Gaussian Concentration Bounds

• Gaussian Concentration Bounds are inequalities that link exponential moment bounds to variance proxies via coordinatewise oscillations, yielding sub-Gaussian tail estimates.
• They are applied across diverse settings including Markov chains, random fields, diffusions, and coding frameworks to govern fluctuations and phase transitions.
• Methodologically, these bounds are established using coupling techniques, martingale decompositions, and Fourier analysis to capture both concentration and anti-concentration phenomena.

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 (Chazottes et al., 2020, Chazottes et al., 16 Feb 2026). 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 (Kartsioukas et al., 2019, Choudhary et al., 24 Jun 2026, Giessing, 2023, Bandeira et al., 2024). 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 $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$ (Chazottes et al., 2020). 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 $\|\underline{\delta}(f)\|_2^2$ (Chazottes et al., 2020).

For random fields on $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$0, the analogous definition is

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$1

for every local $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$2 with bounded differences and every $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$3, where $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$4 is the single-site oscillation and $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$5 is the $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$6-norm of the oscillation vector (Chazottes et al., 16 Feb 2026). This is equivalent, up to constants, to

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$7

which again identifies $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$8 as the intrinsic concentration constant of the field (Chazottes et al., 16 Feb 2026).

A distinct but related formulation appears in anti-concentration theory through the Lévy concentration function

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$9

or equivalently over all intervals of length at most $\mu$0 (Choudhary et al., 24 Jun 2026). In that language, Gaussian-type behavior means $\mu$1 is of order $\mu$2, as for a one-dimensional Gaussian, rather than decaying with a non-Gaussian power law or logarithmic correction (Choudhary et al., 24 Jun 2026, Giessing, 2023).

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 $\mu$3 does not depend on the past, then $\mu$4, $\mu$5, and the constant is $\mu$6, recovering McDiarmid’s inequality with optimal constant (Chazottes et al., 2020). For one-step Markov chains, $\mu$7, the Dobrushin coefficient, and the resulting concentration constant is $\mu$8 under $\mu$9 (Chazottes et al., 2020).

The SCUM paper isolates two sharp regimes. Under the oscillation condition

$C>0$0

one has

$C>0$1

and hence

$C>0$2

(Chazottes et al., 2020). Under the variation condition

$C>0$3

the same bounds hold with $C>0$4 replaced by $C>0$5 (Chazottes et al., 2020).

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 (Chazottes et al., 2020). Likewise, there are examples with $C>0$6 or with $C>0$7 but $C>0$8 for which GCB fails (Chazottes et al., 2020). 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 (Chazottes et al., 2020).

Methodologically, the proofs are coupling-based. A martingale decomposition controls increments $C>0$9, and a general coupling theorem yields

$f\in\mathcal L$0

provided the cumulative disagreement probability $f\in\mathcal L$1 is finite (Chazottes et al., 2020). Maximal coupling and renewal-type inequalities then show $f\in\mathcal L$2 or $f\in\mathcal L$3 (Chazottes et al., 2020). 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 $f\in\mathcal L$4 is shift-equivariant, and the coding radius $f\in\mathcal L$5 records the finite, configuration-dependent input window needed to determine the output at the origin (Chazottes et al., 16 Feb 2026). If $f\in\mathcal L$6 is a finitary coding of an i.i.d. field and the coding volume has finite second moment,

$f\in\mathcal L$7

then for every local continuous $f\in\mathcal L$8 with bounded differences,

$f\in\mathcal L$9

(Chazottes et al., 16 Feb 2026). Thus $$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$0 satisfies GCB with

$$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$1

(Chazottes et al., 16 Feb 2026).

Under the short-range factorization property, a first-moment condition suffices. If $$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$2 satisfies SRFP with constant $$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$3, then

$$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$4

so finite first moment of the coding volume already implies GCB (Chazottes et al., 16 Feb 2026). 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 (Chazottes et al., 16 Feb 2026).

The key analytic tool is a refinement of bounded-differences due to Marton. If $$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$5 is local and

$$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$6

then

$$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$7

(Chazottes et al., 16 Feb 2026). In the coding setting, the influences $$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$8 are themselves random because whether input site $$\mathbb{E}_\mu\big[e^{f-\mathbb{E}_\mu f}\big] \le \exp\big( C\,\|\underline{\delta}(f)\|_2^2\big),$$9 affects an output observable depends on the realized coding radii (Chazottes et al., 16 Feb 2026). 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 $\delta_j(f)$0, the unique Gibbs measure $\delta_j(f)$1 satisfies GCB iff $\delta_j(f)$2 (Chazottes et al., 16 Feb 2026). For the random-cluster model with $\delta_j(f)$3, GCB holds in the subcritical phase $\delta_j(f)$4 and fails in the phase with multiple Gibbs states (Chazottes et al., 16 Feb 2026). 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 (Chazottes et al., 16 Feb 2026). 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 $\delta_j(f)$5, the paper on evolution under diffusions asks whether an initial measure $\delta_j(f)$6 satisfying GCB$\delta_j(f)$7 yields $\delta_j(f)$8 with GCB$\delta_j(f)$9 for the Markov semigroup $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$0 (Chazottes et al., 2019). Under assumptions (H1), (H2), and additional bounds involving functions $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$1, Theorem 3.1 states that if $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$2 satisfies GCB$$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$3, then for each $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$4, $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$5 satisfies GCB$$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$6 for some finite $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$7 (Chazottes et al., 2019). 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 (Chazottes et al., 2019).

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: $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$8 uniformly in the step size $$\|\underline{\delta}(f)\|_2^2=\sum_{j=0}^\infty \delta_j(f)^2$$9 (Lemaire et al., 2010). These density bounds are then combined with a modification of the Herbst argument. A dominated measure $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$0 with density $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$1 satisfying $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$2 relative to a reference measure $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$3 that has a log-Sobolev inequality inherits concentration through a bound involving $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$4 (Lemaire et al., 2010).

The resulting Monte Carlo error bound for the Euler scheme is

$$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$5

for Lipschitz $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$6 with $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$7, where $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$8 is explicit and independent of $$\mu\big(|f-\mathbb{E}_\mu f|>u\big) \le 2\exp\left(-\frac{u^2}{4C\|\underline{\delta}(f)\|_2^2}\right),$$9, and $\|\underline{\delta}(f)\|_2^2$0 (Lemaire et al., 2010). The paper also proves matching Gaussian lower bounds under a growth condition on $\|\underline{\delta}(f)\|_2^2$1, showing the concentration is sharp (Lemaire et al., 2010).

For probabilistic cellular automata, the picture is again dynamical but on configuration spaces. If a measure $\|\underline{\delta}(f)\|_2^2$2 on $\|\underline{\delta}(f)\|_2^2$3 satisfies $\|\underline{\delta}(f)\|_2^2$4 and $\|\underline{\delta}(f)\|_2^2$5 is a PCA transition operator with contraction parameter

$\|\underline{\delta}(f)\|_2^2$6

then $\|\underline{\delta}(f)\|_2^2$7 satisfies $\|\underline{\delta}(f)\|_2^2$8, and after $\|\underline{\delta}(f)\|_2^2$9 steps the constant becomes

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$00

where $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$01 is the product-measure GCB constant (Chazottes et al., 7 Jul 2025). If $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$02, the unique stationary measure satisfies $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$03 (Chazottes et al., 7 Jul 2025). 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 (Chazottes et al., 7 Jul 2025). 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

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$04

with independent standard Gaussian $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$05 and square-summable coefficient sequences $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$06 (Choudhary et al., 24 Jun 2026). The object of interest is the Lévy concentration function

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$07

The main theorem gives

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$08

where $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$09 and $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$10 is an explicit piecewise function determined by the weight profile of the $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$11 (Choudhary et al., 24 Jun 2026).

This bound is adaptive. If the linear Gaussian part dominates, it recovers Gaussian-type anti-concentration

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$12

up to constants; if the quadratic part dominates, it yields chi-square-type rates governed by the spectral profile of $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$13 (Choudhary et al., 24 Jun 2026). In particular, the paper recovers the $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$14 behavior for a single chi-square mode, linear $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$15 behavior in certain balanced cases, and $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$16-type behavior in critical sign-cancellation regimes (Choudhary et al., 24 Jun 2026). It does so without assuming finite support of $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$17, lower bounds on $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$18, sign restrictions, or balance conditions beyond square summability (Choudhary et al., 24 Jun 2026).

The proof is purely Fourier analytic. Petrov’s smoothing inequality bounds $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$19 through an integral of $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$20, and the characteristic function factorization

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$21

makes it possible to split the analysis into small-$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$22 Gaussian and large-$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$23 chi-square regimes (Choudhary et al., 24 Jun 2026). This is explicitly contrasted with Carbery–Wright, which always gives an $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$24-type rate, and with Kolmogorov–Rogozin, which is too crude to capture the sharp dependence on $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$25 (Choudhary et al., 24 Jun 2026).

A related but earlier contribution gives improved concentration bounds for finite-dimensional Gaussian quadratic forms of the type

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$26

where $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$27 is monotone (Gallagher et al., 2019). The paper develops Chernoff-style bounds

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$28

with $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$29 and $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$30 expressed through trace-like quantities $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$31, $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$32, and shifted analogues (Gallagher et al., 2019). The key innovation is an optimal quadratic upper envelope for the log-MGF components $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$33 and $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$34, yielding substantially tighter bounds than earlier trace-based inequalities (Gallagher et al., 2019). 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 (Giessing, 2023). If

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$35

then for every $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$36,

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$37

(Giessing, 2023). The same statement holds for $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$38 under non-degenerate marginals (Giessing, 2023). For Gaussian order statistics $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$39 of a finite Gaussian field,

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$40

(Giessing, 2023).

The conceptual content is explicit: the supremum has the same anti-concentration profile, up to constants, as a single Gaussian random variable with variance $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$41 (Giessing, 2023). This differs from classical Borell–TIS concentration, which uses $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$42 to control tails of the supremum. Here the local mass $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$43 is governed by the often much smaller variance of the supremum itself (Giessing, 2023).

To make these bounds useful, the paper also derives variance estimates. If $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$44 and $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$45, then

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$46

(Giessing, 2023). If additionally $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$47 and $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$48, then

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$49

(Giessing, 2023). These show that in weakly correlated, high-entropy settings, the variance is typically of order $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$50, a superconcentration phenomenon already known in other Gaussian extremes contexts (Giessing, 2023).

A different but related notion of Gaussian concentration appears in maxima of Gaussian arrays through relative stability. For a triangular array $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$51 with standard normal marginals, relative stability means

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$52

where $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$53 (Kartsioukas et al., 2019). Uniform relative stability is equivalent to uniformly decreasing dependence for Gaussian arrays (Kartsioukas et al., 2019). The main rate theorem states that if $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$54 and

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$55

then any sequence $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$56 with

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$57

satisfies

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$58

(Kartsioukas et al., 2019). In the i.i.d. case this recovers the $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$59 scale; power-law covariance decay yields $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$60, and logarithmic covariance decay yields $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$61 (Kartsioukas et al., 2019). 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 $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$62 for $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$63 and derives non-uniform tail bounds of the form

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$64

which are finite-sample valid, asymptotically optimal, and sub-Gaussian in $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$65 (Austern et al., 2022). A Wasserstein-based variant gives computable quantile bounds and empirical Berry–Esseen inequalities without prior knowledge of $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$66 (Austern et al., 2022). 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 $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$67, the paper proves

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$68

where $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$69 is an associated free model and $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$70 are covariance parameters (Bandeira et al., 2024). Consequently,

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$71

(Bandeira et al., 2024). 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 (Bandeira et al., 2024).

For high-dimensional sub-Gaussian vectors, another extension studies $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$72 when $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$73 is sub-Gaussian but not necessarily Gaussian or coordinatewise independent (Spokoiny, 2023). The Gaussian benchmark is the Laurent–Massart inequality

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$74

with $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$75 (Spokoiny, 2023). 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 $$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$76 using Laplace approximation, even without Hanson–Wright assumptions (Spokoiny, 2023).

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

$$\log\mathbb E\big(e^{\lambda (f(Y)-\mathbb E f(Y))}\big)\le \frac{C}{2}\,\lambda^2\,\|\delta f\|_2^2,$$77

which are often sharper for the given process but are not sub-Gaussian in the classical sense (Wang et al., 2020). 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 (Chazottes et al., 2020, Chazottes et al., 16 Feb 2026, Chazottes et al., 7 Jul 2025, Chazottes et al., 2019). In others, it refers to Gaussian-type tails or anti-concentration for specialized functionals such as maxima, order statistics, or quadratic forms (Choudhary et al., 24 Jun 2026, Giessing, 2023, Kartsioukas et al., 2019, Gallagher et al., 2019). 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.