General Nonlocal Maps

• General nonlocal maps are operators whose outputs depend on an extended set of values through kernels, reflections, or memory, rather than local neighborhoods.
• They appear in various contexts such as diffusion processes, fractional dynamics, boundary value problems, and integrable systems.
• Their study leverages operator-theoretic structures to ensure stability, regularity, and, in some cases, exact solvability in complex systems.

General nonlocal maps are operators, dynamical rules, or reductions in which the value produced at a point, state, or time step depends on an extended set of values rather than only on an infinitesimal neighborhood or the same point. Across recent arXiv literature, the phrase encompasses kernel-integral operators, discrete convolution maps, boundary trace operators, history-dependent maps with fractional memory, nonlocal reductions in integrable systems, and order-preserving nonlinear maps on Banach or Hilbert spaces. This breadth suggests that “general nonlocal maps” is best understood as a structural category rather than a single canonical formula: the unifying feature is dependence on globally coupled data through kernels, traces, reflections, or spectral symmetries (Tao et al., 2018, Gesztesy et al., 2010, Tarasov, 22 Sep 2025, Mazowiecka et al., 2017, Okulov, 2019, Garcia-Morales, 2013).

1. Formal archetypes

Several mathematically distinct constructions recur under the name.

Setting	Representative form	Domain of dependence
Nonlocal diffusion / graph operator	$$(\mathcal{L}^h Z)_i=\sum_j \omega(X_i,X_j)(Z_j-Z_i)$$	All sampled positions $j$
Continuum nonlocal operator	$$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$	Entire domain $\Omega$
Boundary trace map	$$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$	Entire boundary $\partial\Omega$
Fractional-memory map	$$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$	Entire past history $\{X_1,\dots,X_n\}$
Optical convolution map	$$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$	Entire transverse field
Global CA map	$\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$	Entire lattice configuration
Integrable nonlocal reduction	$j$0 or $j$1	Reflected space-time point

In neural-network form, the original nonlocal block is

$j$2

while a diffusion-form reformulation uses

$j$3

In fractional dynamics, exact discrete maps arise by evaluating solutions of general fractional integral or derivative equations with periodic kicks at kick times. In boundary-value problems, the Dirichlet-to-Neumann map sends boundary data to normal derivatives and is nonlocal because the value at one boundary point depends on the entire boundary datum. In cellular automata, local, nonlocal, and global maps are distinguished by whether one updates sites, neighborhoods, or whole configurations encoded as a single number (Tao et al., 2018, Tarasov, 22 Sep 2025, Gesztesy et al., 2010, Garcia-Morales, 2013).

2. Shared operator-theoretic structure

A large part of the literature formulates nonlocal maps through kernels. For sampled features $j$4, the affinity kernel defines a matrix $j$5. Under finite Frobenius norm, row normalization, symmetry, and nonnegativity,

$j$6

the resulting operator behaves like a symmetric nonlocal diffusion or a normalized Markov kernel. The discrete operator

$j$7

is mean-preserving and dissipative, and $j$8 is positive semidefinite. Its continuum analogue

$j$9

has the same diffusion interpretation and, with square-integrable kernel, is Hilbert–Schmidt (Tao et al., 2018).

The same structural vocabulary appears in general linear elliptic theory. There one studies

$$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$0

with symmetric kernel $$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$1, Lévy-type integrability, and a lower bound by a radial kernel $$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$2. The associated quadratic form defines an energy space with exterior Dirichlet condition $$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$3 on $$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$4, and a symmetrized comparison operator

$$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$5

is used for Talenti-type mass concentration estimates (Ferone et al., 2023).

Boundary operators furnish another operator-theoretic model. For Schrödinger operators with bounded, possibly non-self-adjoint nonlocal interaction $$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$6, the Dirichlet-to-Neumann map

$$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$7

is nonlocal on $$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$8 because $$\mathcal{L}z(x)=\int_\Omega \rho(x,y)(z(y)-z(x))\,dy$$9 depends on the entire function $\Omega$0 through the solution operator and Green kernel. In smooth settings it is a pseudodifferential operator of order $\Omega$1; in the paper’s abstract Lipschitz setting it is handled through trace maps, resolvents, and Fredholm determinants (Gesztesy et al., 2010).

This variety matters because it rules out a narrow identification of nonlocality with Euclidean convolution alone. In the cited literature, nonlocality also includes boundary-to-boundary maps, pairwise off-diagonal operators, and maps defined by space-time reflection.

3. Memory, iteration, and exact discrete-time maps

One major meaning of general nonlocal maps is temporal memory. In general fractional dynamics, the basic operators are Laplace-convolution forms

$\Omega$2

$\Omega$3

with Sonin or Luchko kernel pairs $\Omega$4. After introducing periodic kicks and using the fundamental theorems of general fractional calculus, one evaluates the exact continuous-time solution at $\Omega$5 and obtains a discrete map whose next state depends on all earlier states through kernel weights. For Caputo-type equations, the resulting exact map is

$\Omega$6

The same framework admits power-law, tempered, Bessel-type, Mittag–Leffler, Kummer, error-function, and power-logarithmic kernels, so the memory law need not be power-law (Tarasov, 22 Sep 2025).

A spatially nonlocal but still iterative example is the doubly nonlocal logistic model

$\Omega$7

whose stationary states satisfy the functional map

$\Omega$8

With step kernels, the analysis yields a uniform solution $\Omega$9, cosine-type “wavelike” patterns, and Gaussian two-crest patterns, together with a linear instability criterion

$$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$0

and thermodynamic-like quantities such as an order parameter $$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$1 and Gibbs entropy $$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$2 for pattern and pattern–pattern transitions (Barbosa et al., 2016).

Cellular automata supply a discrete combinatorial notion of nonlocal map. A local rule updates $$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$3 from a neighborhood of range $$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$4, a nonlocal rule updates encoded neighborhood values $$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$5, and a global rule updates the whole configuration encoded by

$$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$6

The universal nonlocal neighborhood map is

$$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$7

and the global characteristic function $$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$8 permits Diophantine approximation of continuous maps such as the logistic map by global cellular automata (Garcia-Morales, 2013).

Optical resonator theory contributes a continuous-space iterative analogue: $$(M_\Omega(z)f)(x)=\int_{\partial\Omega} M_\Omega(z)(x,y)f(y)\,d\sigma(y)$$9 with kernel $\partial\Omega$0 given by a Green function of diffraction, diffusion, or cavity propagation. In suitable thin-slice limits these maps are equivalent to Ginzburg–Landau-type or nonlinear Schrödinger-type PDEs, yet are iterated in discrete time and implemented efficiently by FFT-based convolution (Okulov, 2019).

4. Boundary maps, fractional geometry, and simple nonlinear maps

In PDE spectral theory, Dirichlet-to-Neumann maps are canonical nonlocal boundary maps. For the Dirichlet problem

$\partial\Omega$1

the DtN map is

$\partial\Omega$2

The paper develops identities such as

$\partial\Omega$3

and the determinant reduction

$\partial\Omega$4

which extends the Jost–Pais formula from one-dimensional local Schrödinger operators to higher-dimensional settings with nonlocal interactions (Gesztesy et al., 2010).

In geometric analysis, nonlocal maps are built from the fractional gradient

$\partial\Omega$5

and the corresponding divergence $\partial\Omega$6, with

$\partial\Omega$7

A fractional div–curl quantity has the form

$\partial\Omega$8

with $\partial\Omega$9. The fractional Coifman–Lions–Meyer–Semmes theorem proves that if $$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$0, $$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$1, and $$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$2, then $$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$3. This yields a nonlocal Wente lemma, continuity of half-harmonic maps into spheres, Hölder regularity for critical systems with nonlocal antisymmetric potentials, and Hölder continuity of $$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$4-harmonic maps into homogeneous targets (Mazowiecka et al., 2017).

A different but related use of “general nonlocal maps” appears in global nonlinear analysis. There, one studies maps

$$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$5

between Banach or Hilbert spaces, where $$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$6 may be nonlocal and even non-variational. Under positivity, convexity-like monotonicity, and principal-eigenvector hypotheses based on Krein–Rutman or ground-state nondegeneracy, these maps are shown to be simple in the sense that no point has more than two preimages; with properness and critical-point assumptions, they become global folds. The admissible linear operators include the Laplacian with several boundary conditions, the quantum harmonic oscillator, the hydrogen atom Hamiltonian, and spectral fractional Laplacians (Calanchi et al., 2023).

Symmetrization theory provides a further structural statement for general linear nonlocal maps. For the Dirichlet problem

$$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$7

the rearranged solution $$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$8 is compared with the solution $$X_{n+1}=X_n+\lambda T M(T)\mathcal{G}(Tn,X_n)+\cdots$$9 of the symmetrized problem involving $\{X_1,\dots,X_n\}$0, $\{X_1,\dots,X_n\}$1, and $\{X_1,\dots,X_n\}$2. The result is a Talenti-type mass concentration comparison,

$\{X_1,\dots,X_n\}$3

together with an energy inequality and a parabolic extension obtained by implicit time discretization (Ferone et al., 2023).

5. Integrable nonlocal reductions and spectral maps

Integrable wave theory uses “nonlocal maps” in yet another precise sense: discrete symmetries that map a field at $\{X_1,\dots,X_n\}$4 to reflected and possibly conjugated values. A general vector nonlinear Schrödinger model is

$\{X_1,\dots,X_n\}$5

with $\{X_1,\dots,X_n\}$6. The one-parameter case $\{X_1,\dots,X_n\}$7 covers local and reverse-space models; the two-parameter case $\{X_1,\dots,X_n\}$8 also includes temporal and space-time nonlocality. The reduction

$\{X_1,\dots,X_n\}$9

implements nonlocality directly at the Lax-pair level. The resulting systems possess PT-invariant formulations, self-phase modulation, cross-phase modulation, multi-wave mixing, and in the integrable $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$0 case a Lax pair and infinitely many conservation laws (Yan, 2016).

For scalar nonlocal NLS equations arising from AKNS reductions, the same Riemann–Hilbert solution generates local and nonlocal solitons, but the symmetry constraints on the scattering data differ. In the reverse-space case $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$1, eigenvalues in $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$2 are paired as $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$3 within the same half-plane; in the reverse-time case $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$4, eigenvalues are paired across half-planes as $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$5; in the reverse-space-time case $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$6, upper- and lower-half-plane eigenvalues are essentially independent while eigenvectors are restricted to $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$7. General $$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$8-solitons in all three equations come from the same determinant formula, but their admissible spectral configurations and dynamics differ. A generic feature is repeated collapsing, and multi-solitons need not be nonlinear superpositions of fundamental solitons (Yang, 2017).

The Dbar approach to coupled nonlocal NLS makes the nonlocal map explicit in spectral space. Two Dbar problems are introduced, with potentials related by

$$E_{n+1}(\mathbf r)=\int K(\mathbf r-\mathbf r')\,f(E_n(\mathbf r'))\,d\mathbf r'$$9

and eigenfunctions related by

$\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$0

This gives a coupled nonlocal NLS system and a general reduction to the scalar nonlocal NLS $\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$1. For arbitrary $\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$2, the reduction is encoded by explicit constraints on purely imaginary spectral points $\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$3, $\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$4 and on the moduli of the Dbar weights: $\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$5 together with

$\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$6

These are explicit general nonlocal reduction conditions at the level of scattering data (Wang et al., 2021).

Reverse-time nonlocal rogue waves exhibit the same phenomenon at the level of Darboux maps. For the reverse-time NLS

$\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$7

the Darboux transformation is constrained by the nonlocal eigenfunction map

$\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$8

The resulting rational rogue waves can be bounded for both $\phi_{t+1}=\chi({}^lR_p^r;\phi_t)$9 and $j$00, or can develop collapsing singularities, depending on free parameters. In the reverse-time nonlocal DS equations, a unified binary Darboux transformation for DSI and DSII yields line rogue waves, exploding cross-shaped multi-rogues, and higher-order patterns such as lump–parabola interactions. Many of these have no counterparts in the local equations (Yang et al., 2017).

6. Neural, optical, and computational realizations

In machine learning, nonlocal maps are used to capture long-range dependencies in feature space. The original nonlocal residual block

$j$01

aggregates information from all positions $j$02. Spectrum analysis of trained models shows that when only $j$03–$j$04 original nonlocal blocks are used, most eigenvalues of the symmetric part of the learned weight matrix are negative, indicating damping; with more blocks, many positive eigenvalues appear, implying potential blow-up and training instability. The diffusion-form reformulation

$j$05

matches the discrete nonlocal diffusion operator, inherits mean preservation and dissipativity, admits a Markov jump-process interpretation, and permits deeper stacking under suitable eigenvalue bounds on $j$06 (Tao et al., 2018).

Optical physics furnishes an experimentally motivated class of nonlocal nonlinear maps: $j$07 Here $j$08 is a Green function for diffraction, dispersion, or resonator propagation, and $j$09 is a local nonlinear point map of Feigenbaum-, Ikeda-, Kerr-, or saturable-gain type. In suitable limits these maps are equivalent to Ginzburg–Landau-type, nonlinear Schrödinger, or Gross–Pitaevskii equations. With resonator kernels they reproduce spatial solitons, vortex eigenmodes, vortex–antivortex lattices, and spatial chaos; smooth multimode noise is used to eliminate numerical artifacts and select stable entities (Okulov, 2019).

A plausible unifying implication is that the practical success or analytical tractability of a nonlocal map usually depends less on the mere presence of long-range coupling than on ancillary structure: symmetry, positivity, normalization, compactness or Hilbert–Schmidt bounds, cone invariance, conservation laws, or explicit spectral constraints. Across neural diffusion stages, Talenti-type symmetrization, fractional div–curl systems, AKNS reductions, and Green-function iteration, those structural conditions determine whether nonlocality yields stability, regularization, fold geometry, exact solvability, or singular collapse (Tao et al., 2018, Ferone et al., 2023, Mazowiecka et al., 2017, Yang, 2017).