Find your next learning step
Normal approximation becomes more concrete when the comparison is written as an integration-by-parts identity. A Stein kernel is the random coefficient that makes this identity exact for a target distribution. For a Gaussian it is constant; for other laws it varies with the observation. Conditional variance from a reversible coupling can estimate the same coefficient, while a Poincare inequality controls fluctuation without magically proving Gaussianity.
Open the chapter contents · Probability root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the decision, or read from Chapter 1. The chapters remain open; these grouped routes are a navigable index.
Route 1 · Define the decision
Route 2 · Work the ledgers
Route 3 · Transfer the method
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE THE DECISION
1. Start from the integration-by-parts identity
Let X be centred with variance σ². A Stein kernel τ satisfies E[Xf(X)]=E[τ(X)f′(X)] for a suitable class of smooth test functions. The identity moves the random multiplier X onto a derivative.
For a centred Gaussian, τ(X)=σ² almost surely. Normal approximation therefore asks whether a non-Gaussian kernel behaves nearly like a constant at the scale and for the test functions used.
CHAPTER 2 OF 12 · DEFINE THE DECISION
2. Separate existence, choice and uniqueness
A density on an interval often admits a canonical one-dimensional kernel obtained by integrating its centred density. Discrete, multivariate or constrained laws require a different operator and may admit more than one kernel.
Writing down τ is not enough. State the measure, support, boundary conditions and test-function class that justify the integration by parts.
CHAPTER 3 OF 12 · WORK THE LEDGERS
3. Derive a kernel for a uniform law
Let X be uniform on [−a,a]. Integration by parts shows that τ(x)=(a²−x²)/2: the boundary factor vanishes at ±a, and differentiating τ times the constant density produces −x times that density.
| x | τ(x) when a=√3 | Reading |
|---|---|---|
| 0 | 3/2 | Largest central coefficient |
| ±1 | 1 | Equals the variance |
| ±√3 | 0 | Boundary coefficient vanishes |
Because Var(X)=a²/3, choosing a=√3 standardises the variance to one. The kernel is exact but plainly nonconstant, so the law is not Gaussian.
CHAPTER 4 OF 12 · WORK THE LEDGERS
4. Check the mean of the kernel
Use f(x)=x in the Stein identity. Then E[X²]=E[τ(X)], so every valid kernel in this identity has mean σ² under the stated conditions.
| Quantity | Variance-one uniform | Gaussian N(0,1) |
|---|---|---|
| E[τ(X)] | 1 | 1 |
| Var(τ(X)) | 1/5 | 0 |
| Kernel range | 0 to 3/2 | Constant 1 |
For the uniform example, E[X⁴]=9/5, so Var(τ)=Var((3−X²)/2)=Var(X²)/4=(9/5−1)/4=1/5.
CHAPTER 5 OF 12 · WORK THE LEDGERS
5. Connect kernel fluctuation to normal approximation
The normal Stein equation converts a distributional distance into expectations involving 1−τ(X) after standardisation. Bounds depend on the chosen probability metric and on the derivative bounds of its test functions.
A small E|1−τ(X)| or root-mean-square fluctuation can support a normal-approximation bound. It is a quantified comparison, not a statement that the variables are literally Gaussian.
CHAPTER 6 OF 12 · TRANSFER THE METHOD
6. Recover an approximate kernel from conditional variance
For an exchangeable pair (W,W′), put D=W′−W and suppose E[D|W]=−λW. A Taylor expansion makes E[D²|W]/(2λ) play the role of a conditional Stein coefficient.
The useful error is the fluctuation of that conditional second moment around Var(W), together with regression remainders and large-step terms. The unconditional equality E[D²]=2λVar(W) is only the first ledger check.
CHAPTER 7 OF 12 · TRANSFER THE METHOD
7. Read a Poincare inequality correctly
A Poincare inequality has the form Var(g(X))≤C E[g′(X)²] for suitable g. It controls how functions fluctuate under the law and supplies a spectral-gap constant C.
It does not by itself say τ is constant. For the variance-one uniform law, the sharp interval constant is 12/π², which is greater than one; taking g(x)=x confirms only that C must be at least Var(X)=1.
CHAPTER 8 OF 12 · TRANSFER THE METHOD
8. Connect the nearby coupling routes
Use exchangeable pairs and conditional regression to construct the conditional-variance coefficient. Use zero-bias coupling for a closely related derivative identity.
The guide to coupling bounds and probability metrics keeps the final claim tied to an explicit distance rather than the vague phrase “approximately normal”.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Know the limits and common misconceptions
A kernel with the correct mean does not identify a Gaussian law. A Poincare bound may be loose, and existence or regularity of a kernel must not be assumed outside its stated setting.
Conditional variance is not ordinary variance with a label attached. Its dependence on the observed state is precisely the information that reveals whether the Gaussian coefficient is stable.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Use an evidence and source routine
Record centring and variance, the Stein operator, support and boundary terms, the kernel formula, test-function class, coupling construction, regression factor, conditional second moment, remainder terms, Poincare constant and probability metric.
The Probability Surveys review of univariate Stein comparison develops Stein kernels for general univariate laws. The Electronic Journal of Probability treatment of Stein density methods connects kernels with functional inequalities in broader settings.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Is a Stein kernel a probability density?
No. It is a coefficient in an integration-by-parts identity under the probability law.
Why is the Gaussian kernel constant?
The Gaussian score and derivative identity have a fixed variance coefficient.
Does a smaller Poincare constant always mean a better normal approximation?
No. It controls functional variance; a normal bound still needs the appropriate Stein discrepancy and assumptions.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Continue with a changed-scale check
Repeat the uniform calculation on [−1,1]. Find Var(X), τ(x), E[τ] and Var(τ), then standardise X and confirm that the transformed kernel matches the variance-one ledger.
Continue through the Probability root, the Gold Standard of Education and the How X Works hub.
Top · Contents · Previous · How X Works