---
title: Trace-Dobrushin Lyapunov Exponent
url: https://www.emergentmind.com/topics/trace-dobrushin-lyapunov-exponent
type: topic
---

# Trace-Dobrushin Lyapunov Exponent

Searching arXiv for the cited paper and closely related context papers.
The trace-Dobrushin Lyapunov exponent is a product-level asymptotic invariant for stationary random cocycles of finite-dimensional quantum channels, introduced in the study of trace-norm forgetting and asymptotic replacement for products of completely positive trace-preserving (CPTP) maps. It is defined from the decay of the centered trace-Dobrushin coefficient of channel products, where the latter measures the exact trace-norm contraction on the self-adjoint trace-zero subspace. In the framework developed in "Asymptotic Replacement for Quantum Channel Products with Applications to Inhomogeneous Matrix Product States" [2605.00157], the exponent governs quenched memory loss, forward and pullback convergence to a unique dynamically stationary random replacement channel, and quantitative consequences for deterministic and stationary random inhomogeneous matrix product states (MPS).

## 1. Centered trace-Dobrushin framework

The exponent is built on the centered trace-Dobrushin coefficient of a positive trace-preserving map $\Phi : M_d \to M_d$, where $M_d$ denotes the space of $d \times d$ complex matrices equipped with the trace norm $\|X\|_1 := \operatorname{Tr}|X|$, and
$$
H := \{X \in M_d^{sa} : \operatorname{Tr}X = 0\}.
$$
For such a map,
$$
D_c(\Phi) := \sup\left\{\frac{\|\Phi(X)\|_1}{\|X\|_1} : X \in H,\ X \neq 0\right\},
$$
which is the induced $1 \to 1$ operator norm of $\Phi$ restricted to $(H,\|\cdot\|_1)$ [2605.00157].

Equivalent variational characterizations identify $D_c(\Phi)$ as
$$
D_c(\Phi)
= \sup_{\rho \neq \sigma \in S_d}\frac{\|\Phi(\rho)-\Phi(\sigma)\|_1}{\|\rho-\sigma\|_1}
= \frac12 \sup_{\rho,\sigma \in S_d}\|\Phi(\rho)-\Phi(\sigma)\|_1
= \frac12 \sup \{\|\Phi(E-F)\|_1 : E,F \text{ orthogonal rank-one projections}\},
$$
with $0 \le D_c(\Phi) \le 1$. Moreover, $D_c(\Phi)=0$ if and only if $\Phi$ is a replacement channel of the form $X \mapsto \operatorname{Tr}(X)\tau$ for some state $\tau$. Thus $2D_c(\Phi)$ is the trace-norm diameter of the output state space $\Phi(S_d)$.

For a sequence of channels $(\Phi_n)$, the chronological product is defined for $s<t$ by
$$
\Phi_{t:s}:=\Phi_{t-1}\circ \Phi_{t-2}\circ \cdots \circ \Phi_s.
$$
Then $D_c(\Phi_{t:s})$ is the residual trace-norm memory after the block, since
$$
\sup_{\rho,\sigma}\|\Phi_{t:s}(\rho)-\Phi_{t:s}(\sigma)\|_1 = 2D_c(\Phi_{t:s}).
$$
This centered viewpoint is operational: trace preservation leaves the hyperplane $\operatorname{Tr}X=0$ invariant, and all state differences $\rho-\sigma$ lie in that hyperplane. The coefficient therefore quantifies residual dependence on the input state rather than mere contractivity of the channel on the full operator space.

A central structural fact is submultiplicativity:
$$
D_c(S\circ T)\le D_c(S)D_c(T)
$$
for positive trace-preserving maps $S$ and $T$. Consequently, for $s<u<t$,
$$
D_c(\Phi_{t:s}) \le D_c(\Phi_{t:u})D_c(\Phi_{u:s}),
$$
which yields logarithmic subadditivity along products. This product-level property is the basis for the Lyapunov exponent.

## 2. Definition and ergodic formulation

The trace-Dobrushin Lyapunov exponent is defined for a stationary random CPTP cocycle over an invertible measure-preserving system $(\Omega,\mathcal{F},\mathbb{P},\theta)$, with a measurable assignment $\omega \mapsto \Phi_\omega$ of CPTP maps on $M_d$. For $n\ge 1$,
$$
\Phi_\omega(n):=\Phi_{\theta^{n-1}\omega}\circ \cdots \circ \Phi_\omega,
\qquad
\kappa_n(\omega):=D_c(\Phi_\omega(n)),
\qquad
\kappa_0(\omega):=1.
$$
Writing
$$
F_n(\omega):=\log \kappa_n(\omega)\in[-\infty,0],
$$
submultiplicativity gives
$$
\kappa_{n+m}(\omega)\le \kappa_n(\theta^m\omega)\kappa_m(\omega),
$$
hence
$$
F_{n+m}(\omega)\le F_n(\theta^m\omega)+F_m(\omega).
$$
The trace-Dobrushin Lyapunov exponent is then
$$
\lambda_{TD}(\omega):=\lim_{n\to\infty}\frac{1}{n}\log \kappa_n(\omega)
=\lim_{n\to\infty}\frac{1}{n}F_n(\omega),
$$
whenever the limit exists [2605.00157].

By Kingman’s subadditive ergodic theorem, the limit exists almost surely and defines a $\theta$-invariant random variable $\lambda_{TD}\in[-\infty,0]$. If $\theta$ is ergodic, then $\lambda_{TD}$ is almost surely constant and satisfies
$$
\lambda_{TD}=\inf_{n\ge 1}\frac{1}{n}\,\mathbb{E}[\log \kappa_n].
$$
The same asymptotic quantity arises from pullback products over the past interval $[-n,0]$,
$$
\Phi_{\omega;0:-n}:=\Phi_{\theta^{-1}\omega}\circ \Phi_{\theta^{-2}\omega}\circ \cdots \circ \Phi_{\theta^{-n}\omega},
$$
with
$$
D_c(\Phi_{\omega;0:-n})=\kappa_n(\theta^{-n}\omega),
$$
so the forward and pullback formulations share the same almost-sure Lyapunov exponent.

Conceptually, $\lambda_{TD}$ records the asymptotic exponential decay rate of the exact residual memory coefficient of random channel products. Because it is defined from $D_c$ rather than from a coarser one-step sufficient criterion, it is an intrinsic product-level quantity. This suggests a noncommutative analogue of the role played by classical Dobrushin contraction rates for random Markov products, but the framework is formulated directly in trace norm on the traceless subspace.

## 3. Negativity, forgetting, and asymptotic replacement

The decisive threshold is the almost sure negativity of $\lambda_{TD}$. In the stationary random setting, the following are equivalent: $\lambda_{TD}(\omega)<0$ for $\mathbb{P}$-almost every $\omega$; $\kappa_{\omega;n:0}\to 0$ almost surely, where $\kappa_{\omega;n:0}:=D_c(\Phi_{\omega;n:0})$; and $\kappa_{\omega;0:-n}\to 0$ almost surely. These equivalences identify negative trace-Dobrushin Lyapunov exponent with quenched forward forgetting and quenched pullback forgetting [2605.00157].

This equivalence is the random counterpart of a deterministic replacement theorem. For fixed $s$, decay $D_c(\Phi_{t:s})\to 0$ is equivalent to strong asymptotic replacement from $s$: there exist states $\eta_{t:s}$ such that the replacement channels
$$
R_{t:s}(X):=\operatorname{Tr}(X)\eta_{t:s}
$$
satisfy
$$
\|\Phi_{t:s}-R_{t:s}\|_{1\to 1}\to 0.
$$
Equivalently, for every $\tau\in S_d$,
$$
\|\Phi_{t:s}-R^{(\tau)}_{t:s}\|_{1\to 1}\to 0,
\qquad
R^{(\tau)}_{t:s}(X):=\operatorname{Tr}(X)\Phi_{t:s}(\tau).
$$
The basic quantitative comparison is
$$
D_c(\Phi_{t:s})\le \|\Phi_{t:s}-R^{(\tau)}_{t:s}\|_{1\to 1}\le 4D_c(\Phi_{t:s}).
$$

In the two-sided deterministic setting, pullback forgetting produces a unique boundary state sequence $(\rho_t)_{t\in\mathbb{Z}}$ with $\rho_{t+1}=\Phi_t(\rho_t)$ such that, for each fixed $t$,
$$
\|\Phi_{t:s}-R_t\|_{1\to 1}\to 0
\quad \text{as } s\to -\infty,
\qquad
R_t(X):=\operatorname{Tr}(X)\rho_t.
$$
Moreover,
$$
\rho_t=\lim_{s\to-\infty}\Phi_{t:s}(\tau)
$$
for every $\tau\in S_d$. In the random stationary setting, negative $\lambda_{TD}$ yields the corresponding random object: a unique dynamically stationary random state $\rho:\Omega\to S_d$ satisfying
$$
\Phi_\omega(\rho_\omega)=\rho_{\theta\omega}
\quad \text{almost surely}.
$$
The associated random replacement channel is
$$
R_\omega(X):=\operatorname{Tr}(X)\rho_\omega.
$$

At the state level, the bounds become
$$
\sup_\sigma \|\Phi_{\omega;n:0}(\sigma)-\rho_{\theta^n\omega}\|_1 \le 2\kappa_{\omega;n:0},
\qquad
\sup_\sigma \|\Phi_{\omega;0:-n}(\sigma)-\rho_\omega\|_1 \le 2\kappa_{\omega;0:-n},
$$
while at the operator level,
$$
\|\Phi_{\omega;n:0}-R_{\theta^n\omega}\|_{1\to 1} \le 4\kappa_{\omega;n:0},
\qquad
\|\Phi_{\omega;0:-n}-R_\omega\|_{1\to 1} \le 4\kappa_{\omega;0:-n}.
$$
Accordingly, the exponent does not merely detect convergence of a preferred state; it governs convergence of the entire channel product to a rank-one replacement map.

## 4. Quenched and annealed rates

When $\lambda_{TD}<0$ almost surely, the product coefficients decay exponentially along almost every realization. More precisely, there exist a measurable $\theta$-invariant random variable $\beta:\Omega\to(-\infty,0)$, a full-measure $\theta$-invariant set $\Omega^\ast$, and almost surely finite measurable random variables $C_{+\beta}$ and $C_{-\beta}$ such that for all $\omega\in\Omega^\ast$ and all $n\ge 1$,
$$
\|\Phi_{\omega;n:0}-R_{\theta^n\omega}\|_{1\to 1}
\le 4C_{+\beta}(\omega)e^{\beta(\omega)n},
$$
and
$$
\|\Phi_{\omega;0:-n}-R_\omega\|_{1\to 1}
\le 4C_{-\beta}(\omega)e^{\beta(\omega)n}.
$$
The corresponding state-level bounds hold with factor $2$ instead of $4$ [2605.00157].

These are quenched statements: they hold almost surely for each environment. Annealed estimates require additional assumptions on the channel environment. Defining
$$
\Delta^{op}_n(\omega)
:=
\max\left\{
\|\Phi_{\omega;n:0}-R_{\theta^n\omega}\|_{1\to 1},
\|\Phi_{\omega;0:-n}-R_\omega\|_{1\to 1}
\right\},
$$
the following bounds hold. If the environment is $\varrho$-mixing in the maximal correlation sense, with $\rho_m\to 0$, then for every $p\in\mathbb{N}$ there exists $C_p<\infty$ such that
$$
\mathbb{E}[\Delta^{op}_n]\le C_p n^{-p}.
$$
If the one-step channels are jointly independent, then there exist $C<\infty$ and $\gamma>0$ such that
$$
\mathbb{E}[\Delta^{op}_n]\le Ce^{-\gamma n}.
$$

The deterministic theory contains analogous quantitative clocks. If $\kappa_{tr}(\Phi_j)\le 1-a_j$ with $a_j\in[0,1]$, then
$$
D_c(\Phi_{t:s})\le \prod_{j=s}^{t-1}(1-a_j)\le \exp\!\left(-\sum_{j=s}^{t-1}a_j\right).
$$
If $G_r(s,t)$ counts indices $j$ with $a_j\ge r\in(0,1)$, then
$$
D_c(\Phi_{t:s})\le (1-r)^{G_r(s,t)}.
$$
A further good-block criterion states that if, for some $\ell$, $M$, and $q\in(0,1)$, every window $[r,r+M)$ contains a subblock $[u,u+\ell)$ with $D_c(\Phi_{u+\ell:u})\le q$, then
$$
D_c(\Phi_{t:s})\le q^{\lfloor (t-s)/M\rfloor},
$$
yielding exponential bounds at rate $|\log q|/M$. The significance of this statement is explicit in the source: multi-step contraction may be present even when every one-step coefficient equals $1$.

## 5. Criteria, comparisons, and examples

Several sufficient criteria certify $\lambda_{TD}<0$ without requiring strict positivity at every step. In the ergodic case, if there exists $L\ge 1$ with
$$
\mathbb{P}\{\omega : D_c(\Phi_{\omega;L:0})<1\}>0,
$$
then $\lambda_{TD}<0$. If, with positive probability, the block $\Phi_{\omega;L:0}$ is strictly positive, then $\lambda_{TD}<0$ as well. A Doeblin-type block minorization also suffices: if for some $L$ there exist $\varepsilon(\omega)\ge 0$ and $\tau_\omega\in S_d$ such that
$$
\Phi_{\omega;L:0}(X)\ge \varepsilon(\omega)\operatorname{Tr}(X)\tau_\omega
\quad (\forall X\ge 0),
$$
and $\mathbb{E}[\log(1-\varepsilon)]<0$ in the extended sense, then $\lambda_{TD}<0$; it is enough that $\mathbb{P}\{\varepsilon>0\}>0$. Likewise, a Markov-Dobrushin block lower bound with
$$
B_\omega \le \Phi_{\omega;L:0}(P)
\quad \text{for all pure } P
$$
and $\mathbb{E}[\log(1-\operatorname{Tr}(B_\omega))]<0$ implies negativity [2605.00157].

The framework is explicitly related to classical Dobrushin coefficients. For a classical Markov chain with transition matrix $P$, the classical coefficient is half the total-variation diameter of the image of the probability simplex, equivalently the contraction factor on the centered $\ell^1$ subspace. The quantum coefficient $D_c(\Phi)$ is the exact noncommutative analogue:
$$
D_c(\Phi)=\frac12 \sup_{\rho,\sigma}\|\Phi(\rho)-\Phi(\sigma)\|_1,
$$
and equals the sharp one-step trace-distance contraction coefficient $\eta_{Tr}(\Phi)$. The paper contrasts this exact product-level quantity with one-step sufficient criteria such as Markov-Dobrushin and quantum Doeblin minorization; for a CPTP map $\Phi$,
$$
D_c(\Phi)\le 1-\alpha_{MD}(\Phi),
$$
and the bound can be strict.

Three examples clarify what the exponent does and does not encode. In alternating dephasing, $ \kappa_{tr}(D_Z)=\kappa_{tr}(D_X)=1$ but $\kappa_{tr}(D_X\circ D_Z)=0$, so forgetting can be a genuinely multi-step effect. For the qubit amplitude-damping channel $\Gamma_\gamma$ with $0<\gamma<1$,
$$
D_c(\Gamma_\gamma)=\sqrt{1-\gamma}<1
\quad \text{but} \quad
\alpha_{MD}(\Gamma_\gamma)=0,
$$
and the powers $\Gamma_\gamma^n$ converge to a pure replacement channel, giving
$$
\lambda_{TD}=\log \sqrt{1-\gamma}<0.
$$
This shows replacement mixing without eventual strict positivity. Conversely, there exists a deterministic channel with unique fixed point $\tfrac12 I$ but $\kappa_{tr}(\Phi^n)=1$ for all $n$, so uniqueness of a stationary state does not imply $\lambda_{TD}<0$.

For unital CPTP maps, Appendix B provides a Hilbert-Schmidt sufficient criterion. If $s_0(T)$ is the largest singular value on the complex trace-zero Hilbert-Schmidt subspace, then
$$
\kappa_{\omega;n:0}\le \sqrt{d}\prod_{j=0}^{n-1}s_0(\Phi_{\theta^j\omega}),
$$
hence
$$
\lambda_{TD}\le \limsup \frac1n \sum \log s_0.
$$
If $\mathbb{E}[\log s_0(\Phi_\omega)]<0$ (or the corresponding block condition holds), then $\lambda_{TD}<0$ and the convergence is to the completely depolarizing replacement channel
$$
R(X)=\operatorname{Tr}(X)\, I/d.
$$

## 6. Role in inhomogeneous matrix product states

The trace-Dobrushin Lyapunov exponent enters the theory of inhomogeneous MPS through the left-canonical CPTP gauge. In the Schrödinger picture, placing MPS tensors in left-canonical form means
$$
\sum_i K_i[n]^\ast K_i[n]=I,
$$
so the auxiliary transfer map
$$
\Phi_n(\rho):=\sum_i K_i[n]\rho K_i[n]^\ast
$$
is CPTP. This places the auxiliary dynamics exactly in the centered trace-Dobrushin framework, with product coefficients controlling boundary stability, thermodynamic limits, and correlations [2605.00157].

In the deterministic setting, let
$$
\Theta_{m,n}:=\Phi_{m+1}\circ \cdots \circ \Phi_n.
$$
If for every fixed $q$, $D_c(\Theta_{q,n})\to 0$ as $n\to\infty$, then there exists a unique right-boundary sequence $\{\rho_r\}_{r\ge 1}$ with
$$
\rho_r=\Phi_r(\rho_{r+1}),
$$
and a unique infinite-volume state $\varphi_\infty$ on the quasi-local algebra such that for local $X\in \mathcal{A}[1,m]$,
$$
\varphi_\infty(X)=\operatorname{Tr}(\widetilde{X}_{[1,m]}(\rho_{m+1})).
$$
If $\Psi_n$ is the trace-closed finite-volume MPS vector and $\varphi_n$ the normalized finite-volume state whenever $\langle \Psi_n,\Psi_n\rangle\neq 0$, then for fixed local $X\in\mathcal{A}[1,m]$,
$$
\varphi_n(X)\to \varphi_\infty(X).
$$
Quantitatively, for all large $n$,
$$
|\varphi_n(X)-\varphi_\infty(X)|
\le 16D_H^2\|X\|_\infty D_c(\Theta_{m,n}),
$$
where $D_H=\dim H$.

The same coefficient governs spatial clustering. If $A\in \mathcal{A}[p,q]$ and $B\in \mathcal{A}[r,s]$ with a gap $q+1<r$, then
$$
|\varphi_\infty(AB)-\varphi_\infty(A)\varphi_\infty(B)|
\le 4\|A\|_\infty \|B\|_\infty D_c(\Theta_{q,r-1}).
$$
Thus the auxiliary product coefficient across the gap is the direct control parameter for correlation decay.

In the stationary random setting, the right-tail MPS product becomes a pullback product,
$$
\Theta_{\omega;m,n}=\Phi_{\theta^{-m}\omega;0:-(n-m)}.
$$
If $\lambda_{TD}<0$ almost surely, then there is a unique stationary random boundary state $\rho_\omega$ with
$$
\Phi_\omega(\rho_\omega)=\rho_{\theta\omega},
$$
and a random infinite-volume state
$$
\varphi_{\infty,\omega}(X)
=
\operatorname{Tr}(\widetilde{X}_{\omega;[1,m]}(\rho_{\theta^{-m}\omega}))
$$
for local $X$. For fixed local $X\in\mathcal{A}[1,m]$,
$$
|\varphi_{n,\omega}(X)-\varphi_{\infty,\omega}(X)|
\le
16D_H^2\|X\|_\infty C_{-\beta}(\theta^{-m}\omega)e^{\beta(\omega)(n-m)}
$$
for all large $n$, with $\beta(\omega)<0$ and $C_{-\beta}(\omega)$ finite almost surely. Correlations also cluster exponentially:
$$
|\varphi_{\infty,\omega}(AB)-\varphi_{\infty,\omega}(A)\varphi_{\infty,\omega}(B)|
\le
4\|A\|_\infty\|B\|_\infty C_{-\beta}(\theta^{-q}\omega)e^{\beta(\omega)L},
$$
where $L=r-q-1$. Annealed high-probability and expectation bounds follow from the channel-product estimates: under $\varrho$-mixing, super-polynomial; under independence, exponential in the gap.

A plausible implication is that the trace-Dobrushin Lyapunov exponent functions as the asymptotic boundary-stability rate for random left-canonical MPS with CPTP auxiliary transfer maps. In the source formulation, the same product coefficient that measures trace-norm forgetting for the auxiliary channels also governs infinite-volume limits and spatial clustering.

## 7. Scope, assumptions, and interpretation

The theory is formulated in finite dimension throughout, for CPTP maps in the Schrödinger picture. Measurability of $\omega\mapsto \Phi_\omega$ is assumed, and for Kingman’s theorem no integrability beyond $F_n^+=0$ is needed. Ergodicity of $\theta$ is used only to make $\lambda_{TD}$ almost surely constant and for certain positive-probability criteria; nonergodic variants are handled by two-sided saturation arguments. Mixing assumptions such as $\varrho$-mixing or independence are required only for annealed bounds, not for quenched convergence [2605.00157].

For MPS applications, the left-canonical CPTP gauge is the structural condition ensuring that the auxiliary transfer maps lie in the trace-preserving centered framework. No primitivity or strict positivity is assumed. Eventual strict positivity is sufficient for $\lambda_{TD}<0$, but the amplitude-damping example shows that it is not necessary.

Within this framework, the trace-Dobrushin Lyapunov exponent has a precise interpretation. It is not merely a spectral quantity attached to a single channel, nor simply a certificate of existence of a stationary state. Rather, it is the almost-sure asymptotic exponential rate of the exact residual input-state dependence of random channel products. Negative exponent means that the state-space diameter of the evolved image collapses exponentially, and this collapse is equivalent to convergence of the full product to a unique dynamically stationary random replacement channel. In that sense, the exponent is the random product-level invariant that links centered trace-norm contraction, asymptotic replacement, and boundary formation in inhomogeneous MPS.

Source: https://www.emergentmind.com/topics/trace-dobrushin-lyapunov-exponent