I routinely use Gauss-Hermite as a tool for approximating complex integrals. While I am proficient in its applications, I am not proficient in its development. I am working to understand the weights for functions where $g(x) \sim \mathcal{N}(0,1)$ such as in

$$\int f(x) g(x)dx \approx \sum_{i=1}^kf(x_i)w_i$$

Papers such as this below:


indicate that

$$g(x) = exp(-x_i^2)$$

where I am assuming $x_i$ are the corresponding nodes used in $f(x_i)$. I'm unable to replicate this in R as I show below. The first portion of code output shows the nodes and weights from the R function and then the second line is my attempt to apply the transformation to the nodes to create the weights.

> gauss.quad.prob(10, dist = 'normal')
 [1] -4.8594628 -3.5818235 -2.4843258 -1.4659891 -0.4849357  0.4849357  1.4659891  2.4843258  3.5818235  4.8594628

 [1] 4.310653e-06 7.580709e-04 1.911158e-02 1.354837e-01 3.446423e-01 3.446423e-01 1.354837e-01 1.911158e-02 7.580709e-04 4.310653e-06

> exp(-(gauss.quad.prob(10, dist = 'normal')$nodes^2))
 [1] 5.551438e-11 2.680628e-06 2.087319e-03 1.165862e-01 7.904423e-01 7.904423e-01 1.165862e-01 2.087319e-03 2.680628e-06 5.551438e-11

Clearly, I'm mistaken and fail to understand the development of the weights properly and am looking for some didactic support to further my understanding.

Thank you in advance.


1 Answer 1


Intuitively, quadrature approximation of an integral is replacing the function $f(x)$ under the integral by a close enough step-function $\hat f (x)$ (piecewise constant with a finite number of jumps).

Let $-\infty=z_0 < z_1 < ... <z_{n-1} < z_n=\infty$ (pardon this abuse of notation) be the points of discontinuity of this step function, so that $\forall x \in (z_j, z_{j+1}) : \hat f(x) = f(x_i)$. Then the weights must be equal to the integrals of weighting function $g(x)$ over each of the intervals $(z_j,z_{j+1})$:

\begin{equation} f(x) \approx \hat f(x) \Rightarrow \int_{-\infty}^{\infty} f(x)g(x) dx \approx \sum_{i=1}^n f(x_i) \int_{z_{i-1}}^{z_i} g(x) dx, \end{equation}

meaning that the weights must be a bit "fatter" than just $g(x_i)$, rather an integral of $g(x)$ over some neighborhood of $x_i$.

Mechanically, according to Wikipedia weights are NOT simply the weighting function evaluated in the nodes but are computed by the formula

\begin{equation} w_{i}={\frac {2^{n-1}n!{\sqrt {\pi }}}{n^{2}[H_{n-1}(x_{i})]^{2}}} \end{equation}

where n is the number of nodes in desired approximation and $H_n(x)$ is the n-th Hermite polynomial, which is precisely supposed to evaluate the integral of $g$ in a neighborhood of a node.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.