In [1]:
import numpy as np
import scipy.linalg


For a given matrix $\hat{A}$, we want to find the closest unitary matrix $\hat{U}$, in the sense that the operator norm (aka 2-norm) of their difference should be minimal.

$$\hat{U} = \arg \min_{\tilde{U}} \Vert \hat{A}-\tilde{U} \Vert$$

For example, let's consider the following matrix:

In [2]:
A = np.matrix([
[-1.75900766E-02-1.15354406E-01j, 6.10816904E-03+9.49971160E-03j, 1.79090787E-02+1.33311069E-02j,-1.82163102E-03-8.77682357E-04j],
[-9.77987203E-03+5.01950535E-03j, 8.74085180E-04+3.25580543E-04j,-6.74874670E-03-5.82800747E-03j,-1.95106265E-03+9.84138284E-04j],
[-6.11175534E-03+2.26761191E-02j,-5.04355339E-03-2.57661178E-02j, 2.15674643E-01+8.36337993E-01j, 1.76098908E-02+1.74391779E-02j],
[ 1.51473418E-03+1.07178057E-03j, 6.40793740E-04-1.94176372E-03j,-1.28408900E-02+2.66263921E-02j, 4.84726807E-02-3.84341687E-03j]

In cases where $\hat{A}$ of dimension $n$ is obtained from projecting down a unitary matrix $\hat{U}_m$ of a larger dimension ($\hat{A} = \hat{P} \hat{U}_m \hat{P}$, where $\hat{P}$ is the projector from dimension $m$ to dimension $n$) , one possible way to quantify the "distance from unitarity" is $$d_u(\hat{A}) = 1 - \frac{1}{n} \trace[\hat{A}^\dagger \hat{A}].$$

This situation is common in quantum information, where $\hat{A}$ is the projection of the unitary evolution of a large Hilbert space of dimension $m$ into a small logical subspace of dimension $n$. The quantity $d_u(\hat{A})$ is then simply the population lost from logical subspace.

In [3]:
def delta_uni(A):
    """ Assuming that A is a matrix obtained from projecting a unitary matrix
        to a smaller subspace, return the loss of population of subspace, as a
        distance measure of A from unitarity.
        Result is in [0,1], with 0 if A is unitary.
    return 1.0 - (A.H * A).trace()[0,0].real / A.shape[0]

For the given matrix, we lose about 80% of the population in the logical subspace:

In [4]:

A $\hat{U}$ that minimizes $\Vert \hat{A} - \hat{U}\Vert$ can be calculated via a singular value decomposition (SVD) [Reich]: $$\hat{A} = \hat{V} \hat{\Sigma} \hat{W}^\dagger,$$ $$ \hat{U} = \hat{V} \hat{W}^\dagger.$$

In [5]:
def closest_unitary(A):
    """ Calculate the unitary matrix U that is closest with respect to the
        operator norm distance to the general matrix A.

        Return U as a numpy matrix.
    V, __, Wh = scipy.linalg.svd(A)
    U = np.matrix(
    return U
In [6]:
U = closest_unitary(A)

The SVD also allows to calculate the distance that $\hat{A}$ has from $\hat{U}$. $$ d(\hat{A}, \hat{U}) = \max_i \vert \sigma_i - 1\vert, $$ where $\sigma_i$ are the diagonal entries of $\hat{\Sigma}$ from the SVD. This is a more general measure of "distance from unitarity" than delta_uni

In [7]:
def deltaU(A):
    """Distance to the closes unitary.

    Calculate the operator norm distance \Vert\hat{A} - \hat{U}\Vert between
    an arbitrary matrix $\hat{A}$ and the closest unitary matrix $\hat{U}$
    __, S, __ = scipy.linalg.svd(A)
    d = 0.0
    for s in S:
        if abs(s - 1.0) > d:
            d = abs(s-1.0)
    return d

For matrices obtained from projecting down from a larger Hilbert space, the maximum distance is 1. For general matrices, it can be larger.

In [8]:
deltaU(A) # should be close to 1, we are *very* non-unitary
In [9]:
deltaU(U) # should be zero, within machine precision

We can double check this with the implemenation of the 2-norm in SciPy:

In [10]:
scipy.linalg.norm(A-U, ord=2) # should be equal to deltaU(A)


[Reich] D.M.Reich. "Characterisation and Identification of Unitary Dynamics Maps in Terms of Their Action on Density Matrices" (unpublished)