---
title: Higher Russian Formula
url: https://www.emergentmind.com/topics/higher-russian-formula
type: topic
---

# Higher Russian Formula

Searching arXiv for the specified paper and closely related work on Frolov cubature.
Higher Russian Formula denotes the higher-order “Russian” cubature of Frolov: a lattice-based numerical integration rule on the unit cube whose error is analyzed in function spaces of dominating mixed smoothness with homogeneous boundary condition. In the formulation studied in "The role of Frolov's cubature formula for functions with bounded mixed derivative" [1503.08846], the rule yields upper bounds for worst-case integration errors in Besov spaces \(\mathbf{B}^s_{p,\theta}\) and Triebel–Lizorkin spaces \(\mathbf{F}^s_{p,\theta}\) over the whole range of admissible parameters \((s\ge 1/p)\). The analysis isolates two qualitatively different smoothness regimes: a “large-smoothness” regime in which the bounds are optimal and a “small-smoothness” regime, especially for \(\mathbf{F}^s_{p,\theta}\) with \(1<\theta<p<\infty\) and \(1/p<s\le 1/\theta\), where the upper bounds have a different logarithmic structure and the matching optimality theory remains open [1503.08846].

## 1. Lattice construction and cubature rule

Let \(d\ge 1\). The construction starts from a non-singular Vandermonde matrix \(T\), where \(\{\xi_i\}_{i=1}^d\) are the distinct real roots of an integer polynomial of degree \(d\), so that \(\det T\neq 0\) [1503.08846]. For each integer \(n\ge 1\), one defines
\[
T_n := n^{-1/d}T,
\qquad
X_n := T_n(\mathbb{Z}^d)\cap [0,1)^d.
\]
Because \(\det T_n = n^{-1}\), one has \(\#X_n \simeq n\). The Frolov cubature rule is
\[
Q_n(f) = \frac{1}{n}\sum_{x\in X_n} f(x),
\]
equivalently a cubature rule with uniform weights \(\lambda_i=1/n\) and nodes \(x^i\in X_n\).

The central structural feature is the choice of lattice. Its dual lattice
\[
B_n(\mathbb{Z}^d)=\bigl(T_n^{-1}\bigr)^T(\mathbb{Z}^d)
\]
satisfies the product-size estimate
\[
\prod_{j=1}^d |z_j| \gtrsim n
\qquad
\forall z=(z_1,\dots,z_d)\in B_n(\mathbb{Z}^d)\setminus\{0\}.
\]
This estimate, together with a refined counting lemma for dual-lattice points in dyadic boxes, is the mechanism that makes the cubature rule effective for mixed-smoothness classes. A plausible implication is that the geometry of the dual lattice, rather than only the primal sampling set, determines the achievable asymptotic error order.

## 2. Function-space setting

The analysis is carried out in Besov and Triebel–Lizorkin spaces of dominating mixed smoothness on \(\mathbb{R}^d\), defined by a tensorized local-mean decomposition [1503.08846]. One fixes univariate Schwartz kernels \(\Psi_0,\Psi_1\) satisfying the moment conditions
\[
\widehat{\Psi}_0(\xi)\neq 0 \text{ for } |\xi|<\varepsilon,
\qquad
\widehat{\Psi}_1(\xi)\neq 0 \text{ for } 1<|\xi|<2\varepsilon,
\]
and
\[
D^\alpha \widehat{\Psi}_1(0)=0
\qquad
(0\le \alpha \le L).
\]
For \(j\ge 1\),
\[
\Psi_j(x)=2^{j-1}\Psi_1(2^{j-1}x),
\]
and the multivariate kernels \(\Psi_{\mathbf m}\) are obtained by tensorization.

The Besov space of dominating mixed smoothness \(B^s_{p,\theta}(\mathbb{R}^d)\) consists of all tempered distributions \(f\) such that
\[
\|f\|_{B^s_{p,\theta}}
=
\left(
\sum_{\mathbf m\in \mathbb{N}_0^d}
2^{\,s|\mathbf m|_1\theta}
\|\Psi_{\mathbf m}*f\|_{L^p(\mathbb{R}^d)}^\theta
\right)^{1/\theta}
<\infty,
\]
with the usual \(\ell^\theta\)-modification for \(\theta=\infty\).

The Triebel–Lizorkin space of dominating mixed smoothness \(F^s_{p,\theta}(\mathbb{R}^d)\) is defined by
\[
\|f\|_{F^s_{p,\theta}}
=
\left\|
\left(
\sum_{\mathbf m\in \mathbb{N}_0^d}
2^{\,s|\mathbf m|_1\theta}
|\Psi_{\mathbf m}*f(\cdot)|^\theta
\right)^{1/\theta}
\right\|_{L^p(\mathbb{R}^d)}
<\infty.
\]
When \(\theta=2\) and \(1<p<\infty\), one recovers the mixed-Sobolev space \(W^s_p\).

To impose homogeneous boundary conditions, one sets
\[
A^{s,\circ}_{p,\theta}
=
\{\,f\in A^s_{p,\theta}(\mathbb{R}^d): \operatorname{supp}f\subset [0,1]^d\,\},
\qquad
A=B \text{ or } F.
\]
The stated purpose of this restriction is that Frolov’s lattice sum does not “see” boundary artifacts. This places the cubature problem in a compactly supported mixed-smoothness setting aligned with the unit-cube integration domain.

## 3. Minimal worst-case error and the large-smoothness regime

The \(n\)th minimal worst-case error in a normed class \(\mathcal F\) is
\[
\operatorname{Int}_n(\mathcal F)
=
\inf_{\{x^i,\lambda_i\}_{i=1}^n}
\sup_{\substack{f\in \mathcal F\\ \|f\|_{\mathcal F}\le 1}}
\left|
\int_{[0,1]^d} f(x)\,dx - \sum_{i=1}^n \lambda_i f(x^i)
\right|.
\]
The paper bounds \(\operatorname{Int}_n(A^{s,\circ}_{p,\theta})\) from above by the Frolov rule \(Q_n\) [1503.08846].

Theorem A covers the large-smoothness regime. Let \(1\le p\le \infty\), \(1\le \theta\le \infty\), with \(p<\infty\) in the \(F\)-case, and suppose
\[
s>\max\{1/p,\,1/\theta\}.
\]
Then
\[
\operatorname{Int}_n\bigl(B^{s,\circ}_{p,\theta}\bigr)
\simeq
n^{-s}(\log n)^{(d-1)(1-1/\theta)},
\]
and likewise
\[
\operatorname{Int}_n\bigl(F^{s,\circ}_{p,\theta}\bigr)
\simeq
n^{-s}(\log n)^{(d-1)(1-1/\theta)}.
\]

These formulas identify the asymptotic rate in both the Besov and Triebel–Lizorkin scales. The power of \(n^{-s}\) reflects the mixed smoothness parameter \(s\), while the logarithmic correction depends on \(d\) and the fine-scale exponent \(\theta\). In this regime the upper bounds are optimal in the sense that they cannot be improved by any other cubature formula.

## 4. Small smoothness in the Triebel–Lizorkin scale

Theorem B treats the difficult small-smoothness case in the \(F\)-scale [1503.08846]. Let \(1<p<\infty\), \(1\le \theta < p\), and assume
\[
\frac{1}{p}<s\le \frac{1}{\theta}.
\]
Then the Frolov rule satisfies
\[
\operatorname{Int}_n\bigl(F^{s,\circ}_{p,\theta}\bigr)
\lesssim
\begin{cases}
n^{-s}(\log n)^{(d-1)(1-s)}, & 1/p<s<1/\theta,\\[1ex]
n^{-s}(\log n)^{(d-1)(1-s)}(\log\log n)^{1-s}, & s=1/\theta.
\end{cases}
\]

The paper emphasizes that these upper bounds show a completely different behavior compared to “large” smoothness \(s>1/\theta\). The distinction is not in the principal power \(n^{-s}\), which remains unchanged, but in the exponent of \(\log n\) and the appearance of an additional \(\log\log n\) factor at the critical value \(s=1/\theta\).

For the mixed-Sobolev class \(W^{s,\circ}_p\), obtained by setting \(\theta=2\), the consequences are stated explicitly:
\[
\operatorname{Int}_n(W^{s,\circ}_p)\simeq n^{-s}(\log n)^{(d-1)/2}
\qquad\text{if } s>1/2,
\]
\[
\operatorname{Int}_n(W^{s,\circ}_p)\lesssim n^{-s}(\log n)^{(d-1)(1-s)}
\qquad\text{if } 1/p<s<1/2,
\]
and
\[
\operatorname{Int}_n(W^{s,\circ}_p)\lesssim
n^{-s}(\log n)^{(d-1)/2}(\log\log n)^{1-s}
\qquad\text{if } s=1/2.
\]

A common misconception is that mixed-Sobolev error bounds necessarily retain the same logarithmic exponent across all admissible \(s\). The stated results show otherwise: the logarithmic term changes once the smoothness crosses the threshold \(1/\theta\), and the critical line may introduce a further iterated logarithm.

## 5. Analytical mechanism behind the two regimes

The proof combines Poisson summation, a discrete Calderón reproducing formula, and dual-lattice counting estimates [1503.08846]. There exist Schwartz-class systems \(\{\Lambda_{\mathbf m}\}\) and \(\{\Psi_{\mathbf m}\}\) such that
\[
\sum_{\mathbf m\in \mathbb{N}_0^d}
\widehat{\Lambda}_{\mathbf m}(\xi)\,\widehat{\Psi}_{\mathbf m}(\xi)=1,
\qquad
\forall \xi\in \mathbb{R}^d.
\]
For \(f\) supported in \([0,1]^d\), Poisson summation and reproducing yield the error identity
\[
I(f)-Q_n(f)
=
\sum_{\mathbf m\in \mathbb{N}_0^d}
\sum_{z\in B_n(\mathbb{Z}^d)\setminus\{0\}}
\widehat{\Lambda}_{\mathbf m}(z)\,
\widehat{\Psi_{\mathbf m}*f}(z).
\]

The dual lattice points are partitioned into dyadic boxes, and
\[
Z_n(\mathbf m)=\#\bigl(B_n(\mathbb{Z}^d)\cap I_{\mathbf m}\bigr)
\]
denotes the number of dual points in the box \(I_{\mathbf m}\). The counting lemma gives
\[
Z_n(\mathbf m)=0
\qquad
\text{for } |\mathbf m|_1\le \log_2 n-c,
\]
and
\[
Z_n(\mathbf m)\lesssim 2^{|\mathbf m|_1}/n
\qquad
\text{for } |\mathbf m|_1>\log_2 n-c.
\]

From the error identity one applies Hölder to the inner sum over \(z\), producing a factor \(Z_n(\mathbf m)^{\,1-1/p}\), and then performs a second summation over \(\mathbf m\). The decisive point is the order in which Hölder is applied across the \(\ell_p\) and \(\ell_\theta\) structures. If \(s>1/\theta\), one groups first in the \(\ell_\theta(\mathbf m)\)-norm and then in \(\ell_p\), recovering the rate
\[
n^{-s}(\log n)^{(d-1)(1-1/\theta)}.
\]
If \(s\le 1/\theta\) but still \(s>1/p\), one must switch the order of Hölder to avoid divergence. The result is a different logarithmic exponent, namely \((d-1)(1-s)\), with an additional \(\log\log n\) term when \(s=1/\theta\).

This suggests that the phase transition at \(s=1/\theta\) is not an artifact of proof technique alone. Rather, it is tied to the interaction between mixed smoothness, fine-scale summability, and the combinatorics of dual-lattice occupation numbers.

## 6. Optimality, lower bounds, and open problems

In the large-smoothness case \(s>1/\theta\), matching lower bounds obtained via atomic constructions show that no cubature can do better than
\[
n^{-s}(\log n)^{(d-1)(1-1/\theta)}.
\]
Accordingly, Frolov’s rule is optimal in that regime [1503.08846]. The paper characterizes this as the regime in which the upper bounds are optimal and cannot be improved by any other cubature formula.

In the small-smoothness case \(1/p<s\le 1/\theta\), the upper bounds are believed sharp, but the matching lower bound is still open in full generality. The summary states that it is known for \(d=2\) via Fibonacci-lattice methods. The abstract likewise notes that the optimality for “small” smoothness is open.

The resulting picture is therefore asymmetric. For \(s>\max\{1/p,1/\theta\}\), the theory provides a sharp order of convergence in both \(B\)- and \(F\)-scales. For \(1/p<s\le 1/\theta\) in the Triebel–Lizorkin scale, the upper theory is precise enough to reveal a qualitatively different logarithmic behavior, but the full minimax classification remains incomplete. This leaves the higher Russian Formula as both a definitive optimal construction in one parameter regime and an unresolved benchmark in another.

Source: https://www.emergentmind.com/topics/higher-russian-formula