The Critical Beta-splitting Random Tree IV: Mellin analysis of Leaf Height

David Aldous; Svante Janson

REVIEW 2 major objections 5 minor 22 references

The Critical Beta-splitting Random Tree IV: Mellin analysis of Leaf Height

T0 review · 2 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The expected leaf height in the critical beta-splitting random tree has a complete asymptotic expansion whose exponents are set by the negative roots of the digamma equation, and the same Mellin machinery yields the variance, CLT, and…

desk verdict Mellin expansion of E[D_n] is real and well-proved conditional on the moment identity imported from the companion paper; send it to a referee, with requests to fix the notation and close the self-containedness gap. read the letter →

arxiv 2412.12319 v1 pith:JF6SKZA6 submitted 2024-12-16 math.PR math.CV

David Aldous , Svante Janson This is my paper · ORCID

classification math.PRmath.CV MSC 60C0505C0544A1560G09

keywords criticalbeta-splittingtreeleafheightMellintransformexchangeablerandompartitiondigammafunctionasymptoticexpansioncentrallimittheoremlargedeviations

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

This paper studies the critical $\beta$-splitting model, the random binary tree on $n$ leaves in which a clade of $m$ leaves is split with probability proportional to $1/(i(m-i))$. The authors establish a complete asymptotic expansion for $D_n$, the hop-height of a uniform random leaf: $E[D_n] \sim (6/\pi^2)\log n + c_0 + c_1 n^{-1} + \sum_{j,k} c_{j,k} n^{-|s_j|-k}$, where $s_j$ are the negative roots of $\psi(s)=\psi(1)$ and $c_0=\zeta(3)/\zeta(2)^2+\gamma/\zeta(2)$. This goes beyond earlier results that supplied only the first few terms and had conjectured that only integer powers of $1/n$ would appear. The same Mellin machinery yields the variance, a central limit theorem, and large-deviation exponents for $D_n$, plus asymptotic expansions for the continuous-time height $L_n$ and the total tree length $\Lambda_n$.

What carries the argument

The central object is the infinite measure $\Upsilon$ on $(0,1)$ defined by $\Upsilon=\int_0^\infty \mathcal{L}(P_{t,1})\,dt$, where $P_{t,1}$ is the asymptotic size-biased proportion of the clade of leaf 1 at time $t$ in the infinite limit tree; in practical terms this means $\int_0^1 f\,d\Upsilon = \int_0^\infty E[f(P_{t,1})]\,dt$ for nonnegative $f$. Its Mellin transform is the reciprocal of $\psi(s)-\psi(1)$, and the paper's inversion estimate shows that the density of $\Upsilon$ is $\upsilon(x)=\frac{6}{\pi^2 x}+\sum_i \frac{1}{\psi'(s_i)}x^{|s_i|}+r_N(x)$, with the negative roots $s_i$ of $\psi(s)=\psi(1)$ controlling all non-integer powers. This density estimate, together with the integral identity for $E[D_n]$, is what converts a limit-tree representation into sharp finite-$n$ asymptotics; a parallel Parseval-formula argument reaches the same expansions by shifting complex contours and collecting residues at the corresponding poles.

What would settle it

Run a high-precision simulation of the discrete model for $n$ near $10^6$, estimate $E[D_n]$ to about $10^{-6}$, and compare $E[D_n]-(6/\pi^2)\log n - c_0$ with the predicted first corrections $-3/(\pi^2 n)-0.0943\,n^{-1.567}-1/(2\pi^2 n^2)$; if the remainder instead matches a purely integer-power series, the negative-root spectrum in Theorem 1.1 is wrong.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the expected leaf height is encoded in one infinite measure $\Upsilon$ on $(0,1)$, built from the occupation time of the size-biased clade proportion in the infinite limit tree. The exact identity $E[D_n]=\int_0^1 (1-(1-x)^{n-1})\,d\Upsilon(x)$ holds for every $n$, and $\Upsilon$ has Mellin transform $\int_0^1 x^{s-1}\,d\Upsilon(x)=1/(\psi(s)-\psi(1))$ for $\Re s>1$. Because the denominator vanishes at $s=1$ and at the negative roots $s_j$ of $\psi(s)=\psi(1)$, shifting the Mellin inversion contour turns this representation into an asymptotic expansion labeled by those roots: $E[D_n] \sim (6/\pi^2)\log n + c_0 + c_1 n^{-1} + \sum_{j\ge1}\sum_{k\ge1} c_{j,k}n^{-|s_j|-k}$, with $c_0=\zeta(3)/\zeta(2)^2+\gamma/\zeta(2)$ and $c_1=-3/\pi^2$. The same mechanism extends to higher moments and to the moment generating function, giving the variance, the CLT with mean $6/\pi^2$ and variance $2\zeta(3)/\zeta(2)^3$, and the large-deviation rate function.

Load-bearing premise

The load-bearing premise is the moment identity $E[P_{t,1}^s]=e^{-t(\psi(s+1)-\psi(1))}$ for the infinite limit tree, imported from the companion preprint [3] rather than proved here; if that identity fails or holds only approximately, the Mellin transform of $\Upsilon$, the inversion estimate, and every asymptotic expansion built on them would not follow.

Editorial extensions

If this is right

The expected leaf height has a computable expansion to any order: after the leading $(6/\pi^2)\log n$, the constant is $0.795155660439$ and the $n^{-1}$ coefficient is $-3/\pi^2$, with later coefficients determined by the roots $s_j$.
The variance of $D_n$ is $(2\zeta(3)/\zeta(2)^3)\log n$ plus a known constant, with error $O((\log n)/n)$, so the fluctuations grow only logarithmically.
The centered leaf height converges to a normal distribution with mean $6/\pi^2$ and variance $2\zeta(3)/\zeta(2)^3$ after normalization by $\sqrt{\log n}$.
Large-deviation probabilities for $D_n/(\log n)$ decay as $n^{-\Lambda^*(x)+o(1)}$ with an explicit rate function, with threshold $x_0=6/\pi^2$ at the center.
The same Mellin method gives $E[\Lambda_n]=(6/\pi^2)n+O(n^{-|s_1|})$ for the continuous tree's total length and the earlier occupation-probability asymptotics for the harmonic descent chain.

Reading between the lines

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

This suggests that any fragmentation model whose infinite limit has an explicit size-biased moment function should admit the same Mellin treatment, with the analogue of $1/(\psi(s)-\psi(1))$ determining the spectrum of powers in its height expansion.
The paper leaves the variance of the continuous-time hop-height $L_n$ open, explicitly because no representation of its higher moments analogous to Proposition 4.1 is available; finding such a representation is the natural next step, after which the present machinery would apply directly.
A concrete, untested prediction of the expansion is that the non-integer correction $-0.0943\,n^{-1.567}$ will be visible in high-precision simulations of $E[D_n]$ for $n$ around $10^5$ to $10^6$; observing only integer-power corrections would indicate that the inversion estimate misses essential structure.

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

Here's my read. The main theorems are solid, conditional on one imported identity, and the paper is worth serious engagement. Aldous and Janson push the Mellin program to a sharp level: explicit constant c0, the full expansion with the mixed integer/non-integer power spectrum, the variance constant, MGF asymptotics, and large deviations. The technical core — Lemma 6.1's inversion estimate, the residue expansions in Section 10.1, and the careful handling of conditionally convergent line integrals in Lemma 10.1 — is genuinely good. The trick of differentiating the Mellin transform to gain integrability is useful beyond this paper.

The soft spots are real but manageable. First, everything rests on (2.4), the moment identity imported from companion preprint [3]. The paper does not reprove it, and if it failed, the Mellin transform of Upsilon and all pole expansions collapse. That is a self-containedness gap, not an identified error — the identity is plausible, and the paper's own calculations are internally consistent once you grant it. A referee should ask the authors to either prove (2.4) in an appendix or make the dependency explicit and wait for [3] to clear review. Second, the D_n/L_n notation is swapped in the displayed definitions (1.4)-(1.5): the text defines D_n as hop-height, but (1.4) gives the continuous-time absorption time, and (1.5) gives the discrete-time one. Easy fix, but currently confusing. Third, the claim that this is the first application of Parseval's formula for Mellin transforms to combinatorial probability is overstated — the Mellin toolbox has been used in combinatorial settings for decades. That is a literature comment, not a mathematical flaw.

My overall assessment matches the reader's: the central argument holds up conditionally, the expansions are new relative to prior work, and the citation pattern is appropriate. Self-citation to [3] is legitimate because the companion theorem is the natural source, not a reflexive citation. For a random tree theorist, this is a strong tool; for someone wanting a self-contained proof, you'll want to wait for [3] or ask the authors to close the gap. I'd send it to a serious referee, with the reviewer focusing on the dependency on (2.4) and the notation, not on hunting for a hidden error — the computations are careful and check out.

Referee Report

2 major / 5 minor

Summary. The paper studies the critical beta-splitting random tree and derives sharp asymptotic results for the height of a uniform random leaf. Using the exchangeable partition representation of the infinite limit tree CTCS(∞) from the companion paper [3], the authors define a measure Υ by (4.1) and identify its Mellin transform as 1/(ψ(s)−ψ(1)) via the moment identity (2.4). From this, they obtain exact representations of E[D_n], E[L_n], and E[Λ_n], and then use Mellin inversion and residue calculus to prove a full asymptotic expansion of E[D_n] (Theorem 1.1), the analogous expansion for E[L_n] (Theorem 1.2), the variance asymptotics for D_n (Theorem 1.3), the moment generating function asymptotics (Theorem 1.4), and the resulting CLT and large deviation results (Theorems 1.5 and 1.6). The paper also develops an alternative Parseval-formula method in Section 10 and applies it to higher moments in Section 11.

Significance. If the results hold, they are a substantial advance over the earlier recurrence-method results in [5] and [2]: they give full asymptotic expansions with explicitly computable coefficients, a surprising spectrum of powers n^{-|s_j|-k} determined by the nontrivial roots of ψ(s)=ψ(1), and a new Mellin-transform route through the limit tree. The analytic work is detailed and careful: Lemma 6.1 is proved in full, the Parseval formula of Lemma 10.1 is proved in Appendix A with an explicit Fejér-kernel argument, and exact finite-n formulas such as (4.13) provide partial checks. The main risk to correctness is the dependence on the unproven identity (2.4) imported from the companion preprint [3]; modulo that input, the internal derivations are consistent and the constants are cross-checked numerically.

major comments (2)

[§1.2 and §5.2] The definitions of D_n and L_n are inconsistent with the rest of the paper. Section 1.2 states that D_n is the hop-height of DTCS(n) and L_n is the height of CTCS(n), but equations (1.4)–(1.5) assign D_n to the absorption time of the continuous-time chain and L_n to the absorption time of the discrete-time chain. Moreover, the proof of (4.4) treats D_n as a continuous-time height, Theorem 1.2 gives E[L_n] ~ (3/π^2) log^2 n, which is the discrete hop-height, and equations (5.5)–(5.6) in §5.2 only make sense if D_n is the continuous height and L_n is the discrete hop-height. As written, Theorem 1.1 appears to contradict the stated definition of D_n. This is more than a typo: the authors must relabel the definitions or the equations so that the discrete and continuous quantities are consistently named throughout the paper.
[§2, Eq. (2.4); §4, Eq. (4.8)] The entire Mellin analysis rests on the identity E[P_{t,1}^s] = exp(-t(ψ(s+1)−ψ(1))) imported without proof from [3, Theorem 4.5]. Equation (4.8) is exactly this identity rewritten for the Mellin transform of Υ, and every subsequent residue expansion—in Theorems 7.3, 8.1, 9.1, 11.1, 12.4, and all of their corollaries—uses the poles of 1/(ψ(s)−ψ(1)). The exact formula (4.13) is derived from (4.8), so it is not an independent check. Since [3] is a not-yet-peer-reviewed companion preprint and no proof of (2.4) is included here, the paper should either provide a proof of (2.4) in an appendix or explicitly state that the main theorems are conditional on [3, Theorem 4.5].

minor comments (5)

[§12.1, Eq. (12.10)] The displayed line 'We thus interpret -zΓ(-ρ(-z)) = ψ′(1) for z = 0' contains a sign error in the argument of ρ; it should be -zΓ(-ρ(z)) = ψ′(1) at z = 0.
[§13] In the acknowledgments, 'Bénédicte Haas for his careful explanation' should use 'her' instead of 'his'.
[§11.2] The title 'Asymptotics of Υ k' contains the typo 'analoguous'; it should be 'analogous'.
[§1.3, Theorem 1.1] The statement '(1.6) for some coefficients c_i and c_{j,k}' would benefit from an explicit note that the double sum is interpreted by grouping terms in decreasing order of n-exponent, as already explained in §3.1; adding a cross-reference would prevent confusion.
[§4, Remark 4.3] The exact formula (4.13) is interesting, but the sentence 'we will not use it' could also mention that it depends on (2.4) in the same way as the main analysis, so it does not offer independent verification; this would address a possible reader misconception.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; the paper's asymptotic results reduce to the parameter-free moment identity (2.4) imported from the authors' companion paper, which is a self-citation risk but not circularity.

full rationale

The derivation chain is internally consistent and contains no fitted constants. The central step is (2.4), E[P_{t,1}^s] = exp(-t(psi(s+1)-psi(1))), quoted from [3, Theorem 4.5] and used to compute the Mellin transform (4.8) of the measure Upsilon. Although [3] is a companion preprint by the same authors and (2.4) is load-bearing, it is a parameter-free statement about the infinite limit tree, with assumptions that do not include the target asymptotic expansions for E[D_n], E[L_n], or the variance. The subsequent Mellin inversion (Lemma 6.1), the residue calculus in Sections 7, 10, 11, and the large-deviation analysis via the Gartner-Ellis theorem are all carried out in this paper from (2.4) and standard digamma facts. Constants such as c0 and c1 are computed from psi'(1) and psi''(1), not fitted. Therefore the results do not reduce to their inputs by construction. The only caveat is that a non-peer-reviewed companion result is imported without re-proof; that is a self-containedness and verification risk, not circularity, so the score is low.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

No free parameters are fitted; constants such as c0 are derived from digamma evaluations. The central claim rests on standard complex analysis and on a key theorem imported from the authors' companion paper III. The measure Upsilon is a mathematical construction, not a new physical entity, so no invented entities are listed.

assumptions (6)

domain assumption Exchangeable partition representation of the infinite tree CTCS(infinity) and the moment identity E[P_{t,1}^s] = exp(-t(psi(s+1)-psi(1))).
Invoked at equation (2.4) and used to compute the Mellin transform of Upsilon at (4.8). All subsequent asymptotics depend on this limit structure imported from the companion paper [3].
domain assumption The critical beta-splitting split probabilities q(m,i) proportional to 1/(i(m-i)).
This is the model definition itself, stated at (1.1), and the source of the harmonic descent Markov chain transitions (1.2) and (1.3).
domain assumption The path of a uniform random leaf corresponds to the harmonic descent Markov chain with transitions (1.2) and (1.3).
Used to connect tree heights to absorption times and to derive the occupation probability identities in Proposition 4.1.
standard math Standard digamma function facts, including the root structure of psi(s)=psi(1) in Lemma 3.1 and asymptotic expansions of psi.
These facts provide the spectrum of powers s_i that appear in the asymptotic expansions and are used throughout the residue calculations.
standard math Mellin inversion, Parseval's formula for Mellin transforms, and residue calculus, including Lemmas 6.1 and 10.1.
These tools convert the Mellin transform of Upsilon into the density expansion and then into the finite-n asymptotic expansions.
standard math Gartner-Ellis theorem for large deviations.
Used in Section 12.3 to convert the pointwise MGF limit (12.22) into the large deviation statements of Theorem 1.6.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Critical Beta-splitting Random Tree IV: Mellin analysis of Leaf Height." pith.science (2026). https://pith.science/paper/JF6SKZA6

@misc{pith2026241212319,
  author       = {Pith},
  title        = {Pith review of: The Critical Beta-splitting Random Tree IV: Mellin analysis of Leaf Height},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JF6SKZA6}},
  note         = {Machine review of arXiv:2412.12319}
}

abstract

In the critical beta-splitting model of a random $n$-leaf rooted tree, clades are recursively split into sub-clades, and a clade of $m$ leaves is split into sub-clades containing $i$ and $m-i$ leaves with probabilities $\propto 1/(i(m-i))$. The height of a uniform random leaf can be represented as the absorption time of a certain {\em harmonic descent} Markov chain. Recent work on these heights $D_n$ and $L_n$ (corresponding to discrete or continuous versions of the tree) has led to quite sharp expressions for their asymptotic distributions, based on their Markov chain description. This article gives even sharper expressions, based on an $n \to \infty$ limit tree structure described via exchangeable random partitions in the style of Haas et al (2008). Within this structure, calculations of moments lead to expressions for Mellin transforms, and then via Mellin inversion we obtain sharp estimates for the expectation, variance, Normal approximation and large deviation behavior of $D_n$.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages

[3]

The critical beta-split ting random tree III: The exchangeable partition representation and the fringe tree

David Aldous and Svante Janson. The critical beta-split ting random tree III: The exchangeable partition representation and the fringe tree. arXiv::2412 .09655, 2024

work page 2024
[5]

The Critical Beta-splitting Random Tree: Heights and Related Results

David Aldous and Boris Pittel. The critical beta-splitt ing random tree I: Heights and related results. arXiv:2302.05066, 2023. To appear in Annals of Applied Probability

work page Pith review arXiv 2023
[2]

The critical beta-split ting random tree II: Overview and open problems

David Aldous and Svante Janson. The critical beta-split ting random tree II: Overview and open problems. arXiv 2303.02529v2, 2024

arXiv 2024
[1]

Probability distributions on cladograms

David Aldous. Probability distributions on cladograms . In Random Discrete Structures (Min- neapolis, MN, 1993) , volume 76 of IMA Vol. Math. Appl. , pages 1–18. Springer, New York, 1996

work page 1993
[4]

The harmoni c descent chain

David Aldous, Svante Janson, and Xiaodan Li. The harmoni c descent chain. Electron. Commun. Probab. 29 (2024), paper no. 77, 1–10

work page 2024
[6]

Random Fragmentation and Coagulation Processes , volume 102 of Cambridge Studies in Advanced Mathematics

Jean Bertoin. Random Fragmentation and Coagulation Processes , volume 102 of Cambridge Studies in Advanced Mathematics . Cambridge Univ. Press, Cambridge, 2006

work page 2006
[7]

Large Deviation Techniques and Applications, 2nd ed

Amir Dembo and Ofer Zeitouni. Large Deviation Techniques and Applications, 2nd ed. Springer, New York, 1998

work page 1998
[8]

Analytic Information Theory

Michael Drmota and Wojciech Szpankowski. Analytic Information Theory. From Compression to Learning. Cambridge Univ. Press, Cambridge, UK, 2023

work page 2023

Show all 22 references

[9]

Mellin transforms and asymptotics: harmonic sums

Philippe Flajolet, Xavier Gourdon, and Philippe Dumas. Mellin transforms and asymptotics: harmonic sums. Theoret. Comput. Sci. , 144(1-2):3–58, 1995

1995
[10]

Analytic Combinatorics

Philippe Flajolet and Robert Sedgewick. Analytic Combinatorics. Cambridge Univ. Press, Cam- bridge, UK, 2009

2009
[11]

Regenerative compositions in the case of slow vari- ation: a renewal theory approach

Alexander Gnedin and Alexander Iksanov. Regenerative compositions in the case of slow vari- ation: a renewal theory approach. Electron. J. Probab. , 17:no. 77, 19 pp., 2012

2012
[12]

Continuum tree asymptotics of discrete fragmentations and applications t o phylogenetic models

B´ en´ edicte Haas, Gr´ egory Miermont, Jim Pitman, and M atthias Winkel. Continuum tree asymptotics of discrete fragmentations and applications t o phylogenetic models. Ann. Probab., 36(5):1790–1837, 2008

2008
[13]

The genealogyof self-similar fragmentations with neg- ative index as a continuum random tree

B´ en´ edicte Haas and Gr´ egory Miermont. The genealogyof self-similar fragmentations with neg- ative index as a continuum random tree. Electron. J. Probab. 9(4):57–97, 2004

2004
[14]

Handelsman and John S

Richard A. Handelsman and John S. Lew. Asymptotic expan sion of a class of integral transforms via Mellin transforms. Arch. Rational Mech. Anal. 35 (1969), 382–396. 44 DA VID J. ALDOUS AND SV ANTE JANSON

1969
[15]

A comment on the article ‘The harmon ic descent chain’ by D

Alexander Iksanov. A comment on the article ‘The harmon ic descent chain’ by D. J. Aldous, S. Janson and X. Li. arxiv 2412.06826, 2024

2024 arXiv
[16]

Another proof of CLT for critical be ta-splitting tree

Alexander Iksanov. Another proof of CLT for critical be ta-splitting tree. Unpublished, 2024

2024
[17]

An Introduction to Harmonic Analysis

Yitzhak Katznelson. An Introduction to Harmonic Analysis . John Wiley & Sons, New York, 1968

1968
[18]

Critical beta-splitting, via contrac tion

Brett Kolesnik. Critical beta-splitting, via contrac tion. arXiv 2404.16021, 2024

2024
[19]

Frank W. J. Olver, Daniel W. Lozier, Ronald F. Boisvert, and Charles W. Clark, editors. NIST Handbook of Mathematical Functions . U.S. Department of Commerce, National Institute of Standards and Technology, Washington, DC; Cambridge Univ. Press, Cambridge, 2010. Also available...

2010
[20]

On the contra ction method with degenerate limit equation

Ralph Neininger and Ludger R¨ uschendorf. On the contra ction method with degenerate limit equation. Ann. Probab., 32(3B):2838–2856, 2004

2004
[21]

Asymptotics an d Mellin–Barnes integrals

Richard Bruce Paris and David Kaminski. Asymptotics an d Mellin–Barnes integrals. Cambridge Univ. Press, Cambridge, 2001

2001
[22]

E. C. Titchmarsh, Introduction to the Theory of Fourier integrals . Oxford Univ. Press, Oxford, UK, 1937. Department of Statistics, 367 Evans Hall # 3860, U.C. Berkel ey CA 94720 Email address : aldousdj@berkeley.edu URL: www.stat.berkeley.edu/users/aldous. Department of Mathe...

1937

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [3]

The critical beta-split ting random tree III: The exchangeable partition representation and the fringe tree

David Aldous and Svante Janson. The critical beta-split ting random tree III: The exchangeable partition representation and the fringe tree. arXiv::2412 .09655, 2024

work page 2024

[2] [5]

The Critical Beta-splitting Random Tree: Heights and Related Results

David Aldous and Boris Pittel. The critical beta-splitt ing random tree I: Heights and related results. arXiv:2302.05066, 2023. To appear in Annals of Applied Probability

work page Pith review arXiv 2023

[3] [2]

The critical beta-split ting random tree II: Overview and open problems

David Aldous and Svante Janson. The critical beta-split ting random tree II: Overview and open problems. arXiv 2303.02529v2, 2024

arXiv 2024

[4] [1]

Probability distributions on cladograms

David Aldous. Probability distributions on cladograms . In Random Discrete Structures (Min- neapolis, MN, 1993) , volume 76 of IMA Vol. Math. Appl. , pages 1–18. Springer, New York, 1996

work page 1993

[5] [4]

The harmoni c descent chain

David Aldous, Svante Janson, and Xiaodan Li. The harmoni c descent chain. Electron. Commun. Probab. 29 (2024), paper no. 77, 1–10

work page 2024

[6] [6]

Random Fragmentation and Coagulation Processes , volume 102 of Cambridge Studies in Advanced Mathematics

Jean Bertoin. Random Fragmentation and Coagulation Processes , volume 102 of Cambridge Studies in Advanced Mathematics . Cambridge Univ. Press, Cambridge, 2006

work page 2006

[7] [7]

Large Deviation Techniques and Applications, 2nd ed

Amir Dembo and Ofer Zeitouni. Large Deviation Techniques and Applications, 2nd ed. Springer, New York, 1998

work page 1998

[8] [8]

Analytic Information Theory

Michael Drmota and Wojciech Szpankowski. Analytic Information Theory. From Compression to Learning. Cambridge Univ. Press, Cambridge, UK, 2023

work page 2023