Abstract
Motivated by quantum information theory, we introduce a dynamical random density matrix built out of the sum of \(k \ge 2\) independent unitary Brownian motions. In the large size limit, its spectral distribution equals, up to a normalising factor, that of the free Jacobi process associated with a single self-adjoint projection with trace 1/k. Using free stochastic calculus, we extend this equality to the radial part of the free average of k free unitary Brownian motions and to the free Jacobi process associated with two self-adjoint projections with trace 1/k, provided the initial distributions coincide. In the single projection case, we derive a binomial-type expansion of the moments of the free Jacobi process which extends to any \(k \ge 3\) the one derived in Demni et al. (Indiana Univ Math J 61:1351–1368, 2012) in the special case \(k=2\). Doing so give rise to a non normal (except for \(k=2\)) operator arising from the splitting of a self-adjoint projection into the convex sum of k unitary operators. This binomial expansion is then used to derive a pde satisfied by the moment generating function of this non normal operator and for which we determine the corresponding characteristic curves. As an application of our results, we compute the average purity and the entanglement entropy of the large-size limiting density matrix.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction and motivation
1.1 Random matrices in quantum information theory
Randomness lies at the heart of Shannon’s pioneering work on classical information theory (see the expository paper [29]). It also plays a key role in quantum information theory through the use of techniques from random matrix theory. Actually, the latter open the way to choose typical random subspaces in large-size quantum systems which violate additivity conjectures for minimum output Rényi and von Neumann entropies (see [5] and references therein). Here, typicality is taken with respect to the uniform measure in the compact complex Grassmann manifold or equivalently with respect to the Haar distribution in the group of unitary matrices. Note that this distribution together with Ginibre random matrices also served in [6] to generate random density matrices induced from states in bipartite systems (see [22] for similar constructions of quantum channels).
A natural dynamical version of the Haar distribution in the group of unitary matrices is the so-called unitary Brownian motion [24]. This stochastic process was used in [27] where the authors introduced and studied a random state drawn from the Brownian motion on the complex projective space (the row vector of a unitary Brownian motion up to a phase). There, the main problem was to write explicitly the joint distributions of tuples formed by the moduli of the state coordinates. This problem was entirely solved in [8] using spherical harmonics in the unitary group. To the best of our knowledge, [8, 27] are the only papers where the unitary Brownian motion is used as a random model in quantum information theory, in contrast to the high occurrence of Haar-distributed unitary matrices [5]. Moreover, it is tempting and challenging as well to prove finite-time analogues of important results in quantum information theory proved using Haar unitary matrices and their Weingarten Calculus (as summarized in [5]).
In this paper, we appeal once more to the unitary Brownian motion in order to introduce a stochastic process valued in the space of density matrices [see (1) below]. The large time limit of this process was already constructed in [6] by partially tracing a pure random state in a bipartite quantum system.
1.2 The dynamical density matrix
Let \(N \ge 1\) be a positive integer and consider a bipartite quantum system \(\mathscr {H}_A \otimes \mathscr {H}_B\), where \(\mathscr {H}_A, \mathscr {H}_B,\) are complex N-dimensional Hilbert spaces. If \((e_j^A)_{j=1}^N, (e_j^B)_{j=1}^N\), are the canonical basis of \(\mathscr {H}_A\) and \({\mathscr {H}_{B}}\) respectively, then
is referred to as the Bell or maximally-entangled state. Now, consider \(k \ge 2\) Haar-distributed unitary matrices \(U_{\infty }^1(N), \dots , U_{\infty }^k(N),\) and define the vector \(\psi ^k \in \mathscr {H}_A \otimes \mathscr {H}_B\) by:
Then the partial trace with respect to \(\mathscr {H}_B\) of the pure state associated with \(\psi ^k(N)\) yields the following reduced state:
where \(\text {tr}\) is the trace operator on the space \(\mathbb {M}_N(\mathbb {C})\) of \(N \times N\) complex matrices. Since the Haar distribution is the stationary distribution of the unitary Brownian motion, it is then natural to introduce the following stochastic process valued in the space of density matrices:
where \((U_{t}^j(N))_{t \ge 0}, 1 \le j \le k,\) are k independent unitary Brownian motions. In particular, \((\widetilde{W}_t^k(N))_{t\ge 0}\) interpolates between the completely mixed state \(\text {Id}_N/N\) at \(t=0\) (\(\text {Id}_N\) being the identity matrix of size N) and the stationary state \(\widetilde{W}_{\infty }^k(N)\). Note also that since \((U_{t}^j(N))_{t \ge 0}, 1 \le j \le k,\) are Lévy processes in the unitary group with identical distributions, then
where the second equality follows from [23], Example 3.3.
As we shall now explain, introducing this model is not simply a matter of replacing Haar-distributed matrices by unitary Brownian motions. Indeed, the large-size limit of \(\widetilde{W}_t^2(N)\) for fixed time t bears a close connection to an instance of the so-called free Jacobi process.
1.3 The large size limit of \(\widetilde{W}_{\infty }^k(N)\) and the free Jacobi process
Recall that independent random matrices behave in the large-size limit, under additional law-invariance assumptions, as \(\star \)-free operators (in Voiculescu’s sense) in a tracial non commutative probability space, say \((\mathscr {A}, \tau )\) [26]. For instance, independent Haar-distributed unitary matrices converge strongly and almost surely as \(N \rightarrow \infty \) to Haar-distributed unitary operators (see [4] and references therein). Consequently, the norm of the operator \(\widetilde{W}_{\infty }^k(N)\) converges almost surely as \(N \rightarrow \infty \) to
where \(\{U_{\infty }^j, 1 \le j \le k\}\) is a k-tuple of Haar unitary operators which are \(\star \)-free in \((\mathscr {A}, \tau )\). Note that
since \(\tau (U_{\infty }^j) = 0\) and since \(\star \)-freeness entails \(\tau (U_{\infty }^j(U_{\infty }^m)^{\star }) = \tau (U_{\infty }^j)\tau ((U_{\infty }^m)^{\star }) = 0\) for any \(1 \le j \ne m \le k\). In particular, when \(k=2\), the invariance of the Haar distribution shows further that \(\widetilde{W}_{\infty }^2/2\) is equally distributed as:
where \(\textbf{1}\) stands for the unit of \(\mathscr {A}\). The spectral distribution of this Hermitian operator is known to be the arcsine distribution [18] and coincides also with an instance of the stationary (i.e. \(t=+\infty \)) distribution of the so-called free Jacobi process [7]. At any time \(t > 0\), the latter is the squared radial part of the compression \(PU_tQ\) of a free unitary Brownian motion \((U_s)_{s \ge 0}\) by two orthogonal (self-adjoint) projections P and Q in \((\mathscr {A}, \tau )\) which are \(\star \)-free from \((U_s)_{s \ge 0}\) [7]. Besides, it was shown in [12] that the coincidence alluded to above holds even at any time \(t > 0\): if \(P = Q\) and if \(\tau (P) = 1/2\) then the spectral distribution of the free Jacobi process \(PU_tPU_t^{\star }P\) (in the compressed algebra \((\mathscr {A}, 2\tau )\)) coincides with that of
Since \((U_s)_{s \ge 0}\) is a unitary free Lévy process with respect to the free multiplicative convolution on the unit circle [18], then the spectral distribution of (2) coincides with that of
where \((U_s^1)_{s \ge 0}\) and \((U_s^2)_{s \ge 0}\) are two free copies of \((U_s)_{s \ge 0}\).
On the other hand, it was proved in [1] that \((U_s)_{s \ge 0}\) is the large-size limit of the time-rescaled Brownian motion in the group of unitary matrices. As a matter of fact, (3) is the large-size limit \(N \rightarrow \infty \) of
Up to a scalar random factor, this Hermitian random matrix is nothing else but \((1/2)\widetilde{W}_{t/N}^2(N)\) which converges almost surely and strongly to
where the second equality follows from \(\tau (U_t^1(U_t^2)^{\star }) = \tau (U_t^1)\tau ((U_t^2)^{\star }) = e^{-t}\) [1]. In a nutshell, the free Jacobi process associated with an orthogonal projection with trace 1/2 is, up to a normalising factor, the large-size limit of \(\widetilde{W}_{t/N}^2(N)\).
The above picture extends to any integer \(k \ge 2\) as follows. On the one hand, \(\widetilde{W}_{t/N}^k(N)\) converges strongly and almost surely as \(N \rightarrow \infty \) to the self-adjoint and unit-trace operator [4]:
where \((U_s^j)_{s \ge 0}, 1 \le j \le k\) are free copies of \((U_s)_{s \ge 0}\) in \((\mathscr {A}, \tau )\). In particular, the following diagram commute:
On the other hand, if
and if \(\tau (P) = 1/k\), then Nica and Speicher’s boxed convolution [28] implies that the \(\star \)-moments of \(G_t^k\) in \((\mathscr {A}, \tau )\) coincide with those of \(PU_tP\) in the compressed space \((P\mathscr {A}P, k\tau )\). Consequently, their corresponding Brown measures coincide [26] and so do the spectral distributions of their squared radial parts
and
1.4 Main results
The choice \(P=Q\) is not a restriction and is rather a matter of simplicity. Indeed, since we are dealing with time dynamics (of Burgers-type) instead of stationary \((t = +\infty \)) regimes, we have to match initial data at \(t=0\) in order to obtain equalities between spectral distributions at any fixed \(t > 0\). For instance, the operator displayed in (2) and \(PU_tQPU_t^{\star }P\) share the same spectral distribution in their corresponding probability spaces when \(\tau (P) = \tau (Q) = 1/2\) and provided that their moment sequences at \(t=0\) coincide [20]. More generally, we shall prove using free stochastic calculus that for any \(t > 0\), the moment sequences of \(W_t^k/k^2\) in \((\mathscr {A}, \tau )\) and of \(PU_tQU_t^{\star }P\) in \((P\mathscr {A}P, k\tau )\) satisfy the same recurrence relation when \(\tau (P) = \tau (Q) = 1/k, k \ge 2\). We shall also prove that the moment sequence of \(W_t^k/k^2\) converge as \(k \rightarrow \infty \) to \((e^{-nt})_{n \ge 0}\) for any fixed time t, which contrasts the weak convergence of \(W_{\infty }^k/k\) to the Marchenko–Pastur distribution [6]. This contrast is due to the high complexity of the structure of the \(\star \)-cumulants of \(U_t\) in comparison with those of \(U_{\infty }\) [9].
Back to the case \(P=Q\), the equality between the spectral distributions of \(W_t^k/k^2\) and of \(PU_tPU_t^{\star }P\) under the assumption \(\tau (P) = 1/k\) opens the way to compute the moments of the former by studying those of the latter. Indeed, for any \(n \ge 1\), \(\tau [(W_t^k)^n]\) is a linear combination of \(k^{2n}\) factors of the form
Apart from constant factors, those where any index \(i_j\) occurs at most once may be computed using the multiplicative Lévy property of the free unitary Brownian motions. However, to the best of our knowledge, the contributions of the remaining factors may be only computed using the freeness property. In this respect, the complexity of \(\tau [(W_t^k)^n]\) increase rapidly even for small orders. For that reason, we rather focus on the moments of \(PU_tPU_t^{\star }P\) and our main result (Theorem 2 below) establishes for any \(n \ge 1\) a binomial-type expansion of
as a linear combination of the moments
where \(T_k:= kP - \textbf{1} = T_k^{\star }\) satisfies \(\tau (T_k) = 0\). This expansion extends to any integer \(k \ge 3\) the one proved in [12] for \(k=2\) for which \(T_2 = 2P-\textbf{1}\) is unitary and self-adjoint, which in turn implies that \(T_2U_tT_2U_t^{\star }\) is distributed as \(U_{2t}\). However, for any \(k \ge 3\), \(T_k\) is not even normal: it is the sum of \((k-1)\) unitary operators and satisfies the relation
Of course, the constant term corresponding to \(j=0\) in the obtained binomial-type expansion is nothing else but the nth moment of the spectral distribution of \(PU_{\infty }PU_{\infty }^{\star }P\), and may be expressed for instance as a weighted sum of Catalan numbers. Surprisingly, the higher order coefficients split as
and are derived after a careful and tricky analysis of several inductive relations. Nonetheless, it would be interesting to seek a combinatorial proof explaining both the splitting of these higher order coefficients and the occurrence of the same binomial coefficients as in the \(k=2\) case.
Once the binomial-type expansion derived, we turn it into a relation between the moment generating functions of the free Jacobi process at time \(t > 0\) and of \(T_kU_tT_kU_t^{\star }\). Using the partial differential equation (hereafter pde) satisfied by the former and derived in [7] for arbitrary traces \(\tau (P)\) and \(\tau (Q)\), we derive the one satisfied by the latter and determine its characteristic curves.
At the quantum information theoretical side, our results allow to compute large-size asymptotics of relevant quantities such as the average purity and the entanglement entropy of the dynamical density matrix \(\widetilde{W}_{t/N}^k(N)\). In the stationary regime \(t = +\infty \), the former was computed in [6] (see section III.C. there) and amounts to compute the second moment of the spectral distribution of \(\widetilde{W}_{\infty }^k/k\). For any fixed time \(t > 0\), the second moment \(\tau [(\widetilde{W}_t^k)^2]\) is readily computed from Theorem 2 together with the freeness property. As a by-product, we recover the large-time limit of the asymptotic average purity for fixed k and we derive its large-k limit at any fixed time \(t > 0\). Still in the large-size limit, we derive formulas for the entanglement entropy both in the stationary and in fixed time regimes. Both expressions involve Gauss hypergeometric polynomials and the second one expresses the difference between the entanglement entropy in both regimes as an infinite series of the moments \(\tau [(T_kU_tT_kU_t^{\star })^j], j \ge 0,\).
The paper is organized as follows. In the next section, we discuss the relation between the \(\star \)-moments of \((G_t^k)/k\) and those of the compression \(PU_tP\) when \(\tau (P) = 1/k\). There, we also prove that the moment sequences of the radial parts \((W_t^k)/k^2\) and of \(PU_tQ\) with \(\tau (P) = \tau (Q) = 1/k\) satisfy the same recurrence relation and that their limits as \(k \rightarrow \infty \) is the Dirac mass at \(e^{-t}\). In the third section, we prove the binomial-type formula for the moments of \(J_t\) then turn it into a relation between moment generating functions. Once we do, we deduce a pde for the moment generating function of \(T_kU_tT_kU_t^{\star }\) and determine its characteristic curves. The fourth section is devoted to formulas for the average purity and for the average entanglement entropy of the density matrix in the large-size limit. We also include two appendices where we prove two formulas which we could not find in literature and which we think are of independent interest. The first formula has the merit to express the moments of the stationary distribution of the free Jacobi process corresponding to \(\tau (P) = 1/k\) as a perturbation of those corresponding to \(\tau (P) = 1/2\). In particular, it involves a family of polynomials with integer coefficients in the variable \((k-2)\) and its derivation relies on special properties of the Gauss hypergeometric function. As to the second formula, it expresses the free cumulants of a self-adjoint projection with arbitrary rank as a difference of two Legendre polynomials.
2 Relating \(G^t_k\) and compressions of \(U_t\)
2.1 Compression by a free projection and Brown measure
Given a collection of operators \((a_1, \dots , a_n)\) in a non commutative probability space \((\mathscr {A}, \tau )\), their joint distribution \(\mu _{a_1, \dots , a_n}\) is the linear functional which assigns to any polynomial P in n non commuting indeterminates its trace \(\tau (P(a_1, \dots , a_n))\). In this respect, the Nica–Speicher generalized R-transform [28] allows to relate the joint distribution of the compressed collection \((Pa_1P, \dots , Pa_nP)\) by a free self-adjoint projection P in the compressed algebra to \(\mu _{a_1, \dots , a_n}\). In particular, when \(n=2\) and if \(a_1 = a, a_2 = a^{\star }\) then \(\mu _{a, a^{\star }}\) is given by all the \(\star \)-moments of a and we shall simply refer to it as the distribution of a. The following result shows that if \(\tau (P) = 1/k, k \ge 2,\) then the compression of \((U_t, U_t^{\star })\) by P amounts to summing k free copies of \((U_t, U_t^{\star })\) up to dilation. Though we expect that this result is known among the free probability community, we did not find it written anywhere and we include it here for the reader’s convenience. Note also that it reduces to the Nica–Speicher convolution semi-group when a is self-adjoint.
Proposition 1
Let P be a self-adjoint projection freely independent from \(\{U_t,U_t^\star \}_{t \ge 0}\) with \(\tau (P) = 1/k, k \ge 2,\) and recall the non normal operator:
Then, the distribution of \(PU_tP\) in \((P\mathscr {A}P, k\tau )\) coincides with that of \(G_t^k/k\) in \((\mathscr {A}, \tau )\).
Proof
Given an operator \(a \in \mathscr {A}\), let \(R(\mu _{a,a^{\star }})\) be its generalized R-transform ([28], section 3.9) and recall that it entirely determines the distribution of a. Then, one has on the one hand:
due to the \(\star \)-freeness of \((U_t^j)_{j=1}^k\) [28]. On the other hand, [28, Application 1.11] entails
where the distribution \(\mu _{PU_tP,PU_t^\star P}\) is considered in the compressed space \((P\mathscr {A}P, k\tau )\). \(\square \)
The Brown measure of a non normal operator plays a key role in random matrix theory since it supplies a candidate for the limiting empirical distribution of a non normal matrix ([26], chapter XI). It is defined through the Fuglede–Fuglede–Kadison determinant given for any \(a \in \mathcal {A}\) by:
where \(|a| = (a^{\star }a)^{1/2}\) is the radial part of a and \(\mu _{|a|}\) is its spectral measure. The Brown measure \(\mu _a\) of a is then defined by:
where \(\nabla ^2\) denotes the Laplacian taken in the distributional sense, and is uniquely determined among all compactly-supported measure by its logarithmic potential:
In a tracial non commutative probability space, the Brown measure is fully determined by \(\star \)-moments and one immediately deduces from the previous proposition that the Brown measures of \(G_t^k/k\) and of \(PU_tP\) coincide when \(\tau (P) = 1/k\). In general, the description of the Brown measure of \(PU_tP\) is a quite difficult problem: the main result proved in [10] already provides a Jordan domain containing its support. As a matter of fact, Proposition 1 offers another way to compute the Brown measure of \(PU_tP\) in the particular case \(\tau (P) = 1/k\) relying on operator-valued free probability as explained in [3]. However, it turns out that the computations are already tedious even for \(k=2\) and as such, we postpone them to a future research work.
In the stationary regime \(t =+\infty \), the fact that the R-diagonal operator \(PU_{\infty }P\) and the average of k free Haar unitaries share the same Brown measure is transparent from Haagerup–Laarsen results ([17], examples 5.3 and 5.5) though not being explicitly pointed out there. Indeed, this measure is radial and absolutely continuous with density given by [17]:
with respect to Lebesgue measure \(d\lambda \). Another approach relying on the so-called quaternion free probability may be found in [21].
2.2 Radial parts and beyond
If we consider the radial parts of \(PU_tP\) and of \(G_t^k/k\), then the equality between the moments of \(PU_tPU_t^{\star }P\) and of \(W_t^k/k^2\) may be readily deduced from the moment-cumulant formula for the compression by a free projection (see Theorem 14.10, [29]). Actually, if \((a_1, \dots , a_m)\) is a collection of operators in \(\mathscr {A}\) which is free from P, then
for any indices \(1 \le i_1, \dots , i_n, \le m\). Here NC(n) is the lattice of non crossing partitions, \(|\pi |\) is the number of blocks of the partition \(\pi \in NC(n)\) and \(\kappa _{\pi }\) is the multiplicative functional of free cumulants of blocks of \(\pi \) (see Lectures 10 and 11 in [29] for more details). Specializing (5) with \((a_{i_{2j+1}}, a_{i_{2j+2}}) = (U_t, U_t^{\star }), 0 \le j \le n-1,\) and \(\tau (P) = 1/k\), we get:
On the other hand, the moment-cumulant formula (11.8) in [29] entails:
where we recall that \(G_t^k = U_t^1+\dots +U_t^k\) and \(W_t^k = G_t^k(G_t^k)^{\star }\). But if V is a block of \(\pi \) then \(\kappa _V\) is the sum of terms of the form
where \(\epsilon (1), \dots , \epsilon (2n) \in \{1, \star \}\) and \(1 \le j_1, \dots , j_{2n}, \le k\). All these terms vanish due to the \(\star \)-freeness of \((U_t^j)_{j=1}^k\) except those of the form
for a single index \(1 \le j \le k\). There are k such terms and all give the same contribution
since \(U_t^1, \dots , U_t^k\) have the same spectral distribution as \(U_t\). Consequently, the RHS of (7) and (6) are equal.
More generally, we shall prove below that given two orthogonal projections P and Q which are \(\star \)-free from \((U_t)_{t \ge 0}\), the moments of \(PU_tQU_t^{\star }P\) and those of \(W_t^k/k^2\) coincide provided that \(\tau (P) = \tau (Q) = 1/k\). Our main tool is free stochastic calculus and we refer to [2, 15] for further details on this calculus. To proceed, recall from [1] the stochastic differential equation satisfied by the free unitary Brownian motion \((U_t)_{t \ge 0}\):
where \((X_t)_{t \ge 0}\) is a free additive Brownian motion. Hence, there exists a k-tuple free additive Brownian motions \((X_t^j)_{t \ge 0}, 1 \le j \le k,\) which are free in \(\mathscr {A}\) and such that
With the help of the free Itô formula [1], we shall prove:
Theorem 1
For any \(n \ge 1, t > 0\), setFootnote 1
Then,
where an empty sum is zero.
Proof
Using (8), we get
whence
where \((dG_t^k)((dG_t^k)^{\star })\) stands for the bracket of the semimartingales \(dG_t^k\) and \((dG_t^k)^{\star }\). Since \((X_t^j)_{t \ge 0}\) are assumed free then
so that
Now, borrowing the terminology and the notations of [2], we introduce the bi-processes:
and write:
Consequently, for any \(n \ge 1\), Proposition 4.3.2 in [2] entails:
Taking the expectation with respect to \(\tau \) of both sides and differentiating with respect to the variable t,Footnote 2 we get:
The last (triple) sum yields the following contribution (the summands there do not depend on the indices j, l):
where the last equality follows from the index change \(m \mapsto n-m-2\). Finally, the summands
do not depend on j since \(W_t^k\) and \(G_t^k\) are symmetric (invariant under permutations) and since the unitary operators \(U_t^j, 1 \le j \le k,\) are free and have identical distributions. As a result,
Similarly,
Performing the index change \(m \mapsto n-m-2\) in \(S_2\), we end up with:
Gathering all the contributions above, we obtain (9). \(\square \)
Setting \(r_n(t):= s_n(t)/k^{2n} = \tau [(W_t^k/k^2)^{n}]\), we readily infer from (9):
Corollary 1
For any \(n \ge 1\),
The moment relation (10) is an instance of the one derived in Corollary 6.1 in [7]. More precisely, let
be the free Jacobi process associated with the self-adjoint projections (P, Q). Viewed as an operator in the compressed algebra \((P\mathscr {A}P, \tau /\tau (P))\), its moments
satisfy the following differential system:
where \(\tau (P) = \lambda \theta \in (0,1], \tau (Q) = \theta \in (0,1]\). Consequently, if \(\lambda = 1, \theta = 1/k\) then (10) and (11) coincide and in turn both moment sequences coincide provided that \(m_n(0) = r_n(0)\) for all \(n \ge 0\).
2.3 Limit as \(k \rightarrow \infty \)
Let \(U_{\infty } \in \mathscr {A}\) be a Haar unitary operator and assume that \(U_{\infty }\) is free with \(\{P,Q\}\). If \(\tau (P) = \tau (Q) = 1/k\) then the spectral distribution of
in the compressed algebra \((P\mathscr {A}P, \tau /\tau (P))\) is given by (see e.g. [7], p. 130):
Its pushforward under the dilation \(x \mapsto kx\) is readily computed as
and converges weakly to the Marchenko–Pastur distribution of parameter one [6]:
If we denote \(W_{\infty }^k:= (U_{\infty }^1+\dots + U_{\infty }^k)(U_{\infty }^1+\dots + U_{\infty }^k)^{\star }\) then we can rephrase the weak convergence above as follows: for any \(n \ge 0\),
The normalization by \(k^n\) may be guessed from the moment-cumulant expansion:
since partitions \(\pi \in NC(2n)\) with more than \((n+1)\) blocks have zero contribution. Indeed, in any such partition, at least one block admits an odd number of elements in which case the corresponding free \(\star \)-cumulant vanishes (see [29], Proposition 15.1).
For fixed time \(t > 0\), the situation becomes different since the free \(\star \)-cumulants of \(U_t\) admit a considerably more complicated structure compared with those of \(U_{\infty }\) [9]. In this respect, we can prove the following limiting result under the stronger normalization \(k^2\), which shows that reversing the order of the (k, t) limits in (12) does not lead to a finite limit.
Proposition 2
For any \(n \ge 0, t \ge 0,\)
In particular, the free Jacobi process \((PU_tPU_t^{\star }P)_{t \ge 0}\) with \(\tau (P) = 1/k\) converges weakly as \(k \rightarrow \infty \) to the constant \(e^{-t}\) in the compressed algebra.
Proof
From (6), we readily see that the limit as \(k \rightarrow \infty \) of \(k\tau (J_t^n)\) is given by the (non crossing) partition with 2n blocks. Therefore,
Since \(c_1(U_t) = c_1(U_t^{\star }) = \tau (U_t) = e^{-t/2}\) (see e.g. [12] and references therein), the proposition follows. \(\square \)
3 Analysis of the moments of the free Jacobi process
3.1 Moments binomial-type formula
For sake of simplicity, we restrict our study from now on to the free Jacobi process \((J_t)_{t \ge 0}\) associated with a single projection P. For those interested in a more general treatment of the free Jacobi process associated with a pair of projections with arbitrary traces, we recommend referring to [14,15,16]. Recall from [12] that when \(\tau (P) = 1/2\), the moments of the free Jacobi process are linear combinations of those of \(U_{2t}\). Indeed, it was observed there that
where \(S = 2P-\textbf{1}\) satisfies \(S = S^{\star } = S^{-1}\). Moreover, Lemme 3.8 in [17] together with the semi-group property of \((U_t)_{t \ge 0}\) show that the spectral distributions of \(SU_tSU_t^{\star }\) and of \(U_{2t}\) coincide. More generally, write:
Then \(S_{j,k}\) is a unitary operator satisfying \((S_{j,k})^k = \textbf{1}\) and
where \(\omega _{j,k} = e^{2i\pi j/k}\) is the kth root of unity. Set
Then \(T_k\) is self-adjoint and \((T_k)^2 = k(k-2)P + \textbf{1} = (k-2)T_k + (k-1)\textbf{1}\). In this respect, we shall prove the following generalization of (13):
Theorem 2
For any \(k \ge 2\) and any \(n \ge 1\),
where \(m_n(\infty )\) is the nth moment of \(J_{\infty }\) in \((P\mathscr {A}P, k\tau )\), given by (21) and (34).
The proof of this Theorem relies on the following four key lemmas.
Lemma 1
Let \(a, b \in \mathscr {A}\) be two operators satisfying \(a^2=(k-2)a + (k-1)\textbf{1}\), \(b^2=(k-2) b+ (k-1)\textbf{1}\). Then, the expansion of \([(a+1)(b+1)]^n\) is uniquely written as:
for some integer sequences \(m_n, (c_{n,j}), (d_{n,j}), (e_{n,j}), (f_{n,j})\) satisfying
Proof
For sake of clarity, we shall omit the notation \(\textbf{1}\) in front of the constant terms. Firstly, the uniqueness follows from the fact that the expansion is a reduced expression. Now, since \(a(a+\textbf{1}) = (k-1)(a+\textbf{1})\) then
Multiplying (14) by \((k-1)\) and using the uniqueness of the coefficients, we readily get:
Similarly, \(b(b+\textbf{1}) = (k-1)(b + \textbf{1})\) so that
The uniqueness property again yields:
Combining (15) and (16), the lemma is proved. \(\square \)
According to Lemma 1, we only need to focus on the sequences \((m_n)_n, (c_{n,j})_{1 \le j \le n}\). The former is closely related to the moment sequence \(m_n(\infty )\) of \(\tilde{\mu }_{\infty }\). As to the latter, it satisfies the following relations:
Lemma 2
For any \(2 \le j \le n-1\),
while
Proof
Follows readily from
together with the identities:
\(\square \)
Note that Lemma (1) allows to rewrite (17) and (18) as
where we set \(c_{n,j} = 0, j > n\). Next, we need the following routine computations to prove Lemma 4 below and which give our first formula for \(m_n(\infty )\):
Lemma 3
For any \(n\ge 1\), we have
where \(C_n\) is the nth Catalan number. In particular,
Proof
The expression of \(m_n(\infty )\) follows. \(\square \)
Remark 1
Taking the expectation in (14), we infer that \(m_n(\infty ) = m_n/k^{2n-1}\). Consequently, the last relation may be written as
or equivalently,
This elementary identity will be used in the proof of Lemma 4 below.
Now, set
where an empty sum is zero. Then
Lemma 4
For any \(n\ge 1\), we have
Proof
We proceed by induction: \(K_{1,0} = 2(k-1)c_{1,1} = 2(k-1)\). Next, assume the result is valid up to order n and write (we recall that \(c_{n,j} =0, j > n\)):
Appealing to the induction hypothesis and to the identity (22), we end up with
as desired. \(\square \)
We are now ready to prove Theorem 2.
Proof of Theorem 2
We apply Lemma 1 to \(a = T_k\) and \(b= U_tT_kU_t^{\star }\) and take the expectation with respect to \(\tau \). By the trace property and the fact that \(\tau (T_k) = \tau (U_tT_kU_t^{\star }) = 0\), we have
whence
where
Equivalently, Lemma 1 again entails:
Appealing further to (19), we obtain
with the convention \(K_{n,j} =0, j > n\) and with \(K_{n,0}\) given by (23). Finally, (27) is satisfied by the sequence
as readily seen from the identity
Moreover, Lemma 4 and the obvious value \(K_{n,n} = 1, n \ge 1,\) show that the boundary conditions coincide, whence we deduce:
Noting that
we are done. \(\square \)
Remark 2
(Combinatorial approach) Applying the moment formula with product as entries, it follows that:
When \(\tau (P) = 1/2\), it is known that the free cumulants of P are given by ([29], Exercise 11.35):
It would be interesting to recover (13) using these formulas together with properties of non crossing partitions. More generally, we can prove (see “Appendix B”) that if P is a self-adjoint projection with \(\tau (P) = \alpha \), then
where \((P_n)_{n \ge 0}\) is the family of Legendre polynomials defined through the Gauss hypergeometric function by:
3.2 pde for the moment generating function
Let
be the moment generating functions of \(J_t\) in the compressed space \((P\mathscr {A}P, k\tau )\) and of \(T_kU_tT_kU_t^{\star }\) in \((\mathscr {A}, \tau )\). Both series have positive convergence radii since the corresponding operators are bounded. From Theorem 2, we deduce the following relation:
Corollary 2
For any \(k \ge 2\) and any \(t > 0\),
where
is the moment generating function of \(J_{\infty }\) and
Proof
It is obvious from Theorem 2 that
The expression of \(M_{\infty ,k}\) is already known (see e.g. section 5 in [7]). Now, recall the following result ([25], p.357): if \((a_n)_{n \ge 0}, (b_n)_{n \ge 0}\) are two real sequences satisfying
then
whenever both series converge absolutely. Applying this result with
and noting that
conclude the proof. \(\square \)
From this corollary, we can derive a pde for \(\rho _{t,k}\):
Proposition 3
The moment generating function \(\rho _{t,k}(z)\) satisfies the pde:
in a neighborhood of the origin with the initial condition:
Proof
Let
be the Cauchy–Stieltjes transform of the free Jacobi process \(PU_tPU_t^{\star }P\) with \(\tau (P) = 1/k\), and recall from [7] that it satisfies the pde:
Then the variable change \(z \mapsto 1/z\) shows that
Now set \(R_{t,k}:= M_{t,k} - M_{\infty ,k}\), then we further get:
where the first equality follows from the fact that \(M_{\infty ,k}\) is a stationary solution of the pde (32):
and the second one follows from (30). Noting that
it follows that the map \((t,z) \mapsto \rho _{t,k}(\alpha (z))\) satisfies:
Next, \(\alpha \) is a one-to-one holomorphic map in \(\mathbb {C}\setminus [1,\infty [\) onto the open unit disc with inverse given by:
Moreover
whence
The sought pde satisfied by \(\rho _{t,k}(z)\) follows after few computations. Finally,
Letting \(h_n:= \tau [(T_k)^n], n \ge 0,\) then we can easily prove using the relation \(T_k^2 = (k-2)T_k + (k-1)\textbf{1}\) that
This is a generalized Fibonacci sequence for which a Binet formula already exists [19]:
As a result,
The proposition is proved. \(\square \)
Setting \(\eta _{t,k}(z):= \rho _{t,k}(e^{t}z)\) then
yielding
In particular, if \(k=2\) then
while
In this case, it is known that \(\eta _{t,2}\) is the moment generating function of the free unitary Brownian motion \(e^tU_{2t}\) [12]:
where \(L_{n-1}^{(1)}\) is the \((n-1)\)th Laguerre polynomial of parameter one. More generally (i.e. for \(k\ge 3\)), the analysis of the characteristic curves of (31) is quite involved as shown in the following paragraph.
3.3 Characteristic curves of the pde
Denote
so that the pde (31) reads:
Elementary transformations show that \(\tilde{\rho }_{t,k}(z):= [\rho _{t,k}(z)]/(k-1)\) satisfies
Let z be fixed in a neighborhood of the origin. Then a characteristic curve starting at z is locally the unique solution of the Cauchy problem:
where we set \(f_k(t):= \tilde{\rho }_{t,k}(z_k(t))\). Along such curve, it holds that:
Now, set
and note that H is an involution \((H^{-1} = H)\), \(H'(u) = -(H(u)-1)^2/2\) and
Then the curve defined by \(y_k(t):= H(z_k(t))\) solves locally around \(-1\) the Cauchy problem:
Besides,
Consequently,
or equivalently
This equation is integrable and yields:
Written differently leads to the functional equation:
where we simply wrote
Setting \(g_k(0):= \tilde{\lambda }_k(y)f_k^2(0) + f_k(0)\), then one has locally:
where the principal determination of the square root is considered. It follows that:
Now, consider the indefinite integral
for two indeterminates (A, B) independent of the variable u. Then
provided the square root is well-defined (we can take any determination of the logarithm). Taking \(A = 1+ k^2g_k(0), B = (k-2)^2g_k(0)\), one gets
or after exponentiating this identity:
Noting that
then \(A-By^2 = 1 + 4g_k(0) \tilde{\lambda }_k(y) = y^2\) which in turn entails:
If we denote the LHS of the second equality \(\xi _{2t}(\sqrt{A-B})\), then lengthy computations yield:
The map \((-\xi _{2t})\) is the inverse of the Herglotz transform \(1+2\eta _{t,2}\) of the spectral distribution of \(U_{2t}\) in a neighborhood of \(u=1\). As a matter of fact, the map
is a locally invertible in a neighborhood of the origin \(u=0\). Let \(\zeta _{2t}\) be its inverse then
whence we end up with:
Remark 3
If \(k=2\) then
Thus, it holds locally:
For fixed \(t > 0\), the map \(z \mapsto z_2(t)\) is known as the \(\Sigma \)-transform of the spectral distribution of \(U_{2t}\) [1].
4 Application to quantum information: asymptotics of the average purity and of the entanglement entropy
Starting from a pure state \(\psi \) in a bipartite quantum system \(\mathscr {H}_A \otimes \mathscr {H}_B\), the reduced state corresponding to \(\mathscr {H}_A\), say \(\rho _A \in \mathbb {M}_N(\mathbb {C})\), is in general mixed. The average purity of \(\rho _A\) is then defined by \(\text {tr}(\rho _A^2)\) and ranges in the interval [1/N, 1] whose endpoints correspond respectively to the completely mixed and to pure states.
As far as the dynamical density matrix \(\widetilde{W}_t^k(N)\) is concerned, we can use Theorem (2) together with the freeness property to compute the large-size asymptotic of its average purity (after time-rescaling \(t \mapsto t/N\) in order to make the matrix model converge). Straightforward computations then yield the following expression:
Letting \(t \rightarrow + \infty \), we recover the stationary asymptotic average purity \([2-(1/k)]/N\) computed in [6]. Fixing \(t > 0\) and letting \(k \rightarrow +\infty \), we rather get:
The entanglement entropy of the bipartite system is defined as the von Neumann entropy:
where \(\ln (\rho )\) is defined through functional calculus. Since \(\widetilde{W}_{t/N}^k(N)\) converges in operator norm to \(\widetilde{W}_{t}^k\) then we define the asymptotic entanglement entropy of \(\widetilde{W}_{t/N}^k(N)\) as \(N \rightarrow \infty \) by:
where we use again the functional calculus. To the best of our knowledge, the asymptotic entanglement entropy of \(\widetilde{W}_{\infty }^k\) has never been computed though it is given by the integral:
Expanding
one obtains (using Tonelli Theorem) the following expression of the asymptotic entanglement entropy in the stationary regime \(t=+\infty \):
where \({}_2F_1\) is the Gauss hypergeometric function and the second equality follows from its Euler integral representation ([13], p. 59). In particular, when \(k=2\), the corresponding entropy reduces to (one either uses the Beta integral or equivalently the Gauss hypergeometric Theorem):
For fixed time \(t > 0\), the spectral distribution of \(J_t\) (and in turn of \(\widetilde{W}_t^k\)) is far from being as simple as in the stationary regime, unless \(k=2\). Nonetheless, we can compute the asymptotic entanglement entropy of \(\widetilde{W}_t^k\) appealing to Theorem 2 and using the expansion
which follows readily from
To this end, we need the following lemma:
Lemma 5
For any \(m \ge 2\),
Proof
We use the binomial Theorem together with Theorem 2 to write
Now the duplication formula for the Gamma function entails:
so that
where we used the Pochhammer symbol:
Since \((j-m)_n = 0\) for any \(n > m-j, 1 \le j \le m\), then we recognize the Gauss hypergeoemetric (terminating) series:
and the lemma is proved. \(\square \)
From this lemma, it follows that the asymptotic entanglement entropy may be written as:
Equivalently,
In particular, when \(k=2\), the Gauss hypergeometric Theorem entails:
whence
Moreover, \(T_2U_tT_2U_t^{\star }\) is unitary and is distributed as \(U_{2t}\) so that:
where \(L_{j-1}^{(1)}\) is the Laguerre polynomial of index one.
5 Concluding remarks
In this paper, we introduced a dynamical random density matrix built out of \(k \ge 2\) independent unitary Brownian motions whose large size limit has, up to a normalization, the same moments as those of the free Jacobi process \(PU_tPU_t^{\star }P\) (in the compressed algebra) subject to \(\tau (P) = 1/k\). Motivated by our previous results proved in [12] valid for \(k=2\), we derived for any \(k \ge 2\) a binomial-type expansion for these moments where the non normal operator \(T_k\) (except when \(k=2\)) plays a key role and extends the orthogonal symmetry \(S = 2P-\textbf{1}\) associated with P. We also provided expressions for the average purity and for the entanglement entropy of the large-size limiting density matrix at fixed time \(t>0\). Our expression for this entropy involves the moments \(\tau [(T_kU_tT_kU_t^{\star })^j], j \ge 0,\) which do not seem to admit a simple form unless \(k=2\), as suggested by the analysis of the characteristic curves of the pde satisfied by their generating function.
On the other hand, the Lebesgue decomposition of the spectral distribution of \(J_t\) for arbitrary orthogonal projections (P, Q) was obtained in [14]. Specializing Theorem 1.1. there to \(P=Q\) with \(\tau (P) = 1/k\), we deduce that the corresponding distribution admits a density (with respect to Lebesgue measure) of the form:
where \(g^k_t\) is the density of the spectral distribution of the unitary operator \(SU_tSU_t^{\star }, \tau (S) = 2\tau (P) - 1 = (2-k)/k\). Similarly, \(g_t^k\) admits a very complicated expression unless \(k=2\).
Finally, it was recently proved in [11] (see eq. (3) there) that
for any self-adjoint projections (P, Q) with arbitrary traces \(\tau (P), \tau (Q), \in (0,1)\). In particular, if \(Q = U_tPU_t^{\star }\) and \(\tau (P) = 1/k\) then we readily deduce:
Consequently, the moments of the free Jacobi process associated with \(\textbf{1}-P\) may be deduced from those of the free Jacobi process associated with P:
Notes
We omit the dependence on k for sake of clarity.
The state \(\tau \) is tracial and all the processes are continuous in the strong topology.
References
Biane, P.: Free Brownian motion, free stochastic calculus and random matrices. In: Voiculescu D.V. (ed.) Free Probability Theory. Fields Institute Communications, vol. 12, pp. 1–19. American Mathematical Society, Providence (1997)
Biane, P., Speicher, R.: Stochastic calculus with respect to free Brownian motion and analysis on Wigner space. Probab. Theory Relat. Fields 112(3), 373–409 (1998)
Belinschi, S.T., Sniady, P., Speicher, R.: Eigenvalues of non-Hermitian random matrices and Brown measure of non-normal operators: Hermitian reduction and linearization method. Linear Algebra Appl. 537, 48–83 (2018)
Collins, B., Dahlqvist, A., Kemp, T.: The spectral edge of unitary Brownian motion. Probab. Theory Relat. Fields 170(1–2), 49–93 (2018)
Collins, B., Nechita, I.: Random matrix techniques in quantum information theory. J. Math. Phys. 571(015215), 34 (2016)
Zyczkowski, K., Penson, K.A., Nechita, I., Collins, B.: Generating random density matrices. J. Math. Phys. 52(6), 062201 (2011)
Demni, N.: Free Jacobi process. J. Theor. Probab. 21, 118–143 (2008)
Demni, N.: Distributions of truncations of the heat kernel on the complex projective space. Ann. Math. Blaise Pascal 21(2), 1–20 (2014)
Demni, N., Guay-paquet, M., Nica, A.: Star-cumulants of the free unitary Brownian motion. Adv. Appl. Math. 69, 1–45 (2015)
Demni, N., Hamdi, T.: Support of the Brown measure of the product of a free unitary Brownian motion by a free self-adjoint projection. J. Funct. Anal. 282(6), 109362 (2022)
Demni, N., Hamdi, T.: Relating moments of self-adjoint polynomials in two orthogonal projections. Available on arXiv
Demni, N., Hamdi, T., Hmidi, T.: Spectral distribution of the free Jacobi process. Indiana Univ. Math. J. 61, 1351–1368 (2012)
Erdelyi, A., Magnus, W., Oberhettinger, F., Tricomi, F. G.: Higher Transcendental Functions. Vol. I, Based, in Part, on Notes Left by Harry Bateman, xxvi+302. McGraw-Hill Book Co., Inc., New York (1953)
Hamdi, T.: Spectral distribution of the free Jacobi process, revisited. Anal. PDE 11(8), 2137–2148 (2018)
Hamdi, T.: Liberation, free mutual information and orbital free entropy. Nagoya Math. J. 239, 1–27 (2018)
Hamdi, T.: Free mutual information for two projections. Complex Anal. Oper. Theory 12(7), 1697–1705 (2018)
Haagerup, U., Larsen, F.: Brown’s spectral distribution measure for R-diagonal elements in finite von Neumann algebras. J. Funct. Anal. 176(2), 331–367 (2000)
Hiai, F., Petz, D.: The Semicircle Law, Free Random Variables, and Entropy. American Mathematical Society, Providence (2000)
Horzum, T., Kocer, E.G.: On some properties of Horadam polynomials. Int. Math. Forum. 4(25–28), 1243–1252 (2009)
Izumi, M., Ueda, Y.: Remarks on free mutual information and orbital free entropy. Nagoya Math. J. 220, 45–66 (2015)
Jarosz, A.: Summing free unitary random matrices. Phys. Rev. E 84, 011146 (2011)
Kukulski, R., Nechita, I., Pawela, L., Puchala, Z., Zyczkowski, K.: Generating random quantum channels. J. Math. Phys. 62(6), 062201 (2021)
Lévy, T.: Schur-Weyl duality and the heat kernel measure on the unitary group. Adv. Math. 218(2), 537–575 (2008)
Liao, M.: Lévy processes in Lie groups. In: Cambridge Tracts in Mathematics, vol. 162, x+266. Cambridge University Press, Cambridge (2004)
Manocha, H.L., Srivastava, H.M.: A treatise on generating functions. In: Ellis Horwood Series: Mathematics and its Applications (1984)
Mingo, J. A., Speicher, R.: FreMingo, J. A., Speicher, R.: Free probability and random matrices. Fields Institute Monographs, 35. Springer, New York; Fields Institute for Research in Mathematical Sciences, Toronto, ON, (2017)
Nechita, I., Pellegrini, C.: Random pure quantum states via unitary Brownian motion. Electron. Commun. Probab. 18(27), 13 (2013)
Nica, A., Speicher, R.: On the multiplication of free N -tuples of noncommutative random variables. Am. J. Math. 118(4), 799–837 (1996)
Nica, A., Speicher, R.: Lectures on the Combinatorics of Free Probability. London Mathematical Society Lecture Note Series, vol. 335 (2006)
Verdú, S.: Fifty years of Shannon theory. IEEE Trans. Inf. Theo. 44(6), 2057–2078 (1998)
Acknowledgements
The authors extend their appreciation to the Deputyship for Research & Innovation, Ministry of Education, Saudi Arabia for funding this research work through the project number (QU-IF-2-4-1-26574). The authors also thank to Qassim University for technical support.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
On behalf of all authors, the corresponding author states that there is no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A. Moments of stationary distribution
In this appendix, we derive another expression of
To the best of our best knowledge, formula (34) below never appeared in literature. Compared to (21), it has the merit to separate the case \(k=2\) corresponding to the arcsine distribution from other values \(k \ge 3\). Our main ingredients are two properties satisfied by the Gauss hypergeometric function.
To proceed, perform the variable change \(x = 4(k-1)y/k^2\) to write:
Here, the second equality follows from the Euler integral representation of the Gauss hypergeometric function, the third and fourth ones follow from the variational formula (25), p.102 and formula (6), p.101 in [13]. Using direct computations, we readily see that
for some polynomial of degree n. For instance
Consequently,
which reduces for \(k=2\) to the known formula:
Appendix B. Free cumulants of an orthogonal projection
The first part of the proof of (29) is a routine computation in free probability theory and we refer the reader to [18] for further details on this machinery. Start with the Cauchy transform of P:
Next, consider the equation
for y lying in a neighborhood of zero. Then the K-transform of P reads:
and in turn, its R-transform is given by
It remains to write down the Taylor expansion of the function:
To this end, we appeal to the generating series of Legendre polynomials:
Indeed, setting \(\beta := 1-2\alpha \in [-1,1]\), one has
so that
where the last equality follows from the recurrence relation:
Extracting the Taylor coefficients of \(f_{\alpha }\) and recalling the definition
we get (29).
Note that since Legendre polynomials are orthogonal with respect to the uniform distribution in \([-1,1]\), they are parity preserving. In particular,
so that one recovers (28) after some computations.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Hamdi, T., Demni, N. Summing free unitary Brownian motions with applications to quantum information. Lett Math Phys 113, 80 (2023). https://doi.org/10.1007/s11005-023-01702-x
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s11005-023-01702-x