On the subgaussian comparison theorem

Ramon van Handel

On the subgaussian comparison theorem

T0 review · 0 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Any 1-subgaussian random vector is dominated in convex order by a universal constant times a standard Gaussian.

desk verdict A short, honest note that upgrades Talagrand's subgaussian comparison to convex-order domination; the proof is sound and the flagged integrability gap is not real. read the letter →

arxiv 2512.18588 v2 pith:EF4HMHN2 submitted 2025-12-21 math.PR

Ramon van Handel This is my paper · ORCID

classification math.PR MSC 60E1560G15

keywords 1-subgaussianvectorconvexordersubgaussiancomparisonmajorizingmeasuretensorizationGaussiandominationrandomprocesssupremafunctions

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 proves that every zero-mean subgaussian random vector — one whose one-dimensional projections have tails no heavier than a standard Gaussian — is dominated, in the convex order, by a universal constant times a standard Gaussian. That means every convex function of the vector has expectation no larger than the corresponding expectation for the scaled Gaussian. This strengthens the classical subgaussian comparison theorem, which covered only positively homogeneous convex functions. The proof works through a general statement for random processes and rests on a tensorization principle that reduces the comparison to stationary processes, bypassing the heavy machinery of majorizing measures.

What carries the argument

The central machinery is the coupling functional F(X, μ) = sup over couplings of E[X_Z], where Z has distribution μ on the index set; it encodes expected suprema through the identity E sup_t (X_t + m_t) = sup_μ (F(X, μ) + ∫ m_t dμ). The key step is a tensorization principle that expresses F(X, μ) as the limit of expected suprema of an auxiliary stationary process obtained by averaging i.i.d. copies of X over sequences whose empirical distribution is μ. Because the auxiliary process is stationary, the classical chaining bound for stationary Gaussian processes applies, yielding F(X, μ) ≤ c F(G, μ) and hence the desired comparison.

What would settle it

Take a two-point index set, let Y be a symmetric heavy-tailed variable with infinite first moment and Z an independent standard normal; set X_1 = Y and X_2 = Y + Z. The increment X_1 - X_2 is Z, so the subgaussian increment assumption holds, but E sup_t X_t is infinite, showing the process-level theorem needs an explicit integrability hypothesis.

Watch

Extended reading notes

Core claim

The central discovery is that if X is a centered random vector in R^n and P(|<v,X>| > x) ≤ 2 exp(-x^2/2) for every unit vector v, then E[f(X)] ≤ E[f(cG)] for every convex f, with a universal constant c and G ~ N(0,I_n). By a classical theorem on couplings, this is equivalent to constructing X together with G so that X = c E[G|X]. The proof is obtained from a more general process-level comparison: for any finite index set, any process with subgaussian increments dominated by a Gaussian process has its shifted supremum bounded by the shifted Gaussian supremum times c. The argument avoids the machinery of majorizing measures by using a certain coupling functional and a tensorization principle t

Load-bearing premise

The proof requires each marginal X_t to have a finite first moment, so that the coupling functional and the expected supremum are well-defined; subgaussian increment tails alone do not force this, so the stated theorem silently assumes this integrability.

Editorial extensions

If this is right

Every 1-subgaussian vector can be coupled with a standard Gaussian so that the vector is exactly the conditional expectation of the Gaussian given it, a property much stronger than tail domination.
The process-level version compares shifted suprema: for any real shifts m_t, E sup_t (X_t + m_t) ≤ E sup_t (c G_t + m_t), covering non-centered processes and many empirical-process settings.
Because convex functions are supremums of affine functions, the comparison transfers automatically to every convex functional, including norms, maximum of linear forms, and envelope functions.
The proof yields a new, elementary route to the majorizing measure theorem: only stationarity-based chaining plus tensorization is needed, avoiding the original geometric construction.

Reading between the lines

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

If the universal constant c is made explicit, the comparison could transfer sharp Gaussian concentration inequalities to subgaussian vectors for all convex Lipschitz functions, likely with near-optimal dimension dependence.
The tensorization recipe suggests a general strategy: to compare a non-stationary process to a target, one can first compare stationary averages; this may extend to target processes beyond Gaussians, such as other symmetric or exchangeable ensembles.
The integrability gap in the proof implies that the cleanest infinite-dimensional formulation will need an explicit finite-moment condition on the marginals; a counterexample with heavy-tailed shifts shows that the tail assumption alone is insufficient.

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

Good to see this. The note does what it says: proves that any 1-subgaussian vector in R^n is dominated in convex order by a universal constant times a standard Gaussian. That is a real strengthening of Talagrand's theorem, which only covered 1-homogeneous convex functions. The proof is clean: Liu's tensorization reduces Fernique's functional for arbitrary subgaussian processes to the stationary case, and the classical Dudley–Fernique bound finishes the job. Theorem 1.3, with the offsets m_t, is a useful extra.

The exposition is a genuine plus. The paper is upfront about which parts are new—the convex-ordering statements—and which are exposition of Liu's tensorization principle. The proof of Proposition 3.1 is self-contained and correct. Corollary 1.2, the coupling X = cE[G|X], follows by Strassen and is a nice payoff.

The reader's conditional verdict flags a missing integrability assumption in Theorem 1.3. I think that concern does not land. 'Centered random process' should mean E[X_t]=0 for every t, which on a finite index set implies max_t E|X_t| < infinity, exactly what Proposition 3.1 needs. So the proof is fine as written. That said, stating the assumption explicitly would cost one line and would preempt the ambiguity; worth doing in revision.

The larger soft spot, if you want one, is that this is a short expository note built on a recent arXiv paper by Liu. The step from Liu's tensorization to convex order is elementary once you see it. That is not a criticism—a good deal of mathematical progress is noticing that an existing tool proves a stronger statement than the author claimed—but it does mean the novelty is 'new corollary plus clean exposition' rather than a new method.

Bottom line: if you work on subgaussian processes, convex order, or high-dimensional concentration, this is worth reading and citing. The proof is clear enough to be used directly. I'd send it to a competent referee; the main thing to check is whether the convex-order step from Theorem 1.3 to Theorem 1.1 is fully justified for all convex functions—it is, via finite maxima of affine functions and monotone convergence. Minor revision, no serious objections.

Referee Report

0 major / 4 minor

Summary. The paper proves that every centered 1-subgaussian random vector X in R^n satisfies E[f(X)] ≤ E[f(cG)] for every convex function f, where G is a standard Gaussian vector and c is a universal constant. This is obtained as a consequence of a more general comparison theorem for random processes (Theorem 1.3): for any centered subgaussian process (X_t) whose increments are dominated by those of a centered Gaussian process (G_t), one has E[sup_t (X_t+m_t)] ≤ E[sup_t (cG_t+m_t)] for any shift (m_t). The proof combines a tensorization principle of J. Liu, which reduces Fernique's functional F(X,μ) to the expected supremum of a stationary process, with the classical Dudley–Fernique chaining bounds for stationary processes. The paper also deduces a Strassen-type coupling X = c E[G|X] as a corollary.

Significance. If correct, Theorem 1.1 is a substantial strengthening of Talagrand's subgaussian comparison theorem, which only covers 1-homogeneous convex functions. The proof is short, self-contained modulo standard chaining results, and gives a clear probabilistic interpretation of Fernique's functional through Liu's tensorization idea. The tensorization principle is proved in full, and the dependence on Liu's recent work is explicitly acknowledged. The argument is non-circular and uses only classical tools (Dudley–Fernique, Birkhoff's theorem, Strassen's theorem). The result is likely to be of interest to researchers in Gaussian processes, empirical process theory, and concentration of measure.

minor comments (4)

[Section 3.1 (Proposition 3.1)] In the statement of Proposition 3.1, X_t is defined 'for every M∈N and t∈T^M', but the set T_N(μ) in the displayed limit consists of sequences of length NK (with μ∈P_K). The proof sets M=NK. Please make this explicit by writing M=NK or by defining T_{N,K}(μ) to avoid a confusing mismatch.
[Section 1.1 / Theorem 1.3] Theorem 1.3 calls X a 'centered random process' but does not define the term. Proposition 3.1 requires max_t E|X_t|<∞, and this is indeed implied if 'centered' means E X_t=0 with finite first moment. Add a parenthetical definition so readers do not mistakenly think the tail-bound condition alone suffices; this would also fully close the apparent integrability gap.
[Lemma 4.1] The last displayed estimate has a factor-of-2 discrepancy: the preceding bounds yield |F(P_X,μ)-F(P_X,μ')| ≤ 2r||μ-μ'||_{TV} + 2E[||X||1_{||X||>r}], not r||μ-μ'||_{TV} + 2E[...] as written. Since r is arbitrary, the asserted continuity in total variation is unaffected, but the displayed inequality should be corrected.
[Section 4 (proof of Theorem 1.3)] In the chain of inequalities, the limits are taken as n→∞, while Proposition 3.1 uses N→∞; the notations n, N, M are used interchangeably. Standardize the notation to avoid confusion. The constant C from the tensorized tail bound is absorbed into the final universal constant; this is fine but could be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proof reduces to an independently established stationary comparison theorem and a tensorization lemma proved in the note.

full rationale

The derivation chain for Theorem 1.3 is not circular. It uses: (i) the identity E[sup_t(X_t+m_t)] = sup_mu(F(X,mu)+∫m_t dmu), which is a definitional reformulation; (ii) Proposition 3.1, proved in the note via Birkhoff's theorem and W1-continuity, representing F(X,mu) as a limit of expected suprema of tensorized processes; (iii) a direct tail-bound calculation showing the tensorized subgaussian process inherits the same subgaussian tails (up to a constant) with respect to the tensorized Gaussian metric; and (iv) the classical Dudley–Fernique Theorem 2.1 for stationary Gaussian comparison, cited to Ledoux rather than to the author. Theorem 2.1 is an external benchmark, not a restatement of the theorem being proved: it is the special case of Theorem 1.3 with m_t=0 and stationary G, but the proof does not derive Theorem 2.1 from Theorem 1.3, so using it as a lemma is legitimate generalization rather than circularity. No parameter is fitted to the conclusion, no conclusion is assumed under another name, and no load-bearing self-citation or imported uniqueness theorem appears. The integrability condition max_t ||X_t||_1<∞ required by Proposition 3.1 is implied by the stated centeredness hypothesis on the finite index set (and is automatic for the subgaussian vector application), so the proof's assumptions cover its use. Hence there are no circular steps.

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

The proof rests on classical probability theorems (Dudley–Fernique, Birkhoff, Strassen) and on Liu's tensorization principle, which the note proves in full. No free parameters are fitted. The only non-explicit ingredient is the integrability of the random process.

assumptions (5)

standard math Dudley–Fernique theorem for stationary Gaussian processes (Theorem 2.1)
Used as a black box in the proof of Theorem 1.3 to compare the expected supremum of the tensorized subgaussian process to that of the tensorized stationary Gaussian process. Cited to [2] and [5].
standard math Birkhoff's theorem: doubly stochastic matrices are convex combinations of permutation matrices
Used in the proof of Proposition 3.1 to reduce the supremum over all couplings of the empirical measure with µ to the supremum over permutations.
standard math Strassen's theorem: convex-order domination is equivalent to the existence of a martingale coupling
Used to derive Corollary 1.2 from Theorem 1.1.
standard math Monotone convergence and approximation of any convex function by an increasing sequence of finite maxima of affine functions
Used to pass from the finite-maximum case in Theorem 1.3 to all convex functions in Theorem 1.1.
domain assumption Integrability of the process: max_t E|X_t| < ∞
Required by Proposition 3.1, but not stated in Theorem 1.3. Without it, Fernique's functional F(X,µ) and E[sup(X_t+m_t)] may be undefined. This is the weakest technical assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On the subgaussian comparison theorem." pith.science (2026). https://pith.science/paper/EF4HMHN2

@misc{pith2026251218588,
  author       = {Pith},
  title        = {Pith review of: On the subgaussian comparison theorem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EF4HMHN2}},
  note         = {Machine review of arXiv:2512.18588}
}

abstract

The aim of this expository note is to prove that any $1$-subgaussian random vector is dominated in the convex ordering by a universal constant times a standard Gaussian vector. This strengthens Talagrand's celebrated subgaussian comparison theorem. The proof combines a tensorization argument due to J. Liu with ideas that date back to the work of Fernique.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

A Bayesian Proof of the Bernoulli Theorem
math.PR 2026-08 conditional novelty 8.0 of 10

For any finite set T and law mu on T, the largest expected Bernoulli-process value under coupling with mu is, up to universal constants, the rate-distortion integral of RD_mu(t) and the ell-1-plus-Gaussian decompositi...
On Talagrand's Convexity Conjecture
math.PR 2026-05 unverdicted novelty 8.0 of 10

Any random vector in R^n dominated in convex order by a standard Gaussian is the sum of three standard Gaussians, solving Talagrand's convexity problem and a weak combinatorial analogue.
Sharp Frobenius-Norm Concentration for Sample Moment Tensors
math.PR 2026-08 accept novelty 7.0 of 10

Sharp dimension-free Frobenius-norm concentration inequalities for sample moment tensors, with matching two-sided Gaussian estimates and an even/odd parity effect.

Reference graph

Works this paper leans on

12 extracted references · 1 linked inside Pith · cited by 3 Pith papers

[1]

Csiszár and J

I. Csiszár and J. Körner.Information theory. Cambridge University Press, Cambridge, second edition, 2011. Coding theorems for discrete memoryless systems

2011
[2]

Fernique

X. Fernique. Regularité des trajectoires des fonctions aléatoires gaussiennes. InÉcole d’Été de Probabilités de Saint-Flour, IV-1974, volume 480 ofLecture Notes in Math., pages 1–96. Springer, 1975

1974
[3]

Fernique

X. Fernique. Évaluations de processus gaussiens composés. InProbability in Banach spaces (Proc. First Internat. Conf., Oberwolfach, 1975), volume 526 ofLecture Notes in Math., pages 67–83. Springer, 1976

1975
[4]

Fernique

X. Fernique. Évaluations de certaines fonctionnelles associées à des fonctions aléatoires gaussi- ennes.Probab. Math. Statist., 2(1):1–29, 1981

1981
[5]

M. Ledoux. Isoperimetry and Gaussian analysis. InLectures on probability theory and statis- tics (Saint-Flour, 1994), volume 1648 ofLecture Notes in Math., pages 165–294. Springer, 1996

1994
[6]

J. Liu. Simple and sharp generalization bounds via lifting, 2025. arxiv:2508.18682v3

arXiv 2025
[7]

Strassen

V. Strassen. The existence of probability measures with given marginals.Ann. Math. Statist., 36:423–439, 1965

1965
[8]

Talagrand

M. Talagrand. Regularity of Gaussian processes.Acta Math., 159(1-2):99–149, 1987

1987

Show all 12 references

[9]

Talagrand.Upper and lower bounds for stochastic processes—decomposition theorems, volume 60 ofErgebnisse der Mathematik und ihrer Grenzgebiete

M. Talagrand.Upper and lower bounds for stochastic processes—decomposition theorems, volume 60 ofErgebnisse der Mathematik und ihrer Grenzgebiete. Springer, 2nd edition, 2021

2021
[10]

S. A. van de Geer.Applications of empirical process theory, volume 6 ofCambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 2000

2000
[11]

Vershynin.High-dimensional probability, volume 47 ofCambridge Series in Statistical and Probabilistic Mathematics

R. Vershynin.High-dimensional probability, volume 47 ofCambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 2018

2018
[12]

Villani.Topics in optimal transportation, volume 58 ofGraduate Studies in Mathematics

C. Villani.Topics in optimal transportation, volume 58 ofGraduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2003. Department of Mathematics, Princeton University, Princeton, NJ 08544, USA Email address:rvan@math.princeton.edu

2003

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Csiszár and J

I. Csiszár and J. Körner.Information theory. Cambridge University Press, Cambridge, second edition, 2011. Coding theorems for discrete memoryless systems

2011

[2] [2]

Fernique

X. Fernique. Regularité des trajectoires des fonctions aléatoires gaussiennes. InÉcole d’Été de Probabilités de Saint-Flour, IV-1974, volume 480 ofLecture Notes in Math., pages 1–96. Springer, 1975

1974

[3] [3]

Fernique

X. Fernique. Évaluations de processus gaussiens composés. InProbability in Banach spaces (Proc. First Internat. Conf., Oberwolfach, 1975), volume 526 ofLecture Notes in Math., pages 67–83. Springer, 1976

1975

[4] [4]

Fernique

X. Fernique. Évaluations de certaines fonctionnelles associées à des fonctions aléatoires gaussi- ennes.Probab. Math. Statist., 2(1):1–29, 1981

1981

[5] [5]

M. Ledoux. Isoperimetry and Gaussian analysis. InLectures on probability theory and statis- tics (Saint-Flour, 1994), volume 1648 ofLecture Notes in Math., pages 165–294. Springer, 1996

1994

[6] [6]

J. Liu. Simple and sharp generalization bounds via lifting, 2025. arxiv:2508.18682v3

arXiv 2025

[7] [7]

Strassen

V. Strassen. The existence of probability measures with given marginals.Ann. Math. Statist., 36:423–439, 1965

1965

[8] [8]

Talagrand

M. Talagrand. Regularity of Gaussian processes.Acta Math., 159(1-2):99–149, 1987

1987