Nearest matrix

A common optimization problem that arises is to find the nearest matrix to another fixed matrix with respect to the quantum relative entropy. We show how some of these problems can be solved in QICS below.

Nearest correlation matrix

We first consider a toy example, in which we are interested in finding the closest correlation matrix to a given positive semidefinite matrix $C\in {\mathbb{S}}^n$ in the quantum relative entropy sense.

Correlation matrices are characterized by being a real positive semidefinite matrices with diagonal entries all equal to one. Therefore, the closest correlation matrix to $C$ can be found by solving the following problem

$$\begin{array}{r}\begin{array}{rlrl}\underset{Y\in {\mathbb{S}}^n}{min}& & & S(C\Vert Y)\\ \text{subj. to}& & & Y_{ii}=1i=1,\dots ,n\\ & & & Y⪰0.\end{array}\end{array}$$

We show how this problem can be solved using QICS below.

NativePICOS

import numpy as np

import qics
from qics.vectorize import vec_dim, mat_to_vec, eye

np.random.seed(1)

n = 5

vn = vec_dim(n)
cn = vec_dim(n, compact=True)

# Generate random positive semidefinite matrix C
C = np.random.randn(n, n)
C = C @ C.T / n
C_cvec = mat_to_vec(C, compact=True)

# Model problem using primal variables (t, X, Y)
# Define objective function
c = np.block([[1.0], [np.zeros((vn, 1))], [np.zeros((vn, 1))]])

# Build linear constraints
diag = np.zeros((n, vn))
diag[np.arange(n), np.arange(0, vn, n + 1)] = 1.

A = np.block([
    [np.zeros((cn, 1)), eye(n),            np.zeros((cn, vn))],  # X = C
    [np.zeros((n, 1)),  np.zeros((n, vn)), diag              ]   # Yii = 1
])  # fmt: skip

b = np.block([[C_cvec], [np.ones((n, 1))]])

# Define cones to optimize over
cones = [qics.cones.QuantRelEntr(n)]

# Initialize model and solver objects
model = qics.Model(c=c, A=A, b=b, cones=cones)
solver = qics.Solver(model)

# Solve problem
info = solver.solve()

Y_opt = info["s_opt"][0][2]
print("The nearest correlation matrix to")
print(C)
print("is")
print(Y_opt)

The nearest correlation matrix to
[[ 1.03838024 -0.9923943   1.03976304 -0.1516761  -0.54476511]
 [-0.9923943   1.81697119 -1.42389728  0.55339008  0.7559633 ]
 [ 1.03976304 -1.42389728  1.58376462 -0.0650662  -0.6859653 ]
 [-0.1516761   0.55339008 -0.0650662   0.47031665  0.1535909 ]
 [-0.54476511  0.7559633  -0.6859653   0.1535909   0.87973255]]
is
[[ 1.         -0.68153866  0.77530944 -0.15666279 -0.52667092]
 [-0.68153866  1.         -0.7916631   0.5004458   0.54809203]
 [ 0.77530944 -0.7916631   1.          0.05716203 -0.52919618]
 [-0.15666279  0.5004458   0.05716203  1.          0.16929971]
 [-0.52667092  0.54809203 -0.52919618  0.16929971  1.        ]]

Relative entropy of entanglement

Entanglement is an important resource in quantum information theory, and therefore it is often useful to characterize the amount of entanglement possessed by a quantum state. This can be characterized by the distance (in the quantum relative entropy sense) between a given bipartite state and the set of separable states.

In general, the set of separable states is NP-hard to describe. Therefore, it is common to estimate the set of separable states using the positive partial transpose (PPT) criteria [1], i.e., if a quantum state ${\rho }_{AB}\in {\mathbb{H}}^{n_A{n}_B}$ is separable, then it must be a member of

$$\mathsf{PPT}=\{{\rho }_{AB}\in {\mathbb{H}}^{n_A{n}_B}:T_B({\rho }_{AB})⪰0\},$$

where ${\mathcal{T}}_B$ denotes the partial transpose with respect to subsystem $B$. Note that in general, the PPT crieria is not a sufficient condition for separability, i.e., there exists entangled quantum states which also satisfy the PPT criteria. However, it is a sufficient condition when $n_A=n_B=2$, or $n_A=2,n_B=3$.

Given this, the relative entropy of entanglement of a quantum state ${\rho }_{AB}$ is given by

$$\begin{array}{r}\begin{array}{rlrl}\underset{{\sigma }_{AB}\in {\mathbb{H}}^{n_A{n}_B}}{min}& & & S({\rho }_{AB}\Vert {\sigma }_{AB})\\ \text{subj. to}& & & \text{tr}[{\sigma }_{AB}]=1\\ & & & {\mathcal{T}}_B({\sigma }_{AB})⪰0\\ & & & {\sigma }_{AB}⪰0.\end{array}\end{array}$$

We show how we can solve this problem in QICS below.

NativePICOS

import numpy as np

import qics
from qics.quantum import partial_transpose
from qics.quantum.random import density_matrix
from qics.vectorize import eye, lin_to_mat, mat_to_vec, vec_dim

np.random.seed(1)

n1 = 2
n2 = 3
N = n1 * n2

vN = vec_dim(N, iscomplex=True)
cN = vec_dim(N, iscomplex=True, compact=True)

# Generate random quantum state
C = density_matrix(N, iscomplex=True)
C_cvec = mat_to_vec(C, compact=True)

# Model problem using primal variables (t, X, Y, Z)
# Define objective function
c = np.block([[1.0], [np.zeros((vN, 1))], [np.zeros((vN, 1))], [np.zeros((vN, 1))]])

# Build linear constraints
trace = lin_to_mat(lambda X: np.trace(X), (N, 1), True)
ptranspose = lin_to_mat(lambda X: partial_transpose(X, (n1, n2), 1), (N, N), True)

A = np.block([
    [np.zeros((cN, 1)), eye(N, True),       np.zeros((cN, vN)), np.zeros((cN, vN))],  # X = C
    [np.zeros((1, 1)),  np.zeros((1, vN)),  trace,              np.zeros((1, vN)) ],  # tr[Y] = 1
    [np.zeros((cN, 1)), np.zeros((cN, vN)), ptranspose,         -eye(N, True)     ]   # T2(Y) = Z
])  # fmt: skip

b = np.block([[C_cvec], [1.0], [np.zeros((cN, 1))]])

# Input into model and solve
cones = [
    qics.cones.QuantRelEntr(N, iscomplex=True),     # (t, X, Y) ∈ QRE
    qics.cones.PosSemidefinite(N, iscomplex=True),  # Z = T2(Y) ⪰ 0
]

# Initialize model and solver objects
model = qics.Model(c=c, A=A, b=b, cones=cones)
solver = qics.Solver(model)

# Solve problem
info = solver.solve()

Bregman projection

A Bregman projection is a generalization of a Euclidean projection, which is commonly used in first-order optimization algorithms called Bregman proximal methods. As an example, the Bregman projection corresponding to the quantum relative entropy (see, e.g., [2]) of a point $Y\in {\mathbb{H}}_{+}^n$ onto the set of density matrices is the solution to

$$\begin{array}{r}\begin{array}{rlrl}\underset{X\in {\mathbb{H}}^n}{min}& & & S(X\Vert Y)-\text{tr}[X-Y]\\ \text{subj. to}& & & \text{tr}[X]=1\\ & & & X⪰0.\end{array}\end{array}$$

We can show that the explicit solution to this is given by $X=Y/\text{tr}[Y]$, which we can use to validate the solution given by QICS.

Note

The Bregman projection problem fixes the second argument of the quantum relative entropy, and optimizes over the first argument. This is as opposed to the first two examples which fix the first argument and optimize over the second. In this case, we can model the problem using qics.cones.QuantEntr, which allows QICS to solve problems much faster than if we modelled the problem using qics.cones.QuantRelEntr.

NativePICOS

import numpy as np
import scipy as sp

import qics
from qics.vectorize import lin_to_mat, vec_dim, mat_to_vec

np.random.seed(1)

n = 5
vn = vec_dim(n, iscomplex=True)

# Generate random positive semidefinite matrix Y to project
Y = np.random.randn(n, n) + np.random.randn(n, n)*1j
Y = Y @ Y.conj().T
trY = np.trace(Y).real

# Model problem using primal variables (t, u, X)
# Define objective function
c = np.block([[1.0], [0.0], [mat_to_vec(-sp.linalg.logm(Y) - np.eye(n))]])

# Build linear constraints
trace = lin_to_mat(lambda X: np.trace(X), (n, 1), iscomplex=True)

A = np.block([
    [0.0, 1.0, np.zeros((1, vn))],  # u = 1
    [0.0, 0.0, trace            ]   # tr[X] = 1
])

b = np.array([[1.0], [1.0]])

# Define cones to optimize over
cones = [qics.cones.QuantEntr(n, iscomplex=True)]

# Initialize model and solver objects
model = qics.Model(c=c, A=A, b=b, cones=cones, offset=trY)
solver = qics.Solver(model)

# Solve problem
info = solver.solve()

analytic_X_opt = Y / trY
numerical_X_opt = info["s_opt"][0][2]

print("Analytic solution:")
print(np.round(analytic_X_opt, 3))
print("Numerical solution:")
print(np.round(numerical_X_opt, 3))

Analytic solution:
[[ 0.147+0.j    -0.083+0.043j  0.108+0.018j -0.005+0.065j -0.085+0.042j]
 [-0.083-0.043j  0.241+0.j    -0.186+0.029j  0.049+0.022j  0.046-0.03j ]
 [ 0.108-0.018j -0.186-0.029j  0.266+0.j     0.071-0.015j -0.053+0.038j]
 [-0.005-0.065j  0.049-0.022j  0.071+0.015j  0.14 +0.j    -0.013+0.005j]
 [-0.085-0.042j  0.046+0.03j  -0.053-0.038j -0.013-0.005j  0.205+0.j   ]]
Numerical solution:
[[ 0.147+0.j    -0.083+0.043j  0.108+0.018j -0.005+0.065j -0.085+0.042j]
 [-0.083-0.043j  0.241+0.j    -0.186+0.029j  0.049+0.022j  0.046-0.03j ]
 [ 0.108-0.018j -0.186-0.029j  0.266+0.j     0.071-0.015j -0.053+0.038j]
 [-0.005-0.065j  0.049-0.022j  0.071+0.015j  0.14 +0.j    -0.013+0.005j]
 [-0.085-0.042j  0.046+0.03j  -0.053-0.038j -0.013-0.005j  0.205+0.j   ]]

References

1. “Separability of mixed states: necessary and sufficient conditions,” M. Horodecki, P. Horodecki, and R. Horodecki, Physics Letters A, vol. 223, no. 1, pp. 1–8, 1996.

2. “A Bregman proximal perspective on classical and quantum Blahut-Arimoto algorithms,” K. He, J. Saunderson, and H. Fawzi, IEEE Transactions on Information Theory, vol. 70, no. 8, pp. 5710-5730, Aug. 2024.