1: %% ver. 1 for nlin arxive, K. 20 December 2001
2: %% ver 2 for nlin arXive, K. 18 June 2002
3: %% ver 3.0 for nlin arxive K 28 August 2002
4:
5: %%%%%
6: \documentstyle[amsmath,amssymb,aps,epsf,eqsecnum,graphicx]{revtex}
7: %%%%%
8: \begin{document}
9: \title{Random unistochastic matrices}
10: \author
11: {Karol {\.Z}yczkowski$^{1,2}$, Marek Ku{\'s}$^1$, Wojciech
12: S{\l}omczy{\'n}ski$^3$ and Hans--J{\"u}rgen Sommers$^4$}
13: \address{$^1$Centrum Fizyki Teoretycznej, Polska
14: Akademia Nauk, \\ Al. Lotnik{\'o}w 32/44, 02-668 Warszawa, Poland}
15: \address{$^2$ Instytut Fizyki im. M. Smoluchowskiego,
16: Uniwersytet Jagiello{\'n}ski, ul. Reymonta 4, 30-059 Krak{\'o}w, Poland}
17: \address{$^3$ Instytut Matematyki, Uniwersytet Jagiello{\'n}ski,
18: ul. Reymonta 4, 30-059 Krak{\'o}w, Poland}
19: \address{$^4$ Fachbereich 7 Physik, Universit{\"a}t Essen,
20: 45117 Essen, Germany}
21: %\date{\today}
22: \date{ August 28, 2002}
23: \maketitle
24:
25: \begin{abstract}
26: An ensemble of random unistochastic (orthostochastic) matrices is defined by
27: taking squared moduli of elements of random unitary (orthogonal) matrices
28: distributed according to the Haar measure on $U(N)$ (or $O(N)$,
29: respectively).
30: An ensemble of symmetric unistochastic matrices is obtained
31: with use of unitary symmetric matrices pertaining to the circular orthogonal
32: ensemble. We study the distribution of complex eigenvalues of
33: bistochastic, unistochastic and orthostochastic
34: matrices in the complex plane. We compute averages
35: (entropy, traces) over the ensembles of unistochastic matrices
36: and present inequalities concerning the entropies
37: of products of bistochastic matrices.
38: \end{abstract}
39:
40: \medskip
41:
42: \pacs{PACS numbers: 05.45.Mt, 02.50.Ga, 02.10.Ud}
43: %%\footnote{Key words ?}
44: %%\footnote{MR ?}
45:
46: \section{Introduction}
47:
48: Consider a square matrix $B$ of size $N$ containing non-negative entries.
49: It is called {\sl stochastic} if each row sums to unity ($\sum_{i=1}^N
50: B_{ij}=1$ for $j=1,\dots,N$) (for information about properties of such
51: matrices consult \cite{CZ98,FW79}). If, additionally, each of its columns sums
52: to unity, i.e., $\ \sum_{i=1}^N B_{ji}=1$ for $i=1,\dots,N$, it is called
53: {\sl bistochastic} (or {\sl doubly stochastic}).
54:
55: Bistochastic matrices emerge in several physical problems. They are used in the
56: theory of majorization \cite{MO79,AU82,An89,Bh97},
57: angular momentum \cite{Lo97}, in transfer problems,
58: investigations of the Frobenius-Perron operator,
59: and in characterization of completely positive maps acting in the space
60: of density matrices \cite{St02}. We shall denote by
61: $\Omega_N^S$ (resp. $\Omega_N^B$) the sets of stochastic (resp. bistochastic)
62: matrices of size $N$.
63:
64: A matrix $B$ is called {\sl orthostochastic}, if there exists an orthogonal
65: matrix $O$, such that $B_{ij}=O_{ij}^2$ for $i,j=1,\dots,N$. Analogously, a
66: matrix $B$ is called {\sl unistochastic}
67: ({\sl unitary-stochastic})\footnote{This notation is not unique: in some
68: mathematical papers (e.g. \cite{PT87}) unistochastic matrices are called
69: orthostochastic}, if there exists a unitary matrix $U$, such that
70: $B_{ij}=|U_{ij}|^2$ for $i,j=1,\dots,N$. Due to unitarity (orthogonality)
71: condition every unistochastic (orthostochastic) matrix is bistochastic. These
72: four sets of nonnegative matrices are related by the following inclusion:
73: $\Omega_N^O \subseteq \Omega_N^U \subseteq \Omega_N^B \subset \Omega_N^S$,
74: where $\Omega_N^U$ and $\Omega_N^O$ represent the sets of unistochastic
75: (orthostochastic) matrices. For $N>2$ all three inclusions are proper
76: \cite{MO79}.
77:
78: Unistochastic matrices appear in analysis of models describing the time
79: evolution in quantum graphs \cite{KS97,Ta00,PZK01,Ta01,PZ01} and in
80: description of non-unitary transformations of density matrices
81: \cite{Ni99b,Ni00}. Moreover, the theory of majorization and unistochastic
82: matrices plays a crucial role in recent research on local manipulations
83: with pure states entanglement \cite{Ni99} or in
84: characterizing the interaction costs of non--local
85: quantum gates \cite{HVC02}.
86:
87: In this work we analyse the structure of the set of bistochastic (unistochastic)
88: matrices of a fixed size and investigate the support of their spectra.
89: Knowledge of any constraints for the localisation of the eigenvalues of such a
90: matrix is of a direct physical importance, since the moduli of the largest
91: eigenvalues determine the rate of relaxation to the invariant state of the
92: corresponding system. We define the notion of entropy of
93: bistochastic matrices and
94: prove certain inequalities comparing the initial and
95: the final entropy of any probability vector subjected to a Markov chain
96: described by an arbitrary bistochastic matrix. A related inequality
97: concerns the entropy of the product of two bistochastic matrices.
98:
99: Moreover, we define physically motivated ensembles of random unistochastic
100: matrices and analyse their properties. As usual, the term {\sl ensemble}
101: denotes a pair: a space and a probability measure defined on it (for example
102: the {\sl circular unitary ensemble} (CUE) represents the group $U(N)$ of
103: unitary matrices of size $N$ with the Haar measure \cite{Me91}). Since any
104: unitary matrix determines a unistochastic matrix, the Haar measure on the
105: unitary group $U(N)$ induces uniquely the measure in the space of
106: unistochastic matrices $\Omega_N^{U}$, which we shall denote by $\mu_U$.
107: Analogously, we shall put $\mu_O$ for the measure in $\Omega_N^{O}$ induced by
108: the Haar measure on the orthogonal group $O(N)$. In the sequel we shall use
109: the names {\sl unistochastic ensemble} (resp. {\sl orthostochastic ensemble})
110: for the pair $\{\Omega_N^{U}, \mu_U \}$ (resp. $\{\Omega_N^{O}, \mu_O \}$).
111:
112: We compute certain averages with respect to these ensembles.
113: Related results were presented recently by Berkolaiko \cite{Be01} and
114: Tanner \cite{Ta01}. In the latter paper the author defines unitary stochastic
115: ensembles, which have a different meaning: they consist of unitary matrices
116: corresponding to a given unistochastic matrix.
117:
118: This paper is organised as follows. In section~II we review properties of
119: stochastic matrices. In particular we analyse the support of spectra of
120: random bistochastic and stochastic matrices in the unit circle. In section~III
121: some results concerning majorization, ordering, and entropies of bistochastic
122: matrices are presented. In particular we prove subadditivity of entropy for
123: bistochastic matrices. Section~IV is devoted to the ensembles of
124: orthostochastic and unistochastic matrices; we investigate the support of
125: their spectra, compute the entropy averages, the average traces, and
126: expectation values of the moduli of subleading eigenvalues. Some open
127: problems are listed in section~V.
128: Analysis of certain families of unistochastic matrices
129: and calculation of the averages with respect to the
130: unistochastic ensemble is relegated to the appendices.
131:
132:
133: \section{Stochastic and bistochastic matrices}
134:
135: \subsection{General properties}
136:
137: In this section we provide a short review of properties of stochastic and
138: bistochastic matrices. The set of bistochastic matrices of size $N$ can be
139: viewed as a convex polyhedron in $R^{N^2}$. There exist $N!$ permutation
140: matrices of size $N$, obtained by interchanging the rows (or columns) of the
141: identity matrix. Due to the Birkhoff theorem, any bistochastic matrix can be
142: represented as a linear combination of permutation matrices. In other words
143: the set of bistochastic matrices is the convex hull of the set of permutation
144: matrices. By the Caratheodory theorem it is possible to use only $N^2-1$
145: permutation matrices to obtain a given bistochastic matrix as their convex
146: combination \cite{MO79}. Farahat and Mirsky showed that in this combination it
147: is sufficient to use $N^2-2N+2$ permutation matrices only, but this number
148: cannot be reduced any further \cite{FM60}. The dimension of the set of
149: bistochastic matrices is $(N-1)^2$. The volume of the polyhedron of bistochastic
150: matrices was computed by Chan and Robbins \cite{CR99}.
151:
152: Due to the Frobenius--Perron theorem any stochastic matrix has at least one
153: eigenvalue equal to one, and all others located at or inside the unit circle.
154: The eigenvector corresponding to the eigenvalue $1$ has all its components real
155: and non--negative.
156: For bistochastic matrices the corresponding eigenvector consists of $N$
157: components equal to $1/N$ and is called uniform. A stochastic matrix $S$ is
158: called {\sl reducible} if it is block diagonal, or if there exists a
159: permutation $P$ which brings it into a block structure,
160: \begin{equation}
161: S'=PSP^{-1} =
162: \left[
163: \begin{array}{cc}
164: A_1 & 0 \\
165: C & A_2 \\
166: \end{array}
167: \right] ,
168: \label{reduci}
169: \end{equation}
170: where $A_i$ are square matrices of size $N_i<N$ for $i=1,2$, $N = N_1+N_2$.
171: It is called {\sl decomposable} if one can find two permutation matrices $P$
172: and $Q$ such that $PSQ$ has the above form. Matrix $S$ is {\sl irreducible}
173: ({\sl indecomposable}) if no such matrix $P$ (matrices $P$ and $Q$) exists
174: (exist) \cite{MO79,Me89}. An irreducible stochastic matrix cannot have two
175: linearly independent vectors with all components nonnegative. Any reducible
176: bistochastic matrix is {\sl completely reducible} \cite{Me89}, i.e., the
177: matrix $C$ in (\ref{reduci}) is equal to zero.
178:
179: A stochastic matrix $S$ is called {\sl primitive} if there exists only one
180: eigenvalue with modulus equal to one. If $S$ is primitive, then $S^k$ is
181: irreducible for all $k\ge 1$ \cite{MO79}. Note that the permutation matrices
182: $P$ with $\operatorname*{tr}P=0$ are irreducible, but non-primitive, since
183: $P^N$ equal to identity is reducible.
184: For any primitive stochastic there exist a natural number $k$
185: such that the power $S^k$ has all entries positive.
186: The fact that all
187: the moduli of eigenvalues but one are smaller than unity
188: implies the convergence $\lim_{k\to\infty} B^k=B_*$.
189: Here $B_*$ denotes the uniform bistochastic matrix ({\sl van
190: der Waerden matrix}) with all elements equal to $1/N$. Its spectrum consists
191: of one eigenvalue equal to one and $N-1$ others equal to zero. The matrix
192: $B_*$ saturates the well known van der Waerden inequality \cite{MO79}
193: concerning the permanent of the bistochastic matrices: ${\rm per}B \ge
194: N!/N^N$, and hence is sometimes called the {\sl minimal} bistochastic matrix
195: \cite{Eg81,Me89}.
196:
197: Each bistochastic matrix of size $N$ may represent a transfer process at an
198: oriented graph consisting of $N$ nodes. If a graph is disjoint or consists of
199: a Hamilton cycle (which represents a permutation of all $N$ elements),
200: the bistochastic matrix is not primitive, and the modulus of the subleading
201: eigenvalue (the second largest) is equal to unity.
202:
203:
204: If matrices $A$ and $B$ are bistochastic, its product $C=AB$ is also
205: bistochastic. However, the set of bistochastic matrices does not form a
206: group, since in general the inverse matrix $A^{-1}$ is not bistochastic (if it
207: exists). For any permutation matrix $P$ its inverse $P^{-1}=P^T$ is
208: bistochastic and the eigenvalues of $P$ and $P^{-1}$ are equal and belong to
209: the unit circle.
210:
211:
212: \subsection{Spectra of stochastic matrices}
213:
214: A stochastic matrix contains only non-negative entries and due to the
215: Frobenius--Perron theorem its largest eigenvalue is real. This leading
216: eigenvalue is equal to unity, since its spectral radius is bounded by
217: the largest and the smallest sum of its rows, all of which are equal to $1$.
218: In the simplest case of permutation matrices the spectrum consists of some
219: roots of unity. The eigenvalues of permutation matrices consisting of only one
220: cycle of length $k$ are exactly the $k$-th roots of unity.
221:
222: Upper bounds for the size of the other eigenvalues are given in \cite{Sc76}.
223: Let $M$ denote the largest element of a stochastic matrix and $m$ the smallest.
224: Then the radius $r_2$ of a subleading eigenvalue satisfies
225: \begin{equation}
226: r_2 \le \frac{M - m}{M+m} \quad .
227: \label{boundmm}
228: \end{equation}
229:
230: From this bound it follows that all subleading eigenvalues of the
231: van der Waerden uniform matrix $B_*$ vanish. Another simple bound of this
232: kind for a matrix of size $N$ reads \begin{equation}
233: r_2 \le {\rm min}\{ NM-1, 1-Nm \} \quad .
234: \label{boundm2}
235: \end{equation}
236:
237: For any stochastic matrix the characteristic polynomial is real, so
238: we may expect a clustering of the eigenvalues of random stochastic matrices
239: at the real line. This issue is related to the result of Kac, who showed
240: that the number of real roots of a polynomial of order $N$ with random real
241: coefficients scales asymptotically like $\ln N$ \cite{Ka59}.
242: The spectrum of a stochastic matrix is
243: symmetric with respect to the real axis. Thus for $N=2$ all eigenvalues are
244: real and the support of the spectrum of the set of stochastic matrices reduces
245: to the interval $[-1,1]$.
246:
247: %\vskip -2.0cm
248: \begin{figure} [htbp]
249: %\vskip -1.0cm
250: \begin{center}
251: \
252: \includegraphics[width=10.0cm,angle=0]{ZKSSfig1.ps}
253: \vskip 0.5cm
254: \caption{
255: Eigenvalues of $3000$ random stochastic matrices of size
256: a) N=3, b) N=4. Panels
257: c) and d) show the spectra of bistochastic matrices of size $3$ and $4$.
258: Thick solid lines represent the bounds (\ref{arcs}) of Karpelevich.}
259: \label{fig1}
260: \end{center}
261: \end{figure}
262: %\vskip -1.0cm
263:
264: \medskip
265:
266: Let $Z_k$ denote a regular polygon (with its interior) centred at $0$ with one of its $k$
267: corners at $1$. The corners are the roots of unity of order $k$. Let ${\bar
268: E}_N$ represents the convex hull of $E_N=\bigcup_{k=2}^N Z_k$, or, in other
269: words, the polygon constructed of all $k$-th roots of unity, with
270: $k=2,\dots,N$. It is not difficult to show that the support $\Sigma_n^S$ of
271: the spectrum of a stochastic matrix of order $N$ is contained in ${\bar E}_n$
272: - see e.g. a concise proof of Schaefer \cite{Sc76}, p.15 (originally
273: formulated for bistochastic matrices). The $k$-th roots of the unity -- the
274: corners of the regular $k$--polygon, represent eigenvalues of non-trivial
275: permutation matrices of size $k$. For $N=3$ this polygon becomes a deltoid
276: (dotted lines in Fig.~1b), while for $N=4$ a non-regular hexagon.
277: However, this set
278: is larger than required: as shown by Dimitriew and Dynkin \cite{DD46} for
279: small $N$ and later generalised and improved by Karpelevich \cite{Ka51} for an
280: arbitrary matrix size, the support $\Sigma_N^S$ of the spectrum of
281: stochastic matrices forms, in general, a set which is not convex. For a
282: recent simplified proof of these statements consult papers by Djokovi{\v c}
283: \cite{Dj90} and Ito \cite{It97}. For instance, the support $\Sigma_3^S$
284: consists of a horizontal interval and the equilateral triangle (see Fig.~1a),
285: while for $\Sigma_4^S$ four sides of the hexagon should be replaced by the
286: arcs, interpolating between the roots of the unity, and given by the solutions
287: $\lambda =\left| \lambda \left( t\right) \right| e^{i\phi \left( t\right) }$
288: with $t\in \left[ 0,1\right]$ of
289:
290: \begin{eqnarray}
291: \lambda^4+(t-1)\lambda -t=0
292: \quad \text {for which} \quad |\phi(t)|\in [\pi/2,2\pi/3], \nonumber
293: \end{eqnarray}
294: and
295: \begin{eqnarray}
296: \lambda^4-2t\lambda^2+(2t-t^2-1)\lambda(t-1)+t^2=0 \quad
297: \text {for which} \quad |\phi(t)|\in [2\pi/3,\pi].
298: \label{arcs}
299: \end{eqnarray}
300:
301: \subsection{Spectra of bistochastic matrices}
302:
303: The {\sl spectral gap} of a stochastic matrix is defined as $1-r_2$, where $r_2$ denotes
304: the modulus of the subleading eigenvalue \cite{Be01} (note that in
305: \cite{Ta01} the gap is defined as $-\ln r_2$). This quantity is relevant for
306: several applications since it determines the speed of the relaxation to the
307: equilibrium of the dynamical system for which $B$ is the transition matrix.
308: Analysing the spectrum of a bistochastic matrix it is also interesting to
309: study the distance of the closest eigenvalue to unity. Since for any
310: bistochastic matrix $1$ is its simple eigenvalue if and only if the matrix is
311: irreducible, some information on the spectrum may be obtained by introducing a
312: measure of irreducibility. Such a strategy was pursued by Fiedler \cite{Fi95},
313: who defined for any bistochastic matrix $B$ the following quantities
314:
315: \begin{equation}
316: \mu(B):= {\rm min}_A \sum_{i\in A} \sum_{j\in {\bar A}} B_{ij}
317: \quad \quad {\rm and} \quad \quad
318: \nu(B):= {\rm min}_A \sum_{i\in A} \sum_{j\in {\bar A}} \frac{B_{ij}}{n(N-n)},
319: \label{munu}
320: \end{equation}
321: where $A$ is a proper subset of indices $I=\{1,2,\ldots,N \}$ containing $n$ elements,
322: $1\le n < N$, and $\bar A$ denotes its complement such that $A\cup {\bar
323: A}=I$. Observe that $\mu$ measures the minimal total weight of the
324: `non-diagonal' block of the matrix, which equals to zero for reducible
325: matrices, while $\nu$ is averaged over the number of $n(N-n)$ elements, which
326: form such a block. Fiedler \cite{Fi95} used these quantities to establish the
327: bounds for the subleading eigenvalue ${\lambda}_2$
328: \begin{equation} |1-\lambda_2|
329: \ge 2 \mu(B) (1-\cos\frac{\pi}{N}) \quad \quad {\rm and} \quad \quad
330: |1-\lambda_2| \ge 2\nu (B).
331: \label{munures}
332: \end{equation}
333:
334: Since $\Omega_N^B \subseteq \Omega_N^S$, the supports of the spectra
335: fulfil $\Sigma_N^B \subseteq \Sigma_N^S$. Moreover, our numerical analysis
336: suggests that for $N\ge 4$ this inclusion is proper. In fact, they do not
337: contradict an appealing conjecture \cite{FW79}, that the support $\Sigma_N^B$
338: of the spectra of bistochastic matrices is equal to the set theoretical sum
339: $E_N$ of regular $k-$polygons $Z_k$ ($k=2,\ldots,N$), whose points are the
340: consecutive roots of the unity. Numerical results obtained for random matrices
341: chosen according to the uniform measure in the $(N-1)^2-$dimensional space of
342: bistochastic matrices are shown in Fig.~1c~and~1d.
343:
344: It is not difficult to show that these polygons are indeed contained in the
345: set $\Sigma_N^B$, i.e., for each $\lambda \in \bigcup_{k=2}^N Z_k$ there
346: exists an $N \times N$ bistochastic matrix $B$ such that $\lambda$ belongs to
347: its spectrum, $\operatorname*{Sp}(B)$. First note that the support $\Sigma_{N-1}^B$ is
348: included in $\Sigma_N^B$. Hence, it is enough to show that $Z_N \subset
349: \Sigma_N^B$. Let us start from the following simple observation: if $A,B \in
350: \Omega_N^B$ commute, then the corresponding eigenspaces are equal, and if
351: $\lambda \in \operatorname*{Sp}(A), \mu \in \operatorname*{Sp}(B)$ are the
352: corresponding eigenvalues, then $x\lambda+(1-x)\mu \in
353: \operatorname*{Sp}(xA+(1-x)B)$ for each $x \in [0,1]$.
354:
355: In the sequel $P_{(i^1_1 \dots i^1_{k_1}) \dots (i^m_1 \dots i^m_{k_m})}$ will denote
356: the matrix, which corresponds to the permutation consisting of $m$ cycles:
357: $(i^1_1 \to \dots \to i^1_{k_1}), \dots, (i^m_1 \to \dots \to i^m_{k_m})$, where
358: $k_1 + \dots + k_m = N$. Let $w_N$ be a vector whose coordinates are the
359: consecutive $N$th-roots of the unity, i.e., $w_N=\left( 1,\exp (2\pi
360: i/N\right) ,\ldots ,\exp \left( 2\pi i\left( N-1\right) /N\right)$. Let
361: $k=0,\ldots ,N-1$. Then matrices $P_{\left( 1\ldots N\right) }^{N-k}$ and
362: $P_{\left( 1\ldots N\right) }^{N-\left( k+1\right) }$ commutes, and $P_{\left(
363: 1\ldots N\right) }^{N-k}w_{N}=\exp (2\pi ki/N) w_{N}$, $P_{\left( 1\ldots
364: N\right) }^{N-\left( k+1\right) }w_{N}=\exp (2\pi \left( k+1\right) i/N)
365: w_{N}$. Hence the edges joining the complex numbers: $\exp (2\pi ki/N)$ and
366: $\exp (2\pi \left( k+1\right) i/N)$ for $k=0,\ldots ,N-1$ are contained in
367: $\Sigma_N^B$, and so the whole boundary of the polygon $Z_N$. Furthermore,
368: taking linear combinations of a bistochastic matrix with an eigenvalue at the
369: edge of the polygon $Z_N$ and the identity matrix ${\mathbb I}_N$ we may
370: generate lines in $\Sigma_N^B$ joining this boundary with point $1$. It
371: follows therefore, that the entire inner part of each polygon constitutes a
372: part of the support of the spectrum of the set of bistochastic matrices.
373:
374: Let us now analyse in details the case $N=3$. $P_{(123)}$ denotes the $3 \times 3$
375: matrix representing the permutation $1\to 2 \to 3 \to 1$. Its third power is
376: equal to identity, $P_{(123)}^3=P_{(1)(2)(3)}={\mathbb I}_3$, while
377: $P_{(123)}^2=P_{(123)}^{-1}=P_{(132)}$. Two edges of the equilateral triangle
378: joined in unity are generated by the spectra of $xP_{(123)}+(1-x){\mathbb
379: I}_3$ - see Fig.~2b. The third vertical edge is obtained from spectra of
380: another interpolating family $xP_{(123)}+(1-x)P_{(132)}$ - see Fig.~2a. The
381: boundary of $\Sigma_3^B$ is obtained from spectra of the members of the convex
382: polyhedron of the bistochastic matrices of size $3$. On the other hand, the
383: permutations $P_{(123)}$ and $P_{(12)(3)}$ do not commute, and the spectra of
384: their linear combination form a curve inside $\Sigma_3^B$ - see Fig.~2c. To
385: show that there are no points in the set $\Sigma_N^B$ outside $E_3$ note that
386: $Z_2 \cup Z_3 = E_3 \subseteq \Sigma_3^B \subseteq \Sigma_3^S = E_3$.
387:
388: Let us move to the case $N=4$. Two pairs of sides of the square forming $Z_4$
389: constructed of commuting permutation matrices are shown in Fig.~2d and
390: Fig.~2e, while Fig.~2f presents the spectra of a combination of noncommuting
391: permutation matrices, which interpolate between $3$ and $4$--permutations.
392: This illustrates the fact that $E_4$ is contained in the set $\Sigma_4^B$, but
393: we have not succeeded in proving that both sets are equal.
394:
395: Analysis of the support of the spectra of stochastic and bistochastic matrices
396: can be thus summarised by
397: \begin{equation}
398: \bigcup_{k=2}^N Z_k = E_N \subseteq \Sigma_N^B \subseteq \Sigma_N^S
399: ={\tilde E}_N \subset {\bar E}_N,
400: \label{sigmabisto}
401: \end{equation}
402: where ${\tilde E}_N$ is a concave hull of the set--theoretical sum $E_N$
403: of the regular polygons $Z_k$ supplemented by the area bounded by the Karpelevich's
404: interpolation curves \cite{Ka51,Dj90,It97}, whereas ${\bar E}_N$ is the
405: closed convex hull of $E_N$.
406:
407: %\vskip -1.0cm
408: \begin{figure} [htbp]
409: %\vskip -1.0cm
410: \begin{center}
411: \
412: \includegraphics[width=12.0cm,angle=0]{ZKSSfig2.ps}
413: \vskip 0.5cm
414: \caption{
415: Eigenvalues of families of bistochastic matrices, which produce the boundary of the set $\Sigma_N^B$. They are constructed of linear combinations,
416: $xP_a+(1-x)P_b$, of permutation matrices of size $N=3$:
417: a) $(P_{(123)}$ and $P_{(132)})$,
418: b) $(P_{(123)}$ and $I_3$),
419: c)~$(P_{(12)(3)}$ and $I_2)$,
420: and of size $N=4$:
421: d) $(P_{(1432)}$ and $P_{(13)(24)})$,
422: e) $(P_{(1432)}$ and $I_4) $,
423: f) $(P_{(1342)}$ and $P_{(132)(4)})$.
424: Panels a), b), d), and f) show spectra of combinations of commuting,
425: and c) and f) of noncommuting matrices.}
426: \label{fig2}
427: \end{center}
428: \end{figure}
429: %\vskip -2.0cm
430:
431: \medskip
432:
433:
434: \section{Majorization and entropy of bistochastic matrices}
435:
436: \subsection{Majorization}
437:
438: Consider two vectors $\vec{x}$ and $\vec{y}$, consisting of $N$ non-negative components each. Let us assume they are normalised in a sense that $\sum_{i=1}^{N}x_{i}=\sum_{i=1}^{N}y_{i}=1$. The theory of {\sl majorization} introduces a partial order in the set of such vectors \cite{MO79}. We say that $\vec{x}$ {\sl is majorized} by $\vec{y}$, written ${\vec{x}} \prec {\vec{y}}$, if
439: \begin{equation}
440: \sum_{i=1}^{k}x_{i}\leq \sum_{i=1}^{k}y_{i} \label{major}
441: \end{equation}
442: for every $k=1,2,\dots ,N$, where we ordered the components of each vector in the decreasing order, $x_{1}\geq \ldots \geq x_{N}$ and $y_{1}\geq \ldots \geq y_{N}$. Vaguely speaking, the vector $\vec{x}$ is more `mixed' than the vector, $\vec{y}$.
443:
444: A following theorem by Hardy, Littlewood, and Polya applies the bistochastic
445: matrices to the theory of majorization \cite{MO79}.
446:
447: {\bf Theorem 1. (HLP)} {\sl For any vectors $\vec x$ and $\vec y$, with sum of
448: their components normalised to unity, ${\vec x} \prec {\vec y}$ if and only if}
449: \begin{equation}
450: {\vec x}=B{\vec y} \quad {\rm for \quad some
451: \quad bistochastic \quad matrix} \quad B.
452: \label{bimaj}
453: \end{equation}
454: It was later shown by Horn \cite{Ho54} that in the above theorem the word `bistochastic'
455: can be replaced by `orthostochastic'. In general, the orthostochastic matrix
456: $B$ satisfying the relation ${\vec x}=B{\vec y}$ is not unique.
457:
458: The functions $f$, which preserve the majorization order:
459: \begin{equation}
460: {\vec x} \prec {\vec y} \quad {\rm implies } \quad f(\vec x) \le f(\vec y).
461: \label{shur}
462: \end{equation}
463: are called {\sl Schur convex} \cite{MO79}. Examples of Schur convex functions include $h_q({\vec x})=\sum_{i=1}^N x_i^q$ for any $q\ge1$, and $h({\vec x})=\sum_{i=1}^N x_i \ln x_i$.
464:
465: The degree of mixing of the vector $\vec x$ can be characterised by its Shannon entropy $H$ or
466: the generalised R\'{e}nyi entropies $H_q$ ($q \ge 1$):
467:
468: \begin{eqnarray}
469: H ({\vec x}) = -\sum_{i=1}^N x_i \ln x_i , \nonumber \\
470: H_q ({\vec x}) = {\frac{1 }{1-q}} \ln \Bigl( \sum_{i=1}^N x_i^q \Bigr).
471: \label{renyi}
472: \end{eqnarray}
473: In the limiting case one obtains $\lim_{q\to 1} H_q({\vec x})= H({\vec x})$.
474: If $\vec x$ represents the non-negative eigenvalues of a density matrix, $H$
475: is called the von Neumann entropy. Let ${\vec x}=B{\vec y}$. Due to the
476: Schur-convexity we have $H_q({\vec x}) \ge H_q({\vec y})$, since we changed
477: the sign and reversed the direction of inequality. An interesting application
478: of the theory of majorization in the space of density matrices representing
479: mixed quantum states is recently provided by Nielsen \cite{Ni99b}. Consider a
480: mixed state $\rho$ with the spectrum consisted of non-negative eigenvalues
481: $\lambda_{1} \geq \ldots \geq \lambda_{N}$, which sum to unity. This state may
482: be written as a mixture of $N$ pure states,
483: $\rho=\sum_{i=1}^N p_i |\psi_i\rangle \langle \psi_i|$ if, and only if, ${\vec
484: p} \prec {\vec \lambda} = (\lambda_1,\ldots,\lambda_N)$, where ${\vec p} =
485: \left(p_1,\ldots,p_k \right)$ (if both vectors have different length, the
486: shorter is extended by a sufficient number of extra components equal to zero).
487: This statement is not true if the pure states states
488: $|\psi_i\rangle $ are required to be distinct \cite{BE02}.
489:
490:
491: Consider now the non-unitary dynamics of the density operators given by
492: $\rho\to \rho'=\sum_{j=1}^{L} q_j U_j \rho U_j^{\dagger}$. This process,
493: called {\sl random external fields} \cite{AL87}, is described by $L$
494: unitary operations $U_j$ ($j=1,\dots,L$) and non-negative probabilities
495: satisfying $\sum_{j=1}^L q_j=1$. Denoting respective spectra, ordered
496: decreasingly, by $\vec \lambda$ and ${\vec \lambda}'$ one can find
497: unistochastic matrix $B$ such that ${\vec \lambda}'=B{\vec \lambda}$, so due
498: to the HLP theorem we have $\vec{\lambda}' \prec {\vec \lambda}$
499: \cite{Ul71,Ni00}. Therefore, after each iteration the mixed state becomes more
500: mixed and its von Neumann entropy is non-decreasing.
501:
502:
503: \subsection{Preorder in the space of bistochastic matrices}
504:
505: A bistochastic matrix $B$ acting on a probability vector $\vec{x}$ makes it
506: more mixed and increases its entropy. To settle which bistochastic matrices
507: have stronger mixing properties one may introduce a relation (preordering) in
508: the space of bistochastic matrices \cite{Sh52} writing
509:
510: \begin{equation}
511: B_{1}\prec B_{2}\quad\text{iff}\quad B_{1}=BB_{2} \text{ for some bistochastic
512: matrix }B \text{ .}
513: \label{maj3}%
514: \end{equation}
515: We have already distinguished some bistochastic matrices: permutation
516: matrices $P$ with only $N$ non-zero elements, and the uniform van der Waerden
517: matrix $B_{\ast}$, with all its $N^{2}$ elements equal to $1/N$.
518: For an arbitrary bistochastic matrix $B$ and for all permutation matrices
519: $P$ we have $B_{\ast}=B_{\ast}B$ and $B=\left( BP^{-1}\right) P$, and hence
520: $B_{\ast}\prec B\prec P$. The relation $B_{1}\prec B_{2}$ implies
521: $B_{1}\vec{y}\prec B_{2}\vec{y}$ for every probability vector $\vec{y}$, but
522: the converse is not true in general (see \cite{Sh54} and \cite{Sc58}).
523:
524:
525: \subsection{Entropy of bistochastic matrices }
526:
527: To measure mixing properties of a bistochastic matrix $B$ of size $N$
528: one may consider its entropy. We define \textsl{Shannon entropy} of $B$ as
529: the mean entropy of its columns (rows), which is equivalent to
530:
531: \begin{equation}
532: H(B):=-\frac{1}{N}\sum_{i=1}^{N} \sum_{j=1}^{N}B_{ij}\ln B_{ij} \text{ .}
533: \label{sha}
534: \end{equation}
535: As usual in the definitions of entropy we set $0\ln0=0$, if necessary.
536: The entropy changes from zero for permutation matrices to $\ln N$ for the
537: uniform matrix $B_{\ast}$.
538:
539: Due to the majorization of each column vector,
540: the relation $C\prec B$ implies $H(B)\leq H(C)$, but
541: the converse is not true.
542: To characterise quantitatively the effect of entropy increase under the action
543: of a bistochastic matrix $B$, let us define the {\textsl{weighted entropy}}
544: of matrix $B$ with respect to a probability vector
545: $\vec{y}=\left(y_{1},\ldots,y_{N}\right) $:
546:
547: \begin{equation}
548: H_{\vec{y}}(B):=\sum_{k=1}^{N}y_{k}H(B_{k})=-\sum_{k=1}^{N}%
549: y_{k}\sum_{j=1}^{N}B_{jk}\ln B_{jk} \text{ ,}
550: \label{entwei}%
551: \end{equation}
552: where $B_k$ is a probability vector defined by $B_{k}:=\left(B_{1k},\ldots,B_{Nk}\right)$
553: for $k=1,\ldots,N$. In this notation $H(B) = H_{e_*}(B)$, where $e_* =
554: (1/N,\ldots,1/N)$. The weighted entropy allows one to write down the following
555: bounds for $H(B\vec{y})$
556:
557: \begin{equation}
558: \max\left\{ H(\vec{y}),H_{\vec{y}}(B)\right\} \leq
559: H(B\vec{y})\leq H(\vec{y})+H_{\vec{y}%
560: }(B)\text{ .}\label{ineq2}%
561: \end{equation}
562: the proof of which is provided elsewhere \cite{Sl01}. These bounds have certain
563: implications in quantum mechanics. For example, if a non-unitary evolution
564: of the density operator under the action of random external field is
565: considered \cite{Ni99b}, they tell us how much the von Neumann entropy of the
566: mixed state may grow during each iteration.
567:
568: Using the above proposition we shall show that the entropy of bistochastic matrices
569: is subadditive. Namely, the following theorem holds:
570:
571: \medskip
572:
573: {\bf Theorem 2.} {\sl Let $A$ and $B$ be two bistochastic matrices. Then}
574: \begin{equation}
575: \max\left\{ H(A),H(B)\right\} \leq H(AB)\leq H(A)+H(B)\text{ ,}
576: \label{ineq3}
577: \end{equation}
578: {\sl and, analogously,}
579: \begin{equation}
580: \max\left\{ H(A),H(B)\right\} \leq H(BA)\leq H(A)+H(B)\text{ .}%
581: \label{ineq4}
582: \end{equation}
583: %\end{theorem}
584:
585: %\begin{proof}
586: {\bf Proof.}
587: We put $C:=AB$ and consider stochastic vectors
588: $\vec{y}_n :=\left( c_{1n},\ldots,c_{Nn}\right) $ for $n=1,\ldots,N$ (the
589: columns of the matrix $C$). Applying (\ref{ineq2}) we get \[\max\left(
590: H_{\vec{y}_{n}}(B),H\left( \vec{y}_n \right) \right) \leq H\left(
591: B\vec{y}_n\right) \leq H_{\vec{y}_n}(B)+H\left( \vec{y}_n\right) .\] Hence
592: \[\max\left({\textstyle\sum\nolimits_{k=1}^{N}} c_{kn}%
593: {\textstyle\sum\nolimits_{j=1}^{N}} \eta\left( b_{jk}\right) ,
594: {\textstyle\sum\nolimits_{k=1}^{N}}
595: \eta\left( c_{kn}\right) \right) \leq
596: {\textstyle\sum\nolimits_{k=1}^{N}}
597: \eta\left(
598: {\textstyle\sum\nolimits_{j=1}^{N}}
599: b_{kj}c_{jn}\right)
600: \]
601: and
602: \[
603: {\textstyle\sum\nolimits_{k=1}^{N}}
604: \eta\left(
605: {\textstyle\sum\nolimits_{j=1}^{N}}
606: b_{kj}c_{jn}\right) \leq
607: {\textstyle\sum\nolimits_{k=1}^{N}}
608: c_{kn}%
609: {\textstyle\sum\nolimits_{j=1}^{N}}
610: \eta\left( b_{jk}\right) +
611: {\textstyle\sum\nolimits_{k=1}^{N}}
612: \eta\left( c_{kn}\right) \text{ ,}
613: \]
614: where $\eta\left( x\right) =-x\ln x$ for $x>0$. Multiplying the above
615: equalities by $1/N$ and summing over $n=1,\ldots,N$ we get (\ref{ineq3}).
616: Setting $C=BA$ we obtain (\ref{ineq4}) in an analogous way. $\Box$
617: %\end{proof}
618:
619: \bigskip
620:
621: \section{Ensembles of random unistochastic matrices}
622: \label{secIV}
623:
624: \subsection{Unistochastic matrices}
625:
626: To demonstrate that a given bistochastic matrix $B$ is unistochastic
627: one needs to find unitary matrix $U$ such that $B_{ij}=|U_{ij}|^2$.
628: In other words one needs to find a solution of the
629: coupled set of nonlinear equations enforced by the unitarity
630: $UU^{\dagger}=U^{\dagger}U={\mathbb I}$,
631: \begin{equation}
632: \sum_{j=1}^N \sqrt{B_{jk}B_{jl}} \exp(i \phi_{jk}- \phi_{jl}) = \delta_{kl}
633: \ {\rm ~~for ~~ all ~~} 1\leq k<l\leq N
634: \label{unit1}
635: \end{equation}
636: for the unknown phases of each complex element
637: $U_{jk}=\sqrt{B_{jk}} e^{i \phi_{jk}}$.
638: The diagonal constraints for $k=l$ are fulfilled,
639: since $B$ is bistochastic.
640:
641: For $N=2$ all bistochastic matrices are orthostochastic
642: (see eg. Eq.(\ref{ttransform})),
643: and so $\Omega_2^B=\Omega_2^U=\Omega_2^O$. This is not the case for higher
644: dimensions, for which there exist bistochastic matrices which are not
645: unistochastic, and so $\Omega_N^U \varsubsetneq \Omega_N^B$ for $N \geq 3$.
646: Thus $\Omega_N^U$ is not a convex set for $N \geq 3$, since it contains all
647: the permutation matrices and is smaller than their convex hull, which,
648: according to the Birkhoff theorem, is equal to $\Omega_N^B$ . Simple examples
649: of bistochastic matrices which are not unistochastic were already
650: provided (for $N=3$) by Schur and Hoffman \cite{MO79},
651:
652: \begin{equation}
653: B_{1}=\frac{1}{2} \left[
654: \begin{array}{ccc}
655: 1 & 1 & 0 \\
656: 1 & 0 & 1 \\
657: 0 & 1 & 1
658: \end{array}
659: \right] ,\text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }B_{2}=\frac{1}{6} \left[
660: \begin{array}{ccc}
661: 0 & 3 & 3 \\
662: 3 & 1 & 2 \\
663: 3 & 2 & 1
664: \end{array}
665: \right] .
666: \label{matkontr}
667: \end{equation}
668:
669: %\vskip -2.0cm
670: \begin{figure} [htbp]
671: %\vskip -1.0cm
672: \begin{center}
673: \
674: \includegraphics[width=10.0cm,angle=0]{ZKSSfig3.eps}
675: \vskip 0.5cm
676: \caption{
677: Chain rule for unistochasticity for $N=3$: (see (\ref{necstouni1}) and
678: (\ref{necstouni2})), a) the longest link $L_1 > L_2+L_3$ so the matrix
679: $B$ is not unistochastic, b) condition for orthostochasticity $L_1 = L_2+L_3$,
680: c) weaker condition for unistochasticity $L_1 \leq L_2+L_3$.}
681: \label{fig2b}
682: \end{center}
683: \end{figure}
684: %\vskip -1.0cm
685:
686: \medskip
687:
688: To see that there exists no corresponding unitary matrix $U$
689: we shall analyse constraints implied by the unitarity.
690: Define vectors containing square roots of the column (row) entries, e.g.,
691: ${\vec v}_k:=\{ \sqrt{B_{k1}}, \sqrt{B_{k2}},..., \sqrt{B_{kN}} \}$.
692: The scalar products of any pair of any two different vectors
693: $\vec{v}_{k}\cdot \vec{v}_{l}$ consists of $N$ terms,
694: $L_n=\sqrt{ B_{kn}B_{ln}}$, $n=1,...,N$.
695: In the case of $B_1$ from (\ref{matkontr})
696: the scalar product related to the two first columns
697: $\left(\vec{v}_{1},\vec{v}_{2}\right)$ consists of three terms
698: $L_1=1$, $L_2=L_3=0$, which do not satisfy the triangle inequality.
699: Thus it is not possible to find the corresponding
700: phases $\phi_{jk}$ satisfyig (\ref{unit1}).
701: This observation allows us to
702: obtain a set of {\sl necessary} conditions, a bistochastic $B$ must satisfy to
703: be unistochastic \cite{PZK01}:
704:
705: \begin{equation}
706: \max_{m=1,\ldots,N} \sqrt{B_{mk}B_{ml}}~\le ~ \frac{1}{2}
707: \sum_{j=1}^N \sqrt{B_{jk}B_{jl}} \ ,{\rm ~~for ~~ all ~~} 1\leq k<l\leq N
708: \label{necstouni1}
709: \end{equation}
710: and
711: \begin{equation}
712: \max_{m=1,\ldots,N} \sqrt{B_{km}B_{lm}}~\le ~ \frac{1}{2}
713: \sum_{j=1}^N \sqrt{B_{kj}B_{lj}} \ , {\rm ~~for ~~ all ~~} 1\leq k<l\leq N \ .
714: \label{necstouni2}
715: \end{equation}
716:
717:
718: We shall refer to the above inequalities as to a `chain rule': the longest
719: link $L_1$ of a closed chain cannot be longer than the sum of all other links
720: $L_2+...+L_N$ - see Fig. \ref{fig2b}.
721: The set of $N(N-1)/2$ conditions (\ref{necstouni1}) (resp. (\ref{necstouni2}))
722: treats all possible pairs of the columns (resp. rows) of $B$. Only for
723: $N=3$ the both sets of conditions are equivalent
724: (since the equations (\ref{unit1}) for the phases may be separated \cite{Pa02}),
725: but for $N\ge 4$ there exist matrices which
726: satisfy only one class of the constraints.
727: For example, a bistochastic matrix
728: \begin{equation}
729: B_4 = \frac{1}{100} \left[
730: \begin{array}{rrrr}
731: $1$\ & 17 \ & 25 \ & 57 \\
732: $38$\ & 38 \ & 24 \ & 0 \\
733: $42$\ & 5 \ & 29 \ & 24 \\
734: $19$\ & 40 \ & 22 \ & 19
735: \end{array}
736: \right],
737: \end{equation}
738: found by Pako{\'n}ski \cite{Pa02},
739: satisfies the row conditions (\ref{necstouni2}), but violates the
740: column conditions (\ref{necstouni1}), so cannot be unistochastic
741: (the third term of the product of the roots of the first and the fourth column
742: is larger than the sum of the remaining three terms).
743: It is
744: easy to see that for an arbitrary $N$ these necessary conditions are satisfied
745: by any bistochastic $B$ sufficiently close to the van der Waerden matrix $B_*$,
746: for which all links of the chain are equal, $L_i=1/N$. It is then tempting to
747: expect that there exists an open vicinity of $B_*$ included in $\Omega_N^U$,
748: i.e., that $B_*$ lies in the interior of $\Omega_N^U$ and, consequently, that
749: $\Omega_N^U$ has positive volume. Certain conditions {\sl
750: sufficient} for unistochasticity were already found by Au-Yeung and Cheng
751: \cite{AYC91}, but they do not answer the question concerning the volume
752: of $\Omega_N^U$. On the other hand, it is well known that the set of
753: unistochastic matrices is connected and compact \cite{He78}, and is not dense
754: in the set of bistochastic matrices \cite{Dj66}.
755:
756: To analyse properties of bistochastic matrices it is convenient to introduce
757: so called $T$-transforms, which in a sense reduce the problem to two
758: dimensions. The $T$-transform acts as the identity in all but two dimensions,
759: in which it has a common form of an orthostochastic matrix
760:
761: \begin{equation}
762: {\tilde T}\left( \varphi \right) = \left[
763: \begin{array}{cc}
764: \cos^2 \varphi & \sin^2 \varphi \\
765: \sin^2 \varphi & \cos^2 \varphi
766: \end{array}
767: \right]
768: { \quad \quad \rm such \quad that} \quad \quad
769: {\tilde O}_2 \left( \varphi \right) = \left[
770: \begin{array}{cc}
771: \cos \varphi & \sin \varphi \\ - \sin \varphi & \cos \varphi
772: \end{array}
773: \right]
774: \label{ttransform}
775: \end{equation}
776: is orthogonal.
777: Any matrix $B$ obtained as a sequence of at most $(N-1)$ $T-$transforms,
778: $B = T_{N-1} \cdots T_2 T_1$, where each $T_k$ acts in the
779: two-dimensional subspace spanned by the base vectors $k$ and $k+1$,
780: is orthostochastic. To show this it is enough to observe that
781: each element of $B$ is a product of
782: non-trivial elements of the transformations $T_k$.
783: Hence taking its square root
784: and adjusting the signs one may find a corresponding orthogonal matrix defined
785: by the products of the elements of ${\tilde O}_k$ \cite{MO79,Bh97,Ni99}.
786: Although any product of an arbitrary number of T-transforms
787: satisfies the chain--links conditions \cite{BE02},
788: for $N\ge 4$ there exist products of a finite number of
789: T-transforms (also called {\sl pinching matrices}) which are not
790: unistochastic \cite{PT87}. In the same paper it is also shown that
791: there exist unistochastic matrices which cannot be written
792: as a product of $T$-transforms.
793:
794: Consider a unistochastic matrix $B$ and the set ${\cal U}_B \subset U(N) $
795: of all unitary matrices corresponding to $B$ in the sense that
796: $B_{ij}=|U_{ij}|^2$ for $i,j=1,\ldots,N$. Such sets endowed with
797: appropriate probability measures play a role in the theory of quantum graphs
798: \cite{Ta00,PZK01,Ta01} and were called {\sl unitary stochastic ensembles} by
799: Tanner \cite{Ta01}. It is easy to see that these sets are invariant under the
800: operations $U \to V_{1}UV_{2}$, where $V_1$ and $V_2$ are arbitrary diagonal
801: unitary matrices. The dimensionality arguments suggest that, having $U \in
802: {\cal U}_B$ fixed, each element of ${\cal U}_B$ can be obtained in this way
803: \cite{false}. However, in general this conjecture is false and for certain
804: bistochastic matrices $B$ the set ${\cal U}_B$ is larger. This is shown in
805: Appendix \ref{sec:unibist}, in which a counterexample for $N=4$ is provided.
806:
807:
808: \subsection{Definition of ensembles}
809:
810: To analyse random unistochastic matrices one needs to specify a probability
811: measure in this set. As it was already discussed in the introduction,
812: unistochastic (USE) and orthostochastic (OSE) ensembles can be defined with
813: help of the Haar measure on the group of unitary matrices $U(N)$ and
814: orthogonal matrices $O(N)$, respectively \cite{Be01}. In other words the
815: bistochastic matrices
816:
817: \begin{equation}
818: B^U_{ij}:=|U_{ij}|^2, \quad {\rm and} \quad B^O_{ij}:=(O_{ij})^2,
819: \label{uniort}
820: \end{equation}
821: pertain to USE and OSE respectively, if the matrices $U$ and $O$ are generated
822: with respect to the Haar measures on the unitary (orthogonal) group.
823:
824: Dynamical systems with time reversal symmetry are described by unitary symmetric
825: matrices \cite{Haake}. The ensemble of these matrices, defined by $W=UU^T$,
826: is invariant with respect to orthogonal rotations, and is called {\sl circular
827: orthogonal ensemble} (COE). In an analogous way we may thus define the
828: following three ensembles of symmetric bistochastic matrices (SBM)
829:
830: \begin{equation}%
831: \begin{tabular}
832: [c]{llll}%
833: a)\qquad & $S_{1}:=BB^{T};$ & $\quad\mathrm{so}\quad$ & $S_{ij}:=\sum
834: _{k=1}^{N}|U_{ik}|^{2}|U_{jk}|^{2}$,$\medskip$\\
835: b)\qquad & $S_{2}:={\frac{1}{2}}(B+B^{T});$ & $\quad\mathrm{so}\quad$ &
836: $S_{ij}={\frac{1}{2}}(|U_{ij}|^{2}+|U_{ji}|^{2})$,$\medskip$\\
837: c)\qquad & $S_{3}:=|W_{ij}|^{2}=|(UU^{T})_{ij}|^{2};$ & $\quad\mathrm{so}%
838: \quad$ & $S_{ij}:=|\sum_{k=1}^{N}U_{ik}U_{jk}|^{2}, \medskip\medskip$
839: \end{tabular}
840: \label{bisym}
841: \end{equation}
842:
843: where bistochastic matrices $B$ are generated according to USE
844: (or, equivalently, unitary matrices $U$ are generated according to CUE).
845:
846:
847: \subsection{Spectra of random unistochastic matrices}
848:
849: Since the sets of the bi-, uni--, and (ortho--)stochastic matrices are related by
850: the inclusion relations: \linebreak $\Omega_N^O \subseteq \Omega_N^U \subseteq
851: \Omega_N^B \subset \Omega_N^S$, analogous relations must hold for the
852: supports of their spectra, $\Sigma_N^O \subseteq \Sigma_N^U \subseteq
853: \Sigma_N^B \subseteq \Sigma_N^S$. For $N=2$ the spectrum of a bistochastic
854: matrix must be real. The subleading eigenvalue $\lambda_2 = 2B_{11}-1$, which
855: allows one to obtain the distributions along the real axis.
856: For USE the respective density is constant, $P_r(\lambda)=1/4$ for
857: $\lambda\in (-1,1)$, while for OSE it is given by the formula,
858: $P_r(\lambda)=1/(2\pi \sqrt{1-\lambda^2})$.
859: These densities are normalized to $1/2$, since for any
860: random matrix its leading eigenvalue is equal to unity.
861:
862:
863: For larger matrices we generated random unitary (orthogonal) matrices with respect to the Haar measure by means of the Hurwitz parameterisation \cite{Hu97} as described in \cite{ZK94,PZK98}. Squaring each element of the matrices
864: generated in this way we get random matrices typical for USE (OSE). In
865: general, the density of the spectrum of random uni--, (ortho--)stochastic matrices
866: may be split into three components:
867:
868: a) two-dimensional density with the domain forming a subset $\Sigma_N^U$ ($\Sigma_N^O$)
869: of the unit circle;
870:
871: b) one-dimensional density at the real line described by the function $P_r(x)$ for $x\in [-1,1]$,
872:
873: and
874:
875: c) the Dirac delta $\frac{1}{N} \delta(z-1)$ describing the leading
876: eigenvalue, which exists for every unistochastic matrix.
877:
878: %\vskip -1.0cm
879: \begin{figure} [htbp]
880: %\vskip -1.0cm
881: \begin{center}
882: \
883: \includegraphics[width=10.0cm,angle=0]{ZKSSfig4.ps}
884: \vskip 0.5cm
885: \caption{
886: Spectra of random unistochastic matrices of size a) $N=3$ ($1200$ matrices)
887: and b) $N=4$ ($800$ matrices); spectra of random orthostochastic matrices of
888: size c) $N=3$ ($1200$ matrices) and d) $N=4$ ($800$ matrices). Thin lines
889: denote $3$-- and $4$--hypocycloids, while the thick lines represent the
890: $3$--$4$ interpolation arc (part of it is shown in Fig.~2f). } \end{center}
891: \end{figure}
892: %\vskip -2.0cm
893:
894: %\medskip
895:
896: Basing on the numerical results presented in Fig.~4a, we conjecture that
897: $\Sigma_3^U=\Sigma_3^O$ and consists of a real interval (already present for
898: $N=2$) and the inner part of the $3$--hypocycloid. This curve is drawn by a
899: point at a circle of radius $1/3$, which rolls (without sliding) inside the
900: circle of radius $1$. The parametric formula reads
901:
902: \begin{equation}
903: \left\{
904: \begin{array}
905: [c]{ccc}
906: x & = & {\frac{1}{3}}(2\cos\phi+\cos2\phi),\medskip\\
907: \ y & = & {\frac{1}{3}}(2\sin\phi-\sin2\phi),
908: \end{array}
909: \right. \label{hyper3}
910: \end{equation}
911: where $\phi\in [0,2\pi)$.
912:
913: To find the unistochastic matrices with spectra at the cycloid consider a
914: two--parameter family of combinations of permutation matrices
915: $a^2{\mathbb I}+b^2P+c^2P^2$ with $a^2+b^2+c^2=1$. Here $P$ represents the
916: nontrivial $3$ elements permutation matrix, $P=P_{(123)}$, so $P^3={\mathbb I}$.
917: One--parameter family of these bistochastic matrices, which satisfy the
918: condition for unistochasticity, produce spectra located along the entire
919: hypocycloid. Consider the matrices
920:
921: \begin{equation}
922: O_{3}(\varphi): =
923: \left[
924: \begin{array}{ccc}
925: a & b & c \\
926: c & a & b \\
927: b & c & a
928: \end{array}
929: \right]
930: \quad \quad {\rm and} \quad \quad
931: B_{3}(\varphi): =
932: \left[
933: \begin{array}{ccc}
934: a^2 & b^2 & c^2 \\
935: c^2 & a^2 & b^2 \\
936: b^2 & c^2 & a^2
937: \end{array}
938: \right] ,
939: \label{hipo3}
940: \end{equation}
941: where their elements depend on a single parameter $\varphi \in [0,2\pi)$ and
942:
943: \begin{equation}
944: \left\{
945: \begin{tabular}{lll}
946: $a=$ & $a(\varphi )=$ & $-\frac{1}{3}(1+2\cos \varphi ) \text{ ,}$ \\
947: $b=$ & $b(\varphi )=$ & $\frac{1}{3}(\cos \varphi -1)+\frac{1}{\sqrt{3}}\sin
948: \varphi \text{ ,}$ \\
949: $c=$ & $c(\varphi )=$ & $\frac{1}{3}(\cos \varphi -1)-\frac{1}{\sqrt{3}}\sin
950: \varphi \text{ .}$%
951: \end{tabular}
952: \right.
953: \label{hyper3bi}
954: \end{equation}
955: It is easy to see that $a^2+b^2+c^2=1$ and $ab+bc+ca=0$, so $O_3$ is orthogonal
956: and $B_3$ is orthostochastic. Simple algebra shows that the spectrum of
957: $B_3(\varphi)$ forms the $3$--hypocycloid given by (\ref{hyper3}) - see Appendix
958: \ref{sec:hypoc}.
959:
960: An alternative approach, based on exponentiation of permutation matrices,
961: leads to unistochastic matrices with spectrum on a hypocycloid. Let
962: $P_{N}:=P_{(12\cdots N)}$ be the nontrivial permutation matrix of size $N$.
963: Since $P_{N}$ is unitary, so is its power $(P_{N})^{\alpha}$. We define it
964: for an arbitrary real exponent by
965: $(P_{N})^{\alpha}:=U^\dagger D^\alpha U$,
966: where $U$ is an unitary matrix diagonalizing $P_N$
967: and $D$ represents the diagonal matrix of its eigenvalues --
968: it is assumed that the eigenphases of such a matrix
969: belong to $[0,2\pi)$. Since $P_{N}^0={\mathbb I}_N$, one may
970: expect that defining the corresponding bistochastic matrices and varying the
971: exponent $\alpha$ from zero to unity one obtains an arc of a hypocycloid.
972: This fact is true as shown in Appendix \ref{sec:hypocun}, in which we
973: prove a general result, valid for an arbitrary matrix size.
974:
975: \medskip
976:
977: {\bf Proposition 3.} {\sl
978: The spectra of the family of unistochastic matrices of size $N$}
979:
980: \begin{equation}
981: B^{(N,\alpha)}_{ij}:= \bigr| \bigr(P_{N}^{\alpha}\bigl)_{ij}\bigl|^2
982: {\rm \quad with \quad}
983: \alpha \in [ 0, N/ 2],
984: \label{hypenu}
985: \end{equation}
986: {\sl generate the $N$-hypocycloid (and
987: inner diagonal hypocycloids of the
988: radius ratio $k/N$ with $k=2,\ldots,N-1$).}
989:
990: \begin{figure} [htbp]
991: \begin{center}
992: \
993: \includegraphics[width=12.0cm,angle=0]{ZKSSfig5.ps}
994: \vskip 0.5cm
995: \caption{
996: Spectra of a family of bistochastic matrices
997: $B^{(N,\alpha)}$ with $\alpha \in [0,N/2]$ obtained by exponentiation of
998: the permutation matrix $P_{N}$ plotted for a) $N=5$ (hypocycloids $5$ and
999: $5/2$), b) $N=6$ (hypocycloids $6$, $3=6/2$ and $2=6/3$), and c) $N=7$
1000: (hypocycloids $7$, $7/2$ and $7/3$). }
1001: \end{center}
1002: \end{figure}
1003:
1004: Let $H_N$ denote the set bounded by $N-$hypocycloid. Fig.~4c suggests that
1005: $H_3$ is contained in $\Sigma_3^O$. This conjecture may be supported by
1006: considering the spectra interpolating between the origin $(0,0)$ and a
1007: selected point on the hypocycloid. To find such a family we shall use the
1008: Fourier matrix $F^{(N)}$ of size $N$ with elements $F^{(N)}_{kl}:=\exp(-2kl\pi
1009: i/N)/\sqrt{N}$. Since amplitudes of all elements of this matrix are equal, the
1010: corresponding bistochastic matrix equals to the van der Waerden matrix $B_*$
1011: for which all subleading eigenvalues vanish. The matrix $P_{N}^{\alpha}$
1012: generates a unistochastic matrix $B^{(N,\alpha)}$ with an eigenvalue at the
1013: hypocycloid, so the family of unistochastic matrices related to
1014: $\bigl(P_{N}^{\alpha}\bigr)^{\beta} (F^{(N)})^{1-\beta}$ provides an
1015: interpolation between the origin and a selected point on the hypocycloid - see
1016: Fig.~6. In other words, we conjecture that the spectra of a two parameter
1017: family of unistochastic matrices obtained from $(P_{N})^{a} (F^{(N)})^{b}$
1018: explore the entire $H_{N}$.
1019:
1020: Numerical results obtained for $N=3$ random uni--, (ortho--)stochastic matrices allow us to
1021: claim that there are no complex eigenvalues outside the 3-hypocycloid, so
1022: that $\Sigma_3^U=\Sigma_3^O=H_2\cup H_3$.
1023: Interestingly, $H_3$ - the $3$-hypocycloid and its interior,
1024: determines the set of all unistochastic matrices which
1025: belong to the cross section of $\Omega_3^B$ defined by the plane spanned by
1026: $P_3, P_3^2$ and $P_3^3={\mathbb I}$ \cite{BE02}, see also Appendix
1027: \ref{sec:hypoc}.
1028:
1029:
1030:
1031: \begin{figure} [htbp]
1032: \begin{center}
1033: \
1034: \includegraphics[width=10.0cm,angle=0]{ZKSSfig6.ps}
1035: \vskip 0.5cm
1036: \caption{
1037: Spectra of a family of bistochastic matrices obtained be squaring
1038: elements of unitary matrices $\bigl(P_{N}^{\alpha}\bigr)^{\beta}
1039: (F^{(N)})^{1-\beta}$ with $\beta\in [0,1]$ and
1040: a) $N=3$, the curves are labelled by $\alpha=0,1/8,2/8,\ldots,3/2$,
1041: b) $N=4$, $\alpha=0,1/8,2/8,\ldots,2$.
1042: For reference we plotted the hypocycloids with a thin line.
1043: }
1044: \end{center}
1045: \end{figure}
1046:
1047: Numerical results for spectra of random uni--, (ortho--)stochastic matrices for $N=4$ are shown
1048: in Fig.~4b~and~4d. The set $\Sigma_4^U$ includes the entire set $\Sigma_3^U$,
1049: but also the $4$--hypocycloid, sometimes called {\sl asteroid}, and formed
1050: by the solutions of the equation
1051: \begin{equation}
1052: x^{2/3} + y^{2/3} =1.
1053: \label{asteroid}
1054: \end{equation}
1055:
1056: More generally, the spectra of orthostochastic matrices of size $N$ contain
1057: the $N$--hypocycloid $H_N$
1058:
1059: \begin{equation}
1060: \left\{
1061: \begin{array}{ccc}
1062: x&=&{\frac{1 }{N}} [(N-1) \cos \phi +\cos (N-1) \phi], \\
1063: y&=&{\frac{1 }{N}} [(N-1) \sin \phi -\sin (N-1) \phi],
1064: \end{array}
1065: \right.
1066: \label{hyperN}
1067: \end{equation}
1068: with corner at $1$, as it is proved in Appendix \ref{sec:hypoc}. The set $\Sigma^U_N$ contains of
1069: the sets bounded by hypocycloids of a smaller dimension, which we denote by
1070: $G_N: = \bigcup_{k=2}^N H_k$.
1071:
1072: However, as seen in Fig.~4b there exist some eigenvalues of unistochastic or
1073: even orthostochastic matrices outside this set.
1074: In fact one needs to find interpolations between roots of unity of different
1075: orders (the corners of $k$ and $n$--hypocycloid) based on the families of
1076: orthostochastic matrices. Such a family interpolating between the corners of
1077: $H_3$ and $H_4$ is plotted in Fig.~2f, since these bistochastic matrices are
1078: orthostochastic. A general scheme of finding the required interpolations is
1079: based on the fact that, if $O(\varphi)$ is a family orthogonal matrices and
1080: $B_{ij}(\varphi)=O_{ij}^2(\varphi)$ is the corresponding family of
1081: orthostochastic matrices, than another such family is produced by the
1082: multiplication by some permutation matrix $P$: $O'(\varphi) = O(\varphi)P$.
1083: For example, if the matrix $O_4(\varphi)$ of size $4$ contains elements
1084: $O_{11}=O_{22}=1$ and the block diagonal matrix $O_2(\varphi)$ (see
1085: (\ref{ttransform})), then the squared elements of the orthogonal matrices
1086: $O_4(\varphi)P_{(1234)}$ provide orthostochastic matrices, the spectra of
1087: which form the $3$--$4$ interpolating curve contained in $E_4$ and plotted in
1088: Fig.~2f~and~4d. To obtain orthostochastic matrices, the spectra of which
1089: provide the curve which joins both complex corners of $H_3$ and does not
1090: belong to it (see Fig 4d), one should take the orthogonal matrix $O_{3,1}$ of
1091: size $4$ with $O_{11}=1$ which contains the block diagonal matrix
1092: $O_3(\varphi)$ (\ref{hipo3}), and create bistochastic matrices out of
1093: $O_{3,1}(\varphi)P_{(12)(34)}$. The exact form of these families of
1094: orthostochastic matrices is provided in Appendix \ref{sec:hypoc2}.
1095:
1096:
1097:
1098: Let ${\tilde G}_N$ denote the set $G_N$ extended by adding the regions bounded
1099: by the interpolations between all neighbouring corners of $H_k$ and
1100: $H_n$ with $k<n \le N$ constructed in an analogous way as for $N=4$.
1101: The sequence of hypocycloids with the neighbouring corners
1102: is given by the order of fractions present in the
1103: {\sl Farey sequence} and {\sl Farey tree} \cite{Schuster}.
1104: Our investigation of the support of the spectra
1105: of unistochastic and orthostochastic matrices may be concluded by a relation
1106: analogous to (\ref{sigmabisto})
1107:
1108: \begin{equation}
1109: \bigcup_{k=2}^N H_k :=G_N \subset \Sigma_N^O \simeq \Sigma_N^U
1110: \simeq {\tilde G}_N \subset E_N=:\bigcup_{k=2}^N Z_k,
1111: \label{sigmaunisto}
1112: \end{equation}
1113: where the sign $\simeq$ represents the conjectured equality. Numerical results
1114: suggest that the support $\Sigma_N$ is the same for both ensembles: USE and
1115: OSE. However, the density of eigenvalues $P(z)$ is different for uni-- and
1116: (ortho--)stochastic ensembles. The repulsion of the complex eigenvalues from the
1117: real line is more pronounced for the orthostochastic ensemble, as
1118: demonstrated in Fig~4b~and~4d.
1119:
1120: The larger $N$, the better the domains $\Sigma_N^O$ and $\Sigma_N^U$ fills
1121: the unit disk. On the other hand, the eigenvalues of a large modulus are
1122: unlikely; the density $P(z)$ is concentrated in a close vicinity of the origin,
1123: $z=0$. Moreover, in this case the weight of the singular part of the density
1124: at the real line, decreases with the matrix size $N$. To characterise the
1125: spectrum qualitatively we analysed the densities of the distributions $P_k(r)$
1126: of the moduli of the largest eigenvalues ${\lambda}_k$. The results obtained
1127: for random unistochastic matrices are shown in Fig.~7a. Fig.~7b shows
1128: analogous data for the singular values $\sigma_i$ of $B$ -- per definition
1129: square roots of the real eigenvalues of the symmetric matrices $BB^T$
1130: \cite{HJ98}. Since the singular values bound moduli of eigenvalues from above
1131: \cite{HJ98}, the distribution $P(\sigma)$ is localised at larger values than
1132: $P(r)$. Thus the expectation values satisfy $\langle r_k\rangle <
1133: \langle\sigma_k\rangle$ as shown in Fig.~7c~and~7d. The modulus of the second
1134: eigenvalue decreases with matrix size as $N^{-1/2}$. These results are
1135: consistent with the recent work of Berkolaiko \cite{Be01}, who suggested
1136: describing the distribution $P_2(r)$ by the generalised extreme value
1137: distribution \cite{LLR83}.
1138:
1139: \vskip -3.9cm
1140: \begin{figure} [htbp]
1141: \begin{center}
1142: \
1143: \includegraphics[width=12.0cm,angle=0]{ZKSSfig7.ps}
1144: \vskip -2.9cm
1145: \caption{
1146: Unistochastic ensemble. Distribution of the modulus of the
1147: second $(\triangle)$, third $(\square)$, fourth $(\Diamond)$ and fifth
1148: $(\times)$ largest eigenvalues a), and singular value b) (lines are drawn to
1149: guide the eye). Expectation value $\langle r_2\rangle$ for $k=2,3,4,5$ of the
1150: subleading c) eigenvalues, and d) singular values as a function of the matrix
1151: size $N$. Exponential fit, represented by solid lines, give $\langle
1152: r_2\rangle \approx \exp(-0.503 )$ and $\langle \sigma_2 \rangle \approx
1153: \exp(-0.446)$ }.
1154: \end{center}
1155: \end{figure}
1156:
1157: \subsection{Average entropy}
1158:
1159: We can compute the mean entropy averaged over the ensembles of unistochastic or
1160: orthostochastic matrices. Clearly we have
1161:
1162: \begin{equation}
1163: \left\langle H\right\rangle _{USE}=N\cdot\left\langle -\left| U_{ij}\right|
1164: ^{2}\ln\left| U_{ij}\right| ^{2}\right\rangle _{CUE}
1165: \end{equation}
1166: and
1167:
1168: \begin{equation}
1169: \left\langle H\right\rangle _{OSE}=N\cdot\left\langle -\left| O_{ij}\right|
1170: ^{2}\ln\left| O_{ij}\right| ^{2}\right\rangle _{COE} \text{ ,}
1171: \end{equation}
1172: where $U_{ij}$ and $O_{ij}$ are entries of unitary or orthogonal matrices, respectively.
1173: The exact formulae for the above averages were obtained by Jones in
1174: \cite{Jo90,Jo91}. Using these results we get
1175:
1176: \begin{equation}
1177: \left\langle H\right\rangle _{USE}=\Psi(N+1)-\Psi(2)=\sum_{k=2}^{N}\frac{1}%
1178: {k}\label{shauni}%
1179: \end{equation}
1180: and
1181:
1182: \begin{equation}
1183: \left\langle H\right\rangle _{OSE} = \Psi(N/2+1)-\Psi(3/2)=\left\{
1184: \begin{tabular}
1185: [c]{ll}%
1186: $2\ln2-2+\sum\nolimits_{l=1}^{k}\frac{1}{l}$ & \qquad $N=2k$\\
1187: $2\sum\nolimits_{l=1}^{k}\frac{1}{2l+1}$ & \qquad $N=2k+1$%
1188: \end{tabular}
1189: \right. \text{ .}
1190: \label{shaorto}%
1191: \end{equation}
1192: where $\Psi$ denotes the digamma function. For large $N$ the digamma function behaves logarithmically, $\Psi(N)\sim\ln N$, and both averages display similar asymptotic behaviour $\left\langle H\right\rangle _{USE}\approx\ln N-1+\gamma$ and $\left\langle H\right\rangle _{OSE}\approx\ln N-2+\gamma+\ln2\text{
1193: ,}$ where $\gamma\approx0.577$ stands for the Euler constant. Note that both averages
1194: are close to the maximal value $\ln N$, attained for orthostochastic matrices
1195: corresponding to $B_{\ast}$, and their difference $\left\langle H\right\rangle
1196: _{USE}-\left\langle H\right\rangle _{OSE}$ converges to $1-\ln2\approx0.307$.
1197:
1198: \subsection{Average traces}
1199:
1200: Traces of consecutive powers of bistochastic matrices, ${\rm tr}B^n$, are
1201: quantities important in applications related to {\sl quantum chaos} \cite{KS97}.
1202: In the following we evaluate the traces
1203: $t_{N,n}:= \langle {\rm tr}B^n \rangle_{\rm USE}$, averaged over the
1204: unistochastic ensemble.
1205:
1206: The spectrum of a bistochastic matrix size $2$ is real and may be written as
1207: $\operatorname*{Sp}(B)=\{1,y\}$ with $y \in [-1,1]$. For the unistochastic
1208: ensemble $y=\cos 2\theta$, where the angle $\theta$ is distributed uniformly,
1209: $P(\theta )= 2/ \pi$ for $ \theta \in [ 0, \pi /2]$. The average traces may be
1210: readily expressed by the moments
1211: $\langle y^{k}\rangle$ for $k \in \mathbb{N}$, namely,
1212:
1213: \begin{eqnarray}
1214: t_{2,2k+1} &=&1+\langle y^{2k+1}\rangle =1 \text{ ,}
1215: \nonumber \\
1216: t_{2,2k} &=&
1217: 1+\langle y^{2k}\rangle =\frac{2k+2}{2k+1} \text{ .}
1218: \label{tmk}
1219: \end{eqnarray}
1220:
1221: Since the average size of all the diagonal elements must be equal,
1222: $\langle |U_{ii}|^{2}\rangle=1/N$, so the average trace $ t_{N,1} =\langle
1223: \sum_{i=1}^{N}|U_{ii}|^{2}\rangle =1$ and does not depend on the matrix size.
1224: Using the results of Mello \cite{Me90}, who computed several averages over the
1225: Haar measure on the unitary group $U(N)$, we derive in Appendix
1226: \ref{sec:traces} the following formulae
1227:
1228: \begin{equation}
1229: t_{N,2} = 1+\frac{1}{N+1},
1230: ~~~~~~ {\rm and } ~~~~~~
1231: t_{N,3} =
1232: \left\{
1233: \begin{array}{ccc}
1234: 1 ~~ & ~~{\rm for}~~ & ~~ N=2 \ , \\
1235: 1+\frac{2}{N^{2}+3N+2} ~~ & ~~{\rm for} ~~ &~~ N \geq 3\ .
1236: \end{array}
1237: \right.
1238: \label{tN23}
1239: \end{equation}
1240: Analogous expressions for larger $N$ may be explicitly written down as functions
1241: of the Mello averages, but it is not simple to put them into a transparent
1242: form. Numerical results support a conjecture that for arbitrary $N$ the
1243: average traces tend fast to unity and the difference $t_{N,n}-1$ behaves as
1244: $N^{1-n}$. This fact is related to the properties of the spectra discussed
1245: above: for large $N$ the spectrum is concentrated close to the center of the
1246: unit circle, so the contribution of all the subleading eigenvalues to the
1247: traces becomes negligible. Related results on average traces of the symmetric
1248: unistochastic matrices $BB^T$ are provided in \cite{Be01}.
1249:
1250: \bigskip
1251:
1252: \section{Closing Remarks}
1253:
1254:
1255: In this paper we define the entropy for bistochastic matrices and proved its subadditivity. We also analysed special classes of bistochastic matrices: the unistochastic and orthostochastic matrices, and introduced probability measures on these sets. We found a characterization of complex spectra of unistochastic matrices in the unit circle and discussed the size of their subleading eigenvalue, which determines the speed of the decay of correlation in the dynamical system described by the matrices under consideration. Unistochastic matrices find diverse applications in different branches of physics and it is legitimate to ask, how a physical system behaves, if it is described by a random
1256: unistochastic matrix \cite{Be01}. Since we succeeded in computation of mean values of some quantities (traces, entropy) averaged over the ensemble of
1257: unistochastic matrices, our results provide some information concerning this issue.
1258:
1259: On the other hand, several problems concerning bi--, uni--, and (ortho--)stochastic
1260: matrices remain open and we conclude this paper listing some of them:
1261:
1262: a) Prove the conjecture that the union of the spectra of bistochastic matrices
1263: $\Sigma_N^B$ is equal to the sum of regular polygons $E_N$, i.e.,
1264: $\Sigma_N^B= E_N $.
1265:
1266: b) Prove that the support $\Sigma_3^U$ of spectra of unistochastic matrices of size
1267: $N=3$ contains exactly the interval $[-1,1]$ and the set bounded by the
1268: $3$--hypocycloid $H_3$, i.e., $\Sigma_3^U=G_3$.
1269:
1270: c) Find an analytical form of the boundaries of the set ${\tilde G}_N$ obtained of
1271: interpolations between the neighbouring corners of $k$--hypocycloids with
1272: $k\le N$.
1273:
1274: d) Show that the support $\Sigma_N^U$ of the spectra of unistochastic matrices contains
1275: ${\tilde G}_N$ and check whether both sets are equal, i.e., $\Sigma_N^U =
1276: {\tilde G}_N$.
1277:
1278: e) Show that the supports of spectra of uni-- and (ortho--)stochastic matrices
1279: are the same, i.e., $\Sigma_N^U=\Sigma_N^O$.
1280:
1281: f) Calculate probability distribution $P_N(z)$ of complex
1282: eigenvalues ensembles for uni--(ortho--)stochastic matrices.
1283:
1284: g) Compute the expectation value of the subleading
1285: eigenvalue $\langle r_2\rangle$, averaged over uni--(ortho--)stochastic
1286: ensemble.
1287:
1288: h) Calculate averages over the unistochastic ensemble of other quantities
1289: characterizing ergodicity of stochastic matrices, including ergodicity
1290: coefficients analysed by Seneta \cite{Se93} and entropy contraction
1291: coefficient introduced by Cohen et al. \cite{CIRRSZ93}.
1292:
1293: i) Find necessary and sufficient conditions for a bistochastic matrix to be unistochastic.
1294:
1295: j) Provide a full characterization of the set of all unitary matrices $U$,
1296: which lead to the same unistochastic matrix $B$, i.e., $B_{ij}=|U_{ij}|^2$
1297: for $i,j=1,\dots,k$.
1298:
1299: \bigskip
1300:
1301: This paper is devoted to the memory of late Marcin Po{\'z}niak,
1302: with whom we enjoyed numerous fruitful discussions
1303: on the properties of bistochastic matrices several years ago.
1304: We are thankful to Prot Pako{\'n}ski for a fruitful interaction
1305: and also acknowledge helpful remarks of I.~Bengtsson,
1306: G.~Berkolaiko, G.~Tanner, and M.~Wojtkowski.
1307: Financial support by Komitet Bada{\'n}
1308: Naukowych under the grant 2P03B-072~19 and
1309: the Sonderforschungsbereich `Unordung und grosse Fluktuationen'
1310: der Deutschen Forschungsgemeinschaft is gratefully acknowledged.
1311:
1312: \appendix
1313: \section{Unistochastic matrices stemming from a given bistochastic matrix}
1314: \label{sec:unibist}
1315:
1316: Let us consider two unitary $N\times N$ matrices $U$ and $W$ such that for all
1317: $i,j=1,\ldots,N$,
1318:
1319: \begin{equation}\label{uw}
1320: |U_{ij}|^2=|W_{ij}|^2 \text{ ,}
1321: \end{equation}
1322: i.e., that the corresponding unistochastic matrices are the same. It is obvious that this happens if $W=V_1UV_2$ with $V_1$ and $V_2$ unitary diagonal. However the converse statement, i.e., that (\ref{uw}) implies existence of two diagonal unitary matrices $V_1$ and $V_2$ such that $U=V_1WV_2$, is false \cite{false}. The plausibility of such conjecture is based on the following dimensional argument: we have $N^2$ real numbers $u_{ij}:=|U_{ij}|$ fulfilling $2N-1$ relations stemming from the normalisation of the rows:
1323:
1324: \begin{equation} \sum_{j=1}^N u_{ij}^2=1, \quad i=1,\ldots,N \text{ ,}
1325: \label{rows}
1326: \end{equation}
1327: and the columns
1328:
1329: \begin{equation}
1330: \sum_{i=1}^N u_{ij}^2=1, \quad j=1,\ldots,N
1331: \label{cols}
1332: \end{equation}
1333: of the unitary matrix $U$. The number of independent relations is less by one than the total number of equations in (\ref{rows}) and (\ref{cols}) since summing all equations in (\ref{rows}) over $i$ gives the same as summing all
1334: equations in (\ref{cols}) over $j$, namely $N=\rm{Tr}(U^{\dagger}U)$.
1335: On the other hand the left and right multiplications by unitary
1336: diagonal $V_1$ and $V_2$ introduce exactly $2N-1$ parameters (here the number
1337: of the independent parameters is diminished by one from the number of
1338: non-zero elements of both $D_1$ and $D_2$ since in the resulting matrix only
1339: the differences of eigenphases of $D_1$ and $D_2$ appear, so we can always
1340: put one of the eigenphases of $D_1$, say, to zero without changing the result
1341: of the transformation $W\mapsto V_1WV_2$). The simplest counterexample we know involves the following unitary matrices $U$ and $W$
1342:
1343: \begin{equation}
1344: U = \frac{1}{2} \left[
1345: \begin{array}{rrrr}
1346: 1 & 1 & 1 & 1 \\
1347: 1 & -1 & -e^{i\alpha} & e^{i\alpha} \\
1348: 1 & -1 & e^{i\alpha} & -e^{i\alpha} \\
1349: 1 & 1 & -1 & -1
1350: \end{array}
1351: \right], \quad
1352: W = \frac{1}{2} \left[
1353: \begin{array}{rrrr}
1354: 1 & 1 & 1 & 1 \\
1355: 1 & -1 & -1 & 1 \\
1356: 1 & -e^{i\beta} & e^{i\beta} & -1 \\
1357: 1 & e^{i\beta} & -e^{i\beta} & -1
1358: \end{array}
1359: \right] \text{ ,}
1360: \end{equation}
1361: with $|U_{ij}|^2=|W_{ij}|^2=1/4$. It is a matter of simple explicit
1362: calculations to show that there are no unitary diagonal $V_1$
1363: and $V_2$ fulfilling $W=V_1 U V_2$,
1364: if $\alpha, \beta \in [0,2\pi]$ and $\alpha\beta \ne 0$.
1365:
1366: \section{Orthostochastic matrices with spectrum at hypocycloids}
1367: \label{sec:hypoc}
1368:
1369: In this appendix we construct orthostochastic matrices with spectra on
1370: $N$-hypocycloids. Consider the following $N\times N$ permutation matrix
1371:
1372: \begin{equation}
1373: P:=\left[
1374: \begin{array}{cccccc}
1375: 0 & 1 & 0 & \ldots & 0 & 0 \\
1376: 0 & 0 & 1 & \ldots & 0 & 0 \\
1377: \vdots & \vdots & \vdots & & \vdots & \vdots \\
1378: 0 & 0 & 0 & \ldots & 0 & 1 \\
1379: 1 & 0 & 0 & \ldots & 0 & 0 \\
1380: \end{array}
1381: \right]
1382: \text{ ,}
1383: \label{P}
1384: \end{equation}
1385: where for simplicity the dimensionality index $N$ has been omitted.
1386: We have $P^N = \mathbb{I}$, $P^K \ne \mathbb{I}$ for $K<N$ and the
1387: eigenvalues of $P$ equal $\exp\left(\frac{2\pi i}{N}k\right)$,
1388: $k=0,1,\ldots,N-1$.
1389:
1390: We shall discuss separately the cases of odd and even $N$.
1391:
1392: First let $N=2K+1$ and
1393:
1394: \begin{equation}\label{OB}
1395: O:=\sum_{j=0}^{N-1}a_jP^j,\quad B:=\sum_{j=0}^{N-1}a_j^2P^j .
1396: \end{equation}
1397: Observe that, since $P^m$ and $P^n$ do not have common non-zero entries for $m\ne n$, $0\le m,n\le N-1$, the elements of $B$ are squares of the corresponding elements of $O$. The eigenvalues of $O$ and $B$ read, respectively,
1398:
1399: \begin{eqnarray}
1400: \Lambda_k&=&\sum_{j=0}^{N-1}a_j\exp\left(\frac{2\pi i}{N}kj\right),\quad
1401: k=0,1,\ldots,N-1 \label{eigO} \\
1402: \lambda_k&=&\sum_{j=0}^{N-1}a_j^2\exp\left(\frac{2\pi i}{N}kj\right),\quad
1403: k=0,1,\ldots,N-1.
1404: \label{eigB}
1405: \end{eqnarray}
1406: Inverting the discrete Fourier transforms in (\ref{eigO}) we obtain
1407: \begin{equation}\label{aj2}
1408: a_j=\frac{1}{N}\sum_{k=0}^{N-1}\Lambda_k\exp\left(-\frac{2\pi
1409: i}{N}kj\right),\quad
1410: k=0,1,\ldots,N-1,
1411: \end{equation}
1412: which, upon substituting to (\ref{eigB}), gives the eigenvalues of $B$ in terms of the eigenvalues of $O$
1413:
1414: \begin{eqnarray}\label{main1}
1415: \lambda_k=\frac{1}{N^2}\sum_{j=0}^{N-1}\sum_{l=0}^{N-1}\sum_{r=0}^{N-1}\Lambda_l
1416: \Lambda_r\exp\left(-\frac{2\pi i}{N}(l+r-k)j\right)=
1417: \frac{1}{N}\sum_{l=0}^{N-1}\sum_{r=0}^{N-1}\Lambda_l\Lambda_r\delta_{r,k-l}=
1418: \frac{1}{N}\sum_{l=0}^{N-1}\Lambda_l\Lambda_{k-l},
1419: \end{eqnarray}
1420: where the indices are counted modulo $N$.
1421:
1422: Our aim is now to find a family of orthogonal matrices $O(\phi)$ such
1423: that when $\phi$ changes (from $0$ to $2\pi$, say) the eigenvalue
1424: $\lambda_0(\phi)$ renders the $N$-hypocycloid $H_N$ in the complex plane, i.e.
1425:
1426: \begin{equation}\label{lambad0}
1427: \lambda_0(\phi)=\frac{1}{N}\left[(N-1)e^{i\phi}+e^{-i(N-1)\phi}\right] .
1428: \end{equation}
1429: First observe that the desired result is achieved if
1430:
1431: \begin{equation}\label{equi}
1432: \Lambda_k=\exp[i(k-K)\phi], \quad k=0,1,\ldots,2K=N-1.
1433: \end{equation}
1434: Indeed
1435:
1436: \begin{eqnarray}
1437: \lambda_0&=&\frac{1}{N}\sum_{l=0}^{N-1}\Lambda_l\Lambda_{-l}
1438: =\frac{1}{N}\left(\Lambda_0^2+\sum_{l=1}^{N-1}\Lambda_l\Lambda_{-l}\right)
1439: =\frac{1}{N}\left(\Lambda_0^2+\sum_{l=1}^{N-1}\Lambda_l\Lambda_{N-l}\right)
1440: \nonumber \\
1441: &=&\frac{1}{N}\left(e^{-2iK\phi}+\sum_{l=1}^{N-1}e^{i(l-K)\phi}e^{(N-l-K)\phi}\right)
1442: =\frac{1}{N}\left(e^{-i(N-1)\phi}+(N-1)e^{i(N-2K)\phi}\right)
1443: \\
1444: &=&\frac{1}{N}\left[(N-1)e^{i\phi}+e^{-i(N-1)\phi}\right].
1445: \end{eqnarray}
1446:
1447: In order to construct an orthogonal matrix $O$ with the spectrum (\ref{equi}) it is enough to find an antisymmetric $A$ with the eigenvalues
1448:
1449: \begin{equation}\label{equia}
1450: \mu_k=i(k-K),\quad k=0,1,\ldots,2K.
1451: \end{equation}
1452: If $A$ is a polynomial in $P$ then so is $O=\exp(A\phi)$, moreover $O$ is orthogonal and has the desired spectrum (\ref{equi}). Let us thus write
1453:
1454: \begin{equation}\label{A}
1455: A:=\sum_{j=1}^K \alpha_j\left(P^j-P^{N-j}\right),
1456: \end{equation}
1457: which is clearly antisymmetric, with the eigenvalues
1458:
1459: \begin{equation}\label{eigA}
1460: \mu_k=\sum_{j=1}^K \alpha_j\left[\exp\left(\frac{2\pi i}{N}jk\right)
1461: -\exp\left(\frac{2\pi i}{N}(N-j)k\right)\right]=
1462: 2i\sum_{j=1}^K \alpha_j\sin\left(\frac{2\pi}{2K+1}kj\right),
1463: \quad k=0,1,\ldots,2K.
1464: \end{equation}
1465: Obviously $\mu_0=0$ and $\mu_{2K+1-k}=\mu_k$, $k=1,2,\ldots,K$. To fulfil (\ref{equia}) it is thus enough that
1466:
1467: \begin{equation}\label{rnie}
1468: 2\sum_{j=1}^K \alpha_j\sin\left(\frac{2\pi}{2K+1}kj\right)=k,\quad
1469: k=1,2,\ldots,K.
1470: \end{equation}
1471: Using
1472:
1473: \begin{equation}\label{ortsin}
1474: \sum_{j=1}^K \sin\left(\frac{2\pi}{2K+1}kj\right)
1475: \sin\left(\frac{2\pi}{2K+1}mj\right)
1476: =\frac{2K+1}{4}\,\delta_{mk},
1477: \end{equation}
1478: we solve (\ref{rnie}) for $\alpha_j$,
1479:
1480: \begin{equation}
1481: \alpha_j=\frac{2}{2K+1}\sum_{k=1}^Kk\sin\left(\frac{2\pi}{2K+1}kj\right) .
1482: \label{aj2k1}
1483: \end{equation}
1484:
1485: For $N$ even, $N=2K$, the construction is very similar. We introduce $E:=\text{diag}(1,1,\ldots,1,-1)$ and define $\tilde{P}:=EP$ which has eigenvalues $\exp\left(\frac{2\pi i}{N}(k+\frac{1}{2})\right)$, $k=0,1,\ldots,N-1$, and construct the matrix $O$ as a polynomial in $\tilde{P}$ rather than in $P$
1486:
1487: \begin{equation}\label{Otilde}
1488: \tilde{O}:=\sum_{j=0}^{N-1}a_j\tilde{P}^j ,
1489: \end{equation}
1490: with the eigenvalues:
1491:
1492: \begin{equation}\label{Lambdatilde}
1493: \tilde\Lambda_k=\sum_{j=0}^{N-1}a_j\exp\left(\frac{2\pi i}{N}
1494: \left(k+\frac{1}{2}\right)j\right),\quad k=0,1,\ldots,N-1
1495: \end{equation}
1496: The matrix $B$ of the squared elements of $O$ is, as previously, the following polynomial in $P$,
1497:
1498: \begin{equation}\label{Btilde}
1499: B:=\sum_{j=0}^{N-1}a_j^2P^j .
1500: \end{equation}
1501: Using exactly the same method as above we express the eigenvalues $\lambda_k$ of $B$ are given in terms of $\Lambda_j$
1502:
1503: \begin{equation}\label{main2}
1504: \lambda_k=\frac{1}{N}\sum_{l=0}^{N-1}\tilde\Lambda_l\tilde\Lambda_{k-l+1}.
1505: \end{equation}
1506: Now if
1507:
1508: \begin{equation}\label{equiA2}
1509: \tilde\Lambda_k=\exp\left[i\left(k-K+\frac{1}{2}\right)\phi\right],
1510: \quad k=0,1,\ldots,2K-1=N-1,
1511: \end{equation}
1512: then
1513:
1514: \begin{eqnarray}\label{hypoeven}
1515: \lambda_{N-1}&=&\frac{1}{N}\sum_{l=0}^{N-1}\tilde\Lambda_l\tilde\Lambda_{N-l}
1516: =\frac{1}{N}\left(\tilde\Lambda_0^2+\sum_{l=1}^{N-1}
1517: \tilde\Lambda_l\tilde\Lambda_{N-l}\right)
1518: =\frac{1}{N}\left(e^{-i(2K-1)\phi}
1519: +\sum_{l=1}^{N-1}e^{i(l-K+1/2)\phi}e^{(N-l-K+1/2)\phi}\right)
1520: \nonumber \\
1521: &=&\frac{1}{N}\left(e^{-i(N-1)\phi}+(N-1)e^{i(N-2K+1)\phi}\right)
1522: =\frac{1}{N}\left[(N-1)e^{i\phi}+e^{-i(N-1)\phi}\right].
1523: \end{eqnarray}
1524: In full analogy with the case of odd $N$ we look for an antisymmetric matrix $\tilde{A}$ with the eigenvalues
1525:
1526: \begin{equation}\label{eigA2}
1527: \tilde\mu_k=i\left(k-K+\frac{1}{2}\right),
1528: \end{equation}
1529: in the form of a polynomial in $\tilde{P}$
1530:
1531: \begin{equation}\label{Atilde}
1532: \tilde{A}
1533: :=2^{1/2}\tilde\alpha_K\tilde P^K+\sum_{j=1}^{K}
1534: \tilde\alpha_j\left(\tilde{P}^j+\tilde{P}^{N-j}\right).
1535: \end{equation}
1536: Since, as it is easy to check, $(\tilde{P}^j)^T=-\tilde{P}^{N-j}$ for $j=1,2,\ldots,K$, the matrix $A$ is indeed antisymmetric for arbitrary real $\alpha_j$, $j=1,2,\ldots,K$. The eigenvalues of $A$ read:
1537:
1538: \begin{equation}\label{eigAtilde}
1539: \tilde{\mu}_k=2^{1/2}i(-1)^k\alpha_K+
1540: 2i\sum_{j=1}^{K-1}\tilde\alpha_j
1541: \sin\left[\frac{\pi}{K}\left(k+\frac{1}{2}\right)j\right],\quad
1542: k=0,1,\ldots,2K-1.
1543: \end{equation}
1544: Using arguments similar to those in the odd $N$ case, we conclude that the choice
1545:
1546: \begin{eqnarray}\label{altilde}
1547: \tilde\alpha_j&=&\frac{1}{K}\sum_{k=0}^{K-1}
1548: \sin\left[\frac{\pi}{K}\left(k+\frac{1}{2}\right)j\right]\left(k-K+\frac{1}{2}\right),
1549: \quad j=1,2,\ldots,K-1 \\
1550: \tilde\alpha_K&=&-\frac{\sqrt{2}}{4},
1551: \end{eqnarray}
1552: leads to the desired result (\ref{eigA2}) and, consequently, the $2K$-hypocycloid (\ref{hypoeven}).
1553:
1554: \section{Unistochastic matrices with spectrum at hypocycloids}
1555: \label{sec:hypocun}
1556:
1557: The above constructed matrices with spectra on $N$-hypocycloids were {\em orthostochastic}.
1558: If the desired matrix should be merely {\em unistochastic}, but not
1559: necessarily orthostochastic, the construction is even simpler. To this end let
1560: us consider the matrix
1561:
1562: \begin{equation}\label{Palpha}
1563: P^\alpha:=U^\dagger D^\alpha U,
1564: \end{equation}
1565: where $U$ is an unitary matrix diagonalizing $P$, where $P$ is given by (\ref{P})), and $D$ is a diagonal matrix
1566:
1567: \begin{equation}\label{D}
1568: D:=\mathrm{diag}\left(1,e^{2\pi i/N}, e^{4\pi i/N},\ldots,e^{2(N-1)\pi
1569: i/N}\right)
1570: \end{equation}
1571: with the eigenvalues of $P$ on the main diagonal. Hence, consequently
1572:
1573: \begin{equation}\label{Dalpha}
1574: D^\alpha=\mathrm{diag}(\Lambda_0,\Lambda_1,\ldots,\Lambda_{N-1}),
1575: \end{equation}
1576: where $\Lambda_k:=\exp(2k\alpha\pi i/N)$, $k=0,1,\ldots,N-1$, are the eigenvalues of $P^\alpha$. Obviously $P^\alpha$ is unitary, and as a function of $P$ can be written in the form $P^\alpha:=\sum_{j=0}^{N-1}a_jP^j$, which gives for the eigenvalues
1577:
1578: \begin{equation}\label{}
1579: \Lambda_k=e^{2k\alpha\pi i/N}=\sum_{j=0}^{N-1}a_j\exp\left(\frac{2\pi
1580: i}{N}kj\right),\quad
1581: k=0,1,\ldots,N-1.
1582: \end{equation}
1583: As previously we obtain the coefficients $a_j$ by inverting the discrete Fourier transform
1584:
1585: \begin{equation}\label{aju}
1586: a_j=\frac{1}{N}\sum_{k=0}^{N-1}\Lambda_k\exp\left(-\frac{2\pi i}{N}kj\right),\quad
1587: k=0,1,\ldots,N-1.
1588: \end{equation}
1589: The associated unistochastic matrix reads thus
1590:
1591: \begin{equation}\label{Balpha}
1592: B:=\sum_{j=0}^{N-1}|a_j|^2P^j.
1593: \end{equation}
1594: and has as the eigenvalues
1595: \begin{equation}\label{lambdaku}
1596: \lambda_k=\sum_{j=0}^{N-1}|a_j|^2\exp\left(\frac{2\pi i}{N}kj\right),\quad
1597: k=0,1,\ldots,N-1,
1598: \end{equation}
1599: i.e.
1600:
1601: \begin{eqnarray}\label{main1u}
1602: \lambda_k=\frac{1}{N^2}\sum_{j=0}^{N-1}\sum_{l=0}^{N-1}\sum_{r=0}^{N-1}\Lambda_l
1603: \Lambda_r^\ast\exp\left(-\frac{2\pi i}{N}(l+r-k)j\right)=
1604: \frac{1}{N}\sum_{l=0}^{N-1}\sum_{r=0}^{N-1}\Lambda_l\Lambda_r^\ast\delta_{r,l-k}=
1605: \frac{1}{N}\sum_{l=0}^{N-1}\Lambda_l\Lambda_{l-k}^\ast,
1606: \end{eqnarray}
1607: hence
1608:
1609: \begin{eqnarray}\label{lambda1u}
1610: \lambda_1=\frac{1}{N}\left(\Lambda_0\Lambda_{N-1}^\ast +
1611: \sum_{l=1}^{N-1}\Lambda_l\Lambda_{l-k}^\ast\right)=\frac{1}{N}\left(e^{-2(N-1)\alpha\pi
1612: i/N}+(N-1)e^{2\alpha\pi i/N}\right),
1613: \end{eqnarray}
1614: which renders the $N$-hypocycloid for $\alpha\in[0,N/2]$. Similarly,
1615: the further eigenvalues $\lambda_k$ generate the inner hypocycloids
1616: (e.g. 3-hypocycloid for $N=6$ -- see Fig. 5), what
1617: proves Proposition 3.
1618:
1619: Observe that for $N=3$ the permutation matrices
1620: $P,P^2$ and $P^3={\mathbb I}_3$ form an equilateral
1621: triangle (in sense of the Hilbert--Schmidt distance,
1622: which is induced by the Frobenius norm of a matrix,
1623: $||P||:=\sqrt{PP^{\dagger}}$). Comparing Eq. (C6) and (C7) with $k=1$ we see
1624: that both quantities have the same structure and the same dependence
1625: on the coefficients $a_j$, which are implicit functions of the
1626: parameter $\alpha$. Therefore, varying this parameter
1627: we obtain the very same curves in two entirely different spaces:
1628: the eigenvalue $\lambda_1=\lambda_1(\alpha)$
1629: provides the $3$--hypocycloid in the plane of complex spectra,
1630: while the family of unistochastic matrices $B=B(\alpha)$
1631: forms the same hypocycloid in the two--dimensional cross-section of the
1632: four--dimensional body of $N=3$ bistochastic matrices determined by
1633: $P,P^2$ and $P^3$.
1634:
1635: \section{Interpolation between corners of two hypocycloids}
1636: \label{sec:hypoc2}
1637:
1638: In this appendix we provide a discuss a family of unistochastic
1639: matrices of size $4$, the spectra of which are not contained in the sum of
1640: $2$, $3$ and $4$--hypocycloids. To find an interpolation between the corners
1641: of $3$ and $4$--hypocycloids consider the orthogonal matrix
1642: $\tilde{O}_{3,4}(\varphi)=O_4(\varphi)P_{1234}$, as defined in section
1643: \ref{secIV},
1644:
1645: \begin{equation}
1646: \tilde{O}_{3,4}(\varphi) :=
1647: \left[
1648: \begin{array}{cccc}
1649: 1 & 0 & 0 & 0 \\
1650: 0 & 1 & 0 & 0 \\
1651: 0 & 0 & \cos \varphi & \sin \varphi \\
1652: 0 & 0 & -\sin \varphi & \cos \varphi
1653: \end{array}
1654: \right] \left[
1655: \begin{array}{cccc}
1656: 0 & 1 & 0 & 0 \\
1657: 0 & 0 & 1 & 0 \\
1658: 0 & 0 & 0 & 1 \\
1659: 1 & 0 & 0 & 0
1660: \end{array}
1661: \right]
1662: \allowbreak =\allowbreak
1663: \left[
1664: \begin{array}{cccc}
1665: 0 & 1 & 0 & 0 \\
1666: 0 & 0 & \cos \varphi & \sin \varphi \\
1667: 0 & 0 & -\sin \varphi & \cos \varphi \\
1668: 1 & 0 & 0 & 0
1669: \end{array}
1670: \right]
1671: \label{inter34a}
1672: \end{equation}
1673: For $\varphi$ varying in $[0,\pi/2]$ this family interpolates between
1674: a four--elements permutation $P_{1234}$ and a matrix, the absolute values of
1675: which represent a three--elements permutation $P_{124,3}$. Thus the spectra of
1676: the corresponding bistochastic matrices give an interpolation between the
1677: third and the fourth roots of identity. As shown in Fig.~4 this interpolation
1678: is located {\sl outside} the hypocycloids $H_3$ and $H_4$, so the support
1679: $\Sigma_4^U$ is larger than their sum $G_4$. Repeating the argument with the
1680: multiplicative interpolation between any such a matrix and the Fourier matrix
1681: $F^{(4)}$ we conclude that all points {\sl inside} the set bounded by this
1682: interpolation belong to the support $\Sigma_4^U$. An analogous scheme allows
1683: us to find an $(N-1)\longleftrightarrow N$ interpolation. For example, the
1684: family of orthogonal matrices $\tilde{O}_{4,5}(\phi)=O_5(\phi)P_{12345}$ of
1685: size $5$ gives an interpolation between $1^{1/5}$ and $i=1^{1/4}$.
1686:
1687: To find the missing $N=4$ interpolation for the negative real part of the
1688: eigenvalues consider a permutation of the orthogonal matrix which contains
1689: the block $O_3$ responsible for the $3$--hypocycloid,
1690:
1691: \begin{equation}
1692: \tilde{O}_{3,2}:=
1693: \left[
1694: \begin{array}{cccc}
1695: 1 & 0 & 0 & 0 \\
1696: 0 & a & b & c \\
1697: 0 & c & a & b \\
1698: 0 & b & c & a
1699: \end{array}
1700: \right] \left[
1701: \begin{array}{cccc}
1702: 0 & 1 & 0 & 0 \\
1703: 1 & 0 & 0 & 0 \\
1704: 0 & 0 & 0 & 1 \\
1705: 0 & 0 & 1 & 0
1706: \end{array}
1707: \right]
1708: \allowbreak =\allowbreak
1709: \left[
1710: \begin{array}{cccc}
1711: 0 & 1 & 0 & 0 \\
1712: a & 0 & c & b \\
1713: c & 0 & b & a \\
1714: b & 0 & a & c
1715: \end{array}
1716: \right],
1717: \label{inter34b}
1718: \end{equation}
1719: where its elements $a$, $b$, and $c$ are function of the angle $\varphi$ as given in (\ref{hyper3bi}). Then the spectra of the corresponding orthostochastic matrices, obtained by squaring the elements of the above orthogonal matrices,
1720:
1721: \begin{equation}
1722: \tilde{B}_{3,4}(\varphi):=
1723: \left[
1724: \begin{array}{cccc}
1725: 0 & 1 & 0 & 0 \\
1726: 0 & 0 & \cos ^{2}\varphi & \sin ^{2}\varphi \\
1727: 0 & 0 & \sin ^{2} \varphi & \cos ^{2}\varphi \\
1728: 1 & 0 & 0 & 0
1729: \end{array}
1730: \right] \quad {\rm and} \quad
1731: \tilde{B}_{3,2}(\varphi):=
1732: \left[
1733: \begin{array}{cccc}
1734: 0 & 1 & 0 & 0 \\
1735: a^{2} & 0 & c^{2} & b^{2} \\
1736: c^{2} & 0 & b^{2} & a^{2} \\
1737: b^{2} & 0 & a^{2} & c^{2}
1738: \end{array}
1739: \right]
1740: \label{inter34c}
1741: \end{equation}
1742: provide the required interpolations located outside hypocycloids
1743: $H_3$ and $H_4$ (see Fig.~4b~and~4d). For larger dimensionality an analogous
1744: construction has to be performed to get the interpolations between the
1745: neighbouring roots of identity.
1746:
1747: \section{Average traces of unistochastic matrices}
1748: \label{sec:traces}
1749:
1750: To evaluate the average traces we rely on the results of Mello \cite{Me90}, who computed various averages, $\langle . \rangle$, over the Haar measure on the unitary group $U(N)$. In particular, he found average values of the following quantities constructed of elements $U_{ab}$ of a unitary matrix of size $N$
1751:
1752: \begin{equation}
1753: Q_{b_1 \beta_1, \cdots ,b_m\beta_m}^{a_1 \alpha_1, \cdots ,a_k\alpha_k}
1754: :=
1755: \langle (U_{b_1\beta_1}\cdots U_{b_m\beta_m})
1756: (U_{a_1\alpha_1}\cdots U_{a_k\alpha_k})^{\ast }\rangle.
1757: \label{Mell1}
1758: \end{equation}
1759: \bigskip
1760: Mean trace of a squared unistochastic matrix, defined by $B_{ij}=|U_{ij}|^2$, reads
1761:
1762: \begin{equation}
1763: t_{N,2}:=\langle {\rm Tr} B^{2}\rangle_{USE}
1764: =
1765: \langle \sum_{i,k=1}^{N}(U_{ik}U_{ik}^{\ast })(U_{ki}U_{ki}^{\ast
1766: }) \rangle_{U(N)}
1767: =\sum_{i,k=1}^{N}Q_{ik,ki}^{ik,ki} =
1768: NQ_{11,11}^{11,11}+N(N-1)Q_{12,21}^{12,21},
1769: \end{equation}
1770: since the symmetry of the problem allowed us to group together the terms according to the number of different indices. Using the results of Mello
1771: $Q_{11,11}^{11,11}=2/(N(N+1))$ and $Q_{12,21}^{12,21}=1/(N(N+1))$ we get $t_{N,2}=(N+2)/(N+1)$.
1772:
1773: The mean trace of $B^3$ reads
1774:
1775: \begin{equation}
1776: t_{N,3} :=\langle {\rm Tr} B^{3}\rangle_{USE}=
1777: NQ_{11,11,11}^{11,11,11}+3N(N-1)Q_{11,12,21}^{11,12,21}+
1778: N(N-1)(N-2)Q_{12,23,31}^{12,23,31}.
1779: \label{an3}
1780: \end{equation}
1781:
1782: The data in Mello's paper allow us to find
1783: $Q_{11,11,11}^{11,11,11}=6/[N(N+1)(N+2)]$,
1784: $Q_{11,12,21}^{11,12,21}=1/[(N+2)(N^2-1)]$,
1785: and
1786: $Q_{12,23,31}^{12,23,31}= (N^2-2)/[(N(N^2-1)(N^2-4)]$,
1787: where the last term (with three different indices) is present only for $N\ge 3$.
1788: Substituting these averages into (\ref{an3}) we arrive with the result (\ref{tN23}).
1789:
1790: In the general case of arbitrary $n$ we may write a formula
1791:
1792: \begin{equation}
1793: t_{N,n} :=\langle {\rm Tr} B^{n}\rangle_{USE}
1794: =\sum_{i_1,\dots i_n}^{N} Q^{i_1i_2,i_2i_3, \dots ,i_{n-1}i_n,i_n i_1}%
1795: _{i_1i_2,i_2i_3, \dots ,i_{n-1}i_n, i_n i_1},
1796: \label{annn}
1797: \end{equation}
1798: which is explicit, but not easy to simplify. An analogous computation for the ensemble of symmetric unistochastic matrices may be based on results of Brouwer and Beenakker \cite{BB96}, who computed the averages (\ref{Mell1}) for COE.
1799:
1800: \begin{references}
1801:
1802: \bibitem{CZ98} Cohen J E and Zbaganu Gh 1998
1803: {\sl Comparison of Stochastic Matrices}
1804: (Basel: Birkhauser)
1805:
1806: \bibitem{FW79} Fritz F-J, Huppert B and Willems W 1979
1807: {\sl Stochastische Matricen}
1808: (Berlin: Springer-Verlag)
1809:
1810: \bibitem{MO79} Marshall A W and Olkin I 1979
1811: {\sl Inequalities: Theory of Majorization and its Applications}
1812: (New York: Academic Press)
1813:
1814: \bibitem{AU82} Alberti P M and Uhlmann A 1982
1815: {\sl Stochasticity and Partial Order}
1816: (Berlin: DL Verlag Wiss.)
1817:
1818: \bibitem{An89} Ando T 1989 Linear Algebra Appl. {\bf 118} 163
1819:
1820: \bibitem{Bh97} Bhatia R 1997 {\sl Matrix Analysis}
1821: (New York: Springer-Verlag)
1822:
1823: \bibitem{Lo97} Louck J D 1997 {\sl Found. Phys.} {\bf 27} 1085
1824:
1825: \bibitem{St02} Stry{\l}a J 2002
1826: Stochastic quantum dynamics,
1827: {\sl arXiv preprint} quant-ph/0204161
1828:
1829:
1830: \bibitem{KS97} Kottos T and Smilansky U 1997
1831: {\sl Phys. Rev. Lett.} {\bf 79} 4794
1832:
1833: \bibitem{Ta00} Tanner G 2000
1834: {\sl J. Phys. A} {\bf 33} 3567
1835:
1836: \bibitem{PZK01} Pako{\'n}ski P, {\.Z}yczkowski K and Ku{\'s} M 2001
1837: {\sl J. Phys. A} {\bf 34} 9303
1838: % {\sl preprint} arXiv nlin.CD/0011050
1839:
1840: \bibitem{Ta01} Tanner G 2001
1841: {\sl J. Phys. A} {\bf 34} 8485
1842: % Lanl preprint nlin.CD/0104014
1843:
1844: \bibitem{PZ01} Pako{\'n}ski P, Tanner G and {\.Z}yczkowski K 2001
1845: Families of line-graphs and their quantization,
1846: {\sl arXiv preprint} nlin.CD/0110043
1847:
1848: \bibitem{Ni99b} Nielsen M A 2000
1849: % Probability distributions consistent with a mixed state
1850: {\sl Phys. Rev. A} {\bf 62} 052308
1851: % LANL preprint quant-ph/9909020 % major
1852:
1853: \bibitem{Ni00} Nielsen M A 2001
1854: % Characterizing mixing and measurement in quantum mechanics
1855: {\sl Phys. Rev. A} {\bf 63} 022114
1856: % LANL preprint quant-ph/0008073 % nowy
1857:
1858: \bibitem{Ni99} Nielsen M A 1999
1859: {\sl Phys. Rev. Lett.} {\bf 83} 436
1860: % entangl. transf and majoriz.
1861:
1862: \bibitem{HVC02} Hammerer K, Vidal G and Cirac J I
1863: 2002 Characterization of non-local gates,
1864: {\sl arXiv preprint} quant-ph/0205100
1865:
1866: \bibitem{Me91} Mehta M L 1991 {\sl Random Matrices}, II ed.
1867: (New York: Academic)
1868:
1869: \bibitem{Be01} Berkolaiko G 2001 {\sl J. Phys. A} {\bf 34} L319
1870: % Lanl preprint nlin.CD/0104009
1871:
1872: \bibitem{FM60} Farahat H K and Mirsky L 1960
1873: {\sl Proc. Cambridge Philos. Soc.} {\bf 56} 322
1874:
1875: \bibitem{CR99} Chan C S and Robbins D P 1999
1876: % On the volume of the polytope of doubly stochastic matrices
1877: {\sl Exp. Math.} {\bf 8} 291
1878:
1879: \bibitem{Me89} Mehta M L 1989 {\sl Matrix Theory}, II ed.
1880: (Delhi: Hindustan Publishing Corporation)
1881:
1882: \bibitem{Eg81} Egorychev G P 1981
1883: {\sl Adv. Math.} {\bf 42} 299
1884:
1885: \bibitem{Sc76} Schaefer H H 1974 {\sl Banach Lattices and Positive
1886: Operators} (Berlin: Springer--Verlag)
1887:
1888: \bibitem{Ka59} Kac M 1959 {\sl Probability and Related Topics}
1889: (New York: Interscience)
1890:
1891: \bibitem{DD46} Dimitriev N and Dynkin E 1946
1892: {\sl Izvestia Acad. Nauk SSSR, Seria Mathem.} {\bf 10} 167
1893:
1894: \bibitem{Ka51} Karpelevich F I 1951
1895: {\sl Izvestia Acad. Nauk SSSR, Seria Mathem.} {\bf 15} 361
1896:
1897: \bibitem{Dj90} Djokovi{\v c} D {\v Z} 1990
1898: {\sl Linear Algebra Appl.} {\bf 142} 173
1899:
1900: \bibitem{It97} Ito H 1997 {\sl Linear Algebra Appl.} {\bf 267} 241
1901: % A new statement about the theorem determining the region of eigenvalues of
1902: % stochastic matrices
1903:
1904: \bibitem{Fi95} Fiedler M 1995 {\sl Linear Algebra Appl.} {\bf 214} 133
1905:
1906: \bibitem{Ho54} Horn A 1954 {\sl Amer. J. Math.} {\bf 76} 620
1907:
1908: \bibitem{BE02} Bengtsson I and Ericsson {\AA} 2002
1909: How to mix a density matrix,
1910: {\sl arXiv preprint} quant-ph/0206169
1911:
1912: \bibitem{AL87} Alicki R and Lendi K 1987
1913: {\sl Quantum Dynamical Semigroups and Applications},
1914: Lecture Notes in Physics vol. 286, (Berlin: Springer-Verlag)
1915:
1916: \bibitem{Ul71} Uhlmann A 1971
1917: {\sl Wiss. Z. Karl-Marx-Univ. Leipzig} {\bf 20} 633
1918:
1919: \bibitem{Sh52} Sherman S 1952 {\sl Proc. Amer. Math. Soc. } {\bf 3}, 511
1920:
1921: \bibitem{Sh54} Sherman S 1954 {\sl Proc. Amer. Math. Soc. } {\bf 5}, 998
1922:
1923: \bibitem{Sc58} Schreiber S 1958 {\sl Proc. Amer. Math. Soc. } {\bf 9}, 350
1924:
1925: \bibitem{Sl01} S{\l}omczy{\'n}ski W 2002
1926: {\sl Open Sys. \& Information Dyn.}, {\sl in press}
1927:
1928: \bibitem{Pa02} Pako{\'n}ski P 2002
1929: Ph.D thesis, Jagiellonian University, Cracow, {\sl (unpublished)}
1930:
1931: \bibitem{AYC91} Au-Yeung Y--H and Cheng C--M 1991
1932: {\sl Linear Algebra Appl.} {\bf 150} 243
1933:
1934: \bibitem{He78} Heinz T F 1978
1935: {\sl Linear Algebra Appl.} {\bf 20} 265
1936:
1937: \bibitem{Dj66} Djokovi{\v c} D {\v Z} 1966
1938: {\sl Amer. Math. Monthly} {\bf 73} 633
1939:
1940: \bibitem{PT87} Poon Y-T and Tsing N-K 1987
1941: Linear and Multilinear Algebra {\bf 21} 253
1942:
1943: \bibitem{false} Roos M, 1964 {\sl J. Math. Phys.} {\bf 5} 1609;
1944: ibid. 1965 {\bf 6} 1354
1945:
1946: \bibitem{Haake} Haake F 2000 {\sl Quantum Signatures of Chaos}, II ed.
1947: (Berlin: Springer-Verlag)
1948:
1949: \bibitem{Hu97} Hurwitz A 1897
1950: {\sl Nachr.\ Ges.\ Wiss.\ G\"ott.\ Math.--Phys.\ Kl.} 71-90
1951:
1952: \bibitem{ZK94} {\.Z}yczkowski K and Ku{\'s} M 1994
1953: {\sl J. Phys. A} {\bf 27} 4235
1954:
1955: \bibitem{PZK98} Po{\'z}niak M, {\.Z}yczkowski K, and Ku{\'s} M 1998
1956: {\sl J. Phys. A} {\bf 31} 1059
1957:
1958:
1959: \bibitem{Schuster} Schuster H G 1988 {\sl Deterministic Chaos}, II ed.
1960: (Weinheim: VCH Verlagsgeselschaft)
1961:
1962: \bibitem{HJ98} Horn R and Johnson C 1985 {\sl Matrix Analysis},
1963: (Cambridge: Cambridge University Press)
1964:
1965: \bibitem{LLR83} Leadbetter M R, Lindgren G and Rootzen H 1983
1966: {\sl Extremes and Related Properties of Random Sequences and Series}
1967: (New York: Springer-Verlag)
1968:
1969: \bibitem{Jo90} Jones K R W 1990 {\sl J.~Phys. A} {\bf 23} L1247
1970:
1971: \bibitem{Jo91} Jones K R W 1991 {\sl J.~Phys. A} {\bf 24} 1237
1972:
1973: \bibitem{Me90} Mello P A 1990 {\sl J. Phys. A} {\bf 23} 4061
1974:
1975: \bibitem{Se93} Seneta E 1993 {\sl Linear Algebra Appl.} {\bf 191} 245
1976:
1977: \bibitem{CIRRSZ93} Cohen J E, Iwasa Y, Rautu Gh, Ruskai~M~B, Seneta~E, and
1978: Zbaganu~Gh 1993 {\sl Linear Algebra Appl.} {\bf 179} 211
1979:
1980: \bibitem{BB96} Brouwer P W and Beenakker C W J 1996
1981: {\sl J. Math. Phys.} {\bf 37} 4904
1982:
1983: \end{references}
1984:
1985: \end{document}
1986:
1987: