--- title: Korobov Functions in Approximation & QMC url: https://www.emergentmind.com/topics/korobov-functions type: topic --- # Korobov Functions in Approximation & QMC 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 [1212.6160] [2503.07976] [1604.04328] [1606.07911]. ## 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 $\sum_{n=1}^N e^{2\pi i a b^n/m}$ | Special numbers, differential identities, analytic number theory | In the periodic RKHS literature, Korobov spaces are Fourier-defined Hilbert spaces on $[0,1]^s$ or $\mathbb T^d$, 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 [1411.2715] [1212.6160] [2501.11275] [1606.07911]. 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 $[0,1]^s$ with kernel \[ K_{\alpha}(\boldsymbol{x},\boldsymbol{y}) = \sum_{\boldsymbol{k}\in\mathbb{Z}^s} w_{\alpha,\boldsymbol{\gamma}}(\boldsymbol{k}) \,e^{2\pi i\,\boldsymbol{k}\cdot(\boldsymbol{x}-\boldsymbol{y})}, \qquad w_{\alpha,\boldsymbol{\gamma}}(\boldsymbol{k}) = \prod_{j=1}^s w_{\alpha,\gamma_j}(k_j), \] where \[ w_{\alpha,\gamma_j}(k_j)= \begin{cases} 1, & k_j=0,\\ \gamma_j |k_j|^{-2\alpha}, & k_j\neq 0. \end{cases} \] The corresponding inner product is expressed through Fourier coefficients, \[ \langle f,g\rangle_{H_{K_\alpha}} = \sum_{\boldsymbol{k}\in\mathbb{Z}^s} \widehat f(\boldsymbol{k})\,\overline{\widehat g(\boldsymbol{k})}\, w_{\alpha,\boldsymbol{\gamma}}(\boldsymbol{k})^{-1}. \] Here $\alpha$ is the smoothness parameter and $(\gamma_j)_{j\in\mathbb N}$ are product weights controlling the relative importance of variables. For integer $\alpha$, this corresponds to square-integrable partial mixed derivatives up to order $\alpha$ in each variable, and by Parseval’s identity the description extends to all real $\alpha>1/2$. The threshold $\alpha>1/2$ is essential: if $\alpha\le 1/2$, the space can contain discontinuous functions, so point evaluation is not well behaved and the RKHS framework breaks down [1411.2715]. A closely related torus formulation uses the Korobov function \[ \kappa_{r,d}(x)=\sum_{\mathbf j\in\mathbb Z^d} A_{\mathbf j}\,\chi_{\mathbf j}(x), \qquad A_{\mathbf j}=\prod_{\ell=1}^d a_{j_\ell}, \] with \[ a_j= \begin{cases} |j|^{-r}, & j\in\mathbb Z\setminus\{0\},\\ 1, & j=0. \end{cases} \] Because the coefficients factor, $\kappa_{r,d}(x)=\prod_{\ell=1}^d \kappa_r(x_\ell)$. The corresponding Korobov space is defined by convolution, \[ K_p^r(\mathbb T^d):=\{f:\ f=\kappa_{r,d}*g,\ g\in L_p(\mathbb T^d)\}, \] with norm $\|f\|_{K_p^r(\mathbb T^d)}:=\|g\|_{L_p(\mathbb T^d)}$. In the Hilbert case $p=2$, $r>1/2$, the kernel is translation invariant, \[ K(x,y)=\kappa_{r,d}(x-y), \] and is the reproducing kernel of $K_2^r(\mathbb T^d)$ [1212.6160]. These two formulations are equivalent in spirit but not identical in notation. One emphasizes an RKHS on $[0,1]^s$ with explicit Fourier weights; the other emphasizes convolution with a specific translation kernel on $\mathbb T^d$. Both encode periodic mixed smoothness coordinatewise. ## 3. Refined smoothness scales: endpoint, logarithmic, and analytic variants The endpoint regime $\alpha\downarrow 1/2$ is delicate for classical Korobov spaces. To capture smoothness just above this threshold, the $\log$-Korobov space replaces algebraic decay by the borderline factor \[ r_{\mu,\gamma_j}(k_j)= \begin{cases} 1, & k_j=0,\\[2mm] \gamma_j |k_j|^{-1}\bigl(\log(\kappa |k_j|)\bigr)^{-\mu}, & k_j\neq 0, \end{cases} \] with $\mu>1$, and defines the kernel \[ K_{\log,\mu}(\boldsymbol{x},\boldsymbol{y}) = \sum_{\boldsymbol{k}\in\mathbb Z^s} r_{\mu,\boldsymbol{\gamma}}(\boldsymbol{k})\, e^{2\pi i\,\boldsymbol{k}\cdot(\boldsymbol{x}-\boldsymbol{y})}. \] The parameter $\mu>1$ 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 \[ e(H_{K_{\log,\mu}},P) = \mathcal O\!\left(N^{-1/2}(\log N)^{-\mu(1-\lambda)/2}\right), \qquad \frac1\mu<\lambda\le 1. \] In the weighted setting, dimension-independent error bounds are obtained under \[ \sum_{j=1}^\infty \gamma_j \max\{1,\log \gamma_j^{-1}\}^{\mu(1-\lambda)}<\infty, \qquad \frac1\mu<\lambda\le 1. \] The same theory carries over to $\log$-cosine spaces via the tent transform $\rho(x)=1-|2x-1|$ [1411.2715]. A different refinement replaces algebraic weights by exponential ones. For analytic periodic functions, weighted Korobov spaces use Fourier weights \[ \omega_h = w^{\sum_{j=1}^s a_j |h_j|^{b_j}}, \qquad w\in(0,1), \] with $a_j\ge 1$, $b_j\ge 1$, and $a_1\le a_2\le\cdots$. Then \[ \|f\|_{H(K_{s,\boldsymbol a,\boldsymbol b})}^2 = \sum_{h\in\mathbb Z^s} |\widehat f(h)|^2\,w^{-\sum_{j=1}^s a_j |h_j|^{b_j}}. \] For multivariate $L_2$-approximation, exponential convergence always holds, with optimal rate \[ p^*(s)=\frac{1}{B(s)}, \qquad B(s):=\sum_{j=1}^s \frac1{b_j}, \] and uniform exponential convergence holds iff \[ B=\sum_{j=1}^\infty \frac1{b_j}<\infty. \] Moreover, strong polynomial tractability with uniform exponential convergence holds iff \[ \sum_{j=1}^\infty \frac1{b_j}<\infty \quad\text{and}\quad a^*:=\liminf_{j\to\infty}\frac{\log a_j}{j}>0, \] with exponent bounds \[ \max\!\left(B,\frac{\log 3}{a^*}\right)\le \tau^*1$, and also for $p=2$, $r>1/2$. In the Hilbert case, the best possible generator cannot improve the dominant algebraic factor: the lower and upper bounds are \[ n^{-r}(\log n)^{(d-2)r} < M_n(U_2^r(\mathbb T^d))_2 < n^{-r}(\log n)^{(d-1)r}. \] Thus approximation by translates of $\kappa_{r,d}$ is near-optimal up to one logarithmic power [1212.6160]. For weighted Korobov spaces of finite smoothness with product weights, the tractability picture for worst-case $L_2$-approximation is complete. If \[ s_\gamma:=\inf\left\{\kappa>0:\ \sum_{j=1}^\infty \gamma_j^\kappa<\infty\right\}, \] then for all continuous linear information, strong polynomial tractability holds iff $s_\gamma<\infty$, and the exponent is \[ p^*(\Lambda^{\rm all})=2\max\{s_\gamma,1/a\}. \] For the same information class, quasi-polynomial tractability, uniform weak tractability, and weak tractability are equivalent and hold iff $\inf_{j\ge 1}\gamma_j<1$. For standard information, the criteria are stricter: weak tractability holds iff \[ \lim_{s\to\infty}\frac1s\sum_{j=1}^s \gamma_j=0, \] and uniform weak tractability holds iff \[ \lim_{s\to\infty}\frac{1}{s^\sigma}\sum_{j=1}^s \gamma_j=0 \qquad \forall\,\sigma\in(0,1]. \] In the $L_\infty$ setting, the theory is less complete; strong polynomial tractability is characterized by $s_\gamma<1$, polynomial tractability by $t_\gamma<1$, but for weaker notions there remain gaps between necessary and sufficient conditions [2102.01449] [2201.09940]. 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 $a_j\to\infty$, with complexity bound \[ N(\varepsilon,d)\le 4\,\varepsilon^{-2\lambda} \prod_{j=1}^d \left(1+2A_\lambda w^{\lambda a_j}\right), \qquad \lambda\in(0,1], \] thereby supplying an explicit algorithm where earlier results were only existential [2010.03286]. For approximation rather than integration, multiple rank-1 lattices extend near-optimal deterministic $L_\infty$ rates to the low-smoothness regime $1/2<\alpha\le 1$ and to general weights, while random shifts yield nearly optimal worst-case root mean squared $L_2$ error. A sufficient condition for strong polynomial tractability is \[ \sum_{\substack{u\subset \mathbb N\\ |u|<\infty}} \gamma_u^\lambda\,(2\zeta(\alpha\lambda))^{|u|}<\infty \qquad \text{for some } \lambda\in(1/\alpha,2) \] [2601.20290]. A different single-lattice method augments one rank-1 lattice by $\mathcal O((\log N)^d)$ carefully chosen shifts and a least-squares reconstruction within frequency fibers; it achieves the optimal worst-case $L_\infty$ rate $\mathcal O(N^{-\alpha+1/2+\varepsilon})$ and the optimal randomized $L_2$ rate $\mathcal O(N^{-\alpha+\varepsilon})$ [2511.09071]. ## 5. Modern constructive approximators and statistical learning Recent approximation theory uses Korobov functions as benchmark classes for constructive deep-learning architectures. For $f\in X^{2,\infty}([0,1]^d)$, deep ReLU networks achieve nearly optimal “super-convergence” rates: \[ \|f-k\|_{L_p(\Omega)} \le C\,N^{-4}L^{-4}(\log_2 N)^{d-1}(\log_2 L)^{d-1}, \qquad 1\le p\le\infty, \] and \[ \|f-k\|_{H^1([0,1]^d)} \le C\,N^{-2}L^{-2}(\log_2 N)^{d-1}(\log_2 L)^{d-1}. \] VC-dimension-based lower bounds show these rates are essentially sharp up to arbitrarily small losses [2311.04779]. Higher-order constructions extend this picture. For $f\in K_p^{m+1}(\Omega)$, deep ReLU CNNs attain \[ N^{-m-1}\big(\log_2 N\big)^{(m+2)(d-1)} \] up to constants depending on $m,d$, with depth bounded by \[ L\leq C_s d^4 m^3 N (\log_2 N), \] while another constructive theory obtains $L_p$ error of order $2m$ and $W_p^1$ error of order $2m-2$ in terms of network width and depth, up to logarithmic factors [2501.11275] [2507.10345]. CNN-specific formulations often use the mixed-derivative space \[ X^{r,p}(\Omega):= \Big\{f\in L^p(\Omega): f|_{\partial\Omega}=0,\ \partial^{\boldsymbol\alpha} f\in L^p(\Omega)\ \text{for }|\boldsymbol\alpha|_\infty\le r\Big\}. \] In the two-dimensional matrix-input setting with effective dimension $D=d^2$, a multi-channel 2D convolutional architecture with zero padding, ReLU activations, and a fully connected readout constructively approximates $X^{2,p}(\Omega)$ functions. For sufficiently large $N$, there exists \[ h\in \mathcal H^{W,L}_{2k+1}(\mathbb R^{d\times d}) \] with \[ W=2Nd^2,\qquad L=2(2\lceil\log_2 N\rceil+3)\lceil\log_2 d\rceil+6d, \] such that \[ \|f-h\|_{L^p(\Omega)} \le \frac{4}{2^{(1-\frac1p)d^2}\,(\log_2 N)^{(3-\frac1p)(d^2-1)}\;N^{-(2-\frac1p)}}. \] The construction is near-optimal in the continuous weight selection model [2503.07976]. Symmetry further changes the approximation regime. For permutation-symmetric Korobov functions in $X^{2,2}_{\mathrm{sym}}(\Omega)$, symmetric squared ReLU networks achieve the optimal rate \[ \|f-\phi_{\mathrm{ReLU}_2}\|_E \le C |f|_{2,2} m^{-1}, \] and both the convergence rate and the constant prefactor scale at most polynomially with respect to the ambient dimension [2511.12398]. 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 \[ O\!\left(2^{n+d}n^{d-1}(n+d\log_2 n)\right) \] and width \[ O(n+d\log_2 n) \] for the sparse-grid level $n$ [2404.14570]. 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 \[ M^{-1/(1+1/\alpha)} \] and, for the smoother RKHS associated with $\tilde K_\alpha$, arbitrarily close to \[ M^{-1/(1+1/(2\alpha))}. \] For nonperiodic densities on $\mathbb R^d$, the periodic scaled Korobov kernel method wraps the density by a modulo operation, estimates the periodic surrogate in a scaled Korobov space, and achieves \[ \mathcal O\!\left(M^{-1/(1+1/(2\alpha)+\epsilon)}\right) \] for densities with smoothness of order $\alpha$ and exponential decay [2108.12699] [2506.15419]. Shallow ReLU approximation on Korobov space also enters classification theory: for $X^{2,p}([-1,1]^d)$, explicit $L_p$ approximation rates are combined with convex $\eta$-norm losses and Tsybakov noise conditions to derive excess misclassification bounds [2509.22748]. ## 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 \[ \frac{\lambda t}{(1+t)^{\lambda}-1}e^{xt} = \sum_{n\ge 0} K_n(\lambda,x)\frac{t^n}{n!}, \] and the Korobov numbers are $K_n(\lambda)=K_n(\lambda,0)$. Writing \[ F(t)=\frac{1}{(1+t)^\lambda-1}, \] one obtains the nonlinear differential identity \[ F'=\frac{-\lambda}{1+t}(F+F^2), \] and more generally \[ F^{(N)}=\frac{(-1)^N\lambda}{(1+t)^N}\sum_{i=1}^{N+1}a_{i-1}(N)F^i, \] with explicitly recurrent coefficients $a_j(N)$. In the limit $\lambda\to 0$, \[ \lim_{\lambda\to 0}\lambda F(t)=\frac{1}{\log(1+t)}, \] which yields closed formulas for derivatives of $1/\log(1+t)$ [1604.04328]. In analytic number theory, Korobov-type exponential sums are \[ S(N;a,b,m)=\sum_{n=1}^N e^{2\pi i a b^n/m}, \] with $a,b,m\in\mathbb Z$, typically $\gcd(b,m)=1$. 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 \[ \exp\!\left(\frac{\log m}{\log\log m-3\log\log\log m}\right)\le N, \] and proves bounds of the shape \[ \left|\sum_{n=1}^N e^{2\pi i a b^n/m}\right| \ll N\exp\!\big(-c(\log\log m)^{3/2}\big). \] Applications include block-frequency asymptotics for the digits of rational numbers and normal-number constructions [1606.07911]. 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. Source: https://www.emergentmind.com/topics/korobov-functions