Korobov Functions in Approximation & QMC

• Korobov functions are defined in multiple frameworks, including periodic RKHS with dominating mixed smoothness and mixed-Sobolev classes with vanishing boundary traces.
• They facilitate advanced numerical methods such as lattice rule constructions in QMC and deep learning approximations with near-optimal convergence rates.
• In analytic number theory, Korobov functions also serve as generating functions, underpinning differential recurrences and exponential sum estimates.

Korobov functions are not a single universally fixed object. In approximation theory and quasi-Monte Carlo (QMC), the term most often refers either to elements of a Korobov space—periodic functions with dominating mixed smoothness—or to the associated translation kernel $\kappa_{r,d}$ that generates such a space by convolution. In a separate approximation-theoretic line, it denotes mixed-Sobolev classes on cubes with vanishing boundary traces, such as $X^{r,p}(\Omega)$ or $K_p^k(\Omega)$. Distinct analytic and number-theoretic usages concern Korobov generating functions, Korobov numbers, and Korobov-type exponential sums (Dung et al., 2012, Fang et al., 11 Mar 2025, Kim et al., 2016, Vandehey, 2016).

1. Terminological scope

The phrase “Korobov function” appears in several technical settings.

Usage	Representative definition	Setting
Periodic RKHS element	$f\in H(K_\alpha)$ with Fourier-weighted norm	QMC, lattice rules, tractability
Korobov kernel function	$$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$	Translate approximation on $\mathbb T^d$
Mixed-derivative cube class	$f\in X^{r,p}(\Omega)$ or $K_p^k(\Omega)$, with $f|_{\partial\Omega}=0$	DNN, CNN, quantum approximation
Generating-function / exponential-sum usage	$F(t)=((1+t)^\lambda-1)^{-1}$ or $X^{r,p}(\Omega)$0	Special numbers, differential identities, analytic number theory

In the periodic RKHS literature, Korobov spaces are Fourier-defined Hilbert spaces on $X^{r,p}(\Omega)$1 or $X^{r,p}(\Omega)$2, and smoothness is encoded by algebraic or exponential decay of Fourier coefficients. In the mixed-derivative cube literature, the emphasis shifts to weak mixed partial derivatives up to a prescribed order in each coordinate, together with homogeneous boundary conditions. In the special-function and exponential-sum literature, the term no longer denotes a smoothness class at all, but instead refers to generating functions or rational exponential sums associated with Korobov’s name (Dick et al., 2014, Dung et al., 2012, Li et al., 20 Jan 2025, Vandehey, 2016).

This suggests that “Korobov function” is best interpreted through its ambient framework—RKHS, mixed-Sobolev approximation class, kernel translate, or generating function—rather than through a single invariant definition.

2. Periodic Korobov spaces and the Korobov kernel

In the standard weighted periodic setting, the Korobov space is a reproducing kernel Hilbert space of periodic functions on $X^{r,p}(\Omega)$3 with kernel

$X^{r,p}(\Omega)$4

where

$X^{r,p}(\Omega)$5

The corresponding inner product is expressed through Fourier coefficients,

$X^{r,p}(\Omega)$6

Here $X^{r,p}(\Omega)$7 is the smoothness parameter and $X^{r,p}(\Omega)$8 are product weights controlling the relative importance of variables. For integer $X^{r,p}(\Omega)$9, this corresponds to square-integrable partial mixed derivatives up to order $K_p^k(\Omega)$0 in each variable, and by Parseval’s identity the description extends to all real $K_p^k(\Omega)$1. The threshold $K_p^k(\Omega)$2 is essential: if $K_p^k(\Omega)$3, the space can contain discontinuous functions, so point evaluation is not well behaved and the RKHS framework breaks down (Dick et al., 2014).

A closely related torus formulation uses the Korobov function

$K_p^k(\Omega)$4

with

$K_p^k(\Omega)$5

Because the coefficients factor, $K_p^k(\Omega)$6. The corresponding Korobov space is defined by convolution,

$K_p^k(\Omega)$7

with norm $K_p^k(\Omega)$8. In the Hilbert case $K_p^k(\Omega)$9, $f\in H(K_\alpha)$0, the kernel is translation invariant,

$f\in H(K_\alpha)$1

and is the reproducing kernel of $f\in H(K_\alpha)$2 (Dung et al., 2012).

These two formulations are equivalent in spirit but not identical in notation. One emphasizes an RKHS on $f\in H(K_\alpha)$3 with explicit Fourier weights; the other emphasizes convolution with a specific translation kernel on $f\in H(K_\alpha)$4. Both encode periodic mixed smoothness coordinatewise.

3. Refined smoothness scales: endpoint, logarithmic, and analytic variants

The endpoint regime $f\in H(K_\alpha)$5 is delicate for classical Korobov spaces. To capture smoothness just above this threshold, the $f\in H(K_\alpha)$6-Korobov space replaces algebraic decay by the borderline factor

$f\in H(K_\alpha)$7

with $f\in H(K_\alpha)$8, and defines the kernel

$f\in H(K_\alpha)$9

The parameter $$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$0 ensures summability of the Fourier weights while still allowing regularity weaker than Hölder continuity; a univariate example in the space is uniformly continuous but not Hölder continuous. For rank-1 lattice rules with component-by-component (CBC) generating vector construction, the worst-case integration error satisfies

$$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$1

In the weighted setting, dimension-independent error bounds are obtained under

$$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$2

The same theory carries over to $$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$3-cosine spaces via the tent transform $$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$4 (Dick et al., 2014).

A different refinement replaces algebraic weights by exponential ones. For analytic periodic functions, weighted Korobov spaces use Fourier weights

$$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$5

with $$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$6, $$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$7, and $$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$8. Then

$$\kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d}A_{\mathbf j}\chi_{\mathbf j}(x)$$9

For multivariate $\mathbb T^d$0-approximation, exponential convergence always holds, with optimal rate

$\mathbb T^d$1

and uniform exponential convergence holds iff

$\mathbb T^d$2

Moreover, strong polynomial tractability with uniform exponential convergence holds iff

$\mathbb T^d$3

with exponent bounds

$\mathbb T^d$4

For $\mathbb T^d$5-approximation with exponential weights, the qualitative tractability results are the same for all linear information and standard information, and $\mathbb T^d$6-EC-weak tractability holds iff $\mathbb T^d$7 (Dick et al., 2012, Kritzer et al., 2016).

4. Approximation, tractability, and lattice-based reconstruction

Approximation by translates of the Korobov kernel is a classical problem on $\mathbb T^d$8. For the unit ball $\mathbb T^d$9 in $f\in X^{r,p}(\Omega)$0 and the generator $f\in X^{r,p}(\Omega)$1, sparse-grid Smolyak constructions yield

$f\in X^{r,p}(\Omega)$2

for $f\in X^{r,p}(\Omega)$3, $f\in X^{r,p}(\Omega)$4, $f\in X^{r,p}(\Omega)$5, and also for $f\in X^{r,p}(\Omega)$6, $f\in X^{r,p}(\Omega)$7. In the Hilbert case, the best possible generator cannot improve the dominant algebraic factor: the lower and upper bounds are

$f\in X^{r,p}(\Omega)$8

Thus approximation by translates of $f\in X^{r,p}(\Omega)$9 is near-optimal up to one logarithmic power (Dung et al., 2012).

For weighted Korobov spaces of finite smoothness with product weights, the tractability picture for worst-case $K_p^k(\Omega)$0-approximation is complete. If

$K_p^k(\Omega)$1

then for all continuous linear information, strong polynomial tractability holds iff $K_p^k(\Omega)$2, and the exponent is

$K_p^k(\Omega)$3

For the same information class, quasi-polynomial tractability, uniform weak tractability, and weak tractability are equivalent and hold iff $K_p^k(\Omega)$4. For standard information, the criteria are stricter: weak tractability holds iff

$K_p^k(\Omega)$5

and uniform weak tractability holds iff

$K_p^k(\Omega)$6

In the $K_p^k(\Omega)$7 setting, the theory is less complete; strong polynomial tractability is characterized by $K_p^k(\Omega)$8, polynomial tractability by $K_p^k(\Omega)$9, but for weaker notions there remain gaps between necessary and sufficient conditions (Ebert et al., 2021, Ebert et al., 2022).

Lattice-based algorithms provide constructive high-dimensional approximations. For numerical integration in analytic weighted Korobov spaces, Korobov lattice rules achieve exponential-weak tractability exactly under the sharp condition $f|_{\partial\Omega}=0$0, with complexity bound

$f|_{\partial\Omega}=0$1

thereby supplying an explicit algorithm where earlier results were only existential (Pillichshammer, 2020). For approximation rather than integration, multiple rank-1 lattices extend near-optimal deterministic $f|_{\partial\Omega}=0$2 rates to the low-smoothness regime $f|_{\partial\Omega}=0$3 and to general weights, while random shifts yield nearly optimal worst-case root mean squared $f|_{\partial\Omega}=0$4 error. A sufficient condition for strong polynomial tractability is

$f|_{\partial\Omega}=0$5

(Cai et al., 28 Jan 2026). A different single-lattice method augments one rank-1 lattice by $f|_{\partial\Omega}=0$6 carefully chosen shifts and a least-squares reconstruction within frequency fibers; it achieves the optimal worst-case $f|_{\partial\Omega}=0$7 rate $f|_{\partial\Omega}=0$8 and the optimal randomized $f|_{\partial\Omega}=0$9 rate $F(t)=((1+t)^\lambda-1)^{-1}$0 (Cai et al., 12 Nov 2025).

5. Modern constructive approximators and statistical learning

Recent approximation theory uses Korobov functions as benchmark classes for constructive deep-learning architectures. For $F(t)=((1+t)^\lambda-1)^{-1}$1, deep ReLU networks achieve nearly optimal “super-convergence” rates: $F(t)=((1+t)^\lambda-1)^{-1}$2 and

$F(t)=((1+t)^\lambda-1)^{-1}$3

VC-dimension-based lower bounds show these rates are essentially sharp up to arbitrarily small losses (Yang et al., 2023). Higher-order constructions extend this picture. For $F(t)=((1+t)^\lambda-1)^{-1}$4, deep ReLU CNNs attain

$F(t)=((1+t)^\lambda-1)^{-1}$5

up to constants depending on $F(t)=((1+t)^\lambda-1)^{-1}$6, with depth bounded by

$F(t)=((1+t)^\lambda-1)^{-1}$7

while another constructive theory obtains $F(t)=((1+t)^\lambda-1)^{-1}$8 error of order $F(t)=((1+t)^\lambda-1)^{-1}$9 and $X^{r,p}(\Omega)$00 error of order $X^{r,p}(\Omega)$01 in terms of network width and depth, up to logarithmic factors (Li et al., 20 Jan 2025, Li et al., 14 Jul 2025).

CNN-specific formulations often use the mixed-derivative space

$X^{r,p}(\Omega)$02

In the two-dimensional matrix-input setting with effective dimension $X^{r,p}(\Omega)$03, a multi-channel 2D convolutional architecture with zero padding, ReLU activations, and a fully connected readout constructively approximates $X^{r,p}(\Omega)$04 functions. For sufficiently large $X^{r,p}(\Omega)$05, there exists

$X^{r,p}(\Omega)$06

with

$X^{r,p}(\Omega)$07

such that

$X^{r,p}(\Omega)$08

The construction is near-optimal in the continuous weight selection model (Fang et al., 11 Mar 2025).

Symmetry further changes the approximation regime. For permutation-symmetric Korobov functions in $X^{r,p}(\Omega)$09, symmetric squared ReLU networks achieve the optimal rate

$X^{r,p}(\Omega)$10

and both the convergence rate and the constant prefactor scale at most polynomially with respect to the ambient dimension (Lu et al., 16 Nov 2025).

Quantum approximation gives another constructive realization. Using sparse-grid expansions, quantum signal processing (QSP), and linear combination of unitaries (LCU), quantum circuits approximate Korobov functions by implementing the Chebyshev representation of hat functions and summing tensor-product terms. The resulting circuit has depth

$X^{r,p}(\Omega)$11

and width

$X^{r,p}(\Omega)$12

for the sparse-grid level $X^{r,p}(\Omega)$13 (Aftab et al., 2024).

Statistical estimation also exploits Korobov structure. In weighted Korobov RKHSs, kernel density estimation based on a regularized variational problem yields dimension-independent mean integrated squared error (MISE) bounds; under suitable smoothness assumptions, the rate is arbitrarily close to

$X^{r,p}(\Omega)$14

and, for the smoother RKHS associated with $X^{r,p}(\Omega)$15, arbitrarily close to

$X^{r,p}(\Omega)$16

For nonperiodic densities on $X^{r,p}(\Omega)$17, the periodic scaled Korobov kernel method wraps the density by a modulo operation, estimates the periodic surrogate in a scaled Korobov space, and achieves

$X^{r,p}(\Omega)$18

for densities with smoothness of order $X^{r,p}(\Omega)$19 and exponential decay (Kazashi et al., 2021, Ye et al., 18 Jun 2025). Shallow ReLU approximation on Korobov space also enters classification theory: for $X^{r,p}(\Omega)$20, explicit $X^{r,p}(\Omega)$21 approximation rates are combined with convex $X^{r,p}(\Omega)$22-norm losses and Tsybakov noise conditions to derive excess misclassification bounds (Liu, 26 Sep 2025).

6. Generating functions, Korobov numbers, and Korobov-type exponential sums

A separate analytic tradition uses “Korobov functions” for generating functions associated with Korobov numbers and polynomials. The generating function for the Korobov polynomials of the first kind is

$X^{r,p}(\Omega)$23

and the Korobov numbers are $X^{r,p}(\Omega)$24. Writing

$X^{r,p}(\Omega)$25

one obtains the nonlinear differential identity

$X^{r,p}(\Omega)$26

and more generally

$X^{r,p}(\Omega)$27

with explicitly recurrent coefficients $X^{r,p}(\Omega)$28. In the limit $X^{r,p}(\Omega)$29,

$X^{r,p}(\Omega)$30

which yields closed formulas for derivatives of $X^{r,p}(\Omega)$31 (Kim et al., 2016).

In analytic number theory, Korobov-type exponential sums are

$X^{r,p}(\Omega)$32

with $X^{r,p}(\Omega)$33, typically $X^{r,p}(\Omega)$34. These were described as “rational exponential sums containing an exponential function.” Classical Korobov bounds are nontrivial only for comparatively long sums, but a later differencing method lowers the modulus rather than the algebraic degree. For moduli with prime factors in a fixed finite set, the method yields nontriviality once

$X^{r,p}(\Omega)$35

and proves bounds of the shape

$X^{r,p}(\Omega)$36

Applications include block-frequency asymptotics for the digits of rational numbers and normal-number constructions (Vandehey, 2016).

These usages are technically distinct from the Korobov-space literature. They concern generating functions, differential recurrences, and incomplete exponential sums rather than approximation classes of mixed smoothness. The shared nomenclature therefore reflects lineage rather than a single common formal object.