Hanson-Wright Inequality

Why This Matters

Scalar concentration inequalities (Hoeffding, Bernstein) control linear functions of independent random variables. sums of the form $\sum_i a_i X_i$. But many quantities in statistics and machine learning are quadratic: sample covariance entries $\frac{1}{n}\sum_i X_i X_j$ for $i \neq j$, kernel evaluations $k(x, x') = \phi(x)^\top \phi(x')$, chi-squared statistics, and second-order U-statistics. For these, you need concentration of the quadratic form $X^\top A X$ where $X$ is a random vector with independent entries.

The Hanson-Wright inequality is the definitive tool for this. It gives a two-term bound that captures two different regimes of deviation, and it is tight up to constants.

Mental Model

Consider the quadratic form $X^\top A X$ where $X \in \mathbb{R}^n$ has independent sub-Gaussian entries. This is a sum of terms $A_{ij} X_i X_j$ involving products of random variables, not just single variables. Products are harder to control because the tails are heavier (the product of two sub-Gaussians is sub-exponential, not sub-Gaussian).

The Hanson-Wright bound says: the deviation of $X^\top A X$ from its expectation is controlled by two terms:

1. Frobenius term $\|A\|_F$: dominates for small deviations, behaves like Gaussian concentration (sub-Gaussian tail $e^{-ct^2}$)
2. Operator term $\|A\|_{\text{op}}$: dominates for large deviations, behaves like sub-exponential concentration (tail $e^{-ct}$)

The transition between regimes happens at $t \approx \|A\|_F^2 / \|A\|_{\text{op}}$.

Formal Setup and Notation

Let $X = (X_1, \ldots, X_n) \in \mathbb{R}^n$ be a random vector with independent, centered, sub-Gaussian entries: $\mathbb{E}[X_i] = 0$ and $\|X_i\|_{\psi_2} \leq K$ for all $i$.

Let $A \in \mathbb{R}^{n \times n}$ be a fixed (deterministic) matrix.

Definition

Quadratic Form

The quadratic form associated with matrix $A$ and random vector $X$ is:

$X^\top A X = \sum_{i,j=1}^n A_{ij} X_i X_j$

Its expectation is $\mathbb{E}[X^\top A X] = \sum_i A_{ii} \mathbb{E}[X_i^2] = \text{tr}(A \cdot \text{diag}(\mathbb{E}[X_i^2]))$. For isotropic $X$ ($\mathbb{E}[X_i^2] = 1$), this simplifies to $\text{tr}(A)$.

Definition

Frobenius Norm

The Frobenius norm of $A$ is $\|A\|_F = \sqrt{\sum_{i,j} A_{ij}^2} = \sqrt{\text{tr}(A^\top A)}$. It measures the total "mass" of $A$ across all entries. In the Hanson-Wright bound, $\|A\|_F$ controls the variance of the quadratic form: the Gaussian-regime fluctuations scale like $\|A\|_F$.

Definition

Operator Norm

The operator norm is $\|A\|_{\text{op}} = \sup_{\|v\|=1} \|Av\|$. It measures the maximum directional stretch. In Hanson-Wright, $\|A\|_{\text{op}}$ controls the extreme-regime tail: large deviations are governed by the single largest singular value of $A$.

Core Relationship Between Norms

For any $n \times n$ matrix $A$:

$\|A\|_{\text{op}} \leq \|A\|_F \leq \sqrt{n} \|A\|_{\text{op}}$

The gap between the two norms measures how "spread out" the matrix is. If $A$ has rank 1, $\|A\|_F = \|A\|_{\text{op}}$ and the two terms in Hanson-Wright are comparable. If $A = I_n$ (identity), $\|A\|_F = \sqrt{n}$ while $\|A\|_{\text{op}} = 1$, and the Frobenius term dominates for all reasonable deviations.

Main Theorems

Theorem

Hanson-Wright Inequality

Statement

Let $X = (X_1, \ldots, X_n)$ have independent, centered, sub-Gaussian components with $\|X_i\|_{\psi_2} \leq K$. For any $n \times n$ matrix $A$ and any $t > 0$:

$\mathbb{P}\!\bigl(|X^\top A X - \mathbb{E}[X^\top A X]| \geq t\bigr) \leq 2\exp\!\Bigl(-c \min\!\Bigl(\frac{t^2}{K^4 \|A\|_F^2},\; \frac{t}{K^2 \|A\|_{\text{op}}}\Bigr)\Bigr)$

where $c > 0$ is a universal constant.

Intuition

The bound has two regimes:

Small deviations ($t \lesssim K^2 \|A\|_F^2 / \|A\|_{\text{op}}$): The $t^2 / \|A\|_F^2$ term is smaller, so the bound is $\exp(-ct^2/(K^4\|A\|_F^2))$. This is a sub-Gaussian tail. It arises because for small deviations, the quadratic form behaves like a sum of many weakly correlated terms, and the CLT-like behavior dominates.

Large deviations ($t \gtrsim K^2 \|A\|_F^2 / \|A\|_{\text{op}}$): The $t / \|A\|_{\text{op}}$ term is smaller, so the bound is $\exp(-ct/(K^2\|A\|_{\text{op}}))$. This is a sub-exponential tail. It arises because extreme deviations are driven by the largest eigenvalue direction of $A$, where the quadratic form is $\lambda_{\max}(A) \cdot X_v^2$ for a single sub-Gaussian variable $X_v = v^\top X$.

The crossover at $t^* = K^2 \|A\|_F^2 / \|A\|_{\text{op}}$ is where the two terms are equal. Below $t^*$, Gaussian behavior; above $t^*$, exponential behavior.

Proof Sketch

The proof proceeds in three steps:

Step 1 (Decoupling). Replace $X^\top A X$ with a decoupled form $\tilde{X}^\top A Y$ where $Y$ is an independent copy of $X$. The decoupling inequality states that for symmetric $A$:

$\mathbb{P}(|X^\top A X - \mathbb{E}[X^\top A X]| \geq t) \leq C \cdot \mathbb{P}(|X^\top A Y| \geq ct)$

This reduces the problem from a quadratic form (products $X_i X_j$) to a bilinear form (products $X_i Y_j$), which is easier to handle because $X$ and $Y$ are independent.

Step 2 (Condition and apply Hoeffding). Condition on $Y$. Then $X^\top A Y = \sum_i X_i (AY)_i$ is a sum of independent sub-Gaussian variables with variance proxy controlled by $\|AY\|^2$. Apply sub-Gaussian concentration to get a bound in terms of $\|AY\|$.

Step 3 (Control $\|AY\|$). Use the bound $\|AY\| \leq \|A\|_{\text{op}} \|Y\|$ for the operator-norm term and $\mathbb{E}[\|AY\|^2] = \|A\|_F^2$ (when $Y$ is isotropic) for the Frobenius term. Combine using a case split on whether $\|Y\|$ is typical or large.

Why It Matters

The Hanson-Wright inequality is essential whenever you work with second-order statistics of random vectors:

• Chi-squared concentration: $\|X\|^2 = X^\top I X$, so $A = I$ and the bound gives $\mathbb{P}(|\|X\|^2 - n| \geq t) \leq 2\exp(-c\min(t^2/n, t))$. The Frobenius regime ($t^2/n$) dominates for $t \leq n$; the operator regime ($t$) dominates for $t \geq n$.

• Random kernel evaluations: inner products $x^\top y$ for random $x, y$ are quadratic in the joint vector $(x, y)$.

• Covariance estimation: off-diagonal entries of $\hat{\Sigma}$ involve terms like $\frac{1}{n}\sum_k X_k^{(i)} X_k^{(j)}$, which are quadratic forms.

• Second-order chaos: any polynomial of degree 2 in sub-Gaussian variables.

Failure Mode

The constant $c$ in the inequality is universal but unspecified. for precise numerical bounds in applications, you may need to track it through the proof. Also, Hanson-Wright requires independent entries; for dependent sub-Gaussian vectors, you need modified versions (e.g., for vectors with sub-Gaussian norm bounds but dependent entries, the inequality may still hold but with $K$ replaced by the sub-Gaussian norm of the entire vector).

Lemma

Decoupling Inequality for Quadratic Forms

Statement

Let $X$ have independent centered entries and let $A$ be a symmetric matrix with $A_{ii} = 0$. Let $Y$ be an independent copy of $X$. Then for all convex, increasing functions $\Phi$ on $[0, \infty)$:

$\mathbb{E}\!\Bigl[\Phi\!\bigl(|X^\top A X|\bigr)\Bigr] \leq \mathbb{E}\!\Bigl[\Phi\!\bigl(C \cdot |X^\top A Y|\bigr)\Bigr]$

where $C$ is a universal constant.

Intuition

Decoupling replaces the "entangled" quadratic form $\sum_{i \neq j} A_{ij} X_i X_j$ (where each $X_i$ appears in multiple terms) with the "decoupled" bilinear form $\sum_{i,j} A_{ij} X_i Y_j$ (where each $X_i$ and $Y_j$ appear in separate roles). The decoupled version is easier to analyze because once you condition on $Y$, the sum is linear in $X$, and all the standard sub-Gaussian tools apply.

Proof Sketch

The proof uses a symmetrization-style argument. Introduce the decoupled form and use the independence of $X$ and $Y$ together with the symmetry of $A$ to show that the tails of the coupled form are controlled by those of the decoupled form. The universal constant $C$ absorbs a factor from the randomization step.

Why It Matters

Decoupling is the key technical device that makes Hanson-Wright provable. It converts a hard problem (concentrating a quadratic form) into an easier one (concentrating a bilinear form, which is linear once you condition on one factor). This technique appears throughout the theory of U-statistics and chaos processes.

Failure Mode

Decoupling requires the diagonal of $A$ to be zero (or handled separately). The diagonal terms $A_{ii} X_i^2$ are not quadratic in the same sense. they are sums of independent sub-exponential variables and are controlled separately using standard sub-exponential concentration.

Two Regimes Explained

The Hanson-Wright bound can be rewritten in high-probability form. With probability at least $1 - \delta$:

$|X^\top A X - \mathbb{E}[X^\top A X]| \lesssim K^2\!\left(\|A\|_F \sqrt{\log(2/\delta)} + \|A\|_{\text{op}} \log(2/\delta)\right)$

The two terms correspond to the two norms:

Regime	Tail behavior	Dominated by	Example
Gaussian ($t$ small)	$e^{-ct^2}$	$\|A\|_F$	Chi-squared with $t \ll n$
Extreme ($t$ large)	$e^{-ct}$	$\|A\|_{\text{op}}$	Chi-squared with $t \gg n$

Canonical Examples

Example

Chi-squared concentration

Let $X \sim \mathcal{N}(0, I_n)$ and $A = I_n$. Then $X^\top A X = \|X\|^2 \sim \chi^2_n$ with $\mathbb{E}[\|X\|^2] = n$. The norms are $\|I\|_F = \sqrt{n}$ and $\|I\|_{\text{op}} = 1$. Hanson-Wright gives:

$\mathbb{P}(|\|X\|^2 - n| \geq t) \leq 2\exp\!\bigl(-c\min(t^2/n,\; t)\bigr)$

For $t = \epsilon n$ (relative deviation $\epsilon$): the bound is $\exp(-c\epsilon^2 n)$ when $\epsilon \leq 1$ (Gaussian regime) and $\exp(-c\epsilon n)$ when $\epsilon \geq 1$ (extreme regime). This matches the known chi-squared tail behavior.

Example

Random kernel evaluation

Let $x, y \in \mathbb{R}^d$ be independent vectors with i.i.d. sub-Gaussian entries of norm $K$. The inner product $x^\top y = Z^\top A Z$ where $Z = (x, y) \in \mathbb{R}^{2d}$ and $A$ is a $2d \times 2d$ block matrix with off-diagonal blocks $I_d/2$ and zero diagonal blocks.

Then $\|A\|_F = \sqrt{d}/\sqrt{2}$ and $\|A\|_{\text{op}} = 1/2$. Hanson-Wright gives:

$\mathbb{P}(|x^\top y| \geq t) \leq 2\exp\!\bigl(-c\min(t^2/d,\; t)\bigr)$

So random inner products in $\mathbb{R}^d$ concentrate around 0 with fluctuations of order $\sqrt{d}$.

Example

Quadratic form with rank-r matrix

If $A$ has rank $r$ with eigenvalues $\lambda_1 \geq \cdots \geq \lambda_r > 0$, then $\|A\|_F = \sqrt{\sum_i \lambda_i^2}$ and $\|A\|_{\text{op}} = \lambda_1$.

The crossover point is $t^* \approx \sum_i \lambda_i^2 / \lambda_1$. If $A$ is rank-1 ($A = \lambda vv^\top$), then $\|A\|_F = |\lambda| = \|A\|_{\text{op}}$ and the two regimes are identical: $e^{-ct^2/\lambda^2}$ transitions directly to $e^{-ct/|\lambda|}$ at $t \approx |\lambda|$.

Common Confusions

Watch Out

Hanson-Wright is not just Hoeffding applied to products

A naive approach would be to note that each term $A_{ij} X_i X_j$ is sub-exponential (product of two sub-Gaussians) and apply a sub-exponential concentration bound. This gives a bound in terms of $\sum_{ij} A_{ij}^2 = \|A\|_F^2$, which captures the Frobenius regime but misses the tighter operator-norm regime for large deviations. Hanson-Wright is strictly stronger because it also captures the $\|A\|_{\text{op}}$ term through the decoupling argument.

Watch Out

The matrix A need not be symmetric or positive semidefinite

The Hanson-Wright inequality applies to any matrix $A$, not just symmetric or PSD ones. For a general $A$, $X^\top A X = X^\top ((A + A^\top)/2) X$ because $X^\top B X = 0$ for any antisymmetric $B$ (since $X^\top B X$ is a scalar that equals its own negative). So you can always reduce to the symmetric part.

Watch Out

Sub-Gaussian entries, not sub-Gaussian vectors

Hanson-Wright requires the entries of $X$ to be independent sub-Gaussian, not just the vector $X$ to have sub-Gaussian norm. For random vectors with dependent entries (like uniform on the sphere), the standard Hanson-Wright does not apply directly. Modified versions exist but require different techniques (e.g., transportation-cost arguments).

Summary

• Hanson-Wright controls $|X^\top A X - \mathbb{E}[X^\top A X]|$ for sub-Gaussian $X$
• Two-term bound: $\exp(-c\min(t^2/\|A\|_F^2, t/\|A\|_{\text{op}}))$
• Frobenius norm $\|A\|_F$ controls the Gaussian (small deviation) regime
• Operator norm $\|A\|_{\text{op}}$ controls the extreme (large deviation) regime
• Crossover at $t^* = \|A\|_F^2 / \|A\|_{\text{op}}$
• Decoupling is the key proof technique: replace $X^\top A X$ with $X^\top A Y$ using an independent copy $Y$
• For $A = I$: recovers chi-squared concentration
• Applies to random kernel evaluations, covariance estimation, second-order chaos

Exercises

ExerciseCore

Problem

Let $X \in \mathbb{R}^n$ have i.i.d. $\mathcal{N}(0, 1)$ entries and let $A = \frac{1}{n}(11^\top - I)$ where $1$ is the all-ones vector. The quadratic form $X^\top A X = \frac{1}{n}(\sum_i X_i)^2 - \frac{1}{n}\sum_i X_i^2$. Compute $\|A\|_F$ and $\|A\|_{\text{op}}$, and use Hanson-Wright to bound the deviation of $X^\top A X$ from its expectation.

Hint

Reveal Solution

ExerciseAdvanced

Problem

Let $X \in \mathbb{R}^n$ have i.i.d. sub-Gaussian entries with parameter $K$ and let $P$ be the orthogonal projection onto a $k$-dimensional subspace. Use Hanson-Wright to show that $\|PX\|^2 = X^\top P X$ concentrates around $k$ with sub-Gaussian fluctuations of order $\sqrt{k}$.

Hint

Reveal Solution

ExerciseResearch

Problem

The Hanson-Wright inequality gives a bound of order $\|A\|_F\sqrt{\log(1/\delta)}$ in the Gaussian regime. Show that this is tight (up to constants) by computing the variance of $X^\top A X$ when $X \sim \mathcal{N}(0, I)$ and verifying that $\text{Var}(X^\top A X) = 2\|A\|_F^2$ for symmetric $A$.

Hint

Reveal Solution

References

Canonical:

• Hanson & Wright, "A Bound on Tail Probabilities for Quadratic Forms in Independent Random Variables" (1971)
• Rudelson & Vershynin, "Hanson-Wright Inequality and Sub-Gaussian Concentration" (2013)
• Vershynin, High-Dimensional Probability (2018), Chapter 6

Current:

• Wainwright, High-Dimensional Statistics (2019), Chapter 6

• Adamczak, "A Note on the Hanson-Wright Inequality for Random Vectors with Dependencies" (2015)

• Boucheron, Lugosi, Massart, Concentration Inequalities (2013), Chapters 2-6

Next Topics

Building on quadratic form concentration:

• Random matrix theory overview: asymptotic behavior of eigenvalues and eigenvectors of large random matrices
• Kernels and RKHS: random kernel evaluations use Hanson-Wright for concentration of kernel matrices