Thermalization and Information Flow

Thermalization research asks why selected observables approach equilibrium and what information is stored in correlations or in the environment during that process. A subsystem becoming mixed is not by itself proof of thermalization. A nearly stationary local observable is not proof that the full state is Gibbsian.

Required background. Entropy production and partial trace supply the stable definitions. Closed-system ETH questions are assessed separately in thermalization and ETH.

A correlation ledger gives an exact finite-model check

Consider unitary evolution of a finite system and bath from an initially uncorrelated state $\rho_S(0)\otimes\tau_B$, where $\tau_B=e^{-\beta H_B}/Z_B$ is thermal. Set $k_B=1$ and define heat into the system by $Q=-\Delta E_B$. Then

$$\Delta S_S-\beta Q =I(S:B)_t+D(\rho_B(t)\Vert\tau_B).$$

Here $I$ is mutual information and $D$ is quantum relative entropy. The identity follows from conserved joint entropy, the initially product state, and the Gibbs logarithm. Both terms on the right are nonnegative. It uses bath-energy bookkeeping; it does not imply $Q=\Delta E_S$ when interaction energy or external work matters.

Established: under these assumptions, local entropy change, generated correlations, and bath disturbance fit a consistent ledger. Landi and Paternostro review this connection. Active: recovering useful parts of that information from restricted observations, extending the account to initial correlations or realistic reservoirs, and distinguishing true equilibration from a finite-time plateau.

The literature check date is 2026-10-11; the selected review evidence extends through 2025, including Fazio et al.’s open many-body treatment. The displayed finite-model identity is a baseline, not a resolution of generic many-body thermalization.

Local relaxation hides several different mechanisms

A small subsystem can appear stationary because information has spread into inaccessible correlations. A finite bath can later return that information, producing recurrences. An approximately Markovian limit suppresses the relevant memory over the observation window. A conserved quantity or symmetry can prevent relaxation to the naive thermal ensemble.

Choose an observable whose alternatives differ: population relaxation, coherence, a two-time correlation, or response to an intervention. An effective temperature inferred from one population ratio need not predict another observable. Likewise, entropy growth can result from dephasing without energy exchange.

Starter project: close the entropy ledger and find its limits

Use a qubit coupled to a small spin bath initially thermal with respect to its specified bath Hamiltonian. Evolve the joint density matrix exactly and compute system entropy, bath energy, mutual information, and bath relative entropy. Check the identity above numerically and recover the no-coupling limit.

Vary bath size within available memory, coupling strength, and observation time. Plot local relaxation beside recurrences and the two information terms. Use eigenvalue cutoffs carefully in logarithms and test positivity and trace preservation; numerical entropy errors near zero eigenvalues need separate attention.

The deliverable is a reproducible ledger and an explicit time window over which a simpler reduced model is adequate. A failure test is an apparent negative total entropy production caused by inconsistent heat signs, an initially correlated state, or numerical truncation. Do not repair such a discrepancy by redefining the quantity after inspecting the result.

Research directions

An experimental extension asks which lower bounds on the information terms can be inferred from accessible observables without full bath tomography. A theoretical extension asks how an initial-correlation correction changes a proposed monotonicity statement. A computational extension compares a controlled reduced model with exact dynamics before moving to a larger bath.

A useful result specifies its observables and resolution: predicting local populations may be much easier than predicting multi-time statistics. Non-Markovian dynamics owns the intervention-based memory question. Reassess a thermalization claim when a new observable, longer time window, or larger system reveals a previously hidden conserved or recurrent structure.

Entry points and references

Read the entropy-production review for the information ledger and the open many-body notes for scaling and dissipative dynamics. The proposed small-bath project is a training baseline; its extension requires a current novelty check.

References

• R. Fazio, J. Keeling, L. Mazza, and M. Schirò, “Many-Body Open Quantum Systems,” SciPost Physics Lecture Notes 99 (2025). DOI.
• G. T. Landi and M. Paternostro, “Irreversible entropy production: From classical to quantum,” Reviews of Modern Physics 93, 035008 (2021). DOI.