Gaussian Bound for Rademacher Sums

• Gaussian Bound is an explicit upper estimate for the tail probabilities of normalized Rademacher sums, defined by adding a rational correction term to the standard normal tail.
• The bound is asymptotically equivalent to the standard normal tail, ensuring sharpness and optimality by matching the precise 1/x decay in the far tail.
• Its derivation employs advanced techniques including exponential tilting, Berry–Esseen bounds, and computer-assisted verification, with applications extending to bounded independent and self-normalized statistics.

A Gaussian bound, in the setting of normalized Rademacher sums, is an explicit upper estimate for tail probabilities that is genuinely modeled on the standard normal tail rather than on a purely exponential envelope. For

$$S_n:=a_1\varepsilon_1+\cdots+a_n\varepsilon_n,\qquad \varepsilon_i\in\{-1,1\}\ \text{i.i.d.},\qquad a_1^2+\cdots+a_n^2=1,$$

with $$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$, the bound obtained in 2010 gives a uniformly valid inequality of the form “normal tail + explicit correction term.” It is asymptotically equivalent to the standard normal upper tail, confirms a longstanding conjectural picture due to Efron, and is sharp in a best-possible sense. The same framework also yields consequences for centered uniformly bounded independent random variables and for self-normalized statistics relevant to the Student test (Pinelis, 2010).

1. Explicit formulation for normalized Rademacher sums

Let $Z$ denote a standard normal random variable, let $\Phi$ be its distribution function, and let $\varphi$ be the standard normal density. The central result states that for every $x>0$,

$$\mathsf P(S_n>x)\le Q(x):=\mathsf P(Z>x)+\frac{C\,\varphi(x)}{9+x^2} <\mathsf P(Z>x)\Bigl(1+\frac{C}{x}\Bigr),$$

where

$C:=5\sqrt{2\pi e}\,(1-\Phi(1))=14.10\dots.$

The quantity $Q(x)$ is therefore an explicit Gaussian-type upper bound: it is not merely of the same exponential order as the Gaussian tail, but is literally the Gaussian tail augmented by a rational correction factor multiplying $\varphi(x)$ (Pinelis, 2010).

The normalization $$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$0 is essential. It places the Rademacher sum on the same variance scale as the standard normal law and makes a direct comparison to $$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$1 meaningful. The theorem is uniform over all choices of coefficients $$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$2 and all dimensions $$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$3.

2. Asymptotic Gaussian character

The defining feature of the bound is its asymptotic equivalence to the normal tail: $$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$4 or equivalently,

$$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$5

This uses the classical asymptotic relation

$$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$6

Hence the correction term has exactly the right scale to preserve Gaussian-tail asymptotics (Pinelis, 2010).

This distinguishes the result from standard exponential inequalities such as

$$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$7

Such bounds are sub-Gaussian but not asymptotically Gaussian in the relevant sense, because their ratio to $$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$8 grows like a constant times $$\mathsf P(\varepsilon_i=1)=\mathsf P(\varepsilon_i=-1)=\tfrac12$$9. The missing $Z$0 factor is decisive in the far tail. A common misconception is to treat all sub-Gaussian estimates as comparable in precision; in this setting, the theorem shows that matching the Gaussian exponential rate alone is insufficient when one wants a tail bound with the correct normal shape.

3. Sharpness and optimality

The result is sharp in two different senses. First, it is asymptotically sharp as $Z$1, since the bound-to-tail ratio tends to $Z$2. Second, the constant $Z$3 in the additive correction term is best possible for the stated inequality: equality in the first bound occurs at $Z$4 and $Z$5. Within the exact theorem stated above, the constant therefore cannot be improved (Pinelis, 2010).

The paper also situates the theorem relative to earlier multiplicative Gaussian bounds of the form

$Z$6

for which the best possible constant is reported as $Z$7. That multiplicative form is sharp as a uniform comparison, but it does not reproduce the Gaussian tail asymptotically with ratio $Z$8. The newer bound improves the shape of the estimate for large $Z$9: it no longer controls the Rademacher tail merely up to a fixed multiplicative constant, but tracks the normal tail itself.

Historically, this sharp asymptotic behavior is the aspect connected to Efron’s conjectural picture. The theorem affirms that one can bound normalized Rademacher tails by something genuinely close to the Gaussian tail, not just by an exponential surrogate.

4. Structure of the proof

The proof combines exponential tilting, normal approximation, extremal reduction, and computer-assisted verification. Its first step is an Esscher tilt transform at parameter $\Phi$0, chosen to match the Gaussian optimizer. This rewrites the tail probability in a form better suited to comparison with $\Phi$1. After tilting, a Berry–Esseen bound with explicit constant is used to separate the expression into a normal-like main term and an error term (Pinelis, 2010).

A further reduction introduces the measure

$\Phi$2

which converts the original optimization over coefficient vectors $\Phi$3 into lower-dimensional extremal problems involving integrals against $\Phi$4. The analysis then exploits Tchebycheff/Markov systems and a Carathéodory reduction to decrease the effective dimension of the extremal problem.

Several of the remaining inequalities are highly technical and are verified by symbolic computation in Mathematica, including positivity of certain Wronskians and auxiliary expressions. The argument is therefore both analytic and computer-assisted. A notable feature of the proof is that the asymptotic normal comparison does not arise from a loose large-deviation estimate, but from a carefully organized reduction whose endpoint is explicit and globally uniform in the coefficients.

5. Extensions and applications

One extension concerns sums of general centered uniformly bounded independent random variables. If $\Phi$5 are i.i.d. and satisfy

$\Phi$6

then, via a result of Bentkus, the paper derives

$\Phi$7

where $\Phi$8 is a linear interpolation of $\Phi$9 on a suitable grid (Pinelis, 2010).

A second application concerns self-normalized sums. For

$\varphi$0

the same bound applies under orthant symmetry, in particular for independent symmetric $\varphi$1. Since Student’s $\varphi$2-statistic is a monotone function of $\varphi$3, this yields tail bounds relevant to the Student test.

These applications show that the theorem is not confined to the exact Rademacher model. Its role is partly foundational: Rademacher tails serve as an extremal template from which more general bounded or self-normalized settings can inherit Gaussian-type tail control.

6. Interpretation and place within probability inequalities

The theorem occupies a specific niche among probability inequalities. It is stronger than a conventional sub-Gaussian estimate because it captures the correct tail shape; it is more explicit than a purely asymptotic statement because the correction term is uniform and closed form; and it is more refined than a constant-factor Gaussian comparison because the multiplicative error vanishes at infinity (Pinelis, 2010).

Its statistical significance lies in the fact that many procedures depend not only on exponential decay but on accurate upper-tail calibration. In that respect, the bound provides a bridge between exact combinatorial distributions and Gaussian asymptotics. The Rademacher sum is discrete and highly non-Gaussian at finite $\varphi$4, yet its upper tail is controlled by an explicit function that is simultaneously uniform, asymptotically normal, and sharp.

The broader conceptual lesson is that Gaussian behavior in tail bounds should not be identified merely with the presence of the factor $\varphi$5. In this result, “Gaussian bound” means a comparison with the normal tail at the correct scale, including the crucial $\varphi$6 asymptotic structure. That is the sense in which the theorem confirms Efron’s conjectural Gaussian picture for Rademacher tails and gives the term its precise meaning in this context.