REVIEW
Renormalization group and long-range conditional mutual information in hierarchical models
Quantum Physics · Submitted 2 Sep 2026 · 110 references
Pith's verdict: Well-crafted paper: clean theorem linking ideal RG to bounded CMI, plus exactly solvable models that test the limits. Honest about assumptions. Deserves a serious referee.
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
The paper proves that if a mixed quantum state admits an ideal renormalization group—an exactly reversible coarse-graining by local channels with fixed on-site Hilbert space dimension—then its conditional mutual information is bounded…
The reading
What carries the argument
The central object is the ideal (locally reversible) RG scheme: a sequence of coarse-graining maps $ ho^{(r)} o ho^{(r+1)}$, each a finite-depth circuit of local channels that are correlation-preserving, $I_{ ho'}(X':ar{X}) = I_{ ho}(X:ar{X})$, a condition equivalent to each coarse-graining step being exactly undone by a fine-graining channel acting only on the coarse-grained region. The proof's workhorse is the bounded causal-tube property: because each RG layer is local at its own scale and shrinks linear size by a factor $b>1$, the causal closure of a fixed-width boundary collar is a multiscale tube of uniformly bounded transverse width containing $O(1)$ sites per horizontal cut. Strong subadditivity enters twice—enlarging regions to absorb the tubes can only increase the CMI, and the final bound is $2 s{min} s{ln} s{dim} A''$, $ s{log} s{dim} C''\} s{le} 2n_* s{log} d_*$. For the tree model the machinery is the posterior probability $q(S) = P( s{sigma}=+1|S)$ of the hidden parent spin, which the paper shows is exactly the variable that preserves correlations and obeys an exact recursion $E_{ s{ideal}}(q_1, s{ldots},q_6)$; since $q$ is continuous, the fixed-dimension hypothesis of the theorem fails there, and UV-finiteness is instead proven directly by a strong-subadditivity argument giving $I(A:C|B) s{le} s{log} 2$.
Extended reading notes
Core claim
On the paper's own terms, the core result is a theorem plus two model constructions. The theorem: for fixed-shape annular tripartitions $A|B|C$ in $d$ dimensions, if a state $ ho$ admits an ideal RG—a finite-depth circuit of local correlation-preserving channels at each scale, each step exactly undoable by a channel acting only on the coarse-grained region, with fixed finite on-site dimension $d_*$—then the CMI is bounded by $2 n_* \log d_*$ with $n_*=O(1)$ sites, uniformly in the scaling limit $N_A/N_B o x_A$, $N_C/N_B o x_C$, $N_B o \infty$. The proof removes the RG gates in the interface causal tubes, coarse-grains the bulk without changing the CMI, enlarges regions so each tube lies inside a single region (which can only increase the CMI by strong subadditivity), then completes the tubes and bounds the remainder by the dimension of the surviving $O(1)$ coarse sites. The first model, a majority-rule variant of Dyson's hierarchical Ising model, implements an exactly locally reversible RG at every scale with $T o 3^{1/ u} T$, flowing to the infinite-temperature product state, yet its CMI decays only algebraically in the buffer size, so the Markov length diverges at every finite $T$; sublinear-range circuits connect it two-way to the product state, while polylogarithmic-range circuits cannot. The second model, a tree concatenating the classical SSB repetition rule with the SWSSB parity rule (the classical counterpart of the Bell tree), has vanishing local CMI for $N_A/N_B < 2$ but global CMI equal to $ ext{ln} 2$ for any subextensive buffer; it carries a strong Z$_2$ symmetry yet is stable against noise, in contrast to SWSSB, and at the critical noise strength $p_c acksimeq 0.2802$ the local CMI decays as $N^{- s{alpha}}$ with $ s{alpha} acksimeq 1.5$ while the two-point mutual information vanishes identically.
Load-bearing premise
Everything rests on assuming the state admits an ideal RG at all: a coarse-graining by local channels that can be exactly reversed, on that state, by local channels acting only on the coarse-grained region, with the on-site Hilbert space dimension never growing; the paper constructs its two models precisely to satisfy this condition and does not prove existence for generic states.
Editorial extensions
If this is right
- CMI becomes a genuine long-distance property: given an ideal RG it is invariant when all three regions are rescaled together, so the scaling limit is described by a scaling function $f(N_A/N_B, N_C/N_B, \{t_i N_B^{1/(\nu_i d)}\})$ without finite-size corrections.
- A divergent Markov length does not block locally reversible RG flow: the majority-rule hierarchy connects to the infinite-temperature product state in both directions by sublinear-range circuits, yet is not phase-equivalent to it under the polylogarithmic-range circuit definition, so the two phase classifications genuinely differ.
- Stability of mixed-state phases is tied to the global, not local, Markov property: the SSB–SWSSB tree breaks the global Markov condition while satisfying the local one and is stable against noise, unlike SWSSB, and strong subadditivity then forces the global CMI to be exactly $\log 2$ below $p_c$ and exactly 0 above it.
- The tree's critical point is a one-dimensional example where the two-point mutual information vanishes identically while the CMI decays only polynomially, a pattern not present at equilibrium criticality, where mutual information decays algebraically and CMI is short-ranged.
Reading between the lines
- The paper's learnability discussion points toward a testable prediction: a state whose CMI decays polynomially but whose RG flows to a trivial fixed point (the majority-rule model) should be easier to learn from measurements than a genuinely nonequilibrium critical point with the same Markov-length behavior; whether a single estimator separates the two is left open.
- Read as a resource theory, the theorem says any RG that must branch into multiple copies—Fermi surfaces with branching MERA, or fracton systems—escapes the bound precisely because the effective on-site dimension grows; quantifying a branching number per RG step would extend the UV-finiteness criterion to those states.
- A natural extension is to discretize the continuous posterior variable $q(S)$ of the ideal tree RG onto a finite alphabet; if a finite-alphabet approximation remains approximately correlation-preserving, the theorem's fixed-dimension hypothesis would apply directly to the noisy tree, and the approximation error could be compared with the critical noise strength $p_c$.
- The sharp plateau structure of the global CMI ($\log 2$ for $p < p_c$, 0 for $p > p_c$) acts as an order parameter for a transition invisible to local observables and to two-point correlations, suggesting that CMI measurements on subextensive buffers could serve as a practical detector of the phase boundary.
Original abstract
A departure of a mixed quantum state from a local Gibbs description is generally invisible to local observables but can be detected by the conditional mutual information (CMI). Here we study the relationship between the renormalization group (RG) and CMI, and in particular, how RG constrains CMI. We first show that the CMI between nonadjacent regions $A$ and $C$, conditioned on the buffer region $B$, is UV-finite whenever the state admits a locally reversible RG with a fixed on-site Hilbert space dimension. We then study two hierarchical models that have long-range CMI and yet admit a simple RG description. The first model has a divergent Markov length at every temperature $0<T<\infty$ but nevertheless flows to an infinite-temperature product state under RG. The second model satisfies the local Markov condition while violating the global one and is stable against weak noise. At the critical noise strength, the two-point CMI decays only polynomially as a function of the system size, while the two-point mutual information vanishes.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
A departure of a mixed quantum state from a local Gibbs description is generally invisible to local observables but can be detected by the conditional mutual information (CMI). Here we study the relationship between the renormalization group (RG) and CMI, and in particular, how RG constrains CMI. We first show that the CMI between nonadjacent regions $A$ and $C$, conditioned on the buffer region $B$, is UV-finite whenever the state admits a locally reversible RG with a fixed on-site Hilbert space dimension. We then study two hierarchical models that have long-range CMI and yet admit a simple RG description. The first model has a divergent Markov length at every temperature $0<T<\infty$ but nevertheless flows to an infinite-temperature product state under RG. The second model satisfies the local Markov condition while violating the global one and is stable against weak noise. At the critical noise strength, the two-point CMI decays only polynomially as a function of the system size, while the two-point mutual information vanishes.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | A departure of a mixed quantum state from a local Gibbs description is generally invisible to local observables but can be detected by the conditional mutual information (CMI). | background 1.00 | 0.08 | 0.08 | asserted 0.92 |
| s1 | Here we study the relationship between the renormalization group (RG) and CMI, and in particular, how RG constrains CMI. | contribution 0.97 | 0.75 | 0.08 | asserted 0.58 |
| s2 | We first show that the CMI between nonadjacent regions $A$ and $C$, conditioned on the buffer region $B$, is UV-finite whenever the state admits a locally reversible RG with a fixed on-site Hilbert space dimension. | result 0.98 | 0.93 | 0.30 | asserted 1.00 |
| s3 | We then study two hierarchical models that have long-range CMI and yet admit a simple RG description. | contribution 0.85 | 0.91 | 0.10 | asserted 0.98 |
| s4 | The first model has a divergent Markov length at every temperature $0<T<\infty$ but nevertheless flows to an infinite-temperature product state under RG. | result 0.99 | 0.87 | 0.10 | asserted 1.00 |
| s5 | The second model satisfies the local Markov condition while violating the global one and is stable against weak noise. | result 0.99 | 0.84 | 0.09 | asserted 1.00 |
| s6 | At the critical noise strength, the two-point CMI decays only polynomially as a function of the system size, while the two-point mutual information vanishes. | result 1.00 | 0.85 | 0.15 | asserted 1.00 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Figures reproduced from arXiv:2609.02141 by the authors.
References
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.