larger simulation cells. The geometrical proportions of the confining cylinder, the aspect ratio \(R/L\), were kept constant and at the same time the ring length was adjusted so the polymer volume fraction \(\phi_p=V_p/V\) stays constant at \(\phi_p\approx0.09\). We observe that both the one-dimensional overlap length of the polymer rings and their three-dimensional contact number follow universal curves, after normalisation with respect to the number of monomers, i.e., for \(l_{AB}/n\) and \(N_{AB}/n\). These results are illustrated in Fig. [\[fig-contacts-final\]](#fig-contacts-final){reference-type="ref" reference="fig-contacts-final"}, which is the main result of this work. The ring-ring overlap length follows the \((1-\phi)\)-asymptote typical for the ideal mixing for all the parameters analysed in our simulations. In contrast, the number of contacts \(N_{AB}\) reveals a more delicate dependence. For small confining cylinders the contact number is a monotonically decreasing function (red symbols in Fig. [\[fig-contacts-final\]](#fig-contacts-final){reference-type="ref" reference="fig-contacts-final"}B). For larger simulation cells, the value of \(N_{AB}(\phi)\) exhibits a maximum at \(\phi\approx0.2\) (black and blue symbols in Fig. [\[fig-contacts-final\]](#fig-contacts-final){reference-type="ref" reference="fig-contacts-final"}B). This fraction \(\phi\approx20\%\) is reminiscent of the turning-point value for the non-monotonic dependencies on the fraction of crowding molecules mentioned in the Introduction, namely, of the dimensions of self-avoiding polymers and diffusion-limited chemical reactions. To compute \(l_{AB}/n\) and \(N_{AB}/n\), we averaged over \(M=2\times10^5\), \(2\times 10^5\), and \(8\times10^4\) simulation steps for cylinder radii \(R=3.5\sigma\), \(4.5\sigma\), \(5.5\sigma\), respectively. The simulation time on a standard 3-3.5 GHz core machine for each crowding fraction \(\phi\) presented in Fig. [\[fig-contacts-final\]](#fig-contacts-final){reference-type="ref" reference="fig-contacts-final"} is about 2, 3, and 10 days, respectively. To accumulate reliable statistical information about the ring-ring contacts at relatively large crowding fractions, particularly long simulations are required because of the slower dynamics of inter-ring mixing. Last, at the same volume fraction of crowders \(\phi\), one can expect crowding particles of larger sizes to cause stronger effects on mixing properties of polymer rings. # Conclusions {#sec-conclusions} Based on extensive Langevin dynamics simulations we analysed the behaviour of polymer chains of a circular topology in the presence of external confinement in a cylindrical geometry and internal crowding by molecular crowding agents. The size of the cylindrical confinement with respect to the monomer size and the length of the chains was chosen to represent the situation of two DNA rings in a typical bacillus cell. The crowding agents were represented by thermally agitated, off-lattice mono-disperse hard spheres. We found that the topological constraints restricting the polymer dynamics alter the response of partially intermingled circular polymers compared to linear chains. In highly crowded environments the polymer dynamics was demonstrated to be slowed down significantly. We found that high concentrations of crowding agents facilitate the spatial separation of ring polymers in cylindrical confinement. In addition, we quantified the extent to which the presence of crowding agents slows down the dynamics of the polymer-crowder system. The simulations for chains of varying length demonstrates that our model results are robust and in principle scalable to the dimensions of real bacterial cells. The effect of molecular crowding obtained above are applicable to de-mixing of genome-sized DNA molecules inside bacterial cells as well as to the behaviour of relatively short DNA plasmids confined in natural compartments inside eukaryotic cells. The abundance of macromolecular crowders also offers a robust and non-specific way to tune the amount of DNA-DNA contacts. The dynamics and spatial occurrence of the latter are vital for biological processes such as DNA-DNA recognition and DNA homologous recombination, when the search for the homologous DNA partner in a coil of a long DNA is to be performed. Note that it would be interesting to analyse how the DNA-polymer segregation takes place in bacteria with other than rod-like shapes, such as in nearly planar squarish or spherical bacteria, see the discussion in Ref. . The compartmentalisation of obstacles and polymer chains we observed can also have implications for ageing phenomena of bacterial cells. For *E. Coli* cells, for instance, a localisation of age-related protein aggregates in low-crowding regions near the cell poles and in between of the two DNA nucleoids was observed. This effect was recently quantified by computer simulations at higher degrees of polymer and DNA crowding inside the nucleoids that hinder the diffusion of these protein aggregates. JS thanks W. K. Kim for providing a computer code for analysis of ring swapping dynamics. The authors acknowledge funding from the Academy of Finland (FiDiPro scheme to RM), the German Research Council (DFG Grant CH 707/5-1 to AGC), and the German Federal Ministry for Education and Research (to JS).

{'timestamp': '2014-01-24T02:06:27', 'yymm': '1401', 'arxiv_id': '1401.5932', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5932'}

# Introduction demo file is intended to serve as a "starter file" for IEEE Computer Society journal papers produced under LaTeX using IEEEtran.cls version 1.8 and later. I wish you the best of success. mds December 27, 2012 ## Subsection Heading Here Subsection text here. ### Subsubsection Heading Here Subsubsection text here. # Conclusion The conclusion goes here. # Proof of the First Zonklar Equation Appendix one text goes here. # Appendix two text goes here. # Introduction This demo file is intended to serve as a "starter file" for IEEE conference papers produced under LaTeX using IEEEtran.cls version 1.8 and later. I wish you the best of success. mds December 27, 2012 ## Subsection Heading Here Subsection text here. ### Subsubsection Heading Here Subsubsection text here. # Conclusion The conclusion goes here. # Introduction demo file is intended to serve as a "starter file" for IEEE journal papers produced under LaTeX using IEEEtran.cls version 1.8 and later. I wish you the best of success. mds December 27, 2012 ## Subsection Heading Here Subsection text here. ### Subsubsection Heading Here Subsubsection text here. # Conclusion The conclusion goes here. # Proof of the First Zonklar Equation Appendix one text goes here. # Appendix two text goes here. # Introduction demo file is intended to serve as a "starter file" for IEEE journal papers produced under LaTeX using IEEEtran.cls version 1.8 and later. I wish you the best of success. mds December 27, 2012 ## Subsection Heading Here Subsection text here. ### Subsubsection Heading Here Subsubsection text here. # Conclusion The conclusion goes here. # Proof of the First Zonklar Equation Appendix one text goes here. # Appendix two text goes here. # Introduction demo file is intended to serve as a "starter file" for IEEE Computer Society journal papers produced under LaTeX using IEEEtran.cls version 1.8 and later. I wish you the best of success. mds December 27, 2012 ## Subsection Heading Here Subsection text here. ### Subsubsection Heading Here Subsubsection text here. # Conclusion The conclusion goes here. # Proof of the First Zonklar Equation Appendix one text goes here. # Appendix two text goes here. # Introduction implementations of quantum computers capable of solving problems at a useful scale generally involve at least many thousands of qubits. Whether the algorithm envisioned is based on the quantum circuit model, or on an adiabatic method, there are a number of physical requirements that constrain the design of a large-scale manufactured quantum device. The device architecture must facilitate precise individual qubit control, computationally interesting interaction between qubits, and high fidelity readout of qubit state. Of particular importance are the practical constraints that arise in acheiving these design goals while maintaining the carefully engineered environment required to implement a quantum algorithm. One of the advantages of an approach based on superconducting qubits is that a largely compatible classical electronics technology is known and available in the guise of single-flux-quanta based circuit architectures. The authors have previously presented the architecture, design, and operation of an SFQ based system for controlling and reading out a quantum annealing processor based on flux qubits at the core of D-Wave One system. The authors, as well as a number of other researchers, have gained some experience using this first generation processor. This experience has informed and guided the design of a second generation quantum annealing processor, the D-Wave Two system. In this paper we provide an overview of this processor's architecture and discuss some of the considerations and trade-offs involved in its design. At the root of the design problem is the operation of a quantum annealing processor based on superconducting flux qubits. Many of the processor building blocks are described by Harris, *et al.*. A number of the control terminals (biases) required by the qubits and inter-qubit couplers are discussed by Johnson *et al.*, and these are largely unchanged. The implementation of the control circuitry, the cross section of the fabrication process, and the number of devices used, are among the main differences between the two generations of D-Wave processors. The D-Wave One used current biased SFQ demultiplexer circuitry to address all \(N\) programmable devices on chip, requiring an asymptotically optimal \(O(\log(N))\) number of control lines. This circuitry was designed with consideration for the need to minimize static power dissipated on chip during programming, and to this end used very low bias supply voltages, and very low value shunt resistors. The predicted peak temperature of the junction shunts during programming, about \(500\,\mathrm{mK}\), was sufficiently low to ensure negligible thermally induced bit errors in DAC programming. We found, however, that it took of order \(1\,\mathrm{s}\) for this heat to dissipate sufficiently for the processor to return to \(\sim20\,\mathrm{mK}\), a temperature low enough to run the quantum annealing algorithm and obtain solutions to posed problems with appreciable probability. As typical computation time is \(\sim 20 \mathrm{\mu s}\), this was clearly an unacceptable amount of time to wait to run the algorithm after programming. D-Wave Two control circuitry was designed to eliminate static power dissipation completely and to simplify the design greatly with XYZ addressing scheme requiring \(O(\sqrt[3]N)\) lines. Though not logarithmic, this scaling is sufficiently weak to allow processors significantly larger than the one under discussion to be operated in our existing apparatus. Improvement in performance of the annealing algorithm was also a central design goal of the D-Wave Two. In a fixed temperature environment, performance can be improved by increasing the qubit energy scale. This can be done by decreasing qubit inductance and capacitance. Given constraints on dielectric permittivity and wiring geometry, this is most practically accomplished in our design by reducing the physical length of the qubit wiring. Reduction of length by a factor of two compared with D-Wave One was achieved by adding two metal layers to the fabrication process, for a total of 6 superconducting metal layers. This allowed for an increase of overall processor density by a factor of four. In the first section of this paper we will step back and discuss the requirements of the whole chip and give our rationale behind the chosen hardware graph topology (common between D-Wave One and Two) from the top-down perspective. We will then be in a position to present the chosen implementation of D-Wave Two control circuitry in bottom-up fashion. Read-out infrastructure will be described in a subsequent publication. # Circuit Topology[\[sec:top\]]{#sec:top label="sec:top"} D-Wave quantum annealing processors have evolved through a series of generations subject to the competing pressures exerted by computational complexity and practical implementation. They are designed to solve this problem: given *hardware graph* \(G\), minimize the following quadratic form over discrete variables \(s_i \in \{-1, +1\}\): \[E(\vec{\mathbf s}|\vec{\mathbf h}, \hat{\mathbf J})=\sum_{i\in nodes(G)} {h_i s_i} + \sum_{<i,j>\in edges(G)}{J_{ij} s_i s_j},\] with problem parameters \(h_i,J_{ij}\in \{-1,-7/8,... +7/8, +1\}\) for our current hardware. Our current hardware graph topology, which we named Chimera, was designed to satisfy a number of common-sense requirements related to its intended use for solving optimization problems, subject to a number of physical implementation constraints. ## Requirements 1. **Non-planarity:** We desire to tackle NP-complete problems, so non-planarity of the underlying graph is an important condition to making the corresponding Ising-spin problem NP-complete . A related motivation is that non-planarity is required to establish chains of qubits that cross each other. 2. **The ability to embed complete graphs:** To solve Ising spin glass problems with different topologies using a single processor, it must be possible to map, or *embed* the problem graph into the available hardware graph. This process typically involves using chains, trees, or other connected sub-components of the hardware graph (comprising *physical qubits*, strongly ferromagnetically coupled to each other) to represent a single node in the problem graph (*logical qubit*). In the language of graph theory, we want the hardware graph to have largest possible variety of problem graphs as its *minors* [@Vicky2008]. While embedding arbitrary graphs is computationally hard, one can nonetheless determine how large of a degree \(M\) complete graph \(K_M\) can be embedded in a given size hardware graph, thus guaranteeing that all graphs up to \(M\) nodes are embeddable using a straightforward prescription. 3. **The ability to incorporate on-chip control circuitry:** While a single qubit, or a handful of them, can be precisely controlled with dedicated analog lines driven by room-temperature electronics, integrating more than a few dozen qubits on a single chip requires some on-chip control circuitry. For example, there are currently six control "knobs" for every qubit, required to make the qubits robust to fabrication variability , and one "knob" per coupler. Where possible, we have designed our QA processors to use *static* flux biases applied to target superconducting loops in order to realize most of these knobs. The desired values of flux biases are programmed into individual control devices using a relatively small number of essentially digital control lines that carry signals generated at room temperature. These the control devices combine the functions of persistent memory and digital-to-analog conversion. We call these devices flux DACs, or \(\Phi\)-DAC s. From an architectural viewpoint, each \(\Phi\)-DAC is a relatively macroscopic object with a typical size of \(\sim 10\,\mathrm{\mu m}\). Having several of them attached to a single qubit sets a lower bound on qubit size and influences possible qubit shapes and hardware graph topologies. ## Constraints 1. **Limited qubit fan-out:** From an applications standpoint, the best option would be for each qubit to be connected to all others. However, directly implementing a complete \(K_M\) graph in hardware for an arbitrarily large \(M\) is impractical. Each qubit in our current design can be connected to only a relatively small number of other qubits \(\lesssim 10\) before non-ideal features arise in the qubit response and the coupling energy scale (compared to \(k_b T\), for example), becomes too small. 2. **Minimizing uncoupled qubit/coupler lengths:** To optimize qubit energy scales and coupling strengths, neither qubits nor couplers can be designed to span an arbitrary length (*e.g.*, full size of a large processor matrix). Ideally, all qubit length should be magnetically coupled to connected couplers, and all coupler length to connected qubits. 3. **Minimization of noise pick-up areas and cross-talks:** Flux qubits and couplers are rf SQUIDs, which can be quite sensitive to magnetic fields. Extreme care needs to be taken to minimize their pick-up of undesired disturbances, such as coupling to external flux noise sources or the unintended coupling (*cross-talk*) from a control line to a device it is not intended to control. 4. **2D chip integration:** While it would be nice to be able to grow a processor lattice in all three dimensions, in reality these lattices have to be implemented on the surface of 2D chips. Even if we imagine adding more metal layers to our fabrication process or 3D integration of several chips stacked on top of each other and passing quantum state between them (through, *e.g.*, superconducting backside vias), growing the processor graph along the *physical* third dimension will always be harder than along the 2D chip plane.[^1] 5. **Regularity and the notion of a "unit tile":** While it is in principle possible to arrange qubits in highly irregular structures, in practice, especially while designing chips not tailored to any specific problem graph structures which might arise in a concrete application, we find it convenient (to simplify the design and operation) to introduce the notion of a *unit tile*, which is a smaller structure that can be replicated in both dimensions of the chip plane. ## Chimera topology One of the features of a flux qubit is that (unlike, *e.g.,* qubits based on quantum dots or individual trapped ions) they are essentially macroscopic inductive loops interrupted by Josephson junction(s) and that these qubit body loops can be stretched and routed as needed. The same is true for qubit-to-qubit couplers, except that parametrically they tend to be lower inductance devices, and thus, shorter. With this in mind we examined different arrangements of qubit loops, and eventually settled on the Chimera unit tile topology (used in both D-Wave One and D-Wave Two processors), schematically depicted in Fig. [\[fig:chimeratop\]](#fig:chimeratop){reference-type="ref" reference="fig:chimeratop"}. Each unit tile consists of 8 qubits--4 horizontal and 4 vertical--with couplers between each horizontal/vertical pair. The unit tile is a complete bipartite graph \(K_{4,4}\). Unit tiles can be arranged into larger grid-like structures that fill a plane, and each horizontal qubit can be coupled to the corresponding qubits in the neighboring tiles to the left and right, while each vertical qubit can be coupled to those in the neighboring tiles above and below. How well does the Chimera topology satisfy the requirements and constraints given above? Consider the following: - The Chimera graph is non-planar. Assuming the ability to establish chains of qubits along rows and columns of the processor matrix, there is a straightforward approach to embed complete graphs up to \(4N\) nodes in a \(N \times N\) grid of unit cells (denoted as \(C_N\)). This is illustrated in Fig. [\[fig:chimeraKN\]](#fig:chimeraKN){reference-type="ref" reference="fig:chimeraKN"} for the case of \(N=4\). This approach can be validated based on the following observations: 1. Taking a single \(K_{M,M}\) tile and ferromagnetically coupling pairs of horizontal and vertical qubits along its diagonal (*contracting edges*, which connect them in graph-theoretical language) produces a complete graph \(K_M\). 2. Taking a \(2 \times 2\) array of complete bipartite graphs \(K_{M,M}\) and ferromagnetically coupling pairs of qubits in the same row/column produces a complete bipartite graph \(K_{2M, 2M}\). 3. Taking two complete graphs \(K_M\) and connecting them to two sides of a complete bipartite graph \(K_{M,M}\) produces a complete graph \(K_{2M}\)-every node in \(K_{2M}\) can be coupled to every other node either because they either belong to the same \(K_{M}\) and were coupled anyway or because they belong to different \(K_{M}\)s, in which case there is connection between them through the complete bipartite part. - The Chimera topology was designed to be interleaved with the required control circuitry (which is schematically represented by lighter shaded areas in Figure [\[fig:chimeratop\]](#fig:chimeratop){reference-type="ref" reference="fig:chimeratop"}). In the implementation under discussion, each square "plaquette" formed at the intersection of two qubits contains three \(\Phi\)-DAC s. Generally, the left (right) \(\Phi\)-DAC provides certain type of control to the vertical (horizontal) qubit, while the middle one controls the corresponding coupler, as schematically shown in the bottom-left corner of this diagram. - Almost all of the qubit length is coupled to couplers, and almost all of the coupler length is coupled to qubits, thus maximizing coupled signal strength. Also, implementing the qubit and coupler loops as long and narrow differential microstrip lines (in practice, over a superconducting ground-plane) minimizes noise and parasitic cross-talk pick-up. - Chimera unit tiles can be arranged into arbitrarily large 2D structures (limited only by fabrication yields, die size and available number of IO lines/die pads required to program all \(\Phi\)-DAC s). For example, the D-Wave One processor contained 128 qubits in a \(C_4\) grid (a \(4\times 4\) grid of 8-qubit tiles) and the D-Wave Two processor contains 512 qubits in a \(C_8\) grid (an \(8\times 8\) grid of 8-qubit tiles). While this approach can be generalized to an arbitrary \(K_{M,M}\) unit tile with \(2M\) qubits, our current implementation of \(M=4\) was chosen because (as will be seen later) all of its required \(\Phi\)-DAC s can be fit in a \(5 \times 5\) array of fixed-size plaquettes without too much wasted space, simplifying (manual) layout and managing overall design complexity. Another advantageous feature of the \(M=4\) size is that the number of \(\Phi\)-DAC s used for problem specification (one per qubit and one per coupler, thus giving a total of 32 per tile) is approximately balanced with the number of \(\Phi\)-DAC s used to make qubits robust against fabrication variations (4 per qubit in the D-Wave One generation, 5 in D-Wave Two, giving a total of 32 or 40 per tile, respectively). For smaller unit tile size majority of the \(\Phi\)-DAC s would be of the second variety. # Design and operation of a \(\Phi\)-DAC. The precision desired for setting problem parameters sets the requirements for the range and precision of individual \(\Phi\)-DAC s. Generally, our current implementation requires about 8 bits of dynamic range for individual \(\Phi\)-DAC s, with full ranges varying from several thousandths of a magnetic flux quantum (\(\mathrm{m\Phi_0}\)) to half a \(\Phi_0\) coupled into qubit or coupler control loops, depending on the \(\Phi\)-DAC type. To achieve this dynamic range while minimizing both total area occupied by control circuitry (thus minimizing qubit length and increasing qubit energy scales) and total number of wires needed for programming, in our current design we chose to implement most of our \(\Phi\)-DAC s as two-stage devices, of the kind schematically shown in Figure [\[fig:DAC1\]](#fig:DAC1){reference-type="ref" reference="fig:DAC1"}. Each of the DAC digits (referred to as "most significant" and "least significant" here, or "MSD" and "LSD") is implemented as a SQUID loop into which we can write and store some number of flux quanta \(m\), *e.g.*, \(-8 \lesssim m \lesssim 8\). Individual quanta can be added to or subtracted from the storage loop via an SFQ pulse source, depicted here as a Josephson junction; its structure and operation will be described in section [\[sec:vscrc\]](#sec:vscrc){reference-type="ref" reference="sec:vscrc"}. Both digit storage loops are magnetically coupled into an output device via an inductive ladder. A single flux quantum added to the MSD coil induces \(M_{MSD}/L_{MSD}*\Phi_0\) flux into the top ladder loop, and \(M_{MSD}/L_{MSD}*M_{OUT}/L_{loop0}^{total}*\Phi_0\) into the output device (\(L_{loop0}^{total}\) denotes total inductance of the MSD loop). The output flux increases proportinally with the number of flux quanta added, up to a maximum determined by the device parameters and addressing scheme. There is in general a nonlinear component associated with the junction inductances, but as long as these inductances are small compared to main loop inductances (true for our devices), this correction is negligible. In our example, if MSD loop can store up to 8 flux quanta of either polarity, it can provide 16 distinct values of output flux, or implement a 4-bit DAC. To increase precision, a second stage is added. Here, the effect of a single flux quantum in the LSD loop is further subdivided by a factor of \(L_{DIV}/L_{loop1}^{total}\). If this loop can also provide 16 distinct values of stored flux and the division ratio is 16, one MSD step will be further subdivided into 16 steps of LSD, and the two-stage device is an 8-bit DAC. In practice, of course, we want to guarantee both the total output range and the coverage of an MSD step by the LSD in the presence of fabrication variations, so we need to add some margin to the number of quanta that we can store in both loops. \(\Phi\)-DAC s with different numbers of digits and weights of each digit can be designed using the same principles, but we have found that this two-stage design is sufficient for almost all of our DACs[^2]. ## \(\Phi\)-DAC: Inductive storage and ladder Having covered the basic idea behind our \(\Phi\)-DAC s, we can present a more realistic layout of their implementation on the top of Figure [\[fig:DAC2\]](#fig:DAC2){reference-type="ref" reference="fig:DAC2"}. Large storage inductors (\(\sim 1\, \mathrm{nH}\)) are implemented as stacked spirals (blue and green in the figure), shown here wound in two metal layers, though our real layouts use four layer spirals, with \(0.25\,\mathrm{\mu m}\) line width and spacing design rules. The inductive ladder is implemented as two galvanically connected washers in the bottom metal layer (red), magnetically coupled to their two coils. The horizontal bar between them implements the shared inductance \(L_{DIV}\) of Figure [\[fig:DAC1\]](#fig:DAC1){reference-type="ref" reference="fig:DAC1"}. To minimize unintended coupling between DAC coils and other elements of the circuit, the whole structure is covered by a shielding sky-plane in the top metal layer (dotted diagonal lines). ::: Simple magnetic coupling between the inductive ladder and target device using the microstrip transformer shown in the top-left panel of Figure [\[fig:DAC2\]](#fig:DAC2){reference-type="ref" reference="fig:DAC2"} is sufficient to implement a full range of several tens of \(\mathrm{m\Phi_0}\) into the target device. However, the majority of our DACs (ones that control the compound Josephson junctions of couplers, inductance tuners and persistent current compensators) require a range comparable to half a \(\Phi_0\), since they need to be able to bias their target's corresponding CJJ all the way between its maximum \(I_c\) and fully suppressed. To implement such higher-range control we merge the CJJ loop of the target device with the MSD stage of the inductive ladder, as shown in the top-right panel of Figure [\[fig:DAC2\]](#fig:DAC2){reference-type="ref" reference="fig:DAC2"} (Josephson junctions are shown as yellow circles). An additional complication of this particular structure is that the DAC should be coupled with equal strength into both halves of the target CJJ loop in order to avoid coupling into target device body. To achieve this, the MSD coil is split into two symmetric halves, as shown in the top-right panel of Figure [\[fig:DAC2\]](#fig:DAC2){reference-type="ref" reference="fig:DAC2"}. The simplistic lumped-element model of Figure [\[fig:DAC1\]](#fig:DAC1){reference-type="ref" reference="fig:DAC1"} is not entirely adequate as a complete description of our \(\Phi\)-DAC devices, especially considering the cross-section of an actual device implemented in all six available metal layers (drawn to scale) at the bottom of Figure [\[fig:DAC2\]](#fig:DAC2){reference-type="ref" reference="fig:DAC2"}. The LSD loop couples flux directly into the MSD one and it can reach the output not only via the inductive ladder, but also via this magnetic connection to the (strongely coupled to the output) MSD. In addition, the MSD flux can reach the output directly, not mediated by the washer (and with the sign opposite to the washer-mediated coupling). We treat a complete \(\Phi\)-DAC structure as a three-port device (LSD, MSD and OUT) and, using the 3D inductance extraction program `FastHenry` with superconductor support, extract its complete inductance matrix: \[\left( \begin{matrix} L_{LSD} & M_{LSD, MSD} & M_{LSD, OUT} \\ M_{LSD, MSD} & L_{MSD} & M_{MSD, OUT} \\ M_{LSD, OUT} & M_{MSD, OUT} & L_{OUT} \end{matrix} \right).\] For subsequent analysis we treat the SFQ pulse sources as simple current sources that can produce (up to) \(I_{in}\) (approximately half of their junction critical current \(I_c\), as discussed below) into a large inductive load, and calculate all relevant parameters of our \(\Phi\)-DAC s as shown in Table [\[tab:fdac\]](#tab:fdac){reference-type="ref" reference="tab:fdac"}. After we build a \(\Phi\)-DAC layout model, we iterate over its geometrical parameters to ensure that it fits into the available space, has the required number of bits and range, and that its MSD/LSD division ratio is such that the LSD comfortably spans a single MSD step. ## \(\Phi\)-DAC: SFQ pulse sources[\[sec:vscrc\]]{#sec:vscrc label="sec:vscrc"} Our implementation of an SFQ source is based on perhaps the earliest incarnations of single-flux-quanta circuits: a current biased dc-SQUID made with two shunted junctions. A schematic of two dc-SQUIID SFQ pulse sources feeding the LSD and MSD storage loops of a single \(\Phi\)-DAC is shown in Figure [\[fig:Vsrc\]](#fig:Vsrc){reference-type="ref" reference="fig:Vsrc"}. To operate, one first applies PWR current bias (biasing all junctions to about half of their \(I_c\)). ADDR is then applied, providing an initial flux bias to the dc-SQUID bodies. Ramping TRIG with a polarity that adds to ADDR in, for example, the dc-SQUID comprised of junctions J0 and J1, eventually steers enough current through J0 to exceed its critical current, causing it to "flip" by \(2\pi\) in phase, admitting a single flux quantum into the dc-SQUID loop. TRIG is then decreased, eventually causing J1 to flip. The J0/J1 dc-SQUID is thus returned to its zero flux state, but in the process the phase drop across the LSD inductor has been increased by \(2\pi\)--an SFQ pulse is added to that storage loop. Assuming the LSD inductor is large compared to the dc-SQUID inductance, this process can be repeated until the persistent current stored in the LSD loop becomes comparable to the PWR current, cancelling it, preventing junctions from further flipping. At that point, the \(\Phi\)-DAC loop has reached its maximum SFQ capacity. If one changes the sign of PWR, using the same process one can add single flux quanta of the *opposite* magnetic field direction into this storage loop (or, *subtract* from the ones stored there). Note that the TRIG line is twisted between the dc-SQUIDs J0/J1 and J2/J3, so when it adds to the ADDR prebias for the J0/J1 dc-SQUID, it subtracts from the J2/J3 dc-SQUID, and the J2/J3 SQUID is quiescent. But if one reverses the polarity of the TRIG pulses relative to the ADDR pre-bias, one can operate the J2/J3 dc-SQUID, adding a SFQ to the MSD \(\Phi\)-DAC coil. The relative polarity of ADDR and TRIG allows us to select the \(\Phi\)-DAC stage on which we want to operate. The PWR, ADDR, and TRIG levels are chosen to meet the following criteria: 1. With PWR held at its active level, the state of a \(\Phi\)-DAC changes by exactly one flux quanta per ADDR, TRIG pulse. 2. Each \(\Phi\)-DAC undergoes SFQ transitions only when all three lines addressing that device are active. If two or fewer lines are active, the state of the \(\Phi\)-DAC does not change. If these criteria are met, then a limited number (\(O(\sqrt[3]N)\)) of control lines can address \(N\) \(\Phi\)-DAC s in what we call "XYZ" fashion, discussed further in Section [\[sec:XYZ\]](#sec:XYZ){reference-type="ref" reference="sec:XYZ"}. Here we discuss the process, which we refer to as margining, by which programming levels are chosen to meet the above criteria. \(\Phi\)-DAC state stability is fully determined by \(\Phi_b\) and \(I_b\), where \(\Phi_b\) is the sum of the ADDR and TRIG flux biases and \(I_b\) is the total current biasing the dc-SQUID SFQ pulse source. \(I_b\) includes contributions from PWR and from the current circulating in the main \(\Phi\)-DAC loop due to its flux state. A critical line in \((\Phi_b, I_b)\) space, similar to that of a current biased dc-SQUID, bounds the region in which a flux state is stable. When this line is crossed due to manipulation of PWR, ADDR, and TRIG, a transition will take place. In Figure [\[fig:margining\]](#fig:margining){reference-type="ref" reference="fig:margining"}, the critical line of the zero flux state of the dc-SQUID pulse source is plotted. Crossing this boundary corresponds to the first junction flip in the SFQ pulse sequence described previously. The system will cross the boundary at a point that depends on the main \(\Phi\)-DAC loop flux state. Margining of the PWR, ADDR, and TRIG levels can be understood as a geometric partitioning of this boundary into active regions, in which intended transitions will take place, and forbidden zones, in which transitions that do not meet the margining criteria would occur. PWR, ADDR and TRIG levels are chosen to maximize the size of the active regions while avoiding the forbidden zones. ## \(\Phi\)-DAC reset The protocols described above allow us to add or subtract SFQ to a \(\Phi\)-DAC stage, ending up with a known number of flux quanta when we start from a known state. For realistic operation we must also be able to reliably reset all \(\Phi\)-DAC s into a known state starting from an *unknown* state. To reset a \(\Phi\)-DAC, \(I_{PWR}\) is set to zero. The dc-SQUID pulse source still sees a non-zero current bias, as long as the \(\Phi\)-DAC is in a non-zero flux state. Programming the \(\Phi\)-DAC under these conditions will cause SFQ changes in the main \(\Phi\)-DAC loop that decrease this current bias. Applying ADDR+TRIG pulses with large enough amplitude to reliably drive transitions (larger than the maximum \(\Phi_b\) value of the critical boundary in Figure [\[fig:margining\]](#fig:margining){reference-type="ref" reference="fig:margining"}) will 'de-program' the \(\Phi\)-DAC  one SFQ at a time, until it reaches its lowest energy zero SFQ state for which the circulating current is zero. To reliably reach this zero SFQ state, junction critical current asymmetry must be small: the critical current difference between the two junctions in the dc-SQUID pulse source should be well under \(\Phi_0/L\), where \(L\) is the main loop inductance. Note that the margining criteria are violated during reset. All \(\Phi\)-DACs are reset simultaneously. ## Minimizing \(\Phi\)-DAC footprint As we mentioned in Section [\[sec:top\]](#sec:top){reference-type="ref" reference="sec:top"}, the \(\Phi\)-DAC area is what ultimately sets the size of our processor unit tile. This in turn determined the length of the qubits, and thus their energy scales, ultimately affecting the performance of the annealing algorithm. Minimizing this area is therefore of great importance to us. What matters for a given \(\Phi\)-DAC to achieve its design objectives is the maximum number of single flux quanta that we can store in its MSD and LSD loops (determining maximum range and precision, provided that division ratio is chosen correctly). That, in turn, is proportional to the \(L \times I_c\) product of the storage loop inductance \(L\) and the pulse source junction \(I_c\). How can we minimize an area required to implement both junctions and inductor to achieve constant (and sufficient) \(L \times I_c\) product? Equivalently, how can we maximize this product in given area? One can observe that (to the first order of approximation), inductance of a spiral coil (for fixed number of available layers) is proportional to its area with some proportionality constant \(\alpha\). The same is true for junction area, given a fixed critical current density \(J_c\). If we have unit area available for the whole \(\Phi\)-DAC, and inductance occupies some fraction \(x\) of that, \(L \times I_c = \alpha\, x \times J_c (1-x) \sim x \times (1-x)\), which reaches its maximum at \(x=0.5\), meaning that half of the optimal \(\Phi\)-DAC area is occupied by storage inductance and another half by source junctions. This was the rule that we used for choosing source junction \(I_c\) vs. storage \(L\) (their \(J_c\) was fixed by requirements of the analog qubit and coupler circuitry in a process with only a single available trilayer). Figure [\[fig:3DACs\]](#fig:3DACs){reference-type="ref" reference="fig:3DACs"} is a CAD view of three \(\Phi\)-DAC s within one plaquette of our current processor. Note that the result of this analysis is independent of critical current density. Suppose a second high-\(J_c\) trilayer becomes available for our next generation design, say, \(9\,\mathrm{kA/cm^2}\) (in addition to our current \(250\,\mathrm{A/cm^2}\)), a factor of 36 in \(J_c\). Just replacing the existing junctions with smaller in size and equal critical current would save us less than a factor of 2 in \(\Phi\)-DAC area. If instead \(L\) is decreased and \(I_c\) is increased by a factor of 6 in value, the total area decreases by the same factor, with \(L\times I_c\) product unchanged. ## XYZ-addressing line count[\[sec:XYZ\]]{#sec:XYZ label="sec:XYZ"} We need 72 \(\Phi\)-DAC s to control all the qubits and couplers of a unit tile (6 per qubit, 16 for controlling internal couplers, and 8 for controlling external couplers), for a total of 4608 \(\Phi\)-DAC s for our D-Wave Two 512-qubit processors. To select one of them using cubic XYZ-addressing, we need at least \(\left \lceil 3*\sqrt[3]{4608} \,\right \rceil = 50\) lines, or about 16 lines per dimension. We have arranged all required \(\Phi\)-DAC s for a given tile in 25 3-DAC plaquettes (one plaquette is empty), as shown in the left panel of Figure [\[fig:XYZ\]](#fig:XYZ){reference-type="ref" reference="fig:XYZ"}. One of three \(\Phi\)-DAC s within a plaquette is selected using one of three ADDR lines, with all three sharing a TRIG line, resulting in 15 ADDR and 5 TRIG lines addressing all \(\Phi\)-DAC s within unit tile. The third dimension of addressing is established by separating tile arrays into PWR domains. Our D-Wave Two processors contain an \(8\times 8\) array of unit tiles, split into sixteen \(2 \times 2\) power domains, as shown in the right panel of Figure [\[fig:XYZ\]](#fig:XYZ){reference-type="ref" reference="fig:XYZ"}. All \(\Phi\)-DAC s within one power domain are connected in series and fed by one of 16 PWR lines. 30 ADDR and 10 TRIG lines are reused between power domains, for a total of \(30+10+16=56\) lines used to address all \(\Phi\)-DAC s within a processor matrix. While it is not optimal (because of the difference in the number of ADDR and TRIG lines), it is sufficiently close, and this arrangement allowed us to achieve a more regular layout without having to assign different roles to a single line within the processor fabric (*e.g.*, make a single line work as an ADDR for one DAC and a TRIG for another). # Conclusions We have described how, starting with top-level requirements of a processor implementing a quantum annealing algorithm, we have designed its hardware graph and required control infrastructure, which allowed us to successfully operate processors with up to 512 rf-SQUID qubits using only 56 control lines for problem programming. Figure [\[fig:photo\]](#fig:photo){reference-type="ref" reference="fig:photo"} shows a microphotograph of an active area of a D-Wave Two processor chip. The most important feature of our new \(\Phi\)-DAC design is its zero static power dissipation--unlike traditional SFQ circuitry, which incorporates on-chip resistive current sources tapping a common voltage rail, this design biases all devices serially with a fixed current whose magnitude is set by a *room-temperature* resistor[^3]. The only energy dissipated on-chip is on the order of \(I_c \times \Phi_0\) per flux quantum moved into (or out of) the storage inductor. For a pair of \(55\,\mathrm{\mu A}\) \(\Phi\)-DAC junctions this corresponds to \(0.22\,\mathrm{aJ}\). Complete reprogramming of all 9216 \(\Phi\)-DAC stages moving from-16 to +16 SFQ in their storage loops would dissipate on chip only about \(65\,\mathrm{fJ}\). While the D-Wave One required a post-programming delay of about \(1\, \mathrm{s}\), D-Wave Two can thermalise to \(20\, \mathrm{mK}\) within \(10\, \mathrm{ms}\), a factor of 100 improvement achieved within one processor generation just in this post-programming thermalization time. [^1]: Of course, one can attempt embedding *logically* higher-dimensional structures into planar chips, but this approach soon runs afoul of the second constraint above. [^2]: Two special cases were introduced in D-Wave Two processors: a DAC which biases the qubit CCJJ major loop, for which currently 5 bits of dynamic range is sufficient and it was implemented as a single coil of the same type directly coupled into a target device, and second a very coarse stage for a qubit flux bias DAC, useful to deal with larger local qubit flux offsets. [^3]: This approach can also be viewed as implementing SFQ "current recycling" taken to its ultimate limit.

{'timestamp': '2014-01-23T02:01:11', 'yymm': '1401', 'arxiv_id': '1401.5504', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5504'}

# Introduction Short laser pulses are widely used in microscopy, spectroscopy and micro-machining. They allow the study of elementary processes in real time and, thanks to their high peak power, they are ideal for investigating nonlinear phenomena. However they are difficult to handle, since as they propagate through any medium, including air, they collect phase distortions. This is particularly critical in the case of focusing through high numerical aperture objectives, since the amount of glass contained in such devices is enough to broaden the pulse duration of a femtosecond pulse by several orders of magnitude. There are several techniques which can in principle be used to compensate phase distortions and deliver transform limited pulses to the sample plane of a microscope. The most straightforward are based on prism and grating compressors, which can be used to minimize the quadratic and cubic terms in the phase expansion. Pulse shaping has, in principle, the ability to compensate arbitrary phase distortions, but how to determine which phase to program in the pulse shaper is less straightforward. Genetic algorithms, for instance, require substantial measurement time and computational effort and do not always yield the desired pulse phase profile in a reproducible way. Frequency-resolved optical gating (FROG) is perhaps one of the most popular tools to retrieve to phase of femotosecond pulses, and there are several others ingenious techniques. A significant step towards reliable and reproducible compression of ultra-short laser pulses was achieved by the group of Dantus, with the development of MIIPS. The authors demonstrated the possibility to compensate the phase distortion introduced by a \(0.60 {\rm NA}\) objective by only using MIIPS. However the compensation of higher NA objectives benefits from external compression (for instance by a prism pair) to retain the accessible phase range of the shaper. In this paper we show that MIIPS can be improved by complementing it with an amplitude gate, that is scanned across the spectrum. We demonstrate that this gated version of MIIPS (Fig.[\[fig:GatedMiips\]](#fig:GatedMiips){reference-type="ref" reference="fig:GatedMiips"}) estimates more accurately the spectral phase, without the need to increase the measurement time. It is also less affected by systematic errors originating from higher order phase distorsions and shaper artifacts. The work is organized as follows: in section [2](#sec:Miipsintro){reference-type="ref" reference="sec:Miipsintro"} we give a brief introduction to the MIIPS technique; in section [3](#sec:Miipschallenges){reference-type="ref" reference="sec:Miipschallenges"} we discuss the limitations of the standard implementation of Miips; in section [4](#sec:GatedMiips){reference-type="ref" reference="sec:GatedMiips"} we introduce Gated-Miips, in section [5](#sec:analytical){reference-type="ref" reference="sec:analytical"} we derive analytical models for MIIPS and Gated-MIIPS; in section [6](#sec:NumSimulation){reference-type="ref" reference="sec:NumSimulation"} we illustrate, by numerical simulation, the advantages of Gated-MIIPS and, in section [7](#sec:ExpData){reference-type="ref" reference="sec:ExpData"}, we show experimental data indicating that Gated-MIIPS is better suitable than standard MIIPS in estimating the phase of a femtosecond laser pulse. ::: # Multiphoton Intrapulse Interference Phase {#sec:Miipsintro} MIIPS was designed as a tool to control multi-photon processes and to characterize and compress ultrashort laser pulses. It employs a pulse-shaper setup, in order to estimate the group delay dispersion (GDD) and to compensate arbitrary complex spectral phase distortions. The main concept is that the GDD is correlated with the intensity of Second Harmonic Generation (SHG): at any given frequency the SHG is maximized if the GDD is null. It follows that a transform limited (flat phase) pulse can be obtained by applying suitable phase masks and maximizing the SHG for the whole spectrum. This can be seen from the equation for the SHG at the frequency \(\omega\) in the ideal case (assuming \(\chi^2=1\)) as: \[\label{eq:SHG} {\rm SHG}(2 \omega) = \left| \int_{-\infty}^{+\infty} \left| E(\omega-\Omega) \right| \left| E(\omega+\Omega) \right| \right. \exp\left[ i (\phi(\omega-\Omega)+\phi(\omega+\Omega) \right] {\rm d}\Omega \left.\vphantom{\int} \right|^2\] Where \(E(\omega)\exp(i \phi)\) is the amplitude of the electric field at angular frequency \(\omega\). A list of the symbols used in this articles can be found in Table [1](#tab:ListOfAbbreviations){reference-type="ref" reference="tab:ListOfAbbreviations"}. Since the first two factors of the integrand in Eq.[\[eq:SHG\]](#eq:SHG){reference-type="ref" reference="eq:SHG"} are positive, the SHG is maximized when the argument of the exponential is zero. Expanding the phase in series and retaining the second order, the approximated equation for the SHG becomes: \[\label{eq:SHGapprox} {\rm SHG}(2 \omega) = \left| \int_{-\infty}^{+\infty} \left| E(\omega-\Omega) \right| \left| E(\omega+\Omega) \right| \exp\left( i \frac{\partial^2\phi(\omega) }{ \partial\omega^2} \Omega^2 \right) {\rm d}\Omega \right|^2\] It follows that the SHG is maximized when the group delay dispersion \({\rm GDD } = \ddot{\phi}(\omega)\) is equal to zero. This condition is at the basis of MIIPS, and it is valid as long as the phase can be locally approximated by a second order polynomial. The standard implementation of MIIPS involves adding a sinusoidal modulation to the phase of the laser pulse, typically using a pulse-shaper: \[f(\omega) = \exp \left\{ i \Phi_0 \, \sin\left[ \tau \, (\omega-\omega_0)-\psi \right] \right\} \label{eq:MiipsTestPhase}\] We note that such a sinusoidal phase modulation can also be seen, in the time domain, as the generation of a pulse train. The modulation frequency \(\tau\), which has the units of time, corresponds to the temporal separation between subsequent pulses. The GDD is then estimated by varying the modulation phase \(\psi\) (Fig.[\[fig:GatedMiips\]](#fig:GatedMiips){reference-type="ref" reference="fig:GatedMiips"}) and determining for which values the SHG intensity is locally maximized. Combining Eq.[\[eq:SHG\]](#eq:SHG){reference-type="ref" reference="eq:SHG"} and Eq.[\[eq:MiipsTestPhase\]](#eq:MiipsTestPhase){reference-type="ref" reference="eq:MiipsTestPhase"} and using the prosthaphaeresis formulae, it is possible to show that the SHG of a laser pulse subject to sinusoidal phase modulation is maximized when: \[\phi(\omega-\Omega)+\phi(\omega+\Omega)+2 \Phi_0 \cos (\tau \Omega) \sin \left[ (\omega-\omega_0) \tau-\psi \right] = 0 \label{eq:MiipsCondition}\] which means that the GDD can be estimated, to the second order of accuracy, as: \[\ddot{\phi}(\omega) = \Phi_0\,\tau^2 \, \sin\left[ \tau \, (\omega-\omega_0)-\psi_m(\omega) \right]\] Where \(\psi_m(\omega)\) is the value of the scanning parameter \(\psi\) which maximizes the SHG for the frequency \(\omega\). The phase is then recovered numerically by double integration. For a more extensive description of the implementation of MIIPS and its applications we refer to the works by Dantus, quoted in the references. We note that the zero and first order terms in the Taylor expansion of the phase do not contribute to second harmonic generation (Eq.[\[eq:SHG\]](#eq:SHG){reference-type="ref" reference="eq:SHG"}) and therefore are not measured by MIIPS. The zero order is responsible for the carrier-envelope phase and the first order leads to a translation of the pulse along the time axes. # Challenges of MIIPS {#sec:Miipschallenges} A known limitation of MIIPS is that it only considers the \(2^{\rm nd}\) order correction to the phase at a given frequency, neglecting higher orders in the Taylor expansion. Although this can lead to tolerable errors for not overly distorted pulses, in other cases it can significantly reduce the accuracy of MIIPS as discussed below. Xu et al. showed that the systematic error is of the order of \((\tau/\Delta t)^2/12\), where \(\tau\) is the MIIPS modulation frequency and \(\Delta t\) is the transform limited pulse duration. From the above estimate, it might seem that accurate results could always be obtained for very short pulse duration, just by choosing a very small modulation frequency, and indeed an ideal MIIPS simulation, which does not take into account any of the shaper artifacts, would confirm this conclusion. In Fig.[\[fig:MiipsPerformance\]](#fig:MiipsPerformance){reference-type="ref" reference="fig:MiipsPerformance"} we report several simulated MIIPS scans performed on a \(5.3{\rm fs}\) laser pulse centered at \(\omega_0=2.4{\rm rad}/{\rm fs}\) (\(785 {\rm nm}\)), after propagation through \(10{\rm cm}\) of BK7 glass, which mimic the optical path length of a microscope objective. For each scan we report the residual spectral phase and the SHG spectrum. In order to compare the different modulation parameters we kept the maximum GDD correction constant (see Eq.[\[eq:maxGDD\]](#eq:maxGDD){reference-type="ref" reference="eq:maxGDD"}). As expected, the best performance in terms of maximum SHG and residual phase are obtained for the smaller values of the modulation frequency \(\tau\). In practice the adjustment of the modulation parameters is subtler because, for highly distorted ultrashort pulses, it requires setting the modulation amplitude to very high values, which in turn deteriorates the spectrum by introducing shaper artifacts, like for instance diffraction caused by phase grating and ripples in transmitted spectra. The former can be minimized by employing a double pass 4f pulse shaper setup. The latter can be explained as follows. By definition the spectral phase is given by multiples of \(2\pi\) radians. However due to cross-talk between the SLM pixels, phase and amplitude distortions can be observed in the spectral regions where the phase crosses \(2\pi\), known as wrapping. Therefore, if the amount of wrapping is very high, the spectrum transmitted through the SLM can be too distorted to be accurately characterized. This poses a limitation to the total phase which can be compensated and it is particularly severe for very short pulses (\(<10 {\rm fs}\)) in combination with high-numerical aperture objectives. For the above reasons, a common choice is to set the phase modulation frequency \(\tau\) equal to the transform limited pulse duration \(\Delta t\). This is a good compromise because lower modulation frequency would result in having a too high modulation amplitude, which would cause phase-wrapping artifacts, as we previously discussed. Higher modulation frequencies, on the other hand, are not optimal because the second order polynomial expansion in Eq.[\[eq:SHGapprox\]](#eq:SHGapprox){reference-type="ref" reference="eq:SHGapprox"} loses validity. Indeed, since MIIPS uses a sinusoidal phase mask, it introduces both second (SOD) and fourth order (FOD) corrections, linked by the relation: \({\rm FOD}=-{\rm SOD} \cdot \tau^2\). In a 2D map which represents the SHG intensity as a function of GDD and FOD values (Fig.[\[fig:MiipsTrajectory\]](#fig:MiipsTrajectory){reference-type="ref" reference="fig:MiipsTrajectory"}), MIIPS maximizes the signal along a linear trajectory and from Fig.[\[fig:MiipsTrajectory\]](#fig:MiipsTrajectory){reference-type="ref" reference="fig:MiipsTrajectory"} it can be seen that this maximum does not necessarily correspond to a zero GDD. ## MIIPS Resolution {#miips-resolution .unnumbered} For a defined set of modulation parameters, the maximum GDD which can be compensated by a single MIIPS iteration is: \[\label{eq:maxGDD} {\rm GDD_{max}} = \Phi_0 \, \tau^2\] On the other hand, the minimum GDD which can be compensated depends on the size of the phase steps and on the maximum relative error which can be accepted for the GDD. If \(\Delta \psi\) is the increment of the argument of the sinusoidal modulation, and \(\epsilon = \Delta {\rm GDD}/{\rm GDD}\) is the maximum relative error on the GDD, then it follows that: \[\label{eq:GDDrelerr} \left|{\rm GDD_{min}}\right| = \Phi_0 \, \tau^2 \, \left( 1 + \left( \frac {\epsilon}{\Delta \psi} \right)^2 \right)^{-1/2}\] The dynamical range, defined as the ratio of the maximum and the minimum GDD is therefore: \[D = \left( 1 + \left( \frac{\epsilon}{\Delta \psi} \right)^2 \right)^{1/2} \approx \frac{\epsilon}{\Delta \psi}\] Which leads, for a scan of 1000 points and an accuracy of \(10\%\), to a dynamical range of only 16. This explains why MIIPS needs to be iterated multiple times to compensate the phase of laser pulses. This limitation can be mitigated by limiting the spectral bandwidth (and the range of GDD values) around the frequency at which the GDD is being compensated, as done in Gated-MIIPS discussed below. # Gated-MIIPS {#sec:GatedMiips} We developed Gated-MIIPS to avoid the limitations of MIIPS when compensating large phase distortions caused by broad spectra and high numerical aperture objective. The idea, illustrated in Fig.[\[fig:GatedMiips\]](#fig:GatedMiips){reference-type="ref" reference="fig:GatedMiips"}, is to enable using higher modulation frequencies for the sinusoidal phase, while avoiding systematic errors due to higher order phase terms. In the next sections, however, we will show that Gated-MIIPS is more accurate that MIIPS also with the same choice of modulation parameters. Gated-MIIPS can be readily implemented using a pulse shaper which provides both phase and amplitude modulation. The amplitude modulation is exploited to gate the spectrum around a specific frequency (Fig.[\[fig:GatedMiips\]](#fig:GatedMiips){reference-type="ref" reference="fig:GatedMiips"}c), therefore improving the validity of the second order polynomial expansion of the phase, even in the case of significant higher order contributions, as will be shown below. It then becomes possible to use higher values of the modulation frequency \(\tau\) and lower values of the modulation amplitude \(\Phi_0\) minimizing the phase wrapping. In the case of a Gaussian gate the modulation that is applied by the pulse shaper can be written as a function of the phase terms: \[\begin{split} M(\omega) = \exp \left\{-\left[ \frac{\tau \, (\omega-\omega_0)-\psi} {\sigma} \right]^2 + i \Phi_0 \sin \left[\tau \, (\omega-\omega_0)-\psi \right] \right\} \end{split} \label{eq:gate}\] As it can be seen from Eq.[\[eq:gate\]](#eq:gate){reference-type="ref" reference="eq:gate"}, in Gated-MIIPS a Gaussian amplitude mask of width \(\sigma\) (here called gate) is translated alongside the phase modulation, using the scanning parameter \(\psi\). To improve the accuracy for the GDD values closer to the gate boundaries (\(\pm \Phi_0 \tau^2\)) the Gated-MIIPS signal \(\mathcal{G}(2 \omega)\) is then obtained by dividing the SHG spectrum by the square of the Gaussian amplitude mask. \[\mathcal{G}(2 \omega) = {\rm SHG}(2 \omega) \cdot \exp \left\{ 4\left[ \frac{\tau \, (\omega-\omega_0)-\psi}{\sigma}\right]^2 \right\} \label{eq:GatedMiipssignal}\] By combining Eq.[\[eq:SHG\]](#eq:SHG){reference-type="ref" reference="eq:SHG"} with Eq.[\[eq:gate\]](#eq:gate){reference-type="ref" reference="eq:gate"}, in analogy with what previously discussed for the standard MIIPS case, it can be shown that: \[\begin{split} \mathcal{G}(2 \omega) &= \left|\int_{-\infty}^{+\infty} \left| E(\omega-\Omega) \right| \left| E(\omega+\Omega) \right|\cdot \exp \left( -\frac{2 \tau^2 \Omega^2}{\sigma^2} \right) \cdot \right. \\ &\exp \left\{\right. i \left[\right. \phi(\omega-\Omega)+\phi(\omega+\Omega)+ \left. 2 \Phi_0 \cos (\tau \Omega) \sin \left[\tau (\omega-\omega_0)-\psi\right] \left.\right] \left.\right\} {\rm d}\Omega \vphantom{\int}\right|^2 \end{split} \label{eq:GatedMiips}\] By direct inspection, it can be seen from Eq.[\[eq:GatedMiips\]](#eq:GatedMiips){reference-type="ref" reference="eq:GatedMiips"} that the Gated-MIIPS signal \(\mathcal{G}(2 \omega)\) is maximized by the same condition expressed by the Eq.[\[eq:MiipsCondition\]](#eq:MiipsCondition){reference-type="ref" reference="eq:MiipsCondition"}. Therefore, in the ideal case Gated-MIIPS should provide the same results as the standard MIIPS. The advantage of Gated-MIIPS stems from the fact that the presence of the Gaussian term in Eq.[\[eq:GatedMiips\]](#eq:GatedMiips){reference-type="ref" reference="eq:GatedMiips"} improves the validity of the second order expansion of the spectral phase (Eq.[\[eq:SHGapprox\]](#eq:SHGapprox){reference-type="ref" reference="eq:SHGapprox"}). In analogy with standard MIIPS, the Gated-MIIPS signal can also be expressed analytically for a few simple pulse shapes. For instance, analytical expressions for the SHG signal of a transform limited Gaussian pulse, subject to sinusoidal phase modulation, have been given by Hacker et al. and reproposed by Lozovoy et al.. However, to the best of our knowledge, an analytical expression of the MIIPS signal in the case of a chirped Gaussian pulse has not been shown yet. In the next section we will derive it for the cases of both MIIPS and Gated-MIIPS. Such an expression can be useful for fitting a MIIPS trace without relying on a peak finding algorithm, as it is usually done. Additionally, we will show that an approximated analytic expression for Gated-MIIPS signal can be also be written for a generic pulse shape, provided that the field amplitude varies slowly with respect to the Gaussian gate (Eq.[\[eq:gate\]](#eq:gate){reference-type="ref" reference="eq:gate"}). # Analytical Models for MIIPS and Gated-MIIPS {#sec:analytical} ## MIIPS on a chirped Gaussian Pulse Let us assume a Gaussian pulse given by: \[\label{eq:ChirpedPulse} E(\omega) = \exp \left( -\frac{(\omega-\omega_0)^2} {{\Delta\omega}^2} + i \, \phi(\omega) \right)\] By substituting Eq.[\[eq:ChirpedPulse\]](#eq:ChirpedPulse){reference-type="ref" reference="eq:ChirpedPulse"} into Eq.[\[eq:SHG\]](#eq:SHG){reference-type="ref" reference="eq:SHG"} we obtain: \[\label{eq:MiipsGaussianPulse} \begin{split} {\rm SHG}(2 \omega) &= \exp \left\{-4 \, \frac { \left( \omega-\omega0 \right) ^2}{{ \Delta \omega }^2}\right\} \cdot \\ &\left| \int _{-\infty }^{\infty }\! \exp \left\{-2\,\frac {{\Omega}^2} {{\Delta \omega }^2} +i\, \left[ \phi(\omega-\Omega)+\phi(\omega+\Omega)+ 2\Phi_0\,\cos \left( \tau\,\Omega \right) \sin \left(\tau \, (\omega-\omega_0) - \psi \right)\right] \vphantom{\int} \right\} {{\rm d}\Omega} \right|^{2} \end{split}\] Now if, as in standard MIIPS, we approximate the phase \(\phi(\omega)\) by a second order polynomial, it becomes possible to calculate the integral by using the Jacobi-Anger expansion, given by the formula: \[\label{eq:JacobiAnger} \exp \left( i \, z \cos (\theta) \, \right) = \sum_{n=-\infty}^{+\infty} \left( i^n J_n(z) \exp(i \, n \, \theta) \right)\] Where \(J_n(z)\) is the Bessel function of the first kind. The resulting SHG is given by: \[\label{eq:MiipsFullAnalytical} \begin{split} {\rm SHG}(2 \omega) = \frac {\pi {\Delta\omega}^2} {\sqrt{ {\Delta\omega}^4 {\ddot{\phi}(\omega)}^2 + 4 }} \exp \left(- \frac {4 (\omega-\omega_0)^2} {{\Delta\omega}^2} \right) \cdot \sum_{n=-\infty}^{+\infty} { \exp \left(- \frac{ {\Delta\omega}^2 \tau^2\, n^2 } {{\Delta\omega}^4 {\ddot{\phi}(\omega)}^2+4} \right) {J_n\left( 2 \Phi_0 \sin \left( \tau (\omega-\omega_0)-\psi \right) \right)}^2 } \end{split}\] Eq.[\[eq:MiipsFullAnalytical\]](#eq:MiipsFullAnalytical){reference-type="ref" reference="eq:MiipsFullAnalytical"} can be accurately computed using a limited number of terms, because as \(\left| n \right|\) increases the contribution of each added term decreases exponentially. A simpler formula, accurate to second order in the phase, can be obtained by doing the Taylor expansion of the phase term in Eq.[\[eq:MiipsGaussianPulse\]](#eq:MiipsGaussianPulse){reference-type="ref" reference="eq:MiipsGaussianPulse"}. After some algebraic manipulation one obtains: \[\begin{split} {\rm SHG}(2 \omega,\phi) = \pi {\Delta\omega}^2 \exp \left( -\frac{4(\omega-\omega_0)^2}{{\Delta\omega}^2} \right) \left\{ 4+{\Delta\omega}^4 \left[ \,\ddot{\phi}(\omega)-\tau^2 \Phi_0 \sin (\tau (\omega-\omega_0)-\psi) \right]^2 \right\}^{-1/2} \end{split} \label{eq:MiipsApproxAnalytical}\] By differentiating the above formula with respect to \(\psi\) and letting \({\partial\, {\rm SHG}(2 \omega)}/{\partial\psi}=0\), we find the explicit expression of the GDD in function of the scanning parameter \(\psi\), which was already found in Eq.[\[eq:MiipsCondition\]](#eq:MiipsCondition){reference-type="ref" reference="eq:MiipsCondition"}. ## Analytical expression for Gated-MIIPS For the specific case of a Gaussian pulse, the analytical expression for the Gated-MIIPS signal is very similar to the one we just derived for MIIPS. By substituting Eq.[\[eq:ChirpedPulse\]](#eq:ChirpedPulse){reference-type="ref" reference="eq:ChirpedPulse"} into Eq.[\[eq:GatedMiips\]](#eq:GatedMiips){reference-type="ref" reference="eq:GatedMiips"}, after some algebraic manipulation, one arrives at a formula similar to Eq.[\[eq:MiipsGaussianPulse\]](#eq:MiipsGaussianPulse){reference-type="ref" reference="eq:MiipsGaussianPulse"}, except for the substitution \({\Delta\omega}^2 \rightarrow L^2=({\Delta\omega}^2 \sigma^2)/(\sigma^2+{\Delta\omega}^2 \tau^2)\) inside the integration sign. \[\label{eq:GatedMiipsGaussianPulse} \begin{split} \mathcal{G}(2 \omega) &= \exp \left\{-4 \, \frac { \left( \omega-\omega0 \right) ^2}{{ \Delta \omega }^2}\right\} \cdot \\ & \left| \int _{-\infty }^{\infty }\! \exp \left\{-2\,\frac {{\Omega}^2} {L^2} +i\, \left[ \phi(\omega-\Omega)+\phi(\omega+\Omega)+ 2\Phi_0\,\cos \left( \tau\,\Omega \right) \sin \left[ \tau\,(\omega-\omega_0)- \psi \right]\right] \vphantom{\int} \right\} {{\rm d}\Omega} \right|^{2} \end{split}\] We note that, since it is always \(L<\Delta\omega\), Gated-MIIPS traces are normally ticker that MIIPS traces recorded with the same modulation parameters (see Fig). Using the Jacobi-Anger expansion is then straightforward to derive the analog of Eq.[\[eq:MiipsApproxAnalytical\]](#eq:MiipsApproxAnalytical){reference-type="ref" reference="eq:MiipsApproxAnalytical"}. In the same way, the approximated Gated-MIIPS signal of a chirped Gaussian pulse is given by expressions equivalent to Eq.[\[eq:MiipsApproxAnalytical\]](#eq:MiipsApproxAnalytical){reference-type="ref" reference="eq:MiipsApproxAnalytical"}: \[\begin{split} \mathcal{G}(2 \omega,\phi) = \pi {L}^2 \exp \left( -\frac{4(\omega-\omega_0)^2}{{\Delta\omega}^2} \right) \left\{ 4+{L}^4 \left[ \,\ddot{\phi}(\omega)-\tau^2 \Phi_0 \sin (\tau (\omega-\omega_0)-\psi) \right]^2 \right\}^{-1/2} \end{split} \label{eq:GatedMiipsApproxAnalytical}\] This confirms that, under this approximation, Gated-MIIPS provides the same information as MIIPS. In the next section, however, we will show by numerical simulation that, in realistic conditions where higher order phase terms cannot be neglected, Gated-MIIPS provides a significant advantage over the standard implementation of MIIPS. A more general formula can be derived for an arbitrary shaped pulse, provided that its amplitude vary slowly with respect to the width of the gate \(\sigma\). By expanding in series both the phase term the field amplitude in Eq.[\[eq:GatedMiips\]](#eq:GatedMiips){reference-type="ref" reference="eq:GatedMiips"} one can obtain: \[\label{eq:GatedMiipsGeneral} \begin{split} \mathcal{G}(2 \omega) &= \frac{\pi \sigma^2}{4} \left\{ \left[ \ddot{\phi}(\omega)-\tau^2 \Phi_0 \sin (\tau (\omega-\omega_0)-\psi) \right]^2 \sigma^4 + 4 \tau^4 \right\}^{-3/2} \cdot \\ &\left\{ 4 \sigma^4 \left|E(\omega)\right|^4 \left[ \ddot{\phi}(\omega)-\tau^2 \Phi_0 \sin (\tau (\omega-\omega_0)-\psi) \right]^2 + \left[ \sigma^2 \left( \left|E(\omega)\right| \, \frac{\partial^2\left|E(\omega)\right|}{\partial\omega^2}-\left( \frac{\partial \left|E(\omega)\right|}{\partial\omega} \right)^2 \right)+ 4 \tau^2 \left|E(\omega)\right|^2 \right]^2 \right\} \end{split}\] # Numerical Comparison {#sec:NumSimulation} In order to test the efficacy of the Gated-MIIPS we performed a series of simulated measurements (see Fig.[\[fig:glasscompression10cm\]](#fig:glasscompression10cm){reference-type="ref" reference="fig:glasscompression10cm"}a-b), in which we compared the ability of MIIPS and Gated-MIIPS in correctly estimating the phase introduced by \(10\,\textrm{cm}\) of glass on a \(5.3\,\textrm{fs}\) laser pulse centered at \(\omega_0 = 2.4 \, \textrm{rad/fs}\). The modulation parameters were \(\Phi_0=100\, {\rm rad}\) and \(\tau=10\,\textrm{fs}\), and the gate width was \(\sigma=0.5\,\textrm{rad}\). In Fig.[\[fig:glasscompression10cm\]](#fig:glasscompression10cm){reference-type="ref" reference="fig:glasscompression10cm"}c-d we report, for the two cases, the actual and estimated values of GDD, the SHG spectrum before and after phase correction obtained by a single iteration, the estimated phase and the residual phase. From the traces reported in Fig.[\[fig:glasscompression10cm\]](#fig:glasscompression10cm){reference-type="ref" reference="fig:glasscompression10cm"} it can be seen that the Gated version of MIIPS is superior in both estimating the group velocity dispersion and yielding greater SHG after one iteration. The difference could be even greater if we consider that, to compensate such a considerable GDD, we had to employ a modulation amplitude \(\Phi_0\) of \(100\, \textrm{rad}\), which could cause significant phase-wrapping distortions. In order to reduce phase wrapping, we would have to decrease \(\Phi_0\) and correspondingly increase \(\tau\) (see Eq.[\[eq:maxGDD\]](#eq:maxGDD){reference-type="ref" reference="eq:maxGDD"}). This would reduce the performance of MIIPS, because of the increased contribution from higher order phase terms, as discussed above. Gated-MIIPS, however, would be relatively unaffected. One important parameter that needs to be considered when comparing MIIPS with Gated-MIIPS is the choice ofthe gate width \(\sigma\) with respect to the pulse bandwidth \(\Delta\Omega\) modulation frequency \(\tau\). To some extent narrower gate values yield better results, as shown in Fig. [\[fig:GateDependence\]](#fig:GateDependence){reference-type="ref" reference="fig:GateDependence"}). However, reducing the gate also means decreasing the intensity of the SHG spectra recorded during the Gated-MIIPS scan. Therefore the integration time needs to be adjusted accordingly. The ideal choice of the gate width is then a compromise between phase accuracy and measurement time. As previously discussed, the reason of the success of Gated-MIIPS in compensating the phase introduced by propagation through transparent media lies in the ability of dealing with higher order phase contributions. To highlight this aspect we performed another simulation in which we assumed the presence of only a \(4^\textrm{th}\) order phase term. This situation could mimic the case in which the second order (GDD) term has been pre-compensated by a prism compressor. The results, for a \(5.3\,\textrm{fs}\) laser pulse centered at \(\omega_0 = 2.4 \, \textrm{rad/fs}\) are reported in Fig.[\[fig:FourthOrder\]](#fig:FourthOrder){reference-type="ref" reference="fig:FourthOrder"}. It can be seen that in this specific situation, not only Gated MIIPS better estimates the GDD but that standard MIIPS even lead to a reduction of the SHG intensity. Since pre-compensating the GDD before pulse-shaping is common practice, this result underlines the usefulness of Gated-MIIPS. # Experimental Data {#sec:ExpData} In this section we experimentally compare the ability of MIIPS and Gated-MIIPS in estimating the chirp accumulated by a femtosecond laser pulse after traveling through glass. To perform the experiment, we started by compressing a \(15fs\) laser pulse, centered at 800nm, and focused through a \(1.3{\rm NA}\) microscope objective. Then we introduced additional glass on the optical path by a matched pair of SF10 prisms and performed a single iteration of either MIIPS or Gated-MIIPS. The two prims were in contact, with a thin layer of optical oil between them to minimize reflections at the interface. In both cases the scan parameters were: \(\Phi_0 = 10 {\rm rad}\) and \(\tau = 25 {\rm fs}\), with 401 phase steps and one second integration per step. The gate width was set to \(0.18\,\textrm{rad}\). The measurements were performed using a 4f pulse shaping setup, based on a 128 pixel spatial light modulator. The SHG was obtained using BBO crystals, deposited on a microscope slide. The results, shown in Fig.[\[fig:GatedMiipsexperimen\]](#fig:GatedMiipsexperimen){reference-type="ref" reference="fig:GatedMiipsexperimen"}a, for \(23{\rm mm}\) glass illustrate that the GDD introduced by the glass is better estimated by Gated-MIIPS. The contrast is particularly striking if one compare the resulting SHG signal after one iteration ([\[fig:GatedMiipsexperimen\]](#fig:GatedMiipsexperimen){reference-type="ref" reference="fig:GatedMiipsexperimen"}b): Gated-MIIPS in just one iteration recovers almost all the SHG spectrum of the compressed pulse. This result, together with the numerical simulations reported in the previous section, unambiguously show that Gated-MIIPS is a significant improvement towards a quick and reliable characterization of ultrashort laser pulses. # Conclusions {#sec:conclusions} We introduced an improved scheme for ultrashort laser pulse compression and characterization, which avoids some limitations of MIIPS for severely phase-distorted ultrashort pulses. We demonstrated its effectiveness in giving a more accurate correction to substantial phase distortions, as those caused by propagation in high-NA microscope objectives.

{'timestamp': '2014-03-27T01:08:15', 'yymm': '1401', 'arxiv_id': '1401.5952', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5952'}

# Introduction {#section:Intro} LMC X--3 was one of the first extragalactic X-ray sources to be discovered. It was subsequently identified with a B main-sequence star. With the discovery that the star is in a 1.7-day binary orbit with a massive dark companion, \(M \gtrsim 2.3~M_{\odot}\), LMC X-3 became the first example of an extragalactic stellar-mass black hole. LMC X--3 is an unusual system straddling the boundary between transient and wind-fed black hole binary systems (e.g., @Soria_2001). Its global characteristics, however, show that its mass transfer is governed by Roche-lobe overflow, which places it with the transients. Although it exhibits extreme X-ray variability, spanning roughly three orders of magnitude in luminosity, LMC X--3 is unique among the transients because, with a few notable exceptions, it remains in a disk-dominated thermal state (approximately 90% of the time). In fact, the dominance of the soft thermal component and the relative simplicity of LMC X--3's spectrum are the reasons and both chose the source to benchmark the performance of accretion-disk spectral models. Likewise, our group has used the extensive thermal-state data available for LMC X--3 to firmly establish the existence of a constant inner-disk radius in black hole binary systems. Recently, have identified two long-duration "low" phases, each lasting several months, during which the system was in a hard state. This unusual behavior is likely related to the well-known variability of the source on long timescales. Initially, Cowley et al. suggested a superorbital X-ray period of \(\sim198\) (or possibly \(\sim99\)) days, but subsequent studies of both optical and X-ray variability find a range of superorbital periodicities extending from \(\sim100-500\) days. There is no stable long period in this source, and the superorbital variability appears not to be attributable to the precession of a warped disk or related at all to the orbital dynamics of the system. Instead, the long-term variability is likely produced by changes in the mass accretion rate (@Brocksopp_2001; hereafter B01). Several groups have used the extensive data available for LMC X--3 to search for connections between X-ray variability and optical variability. Notably, B01 find that the X-ray and optical light curves are correlated, with an X-ray lag of 5--10 days, whereas find a somewhat longer \(\sim 20\)-day lag. A complementary body of work has focused on examining very fast (\(\lesssim 10\)s) variability in the hard state for other black-hole binary systems (e.g., @Gandhi_2010 [@Durant_2011]). In these systems, a surprising subsecond anticorrelation between X-ray and optical is commonly observed that is unlikely to be related to X-ray reprocessing. Rather, it appears that this behavior is related to the presence of a jet. In the case of LMC X--3, no jet is expected, particularly for the thermal state relevant here. In the standard \(\alpha\)-disk theory of accretion, one naturally expects the signals between any two wave bands to be correlated with a time delay that corresponds to the viscous time for a parcel of gas to travel from the outer emission region (longer wavelength band) in the disk to the inner (shorter wavelength band). Let us consider an annular region with an outer radius \(R\) and a gap in radius \(\Delta R\) between the two zones that are of interest in this paper, namely, the regions of optical and X-ray emission. Since the smaller radius in this case is negligible compared to \(R\), the viscous time for gas with viscosity \(\nu\) to traverse the distance \(\Delta R \sim R\) is simply, \[t_{\rm visc} = R^2/\nu. \label{eq:tvisc}\] In this paper, we link the optical variability of LMC X--3 to its X-ray variability by using a wealth of optical and X-ray data. More specifically, we employ two extensive optical / infrared (OIR) data sets and, in the X-ray band, we combine data collected daily by *RXTE *'s All-Sky Monitor (ASM) with high-sensitivity data obtained in many hundreds of pointed observations using *RXTE *'s Proportional Counter Array (PCA). # Data Our principal OIR data set was obtained using the ANDICAM on the SMARTS 1.3 m telescope. ANDICAM is a dual-channel imager, which we used to obtain pairs of images of LMC X--3 in \(B\), \(V\), \(I\), or \(J\). The data were reduced and calibrated with respect to several field stars, as described in. The ANDICAM data were supplemented by \(B\) and \(V\) data kindly supplied by C. Brocksopp and described in B01. These data are a subset from a larger sample obtained using the 0.91 m Dutch Telescope of the European Southern Observatory between 1993 and 1999 during a total of 16 different observing runs. Our full data set is derived from 470, 514, 476, and 440 SMARTS observations in \(B, V, I,\) and \(J\), respectively, and 356 and 685 additional ESO observations in \(B\) and \(V\) (B01). We employ just \(\sim 75\%\) of the B01 data set, because we exclusively consider those data collected after 1996 March 7 (MJD \(\geq 50149\)) when *RXTE * operations entered a mature phase. Our *RXTE * X-ray data set is derived from both ASM and PCA observations. The calibrated ASM data using the full bandwidth (2--12 keV) and 1-day time bins were retrieved from the website supported by the ASM/*RXTE* team[^1]. The PCA data, derived from 1598 observations made using the PCU-2 detector, were all reduced and analyzed following closely our standard methods (e.g., @Steiner_2010), but in this case the data were partitioned in somewhat finer time bins, using continuous exposure intervals with a mean of 2 ks (and range 300s to 5000s). Our goal here is relatively modest, namely, to derive reliable estimates of flux and construct an X-ray light curve. The SMARTS OIR light curves and the contemporaneous ASM + PCA X-ray light curve are shown in Figure [\[fig:lc\]](#fig:lc){reference-type="ref" reference="fig:lc"}. ## Unifying PCA and ASM Data Sets {#subsection:mdot} In order to most fully capture the X-ray variability of LMC X--3, we combine the ASM and PCA data sets into a single, unified light curve. Given the complementary cadence and sensitivity of the two instruments, this approach results in a light curve that for our purposes is distinctly superior to using only one of the light curves singly. The method we adopt here for combining the light curves does not critically affect our final results; e.g., the character of our results are unaffected if we use the ASM data alone. We choose to standardize the two data sets by determining for each ASM and PCA observation the mass accretion rate onto the black hole, \({\dot{m}}_X\). We make this choice because LMC X--3 is nearly always in the thermal state, a state in which the stable inner-radius results in a constant efficiency (depending only on the spin parameter), and \({\dot{m}}_X\) is therefore simply proportional to the bolometric luminosity emitted by the disk. The mass accretion rate determined from PCA spectral fits can the be compared to the 2--12 keV PCA flux, providing a scaling relationship that is then used to determine \({\dot{m}}_X\) values for the 2--12 keV flux measurements from ASM data, which cannot be fit for bolometric disk luminosity. Fortunately, when LMC X--3 does occasionally enter the hard state its luminosity is so low that \({\dot{m}}_X \approx 0\) is a quite reasonable approximation for the purposes of this study. We first determine \({\dot{m}}_X\) for the PCA data by fitting the \(\sim 1600\) PCU-2 spectra to a standard disk model that accommodates the presence of a weak Compton component. The complete spectral model in *XSPEC* notation takes the form: tbabs\(\times(\)simpl\(\otimes\)kerrbb2). The disk component kerrbb2 fits for \({\dot{m}}_X\) using fixed--and here, approximate--values for black-hole mass, inclination and distance. simpl models the Compton power law and tbabs the low-energy absorption. Each PCA spectrum thus delivers a value of the accretion rate[^2] At the same time, we compute the *observed* PCA flux \(F_X\) from 2--12 keV relative to the standard flux of the Crab in the same band. As illustrated in Figure [\[fig:mdot\]](#fig:mdot){reference-type="ref" reference="fig:mdot"}, the relationship between \({\dot{m}}_X\) and \(F_X\) obtained in this way is well defined and tightly constrained[^3]. We likewise converted the 2--12 keV ASM count rates into fluxes relative to the Crab (1 Crab = 75 ASM c s\(^{-1}\)) and then interpolated the relationship derived for the PCA in order to establish the relation between flux and \({\dot{m}}_X\) for the ASM. (To accommodate negative ASM count rates, we used an extrapolation of the low-flux PCA data rather than adopting a floor value for \({\dot{m}}_X\).) Having established a common scale for the two detectors, we then used the lowess nonparametric smoothing algorithm to derive a representation of the true, underlying light curve (which is at all points locally determined). Because the errors are grossly different for the PCA and ASM data, we achieve appropriate weighting via Monte-Carlo randomization: A lowess curve fit is derived for each of 1000 random realizations and the median curve is taken as the most representative fit. Here, we used a third-order lowess curve fit and adopted a window length of approximately one month, a choice motivated by the X-ray autocorrelation time (§ [3](#section:xcors){reference-type="ref" reference="section:xcors"}). The precise settings used in fitting the data only affect our results cosmetically. For example, with a window length of 3 days, we obtained the same-though noisier-results. # Correlated Variability {#section:xcors} With the unified X-ray light curve now in hand (i.e., \({\dot{m}}_X (t)\)), we use the unsmoothed data and examine the relationship between X-ray and OIR variability. Specifically, we assess the correlated variability between \({\dot{m}}_X\) and the OIR bands as a function of time lag using the discrete correlation function (DCF) of, with all observation times referred to the Solar System barycenter. A DCF is derived from a pair of time series, \(A_i\) and \(B_j\), by first establishing a set of lags \(\tau\) and cross-fluxes \(\Upsilon\), which are calculated across all paired time differences amongst the data: \[\Delta t_{ij} = t_i-t_j, \label{eq:tlag}\] \[\Upsilon_{ij} = \frac{(A_i-\bar{A})(B_j-\bar{B})}{\sigma_A\sigma_B}, \label{eq:xflux}\] where \(\bar{A}\) and \(\bar{B}\) are the means of series \(A\) and \(B\), respectively. Similarly, \(\sigma_A\) and \(\sigma_B\) are the standard deviations of \(A\) and \(B\). The DCF for time lag \(\tau\) is then \[{\rm DCF}(\tau) = \frac{1}{N(\tau)}\sum_{\Delta t_{ij} \in \tau}\Upsilon_{ij}, \label{eq:dcf}\] where \(N(\tau)\) is the number of \(\Delta t_{ij}\) elements in the bin \(\tau\). We convert the OIR data into fluxes and then standardize all data sets to unity variance and zero mean and then compute a DCF between each OIR light curve and the (unsmoothed) \({\dot{m}}_X\) time series (§ [2.1](#subsection:mdot){reference-type="ref" reference="subsection:mdot"}), binning the time lag data into \(\sim\)`<!-- -->`{=html}1 day intervals. Given this sampling and given data that span a decade, correlated variability is discernible over timescales ranging from days to years. As shown in Figure [\[fig:xcors\]](#fig:xcors){reference-type="ref" reference="fig:xcors"}, the X-ray time lag for the strongest features in the cross correlation of \({\dot{m}}_X\) with the OIR flux in each individual band is \(\approx 2\) weeks. This value is intermediate between those found by B01 and. A possible explanation for the differences is the lag's luminosity dependence discussed in § [6](#section:discussion){reference-type="ref" reference="section:discussion"}. The lag of the X-ray flux is naturally explained as a consequence of the viscous time delay for a density perturbation to propagate from the outer disk, the location of peak OIR emission, to the vicinity of the black hole where the X-rays are produced. Why are the cross-correlation peaks so broad? An examination of the X-ray autocorrelation (Figure [\[fig:xauto\]](#fig:xauto){reference-type="ref" reference="fig:xauto"}) suggests an answer. The autocorrelation timescale over which the X-ray source brightens or dims, several weeks, is similar to the typical lag time of \(\approx 2\) weeks. This is not surprising given that the timescale for the inner disk to change from faint to bright--the autocorrelation time--should be commensurate with the time required for a parcel of gas to travel from the outer disk to the center, i.e., the viscous time. On longer timescales, first at \(\sim100\) days and \(160\) days, other features appear in the autocorrelation function (Figure [\[fig:xauto\]](#fig:xauto){reference-type="ref" reference="fig:xauto"}). As evidenced by alternating positive/negative harmonic peaks, the anti-correlation feature at 160 days, e.g., is likely to be at least partially related to radiative or mechanical feedback in the system, whereby periods of strong accretion inhibit the accretion flow onto the outer disk. This is also evident in the strong anti-correlation signature between OIR and X-ray bands at \(\sim-120\)d lag (where the optical behavior now follows after the X-ray). In the next section, we develop a model for the accretion-driven, \(\approx 2\) week time lag in the context of a simplistic physical picture of the disk-star system in LMC X--3. # The Accretion Model {#section:accmodel} Our purpose is to model the complex OIR light curves of LMC X--3 as a blend of three components of emission attributable to (1) the companion star; (2) a multitemperature thermal disk; and (3) the regions in the disk and star where X-rays are reprocessed. (We need not consider synchrotron emission here, as there is no evidence for a jet in LMC X--3, unlike many other sources.) Our model for the system is shown schematically in Figure [\[fig:schem1\]](#fig:schem1){reference-type="ref" reference="fig:schem1"}. We approximate the net emission \(F(t)\) in a given OIR band \(i\) as a sum of the three components: \[F_{i}(t) = X_i(t) + S_i(t) + H_i(t), \label{eq:flux}\] where \(X_i(t)\) is the OIR disk emission which relates to the later X-ray emission (or \({\dot{m}}_X\)) at time \(t+t_{\rm visc}\); \(S_i\) is the direct flux from the star plus any steady component of disk light; and \(H_i\) is the reprocessed emission from the outer disk and companion star. ## Stellar Component {#subsec:stellarmod} The simplest term, \(S_i\), is a constant plus sinusoidal terms that approximate the asymmetric ellipsoidal variability of the tidally distorted star: \[S_i(t) \equiv C_i + c_{i} \left[ {\rm sin}( 4\pi \phi-\pi/2) + \epsilon e(\phi) \right],\][\[eq:ellipse\]]{#eq:ellipse label="eq:ellipse"} where phase \(\phi \equiv (t-t_0)/P\) and \[e(\phi) \equiv {\rm max}\left(0, {\rm sin}(4\pi \phi + \pi/2)\right) \times \left\{ \begin{array}{l l} 1, & \quad {\rm sin}(2\pi \phi + \pi/4) > 0,\\ -1, & \quad {\rm sin}(2\pi \phi + \pi/4) \leq 0 \end{array} \right. \label{eq:asym}\] allow for the difference in the amplitudes of the two minima. Both \(\epsilon\) and \(c_i\), the normalizations respectively of the asymmetric and sinusoidal terms, are free parameters, as are \(t_0\), \(C_i\), and \(P\). ## Disk Component {#subsec:diskmod} The connection of the term \(X_i(t)\), the variable OIR component of disk emission, to \({\dot{m}}_X\) is obvious. However, \({\dot{m}}_X\) was derived for the inner X-ray-emitting portion of the disk, which is located far from the optical-emitting region, and the mass transfer rates may be different in the two regions. Therefore, in order to compute the OIR disk emission from \({\dot{m}}_X\), our model must include a description of how matter flows from the outer disk to the inner region. This is established via a transfer function which maps between X-ray and OIR regimes. We envision the disk described with such a model as a series of concentric rings. The accretion flow for each individual ring is approximately steady-state, but the steady-state solution varies from ring to ring. The order of the rings are fixed, but they are allowed to stretch or compress in width during inflow. Meanwhile, the observed, viscosity-induced time delay of the X-ray flux relative to the optical will have some dependence on the mass supply of the individual rings. In our generic prescription, we incorporate this dependence as a free power-law scaling on the time delay, which depends on \({\dot{m}}_X\), a choice that is explained further in Section [6](#section:discussion){reference-type="ref" reference="section:discussion"}. We likewise assume that the optical emission has a power-law dependence on \({\dot{m}}_X\): \[X_i(t) \equiv a_{i} \times \left[{\dot{m}}_X(t+\delta t(t)) /10^{18}g\;s^{-1}\right]^\beta, \label{eq:f_lag}\] with \(a_{i}\) and \(\beta\) as free parameters and \(\delta t(t)\) is the viscous lag at time \(t\), defined as \[\delta t (t) \equiv \Delta_i \left(\frac{{\dot{m}}_X(t+\delta t(t))}{< {\dot{m}}_X >} \right)^\psi, \label{eq:lag}\] where \(\Delta_i\) and \(\psi\) are free parameters. A slightly modified prescription for the accretion emission is considered in Appendix [\[section:model2\]](#section:model2){reference-type="ref" reference="section:model2"}. Finally, we allow for a potential scaling error in our conversion from X-ray flux to \({\dot{m}}_X\) by introducing a floating offset \(a_X\) which modifies the zero-point of \({\dot{m}}_X\). For simplicity, rather than introducing a new variable, \({\dot{m}}_X\) will be understood to contain an implicit free zero-point, which is determined in the fit. ## Reprocessed Emission {#subsec:xheatmod} The X-ray heating by the central source, which occurs on a timescale of seconds, is treated as instantaneous. That is, the heating term \(H_i(t)\) responds directly to changes in \({\dot{m}}_X(t)\) (zero time lag). We arbitrarily parameterize the dependence of \(H_i(t)\) on \({\dot{m}}_X\) as a broken power law. This choice allows for variable shielding by the disk as the X-ray luminosity varies. The heating term contains two elements: The primary term, which has no orbital phase dependence, describes X-ray heating of the outer disk; and the secondary term, \(Z(t)\), describes X-ray heating of the companion star. \(Z(t)\) depends on phase and is maximum at superior conjunction of the star. We assume that \(Z(t)\) varies sinusoidally and that its normalization constant \(h_i = h\) is the same for all of the OIR bands. This approximation is valid because for the B-type companion star the OIR bands all lie in the Rayleigh-Jeans part of the spectrum. \[H_i(t) \equiv h~c_{i}\left(1+Z(t)\right) \left\{ \begin{array}{l l l l} ({\dot{m}}_X(t))^{\gamma_1}, & {\dot{m}}_X(t) < \dot{m}_{\rm break}, \\ \dot{m}_{\rm break}^{\gamma_1} \left(\frac{{\dot{m}}_X(t)}{\dot{m}_{\rm break}}\right)^{\gamma_2}, & {\dot{m}}_X(t) > \dot{m}_{\rm break}, \end{array} \right. \label{eq:xheat}\] where \[Z(t) = \eta/2\left(1+{\rm sin}(2\pi\phi-\pi/2)\right), \label{eq:shadow}\] so that \(Z(t)\) varies between 0 and \(\eta\) each orbit, where \(\eta\), \(h\), \(\gamma_1\), \(\gamma_2\), and \(\dot{m}_{\rm break}\) are fit parameters. Our full model consists of a total of 27 free parameters (plus two additional calibration parameters to align the B01 and SMARTS datasets). There are 16 "chromatic" parameters that contain a subscript \(i\) (4 each per OIR band) and there are 11 "gray" parameters that are independent of wavelength (e.g., the orbital period \(P\)). Correlations amongst the fit parameters are discussed and shown in Appendix [\[append:correlations\]](#append:correlations){reference-type="ref" reference="append:correlations"}. # Results {#section:results} The model has been implemented in *python* and applied at once to the composite OIR data set. All fits are computed using the Markov Chain Monte-Carlo (MCMC) routine emcee-hammer in a Bayesian formalism. Because our model is an oversimplification of complex real processes, we expect the quality of the fits to be limited by systematic effects. We therefore adopt 10% and 20% (systematic) error bars for the optical and IR data, respectively, and we ignore the much smaller measurement errors. These round values were chosen by making a preliminary fit to the data and assessing the rms fit residuals. Because the uncertainties were estimated in this utilitarian way, any goodness-of-fit statistics-such as \(\chi^{2}/\nu\)-should be interpreted with reserve. We use a flat prior (i.e., uniform weighting) on the period; for all other parameters, we use informed priors whenever one is evident. Otherwise, we default to flat priors on the logarithm of normalization terms (i.e., scale-independent weighting), and flat priors on shape parameters. The fitting results, which are based on our analysis of all the OIR data, are presented in Table [\[tab:fits\]](#tab:fits){reference-type="ref" reference="tab:fits"}. MCMC directly computes the posterior probability distribution of a model's parameters, but it does not specifically optimize the fit quality. Therefore, the best-fitting value of \(\chi^{2}/\nu\) given in Table [\[tab:fits\]](#tab:fits){reference-type="ref" reference="tab:fits"} has been obtained via other standard optimization methods (e.g., Levenberg--Marquardt, downhill-simplex, etc.), while taking the MCMC results as a starting point. To achieve our fits via emcee-hammer, we used 500 walkers with 15000 steps apiece for a total of 7.5 million MCMC samples. Our analysis and results are based on the final 2.5 million samples. Convergence has been diagnosed following using a stringent criterion of \(\tilde{R} < 1.1\). For our analysis, across all parameters, \(1.049 < \tilde{R} < 1.085\). In Figure [\[fig:panels\]](#fig:panels){reference-type="ref" reference="fig:panels"}, we show for the \(V\)-band light curve (the band with the most data) the best-fit model broken down into its component contributions. The figure makes clear that the total flux is dominated by the constant component of stellar light, while the disk and X-ray-heating component contributions are at most \(\sim 1/2\) and \(\sim 1/4\) as large, respectively (as determined in the fit). The *maximum* contribution of X-ray heating to the total stellar flux is \(h{{\dot{m}}_{\rm break}}^{\gamma_1} \sim 0.16\)%. Meanwhile, the contribution to the stellar flux due to ellipsoidal variability is \(c_i/C_i \approx 5-10\%\) (Section [4.1](#subsec:stellarmod){reference-type="ref" reference="subsec:stellarmod"}), while the asymmetric term is a factor \(\sim 10\) smaller still. B01 conclude that the optical variability of LMC X-3 is due to viscous disk emission rather than X-ray reprocessing, i.e., that reprocessing is of secondary importance. Our results support B01's conclusion. Our fits indicate that as the X-ray flux varies over its full range that the fraction of stellar light to the total light in both the \(B\) and \(V\) bands is \(\sim 80\%\pm10\%\). A consistent value of \(\approx 70\%-90\%\) has been derived independently for a wide range of X-ray luminosities using spectroscopic data (@Jerry_LMCX3). The value of the scaling index \(\beta = 1.31\pm0.08\) (90% confidence), which relates the OIR flux to the time-lagged disk emission \({\dot{m}}_X\) (Section [4.2](#subsec:diskmod){reference-type="ref" reference="subsec:diskmod"}), may indicate that the rates of inflow at the outer and inner radii are not matched. If so, the implication is that mass is lost from the body of the disk, especially at high values of \({\dot{m}}_X\). This scenario is readily explained by the action of disk winds (e.g., @Miller_2006_H1743 [@Luketic_2010; @Ponti_2012; @Neilsen_2012]). Alternatively, as discussed in § [6](#section:discussion){reference-type="ref" reference="section:discussion"}, a standing shock from the accretion stream could be responsible for the mismatch in rates. From the flux densities computed in Table [\[tab:fits\]](#tab:fits){reference-type="ref" reference="tab:fits"}, very rough constraints can be placed on the temperature of the star and of the OIR-emitting region of the disk. However, such calculations are fundamentally limited by the accuracy of the absolute OIR flux calibration, which is marred by 20--30% zero-point calibration differences between SMARTS and B01 data sets. We rely on the SMARTS calibration as our standard rather than that of B01 because the former calibration is based on a larger sample of standard stars that were observed more frequently. In Figure [\[fig:denoised\]](#fig:denoised){reference-type="ref" reference="fig:denoised"}, we show two versions of the OIR light curves folded on LMC X--3's orbital period: On the left are the light curves in their original form, and on the right are the filtered light curves produced by removing the two nonstellar contributions, i.e., the \({\dot{m}}_X\)-induced component \(X(t)\) and the reprocessed X-ray emission \(H(t)\). Plainly, the ellipsoidal variability is much more apparent in the filtered light curves. Because of the long baseline and abundance of data, the signal evident from the filtered light curves (using the complete model) allows one to determine the \(1.7\)-day orbital period to the remarkable precision of two-tenths of a second. (Prior to this work, the best determination of the period has been \(P=1.7048089 \pm 1.1\times10^{-6}\) d (@Song_2010); our result is in good accordance with theirs.) The orbital phase, however, conforms to its prior and is otherwise unconstrained. The quality of our period determination is illustrated in Figure [\[fig:period\]](#fig:period){reference-type="ref" reference="fig:period"} where it is shown to compare favorably with an independent determination based on three decades of spectroscopic velocity data. As indicated in the figure, the two data sets jointly determine the period to a precision of 90 milliseconds (\(P = 1.704808 \pm 1\times 10^{-6}\) d). The raw light curves, by comparison, cannot even deliver a unique orbital period because of strong aliasing. A Lomb-Scargle search shows that the strongest false period is favored 9:1 over the true period. Moreover, the uncertainty in the period determination is a factor \(\sim 3\) worse than achieved using the light curve model. # Discussion {#section:discussion} ## Reprocessing and Self-Shadowing The intensity of the X-ray heating component in the OIR bands is a surprisingly strong function of X-ray luminosity, scaling as \(H_i \propto {\dot{m}}_X^{1.9}\), up to a critical value of \({\dot{m}}_X \approx 6.2 \times 10^{18}\) g/s (roughly \(\sim 50\%\) of the Eddington limit). Above this luminosity, the heating signal drops off rapidly as \({\dot{m}}_X^{-3.4}\) such that at the maximum X-ray luminosity (\(\approx L_{\rm Edd}\)) its intensity is a factor of 10 below its peak. A likely explanation for the initial superlinear rise with luminosity is that the flaring of the disk at higher luminosity increases the solid angle of the reprocessing region, and that this effect dominates over the self-shadowing by the disk. Above \(\sim 50\% L_{\rm Edd}\), the shadowing dominates and the heating signal dwindles rapidly. For the shadowing to be substantial without producing more OIR emission than is observed, the flare in the disk's scale height must move inward as the X-ray luminosity increases. That is, we envision a flared funnel-like region in the inner disk that contracts in around the black hole as luminosity increases and as the flaring becomes more pronounced. The OIR emission from the reprocessing region remains modest, even at the highest luminosities, because the reprocessing excess is relegated to shorter wavelengths and emitted from within a diminished area of the disk. A prediction of this picture is that the pattern of reprocessing shifting to smaller area and shorter wavelength at growing luminosity should be observable in optical / UV for those low-inclination systems which show substantial luminosity variability. As viewed from the X-ray source, the star subtends an angle of \(\sim 37\degr\) (ignoring disk obscuration). The corresponding fraction of the X-ray heating signal that is reprocessed in the face of the star is obtained from the fit: \(\eta/(1+\eta) \approx 40\%\) (Equations [\[eq:xheat\]](#eq:xheat){reference-type="ref" reference="eq:xheat"},[\[eq:shadow\]](#eq:shadow){reference-type="ref" reference="eq:shadow"}). The reprocessed emission from the star can be distinguished from the reprocessed disk emission because the former is modulated at the orbital period. For simplicity, the fraction of the total reprocessed emission contributed by the star is treated as a constant, free parameter. ## Wavelength-Independence of the Lag In canonical thin accretion disk theory, the wavelength of peak emission from the disk scales as \(\lambda_{\rm max}(R) \propto R^{3/4}\), whereas for irradiation-dominated disks the peak wavelength scales roughly as \(\lambda_{\rm max}(R) \propto R^{3/7}\). On this basis, one expects a factor \(\sim 2\) variation in the viscous time lag across the OIR bands. However, as evident from Table [\[tab:fits\]](#tab:fits){reference-type="ref" reference="tab:fits"} and from inspection of Fig. [\[fig:xcors\]](#fig:xcors){reference-type="ref" reference="fig:xcors"}, the OIR time lag is essentially independent of wavelength from the \(B\)-band (\(\approx4400\) Å) to the \(J\)-band (\(\approx12,500\) Å)[^4]. The constancy of the viscous time lag with wavelength implies that the OIR signal arises from a single radius in the disk. This result is incompatible with any simple disk theory for which temperature falls off monotonically with radius, unless one imagines locating the radius of OIR emission at the disk's outer edge. However, the outer edge is obviously ruled by the data because the entire disk would then be so hot that it would outshine the star. ## Shocks and Hotspots in the Outer Disk Instead, the constancy of the time lag with wavelength is naturally explained by the presence of a "hot spot" or similar structure within the outer disk that is bright enough to dominate the OIR lag signal. Doppler modulation tomographic maps of accreting binary systems show clear evidence for such features in other sources (e.g., @Steeghs_2004 [@Calvelo_2009; @Kupfer_2013]). An excellent example is provided by the map of the black hole binary A0620--00, which reveals bright spots embedded in a hot "ring" in the disk (dominated by two large crescent structures; @Neilsen_2008). The ring is located at 45% or 60% of the L1 radius (\(R_{\rm L1}\)) depending on whether one assumes, respectively, a Keplerian or ballistic trajectory for the gas stream. This range in radius exactly brackets the *circularization radius* of the system, the radius at which the angular momentum of the tidal stream matches the angular momentum of the gas in the disk. The circularization radius \(R_{\rm C}\) is located at 50% \(R_{L1}\) for A0620--00, while the truncation radius of the outer disk is predicted to be located at \(\sim 80-90\)% \(R_{L1}\). A schematic diagram of a system like A0620--00 is shown in Figure [\[fig:schem2\]](#fig:schem2){reference-type="ref" reference="fig:schem2"}. The temperatures of these bright spots and rings are poorly constrained in quiescent black-hole binaries, and they are essentially unconstrained in active systems. In order to crudely gauge the temperature of such a spot, we look to the recent work by where for a cataclysmic variable (CV) system they have measured \(T_{\rm hotspot} \approx 30,000\) K. Such hot spots are attributed to shocks produced by the tidal stream of gas from the companion star impinging on the disk. A strong shock (Mach number \(\mathcal{M} > 10\)) can result in up to a hundredfold jump in temperature in the post-shock gas. If the mass flow rate in the stream \(\dot{m_{\rm s}}\) is sufficiently large, a shock that traverses inward from \(R_{\rm out}\) will eventually stall at the circularization radius \(R_C\). For simplicity, we assume that the flow is purely ballistic, i.e., that the velocity is unaffected by \(\dot{m_{\rm s}}\), and that the hotspot luminosity increases as \(\dot{m_{\rm s}}\) increases. We now mention an alternative explanation for the superlinear scaling between OIR and X-ray emission (\(\beta = 1.3\)), which in Section [5](#section:results){reference-type="ref" reference="section:results"} we attributed to disk winds. Here, where the OIR emission originates in a shock, we instead envision that this nonlinear relationship could plausibly result from changes in the structure or strength of the shock caused by variations in \(\dot{m_{\rm s}}\). ## Estimating \(\alpha\)-viscosity The cooling time for the shock-heated gas is negligibly short compared to the viscous timescale. Therefore, the stability and brightness of these structures are related to the instantaneous value of \(\dot{m_{\rm s}}\). The ring itself must be relatively narrow owing to the prompt cooling. The cooled gas relaxes to the disk profile, and then it proceeds inward to the center on the viscous timescale. The outermost disk is always dominated by gas pressure so that the scale height at each radius is given, roughly, by the ratio of the sound speed in the midplane, \(c_s=\sqrt{kT/\mu m_p}\), to the Keplerian velocity in the disk. Using the \(\alpha\) disk approximation for disk viscosity, \[\nu = \alpha c_s z, \label{eq:alphadisk}\] where \(z\) is the scale height of the disk. Then, \[\nu \equiv \alpha c_s^2 \frac{R^{3/2}}{\sqrt{G M}}, \label{eq:visc}\] and the viscous timescale can be written as \[t_{\rm visc} \approx \sqrt{G M R} \left(\frac{\mu m_p}{\alpha k T}\right). \label{eq:tviscx}\] In what follows, recall that we are approximating the evolving disk as a superposition of independent, steady-state solutions (Section [4.2](#subsec:diskmod){reference-type="ref" reference="subsec:diskmod"}). We make two other assumptions, which are eminently reasonable: (1) The disk is strongly illuminated; and (2) \(\dot{m_{\rm s}}\) alone determines the X-ray and OIR emission. Building on these assumptions and our measured value of the time lag, we now derive a rough estimate of \(\alpha\) in the outer disk. We first consider the conditions in the disk and the location of the OIR-emitting hotspots. The strong X-ray irradiation of the disk will modify its structure at large distances from the central source, i.e., \(R > 10^{10}\) cm (@Hartmann_book). Meanwhile, the midplane temperature, being less sensitive to irradiation, will tend to follow the viscous dissipation profile, \(T_{\rm m} \propto R^{-3/4}\). However, the temperature of the irradiated surface layer of the disk will fall off much more slowly, as \(T_{\rm s} \propto R^{-3/7}\), or even as slowly as \(T_{\rm s} \propto R^{-1/3}\). To our knowledge, the vertical temperature gradient in strongly irradiated disks has not been given serious consideration. Generally, \(T_{\rm s}\) and \(T_{\rm m}\) in the outer disk, determined from irradiation and viscous dissipation, respectively, are comparable and within a factor of a few of one another, and it is reasonable to approximate the vertical temperature profile as isothermal (e.g., @Cunningham_1976 [@King_1998]). For the average mass accretion rate onto LMC X--3, we estimate for any reasonable set of disk parameters that to within a factor \(\sim3\) the midplane temperature \(T_{\rm C} \sim 10^5\) K at \(R_{\rm C}\). For this temperature and radius, using Equations [\[eq:alphadisk\]](#eq:alphadisk){reference-type="ref" reference="eq:alphadisk"}-[\[eq:tviscx\]](#eq:tviscx){reference-type="ref" reference="eq:tviscx"} we find to order of magnitude that \(\alpha \sim 0.5\). Our result agrees well with other measurements of \(\alpha\), particularly those obtained for CVs (@King_2007, and references therein). The most reliable of these other measurements, like our estimate for LMC X-3, were derived for irradiated outer disks. ## \(\alpha\) Scales with Luminosity A surprising outcome of our model and analysis is the positive luminosity scaling of the time lag (\(\psi \approx 0.25\); Eq. [\[eq:lag\]](#eq:lag){reference-type="ref" reference="eq:lag"}). Recalling that \(t_{\rm visc} \propto \dot{m}_X^{\psi}\), Eq. [\[eq:tviscx\]](#eq:tviscx){reference-type="ref" reference="eq:tviscx"} gives \(\alpha t_{\rm visc} \propto T^{-1} \propto \dot{m}_X^{-1/4}\). For a constant value of \(\alpha\), one expects to find \(\beta =-0.25\), whereas we find a positive value of \(\beta\). As a bottom line, we find that the viscosity parameter varies with luminosity (or equivalently, mass accretion rate) as \(\alpha \propto {\dot{m}}_X^{-1/2}\). An inverse scaling such as we predict has in fact, already been suggested for LMC X--3 based purely on changes in its disk spectra as the luminosity varies. If this scaling is related to the increase in temperature which results from a higher \({\dot{m}}_X\), then \(\alpha\) may be lower in the inner disk[^5]. However, GRMHD simulations indicate that \(\alpha\) increases as radius decreases, at least for the innermost disk where MHD turbulence is dominant. Of course, there is no reason to expect viscosity to be constant over the disk. In particular, the viscosity may depend on other factors such as the disk's density, temperature or magnetization. provide one possible explanation for the scaling relation we find relating \(\alpha\) and luminosity. These authors find an inverse scaling between \(\alpha\) and the ratio of gas-to-magnetic pressures, \(P_{\rm gas}/P_{\rm mag}\) (a net quantity usually referred to as "plasma \(\beta\)"). Given the scaling, \(\alpha \propto P_{\rm mag}/P_{\rm gas}\), if \(P_{\rm mag}\) is sufficiently insensitive to the accretion flow in the outer disk, then one would expect that as the mass transfer rate (\(\dot{m}\)) varies, when \(\dot{m}\) increases, \(P_{\rm gas}\) will too. A more specific prediction would be beyond scope of this work, but we note qualitatively that this effect very naturally gives rise to an inverse relationship between \(\alpha\) and mass accretion rate, as required by our fit. # Conclusions {#section:concs} Motivated by a roughly two-week time-lagged correlation between the OIR and X-ray light curves of LMC X--3, we develop a new method of analysis that is applicable to active X-ray binary systems. We model the OIR emission deterministically as a combination of accretion emission, X-ray reprocessing, and stellar emission. The components driven by accretion are computed using the X-ray light curve. The model allows accretion signals to be filtered out of the OIR light curves, thereby isolating the stellar component of light. The utility of this technique is demonstrated for LMC X--3, a system which exhibits large-amplitude, but simple, variability in its perennial thermal state. We demonstrate that this method improves the observability of the stellar ellipsoidal light curves and leads to an improved determination of the orbital period. Furthermore, the method allows one to disentangle the OIR component of disk emission from the reprocessed X-ray emission. The time lags in the system are independent of wavelength, which indicates that they originate from a hot ring at a single, fixed radius in the disk. We identify the radius of this ring as the circularization radius where the tidal stream from LMC X--3's B-star companion meets the disk, inducing a bright shock. By interpreting our results through the lens of \(\alpha\)-disk models, we estimate the viscosity in the outer disk: Based upon the average properties of the disk, we estimate \(\alpha \sim 0.5\) to order magnitude. Furthermore, we unexpectedly find that \(\alpha\) diminishes as luminosity increases (\(\alpha \propto {\dot{m}}_X^{-1/2}\)). We speculate that this result may be related to changes in \(P_{\rm gas}/P_{\rm mag}\) related to evolution in the mass accretion rate in the outer disk (e.g, @Bai_Stone_2011). It is a pleasure to thank Poshak Gandhi, Chris Done, Ramesh Narayan, Andy Fabian, and Joey Neilsen for helpful discussions which have improved this work, as well as the anonymous referee, for her/his helpful review. We thank Catherine Brocksopp for making her B01 data set available to us, and Xuening Bai for his input on disk theory. Support for JFS has been provided by NASA Hubble Fellowship grant HST-HF-51315.01. [^1]: http://xte.mit.edu/ASM_lc.html [^2]: The other fit parameter of the disk component, spin, is allowed to vary, but is distributed with modest scatter about a fixed value. A separate paper is forthcoming on the spin of LMC X--3. [^3]: Roughly 0.7% of the data, 11 data points, fall \(>5\sigma\) below the curve. These exceptional spectra show an unusually strong Compton component, more than an order of magnitude brighter than is typical for the disk-dominated states of LMC X--3. [^4]: The wavelength-independence of the lag has been verified by examining sub-intervals of the data as well. [^5]: On the other hand, this dependence cannot be *purely* due to an inherent scaling with temperature, i.e. \(\alpha \propto T^{-2}\), without the \(\alpha\)-disk interpretation breaking down. A weaker scaling with temperature would be compatible.

{'timestamp': '2014-01-23T02:02:43', 'yymm': '1401', 'arxiv_id': '1401.5529', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5529'}

# Introduction Can cooperation in a one-shot Prisoner's Dilemma be justified between rational agents? Rapoport argued in the 1960s that two agents with mutual knowledge of each others' rationality should be able to mutually cooperate. Howard explains the argument thus: > Nonetheless arguments have been made in favour of playing C even in a single play of the PD. The one that interests us relies heavily on the usual assumption that both players are completely rational and know everything there is to know about the situation. (So for instance, Row knows that Column is rational, and Column knows that he knows it, and so on.) It can then be argued by Row that Column is an individual very similar to himself and in the same situation as himself. Hence whatever he eventually decides to do, Column will necessarily do the same (just as two good students given the same sum to calculate will necessarily arrive at the same answer). Hence if Row chooses D, so will Column, and each will get 1. However if Row chooses C, so will Column, and each will then get 2. Hence Row should choose C. Hofstadter described this line of reasoning as "superrationality", and held that knowledge of similar cognitive aptitudes should be enough to establish it, though the latter contention is (to say the least) controversial within decision theory. However, one may consider a stronger assumption: what if each agent has some ability to predict in advance the actions of the other?\ \ This stronger assumption suggests a convenient logical formalism. In the 1980s, Binmore considered game theory between Turing machines which had access to one anothers' Gödel numbers[^1]: > \...a player needs to be able to cope with hypotheses about the reasoning processes of the opponents other than simply that which maintains that they are the same as his own. Any other view risks relegating rational players to the role of the "unlucky" Bridge expert who usually loses but explains that his play is "correct" and would have led to his winning if only the opponents had played "correctly". Crudely, rational behavior should include the capacity to exploit bad play by the opponents.\ > In any case, if Turing machines are used to model the players, it is possible to suppose that the play of a game is prefixed by an exchange of the players' Gödel numbers. Howard and McAfee considered the Prisoner's Dilemma in this context[^2], and each presented an example of an algorithm which would always return an answer, would cooperate if faced with itself, and would never cooperate when the opponent defected. (The solution discussed in both papers was a program that used quining of the source code to implement the algorithm "cooperate if and only if the opponent's source code is identical to mine"; we represent it in this paper as Algorithm [\[cliquebot\]](#cliquebot){reference-type="ref" reference="cliquebot"}, which we call CliqueBot on account of the fact that it cooperates only with the 'clique' of agents identical to itself.)\ \ More recently, Tennenholtz reproduced this result in the context of other research on multi-agent systems, noting that CliqueBot can be seen as a Nash equilibrium of the expanded game where two players decide which code to submit to the Prisoner's Dilemma with source-code swap. This context (called "program equilibrium") led to several novel game-theoretic results, including folk theorems by Fortnow and Kalai, Kalai, Lehrer and Samet, an answer by Monderer and Tennenholtz to the problem of seeking strong equilibria (many-agent Prisoner's Dilemmas in which mutual cooperation can be established in a manner that is safe from *coalitions* of defectors), a Bayesian framework by Peters and Szentes, and more.\ \ However, these approaches have an undesirable property: they restrict the circle of possible cooperators dramatically---in the most extreme case, only to agents that are syntactically identical! (In a moment, we will see that there are many examples of semantically distinct agents such that one would wish one's program to quickly cooperate with each of them.) Thus mutual cooperation is inherently brittle for CliqueBots, and an ecology of such agents would be akin to an all-out war between incompatible cliques.\ \ This problem can be patched somewhat, but not cured, by prescribing a list of agents with whom mutual cooperation is desirable, but this approach is inelegant and requires all of the relevant reasoning to happen outside of the system. We'd like to see agents that can reason on their own somewhat.\ \ A natural-seeming strategy involves simulating the other agent to see what they do when given one's own source code. Unfortunately, this leads to an infinite regress when two such agents are pitted against one another.\ \ One attempt to put mutual cooperation on more stable footing is the model-checking result of van der Hoek, Witteveen, and Wooldridge, which seeks "fixed points" of strategies that condition their actions on their opponents' output. However, in many interesting cases there are several fixed points, or none at all, and so this approach does not correspond to an algorithm as we would like.\ \ Since the essence of this problem deals in counterfactuals---e.g. "what would they do if I did this"---it is worth considering modal logic, which was intended to capture reasoning about counterfactuals, and in particular the Gödel-Löb modal logic **GL** with provability as its modality. (See Boolos and Lindström for some good references on **GL**.) That is, if we consider agents that cooperate if and only if they can prove certain logical formulas, the structure of logical provability gives us a genuine framework for counterfactual reasoning, and in particular a powerful and surprising tool known as Löb's Theorem: We shall see that Löb's Theorem enables a flexible and secure form of mutual cooperation in this context. In particular, we first consider the intuitively appealing strategy "cooperate if and only if I can prove that my opponent cooperates", which we call FairBot. If we trust the formal system used by FairBot, we can conclude that it is unexploitable (in the sense that it never winds up with the sucker's payoff). When we play FairBot against itself (and give both agents sufficient power to find proofs), although either mutual cooperation or mutual defection seem philosophically consistent[^3], it always finds mutual cooperation (Theorem [\[FBFB\]](#FBFB){reference-type="ref" reference="FBFB"})![^4] Furthermore, we can construct another agent after the same fashion which improves on the main deficit of the above strategy: namely, that FairBot fails to correctly defect against CooperateBot.[^5] We call this agent PrudentBot.\ \ Moreover, the underpinnings of this result (and the others in this paper) do not depend on the syntactical details of the programs, but only on their semantic interpretations in provability logic; therefore two such programs can cooperate, even if written differently (in several senses, for instance if they use different Gödel encodings or different formal systems). Accordingly, we define a certain class of algorithms, such as FairBot and PrudentBot, whose behavior can be described in terms of a modal formula, and show that the actions these "modal agents" take against one another can be described purely in terms of these modal formulas. Using the properties of Kripke semantics, one can algorithmically derive the fixed-point solutions to the action of one modal agent against another; indeed, the results of this paper have additionally been checked by a computer program written by two of the authors, hosted at [github.com/klao/provability](github.com/klao/provability).\ \ We next turn to the question of whether a meaningful sense of optimality exists among modal agents. Alas, there are several distinct obstacles to some natural attempts at a nontrivial and non-vacuous criterion for optimality among modal agents. This echoes the impossibility-of-optimality results of Anderlini and Canning on game theory for Turing machines with access to each others' source codes.\ \ All the same, the results on Löbian cooperation represent a formalized version of robust mutual cooperation on the Prisoner's Dilemma, further validating some of the intuitions on "superrationality" and raising new questions on decision theory. The Prisoner's Dilemma with exchange of source code is analogous to Newcomb's problem, and indeed, this work was inspired by some of the philosophical alternatives to causal and evidential decision theory proposed for that problem (see Drescher and Altair ).\ \ A brief outline of the structure of this paper: in Section [2](#formal){reference-type="ref" reference="formal"}, we define our formal framework more explicitly. In Section [3](#playfair){reference-type="ref" reference="playfair"}, we introduce FairBot, prove that it achieves mutual cooperation with itself and cannot be exploited (Theorem [\[FBFB\]](#FBFB){reference-type="ref" reference="FBFB"}); we then introduce PrudentBot, and show that it is also unexploitable, cooperates mutually with itself and with FairBot, and defects against CooperateBot.\ \ In Section [4](#Modal Agents){reference-type="ref" reference="Modal Agents"}, we develop the theory of modal agents, with a focus on showing that their action against one another is well-defined. We also show that a feature of PrudentBot---namely, that it checks its opponent's response against DefectBot---is essential to its functioning: modal agents which do not use third parties cannot achieve mutual cooperation with FairBot unless they also cooperate with CooperateBot. Then, in Section [5](#optimality){reference-type="ref" reference="optimality"}, we discuss several obstacles to proving nontrivial optimality results for modal agents. In Section [6](#PhilosophicalConclusions){reference-type="ref" reference="PhilosophicalConclusions"}, we will explain our preference for PrudentBot over FairBot, and speculate on some future directions, before closing in Section [7](#OpenProblems){reference-type="ref" reference="OpenProblems"} with a list of open problems in this area. # Agents in Formal Logic {#formal} There are two different formalisms which we will bear in mind throughout this paper. The first formalism is that of algorithms, where we can imagine two Turing machines `X` and `Y`, each of which is given as input the code for the other, and which have clearly defined outputs corresponding to the options \(C\) and \(D\). (It is possible, of course, that one or both may fail to halt, though the algorithms that we will discuss will provably halt on all inputs.) This formalism has the benefit of concreteness: we could actually program such agents, although the ones we shall deal with are often very far from efficient in their requirements. It has the drawback, however, that proofs about algorithms which call upon each other are generally difficult and untidy, relying upon delicate bounds on (e.g.) the length of proofs.\ \ Therefore, we will do our proofs in another framework: that of logical provability in certain formal systems. More specifically, the agents we will be most interested in can be interpreted via modal formulas in Gödel-Löb provability logic, which is especially pleasant to work with. This bait-and-switch is justified by the fact that all of our tools do indeed have equivalently useful bounded versions; variants of Löb's Theorem for bounded proof lengths are well-known among logicians. The interested reader can therefore construct algorithmic versions of all logically defined agents in this paper, and with the right parameters all of our theorems will hold for such agents.\ \ In particular, our "agents" will be formulas in Peano Arithmetic, and our criterion for action will be the existence of a finite proof in the tower of formal systems \(\textsf{PA+n}\), where \(\textsf{PA}\) is Peano Arithmetic, and \(\textsf{PA+(n+1)}\) is the formal system whose axioms are the axioms of \(\textsf{PA+n}\), plus the axiom that [PA+n]{.sans-serif} is consistent, i.e. that \(\neg\Box\dots\Box\bot\) with \(n+1\) copies of \(\Box\).\ \ Fix a particular Gödel numbering scheme, and let `X` and `Y` each denote well-formed formulas with one free variable. Then let \(X(Y)\) denote the formula where we replace each instance of the free variable in `X` with the Gödel number of `Y`. If such a formula holds in the standard model of Peano Arithmetic, we interpret that as `X` cooperating with `Y`; if its negation holds, we interpret that as `X` defecting against `Y`. (In particular, we will prove theorems in [PA+n]{.sans-serif} to establish whether the agents we discuss cooperate or defect against one another.) Thus we can regard such formulas of arithmetic as decision-theoretic agents, and we will use "source code" to refer to their Gödel numbers. Of course, it is easy to create `X` and `Y` so that \(X(Y)\) is an undecidable statement in all [PA+n]{.sans-serif} (e.g. the statement that the formal system \(\textsf{PA+}\omega\) is consistent). But the philosophical phenomenon we're interested in can be achieved by agents which do not present this problem, and whose finitary versions in fact always return an answer in finite time.\ \ Two agents which are easy to define and clearly decidable are the agent which always cooperates (which we will call `CooperateBot`, or `CB` for short) and the agent which always defects (which we will call `DefectBot`, or `DB`). In pseudocode: Howard, McAfee and Tennenholtz introduced functionally equivalent agent schemas, which we've taken to calling `CliqueBot`; these agents use quining to recognize self-copies and mutually cooperate, while defecting against any other agent. In pseudocode: By the diagonal lemma, there exists a formula of Peano Arithmetic which implements `CliqueBot`. (The analogous tool for computable functions is Kleene's recursion theorem ; in this paper, we informally use "quining" to refer to both of these techniques.)\ \ `CliqueBot` has the nice property that it never experiences the sucker's payoff in the Prisoner's Dilemma. This is such a clearly important property that we will give it a name: However, `CliqueBot` has a notable drawback: it can only elicit mutual cooperation from agents that are syntactically identical to itself. (If two `CliqueBots` were written with different Gödel-numbering schemes, for instance, they would defect against one another!)\ \ One might patch this by including a list of source codes (or a schema for them), and cooperate if the opponent matches any of them; one would of course be careful to choose only source codes that would cooperate back with this variant. But this is a brittle form of mutual cooperation, and an opaque one: it takes a predefined circle of mutual cooperators as given. For this reason, it is worth looking for a more flexibly cooperative form of agent, one that can deduce for itself whether another agent is worth cooperating with. # Löbian Cooperation {#playfair} A deceptively simple-seeming such agent is one we call `FairBot`. On a philosophical level, it cooperates with any agent that can be proven to cooperate with it. In pseudocode: `FairBot` references itself in its definition, but as with `CliqueBot`, this can be done via the diagonal lemma. By inspection, we see that `FairBot` is unexploitable: presuming that Peano Arithmetic is sound, `FairBot` will not cooperate with any agent that defects against `FairBot`.\ \ The interesting question is what happens when `FairBot` plays against itself: it intuitively seems plausible either that it would mutually cooperate or mutually defect. As it turns out, though, Löb's Theorem guarantees that since the `FairBot`s are each seeking proof of mutual cooperation, they both succeed and indeed cooperate with one another! (This was first shown by Vladimir Slepnev.) However, it is a tidy logical accident that the two agents are the same; we will understand better the mechanics of mutual cooperation if we pretend in this case that we have two distinct implementations, \(\texttt{FairBot}_1\) and \(\texttt{FairBot}_2\), and prove mutual cooperation from their formulas without using the fact that their actions are identical. However, `FairBot` wastes utility by cooperating even with `CooperateBot`. (See Section [6](#PhilosophicalConclusions){reference-type="ref" reference="PhilosophicalConclusions"} for the reasons we take this as a serious issue.) Thus we would like to find a similarly robust agent which cooperates mutually with itself and with `FairBot` but which defects against `CooperateBot`.\ \ Consider the agent `PrudentBot`, defined as follows: # Modal Agents {#Modal Agents} It is instructive to consider `FairBot` and `PrudentBot` as modal statements in Gödel-Löb provability logic (often denoted **GL**). Namely, if we consider the actions of `FairBot` and any other agent `X` against one another, then the definition of `FairBot` is simply \([FB(X)=C]\leftrightarrow \Box [X(FB)=C]\), and the definition of `PrudentBot` is \([PB(X)]\leftrightarrow (\Box[X(PB)]\wedge \Box(\neg\Box\bot\to\neg[X(DB)])).\) There are a number of tools, including fixed-point theorems and Kripke semantics, which work for families of such modal statements; and thus we will define a class of *modal agents* for the purpose of study.\ \ Informally, a modal agent is one whose actions are determined solely by the provability of statements regarding its opponent's actions against itself and against other simpler agents[^6]. That is, if `X` is a modal agent, then there is a modal-logic formula[^7] \(\varphi\) and a fixed set of simpler modal agents \(\texttt Y_1,\dots,\texttt Y_N\) such that, for any opponent \(\texttt Z\), \[\label{modalagent} [X(Z)=C]\leftrightarrow \varphi\left([Z(X)=C],[Z(Y_1)=C],\dots,[Z(Y_N)=C]\right).\] Furthermore, since a modal agent does all of this via provability, the formula \(\varphi\) must be *fully modalized*: all instances of variables must be contained inside sub-formulas of the form \(\Box\psi\).\ \ We must lay some groundwork (following Lindström ) before formally defining the class of modal agents. Write \(\varphi(p_1,\dotsc,p_n)\) to denote a formula \(\varphi\) in the language of [GL]{.sans-serif} whose free (propositional) variables are included in the set \(\{p_1,\dotsc,p_n\}\). (Note that this is different from Lindström's convention, who doesn't display the free variables.) Write \(\Box^+\varphi\) to mean \(\varphi\wedge\Box\varphi\). If \(\texttt X\) and \(\texttt Y\) are agents, we write \([X(Y)=C]\), or \([X(Y)]\) for short, for the application of \(\texttt X\) to the Gödel number of \(\texttt Y\); we interpret this logical formula as the assertion that \(\texttt X\) cooperates when playing against \(\texttt Y\). Accordingly, we say that `X` cooperates with `Y` if \([X(Y)]\) holds in the standard model of [PA]{.sans-serif}. `CooperateBot`, `DefectBot`, `FairBot` and `PrudentBot` are all modal agents, but as we shall see, `CliqueBot` is not.\ \ We now prove three theorems demonstrating that the notion of "modal agent" has good properties. First, we note that it makes no practical difference if we include proofs about the actions \([X(Z)]\) and \([Y_i(Z)]\) in our definition: Next, we show that the actions two modal agents take against each other are described by a fixed point of their modal formulas, as one would expect. In particular, this shows that modal agents' actions against each other depend only on their modal formulas, not on other features of their source code. Finally, we show that modal agents' actions depend only on their opponents' *behavior*, not on other features of their source code. It feels a bit clunky, in some sense, for the definition of modal agents to include references to other, simpler modal agents. Could we not do just as well with a carefully constructed agent that makes no such outside calls (i.e. a modal agent of rank 0)?\ \ Surprisingly, the answer is no: there is no modal agent of rank 0 that achieves mutual cooperation with `FairBot` and defects against `CooperateBot`. In particular: # Obstacles to Optimality {#optimality} It is worthwhile to ask whether there is some meaningful sense of "optimality" for logical agents or modal agents in particular. For many natural definitions of optimality, this is impossible. For instance, there is no `X` such that for all \(\texttt{Y}\), the utility achieved by `X` against \(\texttt{Y}\) is the highest achieved by any `Z` against \(\texttt{Y}\). (To see this, consider \(\texttt{Y}\) defined so that \(Y(Z)=C\) if and only if `Z`\(\neq\)`X`.) More generally, an agent can "punish" or "reward" other agents for any arbitrary feature of their code.\ \ Could we hope that `PrudentBot` might at least be optimal among modal agents in some meaningful sense? As it happens, there are at least three different kinds of impediments to optimality among modal agents, which together make it very difficult to formulate any nontrivial and non-vacuous definition of optimality.\ \ Most directly, for any modal agents \(\texttt{X}\) and \(\texttt{Y}\), either their outputs are identical on all modal agents, or there exists a modal agent `Z` which cooperates against \(\texttt{X}\) but defects against \(\texttt{Y}\). (For an enlightening example, consider the agent `TrollBot` which cooperates with `X` if and only if \(\textsf{PA}\vdash X(DB)=C\).) Thus any nontrivial and nonvacuous concept of optimality must be weaker than total dominance, and in particular it must accept that third parties could seek to "punish" an agent for succeeding in a particular matchup.\ \ Another issue is illustrated by the following agent: That is, `JustBot` cooperates with `X` if and only if `X` cooperates with `FairBot`. (Note that `JustBot` has different source code from `FairBot`; in particular, it can use a hard-cooded reference to `FairBot`'s code, where `FairBot` must use a quine.) Clearly, `JustBot` is exploitable by some algorithm (in particular, consider the non-modal algorithm which cooperates only with the corresponding `FairBot` and with nothing else), but since it is behaviorally equivalent to `FairBot`, by Theorem [\[behavioral\]](#behavioral){reference-type="ref" reference="behavioral"} it cannot be exploited by any modal agent.\ \ A third problem is that a modal agent has a finite amount of predictive power, and it can fail to act optimally against sufficiently complicated or slow-moving opponents. Explicitly, consider the family of modal agents \(\texttt{WaitFairBot}_K\), defined by \[[WaitFairBot_K(X)=C]\leftrightarrow ((\neg\Box^{K+1}\bot)\wedge \Box(\neg\Box^K\bot\to [X(WaitFairBot_K)=C])).\] As it happens[^8], any modal agent which defects against `DefectBot` will fail to achieve mutual cooperation with \(\texttt{WaitFairBot}_K\) for all sufficiently large \(K\).\ \ Despite these reasons for pessimism, we have not actually ruled out the existence of a nontrivial and non-vacuous optimality criterion which corresponds to our philosophical intuitions about "correct" decisions. Additionally, there are a number of ways to depart only mildly from the modal framework (such as allowing quantifiers over agents), and these could invalidate some of the above obstacles. # Philosophical Digressions {#PhilosophicalConclusions} One might ask (on a philosophical level) why we object to FairBot in the first place; isn't it a feature, not a bug, that this agent offers up cooperation even to agents that blindly trust it? We suggest that it is too tempting to anthropomorphize agents in this context, and that many problems which can be interpreted as playing a Prisoner's Dilemma against a CooperateBot are situations in which one would not hesitate to "defect" in real life without qualms.\ \ For instance, consider the following situation: You've come down with the common cold, and must decide whether to go out in public. If it were up to you, you'd stay at home and not infect anyone else. But it occurs to you that the cold virus has a "choice" as well: it could mutate and make you so sick that you'd have to go to the hospital, where it would have a small chance of causing a full outbreak! Fortunately, you know that cold viruses are highly unlikely to do this.[^9] If you map out the payoffs, however, you find that you are in a Prisoner's Dilemma with the cold virus, and that it plays the part of a CooperateBot. Are you therefore inclined to "cooperate" and infect your friends in order to repay the cold virus for not making you sicker?\ \ The example is artificial and obviously facetious, but not entirely spurious. The world does not come with conveniently labeled "agents"; entities on scales at least from viruses to governments show signs of goal-directed behavior. Given a sufficiently broad counterfactual, almost any of these could be defined as a CooperateBot on a suitable Prisoner's Dilemma. And most of us feel no compunction about optimizing our human lives without worrying about the flourishing of cold viruses.[^10]\ \ In a certain sense, PrudentBot is actually "good enough" among modal agents that one might expect to encounter: there are bound to be agents (CooperateBot and DefectBot) whose action fails to depend in any sense upon predictions of other agents' behavior, and other agents (FairBot, PrudentBot, etc) whose action depends meaningfully on such predictions. One should not expect to encounter a TrollBot or JustBot arising naturally! But it's worth pondering if this reasoning can be made formal in any elegant way.\ \ Do these results imply that sufficiently intelligent and rational agents will reach mutual cooperation in one-shot Prisoner's Dilemmas? In a word, no, not yet.[^11] Many things about this setup are notably artificial, most prominently the perfectly reliable exchange of source code (and after that, the intractably long computations that might perhaps be necessary for even the finitary versions).\ \ Nor does this have direct implications among human beings; our abilities to read each other psychologically, while occasionally quite impressive, bear only the slightest analogy to the extremely artificial setup of modal agents. Governments and corporations may be closer analogues to our agents (and indeed, game theory has been applied much more successfully on that scale than on the human scale), but the authors would not consider the application of these results to such organizations to be straightforward, either. The theorems herein are not a demonstration that a more advanced approach to decision theory (i.e. one which does not fail on what we consider to be common-sense problems) is practical, only a demonstration that it is possible. # Open Problems {#OpenProblems} In particular, here are some open problems we have come across in this area: - Is there a natural, nonvacuous, and nontrivial definition of optimality among modal agents? - Are there tractable ways of studying agents which can incorporate quantifiers as well as modal operators? For example, we might consider the non-modal agent `X` such that `X` cooperates with `Y` iff some formal system proves both that `Y` is unexploitable (given the consistency of some other formal system) and that `Y` cooperates with `X`. - What different dynamics arise when we consider the analogues of modal agents in more complicated games than the Prisoner's Dilemma? In particular, there are issues raised by games with more than one "superrational equilibrium", like the Coordination Game. The natural analogues of FairBot and PrudentBot transform any finite game between themselves into a coordination or bargaining game, but do not provide insight on how to resolve those sorts of conflicts. - What happens if we apply probabilistic reasoning rather than provability logic? As this allows for mixed strategies, it introduces all of the complexities of bargaining games, as well as new ones. - What differs in games with more than two players; in particular, what might coordination and bargaining among coalitions look like? It is easier to imagine two agents with each others' source code agreeing to cooperate in a Prisoner's Dilemma than it is to imagine three agents with each others' source code agreeing on how to subdivide a fixed prize (which they lose if they do not have a majority agreeing on an acceptable split).

{'timestamp': '2014-01-23T02:05:32', 'yymm': '1401', 'arxiv_id': '1401.5577', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5577'}

# Introduction Finite convex sets are important in various areas of basic and applied mathematics, and appear in linear state-space processes. One area of application is the first-order DeGroot discrete-time state-space process in which the state of each point of a set \(n\) points, \({\bf{x}}_{i}(k+1),\; i = 1...,n\), is a convex combination of the immediately prior states of all points of the system \({\bf{x}}_{j}(k) \in \mathbb{R}^{n},\; j = 1...,n\). The model was formulated as a mechanism by which consensus might be reached among a set of individuals, and it has become the benchmark model of the literature on opinion dynamics. Its precursors include the models of French and Harary. In addition, the model has become increasingly prominent in control theory. When the mechanism unfolds in an aperiodic irreducible network, the system converges to a single value on the real number line and, more generally for an \({\bf{X}}(k) \in \mathbb{R}^{n \times m}\), to a single location in \(m\)-dimensional space. The second-order convex combination mechanism that is employed in this paper includes the DeGroot mechanism as a special case. It allows the emergence of complex objects composed of discrete points distributed in \(m\)-dimensional space. The emergent configurations of points include all polytopes (geometric objects with flat sides) and all other arrangements of points in the convex hull of the initial \({\bf{X}}(0)\) state of the system. In its application to the literature on opinion dynamics, these objects correspond to feasible patterns of influenced opinions that may include opinion clusters, each with a different mean opinion, and differentiated factions, each with a different consensus on an issue. More broadly, with respect to control theory applications, the model may be employed to move a set of \(n\) points, in discrete steps, from initial positions in \(m\)-dimensional space to any specified configuration of points in the convex hull of \({\bf{X}}(0)\). The control occurs in the framework of a time-invariant network structure with an evolving matrix polynomial that is adjusted by a diagonal matrix of values. The diagonal values, specific to each point of the system, control the values of the walks in the network structure. In an irreducible structure, the detailed typology (its particular configuration of edges) does not constrain the feasible set of specified end-states of the system, nor do the magnitudes of initial distances among the points. # Second-Order Convex Combination Mechanism Let \({\bf{X}}(0) \in \mathbb{R}^{n \times m}\) and \({\bf{V}}(k) \in \mathbb{R}^{n \times n}\). Then \({\bf{X}}(k) = {\bf{V}}(k){\bf{X}}(0)\) is a set of points with \(m\) coordinates in the convex hull of \({\bf{X}}(0)\) for all \({\bf{V}}(k)\) in the domain of nonnegative matrices with rows sums constrained to 1. Such \({\bf{V}}(k)\) are generated by the second order discrete-time state-space process \[{\bf{X}}(k+1) = {\bf{A}}{\bf{W}}{\bf{X}}(k) + ({\bf{I}}-{\bf{A}}){\bf{X}}(0), k=0,1,2..., \label{eq:FJ}\] where \({\bf{W}}\) is a nonnegative matrix with row sums constrained to 1, and \({\bf{A}}\) is a diagonal matrix constrained to \({\bf{0}} \le {\bf{A}} \le {\bf{I}}\). The process presents an evolving matrix polynomial \[\begin{aligned} {\bf{V}}(k) & = {\bf{AW}}{\bf{V}}(k-1) + ({\bf{I}}-{\bf{A}}), \;\; {\bf{V}}(0) = {\bf{I}}, \;\; k=1,2..., \\ & = ({\bf{AW}})^k + \big[{\bf{I}} + {\bf{AW}} + ({\bf{AW}})^2 +... + ({\bf{AW}})^{k-1} \big] ({\bf{I}}-{\bf{A}}),\;k>2 \notag \end{aligned}\] which corresponds to walks in a network structure and preserves each \({\bf{V}}(k)\) as a nonnegative matrix with rows sums of 1: \[\begin{aligned} {\bf{X}}(1) & = {\bf{V}}(1){\bf{X}}(0), \;\;{\bf{V}}(1) = {\bf{AW}} + {\bf{I}}-{\bf{A}}, \notag \\ & = \big[{\bf{AW}} + {\bf{I}}-{\bf{A}}\big]{\bf{X}}(0), \notag \\ {\bf{X}}(2) & = {\bf{V}}(2){\bf{X}}(0),\;\; {\bf{V}}(2) = {\bf{AW}}{\bf{V}}(1) + {\bf{I}}-{\bf{A}}, \notag \\ & = \big[ {\bf{AW}}{\bf{V}}(1) + {\bf{I}}-{\bf{A}} \big] {\bf{X}}(0), \notag \\ & = \big[ ({\bf{AW}})^2 + \big({\bf{I}} + {\bf{AW}} \big)({\bf{I}}-{\bf{A}}) \big] {\bf{X}}(0), \notag \\ & \vdots \notag \\ {\bf{X}}(k) & = {\bf{V}}(k){\bf{X}}(0),\;\; k > 2 \notag \\ & = \big[({\bf{AW}})^k + \big({\bf{I}} + {\bf{AW}} + ({\bf{AW}})^2 +... + ({\bf{AW}})^{k-1} \big)({\bf{I}}-{\bf{A}}) \big] {\bf{X}}(0), \notag \\ & \vdots \notag \\ {\bf{X}}(\infty) & = {\bf{V}}{\bf{X}}(0), \;\; {\bf{V}} \equiv \mathop{\lim }\limits_{k \to \infty } {\bf{V}}(k) \;\; \text{if this limit exists.} \end{aligned}\] Because \({\bf{V}}(1) = {\bf{AW}} + {\bf{I}}-{\bf{A}}\) is row-stochastic, all \({\bf{W}}{\bf{V}}(k-1)\) are row-stochastic, and all \({\bf{V}}(k) = {\bf{AW}}{\bf{V}}(k-1) + ({\bf{I}}-{\bf{A}}), \;\; k>1,\) are row-stochastic. The sequence \(\{ {\bf{V}}(k); k = 0,1... \}\) converges if and only if the \(\mathop{\lim }\limits_{k \to \infty } ({\bf{AW}})^k\) exists. The spectral radius \(\rho({\bf{AW}})\) is controlled by \({\bf{A}}\). - If \({\bf{A}} = {\bf{0}}\), then \({\bf{V}} = {\bf{I}}\), and \({\bf{X}}(\infty) = {\bf{X}}(0)\). - If \({\bf{A}} = {\bf{I}}\), i.e, \({\bf{AW}}\) is stochastic, then \({\mid {\lambda} \mid}_{\text{max}} = 1\) (Frobenius), the sequence \(\{ {\bf{W}}^{k};k=0,1,... \}\) converges, and \(\{ {\bf{V}}(k);k=0,1... \}\) converges, if and only if the eigenvalues of \({\bf{W}}\) for which \(\mid {\lambda} \mid = 1\) are all \(\lambda = 1\). - If \({\bf{0}} \le {\bf{A}} \le {\bf{I}}\), i.e., \({\bf{AW}}\) is sub-stochastic, then \({\mid {\lambda} \mid}_{\text{max}} \le 1\), (Frobenius), the sequence \(\{ ({\bf{AW}})^{k};k=0,1,... \}\) converges, and \(\{ {\bf{V}}(k);k=0,1... \}\) converges (a) if the eigenvalues of \({\bf{AW}}\) for which \(\mid {\lambda} \mid = 1\) are all \(\lambda = 1\) or (b) if \(\lambda = 1\) is not included in the spectrum. In the latter case, \(( {{\bf{I}}-{\bf{AW}}})\) is nonsingular and \({\bf{V}} = { ( {{\bf{I}}-{\bf{AW}}})^{-1}}({ {\bf{I}}-{\bf{A}} })\). - If \({\bf{0}} \le {\bf{A}} < {\bf{I}}\), i.e, \({\bf{AW}}\) is strictly sub-stochastic, then \({\mid {\lambda} \mid}_{\text{max}} < 1\) (Frobenius), the sequence \(\{ ({\bf{AW}})^{k};k=0,1,... \}\) converges, and \(\{ {\bf{V}}(k);k=0,1... \}\) converges to \({\bf{V}} = { ( {{\bf{I}}-{\bf{AW}}})^{-1}}({ {\bf{I}}-{\bf{A}} })\). In this case, \(\mathop{\lim }\limits_{k \to \infty } ({\bf{AW}})^k = {\bf{0}}\), and \({\bf{V}} = \big[ \sum_{k=0}^{\infty} ({\bf{AW}})^k \big]({\bf{I}}-{\bf{A}})\) involves the Newman series. In this framework, if \({\bf{W}}\) is aperiodic and irreducible, and \({\bf{A}} = {\bf{I}}\), then \({\bf{V}} = \mathop{\lim }\limits_{k \to \infty } {\bf{W}}^{k}\). The Perron-Frobenius theorem applies and gives the *simple* object, one point, with the coordinates \({\bf{X}}(\infty)\), in the convex hull of \({\bf{X}}(0)\). If \({\bf{W}}\) is aperiodic and irreducible, and \({\bf{0}} < {\bf{A}} < {\bf{I}}\), then the result \[{\bf{X}}(\infty) = \big( {\bf{I}}-{\bf{AW}} \big)^{-1} ({\bf{I}}-{\bf{A}}) {\bf{X}}(0)\] is a *complex* object of discrete points, with the coordinates \({\bf{X}}(\infty)\). As \({\bf{A}}\) approaches \({\bf{I}}\), these complex objects converge to the simple object of the \({\bf{A}} = {\bf{I}}\) case. More generally, for a particular \(a_{ii}\), \(\mathop{\lim }\limits_{ {a_{ii}} \to 1 } = 1\) and \(\mathop{\lim }\limits_{ {a_{ii}} \to 0 } = 0\). Hence there is no loss of generality, and considerable gain of parsimony, entailed in the restriction \({\bf{0}} < {\bf{A}} < {\bf{I}}\). Every \({\bf{X}}(0)\) is associated with a polytope that is a box of \(m\)-dimensions, defined by the maximum and minimum values of each column of \({\bf{X}}(0)\). All \(i=1..,n\) points of the system might be located at the vertices of the box. The feasible initial coordinates of the \(n\) points of the system include, as special cases, all convex polytopes in \(m\) dimensions. These special cases occur when non-vacuous subsets of the \(n\) points of the system occupy each vertex of a particular type of polytope (e.g., a pentagon) and all points are initially located at these vertices. Holding \({\bf{X}}(0)\) and \({\bf{W}}\) constant, the state-space process will generate complex objects in the minimal convex set (convex hull) of \({\bf{X}}(0)\) for all \(0 < {\bf{A}} < {\bf{I}}\). It is immediately apparent that: - for given \(\{{\bf{W}},{\bf{X}}(0) \}\), the diagonal values of \({\bf{A}}\) control the system's outcome; - for given \(\{ {\bf{W}}, {\bf{0}} < {\bf{A}} < {\bf{I}},{\bf{X}}(\infty) \}\), there exists a unique \[\begin{aligned} {\bf{X}}(0) & = {\bf{V}}^{-1}{\bf{X}}(\infty) \notag \\ & = ({\bf{I}}-{\bf{A}})^{-1} \big( {\bf{I}}-{\bf{AW}} \big) {\bf{X}}(\infty), \label{eq:X0} \end{aligned}\] on which basis the state-space process will converge to the specified complex object \({\bf{X}}(\infty)\); and - for given \(\{ {\bf{W}}, {\bf{X}}(0),{\bf{X}}(\infty) \}\), there may be no feasible \({\bf{0}} < {\bf{A}} < {\bf{I}}\) that satisfies \[{\bf{X}}(\infty)-{\bf{X}}(0) = {\bf{A}} \big[ {\bf{W}}{\bf{X}}(\infty)-{\bf{X}}(0) \big].\] An implication of equation [\[eq:X0\]](#eq:X0){reference-type="ref" reference="eq:X0"} is that, for a given \({\bf{W}}\), there exists an infinite number of \(\{{\bf{A}},{\bf{X}}(0)\}\) combinations with which the state-space process will converge to a specified complex object \({\bf{X}}(\infty)\). I draw on this implication in an analysis of the emergence of specified complex objects via the state-space process. I restrict the analysis to aperiodic irreducible \({\bf{W}}\). Such matrices correspond to strongly connected networks in which at least one path from \(i\) to \(j \ne i\) exists for all ordered \((i,j)\) pairs of nodes. Given a strong structure, a sufficient condition of aperiodicity is the existence of at least one positive value on its main diagonal, which corresponds to at least one positive resistance loop in the network. It is significant that the particular details of the typology of the network associated with \({\bf{W}}\), i.e., its idiosyncratic configuration of edges, do not matter. The detailed typology may be important in other respects, but it is not important with respect to reaching the end state of the process that is the specified complex object. The remainder of paper is organized into two main sections and concludes with a brief discussion: 1. The first section illustrates various complex objects that may be formed under the structural conditions of an irreducible aperiodic \({\bf{W}}\). 2. The second section addresses the particular combinations \(\{{\bf{A}},{\bf{X}}(0)\}\) with which the state-space process will converge to a specified complex object \({\bf{X}}(\infty)\) for a given \({\bf{W}}\). # Complex Objects The complex objects that arise within the polytope associated with \({\bf{X}}(0)\) are \(n\) points with \(m\) coordinates. If undirected edges are added to every \((i,j)\) unordered pair of discrete points, then the object is a simplex with \(\binom {n} {2}\) 1-faces. If a subset of the \(\binom {n} {2}\) possible edges is added, then special cases of emergent objects within the convex hull of \({\bf{X}}(0)\) may be obtained depending on how the edges are drawn. If no edges are added, then the complex object is array of points with coordinates that locate them in a \(m\)-dimensional space. The dimensions of this space may be geographical, or more generally dimensions of point-states on \(m\) variables. In the special case of a \(1\)-dimensional \({\bf{X}}(0)\) of point-states on one variable, the complex object is an array of points on the real number line, and this array may be displayed as a distribution of the number points located in positions (intervals) of the line. I leave the application open and add no edges to the generated complex objects. Figures 1 and 2 are based on the same \({\bf{W}}\) and \({\bf{X}}(0)\). Figure 1 shows (a) a random array of initial positions (squares), (b) the simple object generated in the special case of \({\bf{A}} = {\bf{I}}\) (plus sign), and (c) the reduced compression effect of \({\bf{A}} = 0.80{\bf{I}}\) (solid circles). Figure 2 is based on a two-value \((0.10, 0.80)\) \({\bf{A}}\), which generates the reduced convergence of Figure 1 for a subset of \(n\) and more reduced convergence for the remainder. The \({\bf{A}}\) values control the complex object within the convex hull of \({\bf{X}}(0)\), conditional on \({\bf{W}}\). Figures 3 and 4 display more orderly emergent complex objects that are resultants of particular combinations of \({\bf{A}}\) and \({\bf{X}}(0)\). Fewer initial positions than the size \(n\) of the system appear when, as here, subsets of \(n\) occupy the same initial positions. Figures 5 and 6 illustrate that if the initial space is organized as a polytope (here a triangle and a pentagon), with non-vacuous subsets of \(n\) that occupy each vertex of the polytope, then the emergent complex object of the state-space process is constrained to the subspace of the polytope within the \(m\)-box. Figure 7 displays a large scale \(1\)-dimensional application and its bar chart results. The result bears on Abelson's unsolved problem in the field of opinion dynamics. His investigation of various models showed that a formal explanation of emergent consensus on specific issues is easily obtained, but that a formal explanation of emergent differentiated opinion clusters is a difficult, to his apparent consternation: "Since universal ultimate agreement is an ubiquitous outcome of a very broad class of mathematical models, we are naturally led to inquire what on earth one must assume in order to generate the bimodal outcomes of community cleavage studies". One answer appears to be that such cleavage is based on an initial distribution of opinions on an issue concentrated around a moderate initial position that is then "pulled apart" by a second order interpersonal influence process unfolding in a strong structure. Figure 8 displays the \({a_{ii}} \; i=1...,n\) values that are involved in the process. The cleavage is based on a moderate mass of individuals who vary in their \({a_{ii}}\) values, and extremists with \({a_{ii}}\) values homogeneously near 1. # Target Complex Objects Given \(\{ {\bf{W}},{\bf{X}}(0), {\bf{X}}(\infty) \}\), each diagonal value of \({\bf{A}}\), i.e., \(a_{11}...,a_{nn}\), must satisfy equation [\[eq:X0\]](#eq:X0){reference-type="ref" reference="eq:X0"}. In scalar form, each \(a_{ii}\) must simultaneously satisfy the constraints entailed in \[0 < {a_{ii}} = \frac{{x_{ih}}(\infty)-{x_{ih}}(0)} {\sum_{j=1}^{n}{{w_{ij}} {x_{jh}}(\infty)-{x_{ih}}(0)} } < 1, \;\; i=1...,n; \;\; h=1..,m. \label{eq:scalaraiisol}\] A specified "target" is feasible if and only if, for all \(i=1...,n; h=1..,m\), (a) the denominator is not zero, the numerator and denominators are of the same sign, and the absolute value of the numerator is less than or equal to the absolute value of the denominator. The domain of feasible "targets" is, therefore, subject to strong constraints, and it may be vacuous. In contrast, given \(\{ {\bf{W}}, {\bf{X}}(\infty) \}\), there exists an infinite number of \(\{{\bf{A}},{\bf{X}}(0)\}\) combinations with which the state-space process will converge to a specified complex object \({\bf{X}}(\infty)\). The infinite set of feasible \(\{{\bf{A}} ,{\bf{X}}(0)\}\) combinations is defined by \[{x_{ih}}(0) = \frac{{x_{ih}}(\infty)-{a_{ii}} \sum_{j=1}^{n}{ {w_{ij}} {x_{jh}}(\infty) }}{1-a_{ii}}, \;\; a_{ii} \in (0,1).\] for \(i=1...,n; \;\; h=1..,m\). With an arbitrary \(a_{ii} \in (0,1)\) all values of \({x_{ih}}(0), \; h=1...,m\) are determined for \(i\). The set of arbitrary \(a_{ii},\; i=1...,n\) values fully specifies a feasible \(\{{\bf{A}} ,{\bf{X}}(0)\}\) combination on the basis of which the state-space process will generate the specified \({\bf{X}}(\infty)\). Whether particular classes of combinations have properties that are preferable to others will depend on the application. In this infinite set of feasible \(\{{\bf{A}} ,{\bf{X}}(0)\}\) combinations, an attractive unbiased combination is obtained with \[{\bf{X}}(0) = ({\bf{I}}-{\bf{A}})^{-1} \big( {\bf{I}}-{\bf{AW}} \big) {\bf{X}}(\infty), \;\; {\bf{A}} = (1/2){\bf{I}}.\] This special case presents \[{x_{ih}}(0) = 2{x_{ih}}(\infty)-\sum_{j=1}^{n}{ w_{ij} {x_{jh}}(\infty) },\] for which the values of \({\bf{X}}(0)\) are neither inflated or deflated by the choice of \({\bf{A}}\), and \[{x_{ih}}(\infty) = \frac{ {x_{ih}}(0)-\sum_{j=1}^{n}{ w_{ij} {x_{jh}}(\infty) }}{2}\] for all \(i=1...,n; \;\; h=1..,m\). However, heterogeneous \(a_{11}...,a_{nn}\) values may be theoretically important. An affine transformation of the specified system \({\bf{X}}(\infty) = {{\bf{V}}}{\bf{X}}(0)\) is permissible \[\alpha + \beta {\bf{X}}(\infty) = {{\bf{V}}}(\alpha + \beta {\bf{X}}(0)),\] because the scalars \(\{ \alpha, \beta \}\) pass through the system without altering \({\bf{V}}\). # Discussion Analysis has been restricted to the case of aperiodic irreducible structures. This restriction may be relaxed to include the aperiodic reducible structures covered by a nonsingular \({\bf{I}}-{\bf{AW}}\), i.e, all aperiodic strong, unilateral, weak, and disconnected structures. In strong structures, the details of network typology do not affect the outcome; every node of a strong structure is a point basis of the network (a minimal subset of nodes from which all its nodes of the network are reachable). However, the point bases of unilateral and weak structures may be more complex, and the \(a_{ii}\) dampening values of those nodes that constitute a point basis of the structure are especially important. In general, network typology affects the temporal trajectories of the state-space. Network typology also is an important factor in the sensitivity and vulnerability of complex objects to structural disturbances. Structural disturbances include changes of \({\bf{W}}\), \({\bf{A}}\), or \({\bf{X}}(0)\) during the state-space process. Here, it may be noted that the \(v_{ij}\) of \({\bf{V}}\) correspond to the relative net influence of node \(j\) on node \(i\). Thus, \[{\bf{r}} = \frac{1}{n}{{\bf{V}}^{T}}{\bf{1}}, \;\; {{\bf{r}}^{T}} {\bf{1}} = 1,\] is the average relative net influence of node \(j\) on all nodes of the system, and may taken as a measure of the centrality of node \(j\) in the system. In the special case of \({\bf{A}}={\bf{I}}\), where \({\bf{V}} = \mathop{\lim }\limits_{k \to \infty } {\bf{W}}^k\), \({\bf{r}}\) is the left eigenvector of \({\bf{W}}\) associated with \(\lambda = 1\). In the special case of \({\bf{A}}=\alpha{\bf{I}}, \; \; 0< \alpha < 1\), \[\begin{aligned} \label{eq:TEC} {\bf{r}} & = \frac{1}{n}{{\bf{V}}^{T}}{\bf{1}}, \\ & = \frac{1}{n} ( {\bf{I}}-\alpha {\bf{W}}^{T})^{-1} {\bf{1}} ( 1-\alpha ), \notag \end{aligned}\] which may expressed as follows \[{\bf{r}} = \frac{1-\alpha}{n} + \alpha{\bf{W}}^{T} {\bf{r}} \label{eq:PageRank}\] It is difficult for me to see how the PageRank equation [\[eq:PageRank\]](#eq:PageRank){reference-type="ref" reference="eq:PageRank"} may be treated as an invention when equation [\[eq:TEC\]](#eq:TEC){reference-type="ref" reference="eq:TEC"} was published prior to it. I noted in the introduction that the convex hull of \({\bf{X}}(0)\) defines a space within which the initial coordinates of the \(n\) points of the system include, as special cases, all convex polytopes in \(m\) dimensions. I illustrated this in Figures 5 and 6, where the initial coordinates of the system presented a triangle and pentagon with each vertex occupied by a subset of \(n\), and with all \(i=1...,n\) located in the vertices. In such cases, all \({\bf{A}}\) will generate an end-state \({\bf{X}}(\infty)\) of coordinates, via the state-space process, that locates all \(n\) points within the constrained subspace of the initial polytope. Conversely, it is interesting that solutions for "target" \({\bf{X}}(\infty)\) frequently present a reduced simplex for \({\bf{X}}(0)\) with fewer than \(n\) vertices, i.e, at least one vertex is occupied by multiple points with the same initial positions. I have not addressed the temporal trajectory of \({\bf{X}}(k)\) as the state-space process alters the coordinates of the nodes. Effects of the detailed typology of \({\bf{W}}\) on these movements are of interest. Additionally, the structure of \({\bf{W}}\) (e.g., its point sets, cut sets, and disjoint path redundancies) is significant in determining the sensitivity and vulnerability of the system to in-process structural disturbances that alter the system's end-state. The present work suggests that in-process modifications of the nodal dampening values \({\bf{A}}\), which alter the values of the walks, may be used to readjust the system so that the points arrive at their specified destinations. Every walk of length \(\ell\) in a network involves a path of lessor or equal length. As a particular \(a_{ii}\) approaches zero, all paths involving edges from node \(i\) are deactivated. In an irreducible network, the limits of a particular pair of values \(({a_{ii}},{a_{jj}})\) present four binary combinations that control the existence of effects of the state of \(i\) on the state \(j\), and vice versa, for all \(i\) and \(j\). Generalizing to all \(\binom {n} {r}\) combinations, \({\bf{A}}\) entirely controls the causal structure of direct and indirect influences of nodes' past states on other nodes' future states. A different implementation of the model (equation [\[eq:FJ\]](#eq:FJ){reference-type="ref" reference="eq:FJ"}) couples \({\bf{A}}\) and \({\bf{W}}\) with the assumption \({a_{ii}} = 1-{w_{ii}}\) for all \(i\). With such coupling, each node's resistance value \({w_{ii}}\), i.e., the structural value of its loop, corresponds to the extent to which the node is open or closed with respect to flows from it. Under this assumption \({\bf{A}} = {\bf{I}}\) corresponds to a \({\bf{W}}\) with a zero main diagonal, and \({\bf{A}} = {\bf{0}}\) corresponds to a \({\bf{W}}\) with a main diagonal of ones. Changes of \({\bf{A}}\) induce limited changes of \({\bf{W}}\). The analysis that has been presented in this paper is fundamentally altered by this assumption, i.e., equation [\[eq:scalaraiisol\]](#eq:scalaraiisol){reference-type="ref" reference="eq:scalaraiisol"} is altered. However, the associated analysis is tractable and reinforces the importance of the \({\bf{A}}\) construct in controlling system outcomes. I have not addressed this implementation in the present paper.

{'timestamp': '2014-01-22T02:10:03', 'yymm': '1401', 'arxiv_id': '1401.5339', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5339'}

# Introduction The firewall paradox introduced by Almheiri, Marolf, Polchinski, and Sully (AMPS) is a wonderful tool for sharpening our understanding of black holes. I believe we will emerge from the firewall debate with either a shocking new quantum gravity effect or a sharper understanding of black hole complementarity. The AMPS paradox is essentially a conflict between the following: - To preserve unitarity, the outgoing Hawking radiation must carry information. There must be enough distinct allowed states of the quantum fields near the black hole to allow the radiation to carry information. - A freely falling observer naively expects to encounter a quantum state in the near-horizon zone that differs from the vacuum only at long wavelengths comparable to the Schwarzschild radius of the black hole. So unitarity requires many allowed quantum states for the fields in the near-horizon zone, while free infall requires a state that is almost unique. In the remainder of the paper, I refer to the near-horizon region between the stretched horizon and \(r \approx 3 G M\) simply as "the zone.\" Here I perform an elementary counting of the number of quantum states that are consistent with the naive expectations of an infalling observer. This is a simple quantum field theory analysis along the lines of the work of Unruh many years ago. I find that an infalling observer expects to encounter a flux of ingoing, negative energy excitations of the vacuum in the zone. These negative energy excitations are responsible for reducing the mass, and therefore the entropy, of the black hole, so arranging the correct quantum state for these excitations is sufficient for unitary evaporation of black holes. Doing the analysis in the infalling frame is convenient because this frame filters out the thermal noise that is present in the outside description and naturally picks out a subset of the zone degrees of freedom that may carry information. To roughly describe the states, every outgoing s-wave in the zone has about a 50% chance of making it to infinity and a 50% chance of reflecting off the geometry and returning to the black hole. (The probability is frequency-dependent, but for the frequencies that dominate Hawking radiation the probability of reflection is of order \(1/2\).) It is perfectly consistent with the equivalence principle for any given quantum to be either reflected or transmitted. So the allowed states can be specified roughly by which s-waves reflect off the barrier. Roughly, if we restrict to frequencies of order the temperature, we can specify the state of each ingoing s-wave photon in the zone. A natural question is how many distinct modes can be excited in the zone at a given time. The answer depends somewhat on the choice of time slicing, but I will show that slices of constant Schwarzschild time contain of order \(\log S\) bits of information. One might worry that these information-carrying excitations will violate the equivalence principle, in that freely falling observers will encounter high-energy quanta. In fact, I show that observers who begin their free fall from at least of order the Schwarzschild radius away from the horizon do not encounter high-energy quanta in these states. On the other hand, observers who somehow begin their free fall from a location very close to the horizon \(do\) encounter high-energy excitations, although they are "frozen\" rather than burned. Such observers are highly boosted relative to the natural rest frame determined by the geometry, and the equivalence principle does not protect them. Therefore, this freezing is consistent with the equivalence principle, unlike the "s-wave firewall\" discussed in AMPS, which affects even observers who fall in freely from infinity. Negative energy excitations are possible in quantum field theory, but they must be compensated by nearby regions of positive energy. In this case, the positive energy is behind the black hole horizon. This makes it clear that these ingoing excitations must be created in a manner that is delocalized over the entire zone. So before worrying about delicate questions of unitarity, the description of the infalling observer is that the outgoing Hawking radiation is created in a manner that is delocalized over the entire zone. Very roughly, in the infalling description the region around \(r = 3 GM\) sends positive energy particles out to infinity and negative energy excitations into the black hole. To give some further intuition, from the outside point of view, I imagine that the stretched horizon is constantly having a delicate conversation with the potential barrier, carefully tuning itself so that the correct quanta are reflected off the barrier. Another analogy is to think of the region between the stretched horizon and the barrier as a kind of electromagnetic cavity. The details of the boundary conditions at the stretched horizon and the barrier control the precise spectrum of electromagnetic excitations in the zone. From this point of view, we can think of the negative energy density in the zone as the result of the Casimir effect. In order to use these states to avoid the firewall paradox, one must assume that most degrees of freedom in the zone of an old black hole are entangled with the stretched horizon rather than the early radiation. Consider high angular momentum modes in the zone. These modes are emitted from the stretched horizon, go a short distance into the zone, and then fall back in. It is natural to think that these modes are entangled with the stretched horizon. However, at late times the entire black hole plus zone system (suitably defined) has fewer degrees of freedom than the Hawking radiation that has already been emitted. Therefore, if the dynamics of the black hole are completely scrambling, then the black hole and the zone must separately be almost maximally entangled with the early radiation and not each other. Therefore, we require nontrivial dynamics of the stretched horizon-essentially it should contain special degrees of freedom that are "bound\" to the excitations in the zone and do not mix with the other degrees of freedom. This means that most modes in the zone are entangled with the stretched horizon, while the special quanta that are going to escape to infinity are entangled with the early radiation. While this is not the simplest possibility for the dynamics of the stretched horizon, a slight complication in a system we know nothing about is a small price to pay for preserving the equivalence principle. It is important to clarify which aspects of this paper are speculative, and which are not. The statement that certain freely falling observers encounter negative energy density is not speculative: it is true in effective field theory, and follows from the properties of the Unruh vacuum, as I describe in more detail in section 2. In addition, the estimates of how many states are consistent with the equivalence principle rely only on well-known physics. However, in effective field theory the choice of state for an evaporating black hole \(is\) uniquely determined by the conditions that it is nonsingular at the future horizon and that no flux is incident from infinity. In order to access the different states consistent with the equivalence principle would require, for example, subtle modifications of the reflection coefficient for quantum fields reflecting off the barrier. I am not aware of any argument within effective field theory for such a modification. Therefore, as I describe in much more detail below, ideas about how the information gets into the ingoing quanta remain highly speculative. #### Conflict with effective field theory? "Evolving back\" \(vs.\) "evolving in.\" One major problem remains: if we zoom in close to the horizon and focus on only the low angular momentum modes, there is a clean division between ingoing and outgoing waves. The states I described above contain information in the ingoing modes, while the outgoing modes are exactly thermal. This is perfectly consistent with unitarity of black hole evaporation until one asks \(how\) the ingoing modes got into just the right state to preserve unitarity. The geometry outside the black hole should contain no information because of the no-hair theorem, so if the outgoing modes are exactly thermal then the ingoing modes cannot contain information. Someone has to put the information into the ingoing modes. More concretely, if the outgoing Hawking quanta can be "evolved back\" until they are localized very close to the horizon, as proposed by AMPS, then information must be stored in \(outgoing\) quanta in the zone. But in the vacuum state, outgoing quanta near the horizon should be nearly maximally entangled with their partners behind the horizon, so they cannot carry information. This is the firewall paradox of AMPS. I present two related arguments pointing out what I believe is an important gap in the argument. First, if we literally evolve the field operator corresponding to the outgoing Hawking radiation backward in time, it will evolve to an operator smeared over the entire backward lightcone (or at least its boundary), and not localized near the horizon. Instead of evolving back, we can try to "evolve in.\" In other words, instead of evolving the data backward in time, we can try to evolve it radially inward. Knowing the black-hole S-matrix roughly gives us all the information about the field modes far from the black hole. We can then try to use the data at large r to constrain the data near the black hole. "Evolving in\" is somewhat familiar from AdS/CFT, but it is a much more delicate operation than standard time evolution. Even in classical field theories, the question of evolving data radially inward is nontrivial. I explain a simple, geometric criterion from the theory of partial differential equations to diagnose when the data can be evolved in. This criterion tells us that inward radial evolution breaks down precisely at the outer edge of the zone, \(r = 3 G M\). The data at large \(r\) can be evolved in all the way to the angular momentum barrier, but not into the zone. This is a strong motivation to consider more carefully an essential feature of the AMPS argument: evolving back the outgoing Hawking radiation in order to obtain information about the quantum fields near the horizon. One can try to get around this obstacle by Fourier transforming on the sphere. Then, at least in the free field theory approximation, it \(is\) possible to "evolve in\" mode by mode. However, we now encounter a different obstacle to stating the AMPS paradox: at least for a wide class of black holes, the behind-the-horizon partner of a mode with definite angular momentum does not fit into any single causal patch. Therefore, we can escape from the firewall paradox by the same means we escaped from Hawking's information paradox: black hole complementarity states that effective field theory can be trusted only within a single causal patch. (The geometry of the infalling patch will be described more fully in a forthcoming paper.) #### Mining. An objection to the idea that the zone can contain as few as \(\log S\) information-carrying degrees of freedom is the mining argument. I show that the same analysis described above naturally leads to a larger number of allowed states in the zone when mining equipment is present. The mining equipment naturally changes the infalling vacuum on scales set by the nature of the mining; this again leads to just the right number of allowed states to preserve unitarity. #### Relation to other work. It is difficult to accurately place this work in the context of the flurry of activity following the firewall paper. The firewall argument has been clarified, strengthened, and generalized by among others. This paper has some similarity to the "balanced holography\" proposal of the Verlindes. Part of the analysis, regarding what modes fit into a causal patch, is similar in spirit to the recent analysis of Ilgin and Yang. The work of Giddings and collaborators has some similarities, but here I try to preserve local, effective field theory in the causal patch. W. Kim and collaborators have also pointed out the negative energy experienced by certain freely falling observers near the horizon of an evaporating black hole. Mathur and Turton also propose that certain observers in the zone may be injured, while observers who fall in from infinity are not; their proposal is in the context of fuzz balls. Papadodimas and Raju give an elegant description of how small nonlocalities can conspire to allow for a resolution of the firewall paradox; here I try to avoid introducing any nonlocality. Maldacena and Susskind have made an inspired proposal for when and how violations of effective field theory occur: old black holes develop an Einstein-Rosen bridge connecting them to their Hawking radiation, modifying the usual geometry behind the horizon; these and related ideas are also devoloped in. A different picture of the entanglement emerges in the analysis of Shenker and Stanford. Harlow and Hayden have made a particularly interesting quantum computation argument against the firewall proposal. Many other proposals for escaping the paradox have been made, and the above is an unfair sampling. Here, I do not propose any fabulous new effects in quantum gravity; rather, I suggest an escape from the firewall argument on its own terms. I argue that the old, simple idea that any given observer can have it all-unitarity, the equivalence principle, and effective field theory-is not yet ruled out. # The Quantum State in the Zone ## Which states are consistent with the equivalence principle? We want to ask how unique the state of the quantum fields in the near-horizon zone is. How many degrees of freedom can be excited consistent with the naive expectations of an infalling observer? Roughly, the equivalence principle dictates that a freely falling observer should detect of order one quantum with wavelength of order \(R_s\) as he falls through the zone. For an eternal black hole, or a big, old black hole in AdS, the natural quantum state is the Hartle-Hawking state. Zooming in on the region near the horizon, the size of the 2-sphere is approximately constant, and the geometry looks like a 2-sphere cross 1+1 dimensional Rindler spacetime-that is, a 2-sphere cross 1+1 dimensional Minkowski space. The Hartle-Hawking state, when restricted to the zone, is roughly the Minkowski vacuum. This approximation becomes better and better as we zoom in closer to the horizon. The quantum state near an evaporating black hole is not uniquely determined by the equivalence principle. A cartoon version is as follows. Focus on quanta with frequency set by the temperature of the black hole, \(\omega \sim T \sim 1/{\ensuremath{R_s}}\). For these quanta, high angular momentum modes have an exponentially small chance of escaping to infinity, while s-waves have of order a 50% chance of escaping. If we focus just on the s-waves, we can say that each outgoing quantum has a 50% chance of escaping to infinity, and a 50% chance of falling back into the black hole.[^1] Which particles get reflected back in is a delicate quantum process. A naive infalling observer would just say that any given quantum may or may not be reflected. He thus identifies a family of reasonable quantum states for the zone. A simple way to think about it is to say that every Schwarzschild time, \(\Delta t \sim {\ensuremath{R_s}}\), 2 s-wave photons are emitted from the stretched horizon, one with each helicity. Depending on which photon escapes to infinity, we have a choice of the state for each of the ingoing photons, shown in figure [\[rindler\]](#rindler){reference-type="ref" reference="rindler"}. Another way to think about it is that tunneling through the barrier eliminates half of the ingoing photos that we would have had if the black hole were surrounded by a reflecting mirror. The choice of state in the zone is the choice of which ingoing quanta to erase.[^2] As mentioned in the introduction, there \(is\) a natural choice of the quantum state for an evaporating black hole. Consider the usual Schwarzschild modes, with definite frequency with respect to the Schwarzschild time. Such modes see a potential barrier set by the black hole mass and the angular momentum of the mode in question. The natural orthonormal basis consists of - 1) Modes incident from infinity, which are purely ingoing in the zone - 2) Modes incident from the horizon, which are purely outgoing near infinity The state of the modes of type 1 are determined by the requirement that there is no incoming flux from infinity, while the state of the modes of type 2 (which is actually a density matrix) is determined by the requirement that the future horizon is nonsingular. The statement in this section is simply that subtle deviations in the reflection coefficient with characteristic scale set by the Schwarzschild radius are consistent with the equivalence principle, and lead to the large number of allowed states described here. #### Alternative characterization of states consistent with the equivalence principle. Since it is confusing to identify states by which quanta are missing, consider another thought experiment that probes the uniqueness of the state of the quantum fields in the zone. Suppose we want a black hole in asymptotically flat space not to evaporate. We can compensate the evaporation by throwing in quanta that are roughly similar to the outgoing Hawking radiation. In order that the expectations of an infalling observer, based on a naive application of the equivalence principle, are unaffected by the quanta we throw in, let us throw them in from infinity, roughly one quantum per Schwarzschild time with wavelength set by the size of the black hole. As long as we obey these rough guidelines, we will not be adding any new high-energy features to the black hole geometry. But we can choose the detailed quantum state of the quanta we throw in. Since we are throwing in of order one quantum per light crossing time with wavelength of order the black hole radius, there is of order one bit of information per quantum. For example, one can send in photons and choose the helicity of each photon. Since the region near the horizon is well-described by Rindler space, it is convenient to translate the above description into a purely Rindler space description. In that case, we replace the asymptotic part of the black hole geometry by cutting off Rindler space at a distance \(R_s\) away from the horizon. Then the quanta thrown in from infinity become, in the Rindler description, quanta with wavelength \(R_s\) thrown in by a Rindler observer at a distance \(R_s\) away from the horizon. Figure [\[rindler\]](#rindler){reference-type="ref" reference="rindler"} can now be thought of as a picture of these real ingoing quanta (dotted null lines). We can summarize the picture in a simple way. If we zoom in on a region deep in the zone, far from the angular momentum barrier, and focus on the s-waves, then we expect a unique state for the outgoing modes, but we have a choice of states for the ingoing modes, all of which are consistent with the equivalence principle. There are two immediate questions. Do these information-carrying modes in the zone injure an infalling observer? And is there enough information in them to allow for unitary black hole evaporation? I address these questions in the following two subsections. ## Experience of an infalling observer It turns out that these excitations do not injure an observer who falls in from infinity, or anywhere outside the angular momentum barrier \(r = 3 G M\). However, they do injure a freely falling observer who somehow arranges to begin freely falling from a location very close to the horizon. For example, consider an observer who is lowered by a rope and "dangles\" at constant Schwarzschild \(r\) for some time. It clear that while he is still holding the rope, the observer will detect thermal radiation due to the Unruh effect. But even after letting go of the rope, this observer will encounter a large number of these ingoing quanta before hitting the singularity. This observer is in effect very boosted relative to the frame picked out by the geometry of the black hole. Roughly, quanta with wavelength of order \(R_s\) are emitted every \(R_s\) from a location \(R_s\) away from the horizon, but this observer sees them as high energy quanta because the "source\" is blue-shifted relative to him, as shown in figure [\[rindler\]](#rindler){reference-type="ref" reference="rindler"}. The above description can be made more quantitative by examining the stress-energy tensor in the zone. At any point well away from the singularity, the stress tensor is small in the frame of an observer who freely falls from infinity. The stress tensor near a black hole was described by Unruh and others. There is one new ingredient here relative to Unruh's work. In Unruh's description, the outgoing s-waves are in the Minkowski vacuum, and therefore thermal from the Rindler point of view, while the ingoing s-waves are completely absent-they are in the Rindler vacuum. The Rindler vacuum is again a unique state in the zone. In fact, as discussed above, some of the s-waves should reflect, so the ingoing s-waves should not be purely in the Rindler vacuum, but instead in a state that is a compromise between the Rindler and Minkowski vacua. We can be more quantitative. Computing the stress tensor in quantum field theory involves an ambiguity in regulating the infinities, but the difference in the stress tensor between two states is well-defined. If we go deep in the zone and treat only the s-wave, then the problem is \(1+1\) dimensional Minkowski space. Define coordinates so that \[ds^2 = dX^+ dX^-\] Then for the simplest case, a free massless field, the stress tensor in the Minkowski vacuum \({\ensuremath{|0_M>}}\) differs from the stress tensor in the Rindler vacuum \({\ensuremath{|0_R>}}\) by: \[\begin{aligned} {\ensuremath{<0_M|}} T_{++} {\ensuremath{|0_M>}}-{\ensuremath{<0_R|}} T_{++} {\ensuremath{|0_R>}} &\sim& {1 \over (X^+)^2} \\ {\ensuremath{<0_M|}} T_{--} {\ensuremath{|0_M>}}-{\ensuremath{<0_R|}} T_{--} {\ensuremath{|0_R>}} &\sim& {1 \over (X^-)^2 } \\ T_{+-} &=& 0 \end{aligned}\] (Since the problem is 1+1 dimensional, the stress tensor has dimensions of mass squared.) Now the black hole is like the Minkowski vacuum for the right movers, but it is missing half of the left movers, so deep in the zone we have \[\begin{aligned} {\ensuremath{<BH|}} T_{++} {\ensuremath{|BH>}}-{\ensuremath{<0_M|}} T_{++} {\ensuremath{|0_M>}} &\approx& 0 \\ {\ensuremath{<BH|}} T_{--} {\ensuremath{|BH>}}-{\ensuremath{<0_M|}} T_{--} {\ensuremath{|0_M>}} &\sim&-{1 \over 2 (X^-)^2 }\\ T_{+-} &=& 0 \end{aligned}\] A freely falling observer would naturally compare to the Minkowski vacuum, so the above equation is relevant for the experience of an infalling observer. Note the minus sign: a freely falling observer sees a \(negative\) energy density relative to Minkowski spacetime. Such negative energy densities are known to be possible in quantum field theory and have been constructed in simple examples. Isolated negative energy densities are not possible-a negative energy density must be compensated by a larger nearby positive energy density. (The field must "pay back\" for the negative energy density with a positive energy density, and it must pay interest on the negative energy it borrowed.) In this case, the compensating positive energy is behind the black hole horizon. So before worrying about delicate questions of unitarity, it appears that the natural description of the infalling observer is that the outgoing modes are in the vacuum, but the ingoing modes are in one of a number of allowed states which carry negative energy into the black hole. Roughly, instead of thinking of particle creation at the stretched horizon, the infalling observer thinks of "pairs\" of particles produced in the angular momentum barrier: an ingoing negative energy particle, and an outgoing positive energy particle. Since the negative energy particle is only possible due to compensating positive energy behind the horizon, the creation of the negative energy particle must be described as delocalized over the entire zone, but this is natural because it has wavelength of order \(R_s\) and is created about \(R_s\) from the horizon. When we come to the question of unitarity, it is these pairs of particles that must be carefully managed in order to preserve unitarity in black hole evaporation. Of course this is just a different description of the same physics seen by the external observer. The external observer naturally compares to the Rindler vacuum in the zone, and describes the same process in a more familiar way: the stretched horizon emits quanta, and some of them bounce off and come back into the black hole, while others escape to infinity. No quanta come in from infinity. ## Unitarity According to the description above, the states consistent with the equivalence principle consist, in the infalling description, of ingoing, negative energy excitations emitted from around the angular momentum barrier. Is the freedom in choosing the state of these excitations sufficient to allow for unitary evaporation? In fact it is: these excitations are precisely the thing that causes the black hole mass to decrease. Unitarity means that when old black holes lose mass they must also become less entangled with the environment. This can be arranged by carefully tuning the negative energy excitations. In addition, the positive energy quanta emitted to infinity must also carry information. Now a miracle is still necessary to get these modes into the correct state. I return later to the question of whether this miracle is possible within effective field theory. The above argument for unitarity suffices, but to get a better feel for the information flow it is natural to count the states. How many bits of information are there in the zone at one time? The answer depends somewhat on the choice of time slice; here, let's use the slices of constant Schwarzschild time natural for an outside observer. The evaporation process erases one photon per Schwarzschild time that was supposed to reflect back into the black hole. This erasing process happens at roughly the angular momentum barrier, \(r \approx 3 G M\). We would like to know how long this missing photon would have remained in the zone. The metric of a Schwarzschild black hole is \[ds^2 =-f(r) dt^2 + {dr^2 \over f(r)} dr^2 + r^2 d\Omega_2^2\] with \[f(r) = 1-{{\ensuremath{R_s}} \over r}~.\] Ingoing photons satisfy \[dt =-{dr \over f(r)}\] so the amount of Schwarzschild time \(t\) that elapses between the time the photon would have been reflected off the barrier until it would have reached the stretched horizon can be calculated to be \[t \sim {\ensuremath{R_s}} \log {{\ensuremath{R_s}} \over l_P} \sim {\ensuremath{R_s}} \log S\] where \(S\) is the black hole entropy. This version of the result generalizes to many types of black holes in a variety of dimensions. Since a single quantum takes of order \(\log S\) units of Schwarzschild time to pass through the zone, and the black hole emits one quantum per Schwarzschild time, the zone must contain at least \(\log S\) information-carrying quanta. Therefore, it is consistent with the equivalence principle to expect that \(\log S\) bits of information are needed to characterize the state in the zone. # Conflict with effective field theory? The above analysis shows that, at the level of counting bits of information, there are enough states in the zone so that an infalling observer does not experience a firewall, while unitarity is preserved. But an important question remains. If we focus on the s-waves in a region deep in the zone, the ingoing and outgoing waves do not interact. The discussion above involves encoding information in the ingoing waves, while the outgoing modes remain in a definite state. Having the correct state in the ingoing sector \(is\) sufficient for black hole evaporation to be unitary, but how does the ingoing sector end up in exactly the correct quantum state to ensure unitarity? Is there a violation of effective field theory outside the black hole? I believe this issue remains unresolved. The basic issue is the following. We know that the radiation propagating out to infinity must contain information. For an old black hole, these outgoing Hawking quanta must be entangled with the early radiation, rather than with the remaining black hole. The issue is whether we can somehow "evolve back\" the outgoing Hawking quanta in order to obtain information about modes while they are still deep in the zone. Let me first describe an approximation where I believe it \(is\) possible to evolve back the outgoing Hawking radiation until it is localized deep in the zone, leading to the firewall paradox. - Assume we can focus on the s-waves, and use the approximation that there is no mixing between ingoing and outgoing waves, and no mixing between modes of different angular momentum. In particular, this means that an outgoing s-wave near the horizon will certainly reach infinity. - Assume that the infalling observer can measure the s-wave behind the horizon-that is, the relevant Hawking partner of the outgoing radiation. Given these assumptions, we \(can\) evolve back, yielding the firewall paradox. (Actually, these assumptions are only sufficient for an "s-wave firewall,\" but as AMPS have argued, even an s-wave firewall violates the equivalence principle. In this section I focus on the arguments for such an s-wave firewall, which are the strongest of all the firewall arguments.) In this approximation, there is no mixing between ingoing and outgoing waves, so we can simply propagate the outgoing Hawking photons back along null rays until they are deep in the zone. ## Evolving back \(vs.\) evolving in However, in the real problem there is mixing between ingoing and outgoing modes: outgoing s-waves with frequencies of order the black hole temperature have an order one probability of reflecting off the geometry. (One might think that there is no angular momentum barrier for the s-wave, so it definitely makes it to infinity. However, in simply going from the \(3+1\) action to the \(1+1\) description, one finds a nontrivial radial potential even for the s-wave.) Therefore, literally evolving back the field operator corresponding to the outgoing mode yields an operator spread over the entire backward lightcone of the outgoing radiation, so it is spread over the entire zone and even extends beyond the zone. Physically, however, this seems like a technicality. We know that no radiation is coming in from infinity, so we should be able to ignore at least the part of the operator that is outside the zone. Instead of doing a standard Cauchy evolution backward in time, the natural way to use the outgoing Hawking radiation to learn about the zone is to do "radial evolution\" inwards. This type of radial evolution is familiar from AdS/CFT, but in the presence of a black hole there is a surprising obstacle. We can address this question in the classical limit. Suppose we know the exact field configuration far from the black hole for all time, and we would like to reconstruct the solution near the horizon, but still outside the horizon. This is simply the classical limit of evolving back the Hawking radiation until it is near the horizon. If we ask the question in flat space, then there would be no problem in doing this reconstruction of the interior. Suppose that we know the solution everywhere outside a sphere of radius R for all time. Can we reconstruct the solution inside the sphere? In flat spacetime, the answer is yes. More generally, however, radial evolution is a much more delicate problem than standard time evolution. This difference shows up already at the classical level: solving a known partial differential equation in a fixed background geometry. Mathematicians have addressed this type of problem, called unique continuation: given the solution to a known differential equation in some region \(r > r_*\) for all time, is there a unique continuation of the solution across the boundary \(r = r_*\)? There is a simple, intuitive diagnostic: if there are null geodesics that graze the surface \(r = r_*\) but do not enter the known region \(r > r_*\) where the solution is known, then the continuation is not unique. This was previously discussed in the AdS/CFT context in; see for related work. This result is intuitive to physicists: given a null geodesic, one can construct a solution that is localized arbitrarily close to the geodesic by going to a regime where the geometric optics approximation is good. If there are null geodesics that never enter the "known\" region \(r > r_*\), then one can construct solutions that are arbitrarily close to 0 in the known region \(r>r_*\), but are nonzero in the interior. In a sense, it may be the case that knowing the solution *exactly* in the region \(r>r_*\) does determine the solution in the interior, but knowing the exterior solution to an arbitrary finite precision does not determine the interior solution to \(any\) precision at all. This is usually summarized by saying that there is no *continuous mapping* between exterior data and interior data. It seems to me that continuity is the physically relevant criterion for whether reconstruction is possible. This result is robust, and holds for a wide variety of differential equations. The essential feature is just that the high-frequency behavior of the equation is the same as the wave equation. I return in the next subsection to possible ways around the reconstruction theorem. We can now analyze the behavior of null geodesics in the Schwarzschild solution. If we choose the affine parameterization, null geodesics are extrema of the action \[S = \int d\lambda g_{a b} \dot x^a \dot x^b\] where dot denotes the derivative with respect to the affine parameter \(\lambda\). For metrics of the Schwarzschild form, and choosing the particle to move only in the \(\theta\) direction on the \(S_2\), this becomes \[S = \int d\lambda \left( f(r) \dot t^2-{\dot r^2 \over f(r)}-r^2\dot \theta^2 \right)\] There are two conserved quantities, \(E = f(r) \dot t\) and \(\ell = r^2 \dot \theta\). Plugging these into the equation \(ds^2 = 0\) for null curves gives \[-f(r) \dot t^2 + {\dot r^2 \over f(r)} + r^2 \dot \theta^2 =-{E^2 \over f(r)} + {\dot r^2 \over f(r)} + {\ell^2 \over r^2} = 0\] In other words, the radial motion of the null geodesics can be read off from an effective potential, \[\dot r^2 + V(r) = E^2 \ \ \ \ \ \ {\rm with} \ V(r) = \ell^2 {f(r) \over r^2}\] For a Schwarzschild black hole, the function \(f(r)\) is \[f(r) = 1-{{\ensuremath{R_s}} \over r}\] so the effective potential for null geodesics is \[V(r) = \ell^2 \left( {1 \over r^2}-{r_s \over r^3} \right)\] This function approaches 0 at the horizon and at infinity, and has a maximum at \(r = 3 r_s/2\), or in other words at the location of the unstable circular null orbit, \(r = 3 G M\). Therefore, for any value of the angular momentum \(\ell\), geodesics with small enough energy \(E\) can bounce off the angular momentum barrier and never enter the known region \(r > 3 G M\). Therefore, according to the theorem, the solution at large \(r\) can be uniquely evolved radially inward until we reach the peak of the angular momentum barrier at \(r = 3GM\). So classically evolving Hawking photons inward fails for the near horizon region \(r < 3 G M\). (I return later to possible ways around the theorem, such as restricting to the s-wave.) The reason reconstruction fails is physically clear, and does not require mathematical theorems. Due to the angular momentum barrier, photons are constantly emitted by the stretched horizon, bounce off the angular momentum barrier, and fall back into the black hole. Such photons can interfere constructively or destructively with with photons that do manage to escape the barrier. The existence of photons that are bound inside the angular momentum barrier means that knowing the state of the outgoing Hawking quanta is not enough information to reconstruct the field near the horizon. ## Ways around the no-reconstruction theorem? Despite the theorem, intuitively one should be able to learn \(something\) about the solution inside \(r=3GM\). The various arguments for the firewall paradox including all rely on some technique of evolving back the outgoing Hawking radiation to obtain information about the degrees of freedom in the zone. For example, let us return to the approximation stated at the beginning of this section: simply Fourier transform on the sphere, and evolve radially inward mode by mode. Essentially, this approximation gets around the theorem by using the fact that in free field theory, modes with different angular momentum do not mix, so for any definite angular momentum we have a 1+1 dimensional problem. Suppose we focus on just one angular mode, say the s-wave. In a 1+1 problem we can reverse space and time at the price of changing positive mass squared to negative mass squared. However, even for fields with negative mass squared, the time evolution for a finite time is well-defined. Therefore, the radial evolution across \(r=3GM\) must be possible, mode by mode, in a free field theory. One obvious objection is that the fields we are dealing with must interact at least gravitationally, so in fact we cannot treat the angular momentum modes separately. Quantitatively, however, it may be possible to argue that these interactions are small enough that they can be neglected. If we work in terms of angular modes, we can perturb around the free field case, and it appears that no problems will arise at any order in perturbation theory. However, we do encounter problems with trying to reconstruct any operator with finite support in position space; see for a much more detailed discussion. #### Angular modes inside the horizon may not fit in one causal patch. My favorite objection pertains to the geometry of the causal patch. Suppose that we \(do\) restrict to the s-wave, and that we do succeed in evolving it back until it is close to the horizon. Now we appear to have a problem: on one hand, this near-horizon s-wave must be entangled with the early radiation, but on the other hand it should be entangled with its partner behind the horizon. But the partner of the s-wave is roughly the s-wave behind the horizon, or in any case it is some mode that is spread over the entire 2-sphere. But then we must ask whether this behind-the-horizon partner mode fits inside the causal patch of the infalling observer. There is only a conflict with black hole complementarity if there is a conflict with effective field theory within a single causal patch. The geometry of the causal patch behind the horizon will be described more completely in a separate publication. For Schwarzschild and Reissner-Nordstrom black holes in 3+1 and higher dimensions, the infalling patch does not contain the entire horizon sphere. Therefore, for these black holes, complementarity offers an escape from the firewall paradox. Note that this objection alone is sufficient to escape from the s-wave AMPS argument, without any need for the discussion in the remainder of the paper. Of course, one can build a mode that is *approximately* localized in a finite angular region on the sphere out of relatively low \(l\) modes, so it may seem like a mere technicality that we do not have access to the entire sphere. However, the fact that the reconstruction problem is impossible for any operator that is \(exactly\) localized in a finite region motivates careful consideration of this technicality. For Kerr black holes the question is more subtle because there is no spherical symmetry, and we are still investigating whether the relevant partner mode fits inside the infalling causal patch. Geometrically, there are some 2-spheres that \(do\) fit inside a causal patch, so it is possible that the firewall argument will work in this case, but further analysis is needed. # Mining One urgent question about the plausibility of this proposal has to do with all the other modes in the zone. Suppose we accept for now that for the s-waves there are enough states in the zone to allow for unitarity evaporation, as I have argued above. What now about the other modes in the zone? I am claiming many degrees of freedom in the zone remain entangled with the black hole, rather than the early radiation. This is dangerous because there is a technique for allowing additional modes to escape to infinity called black hole mining. The original mining procedure described by Unruh and Wald involved lowering boxes into the zone to capture the Hawking radiation. Recently, A. Brown showed that in fact any box that obeys the null energy condition is worse than superfluous, and the optimal mining procedure is simply to lower a string into the zone. The string cannot be lowered all the way to the stretched horizon without breaking, but it can be lowered far inside the angular momentum barrier. The tension of the string controls how far it can be lowered without breaking; see figure [\[mining\]](#mining){reference-type="ref" reference="mining"}. The string should couple to the field that we want to extract from the black hole. Then the quantum of interest can be emitted by the black hole, and instead of falling back in due to the angular momentum barrier, it can be absorbed by the string. Once it has been absorbed, it can then travel up the string-simply a 1+1 dimensional problem with no angular momentum barrier. Therefore lowering strings into the horizon allows us to extract high-l quanta from the black hole. The process of emitting a high-l quantum to infinity with the help of a string decreases the mass and entropy of the black hole. Therefore unitarity requires that the quanta absorbed by the string must be entangled with the early radiation and not the black hole horizon. In order to preserve unitarity, it must be the case that lowering a string into the zone \(changes\) the quantum state in such a way that the modes absorbed by the string carry information. In the presence of mining equipment, the zone must carry not only the \(\log S\) bits of information described above, but an additional number of bits that will be absorbed by the mining equipment. As Brown showed, a string with a given tension can be lowered only to a distance above the horizon related to its tension. If \(\mu\) is the tension of the string, then it can be lowered only to the location where the local Hawking temperature is related to the tension \(\mu = T^2\). Therefore, the proper distance \(x\) from the horizon is given by \[x^2 = {1 \over \mu}\] In order for the mining to be effective-for the mining equipment to extract more energy than it emits-the mining apparatus cannot be moved quickly into place, nor can it be quickly turned on and off. As a result, the important question is essentially a static one: how does the quantum state of the black hole differ in the presence of a string designed to mine the black hole? Turning again to the perspective of the infalling observer, by the same argument as above the string will effectively emit negative energy particles with a wavelength set by the location of the endpoint of the string into the black hole. From the outside point of view, the string is absorbing some of the quanta of the field, dominantly at the endpoint of the string. But for the string to be effective at mining, it must be "cold\" and not send quanta into the black hole from infinity. Just as above, the black hole state now differs from the Minkowski vacuum near the horizon due to the missing ingoing quanta that should have been reflected by the angular momentum barrier but instead were absorbed by the string. The black hole is also missing the s-waves that escaped, as before. These additional negative energy excitations cause the black hole to evaporate more quickly. For the same reason as above, the quantum state of these negative energy excitations must be correct in order for unitary evaporation. Infalling observers will notice some additional structure in the quantum state near the mining string. This additional structure may injure some observers, but it does not violate the equivalence principle. The discussion of whether there is a conflict with effective field theory also mirrors the s-wave construction. The same theorem prohibits evolving radially inward; one can again try to avoid the theorem by decomposing in angular modes, but these angular modes still do not fit in any causal patch for Schwarzschild or Reissner-Nordstrom black holes. To summarize, the presence of mining equipment in the zone (say a string) naturally changes the state seen by the infalling observer, with structure on the scale set by the location of the string. The new state still contains enough information-carrying modes in order to allow for unitarity without violating the equivalence principle. # Conclusions and future directions I have presented what I consider an attractive picture of which modes in the near-horizon zone carry information. In this description, even for an old black hole, most modes in the zone remain entangled with the stretched horizon rather than the early radiation. Note that the arguments here still leave a lot of freedom about precisely how many modes in the near-horizon zone are \(not\) entangled with the black hole. In order to allow for unitary evaporation, at least of order \(\log S_{\rm BH}\) modes must be entangled with the early radiation or in a pure state by themselves, and not entangled with the black hole. I have pointed out what I believe is a significant loophole in the arguments that the infalling observer must detect a violation of unitarity, effective field theory, or free infall. There are two main ingredients in this loophole. First, theorems from the theory of partial differential equations strongly suggest an obstacle to evolving the S-matrix data radially inward into the near-horizon zone. This obstacle can be avoided by working in terms of angular momentum modes on the 2-sphere, but this brings us to a second obstacle: at least for the simplest black holes, the angular momentum modes do not fit inside a single causal patch. Therefore the old idea of black hole complementarity, that no single observer will see a violation of effective field theory, has not been ruled out. This loophole essentially allows some number of outgoing modes in the zone to carry information, because their partners behind the horizon do not fit into any single causal patch. As stated, it applies to any mode that is delocalized over the entire sphere-far more than the \(\log S_{\rm BH}\) modes necessary for unitarity. So these arguments do not yet fix the number of information-carrying degrees of freedom in the zone. It is of course reasonable for effective field theory to be violated by a small amount even within one causal patch-the challenge is to identify which observables receive significant corrections. However, the simplest possibility remains that effective field theory is not corrected in any significant way for causal patches with low curvature. Fortunately, the issues I have raised can be addressed using known techniques. We can analyze the geometry of the infalling causal patch in combination with a field theory analysis of which modes should be entangled for a variety of black holes. Some results in this direction will appear in. In addition, there may be other ways of tightening the arguments in favor of the firewall paradox. For example, in spite of the motivating classical theorems, perhaps there is a clear argument that in quantum field theory the outgoing Hawking radiation can be evolved back to constrain the fields near the horizon. I have not addressed the AdS/CFT version of the firewall argument here. However, the obstacles to evolving radially inward discussed here appear in a very similar form in that context.

{'timestamp': '2014-07-22T02:13:56', 'yymm': '1401', 'arxiv_id': '1401.5340', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5340'}

# Preamble # Introduction ## Aim To introduce a model which innately and efficiently represents the salient properties of receptive fields of neurons in the V1 to MT motion processing stream. ### Specific Aim As one traces the neural pathway connecting the input layers of V1 to area MT, neurons become increasingly specialized for motion detection. In ascending order, attributes such as orientation selectivity, direction selectivity, speed tuning, increasing preferred speeds, component spatial frequency selectivity, and pattern (plaid) spatial frequency selectivity are sequentially acquired. Furthermore, the temporal frequency filtering properties of the neuronal population changes in a particular way along this pathway. Specifically, the proportion of bandpass temporal frequency filtering neurons to lowpass temporal frequency filtering neurons increases. Existing receptive field models do not represent this fundamental emergent property. Our specific aim is to introduce and describe a model for the receptive fields of V1 to MT neurons which innately and efficiently represents the aforementioned properties. ## Motivation and Background The detection of moving objects in our environment is both critical to our experience of the world and to our very survival. Imagine a hunter in rapid pursuit of a small agile animal. Both predator and prey detect abrupt changes in velocity occurring over an interval of less than a tenth of a second. The precision of our visual motion detection system is undoubtedly striking. The middle temporal (MT) area of the mammalian brain plays a key role in the visual processing of motion. The importance of area MT in motion processing is supported by a growing wealth of electrophysiological evidence which was initiated by Dubner and Zeki's 1971 report [@duze1971; @ze1974; @ze1980; @vama1981; @mava1983; @mava1983b; @feka1984; @unde1986; @mane1987; @roal1989; @mone1996; @smma2005]. For purposes of motion processing, the main input signals into MT originate in specialized cells of the primary visual cortex (V1). Kuffler's early studies of retinal ganglion cell response properties led to similar studies of simple cortical cells by Hubel and Wiesel. Together, their work generated great interest in neuronal receptive fields [@ku1953; @hu1957; @hu1959; @huwi1959; @ma1980; @da1985]. The functional forms of these receptive fields are at once beautifully simple yet enormously complex. Hence mathematical models have been used in step with electrophysiological studies to advance our understanding of their properties. Initially, focus was predominantly on their spatial structure [@ma1980; @ga1946]. Now, however, their temporal structure is increasingly studied in tandem. In particular, for most neurons in the V1 to MT processing stream, it is now appreciated that spatial and temporal features cannot be studied separately. They are spatiotemporally inseparable entities. In particular, motion is encoded by orientation in the spectral domain, and spatiotemporally-oriented filters are therefore motion detectors . Therefore at first glance it may seem an easy matter to mathematically model motion detecting neurons. The challenge, however, is to develop physiologically sound receptive field models which reflect the hierarchical structure of the motion processing stream. Various models do exist which are spatiotemporally-oriented filters, and are therefore motion detectors from a mathematical standpoint [@adbe1985; @qi1994; @qian1994; @qian1997; @qifr2009]. However, they fail to represent one of the most salient characterizing attributes of the motion processing stream: the lowpass to bandpass distribution of temporal frequency filtering properties along the V1 to MT specialization hierarchy. This feature is discussed further in the next subsection. In addition to the above temporal frequency filter property profile, the neurons of the V1 to MT motion processing stream have certain characteristics which set them apart and prescribe bounds on models of their spatiotemporal receptive field structure. Namely, they are tuned to specific orientations, directions, and speeds [@mone1996]; and they are more likely than other V1 cells to exhibit "end-stopping" phenomena [@scha2001]. Of the V1 cells, the subclass with direct synaptic projections from V1 to MT are the most specialized towards motion and are thought to be the main channels of motion substrates incident into MT neurons. For instance, Churchland et al showed that V1 and MT neurons lose direction selectivity for similar values of spatial disparity, suggesting V1 is MT's source of direction selectivity [@chpr2005]. Additionally, these V1 to MT neurons cluster motion substrate properties such as strong direction selectivity and tuning to high speeds. For example, Foster et al found that highly direction selective neurons in the macaque V1 and V2 were more likely to be tuned to higher temporal frequencies and lower spatial frequencies, i.e. higher speeds [@foga1985]. In other words, neurons that are highly direction selective are more likely to also be tuned to higher speeds. Such neurons are higher up in the V1 to MT hierarchy. Similarly, McLean and Palmer studied direction selectivity in cat striate cortex, and found that space-time inseparable cells were direction selective while space-time separable cells were not direction selective . Again reflecting the hierarchy of the motion processing stream. The distribution of some distinct anatomical and histological features have been shown to correlate with the motion-specialization hierarchy. For instance, Shipp and Zeki's retrograde tracer studies showed that the majority of direct V1 to MT projecting neurons originate within layer 4B [@shze1989]. Based on their results, they suggested that layer 4B contained a functionally subspecialized anatomically segregated group whose members each have direct synaptic connections with MT cells. The exact properties of the V1 to MT projectors are likely to be intermediate between those of cells in V1 layer 4C\(\alpha\) and MT neurons. This is because input into MT appears to be largely constituted of magnocellular predominant streams from 4C\(\alpha\), while parvocellular predominant streams may play a much less role  [@lihu1988; @mane1990; @yasa2001]. Progressive specialization along the V1 to MT motion processing stream is also observable in the morphological characteristics of the neurons. For instance, distinguishing attributes of the V1-MT or V2-MT projectors such as size, arborization patterns, terminal bouton morphology, and distribution have been observed  [@ro1989; @ro1995; @anbi1998; @anma2002]. Sincich and Horton demonstrated that layer 4B neurons projecting to area MT were generally larger than those projecting to layer V2. Overall, neurons in the V1 to MT motion processing stream are highly and progressively specialized. Hence description of their receptive fields requires adequately specialized models which reflect not only their individual attributes, but also the emergent hierarchical properties of the network. Next we discuss one of the most fundamental of such emergent network properties: the particular distribution of temporal frequency filtering types along the stream . ## Temporal Frequency Filtering Property Distribution Hawken et al found that direction-selective cells were mostly bandpass temporal frequency filters, while cells which were not direction-selective were equally distributed into bandpass and lowpass temporal frequency filtering types [@hash1996]. Foster et al found a similar phenomenon in macaque V1 and V2 neurons. V1 neurons were more likely to be lowpass temporal frequency filters, while their more specialized downstream heirs, V2 neurons, were more likely to be bandpass temporal frequency filters [@foga1985]. Less specialized neurons located anatomically upstream (lower down in the hierarchy) are more likely to have lowpass temporal frequency filter characteristics, while more specialized neurons located anatomically downstream (higher up in the hierarchy) are more likely to display bandpass temporal frequency filter characteristics. For example, LGN cells are at best only weakly tuned to direction and orientation ( , @femi2000) and are hence equally distributed into lowpass and bandpass categories. Layer 4B V1 cells and MT cells on the other hand, are almost all direction selective  [@mava1983b] and hence are mostly bandpass temporal frequency filters. We firmly believe this classification is not arbitrary, but instead is a direct manifestation of the particular spatiotemporal structure of the V1 to MT motion processing stream. Consequently, the receptive field model must reflect this increased tendency for bandpass-ness with ascension up the hierarchy. In other words, the representation scheme must be one that is inherently more likely to deliver bandpass-ness to a more specialized cell (such as in layer 4B or MT) and lowpass-ness to a less specialized cell (such as in layer 4A, 4C\(\alpha\), or 4C\(\beta\)). However, none of the existing receptive field models reflect this underlying spatiotemporal structure. In contrast, as will be seen in Section ([3](#sec:Gabor-Einstein Wavelet){reference-type="ref" reference="sec:Gabor-Einstein Wavelet"}) below, the Gabor-Einstein wavelet's wave carrier is a sinc function which directly confers the aforementioned salient property. Individual Gabor-Einstein basis elements are lowpass temporal frequency filters; and bandpass temporal frequency filters can only be obtained by combinations of basis elements. A study by DeAngelis and colleagues also corroborates the above. They found that temporally monophasic V1 cells in the cat were almost always low pass temporal frequency filters, while temporally biphasic or multiphasic V1 cells were almost always bandpass temporal frequency filters [@deoh1993b]. The explanation for this finding is inherent and explicit in the Gabor-Einstein wavelet basis, where bandpass temporal character necessarily results from biphasic or multiphasic combination. All monophasic elements on the other hand, are lowpass temporal frequency filters. However, we will see that according to the model, the converse is not true: i.e. not all lowpass temporal frequency filters are monophasic, and not all multiphasic combinations yield bandpass temporal frequency filters. ## Analogy to Special Relativity Perceptual and electrophysiological studies reveal intriguing similarities between how space and time are mixed in the visual cortex and how they are mixed in the special theory of relativity (STR). For instance, within certain limits, binocular neurons cannot distinguish a temporal delay from a spatial difference. As a result, inducing a monocular time delay by placing a neutral density filter over one eye but not the other, causes a pendulum swinging in a 2D planar space to be perceived as swinging in 3D depth space. This is known as the Pulfrich effect. Also, most neurons in the V1 to MT stream are maximally excited only by stimuli moving at that neuron's preferred speed. They are *speed tuned*. In this subsection, we briefly review and summarize the special theory of relativity. We identify the stroboscopic pulfrich effect and speed tuning as cortical analogues of STR's joint encoding of space and time. And finally, we explain how STR is used in (and motivates) the design of the Gabor-Einstein wavelet. ### Review of Special Relativity In 1905, Albert Einstein proposed the special theory of relativity . He was motivated by one thing: a firm belief that Maxwell's equations of electromagnetism are *laws of nature*. Maxwell's equations are a classical description of light . According to Maxwell, light is the propagation of electric and magnetic waves interweaving in a particular way perpendicular to each other. The speed of propagation in a vacuum is the constant \(c=299,792,458\) m/s. The other governing constraint on STR is the *relativity principle*. It has been around and generally accepted since the time of Galileo, and it states that the laws of nature are true and the same for all non-accelerating observers. Einstein believed Maxwell's equations to be laws of nature, and therefore to satisfy the relativity principle. In other words, Einstein believed the speed of light must be the same to all non-accelerating observers regardless of their velocity relative to each other. For the speed of light to be fixed, something(s) had to yield and be unfixed. What had to yield were the components of the definition of speed, i.e. space and time. Space and time had to become functions of relative velocity and of each other, to ensure the speed of light remains constant to all non-accelerating observers. The transformation rule---of space and time coordinates between two observers moving with constant velocity relative to each other---is called the Lorentz equations. It specifies how space and time are mixed in STR. More specifically, the Lorentz equations are the transformation rule between the (spacetime) coordinates of two inertial reference frames. For two inertial reference frames, \(R\) and \(R'\), moving exclusively in the \(x\) direction with constant velocity \(v\) relative to each other, the Lorentz transformation is given by, \[\begin{array}{l} \displaystyle t' = \gamma\left( t-\frac{v}{c^2} x\right),\\ \displaystyle x' = \gamma( x-vt),\\ \end{array} \label{Lorentz_transform}\] where (\(t',x'\)) are the coordinates of \(R'\), (\(t,x\)) are the coordinates of \(R\), \(c\) is the speed of light, and \(\gamma\) is the Lorentz factor and is given by, \[\gamma = \frac{1}{\sqrt{1-\frac{v^2}{c^2}}}.\] We demonstrate below that the stroboscopic pulfrich effect is analogous to STR in one respect, and speed tuning of neurons is analogous to STR in another respect. ### The Stroboscopic Pulfrich Effect In the classical pulfrich effect, a monocular temporal delay results in a perception of depth in pendular swing . It has a simple geometric explanation. By the time the temporally delayed eye "sees" the pendulum at a given retinal position, the pendulum has advanced to a further position in its trajectory, hence the visual cortex always receives images that are spatially offset between the eyes. The spatial disparity results in stereoscopic depth perception. This form of pulfrich phenomena is termed *classical* to distinguish it from the stroboscopic pulfrich effect. In the stroboscopic pulfrich effect, the pendulum is not a continuously moving bulb, but instead is a strobe light which samples the trajectory of the classical pendulum in time and space [@le1970; @mo1976; @buro1979; @mo1979]. The stroboscopic pulfrich effect cannot be explained by the simple geometric illustration that explains the classical pulfrich effect. This is because depth is still perceived in sequences involving a temporal delay in flash between eyes, but particularly lacking an interocular spatial disparity in the flash . In stereoscopic depth perception, a planar (x,y) difference in retinal image position (\(\Delta x\)) *transforms into* (is perceived as) a displacement along the \(z\) direction. In the stroboscopic pulfrich effect, a temporal difference (\(\Delta t\)) between eyes in retino-cortical transmission of image signal, results in a planar pendulum trajectory being perceived as (transforming into) depth. In STR, the transformation is between two inertial reference frames moving relative to each other. By analogy, we assert that for the stroboscopic pulfrich effect, there is a transformation between the interocular retinal space (\(x_r,t_r\)) and the perceptual depth space (\(x_z,t_z\)). We identify the variables (\(x_r,t_r\)) of the interocular space as the difference between eyes of the corresponding variables in the retino-cortical space i.e. \(x_r\) is the difference between the left and right eye retinal positions of an image; and \(t_r\) is the difference between the right and left eyes in retino-cortical image transmission times. The \((x_r,t_r)\rightarrow(x_z,t_z)\) transformation may take the form \(x_z=\alpha(x_r\pm \tilde{v}t_r)\), were \(\tilde{v}\) is proportional to pendulum velocity and \(\alpha\) is a function of \(\tilde{v}\). Note the similarity between \(x_z\) and the equation for \(x'\) in the Lorentz transformation, \(x'=\gamma(x-vt)\). This analogy suggests that similar mathematical structures underlie STR and the joint encoding of space and time in the visual cortex. The stroboscopic pulfrich effect is not by itself direct evidence of joint encoding of spatial and temporal disparities  [@recu2005; @recu2005b; @recu2005c], however direct electrophysiological evidence does exist for such linkage in cortical encoding [@capa1989; @anoh2001; @pabo2003]. This mathematical similarity influences our design of the Gabor-Einstein wavelet in a particular way which we describe in Section ([2.5](#Gabor_Einstein_Design){reference-type="ref" reference="Gabor_Einstein_Design"}) below. But first, we discuss another neuronal phenomenon, speed tuning, which is analogous to STR in a certain physical sense. ### Speed Tuning Most motion processing stream neurons in V1 or area MT are speed tuned. They have a preferred speed, which is the speed of moving stimuli to which they maximally respond. Other speeds also elicit a response, but with a lower rate of action potential. When presented moving sine wave grating stimuli, truly speed tuned neurons respond maximally to their preferred speed independent of the spatial or temporal frequency of the stimuli. The velocity, \(\begin{equation} \end{equation} where\[is temporal frequency and\)u\(is spatial frequency. In higher dimensions, \begin{equation} \omega = u \end{equation} where\)u\(is the spatial frequency in the\)x\(direction,\)v\(is the spatial frequency in the\)y\(direction,\) ## Design of the Gabor-Einstein Wavelet {#Gabor_Einstein_Design} We constrained the Gabor-Einstein model by requiring the following properties: 1. The minimum possible number of parameters 2. Relativistic-invariance of the wave carrier ### Minimal Number of Parameters In addition to orientation in the spatiotemporal frequency domain, localization in the space, time, spatial frequency, and temporal frequency domains are the essential features of the V1 to MT neuron receptive field. These are the most basic attributes of the receptive field. As such, they correspond with the minimum number of model parameters needed. Specifically, speed tuning, the spatial and the temporal localization envelopes, the amplitude modulation parameter, and the spatial and temporal frequencies must all be represented. If one adopts the notion of a gaussian spatial localizer , the minimum number of parameters add up to eight: four for the spatial envelope and four for the wave carrier. The four necessary spatial envelope parameters are: amplitude factor, \(A\); variance in the \(x\) direction, \(\sigma_x\); variance in the \(y\) direction, \(\sigma_y\); and spatial orientation of envelope, \(\theta_e\). The four necessary wave carrier parameters are: temporal frequency, \(\omega\); spatial frequency in the x direction, \(u\); spatial frequency in the \(y\) direction \(v\), and the phase, \(\phi\). The orientation of the wave carrier \(\theta_s=\tan^{-1}(v/u)\) is not an independent parameter. We use it in computer simulations below only for convenience. The essential feature that is not yet explicitly accounted for in the above parameter tally is the temporal envelope of the receptive field. This can be implemented either through the gaussian envelope or through the wave carrier, and in either case may conceivably involve an additional parameter. The next design criteria, *relativistic-invariance of the wave carrier*, resolves this ambiguity without adding another parameter. ### Relativistic-invariance of the Wave Carrier Given the analogies of spatiotemporal mixing in the visual cortex to spatiotemporal mixing in the special theory of relativity, it is reasonable to conjecture that their endpoints are within proximity of each other. Einstein eventually sought a lexicon in which physical laws are expressed the same way in any two inertial reference frames. That lexicon is called relativistic-invariance (or Lorentz-invariance). By analogy, we ultimately seek a lexicon in which physical laws can be expressed the same way in both the interocular and perceptual spaces. Lorentz-invariance of physical law is the specification (or endpoint) arising out of STR. Hence, requiring Lorentz-invariance of the receptive field's wave carrier is desirable. Moreover, this immediately resolves the above ambiguity regarding implementation of the temporal envelope profile. The temporal envelope must be implemented via the wave carrier for relativistic invariance to hold. And since the receptive field amplitude eventually falls, the sinc function emerges as a natural descriptor. Furthermore, the sinc function has the distinct advantage of not introducing any additional parameters. ## Summary The way space and time are mixed in the visual cortex bears resemblance to the way they are mixed in the special theory of relativity. We demonstrated that the stroboscopic pulfrich effect and speed tuning are cortical analogues of the spacetime mixing mechanics of STR. In STR, the mixture of space and time is described by the Lorentz transform which relates the spacetime coordinates of inertial reference frames moving with constant velocity relative to each other. In the striate and extrastriate motion processing stream, the analogous transformation is between the 3D interocular space (x,y planar inter-retinal space + inter-retinocortical time) and the 3D perceptual space (x,y,z 3D space). Though on the surface, STR and cortical spacetime encoding appear to be unrelated processes, we conjecture and have partly shown a shared underlying mathematical structure. This structure can be exploited to deepen our understanding of cortical encoding of motion, depth, and more fundamentally, the joint encoding of space and time in the visual cortex. Here, we proceeded by requiring relativistic-invariance of the receptive field's wave carrier. We simultaneously required that the model be constrained to the minimum possible number of parameters. Under these constraints, the sinc function with energy-momentum relation as argument emerged as a natural descriptor of the receptive field's wave carrier. Furthermore, the Gabor-Einstein wavelet explains a number of salient physiological attributes of the V1 to MT spectrum. Chief amongst these being the distribution of bandpass to lowpass temporal frequency filter distribution profile; which we postulate is a fundamental manifestation of the way space and time are mixed in the visual cortex. The remainder of this paper is organized as follows: Section ([3](#sec:Gabor-Einstein Wavelet){reference-type="ref" reference="sec:Gabor-Einstein Wavelet"}) introduces the Gabor-Einstein wavelet. Section ([4](#sec:simulations){reference-type="ref" reference="sec:simulations"}) presents computer simulations. Section ([5](#sec:discussion){reference-type="ref" reference="sec:discussion"}) is the discussion. It includes the following subsections: Subsection ([5.1](#subsec:model_framework){reference-type="ref" reference="subsec:model_framework"}) introduces a simple framework for classifying and naming the components and hierarchical levels of receptive field models. Subsection ([5.1](#subsec:model_framework){reference-type="ref" reference="subsec:model_framework"}) also identifies the Gabor-Einstein's place within this larger framework of receptive field modeling. Subsection ([5.2](#subsec:related_work){reference-type="ref" reference="subsec:related_work"}) briefly discusses related work. Subsection ([5.3](#subsec:higher_order){reference-type="ref" reference="subsec:higher_order"}) discusses some phenomena which cannot be explained by models such as the native Gabor-Einstein wavelet in isolation, i.e. models of single neuron receptive fields early in the visual pathway. Specifically, it discusses some higher order phenomena and non-linearities which necessarily arise from neuronal population and network interactions. Section ([6](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}) concludes the paper. # The Gabor-Einstein Wavelet {#sec:Gabor-Einstein Wavelet} In this section, we present the Gabor-Einstein wavelet. It has the following properties: - Like the Gabor function, it is a product of a gaussian envelope and a sinusoidal wave carrier. - It differs from the Gabor function in that its wave carrier is a relativistically-invariant sinc function whose argument is the energy-momentum relation. - Its gaussian envelope contains only spatial arguments. i.e. its spatial envelope is not time-dependent. - It has the minimum possible number of parameters. - Its fourier transform is the product of a mixed frequency gaussian and a temporal frequency step function. - Like the Gabor function, it generates a quasi-orthogonal basis. We define the Gabor-Einstein wavelet as follows, \]\label{Lorentz_Gabor_l} G(t,x,y) = A~\exp\left(-\frac{(x-x_0)^2}{2\sigma^2_x}-\frac{(y-y_0)^2}{2\sigma^2_y}\right)\mbox{sinc}(\omega_0 t-u_0 x-v_0 y+\phi),\[ where the constant multiplier \(A\) is amplitude; \(\sigma_x\) and \(\sigma_y\) are the gaussian variances in the x and y directions; \(x_0\) and \(y_0\) are the respective \(x\) and \(y\) coordinates of the gaussian center; \(\omega_0\), \(u_0\), and \(v_0\) are the frequencies of the wave carrier in the \(t\), \(x\), and \(y\) directions respectively; \(\phi\) is the sinusoid phase; and the *sinc* function is defined as, \]\mbox{sinc}(x) = \frac{\sin(x)}{x}.\[ We have rotated our coordinates by an angle \(\theta\) from a reference state \((x',y')\) to a state \((x,y)_\theta\) which we denote \((x,y)\) for notational simplicity. The transformation, \](x',y')~~~\longrightarrow ~~~(x,y)_\theta :=(x,y), \label{Eqn:envelope_axes_rotate_1}\[ is given by, \]\vectwo{x}{y} = \left( \begin{array}{cc} \cos (\theta) &-\sin (\theta)\\ \sin (\theta) & ~~\cos (\theta)\\ \end{array} \right)\vectwo{x'}{y'}. \label{Eqn:envelope_axes_rotate_2}\[ We proceed here with the following instance of the Gabor-Einstein wavelet, \]\label{Gabor-Einstein} G(t,x,y) = \exp\left(-\frac{x^2}{2\sigma^2_x}-\frac{y^2}{2\sigma^2_y}\right)\mbox{sinc}(t-u_0 x-v_0 y),\[ where we have set \(\omega_0\) equal to one. We can always do so for one reference neuron by simply defining the unit of time as \(1/\omega_0\), the duration the neuron's frame cycle. As we will see, this value, \(1/\omega_0\), is the shortest frame duration to which the neuron can respond. In the above equation, we have also set \(A=1\), and \(\theta=x_0=y_0=\phi=0\). The fourier transform is as follows, \]\label{Gabor-Einstein-Fourier} H(\omega,u,v)=\frac{\sigma_x\sigma_yA}{2}\sqrt{\frac{\pi}{2}}\exp\left(-\frac{\sigma_x^2(u+u_0\omega)^2}{2}-\frac{\sigma_y^2(v+v_0\omega)^2}{2}\right)L(\omega),\[ where \(L(\omega)\) is given by, \]\label{eqn:sign} L(\omega) = \mbox{sign}(1-\omega) + \mbox{sign}(1+\omega)\[ and the *sign* function is defined as, \]\mbox{sign}(x) = \begin{cases} ~~1 & \mbox{if}~x>0,\\ ~~0 & \mbox{if}~x=0, ~\mbox{and}\\-1 & \mbox{if}~ x<0. \end{cases}\[ The above fourier transform was obtained using Mathematica symbolic software. The fourier transform of the general case will likely be challenging to obtain analytically either by hand or symbolic software. Hence, we anticipate numerical methods may have an important role to play. The maximum magnitude of the above response function, Equation ([\[Gabor-Einstein-Fourier\]](#Gabor-Einstein-Fourier){reference-type="ref" reference="Gabor-Einstein-Fourier"}), is attained where the argument of the exponent is zero, i.e. where, \]\label{F_max} \sigma_x^2(u+u_0\omega)^2+\sigma_y^2(v+v_0\omega)^2=0.\[ The neuron's preferred spatial frequency, \(f_0\), is dependent on the temporal frequency, \(\omega\), of the stimulus, and is given by the above bivariate quadratic equation's solution, \]\label{Eqn:f_0} f_0(\omega)=(-u_o\omega,-v_0\omega).\[ The location where the response maximum is attained is a parametrized curve in 3D frequency space. It is given by, \]R(\omega)=(\omega,-u_0\omega,-v_0\omega).\[ Although the location in \((u,v)\) space where the neuron's maximum response is attained depends on \(\omega\), the magnitude of the maximum response is itself a constant. It is given by, \]R_{max} = |H(\omega,f_0(\omega))|= \frac{\sigma_x\sigma_yA}{2}\sqrt{\frac{\pi}{2}}.\[ The half magnitude response is attained at values of \((u,v)\) satisfying, \]\exp\left(-\frac{\sigma_x^2(u+u_0\omega)^2}{2}-\frac{\sigma_y^2(v+v_0\omega)^2}{2}\right) = \frac{1}{2}.\[ Taking the natural logarithm of the above equation yields, \]\label{half_max} \frac{\sigma_x^2(u+u_0\omega)^2}{2~\mbox{ln}(2)}+\frac{\sigma_y^2(v+v_0\omega)^2}{2~\mbox{ln}(2)} = 1.\[ Defining \(\sigma_x^{-1}\sqrt{2~\mbox{ln}(2)}\rightarrow a\) and \(\sigma_y^{-1}\sqrt{2~\mbox{ln}(2)} \rightarrow b\), Equation ([\[half_max\]](#half_max){reference-type="ref" reference="half_max"}) becomes, \]\label{half_max2} \frac{(u+u_0\omega)^2}{a^2}+\frac{(v+v_0\omega)^2}{b^2} = 1.\[ In the above form, one readily recognizes this as the ellipse centered at \((-u_0\omega,-v_0\omega)\), whose principal axes radii are \(a\) and \(b\) in the \(x\) and \(y\) directions respectively. The long axis in the frequency domain is the short axis in the spatial domain and vice versa. Without loss of generality, we can assume that prior to rotation of the spatial axes by angle \(\theta\), the long axis of the receptive field envelope is parallel to the x axis. Then the orientation of the on-off bands, i.e. planes of the spatial wave, are aligned parallel to the long axis of the envelope when the rotation angle, \(\theta\), is related to the polar angle \(\mu\) by the relationship, \]\label{parallel} \theta= \mu.\[ On the other hand, the on-off bands are perpendicular to the long axis of the envelope when, \]\label{perpendicular} \mu=\theta+\frac{\pi}{2}.\(\) In the case of Equation ([\[parallel\]](#parallel){reference-type="ref" reference="parallel"}), the half magnitude frequency bandwidth is readily seen to equal \(2a\), the length of the short axis, while in the case of Equation ([\[perpendicular\]](#perpendicular){reference-type="ref" reference="perpendicular"}), the half magnitude frequency bandwidth equals \(2b\), the length of the short axis. The cases of skewed alignment take on values between \(2a\) and \(2b\) and are also computable from the geometry. Unlike the preferred spatial frequency, the half-magnitude frequency bandwidth is not dependent on the temporal frequency of the stimulus. In summary, the Gabor-Einstein wavelet spatiotemporal receptive field model predicts that the magnitude of a V1 to MT neuron's preferred spatial frequency is linearly dependent on the temporal frequency of the stimulus as shown in Equation ([\[Eqn:f_0\]](#Eqn:f_0){reference-type="ref" reference="Eqn:f_0"}). In the next section, we present some receptive field simulations using the Gabor-Einstein wavelet. # Computer Simulations {#sec:simulations} In this section, we present computer simulations demonstrating the Gabor-Einstein wavelet's properties. We start with simulations which show the model's adherence to basic physiological form. We then present simulations which demonstrate how the Gabor-Einstein model innately represents the temporal frequency filtering property distribution along the V1 to MT neuronal hierarchy. The following notations are used in the figure captions: \(\theta_e\) is the angle of rotation of the axes of the gaussian envelope as described in Equations ([\[Eqn:envelope_axes_rotate_1\]](#Eqn:envelope_axes_rotate_1){reference-type="ref" reference="Eqn:envelope_axes_rotate_1"}) and ([\[Eqn:envelope_axes_rotate_2\]](#Eqn:envelope_axes_rotate_2){reference-type="ref" reference="Eqn:envelope_axes_rotate_2"}). \(\theta_s\) is the angle of rotation of the axes of the wave carrier. ctr is the 2-component vector consisting of the \(x\) and \(y\) coordinates of the gaussian envelope center respectively. \(\sigma=(\sigma_x,\sigma_y)\) is the 2-component vector consisting of the gaussian variances of the envelope in the \(x\) and \(y\) directions respectively. \(\phi\) is phase of the wave carrier. \(t\) is time. And \(\hat{\omega}=(\omega_0,u_0,v_0)\) is the 3-component vector consisting of the temporal, x-spatial, and y-spatial frequencies of the wave carrier respectively. For succinctness, \(\hat{\omega}=1\), for instance, is taken to be equivalent to \(\hat{\omega}=(1,1,1)\). The same short-hand notation applies to the other multi-component vectors. ## Basic Physiological Form Figure ([\[fig:gaussians\]](#fig:gaussians){reference-type="ref" reference="fig:gaussians"}) shows receptive field envelopes with varying aspect ratios and centers; Figure ([\[fig:wavecarriers\]](#fig:wavecarriers){reference-type="ref" reference="fig:wavecarriers"}) shows the sinc function wave carriers at varying times \(t\) and at one different orientation; Figures ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}) and ([\[fig:gabor-einstein-center-3d\]](#fig:gabor-einstein-center-3d){reference-type="ref" reference="fig:gabor-einstein-center-3d"}) show Gabor-Einstein wavelets with varying envelope centers; Figure ([\[fig:gabor-einstein-spatial-freq\]](#fig:gabor-einstein-spatial-freq){reference-type="ref" reference="fig:gabor-einstein-spatial-freq"}) shows Gabor-Einstein wavelets with varying spatial frequencies; and Figure ([\[fig:gabor-einstein-spatial-phase\]](#fig:gabor-einstein-spatial-phase){reference-type="ref" reference="fig:gabor-einstein-spatial-phase"}) shows Gabor-Einstein wavelets of varying phase. In Figure ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}), one sees the anisometry that is conferred by the sinc function wave carrier. The anisometry increases from Fig ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}a) through ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}d), due to the rapid fall off in the sinc function amplitude outside of the central peak over the two half cycles on each side of zero. In contrast, Fig ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}e) and Fig ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}f) appear similar to each other in form, and each look more isometric than Fig ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}d). This illustrates the presence of both isometric and anisometric basis elements which differ only in a single parameter (phase, or time, or envelope center). This expressiveness in the basis should allow for economical representation of a visual scene. Of note, the amplitudes of the more isometric elements Figs ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}e) and ([\[fig:gabor-einstein-center\]](#fig:gabor-einstein-center){reference-type="ref" reference="fig:gabor-einstein-center"}f) are much smaller than those of the more anisometric preceding elements. If desired, this can be readily modulated by the constant amplitude factor parameter. Figure ([\[fig:gabor-einstein-center-3d\]](#fig:gabor-einstein-center-3d){reference-type="ref" reference="fig:gabor-einstein-center-3d"}) shows 3-dimensional plots illustrating the above discussed isometry modulation feature of the Gabor-Einstein basis. Figure ([\[fig:gabor-einstein-spatial-freq\]](#fig:gabor-einstein-spatial-freq){reference-type="ref" reference="fig:gabor-einstein-spatial-freq"}) shows receptive fields with various spatial frequency tuning. Figure ([\[fig:gabor-einstein-spatial-phase\]](#fig:gabor-einstein-spatial-phase){reference-type="ref" reference="fig:gabor-einstein-spatial-phase"}) shows various receptive fields which are phase-shifted relative to each other. For any given spatial location in the receptive field, the peak amplitude of a cycle is phase dependent. This is a feature distinct to the Gabor-Einstein model. And like the anisometry-modulation feature, it confers expressiveness to the basis family. This should allow for efficient representation of natural visual scenes. In the model, time has a similar effect on peak amplitude. And this behavior is in agreement with space-time inseparable receptive field profiles in cat striate cortex, where a time dependent phase drift has been shown [@deoh1993]. ## Temporal Frequency Filter Property Distribution {#subsec:temp_fil_sim} Figure ([\[fig:sinc and fourier\]](#fig:sinc and fourier){reference-type="ref" reference="fig:sinc and fourier"}) shows the essential property which the Gabor-Einstein wavelet inherits from the sinc function. The sinc function's fourier transform is a lowpass filter. Higher frequency sinc functions have wider bandwidth. Bandpass filters are formed by taking the difference between sinc functions of different frequency as illustrated in Figure ([\[fig:sinc and fourier\]](#fig:sinc and fourier){reference-type="ref" reference="fig:sinc and fourier"}). The Gabor-Einstein wavelet's fourier transform has a step function factor inherited from the sinc function. This allows it explain the particular distribution of lowpass to bandpass temporal frequency filter properties of V1 to MT neurons (Foster et al 1985; DeAngelis et al 1993b; Hawken et al 1996) in a manner innately representative of the motion-processing stream's neuronal hierarchy. For simplicity of illustration, Figure ([\[fig:GE_Fourier\]](#fig:GE_Fourier){reference-type="ref" reference="fig:GE_Fourier"}) shows a 2D (x,t) Gabor-Einstein element. Its fourier transform is shown in Figs ([\[fig:GE_Fourier\]](#fig:GE_Fourier){reference-type="ref" reference="fig:GE_Fourier"}a) and ([\[fig:GE_Fourier\]](#fig:GE_Fourier){reference-type="ref" reference="fig:GE_Fourier"}b), while its (x,t) domain representation is shown in Fig ([\[fig:GE_Fourier\]](#fig:GE_Fourier){reference-type="ref" reference="fig:GE_Fourier"}c). The lowpass temporal frequency nature of the fourier representation is apparent, as the support straddles the zero. Figure ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}) shows the construction of temporal frequency bandpass Gabor-Einstein wavelets. As illustrated, the temporal frequency bandpass Gabor-Einstein element results from the difference of two temporal frequency lowpass Gabor-Einstein elements. Fig ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}a) and Fig ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}b) are plots of Equation ([\[Gabor-Einstein-Fourier\]](#Gabor-Einstein-Fourier){reference-type="ref" reference="Gabor-Einstein-Fourier"}), i.e. they are the fourier transform of the Gabor-Einstein wavelet described by Equation ([\[Gabor-Einstein\]](#Gabor-Einstein){reference-type="ref" reference="Gabor-Einstein"}). We label this basis element "\(\omega_b=1\)", meaning it is the zero-centered lowpass temporal frequency filter whose temporal frequency bandwidth is one. Accordingly, Fig ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}c) and Fig ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}d) are plots of the "\(\omega_b=3\)" Gabor-Einstein basis element, i.e. they are plots of the zero-centered lowpass temporal frequency filter whose temporal frequency bandwidth is three. Fig ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}e) and Fig ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}f) plot the bandpass temporal frequency filter basis element obtained by taking the difference of the "\(\omega_b=1\)" and "\(\omega_b=3\)" basis elements. The left-hand column of Figure ([\[fig:GE_striatotemporal spectrum\]](#fig:GE_striatotemporal spectrum){reference-type="ref" reference="fig:GE_striatotemporal spectrum"}) plots the Gabor-Einstein basis elements of Fig ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}), while the right hand column plots the corresponding temporal response probed at the spatial origin of the receptive field. There is an increasing complexity of the temporal waveform structure as one progresses from lowpass to bandpass basis element. All pure Gabor-Einstein basis elements are lowpass temporal frequency filters. On the other hand, the bandpass temporal frequency filter property necessarily results from summation of Gabor-Einstein basis elements. Hence bandpass temporal elements are necessarily complex (not pure). However, not all lowpass temporal frequency filters are pure; and not all basis summations yield bandpass temporal frequency filters. For instance, Figure ([\[fig:GE_sum\]](#fig:GE_sum){reference-type="ref" reference="fig:GE_sum"}) shows the sum of two pure Gabor-Einstein basis elements which yield another lowpass temporal frequency filter. The temporal waveform of the composite element is indeed more complex than that of its two pure constituents, however, it appears less complex than the composite temporal waveform of Fig ([\[fig:GE_band\]](#fig:GE_band){reference-type="ref" reference="fig:GE_band"}). # Discussion {#sec:discussion} The neurons in the V1 to MT dorsal stream are highly specialized towards motion detection. Furthermore, they are embedded in a hierarchical order of increasing motion specialization. The earliest (least specialized) members of the chain are the non-speed tuned V1 cells, while the most specialized members are the MT cells. Naturally, LGN cells precede input layer V1 cells both anatomically as well as in level of motion specialization. Similarly, cells located in the medial superior temporal (MST) area and the parietal lobe are further along than MT cells both anatomically and in motion specialization. We have focused our current study on the V1 to MT spectrum. We set out to obtain a good mathematical representation for the receptive fields of these neurons. The term *good* here implies that the representation should faithfully capture the salient features of the motion specialization spectrum and serve as an effective label or address schema for neurons in the motion processing stream. For instance, it is known that MT neurons receive input from V1 neurons and most directly from V1-to-MT projectors, therefore in a good basis, the representation for an MT neuron should result from some combination of representations of some V1 neuron(s). Furthermore, the resulting MT representation should naturally manifest known attributes of MT neurons. Appropriate comparison and discussion of the current work and related work requires a more precise language than currently available. Hence, we begin this discussion by introducing a simple framework for classifying and naming the components and hierarchical levels of receptive field models. ## Receptive Field Model Classification {#subsec:model_framework} In the literature, there is currently no clear categorization of the various mathematical descriptions of receptive fields. Ironically, the laudable increase in discovery of receptive field attributes and modulation mechanisms may have led to further conceptual entanglement of the mathematical forms of receptive fields. Here, we propose a simple framework to organize and classify receptive field descriptions and briefly review some basic concepts. This should help place the current work, the Gabor-Einstein model, in the appropriate context; and furthermore, should help organize and direct receptive field research. We will identify a receptive field by its constituents and method of construction. A receptive field can be constructed using two essential things: 1. Primary input receptive fields 2. Input combination scheme The primary input receptive field is the response function of the neuron at the lowest level on the hierarchy of study. For instance, when studying the motion-processing stream between V1 and MST, the primary input receptive field is the response function of a V1 input layer neuron. An input combination scheme is a model for the network of neurons projecting onto the study neuron. For instance, a common scheme is linear summation of weighted V1 responses to yield an MT neuron. The input combination scheme can be classified according to the stage at which input processing, if any, occurs. Three classes arise: 1. Pre-processing 2. Intra-processing 3. Post-processing Pre-processing is modification of the input responses prior to combining them. For instance, V1 elements are squared in the motion-energy model, after which they are summed to yield the MT response . Intra-processing is modification of the input responses during the combination process. For example, the Soft MAX scheme [@lafe2004; @ripo1999; @nose1995]. Post-processing is modification of the response after input combination. For instance, divisive normalization schemes [@he1992; @he1993; @sihe1998; @tshu2010]. The divisive normalization scheme, proposed by Heeger in 1992, is a neuron population based model of V1 to MT function. In that model, each neuron's behavior is modulated by that of a population module of which it was a member. Specifically, its manifest response is its native response divided by the net sum of the responses of the other neurons in its population module plus a regularization factor. The model bears resemblance to receptor models in pharmacokinetics, where the regularization factor plays the role of a half-saturation constant. The third categorization attribute is the hierarchical level from which input originates relative to the study neuron. The three following classes arise: 1. Feed-forward 2. Isostratal 3. Feedback Feed-forward implies the input comes in from a lower hierarchical level. For instance V1 projecting to MT. Isostratal implies that the input comes in from the same level. For instance, normalization of MT response by neighborhood neuron pool [@he1992; @he1993; @sihe1998; @tshu2010]. Feedback implies the input comes in from a higher level in the hierarchy. Of note, most receptive fields likely receive contributions from all three types of inputs. In light of the above framework, the Gabor-Einstein model is a primary input receptive field for the V1 to MT motion processing stream. It is compatible with any of the aforementioned input combination schemes, hierarchical input routing schemes, and input processing schemes. The work presented in this paper focused mainly on the primary input receptive field, and for simplicity of illustration, assumed a basic (i.e. linear) input combination scheme. Certain important phenomena cannot be explained at the level of early (e.g. input layer V1) primary input receptive fields alone. Next, we discuss some aspects of such phenomena including end-stopping, a short selection of non-linearity mechanisms, and contrast-modulated speed tuning. ## Related Work {#subsec:related_work} A key distinguishing feature of the Gabor-Einstein model is how it naturally represents the distribution of lowpass to bandpass temporal frequency filtering properties along the V1 to MT motion specialization hierarchy. Ratio models are a class of models with a different objective. The ratio models aim to mechanistically model the way speed tuned units arise from combinations of non-speed tuned bandpass and low pass temporal frequency filters. These models are based on the idea that stimulus temporal frequency is proportional to the ratio of outputs of spatiotemporally separable bandpass to lowpass neuronal outputs [@tosh1973; @adbe1986; @peth2002; @pe2004; @pe2005]. Various summation-based input combination schemes along the V1 to MT stream have been used. Heeger and colleagues proposed a two stage weighted linear summation model in which MT fields resulted from weighted sums of V1 fields . Sereno as well as Nowlan & Sejnowski proposed multilevel neural network reinforcement learning models in which MT neuron response was a function of a linear summation of V1 responses . Tsui et al proposed a model in which MT response was obtained by a soft-max weighted summation of V1 inputs [@tshu2010]. Each of these schemes differs fundamentally in approach from the Gabor-Einstein model. Namely, we focus on the receptive field of a single neuron along the V1 to MT pathway, implicitly representing the network spatiotemporal structure in the inherent attributes of the basis itself. Specifically, the natural representation of temporal frequency filter distribution along the V1 to MT spectrum. In the Gabor-Einstein basis, the basis itself mandates summation as a means of ascending up the specialization hierarchy. This is a gratifying result, and supports the notion that the sinc function is a natural way to describe the V1 to MT spectrum. ## Higher-order Phenomena {#subsec:higher_order} Here, we discuss certain higher order phenomena whose description exceeds the scope of single early neuron receptive fields. ### End-stopping End-stopping is a center-surround type phenomenon in which---for instance---a full length bar stimulus results in a submaximal response, while some shorter bar stimulus generates maximal response. In other words, the excitatory portion of the receptive field is shorter than the full length of the receptive field. Typically, the periphery of the receptive field, i.e. its *end*, elicits suppression (or "stopping"). Hence neurons exhibiting such behavior are said to be "end-stopped". Jones and colleagues found that majority (94%) of V1 cells in the macaque were end-stopped to various degrees . Moreover, the most prominently end-stopped V1 cells have been found in layer 4B which is known to be the dominant projection layer to MT . Furthermore, Tsui et al have argued a role for V1 end-stopping in MT motion integration and solution to the aperture problem [@tshu2010]. It seems therefore that a complete model of V1 to MT neurons must account for end-stopping. End-stopping phenomena would be difficult to explain at the single input layer V1 neuron stage, since it likely involves isostratal interactions. Tsui and colleagues modeled end-stopping by specifying the normalization pool as surround cells arranged along the orientation axis of a center cell. With this spatially-specific implementation of normalization, surround cell activity yielded end-stopping . The response was then fit to a saturation kinetics model equivalent to Heeger's divisive normalization model . Regarding center-surround phenomena such as end-stopping, the Gabor-Einstein wavelet can be considered a *classical receptive field*. A scheme similar to that of Tsui et al can be used to implement end stopping using the Gabor-Einstein wavelet as both center and surround, or in hybrid with other wavelet transform as either center or surround. Of note, other schemes have also been used to implement end-stopping. For instance, Skottun proposed an orientation-modulated cancellation strategy . The Gabor-Einstein wavelet is compatible with this scheme as well. ### Nonlinearity Mechanisms The perception of motion is likely a higher order phenomenon that emerges out of the elaborate nonlinear connection network of which MT is only a part. Some of these non-V1 areas from which MT receives input include LGN  [@sipa2004; @naly2006], superior colliculus [@rogr1990; @bewu2010; @bewu2011], and extrastriate regions [@mava1983; @boun1990; @unde1986]. Some combination of these non-V1 inputs may explain the persistence of visual responsiveness and direction selectivity after V1 lesions or cooling  [@gisa1992; @azpa2003; @roal1989; @rogr1990]. Indeed several researchers have drawn specific attention to the necessity of non-linear models. Moreover, these nonlinearities are not just a feature of late (specialized) visual neurons, but have been demonstrated very early in the visual pathway. For example, Schwartz et al pointed out the need for nonlinear retinal ganglion cell receptive field models . And Rosenberg et al showed that LGN Y cells utilize a nonlinear mechanism to represent interference patterns [@rohu2010]. Fine mechanistic modeling of the single V1 to MT neuron's receptive field's most salient features is certain to yield valuable physiological insight into the myriad non-V1 connections to MT. ### Contrast-Modulated Effects Contrast-modulated speed tuning is likely unexplainable at the level of early primary input receptive fields of single neurons. Existing explanations of such phenomena appear to necessarily involve nonlinear isostratal interactions. For instance, divisive normalization [@he1992; @he1993; @sihe1998]. In general, the effects of contrast on the receptive field of V1 to MT neurons can safely be assumed to be highly modulated by neighboring neurons. Significant data exists to support such population encoding. For instance, DeAngelis et al showed that end and side stopping were strongest along the preferred excitatory orientation, yet superimposed suppressive stimuli was much more broadly tuned than the classical receptive field, suggesting a modulatory pool of neurons . Such neuron pools also likely mediate the effects of contrast on speed tuning. Krekelberg et al found that most MT cells in alert macaques preferred lower speeds for lower stimulus contrast [@krva2006]. Priebe et al found that a significant population of V1 cells have contrast-modulated speed tuning curves [@prli2006]. Livingstone and Conway found that the speed tuning curves of most V1 neurons in alert macaques were dependent on contrast stimuli. Specifically, for lower contrast stimuli, V1 cells shifted their tuning curves to lower speeds; while for higher contrast stimuli, the opposite was observed [@lico2007]. Psychophysical studies also show that both high contrast stimuli and high spatial frequency stimuli are associated with higher perceived speeds [@brmo2011]. The increase in preferred speed with contrast is a phenomenon that likely involves nonlinear isostratal interactions. As such, early single neuron receptive field models in isolation are unable to explain such phenomena. Nonetheless the primary input receptive field is the theoretical and physiological foundation of higher order phenomena. Hence it is essential to have a physiologically sound model such as the Gabor-Einstein wavelet which naturally generates the hierarchical specialization structure of the motion stream. The Gabor-Einstein wavelet can be readily plugged-in to divisive normalization schemes which may be necessary to explain contrast-modulated phenomena. # Conclusion {#sec:conclusion} In this paper, we introduced the Gabor-Einstein Wavelet, a new family of functions for modeling the receptive fields of neurons in the V1 to MT motion processing stream. We showed that the way space and time are mixed in the visual cortex has analogies to the way they are mixed in the special theory of relativity. We therefore constrained the Gabor-Einstein model to a relativistically invariant wave carrier and to the minimum possible number of parameters. These constraints yielded a sinc function wave carrier with energy-momentum relation as argument. The model innately and efficiently represents the temporal frequency filtering property distribution along the motion processing stream. Specifically, on the V1 end of the stream, the neuron population has an equal proportion of lowpass to bandpass temporal frequency filters; whereas on the MT end, they have mostly bandpass temporal frequency filters. From our analysis and simulations, we showed that the distribution of temporal frequency filtering properties along the motion processing stream is a direct effect of the way the brain jointly encodes space and time. We uncovered this fundamental link by demonstrating an analogous mathematical structure between the special theory of relativity and the joint encoding of space and time in the visual cortex. The Gabor-Einstein model and the experiments it motivates will provide new physiological insights into how the brain represents visual information.

{'timestamp': '2014-01-23T02:06:09', 'yymm': '1401', 'arxiv_id': '1401.5589', 'language': 'en', 'url': 'https://arxiv.org/abs/1401.5589'}