Calculate Subgaussian Norm Of A Bounded Variable

Use this premium calculator to estimate the subgaussian size of a bounded random variable using Hoeffding-style bounds. Enter lower and upper support limits, choose whether the variable is centered, and compare the support width, radius from the mean, and variance proxy in one place.

Subgaussian Norm Calculator

Results

How to calculate the subgaussian norm of a bounded variable

The phrase calculate subgaussian norm of a bounded variable usually refers to finding a practical upper bound on how concentrated a random variable is around its mean. In probability, statistics, machine learning, compressed sensing, random matrix theory, and empirical process theory, subgaussian behavior is important because it gives Gaussian-like tail decay without requiring the variable itself to be normally distributed. A bounded random variable is one of the easiest classes of variables to control: if a variable lives in an interval, then its centered version automatically satisfies a subgaussian inequality. That fact is the core idea behind this calculator.

Suppose a random variable X is almost surely bounded in an interval [a, b]. Then the centered variable X – E[X] is subgaussian. A standard result, often called Hoeffding’s lemma, states that for every real number λ,

E[exp(λ(X – E[X]))] ≤ exp(λ²(b – a)² / 8).

This inequality says that the centered bounded variable behaves like a Gaussian with variance proxy (b – a)² / 4. Equivalently, in the common mgf form E[exp(λY)] ≤ exp(λ²σ² / 2), where Y = X – E[X], one may take σ = (b – a) / 2. That quantity is what many practitioners informally call a subgaussian parameter or a convenient subgaussian norm proxy for bounded variables.

What this calculator computes

This page gives you two closely related measurements:

• Centered Hoeffding proxy: σ = (b – a) / 2. This is the standard, robust estimate for X – E[X] when only the support interval is known.
• Support radius around the mean: K = max(|a – μ|, |b – μ|). If you know the mean μ and want the maximum distance from the mean to the support edges, this radius is a useful geometric quantity.

In many applications, the first quantity is the right answer because concentration inequalities are stated for centered variables. The second quantity is still useful, especially when the support is asymmetric around the mean or when you want a direct bound on |X – μ|.

Why bounded variables are automatically subgaussian

A Gaussian random variable has tails that decay like exp(-ct²). A subgaussian random variable has tail behavior of the same general type. If X is bounded, then large deviations are literally impossible beyond the support, so it is no surprise that boundedness implies subgaussianity. The nontrivial part is the exact scale. Hoeffding’s lemma shows that the scale is driven by the interval width b – a. Wider support means weaker concentration; narrower support means stronger concentration.

This is especially important in data science and statistical learning. If a feature, loss, reward, or response variable is clipped, bounded, or naturally restricted to a finite interval, then concentration arguments often become immediate. That is why analysts regularly estimate a subgaussian constant from support bounds even when the full distribution is unknown.

Step-by-step formula

1. Identify the support interval: X ∈ [a, b].
2. Center the variable if needed: let Y = X – E[X].
3. Apply Hoeffding’s lemma: E[exp(λY)] ≤ exp(λ²(b – a)² / 8).
4. Match this to the standard subgaussian mgf form exp(λ²σ² / 2).
5. Conclude that a valid subgaussian proxy is σ = (b – a) / 2.

If you are using another formal definition of the subgaussian norm, such as the Orlicz ψ2 norm, there can be absolute constant-factor differences between conventions. That is normal. The support-width rule remains the main practical estimate, and the constants depend on the exact normalization adopted in a textbook or paper.

Worked examples

Example 1: Bernoulli random variable

Let X be Bernoulli with success probability p. Then X takes values in [0, 1], so the centered variable X – p is subgaussian with Hoeffding proxy σ = (1 – 0)/2 = 0.5. This bound is distribution-free across all values of p, even though the exact concentration can be tighter for some p.

Example 2: Dice roll scaled to probability range

Suppose a variable is known to lie between 2 and 5. Then b – a = 3, so the centered subgaussian proxy is σ = 1.5. The corresponding variance proxy is σ² = 2.25. If the mean is μ = 3, then the radius around the mean is max(|2 – 3|, |5 – 3|) = 2.

Example 3: Survey score on a 1 to 7 scale

A bounded survey item lives in [1, 7]. Then the centered score minus its expectation is subgaussian with proxy σ = 3. This is useful in sample mean concentration, confidence calculations, and finite-sample learning theory when the exact score distribution is unknown but the scale is known.

Comparison table: common bounded variables and their subgaussian proxies

Variable type	Support [a, b]	Width b – a	Hoeffding proxy σ = (b – a)/2	Variance proxy σ²
Bernoulli(p)	[0, 1]	1	0.5	0.25
Rademacher variable	[-1, 1]	2	1	1
Uniform score on 1 to 5 scale	[1, 5]	4	2	4
Standardized bounded feature	[-0.5, 0.5]	1	0.5	0.25
Proportion or rate	[0, 1]	1	0.5	0.25

Why the variance proxy matters

The subgaussian parameter is often used to control sample means. If X1, …, Xn are independent copies of a centered subgaussian variable with proxy σ, then the sample average typically concentrates at scale roughly σ / √n. For bounded data, this means support width directly influences the speed of concentration. A narrower range yields tighter finite-sample guarantees.

For example, if a variable lies in [0, 1], then Hoeffding’s inequality implies P(|X̄ – E[X]| ≥ t) ≤ 2 exp(-2nt²). This is one of the most used concentration inequalities in applied statistics. It powers confidence bounds in online learning, A/B testing, quality control, and survey sampling.

Comparison table: support width versus concentration scale

Support width b – a	Subgaussian proxy σ	Variance proxy σ²	Approximate sample-mean scale σ / √100
0.5	0.25	0.0625	0.025
1	0.5	0.25	0.05
2	1	1	0.10
4	2	4	0.20
10	5	25	0.50

When to use the Hoeffding proxy and when to be more precise

The Hoeffding proxy is excellent when you know only that the variable is bounded. It is conservative, simple, and universal. However, it may not be tight. If you know more about the distribution, you can often sharpen the estimate:

• If the variable is symmetric and tightly concentrated near the mean, the true subgaussian norm may be smaller.
• If the distribution is exactly Bernoulli, one can derive sharper constants depending on p.
• If the variable is not centered, concentration statements usually apply to X – E[X], not to X itself.
• If the variable is merely bounded with heavy mass near one endpoint, the support-width method still works even though it does not exploit distribution shape.

Common mistakes when calculating a subgaussian norm

1. Using the raw interval width as the norm. The Hoeffding mgf parameter is based on (b – a)/2, not on b – a.
2. Forgetting to center the variable. Most subgaussian inequalities are written for X – E[X].
3. Mixing norm conventions. The ψ2 norm, tail-parameter form, and mgf parameter form may differ by absolute constants.
4. Assuming boundedness implies exact Gaussian tails. Boundedness implies subgaussian behavior, but the constants may be conservative.
5. Ignoring asymmetry around the mean. If μ is not the midpoint of [a, b], the radius around the mean can differ from half-width.

Applications in statistics, learning, and data analysis

Bounded-variable subgaussian calculations show up everywhere. In A/B testing, binary conversion outcomes lie in [0, 1], so concentration of sample proportions follows immediately. In reinforcement learning, clipped rewards are bounded, making confidence intervals and exploration bonuses easier to control. In survey design, ordinal scales like 1 to 5 or 1 to 7 produce bounded observations that satisfy finite-sample concentration estimates. In optimization and stochastic gradient methods, clipped gradients or losses are bounded, which can simplify nonasymptotic analysis.

This is why the support interval is often one of the first pieces of information a probabilist or data scientist looks for. Once the interval is known, a valid concentration scale is available even before fitting a detailed model.

Authoritative references

For deeper theory, consult these high-quality resources:

• University of California, Berkeley: Hoeffding inequality notes
• Carnegie Mellon University: Subgaussian random variables lecture notes
• NIST Engineering Statistics Handbook

Bottom line

If you want to calculate the subgaussian norm of a bounded variable quickly and safely, start with the support interval. For a variable bounded in [a, b], the centered variable X – E[X] has a standard Hoeffding-style subgaussian proxy σ = (b – a)/2. This page automates that calculation and also reports the support radius around a supplied mean. For most practical concentration arguments, that support-width method is the right first answer.

Interpretation note: Different books define the subgaussian norm with slightly different constants. The calculator reports the standard bounded-variable mgf proxy derived from Hoeffding’s lemma and a geometric radius around the mean, both of which are highly useful in applications.