Calculate Sub Gaussian Norm Of A Bounded Variable

Use support bounds and an optional known mean to estimate a rigorous sub gaussian parameter for a bounded random variable. This calculator applies the classic Hoeffding range bound and also reports a practical upper bound for the psi-2 norm based on bounded support.

Calculator Inputs

Results

For a random variable X supported on [a, b], Hoeffding’s lemma implies that X – E[X] is sub gaussian with proxy parameter sigma = (b – a) / 2. If you also know the mean mu, then |X – mu| is bounded by max(mu – a, b – mu), which gives a practical psi-2 upper bound of max(mu – a, b – mu) / sqrt(ln 2).

Expert Guide: How to Calculate the Sub Gaussian Norm of a Bounded Variable

When a random variable is bounded, it is automatically well behaved in the tail compared with many unbounded distributions. That fact is one of the most useful tools in probability, statistics, machine learning, empirical process theory, and risk analysis. If you know a variable can never fall below a and can never exceed b, then you can derive a clean concentration guarantee for its centered form. This is where the idea of the sub gaussian norm becomes practical. In applications, people often want a single number that summarizes how sharply a variable concentrates around its mean. For bounded variables, that number can be estimated very reliably from the support alone.

The simplest and most common result is based on Hoeffding’s lemma. If X lies in [a, b] almost surely, then the centered variable X – E[X] satisfies a gaussian-like moment generating function bound. In standard proxy notation, this means it is sub gaussian with parameter

Hoeffding proxy for X – E[X]:
sigma = (b – a) / 2

Equivalent mgf statement:
E exp(lambda (X – E[X])) <= exp(lambda squared times (b – a) squared divided by 8)

This formula is powerful because it does not require a full distribution. You do not need normality, symmetry, or independence to compute the parameter for a single variable. You only need the support width. That is why the range-based calculation appears so often in concentration inequalities for bounded observations, randomized algorithms, online learning, and finite sample confidence analysis.

What does “sub gaussian norm” mean here?

There are two common conventions in the literature. The first is the mgf proxy parameter, often written as sigma, which appears in bounds of the form

E exp(lambda Y) <= exp(lambda squared times sigma squared divided by 2)

for a centered variable Y. Under this convention, a bounded variable on [a, b] gives the clean proxy sigma = (b – a) / 2.

The second convention is the Orlicz psi-2 norm, often written as ||Y||_psi2. One standard definition is the smallest K such that

E exp(Y squared divided by K squared) <= 2

If Y is absolutely bounded by M, then a direct bound is ||Y||_psi2 <= M / sqrt(ln 2). For a bounded variable X with known mean mu and support [a, b], you can use

M = max(mu – a, b – mu)

and therefore obtain

||X – mu||_psi2 <= max(mu – a, b – mu) / sqrt(ln 2)

This second number is often more conservative than the Hoeffding proxy because it is based on an absolute boundedness argument. Still, it is rigorous, simple, and useful when you want a direct psi-2 estimate from deterministic support information.

Step by step method

1. Identify the support. Find real numbers a and b so that X is always between them.
2. Center the variable. For sub gaussian work, the object of interest is usually X – E[X]. If the mean is unknown, the Hoeffding proxy still works from the range alone.
3. Compute the support width. Calculate b – a.
4. Compute the Hoeffding proxy. Set sigma = (b – a) / 2.
5. If the mean is known, tighten the absolute centered bound. Compute M = max(mu – a, b – mu).
6. Compute a bounded-support psi-2 upper bound. Set K = M / sqrt(ln 2).
7. Use the tail inequality if needed. For t greater than or equal to 0, a two-sided Hoeffding bound is P(|X – E[X]| >= t) <= 2 exp(-2 t squared / (b – a) squared).

Why bounded variables are automatically sub gaussian

Boundedness prevents extreme outliers. A gaussian tail decays like exp(-c t squared), and bounded variables decay even faster in a trivial sense because the tail becomes exactly zero once t exceeds the largest possible deviation. The sub gaussian framework does not require the variable to be literally normal. It only requires a normal-like tail envelope or a normal-like mgf inequality. Bounded support is more than enough to imply that behavior.

In practice, this matters because many real measurements are naturally bounded: test scores from 0 to 100, click through rates from 0 to 1, pass fail indicators, normalized sensor outputs, proportions, finite reward signals in reinforcement learning, and trimmed risk metrics. In each case, the support interval gives an immediate concentration control even before you estimate variance from data.

Comparison table: common bounded variables

Variable	Support	Mean	Variance	Hoeffding proxy sigma	Psi-2 upper bound using support
Bernoulli(p) with p = 0.5	[0, 1]	0.5	0.25	0.5	0.5 / sqrt(ln 2) = 0.6006
Bernoulli(p) with p = 0.2	[0, 1]	0.2	0.16	0.5	0.8 / sqrt(ln 2) = 0.9609
Uniform(0, 1)	[0, 1]	0.5	1/12 = 0.0833	0.5	0.5 / sqrt(ln 2) = 0.6006
Rademacher (+1 or -1 with equal chance)	[-1, 1]	0	1	1	1 / sqrt(ln 2) = 1.2011
Uniform(-2, 3)	[-2, 3]	0.5	25/12 = 2.0833	2.5	2.5 / sqrt(ln 2) = 3.0028

Notice the pattern. The Hoeffding proxy depends only on the interval width. That is why Bernoulli(0.2), Bernoulli(0.5), and Uniform(0,1) all share the same proxy sigma = 0.5 despite having different variances. The psi-2 upper bound can improve if the mean sits close to the center of the interval and can become looser when the mean is near an edge.

Comparison table: tail bounds implied by support width

The next table uses the two-sided Hoeffding tail formula for variables on [0,1], where sigma = 0.5. The values below are numerical upper bounds, not exact probabilities. They are useful as conservative concentration guarantees.

Deviation t	Hoeffding upper bound 2 exp(-2 t squared)	Interpretation
0.10	1.9604	Trivial because the formula is capped by 2 for very small deviations.
0.25	1.7650	Still loose, but valid as a worst-case bound across all distributions on [0,1].
0.50	1.2131	At half-range deviation, the concentration starts becoming informative.
0.75	0.6493	For larger deviations, the gaussian-style decay becomes clearer.
1.00	0.2707	The event is already impossible for many centered variables, but the range-only guarantee stays conservative.

How to interpret the result from the calculator

• Hoeffding proxy sigma is the standard range-based sub gaussian parameter for the centered variable.
• Psi-2 upper bound is a deterministic norm bound derived from the largest possible centered magnitude.
• Tail bound at t gives a worst-case upper envelope for the probability of a deviation at least as large as t.
• Centered max deviation M tells you the largest absolute departure from the chosen center or known mean.

If you are doing theoretical work, the Hoeffding proxy is typically the most natural quantity to report when concentration inequalities are written in mgf or tail form. If you are comparing norms in high-dimensional probability or empirical process theory, the psi-2 quantity is often the right language. In many settings, reporting both is best because the two conventions appear side by side in modern texts.

Common mistakes to avoid

• Using X instead of X – E[X]. Sub gaussian statements are usually about a centered variable.
• Confusing variance with sub gaussian norm. A variable can have small variance but still inherit a larger worst-case sub gaussian proxy from its support width.
• Assuming exactness. The range-based value is often a valid upper bound, not necessarily the smallest possible norm.
• Ignoring a known mean. If you know mu, the bounded-support psi-2 estimate can be tightened noticeably.
• Using the tail formula without understanding conservatism. Hoeffding bounds are designed to be distribution-free, so they can be much looser than the true tail for a particular distribution.

Where these formulas come from

The range-based proxy comes from Hoeffding’s lemma, one of the foundational facts in concentration theory. The result shows that any centered bounded variable has a moment generating function controlled by the square of the support width. This directly produces exponential tail bounds and implies sub gaussian behavior. A separate but related fact is that if a variable is bounded in absolute value by M, then the psi-2 norm is bounded by M divided by sqrt(ln 2), because the exponential square moment is then uniformly controlled by the support ceiling.

For deeper reading, see these authoritative academic and government resources:

• University of California, Berkeley notes on tail bounds and concentration
• MIT OpenCourseWare probability theory materials
• NIST Engineering Statistics Handbook

Practical takeaway

If all you know is that a random variable lies in the interval [a, b], you already know enough to produce a robust sub gaussian guarantee. The fastest formula is

sigma = (b – a) / 2

for the centered variable X – E[X]. If you also know the mean mu, then you can compute the tighter deterministic bound

M = max(mu – a, b – mu), and ||X – mu||_psi2 <= M / sqrt(ln 2)

These quantities are especially useful when you are building confidence intervals, proving uniform convergence, setting regularization levels, analyzing bounded rewards, or comparing distribution-free risk guarantees. The calculator above turns those formulas into an immediate numerical result and visualizes the associated tail envelope so that you can move from theory to application quickly and cleanly.

This calculator provides rigorous upper bounds under standard sub gaussian conventions for bounded variables. It is intended for educational and analytical use and does not replace a full distribution-specific derivation when an exact norm is required.