Criticality conditions in the Derrida-Retaux model with a random number of terms

Alexey Lotnikov; Anna Kotova

REVIEW 2 major objections 5 minor 7 references

Criticality conditions in the Derrida-Retaux model with a random number of terms

T0 review · 2 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Moment-generating-function tests give sufficient conditions for positive or zero free energy in the Derrida–Retaux model.

desk verdict A genuine but narrow extension of Derrida–Retaux criticality conditions; the main theorem is plausible and mostly provable, with a real but repairable gap about exponential moments for N. read the letter →

arxiv 2502.02535 v1 pith:OBCVUQCN submitted 2025-02-04 math.PR

Alexey Lotnikov , Anna Kotova This is my paper · ORCID

classification math.PR MSC 60J8060K35

keywords Derrida–Retauxmodelrandomnumberoftermsfreeenergycriticalitymomentgeneratingfunctionsubcriticalandsupercriticalregimeshierarchicalrenormalization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper studies the Derrida–Retaux recursion $X_{n+1}=(X_n^{(1)}+\cdots+X_n^{(N_n)}-a)^+$ with a random number $N_n$ of terms, and asks when the normalized free energy $Q=\lim \mathbb{E}(X_n)/(\mathbb{E}N)^n$ is positive or zero. The authors prove sufficient conditions on the initial distribution of $X_0$ and on the distribution of $N$. The test is expressed through the function $D_0(s,m)=(m-1)sF_0'(s)-aF_0(s)$, built from the moment generating function $F_0$ of $X_0$ and the mean or maximum of $N$. If $D_0((\mathbb{E}N)^{1/a},\mathbb{E}N)>0$ then $Q>0$; if $N$ is bounded by $M$ and $D_0(1+(M-1)/a,M)<0$ then $Q=0$. These are the first criticality conditions of this form for random numbers of terms and extend the known criterion for the fixed-number case.

What carries the argument

The argument runs through the generating functions $F_n(s)=\mathbb{E}(s^{X_n})$ and their exact evolution under the recursion. The key object is the operator $D_n(s)=(m-1)sF_n'(s)-aF_n(s)$, which in expectation form is $(m-1)\mathbb{E}(X_n s^{X_n})-a\mathbb{E}(s^{X_n})$. Lemma 1 shows that when the initial value $D_0$ is positive at a suitable $s$, the corresponding expression grows exponentially in $n$; Lemma 2 shows that $Q=0$ forces the same expression to stay bounded, so positivity of the initial value implies $Q>0$. For the subcritical direction, Lemma 3 gives a monotonicity inequality for $D_n$ under the assumption that the starting value is negative and $s$ is large enough, and Lemma 4 supplies the Chebyshev-type inequality $\mathbb{E}(X_n s^{X_n}) \ge \mathbb{E}(X_n)\mathbb{E}(s^{X_n})$ that converts negativity of $D_n$ into a uniform upper bound on $\mathbb{E}(X_n)$, hence $Q=0$.

What would settle it

Take $a=1$, $P(N=2)=1$, and $X_0$ equal to $0$ or $3$ with probabilities $1/2$ each; then $F_0(2)=4.5$, $F_0'(2)=6$, so $D_0(2)=2\cdot 6-4.5=7.5>0$ and Theorem 1 predicts $Q>0$. Iterating the exact recurrence for the first $n$ steps and checking whether $\mathbb{E}(X_n)/2^n$ stays bounded away from zero would directly confirm or contradict the prediction.

Watch

Extended reading notes

Core claim

The central claim, Theorem 1, is that the sign of a single expression constructed from the initial generating function decides the regime of the model. Define $D_0(s,m)=(m-1)sF_0'(s)-aF_0(s)$, where $F_0(s)=\mathbb{E}(s^{X_0})$. If $D_0((\mathbb{E}N)^{1/a},\mathbb{E}N)>0$, the model is supercritical, $Q>0$, and this holds without any boundedness assumption on $N$. If $N$ is almost surely at most $M$ and $D_0(1+(M-1)/a,M)<0$, the model is subcritical, $Q=0$. The theorem is stated as sufficient conditions: it does not claim to cover all cases where $Q$ has a sign. When the number of terms is fixed, the two thresholds coincide for $a=1$ and the statement reproduces the previously known criterion, while for $a>1$ it gives new sufficient conditions.

Load-bearing premise

The theorem requires the moment generating function $F_0$ and its derivative to be finite at the specific test points, but the model assumptions only guarantee that $\mathbb{E}X_0$ is finite; for heavy-tailed $X_0$ the test expression can be undefined, so the conditions do not apply.

Editorial extensions

If this is right

If the first condition holds, the free energy $Q$ is positive and the normalized expectations $\mathbb{E}(X_n)/(\mathbb{E}N)^n$ decrease to a positive limit.
If $N$ is bounded and the second condition holds, $Q$ is zero and $\mathbb{E}(X_n)$ stays bounded by $a/(\mathbb{E}N-1)$.
For a fixed number of terms, Theorem 2 gives a concrete generating-function test, reducing to the known $a=1$ criterion and producing new conditions for $a>1$.
The first condition applies even when $N$ is unbounded; only the second requires the boundedness $N\le M$.
Because the conditions are sufficient, there remains a range of distributions for which the theorem is silent; the critical case $D_0=0$ is not classified.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The authors' conditions are one-sided tests; a natural next step is to check whether the two thresholds bound the true critical surface, with $D_0=0$ marking a critical window in which finer asymptotics are needed.
When $F_0(s)$ is infinite at the test point, the theorem is not applicable even though $\mathbb{E}X_0$ is finite; in that regime one would need a different normalization or a truncated-moment analogue of $D_0$.
For unbounded $N$, replacing the deterministic bound $M$ by a tail bound might yield an analogous subcritical test, since the proof of condition 2 uses only the inequality $\mathbb{E}(N v^N) \le M\mathbb{E}(v^N)$.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Chen, quick take: this paper extends Derrida–Retaux criticality criteria to a random number of terms N and to a>1 for fixed n. The theorems are new relative to what I know, and the generating-function machinery is the right tool. The proofs are largely sound: Lemma 4 is a neat Chebyshev-type rearrangement, and the monotonicity arguments in Sections 3.2 and 3.3 are convincing once the stated inequalities hold. The mislabeled concavity in Lemma 1 (f1 is convex, not concave) is a genuine typo, but the inequality f2(y) ≥ a(EN)^{y/a} is true for 0≤y≤a, so the proof survives. Credit where due: the paper does not rest on fitted parameters or circular normalization, and it openly cites the a=1 fixed-N antecedents.

The soft spot is real and slightly deeper than the reader flagged. The theorem’s supercritical criterion evaluates D0 at s=(EN)^{1/a}, which requires G(z)=E[z^N] finite at z>1. The model only assumes E N < ∞, and with a heavy-tailed N (e.g., P(N=k)~c/k^3), G(z)=∞ for every z>1. Then F1(s)=∞ for all s>1 and the entire iteration in inequality (5) cannot start. The same issue affects the subcritical side when N is bounded (then G is finite everywhere, so only the initial F0 matters there). The finite-mgf assumption on X0 is also unstated, but the N-mgf issue is the load-bearing one. Both are repairable: add a hypothesis on exponential moments of N, or prove a truncation argument using monotonicity of Q in N. As written, the theorem oversells its scope.

Would I engage? Yes. The core idea is right, the paper is short and readable, and the gap is fixable without changing the architecture. For a referee: send it out, but insist the authors state the finiteness conditions precisely and either prove the heavy-tailed case or restrict to N with finite mgf near 1. If the fix is clean, this is a solid contribution to a small active literature; if not, it is still a useful lemma collection. My verdict mirrors the conditional: accept pending correction of the stated moment hypotheses.

Referee Report

2 major / 5 minor

Summary. The paper studies the Derrida–Retaux recursion X_{n+1} = (X_n^{(1)} + ... + X_n^{(N_{n+1})} - a)_+ with i.i.d. random family sizes N and a nonnegative integer initial value X_0. It defines the free energy Q as the limit of normalized expectations and proves Theorem 1, giving sufficient conditions for supercriticality (Q > 0) when D_0((EN)^{1/a}, EN) > 0 and, for bounded N with N <= M, for subcriticality (Q = 0) when D_0(1 + (M-1)/a, M) < 0. The proof uses moment generating functions, recursive inequalities for D_n(s) = (m-1)sF_n'(s) - aF_n(s), and elementary association inequalities. Theorem 2 specializes the result to a fixed number of terms n, recovering the known a = 1 case and giving a new result for a > 1.

Significance. The proposed conditions are explicit, involve no fitted parameters, and, if valid, provide a practical moment-generating-function test for the phase of the Derrida-Retaux model with random family sizes. The paper is self-contained and the main recursive inequalities in Lemmas 1 and 3 are essentially correct once the finiteness issues discussed below are resolved. The fixed-n specialization in Theorem 2 is a genuine extension of the a = 1 result in [2] to a > 1. The chief weakness is that Theorem 1(1) is not proved in the stated generality because the argument requires G(z) = E[z^N] to be finite at values z > 1, which is not implied by the stated assumptions on N.

major comments (2)

[Theorem 1(1), Section 3.2, Eq. (5)] The proof of the supercriticality criterion is not valid for the stated generality because it silently requires the generating function G(z) = E[z^N] to be finite at z = F_n(s) > 1. For an unbounded N with finite mean and tail P(N = k) ~ c/k^3, G(z) is infinite for every z > 1; then (3) gives F_1(s) = infinity and the quantities in (5) are undefined. The assumptions E X_0 < infinity and P(N > 1) > 0 do not exclude this case. In addition, the condition D_0((EN)^{1/a}, EN) requires F_0((EN)^{1/a}) to be finite, which also is not implied by E X_0 < infinity. The theorem should either add explicit hypotheses, for example F_0((EN)^{1/a}) < infinity and G finite on [1, F_0((EN)^{1/a})], or the proof should establish part (1) by a truncation and monotonicity argument that avoids evaluating G at values where it is infinite.
[Lemma 3, Section 3.3] In the proof of Lemma 3, the chain vG'(v) <= E[N v^N] <= EN * E[v^N] <= M G(v) contains the false inequality E[N v^N] <= EN * E[v^N] for v > 1, because N and v^N are positively correlated. The desired bound vG'(v) <= M G(v) follows directly from the pointwise inequality N <= M, so the lemma is correct but the given justification must be replaced.

minor comments (5)

[Lemma 1, Eq. (6)] The function f_1(y) = a(EN)^{y/a} is convex, not concave. The conclusion f_2(y) >= f_1(y) is nonetheless true because a convex function lies below its chord between the endpoints y = 0 and y = a; the proof should state this correctly.
[Inequality (7), proof of Lemma 3] The left-hand side of (7) is written with s(EN - 1)F'_{n+1}(s) - aF_{n+1}(s), but for D_{n+1} it should be s(M - 1)F'_{n+1}(s) - aF_{n+1}(s).
[Proof of Lemma 3] The sentence 'By assumption y >= 1, s >= 1 + 1/a' contains a typo; the hypothesis of Lemma 3 is s >= 1 + (M-1)/a, and the displayed implication should use this bound.
[Theorem 1 statement] The phrase 'Let it be D_0(s,m) = ...' is awkward; 'Let D_0(s,m) = ...' is the standard wording.
[References] References [5] and [7] appear in the bibliography but are not cited in the body of the paper; the authors should add citations or remove the entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the sufficient conditions are derived from the recurrence via moment-generating-function inequalities, with no fitted parameter renamed as a prediction.

full rationale

No circular steps were identified. Theorem 1 states sufficient conditions in terms of the initial moment generating function F0 and the generating function G of N; the proof defines D_n(s) = (EN-1)sF'_n(s)-aF_n(s) (or with M in the bounded case) and derives the iterative inequalities (5) and (7) directly from the generating-function recurrence (3)-(4). Lemmas 1 and 2, and Lemmas 3 and 4, are proved from distributional inequalities and elementary moment bounds, then combined by contradiction; neither lemma assumes the theorem's conclusion as a premise. The condition D0((EN)^{1/a}, EN)>0 is an explicit test on the initial data, not a quantity fitted to the predicted energy Q. The paper honestly notes that in the deterministic case P(N=n)=1 with a=1, the result reduces to the known result from reference [2]; this is a consistency check, not a renamed input. The only concern raised by a close reading is a regularity gap: for unbounded N and s>1, G(F_n(s)) may be infinite for heavy-tailed N, so the iterative inequality (5) may not be applicable. That is a potential missing hypothesis or proof gap, not circularity, because it does not make the conclusion equivalent to the assumptions by definition. There are no load-bearing self-citations, no fitted inputs called predictions, and no ansatz smuggled in via citation.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No fitted constants and no invented entities. The proof relies on standard probabilistic inequalities, a branching-process fact, and an implicit finiteness condition on the moment generating function of X0.

assumptions (5)

standard math For v >= 1 and i.i.d. copies N1,N2 of N, E[(N1-N2)(v^{N1}-v^{N2})] = 2(E(N v^N) - EN E v^N) >= 0.
Used in Lemma 1 to prove vG'(v) >= EN G(v), the key inequality driving exponential growth of D_n.
domain assumption The depth-k descendant count T_k in the Galton-Watson tree with offspring distribution N satisfies E T_k = (EN)^k.
Used in Lemma 2 to convert a tail bound on X_n into a bound on E(X_n s^{X_n}). Standard branching process fact.
standard math Bernoulli's inequality (1+x)^y >= 1+xy for integer y>=1 and x>=0.
Used in Lemma 3 to show (a-p)(M-1)+a <= a s^{a-p} for s >= 1+(M-1)/a.
standard math For independent copies Y1,Y2 of X_n and s>=1, E[(Y1-Y2)(s^{Y1}-s^{Y2})] >= 0.
Used in Lemma 4 to show E(X_n s^{X_n}) >= E X_n E s^{X_n}.
domain assumption F0(s)=E(s^{X0}) and F0'(s) are finite at s=(EN)^{1/a} and at s=1+(M-1)/a.
The model only assumes E X0 < infinity. The theorem's conditions D0>0 or D0<0 require F0 and F0' to exist at the evaluation points; this is implicit but not stated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Criticality conditions in the Derrida-Retaux model with a random number of terms." pith.science (2026). https://pith.science/paper/OBCVUQCN

@misc{pith2026250202535,
  author       = {Pith},
  title        = {Pith review of: Criticality conditions in the Derrida-Retaux model with a random number of terms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OBCVUQCN}},
  note         = {Machine review of arXiv:2502.02535}
}

abstract

The article considers the Derrida-Retaux model with a random number of terms, i.e. a sequence of integer random variables defined by the relations $ X_{n + 1} = (X_n^{(1)} + X_n^{(2)} + ... + X_n^{(N_n)} - a)^{+}$, $n\ge 0$, where $X_n^{j}$ are independent copies of $X_n$, the values of $N_j$ are independent and identically distributed, $a$ is a positive integer. The energy in the model is defined as $Q:=\lim\limits_{n\to\infty} \frac{\mathbb{E}(X_{n})}{(\mathbb{E}N_1)^{n}}$. We present sufficient conditions (in terms of distributions of $X_0$ and $N_1$) for subcritical ($Q=0$) and supercritical ($Q>0$) regimes of model behavior.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages

[2]

and Shi, Z

Chen, X., Derrida, B., Hu, Y., Lifshits, M. and Shi, Z. A hierarchical renormalization model: some properties and open questions. In: Sojourns in Probability Theory and Statistical Physics - III. Inter- acting Particle Systems and Random Walks, A Festschrift for Charles M. Newman. Ser.: Springer Proc. Math. & Statist. V. Sidoravicius (Ed.), Springer, 2019...

work page 2019
[1]

and Martin, A

Collet, P., Eckmann, J.P., Glaser, V. and Martin, A. (1984). Study of the iterations of a mapping associated to a spin-glass model.Commun. Math. Phys.94, 353–370

work page 1984
[3]

and Retaux, M

Derrida, B. and Retaux, M. (2014). The depinning transition in presence of disorder: a toy model. J. Statist. Phys. 156, 268–290

work page 2014
[4]

and Vannimenus, J

Derrida, B., Hakim, V. and Vannimenus, J. (1992). Effect of disorder on two-dimensional wetting. J. Statist. Phys. 66, 1189–1213

work page 1992
[5]

and Toninelli, F.L

Giacomin, G., Lacoin, H. and Toninelli, F.L. (2010). Hierarchical pinning models, quadratic maps and quenched disorder. Probab. Theory Related Fields 147, 185–216

work page 2010
[6]

and Shi, Z

Hu, Y. and Shi, Z. (2018). The free energy in the Derrida–Retaux recursive model. J. Statist. Phys. 172, 718–741

work page 2018
[7]

Theory Related Fields 148, 159–175

Lacoin, H.(2010).Hierarchicalpinningmodelwithsitedisorder: disorderismarginallyrelevant.Probab. Theory Related Fields 148, 159–175. St. Petersburg State University, Universitetskaya emb. 7/9, 191023, St. Petersburg, Russia E-mail : Kotann2710@mail.ru E-mail : alex.lotnikov@gmail.com 6

work page 2010

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [2]

and Shi, Z

Chen, X., Derrida, B., Hu, Y., Lifshits, M. and Shi, Z. A hierarchical renormalization model: some properties and open questions. In: Sojourns in Probability Theory and Statistical Physics - III. Inter- acting Particle Systems and Random Walks, A Festschrift for Charles M. Newman. Ser.: Springer Proc. Math. & Statist. V. Sidoravicius (Ed.), Springer, 2019...

work page 2019

[2] [1]

and Martin, A

Collet, P., Eckmann, J.P., Glaser, V. and Martin, A. (1984). Study of the iterations of a mapping associated to a spin-glass model.Commun. Math. Phys.94, 353–370

work page 1984

[3] [3]

and Retaux, M

Derrida, B. and Retaux, M. (2014). The depinning transition in presence of disorder: a toy model. J. Statist. Phys. 156, 268–290

work page 2014

[4] [4]

and Vannimenus, J

Derrida, B., Hakim, V. and Vannimenus, J. (1992). Effect of disorder on two-dimensional wetting. J. Statist. Phys. 66, 1189–1213

work page 1992

[5] [5]

and Toninelli, F.L

Giacomin, G., Lacoin, H. and Toninelli, F.L. (2010). Hierarchical pinning models, quadratic maps and quenched disorder. Probab. Theory Related Fields 147, 185–216

work page 2010

[6] [6]

and Shi, Z

Hu, Y. and Shi, Z. (2018). The free energy in the Derrida–Retaux recursive model. J. Statist. Phys. 172, 718–741

work page 2018

[7] [7]

Theory Related Fields 148, 159–175

Lacoin, H.(2010).Hierarchicalpinningmodelwithsitedisorder: disorderismarginallyrelevant.Probab. Theory Related Fields 148, 159–175. St. Petersburg State University, Universitetskaya emb. 7/9, 191023, St. Petersburg, Russia E-mail : Kotann2710@mail.ru E-mail : alex.lotnikov@gmail.com 6

work page 2010