Kolmogorov-Optimal Approximants

• Kolmogorov-optimal approximants are defined as constructions that achieve the minimal worst-case error equal to the Kolmogorov n-width for a target class and norm.
• They are derived through adaptive partitions, singular value decompositions, or smooth superpositions, facilitating optimal nonlinear and low-rank approximations.
• Applications include spectral theory, quantization, and fractal analysis, with explicit algorithms ensuring both theoretical optimality and practical implementation.

A Kolmogorov-optimal approximant is a constructed object—function, operator, distribution, or matrix—that achieves or attains the minimal possible worst-case approximation error with respect to the Kolmogorov n-width for a given class and norm. This concept permeates nonlinear, adaptive, and probabilistic approximation theory, and connects to explicit algorithms both in function spaces and for discrete structures. Kolmogorov-optimality is quantified asymptotically by the Kolmogorov n-width, which measures the best accuracy achievable by any $n$-dimensional (possibly nonlinear) approximant, and is intimately linked to partitions, multifractal analysis, and, in finite-dimensional settings, to best low-rank approximations.

1. Kolmogorov Widths and the Optimization Principle

Given a compact subset $K$ of a normed space $X$, the Kolmogorov n-width $d_n(K;X)$ is defined as

$$d_n(K;X) = \inf_{\dim S = n} \sup_{f \in K} \inf_{g\in S} \|f - g\|_X,$$

where the infimum is taken over all $n$-dimensional subspaces $S\subset X$. The n-width quantifies the minimal maximal deviation achievable by any $n$-dimensional method, linear or nonlinear, for approximating $K$.

An approximant or family of approximants is called "Kolmogorov-optimal" (or more precisely "Kolmogorov-optimal of order $n$") if it achieves error $d_n(K;X)$; that is, the worst-case error of the method equals the Kolmogorov width. For a given class, this is the gold standard for best-possible approximation rates.

In the matrix case, it is explicitly shown that the sequence of Kolmogorov n-widths for the image of the unit ball under a matrix $A$ coincides with the singular values:

$d_n(A(B_2^m)) = \sigma_{n+1},$

and any best rank-$n$ approximation in the spectral norm is Kolmogorov-optimal (Floater et al., 2020).

2. Adaptive Partitions and Asymptotic Rates

For nonlinear and nonuniform approximation, Kolmogorov-optimality is achieved via adaptive partition schemes. Consider the unit cube $Q \subset \mathbb{R}^d$ and a monotone set function $J: \mathcal{D} \to [0,\infty)$ defined on dyadic subcubes $\mathcal{D}$, representing a "badness" or local approximation fidelity measure.

The construction uses an adaptive subdivision: for $x > 1/J(Q_0)$, define the minimal "x-good" partition $G_x$ by subdividing such that, for each cube $Q\in G_x$,

$J(Q) < 1/x \quad \text{but} \quad J(Q^-) \geq 1/x,$

with $Q^-$ the parent of $Q$. This produces a partition $G_x$ of minimal cardinality $M(x)=|G_x|$ with the property $\sup_{Q\in G_x} J(Q)<1/x$ (Kesseböhmer et al., 2023).

The rate of growth $M(x)$ is governed by the critical zero $s^*$ of the partition function

$$\tau_J(q) = \limsup_{n\to\infty} \frac{1}{n\log2} \log\left(\sum_{Q\in\mathcal{D}_n} J(Q)^q\right),$$

and the associated zeta sum

$Z_J(s) = \sum_{Q\in\mathcal{D}} J(Q)^s.$

Here, $s^* = \inf\{s\ge0: Z_J(s)<\infty\}$, which is equivalently the unique solution to $\tau_J(q)=0$.

Under this scheme, the exact asymptotics are

$\lim_{x\to\infty} \frac{\log M(x)}{\log x} = s^*,$

and, dually, optimal $n$-term partitioning achieves local error decay

$\gamma_n \sim C n^{-1/s^*}, \quad n\to\infty,$

where $$\gamma_n = \min\,\{\max_{Q\in P} J(Q):\, P \text{ partition, } |P|\le n\}$$ (Kesseböhmer et al., 2023).

3. Explicit Construction of Kolmogorov-Optimal Approximants

The adaptive partition $G_x$ serves as the geometric scaffold for constructing Kolmogorov-optimal approximants:

• On each cube $Q\in G_x$, take the best local approximation of the target function $f$—for instance, by local averaging or polynomial projection.
• The global approximant $A_n f$ is assembled by superposing these local approximants.

Formally, letting $n \approx |G_{n^{1/s^*}}|$, one obtains

$\|f - A_n f\| \leq n^{-1/s^*},$

matching the order of the Kolmogorov width $d_n \sim n^{-1/s^*}$ (Kesseböhmer et al., 2023).

In the low-rank matrix context, every orthonormal basis in an $n$-dimensional Kolmogorov-optimal subspace $S$ yields a best rank-$n$ approximation $A_n = P_S A$, with extremal error

$\|A-A_n\|_2 = \sigma_{n+1},$

and the manifold of such optimal subspaces is characterized by spectral-positivity and orthogonality constraints (Floater et al., 2020).

For discrete random variables, the Kolmogorov-optimal $m$-approximant to a variable $X$ is the one minimizing Kolmogorov distance $d_K(X,Y)$ over all $Y$ with support size at most $m$. The construction uses a min-max path algorithm in a DAG built from the support of $X$, ensuring provable optimality with respect to the Kolmogorov metric (Cohen et al., 2018).

4. Applications and Variants

Kolmogorov-optimal approximants appear in numerous domains:

• Spectral theory of differential operators: For singular Sturm–Liouville (Krein–Feller) problems, the set function $J(Q) = \mu(Q)^s |Q|^a$ relates to eigenvalue asymptotics via Weyl-type laws $\lambda_k \sim C k^{s^*}$ (Kesseböhmer et al., 2023).
• Quantization of measures: Setting $J(Q)=\mu(Q)^r$ yields the $r$th-order quantization dimension $D_r = s^*$, governing rates of optimal finite-support approximations (Kesseböhmer et al., 2023).
• Sobolev and Besov embeddings: For $J_{\mu,r,p}(Q) = [\mu(Q)]^{2/p} |Q|^{(2r/d)-1}$, the critical exponent $s^*$ controls the decay $d_n \sim n^{-1/s^*}$ for nonlinear widths (Kesseböhmer et al., 2023).
• Low-rank matrix approximations: The best rank-$n$ approximations with respect to the spectral norm coincide with the Kolmogorov widths and allow flexibility beyond classical SVD, including structured or problem-oriented subspaces (Floater et al., 2020).
• Discretized distributions: In probabilistic modeling and simulation, Kolmogorov-optimal $m$-approximants offer minimal worst-case cdf deviation while reducing support cardinality (Cohen et al., 2018).

5. Smooth Kolmogorov-Type Approximants

The classical Kolmogorov-Arnold representation is exact but exhibits pathological lack of smoothness in its single-variable "inner" functions. Recent work constructs approximate Kolmogorov-Arnold superpositions where all inner and outer functions are $C^2$, resulting in explicit, parallelizable, and fully smooth approximants that nonetheless achieve the Kolmogorov-optimal approximation rate for $\alpha$-Hölder continuous targets:

$\|f-\hat{f}\|_\infty \leq C N^{-\alpha/d},$

matching the Kolmogorov width $w_N(\mathcal{H}^\alpha) \sim N^{-\alpha/d}$. This construction leverages translated/dilated $C^2$ shape functions and row-wise $C^2$ interpolation to maintain smoothness throughout (Song et al., 6 Aug 2025).

6. Fractal and Multifractal Considerations

The asymptotic order of Kolmogorov-optimal approximants is determined by multifractal parameters of the underlying set function $J$. Fractal-geometric quantities, such as $\tau_J(0)$ and $\dim_\infty(J)$, provide explicit upper and lower bounds on the critical exponent $s^*$:

$$\frac{\tau_J(0)}{\tau_J(0)-\tau_J(1)} \leq s^* \leq \frac{[\dim_\infty(J) + \tau_J(1)]}{\dim_\infty(J)} \leq \frac{\tau_J(0)}{\dim_\infty(J)} \leq \frac{d}{\dim_\infty(J)}.$$

This relates the performance of Kolmogorov-optimal approximants to the fine geometric structure of the measure or function being approximated, and connects the theory to entropy and partition zeta functions (Kesseböhmer et al., 2023).

7. Algorithmic and Computational Aspects

Kolmogorov-optimal approximants admit constructive algorithms in both finite and infinite-dimensional settings:

• Discrete random variables: A polynomial-time $O(n^2m)$ algorithm constructs the Kolmogorov-optimal $m$-approximant, outperforming linear programming and heuristic binning (Cohen et al., 2018).
• Low-rank matrix approximations: Iterated projection/orthonormalization algorithms converge to the SVD-optimal subspaces, while the entire manifold of optimal subspaces may be efficiently explored, allowing exploitation of structure and flexibility (Floater et al., 2020).
• Adaptive partitions: The partition algorithms generate near-optimal support and storage cost for piecewise polynomial approximations, applicable in high dimensions and for measures with fractal support (Kesseböhmer et al., 2023).
• Smooth superpositions: Explicit recipes using $C^2$ shape functions yield directly implementable, parallelizable two-layer architectures (Song et al., 6 Aug 2025).

These computational strategies ensure that Kolmogorov-optimality is not a purely existential property but a practical one across diverse settings.