Sub-optimality of Gauss–Hermite quadrature and optimality of trapezoidal rule for functions with finite smoothness

1 Introduction

This paper is concerned with a sub-optimality of Gauss–Hermite quadrature and an optimality of the trapezoidal rule.

Given a function $f : R \to R$ , Gauss–Hermite quadrature is arguably one of the standard numerical integration methods to compute the integral

\int_{R} f (x) \frac{1}{\sqrt{2 π}} e^{- x^{2} / 2} d x .

It is a Gauss-type quadrature, i.e., the quadrature points are the zeros of the degree- $n$ orthogonal polynomial associated with the weight function, and the corresponding quadrature weights are readily defined. With the weight function $ρ (x) := \frac{1}{\sqrt{2 π}} e^{- x^{2} / 2}$ , the orthogonal polynomial we have is the (so-called probabilist’s) Hermite polynomial.

Gauss–Hermite quadrature is widely used; here, we just mention spectral methods [Guo.BY_1999_Hermite, mao2017hermite, Canuto.C_etal_2006_book], and applications in aerospace engineering [braun2021satellite, BBM2020book], finance [Brandimarte2006book, FR2008book, GH2010], and physics [SS2016book, Gezerlis2020book]. Nevertheless, the limitation of this method seems to be less known.

We start with numerical results that illustrate this deficiency. Figure 1 shows a comparison of Gauss–Hermite rule and a suitably truncated trapezoidal rule on $R$ . Here the target function is $f (x) = | x |$ , and the integral is with respect to the standard Gaussian measure. For the trapezoidal rule, we integrate $f (x) ρ (x)$ , upon a suitable cut-off of the domain. We discuss the setting of this experiment in more detail at the end of Section 4.

Figure 1: Absolute integration error for $f (x) = | x |$ . Gauss–Hermite rule exhibits a slower error decay than the trapezoidal rule does.

What we observe in Figure 1 is that while the trapezoidal rule achieves $O (n^{- 1.8})$ , Gauss–Hermite quadrature achieves only a slower convergence rate $O (n^{- 1.0})$ , where $n$ is the number of quadrature points. A similar empirical inefficiency of Gauss–Hermite quadrature is reported in a paper by one of the present authors and Nuyens [NS2021], also in a recent paper by Trefethen [T2021], who pointed out the problems of the quadrature weights decaying too fast relative to the widely-spread quadrature points.

In this paper, we prove a sub-optimality of Gauss–Hermite quadrature. More precisely, we establish a sharp lower bound for the error decay of Gauss–Hermite quadrature in the sense of worst case error. Moreover, we show that a suitably truncated trapezoidal rule achieves the optimal rate of convergence, up to a logarithmic factor.

Integrands of our interest are functions with finite smoothness. In this regard, we will work under the assumption that the integrand function lives in the weighted $L^{2}$ -Sobolev space with smoothness $α \in N$ . This function space is widely used; see for example books [B1998, Canuto.C_etal_2006_book, Shen.J_2011_book] and references therein. For $α \in N$ , this space is equivalent to the so-called Hermite space of finite smoothness, which has been attracting attention in high-dimensional computations; see [IG2015_Hermite, DILP2018, gnewuch2021countable] for its use in high-dimensional computations, and see [B1998, 1.5.4. Proposition] together with Section 2 below for the equivalence.

In this setting, we prove that the worst case integration error of Gauss–Hermite quadrature is bounded from below by $n^{- α / 2}$ up to a constant. This rate matches the upper bound shown by Mastroianni and Monegato [MM1994], and thus cannot be improved. Moreover, this rate provides a rigorous verification of the numerical findings made in [DILP2018, Section 4], where the authors computed approximate values of the worst-case error for Gauss–Hermite quadrature in the Hermite space of finite smoothness, and observed the rate $O (n^{- α / 2})$ for $α = 1, 2,$ and $3$ .

It turns out that this rate is merely half of the best possible: if we allow $n$ quadrature-points and weights to be arbitrary, then the best achievable using (linear) quadratures is of the rate $O (n^{- α})$ ; see [DILP2018, Theorem 1] for a precise statement. Dick et al. [DILP2018] also show that a class of numerical integration methods based on so-called (scaled) higher order digital nets achieve the optimal rate $O (n^{- α})$ up to a logarithmic factor in the multi-dimensional setting, including the one-dimensional setting as a special case.

Our results on the trapezoidal rule show that the trapezoidal rule, a method arguably significantly simpler than one-dimensional higher order digital nets, also achieves this optimal rate, up to a logarithmic factor, and thus nearly twice as fast as the error-decay rate of Gauss–Hermite rule. It is also worth mentioning that Gauss–Hermite quadrature requires a nontrivial algorithm to generate quadrature points, whereas for the trapezoidal rule we simply have equispaced points.

For analytic functions, efficiency of the trapezoidal rule is well known; related studies date back at least to the paper by Goodwin in 1949 [Goodwin1949], and this accuracy is not only widely known to, but also still actively studied by contemporary numerical analysts [S1997, MS2001, GilEtAl2007, Waldvogel2011, TW2014, T2021]. Our results show that this efficiency extends to Sobolev class functions, where we do not have tools from complex analysis such as contour integrals. Our proof uses the strategy recently developed by one of the present authors and Nuyens [NS2021] for a class of quasi-Monte Carlo methods.

We now mention other error estimates of Gauss–Hermite quadrature in the literature. Based on results by Freud

[Freud1972], Smith, Sloan, and Opie [SSO1983] showed an upper bound of the rate

O (n^{- α / 2})

for

α

-times continuously differentiable functions whose

α

-th derivative satisfies a suitable growth condition for

α \in N

. Since the weighted Sobolev space seems to be more frequently used, our focus is on this class. Della Vecchia and Mastroianni [DM2003] showed an upper bound of the rate

O (n^{- 1 / 6})

for

ρ

-integrable a.e. differentiable functions whose derivative is also

ρ

-integrable. Moreover, their result implies a matching lower bound in the sense of worst-case error, and thus in this sense their upper bound is sharp. It does not seem to be trivial if their bounds generalise, for example to the order

n^{- α / 6}

or to

n^{- α / 2 + 1 / 3}

, with the

α

-times differentiability of the integrand for

α \geq 2

. In contrast, our results show that, by assuming the square

ρ

-integrability, the rate improves to

O (n^{- α / 2})

, for general

α \in N

Before moving on, we mention that our results motivate further studies of Gauss–Hermite based algorithms in high-dimensional problems. Integration problems in high-dimension arise, for example in computing statistics of solutions of partial differential equations parametrised by random variables. In particular, integration with respect to the Gaussian measure has been attracting increasing attention; see for example

[Graham.I_eta_2015_Numerische, C2018, YK2019, HS2019ML, D2021, dung2022analyticity]. The measure being Gaussian, algorithms that use Gauss–Hermite quadrature as a building block have gained popularity [C2018, D2021, dung2022analyticity]. The key to proving error estimates in this context is the regularity of the quantity of interest with respect to the parameter, and for elliptic and parabolic problems such smoothness, even an analytic regularity, has been shown [BNT2010SIREV, NT2009parabolic, dung2022analyticity]. In contrast, the solutions of parametric hyperbolic systems suffer from limited regularity under mild assumptions [MNT2013_StochasticCollocationMethod, MNT2015_AnalysisComputationElasticWave]. Hence, our results, which show a suboptimality of Gauss–Hermite rule for functions with finite smoothness, caution us and encourage further studies of the use of algorithms based on Gauss–Hermite rule for this class of problems.

Finally, we note that if we have the weight function $e^{- y^{2}}$ instead of $ρ$ , the corresponding orthogonal polynomials are called physicist’s Hermite polynomials. Our results for Gauss–Hermite quadrature can be obtained for these polynomials by simply rescaling our results by $x = \sqrt{2} y$ . Likewise, results for physicist’s Hermite polynomials in the literature, e.g. in [Szego1975], are used throughout this paper.

The rest of this paper is organized as follows. In Section 2 we introduce necessary definitions such as Hermite polynomials, the weighted Sobolev spaces, and the Hermite spaces. We also discuss the norm equivalence between the weighted Sobolev space and the Hermite space. In Section 3 the sub-optimality of Gauss–Hermite quadrature is shown. In particular, we obtain the matching lower and upper bounds for the worst-case error. In Section 4 the optimality of the trapezoidal rule is shown. Finally, Section 5 concludes this paper.

2 Function spaces with finite smoothness

Throughout this paper, we use the weighted space $L_{ρ}^{2} := L_{ρ}^{2} (R)$ , the normed space consisting of the equivalence classes of Lebesgue measurable functions $f : R \to R$ satisfying $∥ f ∥_{L_{ρ}^{2}}^{2} := \int_{R} | f (x) |^{2} ρ (x) d x < \infty$ , where the equivalence relation is given by $f \sim g$ if and only if $∥ f - g ∥_{L_{ρ}^{2}} = 0$ .

2.1 Hermite polynomials

For $k \in N \cup {0}$ , the $k$ -th degree probabilist’s Hermite polynomial is given by

H_{k} (x) = \frac{(- 1)^{k}}{\sqrt{k!}} e^{x^{2} / 2} \frac{d^{k}}{d x^{k}} e^{- x^{2} / 2}, x \in R,

(1)

where they are normalised so that $∥ H_{k} ∥_{L_{ρ}^{2}} = 1$ for all $k \in N \cup {0}$ . The polynomials $(H_{k})_{k \geq 0}$ form a complete orthonormal system for $L_{ρ}^{2}$ .

The following properties are used throughout the paper.

H_{k}^{'} (x) = \sqrt{k} H_{k - 1} (x), k \geq 1;

(2)

	$\frac{d^{τ}}{d x^{τ}} (H_{k} (x) ρ (x))$
		$= (- 1)^{τ} \sqrt{\frac{(k + τ)!}{k!}} H_{k + τ} (x) ρ (x), k \geq 0, τ \geq 0.$		(3)

2.2 Weighted Sobolev space

The function space we consider is the Sobolev space of square integrable functions, the integrability condition of which is imposed by the Gaussian measure.

Definition (Weighted Sobolev space)

For $α \in N$ , the weighted Sobolev space $H_{α}$ (with the weight function $ρ$ ) is the class of all functions $f \in L_{ρ}^{2}$ such that $f$ has weak derivatives satisfying $f^{(τ)} \in L_{ρ}^{2}$ for $τ = 1, \dots, α$ :

H_{α} := {f \in L_{ρ}^{2} ∣ ∣ ∣ ∥ f ∥_{α} := (α \sum τ = 0 ∥ f^{(τ)} ∥_{L_{ρ}^{2}}^{2})^{1 / 2} < \infty} .

$_{■}$

Elements in $H_{α}$ for $α \in N$ are in the standard local Sobolev space $W_{l o c}^{1, 2} (R)$ , and thus admit a continuous representative. In what follows, we always take the continuous representative of $f \in H_{α}$ .

We recall another important class of functions, the so-called Hermite space. For this space we follow the definition of [DILP2018].

Definition (Hermite space with finite smoothness)

For $α \in N$ , the Hermite space with finite smoothness $H_{α}^{H e r m i t e}$ is given by

where $ˆ f (k) = (f, H_{k})_{L_{ρ}^{2}} := \int_{R} f (x) H_{k} (x) ρ (x) d x$ , and

r_{α} (k) := {\begin{matrix} 1, & i f k = 0, (\sum_{τ = 0}^{α} β_{τ} (k))^{- 1}, & i f k \geq 1, \end{matrix} a n d β_{τ} (k) := {\begin{matrix} \frac{k!}{(k - τ)!}, & i f k \geq τ, 0, & o t h e r w i s e . \end{matrix}

$_{■}$

It turns out $H_{α} = H_{α}^{H e r m i t e}$ , where the equality here means the norm equivalence. Hence, results established for the Hermite space $H_{α}^{H e r m i t e}$ can be readily translated to $H_{α}$ up to a constant, which allows us to compare the results on higher order digital nets in [DILP2018] with ours.

A proof of this equivalence of the norm is outlined in [B1998, 1.5.4. Proposition]; see also [DILP2018, Lemma 6]. Here, for completeness we give a proof.

Lemma 1

Let $f \in H_{α}$ with $α \in N$ , then $f \in H_{α}^{H e r m t i t e}$ with the same smoothness parameter $α$ . $_{□}$

Proof

We first prove the claim for $α = 1$ . Assume $f \in H_{1}$ . Let $F (x) := f (x) (ρ (x))^{1 / 2 + ε}$ and $G (x) := H_{k} (x) (ρ (x))^{1 / 2 - ε}$ for $0 < ε < 1 / 2$ . We have

\int_{R} F^{'} (x) ϕ (x) d x = - \int_{R} F (x) ϕ^{'} (x) d x,

for any function $ϕ$ in the space of compactly supported infinitely differentiable functions $C_{c}^{\infty} (R)$ . Since $G$ is in the standard Sobolev space $W^{1, 2} (R)$ , there exists a sequence ${ϕ_{N}}_{N \in N} \subset C_{c}^{\infty} (R)$ that satisfies

∥ G - ϕ_{N} ∥_{L^{2} (R)}^{2} + ∥ G^{'} - ϕ_{N}^{'} ∥_{L^{2} (R)}^{2} = ∥ G - ϕ_{N} ∥_{W^{1, 2} (R)}^{2} \leq 1 / N .

Then, letting $g \in L^{2} (R)$ , the Cauchy–Schwarz inequality implies $\int_{R} | F (x) g (x) | d x \leq ∥ f ∥_{L_{ρ}^{2}} ∥ g ∥_{L^{2} (R)} < \infty$ , and for $F^{'} (x) = f^{'} (x) (ρ (x))^{1 / 2 + ε} - (ε + 1 / 2) x f (x) (ρ (x))^{1 / 2 + ε}$ we also have

	$\int_{R} \| F^{'} (x) g (x) \| d x$	$\leq {(\int_{R} \| f^{'} (x) \|^{2} e^{- x^{2} / 2} d x)}_{R}^{1 / 2} {(\int_{R} \| g (x) \| d x)}_{R}^{1 / 2}$
		$+ sup t \in R \| (ε + 1 / 2)^{2} t^{2} e^{- ε t^{2}} \| {(\int_{R} \| f (x) \|^{2} e^{- x^{2} / 2} d x)}_{R}^{1 / 2} {(\int_{R} \| g (x) \|^{2} d x)}_{R}^{1 / 2} < \infty .$

Therefore both $⟨ F, \cdot ⟩ := \int_{R} F (x) \cdot d x$ and $⟨ F^{'}, \cdot ⟩ := \int_{R} F^{'} (x) \cdot d x$ define a continuous functional on $L^{2} (R)$ . Hence, we have $\int_{R} F^{'} (x) G (x) d x = - \int_{R} F (x) G^{'} (x) d x$ and thus

		$\int_{R} f^{'} (x) H_{k} (x) ρ (x) - (ε + 1 / 2) x H_{k} (x) f (x) ρ (x) d x = \int_{R} F^{'} (x) G (x) d x$
	$=$	$- \int_{R} F (x) G^{'} (x) d x = - (\int_{R} f (x) \sqrt{k} H_{k - 1} (x) ρ (x) - (- ε + 1 / 2) x H_{k} (x) f (x) ρ (x) d x),$

which is equivalent to

	$\int_{R} f^{'} (x)$	$H_{k} (x) ρ (x) = - \int_{R} f (x) \sqrt{k} H_{k - 1} (x) ρ (x) - x H_{k} (x) f (x) ρ (x) d x$		(4)
		$= - \int_{R} f (x) (H_{k} (x) ρ (x))^{'} d x = \int_{R} f (x) \sqrt{k + 1} H_{k + 1} (x) ρ (x) d x,$

where we used $(H_{k} (x) ρ (x))^{'} = - \sqrt{k + 1} H_{k + 1} (x) ρ (x)$ . Hence we obtain

(f^{'}, H_{k})_{L_{ρ}^{2}}^{2} = (k + 1) (f, H_{k + 1})_{L_{ρ}^{2}}^{2},

and

\infty \sum k = 0 (k + 1) (f, H_{k + 1})_{L_{ρ}^{2}}^{2} = \infty \sum k = 0 (f^{'}, H_{k})_{L_{ρ}^{2}}^{2} = ∥ f^{'} ∥_{L_{ρ}^{2}}^{2} < \infty .

This implies $f \in H_{1}^{H e r m i t e}$ since $r_{1}^{- 1} (k) = k + 1$ . For general $τ = 1, . . ., α$ , assuming $f \in H_{τ}$ and by repeating the same argument as above, we have $(f^{(τ)}, H_{k})_{L_{ρ}^{2}}^{2} = (f, H_{k + τ})_{L_{ρ}^{2}}^{2} \prod_{j = 1}^{τ} (k + j)$ for $k \geq 0$ and thus

	$∥ f^{(τ)} ∥_{L_{ρ}^{2}}^{2}$	$= \infty \sum k = 0 (f, H_{k + τ})_{L_{ρ}^{2}}^{2} τ \prod j = 1 (k + j) \geq \frac{1}{τ^{τ}} \infty \sum k = 0 (f, H_{k + τ})_{L_{ρ}^{2}}^{2} (k + τ)^{τ}$
		$= \frac{1}{τ^{τ}} \infty \sum k = τ k^{τ} (f, H_{k})_{L_{ρ}^{2}}^{2} .$

Hence, observing ${lim}_{k \to \infty} r_{τ} (k) k^{τ} = 1$ (see also [DILP2018, p. 687]), we conclude $f \in H_{τ}^{H e r m i t e}$ . $_{■}$

3 Matching bounds for Gauss–Hermite quadrature

In this section, we prove the sub-optimality of Gauss–Hermite quadrature. We first introduce the following linear quadrature of general form

Q_{n} (f) = n \sum j = 1 w_{j} f (ξ_{j})

(5)

with arbitrary $n$ distinct quadrature points on the real line

- \infty < ξ_{1} < ξ_{2} < \dots < ξ_{n} < \infty

and quadrature weights $w_{1}, \dots, w_{n} \in R$ . Gauss–Hermite quadrature $Q_{n}^{G H}$ is given by the points $(ξ_{j}^{G H})_{j = 1, \dots, n}$ being the roots of $H_{n}$ and the weights being $w_{j} = 1 / [H_{n}^{'} (ξ_{j}^{G H})]^{2}$ , $j = 1, \dots, n$ ; see for example [Shen.J_2011_book, Theorem 3.5].

Given a quadrature rule $Q_{n}$ , it is convenient to introduce the notation

e^{w o r} (Q_{n}, H_{α}) := sup 0 \neq f \in H_{α} \frac{| I (f) - Q_{n} (f) |}{∥ f ∥_{α}} .

The quantity $e^{w o r} (Q_{n}, H_{α})$ is commonly referred to as the worst-case error of $Q_{n}$ in $H_{α}$ ; see for example [Dick.J_Kuo_Sloan_2013_ActaNumerica]. Now we can state our aim of this section more precisely: we prove the matching lower and upper bounds on $e^{w o r} (Q_{n}^{G H}, H_{α})$ for Gauss–Hermite quadrature.

3.1 Lower bound

We first derive the following lower bound on $e^{w o r} (Q_{n}, H_{α})$ for the general quadrature (5).

Lemma 2

Let $α \in N$ . For $n \geq 2$ , let

σ := {\begin{matrix} α & if {min}_{j = 1, \dots, n - 1} (ξ_{j + 1} - ξ_{j}) \leq 1. 0 & otherwise. \end{matrix}

(6)

Then, there exists a constant $c_{α} > 0$ , which depends only on $α$ , such that the worst-case error $e^{w o r} (Q_{n}, H_{α})$ of a general function-value based linear quadrature (5) in the weighted Sobolev space $H_{α}$ is bounded below by

	$e^{w o r} (Q_{n}, H_{α}) \geq c_{α} min i = 1, \dots, n - 1 (ξ_{i + 1} - ξ_{i})^{σ + 1 / 2}$
	$\times n - 1 \sum j = 1 e^{- max (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2} {(n - 1 \sum k = 1 e^{- 1_{\geq 0} (ξ_{k} ξ_{k + 1}) min (ξ_{k}^{2}, ξ_{k + 1}^{2}) / 2})}^{- 1 / 2},$

where $1_{\geq 0} (x)$ is equal to 1 if $x \geq 0$ and 0 otherwise. $_{□}$

Proof

The heart of the matter is to construct a function $0 \neq h_{n} \in H_{α}$ such that $h_{n} (ξ_{j}) = 0$ for all $1 \leq j \leq n$ , resulting in $Q_{n} (h_{n}) = 0$ , and that $∥ h_{n} ∥_{α}$ is small but $I (h_{n})$ is large. Define a function $h : R \to R$ by

h (x) := h_{n} (x) := ⎧ ⎪ ⎨ ⎪ ⎩ \begin{matrix} (\frac{x - ξ_{j}}{ξ_{j + 1} - ξ_{j}})^{α} (1 - \frac{x - ξ_{j}}{ξ_{j + 1} - ξ_{j}})^{α} & \begin{matrix} if there % exists j \in {1, \dots, n - 1} such that x \in [ξ_{j}, ξ_{j + 1}], \end{matrix} 0 & \begin{matrix} otherwise. \end{matrix} \end{matrix}

Then, $h$ turns out to fulfill our purpose. This type of fooling function used to prove lower bounds for the worst-case error is called bump function (of finite smoothness), in quasi-Monte Carlo theory; see, for instance, [DHP2015, Section 2.7].

First we show $h \in H_{α}$ . It follows from

{(\frac{x - ξ_{j}}{ξ_{j + 1} - ξ_{j}})}^{α} {(1 - \frac{x - ξ_{j}}{ξ_{j + 1} - ξ_{j}})}^{α} = α \sum ℓ = 0 (- 1)^{ℓ} (\frac{}{α}) ℓ {(\frac{x - ξ_{j}}{ξ_{j + 1} - ξ_{j}})}^{α + ℓ}

that we have

h^{(τ)} (x) = \frac{1}{(ξ_{j + 1} - ξ_{j})^{τ}} α \sum ℓ = 0 (- 1)^{ℓ} (\frac{α}{ℓ}) \frac{(α + ℓ)!}{(α + ℓ - τ)!} {(\frac{x - ξ_{j}}{ξ_{j + 1} - ξ_{j}})}^{α + ℓ - τ}

for $τ = 0, 1, \dots, α$ and any $ξ_{j} < x < ξ_{j + 1}$ . As we have $h^{(τ)} (ξ_{j}) = h^{(τ)} (ξ_{j + 1}) = 0$ for $τ = 0, 1, \dots, α - 1$ , the function $h$ is $(α - 1)$ -times continuously differentiable. Moreover, $h^{(α - 1)}$ is continuous piecewise polynomial and thus weakly differentiable. Also, noting that $1_{\geq 0} (ξ_{j} ξ_{j + 1}) = 0$ only when $ξ_{j} < 0 < ξ_{j + 1}$ and otherwise $1_{\geq 0} (ξ_{j} ξ_{j + 1}) = 1$ , we have

	$\int_{ξ_{j}}^{ξ_{j + 1}} \| h^{(τ)} (x) \|^{2} ρ (x) d x$
	$\leq \frac{e^{- 1_{\geq 0} (ξ_{j} ξ_{j + 1}) min (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2}}{\sqrt{2 π}} \int_{ξ_{j}}^{ξ_{j + 1}} \| h^{(τ)} (x) \|^{2} d x$
	$= \frac{e^{- 1_{\geq 0} (ξ_{j} ξ_{j + 1}) min (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2}}{\sqrt{2 π} (ξ_{j + 1} - ξ_{j})^{2 τ}}$

	$= \frac{e^{- 1_{\geq 0} (ξ_{j} ξ_{j + 1}) min (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2}}{\sqrt{2 π} (ξ_{j + 1} - ξ_{j})^{2 τ - 1}}$
	$\times α \sum ℓ_{1}, ℓ_{2} = 0 \frac{(- 1)^{ℓ_{1} + ℓ_{2}}}{2 (α - τ) + ℓ_{1} + ℓ_{2} + 1} (\frac{α}{ℓ_{1}}) (\frac{α}{ℓ_{2}}) \frac{(α + ℓ_{1})!}{(α + ℓ_{1} - τ)!} \frac{(α + ℓ_{2})!}{(α + ℓ_{2} - τ)!},$

for $τ = 0, 1, \dots, α$ . The last sum over $ℓ_{1}$ and $ℓ_{2}$ does not depend on $j$ . Denoting this sum by $S_{α, τ}$ , we obtain

	$∥ h ∥_{α}^{2}$	$= α \sum τ = 0 \int_{R} \| h^{(τ)} (x) \|^{2} ρ (x) d x = α \sum τ = 0 n - 1 \sum j = 1 \int_{ξ_{j}}^{ξ_{j + 1}} \| h^{(τ)} (x) \|^{2} ρ (x) d x$
		$\leq \frac{1}{\sqrt{2 π}} α \sum τ = 0 S_{α, τ} n - 1 \sum j = 1 \frac{e^{- 1_{\geq 0} (ξ_{j} ξ_{j + 1}) min (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2}}{(ξ_{j + 1} - ξ_{j})^{2 τ - 1}}$
		$\leq \frac{1}{\sqrt{2 π}} α \sum τ = 0 \frac{S_{α, τ}}{{min}_{1 \leq i \leq n - 1} (ξ_{i + 1} - ξ_{i})^{2 τ - 1}} n - 1 \sum j = 1 e^{- 1_{\geq 0} (ξ_{j} ξ_{j + 1}) min (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2}$
		$\leq \frac{1}{\sqrt{2 π} {min}_{1 \leq i \leq n - 1} (ξ_{i + 1} - ξ_{i})^{2 σ - 1}} α \sum τ = 0 S_{α, τ} n - 1 \sum j = 1 e^{- 1_{\geq 0} (ξ_{j} ξ_{j + 1}) min (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2} < \infty .$

This proves $h \in H_{α}$ .

By definition of $h$ , we have $h (ξ_{j}) = 0$ for all $j = 1, \dots, n$ , and thus

Q_{n} (h) = 0.

Moreover, we have

	$I (h)$	$= \int_{R} h (x) ρ (x) d x = n - 1 \sum j = 1 \int_{ξ_{j}}^{ξ_{j + 1}} h (x) ρ (x) d x$
		$\geq \frac{1}{\sqrt{2 π}} n - 1 \sum j = 1 e^{- max (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2} \int_{ξ_{j}}^{ξ_{j + 1}} {(\frac{x - ξ_{j}}{ξ_{j + 1} - ξ_{j}})}^{α} {(1 - \frac{x - ξ_{j}}{ξ_{j + 1} - ξ_{j}})}^{α} d x$
		$= \frac{1}{\sqrt{2 π}} n - 1 \sum j = 1 e^{- max (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2} (ξ_{j + 1} - ξ_{j}) \int_{0}^{1} x^{α} {(1 - x)}^{α} d x$
		$= \frac{(α!)^{2}}{(2 α + 1)! \sqrt{2 π}} n - 1 \sum j = 1 e^{- max (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2} (ξ_{j + 1} - ξ_{j})$
		$\geq \frac{(α!)^{2}}{(2 α + 1)! \sqrt{2 π}} min 1 \leq i \leq n - 1 (ξ_{i + 1} - ξ_{i}) n - 1 \sum j = 1 e^{- max (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2} .$

Using the above results, we obtain

	$e^{w o r} (Q_{n}, H_{α})$	$\geq \frac{\| I (h) - Q_{n} (h) \|}{∥ h ∥_{α}}$
		$\geq \frac{(α!)^{2}}{(2 α + 1)! (2 π)^{1 / 4}} {(α \sum τ = 0 S_{α, τ})}_{α, τ}^{- 1 / 2} min 1 \leq i \leq n - 1 (ξ_{i + 1} - ξ_{i})^{σ + 1 / 2}$
		$\times n - 1 \sum j = 1 e^{- max (ξ_{j}^{2}, ξ_{j + 1}^{2}) / 2} {(n - 1 \sum k = 1 e^{- 1_{\geq 0} (ξ_{k} ξ_{k + 1}) min (ξ_{k}^{2}, ξ_{k + 1}^{2}) / 2})}^{- 1 / 2} .$

Now the proof is complete. $_{■}$

Using the general lower bound in Lemma 2, we obtain the following lower bound on the worst-case error for Gauss–Hermite quadrature.

Theorem 1

Let $α \in N$ . For any $n \geq 2$ , the worst-case error of the Gauss–Hermite quadrature in the weighted Sobolev space $H_{α}$ is bounded from below as

e^{w o r} (Q_{n}^{G H}, H_{α}) \geq C_{α} n^{- α / 2}

with a constant $C_{α} > 0$ that depends on $α$ but independent of $n$ . $_{□}$

Proof

Let $ξ_{j}^{G H}$ , $j = 1, \dots, n$ , be the roots of $H_{n}$ . For any $n \geq 2$ , it holds that

\frac{π}{\sqrt{n + 1 / 2}} < min j = 1, \dots, n - 1 (ξ_{j + 1}^{G H} - ξ_{j}^{G H}) \leq \frac{\sqrt{21 / 2}}{\sqrt{n + 1 / 2}};

(7)

see, for instance, [Szego1975, Eq. (6.31.22)]. Thus, to invoke Lemma 2 we let

σ := {\begin{matrix} α & for n \geq 10, 0 & for 2 \leq n < 10, \end{matrix}

so that the conditions in (6) are satisfied. Also, each node $ξ_{j}^{G H}$ is bounded below and above as follows; see, for instance, [Szego1975, Eq. (6.31.19)]: for $n$ odd, we have $ξ_{(n + 1) / 2}^{G H} = 0$ with the positive zeros satisfying

\frac{j π}{\sqrt{n + 1 / 2}} < ξ_{(n + 1) / 2 + j}^{G H} < \frac{4 j + 3}{\sqrt{n + 1 / 2}} for j = 1, \dots, (n - 1) / 2,

(8)

and for $n$ even,

\frac{(j - 1 / 2) π}{\sqrt{n + 1 / 2}} < ξ_{n / 2 + j}^{G H} < \frac{4 j + 1}{\sqrt{n + 1 / 2}} for j = 1, \dots, n / 2,

(9)

with symmetricity $ξ_{j}^{G H} = - ξ_{n + 1 - j}^{G H}$ , $1 \leq j \leq n$ for $n \geq 2$ odd and even.

Let $n$ be odd. Using the result in Lemma 2, equations (7), and (8), together with the symmetricity of the Hermite zeros, we obtain

	$e^{w o r} (Q_{n}^{G H}, H_{α})$

	$\geq 2 c_{α} {(\frac{π^{2}}{n + 1 / 2})}^{σ / 2 + 1 / 4} (n - 1) / 2 \sum j = 1 e^{- (4 j + 3)^{2} / (2 n + 1)} {⎛ ⎝ 2 + 2 (n - 1) / 2 \sum k = 2 e^{- π^{2} (k - 1)^{2} / (2 n + 1)} ⎞ ⎠}^{- 1 / 2} .$

The sum over $j$ is further bounded below by

	$(n - 1) / 2 \sum j = 1 e^{- (4 j + 3)^{2} / (2 n + 1)}$	$\geq \int_{1}^{(n - 1) / 2 + 1} e^{- (4 x + 3)^{2} / (2 n + 1)} d x$
		$= \frac{\sqrt{n + 1 / 2}}{4} \int_{7 / \sqrt{n + 1 / 2}}^{(2 n + 5) / \sqrt{n + 1 / 2}} e^{- x^{2} / 2} d x$
		$\geq \frac{\sqrt{n + 1 / 2}}{4} \int_{\sqrt{14}}^{11 \sqrt{2} / \sqrt{7}} e^{- x^{2} / 2} d x$
		$= \frac{\sqrt{π (n + 1 / 2)}}{4 \sqrt{2}} (erf (11 / \sqrt{7}) - erf (\sqrt{7})),$

where $erf$ denotes the error function, and the last inequality holds for any odd $n \geq 3$ . The sum over $k$ is further bounded above by

	$(n - 1) / 2 \sum k = 2 e^{- π^{2} (k - 1)^{2} / (2 n + 1)}$	$\leq \int_{1}^{(n - 1) / 2} e^{- π^{2} (x - 1)^{2} / (2 n + 1)} d x$
		$\leq \sqrt{n + 1 / 2} \int_{0}^{\infty} e^{- π^{2} x^{2} / 2} d x = \sqrt{\frac{n + 1 / 2}{2 π}} .$

Using these bounds, we have

	$e^{w o r} (Q_{n}^{G H}, H_{α})$	$\geq 2 c_{α} {(\frac{π^{2}}{n + 1 / 2})}^{σ / 2 + 1 / 4} \frac{\sqrt{π (n + 1 / 2)}}{4 \sqrt{2}} (erf (11 / \sqrt{7}) - erf (\sqrt{7}))$
		$\times {⎛ ⎝ 2 + 2 \sqrt{\frac{n + 1 / 2}{2 π}} ⎞ ⎠}^{- 1 / 2}$
		$\geq c_{α} π^{σ + 1 / 4} \frac{erf (11 / \sqrt{7}) - erf (\sqrt{7})}{2 \sqrt{2} \sqrt{2 + \sqrt{2}}} \frac{1}{(n + 1 / 2)^{σ / 2}}$
		$\geq c_{α} π^{1 / 4} \frac{erf (11 / \sqrt{7}) - erf (\sqrt{7})}{2^{(α + 5) / 2}} \frac{1}{n^{α / 2}} .$

Let $n$ be even. As in the odd case, but now using (9) instead of (8), we obtain

	$e^{w o r} (Q_{n}^{G H}, H_{α})$
	$\geq c_{α} {(\frac{π^{2}}{n + 1 / 2})}^{σ / 2 + 1 / 4}$
	$\times (e^{- (ξ_{n / 2 + 1}^{G H}})^{2} / 2 + 2 n / 2 \sum j = 2 e^{- (ξ_{n / 2 + j}^{G H})^{2} / 2}) (1 + 2 n / 2 \sum k = 2 e^{- (ξ_{n / 2 + k - 1}^{G H})^{2} / 2})^{- 1 / 2}$
	$\geq c_{α} {(\frac{π^{2}}{n + 1 /}}$

Sub-optimality of Gauss–Hermite quadrature and optimality of trapezoidal rule for functions with finite smoothness

Randomizing the trapezoidal rule gives the optimal RMSE rate in Gaussian Sobolev spaces

Optimal functional supervised classification with separation condition

Optimal Complexity and Certification of Bregman First-Order Methods

Central limit theorems for discretized occupation time functionals

Infinite-Variate L^2-Approximation with Nested Subspace Sampling

Optimality of the recursive Neyman allocation

Exponential Weights Algorithms for Selective Learning

1 Introduction

2 Function spaces with finite smoothness

2.1 Hermite polynomials

2.2 Weighted Sobolev space

Definition (Weighted Sobolev space)

Definition (Hermite space with finite smoothness)

Lemma 1

Proof

3 Matching bounds for Gauss–Hermite quadrature

3.1 Lower bound

Lemma 2

Proof

Theorem 1

Proof

Sub-optimality of Gauss–Hermite quadrature and optimality of trapezoidal rule for functions with finite smoothness

Related Research

Randomizing the trapezoidal rule gives the optimal RMSE rate in Gaussian Sobolev spaces

Optimal functional supervised classification with separation condition

Optimal Complexity and Certification of Bregman First-Order Methods

Central limit theorems for discretized occupation time functionals

Infinite-Variate L^2-Approximation with Nested Subspace Sampling

Optimality of the recursive Neyman allocation

Exponential Weights Algorithms for Selective Learning

1 Introduction

2 Function spaces with finite smoothness

2.1 Hermite polynomials

2.2 Weighted Sobolev space

Definition (Weighted Sobolev space)

Definition (Hermite space with finite smoothness)

Lemma 1

Proof

3 Matching bounds for Gauss–Hermite quadrature

3.1 Lower bound

Lemma 2

Proof

Theorem 1

Proof