---
title: Gaussian Concentration Bounds
url: https://www.emergentmind.com/topics/gaussian-concentration-bounds-gcb
type: topic
---

# Gaussian Concentration Bounds

Searching arXiv for recent and foundational papers on Gaussian concentration bounds and related anti-concentration/concentration frameworks.
Gaussian concentration bounds (GCB) are inequalities of the form
\[
\mathbb{E}_\mu[e^{f-\mathbb{E}_\mu f}] \le \exp\big(C\|\underline{\delta}(f)\|_2^2\big)
\]
or, in random-field notation,
\[
\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,
\]
which imply sub-Gaussian deviation estimates for observables with bounded coordinatewise oscillations [2001.06633] [2602.14618]. In the contemporary literature, the term also encompasses closely related Gaussian-type concentration and anti-concentration phenomena for maxima of Gaussian arrays, quadratic forms in Gaussian variables, suprema of Gaussian processes, and spectral functionals of random matrices, where the central object is not always an exponential-moment inequality but still exhibits Gaussian scaling in the relevant deviation or Lévy concentration function [1910.04259] [2606.25441] [2310.12119] [2406.11453]. Across these settings, the governing theme is that Gaussian structure yields quantitative control of fluctuations through variance-type parameters, spectral data, or local oscillation seminorms, while the sharp form of the bound depends strongly on geometry, dependence, and whether one studies concentration or anti-concentration.

## 1. Definitions and canonical formulations

In the setting of stochastic chains of unbounded memory on countable alphabets, a probability measure \(\mu\) satisfies a Gaussian concentration bound if there exists \(C>0\) such that for all \(f\in\mathcal L\),
\[
\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),
\]
where \(\delta_j(f)\) is the coordinate sensitivity and \(\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2\) [2001.06633]. By Chernoff bounds this yields
\[
\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),
\]
so the terminology “Gaussian” refers to sub-Gaussian tails with variance proxy proportional to \(\|\underline{\delta}(f)\|_2^2\) [2001.06633].

For random fields on \(\mathbb Z^d\), the analogous definition is
\[
\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,
\]
for every local \(f:B^{\mathbb Z^d}\to\mathbb R\) with bounded differences and every \(\lambda>0\), where \(\delta_j f\) is the single-site oscillation and \(\|\delta f\|_2\) is the \(\ell^2\)-norm of the oscillation vector [2602.14618]. This is equivalent, up to constants, to
\[
\mathbb P\big(|f(Y)-\mathbb E f(Y)|>u\big) \le 2\exp\Big(-\frac{u^2}{2C\,\|\delta f\|_2^2}\Big),
\]
which again identifies \(C\) as the intrinsic concentration constant of the field [2602.14618].

A distinct but related formulation appears in anti-concentration theory through the Lévy concentration function
\[
Q_Y(\varepsilon) := \sup_{x\in\mathbb R}\mathbb P(x<Y\le x+\varepsilon),
\]
or equivalently over all intervals of length at most \(\varepsilon\) [2606.25441]. In that language, Gaussian-type behavior means \(Q_Y(\varepsilon)\) is of order \(\varepsilon/\sigma\), as for a one-dimensional Gaussian, rather than decaying with a non-Gaussian power law or logarithmic correction [2606.25441] [2310.12119].

## 2. Product measures, Markov chains, and infinite-memory processes

For independent sequences, GCB reduces to classical bounded-differences concentration. In the SCUM framework, if the kernel \(g\) does not depend on the past, then \(\Osc_j(g)=\Var_j(g)=0\), \(\Delta(g)=\Gamma(g)=1\), and the constant is \(C=1/8\), recovering McDiarmid’s inequality with optimal constant [2001.06633]. For one-step Markov chains, \(\Var_0(g)=\Osc_1(g)=d(Q)\), the Dobrushin coefficient, and the resulting concentration constant is \((1-d(Q))^{-2}/8\) under \(d(Q)<1\) [2001.06633].

The SCUM paper isolates two sharp regimes. Under the oscillation condition
\[
\Delta(g):=1-\sum_{j=1}^\infty \Osc_j(g)>0,
\]
one has
\[
\sup_{x\in X^-}\mathbb E_{P^x}\Big[e^{\theta(f-\mathbb E_{P^x}f)}\Big]
\le
\exp\Big(\tfrac{\theta^2}{8}\Delta(g)^{-2}\|\underline{\delta}(f)\|_2^2\Big),
\]
and hence
\[
\sup_{x\in X^-}P^x\big(|f-\mathbb E_{P^x}f|>u\big)
\le
2\exp\Big(-\frac{2u^2}{\Delta(g)^{-2}\|\underline{\delta}(f)\|_2^2}\Big)
\]
[2001.06633]. Under the variation condition
\[
\Gamma(g):=\prod_{j=0}^\infty(1-\Var_j(g))>0,
\]
the same bounds hold with \(\Delta(g)\) replaced by \(\Gamma(g)\) [2001.06633].

These conditions are described as essentially optimal. If a continuous, strongly non-null kernel admits at least two distinct ergodic compatible measures, then none of these ergodic measures satisfies a GCB [2001.06633]. Likewise, there are examples with \(\sum_j\Osc_j(g)\in(1,1+\varepsilon]\) or with \(\sum_j\Var_j(g)^{1+\varepsilon}<\infty\) but \(\sum_j\Var_j(g)=\infty\) for which GCB fails [2001.06633]. This rules out a common misconception that summable weak dependence in a loose sense is sufficient; the paper shows that phase transition and heavy-tailed renewal behavior both obstruct Gaussian concentration [2001.06633].

Methodologically, the proofs are coupling-based. A martingale decomposition controls increments \(V_k\), and a general coupling theorem yields
\[
\mathbb E_\mu[e^{\theta(f-\mathbb E_\mu f)}]\le
\exp\Big(\tfrac{\theta^2}{8}(1+r)^2\|\underline{\delta}(f)\|_2^2\Big),
\]
provided the cumulative disagreement probability \(r\) is finite [2001.06633]. Maximal coupling and renewal-type inequalities then show \(1+r\le \Delta(g)^{-1}\) or \(1+r\le \Gamma(g)^{-1}\) [2001.06633]. This suggests a structural principle: in infinite-memory systems, GCB is governed less by spectral gap technology than by quantitative control of influence propagation under tailored couplings.

## 3. Finitary codings, random fields, and sharp moment criteria

For dependent random fields obtained from i.i.d. inputs, finitary coding provides another route to GCB. A coding map \(\varphi:A^{\mathbb Z^d}\to B^{\mathbb Z^d}\) is shift-equivariant, and the coding radius \(r_\varphi(x)\) records the finite, configuration-dependent input window needed to determine the output at the origin [2602.14618]. If \(Y=\varphi(X)\) is a finitary coding of an i.i.d. field and the coding volume has finite second moment,
\[
\mathbb E[(2r_\varphi(X)+1)^{2d}]<\infty,
\]
then for every local continuous \(f\) with bounded differences,
\[
\log\mathbb E\big[\exp\{\lambda(f(Y)-\mathbb E f(Y))\}\big]
\le
2^d\lambda^2\mathbb E[(2r_\varphi(X)+1)^{2d}]\,\|\delta f\|_2^2
\]
[2602.14618]. Thus \(Y\) satisfies GCB with
\[
C=2^{d+1}\,\mathbb E[(2r_\varphi(X)+1)^{2d}]
\]
[2602.14618].

Under the short-range factorization property, a first-moment condition suffices. If \(\varphi\) satisfies SRFP with constant \(\alpha\in(0,1]\), then
\[
\log\mathbb E\big[\exp\{\lambda(f(Y)-\mathbb E f(Y))\}\big]
\le
3\,\alpha^{-d}\lambda^2\big(\mathbb E[(2r_\varphi(X)+1)^d]\big)^2\,\|\delta f\|_2^2,
\]
so finite first moment of the coding volume already implies GCB [2602.14618]. The paper states that these moment conditions are sharp: in complete generality the second-moment assumption is sharp, while under SRFP the first-moment assumption is sharp [2602.14618].

The key analytic tool is a refinement of bounded-differences due to Marton. If \(g\) is local and
\[
|g(x)-g(x')|\le \sum_{i\in\dep(g)} c_i(x)\,1_{\{x_i\neq x'_i\}},
\]
then
\[
\log\mathbb E\big[e^{\lambda(g(X)-\mathbb E g(X))}\big]
\le
\frac{\lambda^2}{2}\sum_{i\in\dep(g)}\mathbb E(c_i^2(X))
\]
[2602.14618]. In the coding setting, the influences \(c_i(x)\) are themselves random because whether input site \(i\) affects an output observable depends on the realized coding radii [2602.14618]. The GCB constant is then controlled by overlap statistics of these random influence sets.

A major application is that GCB coincides with the uniqueness regime for several lattice models. For the ferromagnetic Ising model on \(\mathbb Z^d\), the unique Gibbs measure \(\nu_\beta\) satisfies GCB iff \(\beta<\beta_c(d)\) [2602.14618]. For the random-cluster model with \(q>1\), GCB holds in the subcritical phase \(p<p_c(q)\) and fails in the phase with multiple Gibbs states [2602.14618]. The same uniqueness-regime criterion is obtained for Potts models, high-color proper colorings, and several one-dimensional processes via finitary codings and return-time conditions [2602.14618]. This establishes a strong link between concentration, coding representations, and phase transitions.

## 4. Diffusions, Euler schemes, and dynamical preservation of GCB

The evolution of Gaussian concentration under stochastic dynamics can be studied either on continuous state spaces via diffusions or on spin systems via probabilistic cellular automata. For diffusions on \(\mathbb R^d\), the paper on evolution under diffusions asks whether an initial measure \(\mu_0\) satisfying GCB\((D_0)\) yields \(\mu_t=\mu_0S_t\) with GCB\((D_t)\) for the Markov semigroup \(S_t\) [1903.07915]. Under assumptions (H1), (H2), and additional bounds involving functions \(m,a,\beta\), Theorem 3.1 states that if \(\mu_0\) satisfies GCB\((D_0)\), then for each \(0<t<T\), \(\mu_t\) satisfies GCB\((D_t)\) for some finite \(D_t\) [1903.07915]. However, the paper also shows that GCB may fail at every positive time, may hold for all finite times but be lost at infinity, or may blow up in finite time, depending on the coefficients [1903.07915].

A different route appears for Euler discretizations of diffusions. The key tool there is that the transition density of the Euler scheme satisfies two-sided Gaussian bounds of Aronson type:
\[
C^{-1}p_{c^{-1}}(t_{j'}-t_j,x,x')\le p^\Delta(t_j,t_{j'},x,x')\le C\,p_c(t_{j'}-t_j,x,x')
\]
uniformly in the step size \(\Delta=T/N\) [1001.1347]. These density bounds are then combined with a modification of the Herbst argument. A dominated measure \(\mu\) with density \(m\) satisfying \(m(x)\le \kappa q(x)\) relative to a reference measure \(\gamma(dx)=q(x)\,dx\) that has a log-Sobolev inequality inherits concentration through a bound involving \(W_1(\mu,\gamma)\) [1001.1347].

The resulting Monte Carlo error bound for the Euler scheme is
\[
\mathbb{P}_x\Bigl(\,\big|E_{MC}(M,\Delta)\big| \ge r + \delta_{C,\alpha(T)}\Bigr) \le 2 \exp\Bigl(-\frac{M}{\alpha(T)}\,r^2\Bigr),
\]
for Lipschitz \(f\) with \(|\nabla f(T,\cdot)|_\infty\le 1\), where \(\alpha(T)\) is explicit and independent of \(\Delta\), and \(\delta_{C,\alpha(T)}=2\sqrt{\alpha(T)\log C}\) [1001.1347]. The paper also proves matching Gaussian lower bounds under a growth condition on \(f\), showing the concentration is sharp [1001.1347].

For probabilistic cellular automata, the picture is again dynamical but on configuration spaces. If a measure \(\mu\) on \(\{-1,+1\}^{\mathbb Z^d}\) satisfies \(GCB(C)\) and \(P\) is a PCA transition operator with contraction parameter
\[
\kappa=\Big(\sum_{A\Subset\mathbb Z^d}|A|\,|r_A|\Big)^2,
\]
then \(\mu P\) satisfies \(GCB(c+C\kappa)\), and after \(n\) steps the constant becomes
\[
C_n=
\begin{cases}
c\,\dfrac{1-\kappa^n}{1-\kappa}+C\kappa^n,& \kappa\ne1,\\[0.5em]
cn+C,& \kappa=1,
\end{cases}
\]
where \(c\) is the product-measure GCB constant [2507.05431]. If \(\kappa<1\), the unique stationary measure satisfies \(GCB(c/(1-\kappa))\) [2507.05431]. The same paper proves that for contractive PCA, GCB for the initial spatial measure implies GCB for the path-space measure on space-time configurations, while at the stationary level a space-time GCB forces uniqueness of the translation-invariant stationary measure [2507.05431]. This suggests that in interacting particle systems, GCB is not only a fluctuation estimate but also a strong structural marker of uniqueness.

## 5. Gaussian quadratic forms and adaptive anti-concentration

Quadratic Gaussian functionals form a central non-Lipschitz class where classical concentration tools are often inadequate. The 2026 paper studies
\[
S=\sum_{k=1}^{\infty}\lambda_k(Z_k^2-1)+\mu_k Z_k,
\]
with independent standard Gaussian \(Z_k\) and square-summable coefficient sequences \(\lambda,\mu\in \ell^2\) [2606.25441]. The object of interest is the Lévy concentration function
\[
Q_S(\varepsilon)=\sup_{x\in\mathbb R}\mathbb P(x<S\le x+\varepsilon).
\]
The main theorem gives
\[
Q_S(\varepsilon)\le 35\Bigg[
\frac{\varepsilon}{(\Lambda_1^2+\|\mu\|_2^2)^{1/2}}
+
\mathbf 1\{\varepsilon\le 2\Lambda_1\}
\exp\!\Big(-\frac14\frac{\|\mu\|_2^2}{\Lambda_1^2}\Big)\,\mathcal I(\varepsilon,\lambda)
\Bigg],
\]
where \(\Lambda_1^2=\sum_k\lambda_k^2\) and \(\mathcal I(\varepsilon,\lambda)\) is an explicit piecewise function determined by the weight profile of the \(\lambda_k\) [2606.25441].

This bound is adaptive. If the linear Gaussian part dominates, it recovers Gaussian-type anti-concentration
\[
Q_S(\varepsilon)\sim \frac{\varepsilon}{\|\mu\|_2}
\]
up to constants; if the quadratic part dominates, it yields chi-square-type rates governed by the spectral profile of \(\lambda\) [2606.25441]. In particular, the paper recovers the \(\varepsilon^{1/2}\) behavior for a single chi-square mode, linear \(\varepsilon\) behavior in certain balanced cases, and \(\varepsilon\log(1/\varepsilon)\)-type behavior in critical sign-cancellation regimes [2606.25441]. It does so without assuming finite support of \(\lambda\), lower bounds on \(|\lambda_k|\), sign restrictions, or balance conditions beyond square summability [2606.25441].

The proof is purely Fourier analytic. Petrov’s smoothing inequality bounds \(Q_S(\varepsilon)\) through an integral of \(|\phi_S(t)|\), and the characteristic function factorization
\[
|\phi_{\lambda_k(Z^2-1)+\mu_k Z}(t)|=
\frac{\exp\!\left(-\frac{2\mu_k^2 t^2}{1+4\lambda_k^2 t^2}\right)}
{(1+4\lambda_k^2 t^2)^{1/4}}
\]
makes it possible to split the analysis into small-\(t\) Gaussian and large-\(t\) chi-square regimes [2606.25441]. This is explicitly contrasted with Carbery–Wright, which always gives an \(\varepsilon^{1/2}\)-type rate, and with Kolmogorov–Rogozin, which is too crude to capture the sharp dependence on \(\varepsilon\) [2606.25441].

A related but earlier contribution gives improved concentration bounds for finite-dimensional Gaussian quadratic forms of the type
\[
Q_f\sim \sum_{i=1}^n f(\eta_i)(Z_i+\delta_i)^2,
\]
where \(f\) is monotone [1911.05720]. The paper develops Chernoff-style bounds
\[
\mathbb P(Q_f>q)\le \min_{t\in(0,1/(2d))}\exp\Bigl(\frac{\nu_f(t)}{2}-(q-\xi)t\Bigr),
\]
with \(\xi\) and \(\nu_f(t)\) expressed through trace-like quantities \(\sum \eta_i\), \(\sum \eta_i^2\), and shifted analogues [1911.05720]. The key innovation is an optimal quadratic upper envelope for the log-MGF components \(L_1\) and \(L_2\), yielding substantially tighter bounds than earlier trace-based inequalities [1911.05720]. A plausible implication is that even within quadratic Gaussian functionals, there is no single “Gaussian concentration” template; sharp behavior depends on whether one studies deviations from the mean or anti-concentration near points, and on how much of the quadratic structure is retained in the analytic argument.

## 6. Suprema, maxima, and order statistics of Gaussian processes

For suprema of separable centered Gaussian processes, the 2023 paper establishes matching two-sided anti-concentration bounds in terms of the variance of the supremum itself [2310.12119]. If
\[
\widetilde Z=\sup_{u\in U}|X_u|<\infty\quad\text{a.s.},
\]
then for every \(\varepsilon>0\),
\[
\frac{\varepsilon/\sqrt{12}}{\sqrt{\mathrm{Var}(\widetilde Z)+\varepsilon^2/12}}
\le
Q(\widetilde Z,\varepsilon)
\le
\frac{\varepsilon\sqrt{12}}{\sqrt{\mathrm{Var}(\widetilde Z)+\varepsilon^2/12}}
\]
[2310.12119]. The same statement holds for \(Z=\sup_{u\in U}X_u\) under non-degenerate marginals [2310.12119]. For Gaussian order statistics \(X_{(k)}\) of a finite Gaussian field,
\[
\frac{\varepsilon/\sqrt{12}}{\sqrt{\mathrm{Var}(X_{(k)})+\varepsilon^2/12}}
\le
Q(X_{(k)},\varepsilon)
\le
\frac{\varepsilon\sqrt{12}}{\sqrt{\mathrm{Var}(X_{(k)})+\varepsilon^2/12}}
\]
[2310.12119].

The conceptual content is explicit: the supremum has the same anti-concentration profile, up to constants, as a single Gaussian random variable with variance \(\mathrm{Var}(\sup X_u)\) [2310.12119]. This differs from classical Borell–TIS concentration, which uses \(\sup_u\mathrm{Var}(X_u)\) to control tails of the supremum. Here the local mass \(Q(\sup X_u,\varepsilon)\) is governed by the often much smaller variance of the supremum itself [2310.12119].

To make these bounds useful, the paper also derives variance estimates. If \(Z=\sup_{u\in U}X_u\) and \(\underline{\sigma}^2:=\inf_{u\in U}\mathrm{Var}(X_u)>0\), then
\[
\mathrm{Var}(Z)\ge \frac{1}{15^2}\left(\frac{\underline{\sigma}}{1+\mathbb E[Z/\underline{\sigma}]}\right)^2
\]
[2310.12119]. If additionally \(\bar\sigma^2:=\sup_u\mathrm{Var}(X_u)<\infty\) and \(|\mathrm{Cor}(X_u,X_v)|\le \rho\), then
\[
\mathrm{Var}(Z)\le 4\bar\sigma^2 \wedge \left[
11^2\bar\sigma^2\rho + 60^2\left(\frac{\bar\sigma}{(\mathbb E[Z/\bar\sigma]-3\sqrt{\log 2})_+}\right)^2
\right]
\]
[2310.12119]. These show that in weakly correlated, high-entropy settings, the variance is typically of order \(1/(\mathbb E Z)^2\), a superconcentration phenomenon already known in other Gaussian extremes contexts [2310.12119].

A different but related notion of Gaussian concentration appears in maxima of Gaussian arrays through relative stability. For a triangular array \(\mathcal E=\{\epsilon_p(i)\}\) with standard normal marginals, relative stability means
\[
\frac{1}{u_p}\max_{1\le i\le p}\epsilon_p(i)\xrightarrow{\mathbb P}1,
\]
where \(u_p=\Phi^{\leftarrow}(1-1/p)\sim \sqrt{2\log p}\) [1910.04259]. Uniform relative stability is equivalent to uniformly decreasing dependence for Gaussian arrays [1910.04259]. The main rate theorem states that if \(\tau(p)\to0\) and
\[
\alpha(p):=\frac{\log N_{\mathcal E}(\tau(p))}{\log p}\to0,
\]
then any sequence \(\delta_p\) with
\[
\delta_p\gg \alpha(p)+\tau(p)+\frac1{\log p}
\]
satisfies
\[
\mathbb P\left(\left|\frac{\max_{i\in[p]}\epsilon_p(i)}{u_p}-1\right|>\delta_p\right)\to0
\]
[1910.04259]. In the i.i.d. case this recovers the \(1/\log p\) scale; power-law covariance decay yields \(\delta_p\gg (\log\log p)/\log p\), and logarithmic covariance decay yields \((\log p)^{-\nu/(\nu+1)}\) [1910.04259]. This line of work concerns concentration around deterministic extreme-value scales rather than GCB in exponential-moment form, but it fits the broader Gaussian concentration program by quantifying the rate at which Gaussian maxima collapse onto their canonical asymptotics.

## 7. Other directions: sample means, matrix models, algorithms, and cautionary distinctions

For bounded i.i.d. variables, efficient Gaussian concentration with finite-sample validity can be built by combining Gaussian approximation and concentration. The 2022 paper studies \(S_n=\sqrt n(\bar W_n-\mu)\) for \(W_i\in[0,R]\) and derives non-uniform tail bounds of the form
\[
\mathbb P(S_n>\sigma u)-\Phi^c(u)
\le
\frac{c_1R}{\sigma\sqrt n}\exp\!\Big(-c_2\frac{u^2\sigma^2}{R^2}\Big),
\qquad
u\ge \frac{R}{\sqrt n},
\]
which are finite-sample valid, asymptotically optimal, and sub-Gaussian in \(u\) [2208.09922]. A Wasserstein-based variant gives computable quantile bounds and empirical Berry–Esseen inequalities without prior knowledge of \(\sigma\) [2208.09922]. This suggests that for mean estimation, GCB and Gaussian approximation need not be competing paradigms; they can be hybridized to achieve both exact validity and asymptotically optimal width.

In random matrix theory, “matrix concentration inequalities and free probability” extend Gaussian concentration ideas to spectra of Gaussian matrices. For a self-adjoint Gaussian random matrix \(X\), the paper proves
\[
\mathbb P\Big[ \mathrm d_H(sp(X),sp(X_{\mathrm{free}}))
>
C\tilde v(X)(\log d)^{3/4}+C\sigma_*(X)t
\Big]
\le e^{-t^2},
\]
where \(X_{\mathrm{free}}\) is an associated free model and \(\tilde v,\sigma_*\) are covariance parameters [2406.11453]. Consequently,
\[
\mathbb P\Big[
|\lambda_{\max}(X)-\lambda_{\max}(X_{\mathrm{free}})|
>
C\tilde v(X)(\log d)^{3/4}+C\sigma_*(X)t
\Big]
\le e^{-t^2}
\]
[2406.11453]. These are sub-Gaussian spectral-edge deviations but centered at the free model rather than the mean, and they are sharp enough to determine phase transitions for outliers in nonhomogeneous random matrices [2406.11453].

For high-dimensional sub-Gaussian vectors, another extension studies \(\|QX\|^2\) when \(X\) is sub-Gaussian but not necessarily Gaussian or coordinatewise independent [2305.07885]. The Gaussian benchmark is the Laurent–Massart inequality
\[
\mathbb P\Big( \|\xi\|^2-\mathrm{tr}(B)>2\sqrt{x\,\mathrm{tr}(B^2)}+2x\|B\|\Big)\le e^{-x},
\]
with \(\xi\sim \mathcal N(0,B)\) [2305.07885]. The paper argues that under local smoothness of the log-MGF and effective-dimension assumptions, one can recover nearly Gaussian bounds of the same form for \(\|QX\|^2\) using Laplace approximation, even without Hanson–Wright assumptions [2305.07885].

By contrast, concentration results for probabilistic programs studied through exponential supermartingales generally yield exponential or sub-exponential tails rather than genuinely Gaussian ones. The paper explicitly contrasts this with Azuma–Hoeffding and notes that its synthesis of exponential supermartingales yields bounds such as
\[
\Pr(C\ge d)\le \alpha^{X_0-K}\beta^{-d},
\]
which are often sharper for the given process but are not sub-Gaussian in the classical sense [2008.00425]. This is an important distinction: not every “concentration bound” in the Gaussian tradition is itself Gaussian in scaling, even if it is derived via moment generating functions.

A final caution concerns terminology itself. In some papers, “Gaussian concentration bounds” means an exponential-moment inequality for bounded-difference observables [2001.06633] [2602.14618] [2507.05431] [1903.07915]. In others, it refers to Gaussian-type tails or anti-concentration for specialized functionals such as maxima, order statistics, or quadratic forms [2606.25441] [2310.12119] [1910.04259] [1911.05720]. The unifying idea is Gaussian scaling, but the operative object may be an MGF bound, a tail bound, a concentration function, or a variance-controlled spectral estimate. This suggests that “GCB” is best understood as a family of structurally related principles rather than a single canonical inequality.

Source: https://www.emergentmind.com/topics/gaussian-concentration-bounds-gcb