Convolution/Integral Memory Kernels

• Convolution/integral memory kernels are functions that mediate nonlocal interactions by encoding hereditary effects via convolution or integral transforms.
• They exhibit key properties such as causality, complete monotonicity, and localized support, enabling robust modeling in fractional PDEs and stochastic processes.
• Efficient numerical strategies, including convolution-product expansions and hierarchical compression, leverage these kernels for scalable simulations in large-scale systems.

A convolution/integral memory kernel is a function that mediates nonlocal interactions in integral operators, mapping an input function or process to an output via convolution or integral transformation, often encoding “memory” or hereditary effects. These structures are central across applied mathematics, analysis, stochastic processes, signal processing, and computational physics, manifesting in integral equations, Volterra and Fredholm operators, time-fractional PDEs, operator compression, and machine learning.

1. Mathematical Structure and Classification

The prototypical integral operator with a memory kernel is

$(H u)(x) = \int_\Omega K(x,y) \, u(y) \, dy$

for $u\in L^2(\Omega)$, where $\Omega$ is a domain in $\mathbb{R}^d$ and $K(x,y)$ is called the kernel. The kernel may be translation-invariant ($K(x,y)=G(x-y)$, convolution kernels), time-variant ($K(x,y)=T(x, x-y)$), or allow for more general spatial, temporal, or even stochastic dependence.

Types of Memory Kernels:

• Volterra (causal) kernels: $K(t,s)$ with support $0\leq s \leq t$.
• Convolution kernels: $k(t-s)$ (time-invariant), in $u\in L^2(\Omega)$0 or other functional spaces.
• Time-varying impulse response (TVIR): $u\in L^2(\Omega)$1, with regularity and compact support in $u\in L^2(\Omega)$2 (Escande et al., 2016).
• Symmetrical Sonin pairs: Two kernels $u\in L^2(\Omega)$3, $u\in L^2(\Omega)$4 connected via integral identities, often of special function form (Luchko, 2023).

Kernels are further classified by smoothness (Sobolev or $u\in L^2(\Omega)$5 regularity), localized support (e.g., impulse response width $u\in L^2(\Omega)$6), decay properties, monotonicity, complete monotonicity, and admissibility in operator compression frameworks.

2. Analytical Properties and Characterizations

Regularity and Support:

• For Hilbert–Schmidt operators: $u\in L^2(\Omega)$7 uniformly in $u\in L^2(\Omega)$8, $u\in L^2(\Omega)$9; compact (or windowed) support in the variable $\Omega$0 ($\Omega$1) ensures efficient representation (Escande et al., 2016).

Causality and Monotonicity:

• Causality: $\Omega$2 for $\Omega$3 (i.e., effect cannot precede cause).
• Complete monotonicity: $\Omega$4 for all $\Omega$5, characterized via Bernstein’s theorem as positive Laplace mixtures of exponentials; equivalent to nonnegativity preservation in convolution outputs (Alfonsi, 2023).

Nonnegativity Preservation:

• A kernel $\Omega$6 is nonnegativity-preserving if the convolution with any sequence of nonnegative inputs yields a nonnegative output. Necessary and sufficient conditions (recursive $\Omega$7 positivity, resolvent monotonicity) are developed, with complete monotonicity sufficing for preservation (Alfonsi, 2023).

Sonin Kernels and Hypergeometric Structure:

• Sonin pairs $\Omega$8 linked by convolution identities exhibit hypergeometric structure:

$\Omega$9

with explicit Laplace transform relationships (Luchko, 2023).

3. Numerical Approximation and Fast Algorithms

Convolution-Product Expansions:

• Given $\mathbb{R}^d$0, approximated in a separable form:

$\mathbb{R}^d$1

leading to efficient operator approximation:

$\mathbb{R}^d$2

where $\mathbb{R}^d$3 is convolution, and $\mathbb{R}^d$4 is pointwise multiplication. The approximation error in Hilbert–Schmidt norm decays as $\mathbb{R}^d$5, with $\mathbb{R}^d$6 the smoothness index of $\mathbb{R}^d$7 (Escande et al., 2016).

Expansion Type	Basis Choice	Optimal Rate
Fourier	$\mathbb{R}^d$8	$\mathbb{R}^d$9
Splines	Degree $K(x,y)$0	$K(x,y)$1
Wavelets	Daubechies, $K(x,y)$2	$K(x,y)$3

Wavelet-based representations yield sparsity and adaptivity to local structure, enabling reduced computational complexity $K(x,y)$4 for discretized grids (Escande et al., 2016).

Hierarchical Compression (H-/H²-matrices, FFM):

• Fast and oblivious convolution quadrature algorithms block-diagonalize the convolution sum via multilevel dyadic partitioning, achieving $K(x,y)$5 or $K(x,y)$6 complexity and $K(x,y)$7 memory (Dölz et al., 2021).
• The Fast Free Memory method (FFM) employs descent-only, octree-based compression with tensorized Chebyshev interpolation and ACA, for $K(x,y)$8 storage and $K(x,y)$9 (non-oscillatory) to $K(x,y)=G(x-y)$0 (oscillatory) cost, scaling to $K(x,y)=G(x-y)$1 (Aussal et al., 2019).

Method	Storage	Computational Cost
Direct	$K(x,y)=G(x-y)$2	$K(x,y)=G(x-y)$3
FFT-based	$K(x,y)=G(x-y)$4	$K(x,y)=G(x-y)$5
FOCQ/H-matrix	$K(x,y)=G(x-y)$6	$K(x,y)=G(x-y)$7
FFM	$K(x,y)=G(x-y)$8	$K(x,y)=G(x-y)$9

GPU kernels for convolution operations in deep networks maximize memory efficiency by shared-memory tiling, double-buffering, and arithmetic intensity metrics, far surpassing generic libraries such as cuDNN for large batches (Chang et al., 2022).

4. Special Kernel Classes and Applications

Symmetrical Sonin and Hypergeometric Kernels:

• Classes based on Wright, Prabhakar, and Horn-type functions encapsulate various physical memory behaviors from simple exponential relaxation to fractional power laws and Gaussian-type decay (Luchko, 2023).

Nonnegativity-Preserving and Completely Monotone Kernels:

• Crucial for stochastic Volterra equations, Hawkes processes, rough volatility, and preservational dynamics; positive mixtures of decaying exponentials are canonical (Alfonsi, 2023).

Learning and Estimation:

• Data-adaptive RKHS methods automatically construct a reproducing kernel tailored to the observable data and the action of the integral/convolution operator, yielding an “automatic” basis and regularization that adapts to operator ill-posedness (Li et al., 16 Jul 2025). This framework outperforms generic ridge or Gaussian process regularizations in inverse kernel learning.

Operator Compression and PDEs:

• Approximation techniques for memory kernels are widely leveraged in large-scale simulations for signal/image processing, spatially varying filtering, boundary integral equations, and time-fractional PDEs. Hierarchical-matrix or FFM-based acceleration is essential in high-fidelity BIE computations and PDE solvers (Aussal et al., 2019).

5. Fundamental Identities and Theoretical Guarantees

Integral and Convolution Identities:

• Sonin kernel pairs satisfy

$K(x,y)=T(x, x-y)$0

providing a rigorous framework for left-inverse or fractional integration/derivation (Luchko, 2023).

• Left-inverse relationships: For sufficiently smooth $K(x,y)=T(x, x-y)$1,

$K(x,y)=T(x, x-y)$2

and

$K(x,y)=T(x, x-y)$3

where $K(x,y)=T(x, x-y)$4 and $K(x,y)=T(x, x-y)$5 denote fractional integration and differentiation.

Error Estimates and Approximation Rates:

• For linear subspace approximations of kernel operators by convolution-product expansions,

$K(x,y)=T(x, x-y)$6

with $K(x,y)=T(x, x-y)$7 the Sobolev regularity of the TVIR, and constants dictated by window width $K(x,y)=T(x, x-y)$8 and subspace choice (Escande et al., 2016).

• For fast matrix-free compression,

$K(x,y)=T(x, x-y)$9

for step size $K(t,s)$0 and hierarchical compression tolerance $K(t,s)$1 (Dölz et al., 2021).

Operator-Theoretic Optimality

No rank-$K(t,s)$2 convolution-product expansion improves the $K(t,s)$3 rate uniformly over kernels of Sobolev regularity $K(t,s)$4, verifying theoretical optimality for a broad class of integral-memory operators (Escande et al., 2016).

6. Connections to Fractional Calculus, Stochastic Analysis, and Nonlocal Models

• Abel and Prabhakar kernels, along with the Riemann–Liouville and Caputo memory kernels, appear as special cases in fractional integration/differentiation theory, encoding long-range decay and anomalous transport (Luchko, 2023).
• Nonnegativity-preserving kernels, notably completely monotone functions, guarantee order preservation and convex invariance for stochastic Volterra equations and rough stochastic differential equations (Alfonsi, 2023).
• Memory kernels underpin nonlocal models in viscoelasticity, anomalous diffusion, mean-field aggregation, nonlocal PDEs, and stochastic finance.

7. Outlook and Research Directions

• Adaptive, kernel-independent fast algorithms continue to scale the efficient simulation and inversion of convolution/integral memory systems to billion-degree problems, as in FFM and H²-based solvers (Aussal et al., 2019, Dölz et al., 2021).
• The systematic construction of data-adaptive reproducing kernel Hilbert spaces for nonparametric kernel estimation opens new directions for learning integral operators directly from observational data, sidestepping manual kernel choice and tuning (Li et al., 16 Jul 2025).
• The synthesis of generalized Sonin/hypergeometric kernels accommodates multi-scale, fractional, and stretched-exponential memory effects in diverse physical and stochastic models (Luchko, 2023).
• Comprehensive characterizations of nonnegativity and monotonicity ensure robust modeling in stochastic and deterministic systems constrained by invariance or positivity.

Convolution/integral memory kernels remain a mathematically rich and computationally essential component in contemporary analysis, scientific computing, and data-driven operator learning.