I have a completely positive map $T$ and a sequence of $d\times d$ states $S_1,S_2,\ldots$ obtained by applying $T$ repeatedly to the identity matrix.

I'm interested in quantifying what happens to purity of $S_i$ as $i$ grows. In particular, is it possible to quantify how fast the purity grows with $i$ and what it converges to, from the spectrum of the Choi matrix?

Toy example, taking channel $T$ with the following Kraus operators $$ \left( \begin{array}{cc} 0 & \frac{1}{\sqrt{2}} \\ \frac{1}{\sqrt{2}} & 0 \\ \end{array} \right),\left( \begin{array}{cc} 0 & 0 \\ 0 & \frac{1}{\sqrt{2}} \\ \end{array} \right)$$

I'm seeing Choi matrix have eigenvalues $1,\frac{1}{2}$

And the purity over time below:


  • 1
    $\begingroup$ Interesting question. Note that you shouldn't expect the purity to increase (or decrease) in general; any unital channel will leave the purity exactly where it is in this case. $\endgroup$
    – forky40
    Dec 22, 2023 at 0:38
  • $\begingroup$ Isn't purity the reciprocal of Schatten-2 norm? This one says unital positive maps are contractive wrt to Schatten-2 norm, so purity would grow -- arxiv.org/pdf/math-ph/0601063.pdf $\endgroup$ Dec 22, 2023 at 6:15
  • $\begingroup$ Why do you expect it to have some connection with the spectrum of the Choi matrix? Also, an unital map preserves the identity operator. $\endgroup$
    – narip
    Dec 22, 2023 at 6:18
  • 1
    $\begingroup$ @narip (ChatGPT told me there is connection) background on question here $\endgroup$ Dec 22, 2023 at 18:01
  • 1
    $\begingroup$ Another possibly useful toy example is the replacement channel $T(X) = \text{Tr}(X) \sigma$ for any state $\sigma$. Then the spectrum of the Choi matrix is proportional to the spectrum of $\sigma$. This a channel for which $T \circ T = T$ so $S_i = S_{i+1}$, but the spectrum can take on a lot of different behaviors... $\endgroup$
    – forky40
    Dec 22, 2023 at 18:56

1 Answer 1


TL;DR: For many channels, as the number $i$ of iterated applications of the channel grows, the state $T^i(\rho_0)$ converges to a fixed point of $T$. Then, the purity of $T^i(\rho_0)$ converges to the purity of a fixed point of $T$, not to an eigenvalue of its Choi matrix.

Example where purity is outside Choi spectrum

Let $\rho$ denote a quantum state with eigenvalues $p_1,\dots,p_d$ and define $\mathcal{C}_\rho(X):=\rho\,\mathrm{tr}(X)$, the state preparation channel for $\rho$. The purity of $\rho$ is $\gamma=\sum_ip_i^2$, but the eigenvalues of the Choi matrix $J(\mathcal{C}_\rho)$ coincide with those of $\rho$. But generally $\gamma\notin\{p_1,\dots,p_d\}$.

Contractive channels

By Banach fixed-point theorem quantum channels that are contractive maps, such as non-trivial depolarizing and amplitude damping channels, have exactly one fixed point $\rho_{\text{fix}}$. In this case, the purity of $T^i(\rho_0)$ tends to $\mathrm{tr}(\rho_{\text{fix}}^2)$ independently of the initial state.

Non-contractive channels

The situation is more complicated for non-contractive channels. If $T$ is non-contractive, then $T^i(\rho_0)$ may fail to converge, as happens in the case of the bit-flip channel acting on $|0\rangle$.

Nevertheless, by Schauder fixed-point theorem, every channel, contractive or otherwise, has at least one fixed-point. Moreover, when $T^i(\rho_0)$ does converge, then it necessarily converges to a fixed point$^1$. However, Banach fixed point theorem no longer applies and the fixed point may not be unique. For example, every point on the line connecting the poles of the Bloch sphere is a fixed point of the phase damping channel. In general, which fixed-point $T^i(\rho_0)$ converges to depends on $\rho_0$.

Non-contractive qubit channels

The situation is a little simpler if $T$ acts on a qubit and when $\rho_0$ is the maximally mixed state. In this case, if $T$ is non-contractive then the maximally mixed state $\frac{I}{2}$ is necessarily$^2$ among the fixed points of $T$. Therefore, the purity of $T^i\left(\frac{I}{2}\right)$ is $\frac12$.

$^1$ This can be deduced from the continuity argument similar to that in the proof of Banach fixed-point theorem: $$ \rho_*=\lim_{i\to\infty}T^i(\rho_0)=T\left(\lim_{i\to\infty}T^{i-1}(\rho_0)\right)=T(\rho_*). $$
$^2$ We can make a simple geometric argument. Suppose $T$ is non-contractive. For contradiction, it is sufficient to show that $T$ shrinks the trace distance between any pair of antipodal points. If $T$ displaces the center of the Bloch sphere $B(O,1)$ from $O$ to $A\ne O$, then more than half of the surface of the ball $B(A,1)$ lies outside of $B(O,1)$. However, the image $T(B(O,1))$ must lie within $B(A,1)$. Therefore, distance between any two antipodal points is necessarily reduced by the action of $T$.

  • $\begingroup$ Thanks for the in-depth reply. I'm wondering if there is a way to link asymptotic purity to properties of channel $T$. For instance, if $T$ is a diagonal matrix, then asymptotic purity is 1. This means that spectrum of T does not determine asymptotic purity either, so what determines it? $\endgroup$ Jan 9 at 21:20
  • $\begingroup$ Asymptotic purity is related to fixed-points of the channel, but has a connection to the eigendecomposition of a certain matrix, too. However, it is not the Choi matrix $J(T)$ of $T$, but simply the matrix $K(T)$ representing linear (super)operator $T$ in some basis. Indeed, $x$ is a fixed-point of linear operator $A$ iff $x$ is the eigenvector of $A$ corresponding to eigenvalue $1$. Every channel has eigenvalue $1$, but they differ in the corresponding eigenstate. If the maximally mixed state $I/d$ is an eigenstate of $T$ corresponding to eigenvalue $1$, then asymptotic purity is $1/d$. $\endgroup$ Jan 9 at 23:50
  • $\begingroup$ We can detect this by inspecting $K(T)$ in the Pauli basis (ordered with identity first): $I/d$ is a fixed-point if and only if the first column is $[1,0,\dots,0]$. Also, if the eigenspace $V_1$ corresponding to eigenvalue $1$ is one-dimensional, then the quantum state $\rho_*$ that spans $V_1$ is the sole fixed-point of $T$, so the asymptotic purity is the purity of $\rho_*$ (regardless of the initial state). In fact, since $1$ is the largest eigenvalue of $T$, this is a special case of the power method. $\endgroup$ Jan 9 at 23:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.