arXiv:2610.36058
Universal Scaling of Conditional Mutual Information in Classical Stochastic Dynamics
Songlei Wang, Chenguang Liang, Zhi-Cheng Yang
Abstract
Conditional mutual information (CMI) characterizes dependence between two regions conditioned on an intervening buffer. Its characteristic decay length, the Markov length, has recently been identified as a diverging length scale at mixed-state phase transitions. Here, we investigate CMI dynamics in one-dimensional classical stochastic models and propose a universal scaling form for systems whose Markov length diverges with time. We derive this form in the inactive phase of parity-conserving directed percolation, where domain-wall coarsening generates a growing length scale and a self-similar probability distribution. We introduce a Hankel-based truncation scheme that enables matrix-product-state simulations at long times and over large spatial regions. Using this method, we verify the proposed scaling both in the inactive phase and at the nonequilibrium critical point, where the dynamical critical exponent governs Markov-length growth. More broadly, the same scaling form captures cases in which collective inference sets the diverging Markov length. Despite these distinct mechanisms, the CMI between fixed microscopic regions exhibits a common scaling dimension of two. Finally, we extend this framework to temporal CMI, establishing an analogous relation between the strength of conditional dependence in time domain and its characteristic memory time.
AI Summary
The paper asks how conditional mutual information (CMI) scales when the spatial Markov length, or its temporal analogue, grows without bound in classical stochastic dynamics. For fixed microscopic endpoint regions separated by a buffer $B$, it proposes $I_t(A:C\mid B)=\ell(t)^{-2}f(|B|/\ell(t))$. An exponential tail of $f$ identifies $\ell(t)$ with the Markov length. In one-dimensional parity-conserving directed percolation, self-similar domain-wall statistics give an interval-entropy scaling form whose second finite difference yields the inverse-square CMI prefactor. In the inactive phase, simulations support $\ell(t)\sim\rho(t)^{-1}\sim t^{1/2}$. At the estimated critical branching rate $Q_c\approx0.135$, numerical collapse instead supports $\ell(t)\sim t^{1/z}$ with $z=1.741$. A length-discounted Hankel truncation scheme compresses probability iMPS using finite-word probabilities rather than Schmidt weights. In a directed-percolation benchmark, it extends reliable Markov-length extraction from approximately $t=900$ to $t=1800$. For temporal CMI in stationary one-dimensional SSEP, Gaussian fluctuating hydrodynamics and Toeplitz-determinant asymptotics predict $I(t,\infty)\sim1/(32t^2)$. Transverse tensor-network calculations agree with this prefactor. Combined with a finite-size scaling ansatz, this implies a size-prefactor exponent $\beta=2z=4$, rather than a previously estimated $3.2(1)$. The general scaling form remains a proposal rather than a theorem: its derivation uses self-similar coarsening, critical support is numerical, and the temporal prediction uses a Gaussian approximation. Finite simulation windows limit conclusions about additional DP exponents. Applicability to other transport classes, broken translation invariance, higher dimensions, and genuinely quantum dynamics remains open.
Error found: The Discussion attributes inverse-square CMI scaling to translation invariance and fixed endpoint sizes without stating the additional interval-entropy scaling assumption needed for that conclusion. The temporal coefficient 1/32 is established within a Gaussian approximation, while the finite-size relation beta = 2z additionally assumes a scaling ansatz not independently verified by the reported infinite-system calculation. For the specified two-spin endpoints, the inactive-phase Taylor expansion gives f(x)=-4h''(x), rather than the stated f(x)=-h''(x).
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 7+18 pages, 12 figures