betaFeedback

WorksPhysics and Astronomy

Reconstructing Meson Parton Distributions from Mellin Moments with Artificial Neural Networks

Abstract

We address the inverse problem of reconstructing a meson parton distribution function (PDF) or parton distribution amplitude (PDA) at a fixed scale from a finite set of Mellin moments. At finite moment order this map is ill posed in the Hadamard sense: the moments do not determine a unique, stable pointwise xx-dependence, and small data perturbations can produce large changes in the reconstruction. We regularize the problem by restricting the admissible function class through constraints justified by first principles—support on x∈[0,1]x\in[0,1], endpoint vanishing, unit normalization, and, for the pion, the exact x↔1−xx\leftrightarrow 1-x symmetry—and by complementing these hard constraints with differentiable penalties that suppress the spurious multiple maxima and central concavity changes typical of ill-posed inversions. We parameterize the unknown function with two interchangeable neural cores, a multilayer perceptron (MLP) and a Kolmogorov–Arnold network (KAN), and find that the KAN reconstructs the target distributions more accurately, most markedly in the asymmetric (kaon) case. We attribute this advantage to the locality of the B-spline basis underlying the KAN, which localizes corrections during moment matching and damps the global oscillations to which the MLP is prone. Benchmarks against synthetic targets, using both ten and five input moments, support these conclusions.

Introduction

Parton distribution functions (PDFs) and parton distribution amplitudes (PDAs) encode, respectively, the longitudinal momentum structure of hadrons in inclusive processes and the leading-twist light-cone valence structure relevant to exclusive ones [1, 2, 3]. In several nonperturbative frameworks—most notably lattice QCD and Dyson–Schwinger/Bethe–Salpeter approaches (DSE/BSE), see, e.g., [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]—one does not access the full xx-dependence directly but rather a finite set of Mellin moments at a scale μ\mu,

Mn(μ)≡∫01xnq(x,μ) dx,n=0,1,…,N,(1)M_n(\mu) \equiv\int_0^1 x^n q(x,\mu)\,\mathrm{d}x,\qquad n=0,1,\ldots,N, \tag*{(1)}

where qq denotes generically a valence PDF or a PDA. The goal of this work is to reconstruct q(x,μ)q(x,\mu) on x∈[0,1]x \in[0,1] from {Mn(μ)}n=0∞\{M_n(\mu)\}_{n=0}^{\infty} with as little functional bias as possible.

Reconstructing qq from finitely many moments is an inverse problem of intrinsically ill-defined character. Hadamard well-posedness requires existence, uniqueness, and stability [20, 21, 22]; finite-NN moment inversion generically violates at least two of the three: infinitely many functions share the same first NN moments, and small moment perturbations may induce large variations in the reconstructed xx-profile [23]. In practice, flexible reconstructions tend to develop unphysical oscillations, extra bumps (multiple maxima), or curvature changes near the peak that are not supported by meson-distribution phenomenology, precisely because nothing in the finite moment data itself rules them out.

A natural response is to postulate a rigid parametric form—a Beta-like ansatz with a handful of shape parameters—and fit only those parameters to the moments. This controls the pathology at the cost of introducing a different, harder-to-quantify bias: the reconstruction can only ever look like a member of the assumed family, however inappropriate that family turns out to be for the true q(x)q(x). Entropy-based approaches offer an alternative means of regularizing the reconstruction, see, e.g., [24]. However, as with a prescribed functional ansatz, the resulting distribution is not determined by the moments alone, but also reflects the additional criterion used to select among the admissible solutions. We instead let a flexible neural network stand in for q(x)q(x) and ask how little, rather than how much, needs to be assumed. This is where the choice of network matters: an ordinary multilayer perceptron (MLP), with enough width or depth, can approximate any continuous function on [0,1][0,1] arbitrarily well, a guarantee formalized by the universal approximation theorem [25, 26]. The Kolmogorov–Arnold network (KAN), the second architecture used in this work, rests on an analogous but distinct guarantee, the Kolmogorov–Arnold representation theorem [27], which shows that every continuous multivariate function decomposes exactly into a finite sum of compositions of continuous univariate functions. In neither case is the theorem itself the algorithm we use—practical networks of finite size only approximate the guarantee—but both theorems play the same conceptual role here: they certify that, before any constraint is imposed, the network is in principle capable of representing the true q(x)q(x), whatever its shape. The bias in our reconstruction is therefore not the implicit bias of an insufficiently expressive functional form, but exactly the explicit, physically motivated bias we choose to impose and nothing more. This is the precise sense in which we call the fit unbiased.

That bias is imposed through a short list of constraints, detailed in § 2, which encode what is actually known about q(x)q(x) independently of the moment data: the kinematic support and normalization of the distribution, its behavior at x=0,1x=0,1, the exchange symmetry of the pion, and—on somewhat different, statistical grounds—the expectation that q(x)q(x) has a single maximum and no reversal of curvature near it. We enforce a constraint by hard-wiring it into the network architecture whenever an exact closed-form realization is available, and fall back to a soft, differentiable penalty added to the loss only when it is not. This mirrors a broader distinction in the physics-informed neural network literature [28], where physical laws are typically imposed as soft penalties on the loss (e.g. residuals of a governing PDE); we prefer hard constraints wherever possible, since a property enforced by construction cannot be traded off against the moment-matching term during training and holds exactly for every value of θ\theta. The same combination—a flexible neural core dressed with a physically motivated, hard-wired prefactor—was pioneered in high-energy phenomenology by the NNPDF collaboration’s neural-network parametrization of parton distributions, in which a preprocessing factor xα(1−x)βx^\alpha(1-x)^\beta multiplying the network output controls the Regge and quark-counting endpoint behavior while the network itself remains free to fit everything else [29, 30]; the endpoint prefactors of § 3.3 follow the same logic. Neural-network parametrizations of this kind have since been extended beyond collinear PDFs to generalized parton distributions (GPDs), where the same nonparametric, bias-minimizing philosophy has been used to model the additional skewness and momentum-transfer dependence while respecting polynomiality and positivity [31, 32], and, most recently, to learn parton distributions directly from lattice-QCD matrix elements within a single neural-network framework spanning collinear PDFs and zero-skewness GPDs [33]. § 2 specifies and justifies each constraint; § 3 introduces the two network cores used to realize them, together with the elements of spline theory needed to follow the KAN construction (collected in Appendix § 7). We find that the KAN systematically improves the reconstructions, especially for asymmetric distributions such as the kaon’s, and we argue in § 3 that the locality of its B-spline activations is the likely reason.

Physical constraints

Define the forward operator mapping a function to its first NN moments,

A:q↦M∈RN,(Aq)n=∫01xnq(x) dx,n=1,…,N.(2)\mathcal{A}:q\mapsto M\in\mathbb{R}^{N},\qquad(\mathcal{A}q)_n=\int_{0}^{1}x^nq(x)\,\mathrm{d}x,\quad n=1,\ldots,N. \tag*{(2)}

At finite NN the operator A\mathcal{A} has an infinite-dimensional nullspace: there exist infinitely many δq≠0\delta q\ne0 with Aδq=0\mathcal{A}\delta q=0, so the reconstruction is non-unique, and it is unstable even after non-uniqueness is reduced. Restricting the admissible function class through properties that hold independently of the moment data is the standard remedy, and it is most useful when each restriction can be tied as directly as possible to the physics of the quantity being reconstructed rather than to convenience.

The variable xx is a light-cone momentum fraction and is kinematically confined to [0,1][0,1]. The unit-area condition

∫01q(x) dx=1(3)\int_{0}^{1}q(x)\,\mathrm{d}x=1 \tag*{(3)}

has a direct meaning in both settings we consider: a leading-twist PDA is normalized to the meson decay constant, i.e. to unity once the constant is factored out [2, 3], while a single-flavor valence PDF integrates to the valence-quark number, again unity for a meson.

A second constraint concerns the behavior at the two endpoints of this same interval. For a leading-twist PDA the vanishing q(0)=q(1)=0q(0)=q(1)=0 is rigorous: the asymptotic distribution is φas(x)=6x(1−x)\varphi_{\mathrm{as}}(x)=6x(1-x), and corrections retain the linear endpoint suppression [2, 3]. For a valence-quark PDF, the endpoint condition q(1)=0q(1)=0 is likewise required [34, 35, 36]: x=1x=1 corresponds to the struck quark carrying the entire meson momentum, leaving the remaining partons with zero momentum; consequently, the corresponding endpoint configuration has vanishing probability. Consistent with this behavior, at resolution scales comparable to or below the proton mass, quark-counting rules predict the large-xx behavior q(x)∼(1−x)βq(x)\sim(1-x)^\beta, with β=2\beta=2 in the partonic limit of QCD [34, 35, 36]. As the resolution scale increases, QCD evolution leads to a larger value of β\beta [37, 38, 39, 40]. This pattern is supported by lattice QCD, data-driven methods, continuum approaches, and symmetry-driven studies (see, e.g., Refs. [6, 7, 8, 9, 10, 11, 12, 13, 14, 24, 41, 42, 43, 44, 45]). The behavior at x→0x\to0 is more delicate: Regge phenomenology gives q(x)∼x−αq(x)\sim x^{-\alpha} for valence distributions [46], which need not vanish—it can rise while remaining integrable (α<1\alpha<1)—and the constraint fails outright for sea and gluon distributions, which we do not consider here. However, we choose to work at the so-called hadron scale, at which the fully dressed valence quarks constitute the sole degrees of freedom, encoding all properties of the hadron, and hence share the entire hadron momentum; see, e.g., Refs. [6, 7, 8, 9, 10, 11, 12, 13, 14]. Then, at this scale, the two-body valence structure of the meson relates the quark q(x)q(x) and antiquark hˉ(x)\bar{h}(x) distributions as:

hˉ(x)=q(1−x).(4)\bar{h}(x)=q(1-x). \tag*{(4)}

Therefore, the large-xx behavior of the quark distribution determines the small-xx behavior of the antiquark distribution, and vice versa. The endpoint condition q(1)=0q(1)=0 thus implies hˉ(0)=0\bar{h}(0)=0, while hˉ(1)=0\bar{h}(1)=0 sets q(0)=0q(0)=0. In other words, at the hadron scale q(0)=q(1)=0q(0)=q(1)=0 is also a rigorous constraint for a valence PDF, as x=0x=0 also corresponds to a configuration of vanishing probability.

Furthermore, for those mesonic systems made of a valence quark and antiquark of the same flavor (h=qh=q), such as quarkonia, or specifically for the pion in the light-flavor sector and isospin-symmetric limit, one is left with a valence PDF at the hadron scale which obeys

q(x)=q(1−x),(5)q(x)=q(1-x), \tag*{(5)}

implying as a third constraint the exchange symmetry under x↔1−xx\leftrightarrow1-x, which reflects the equal sharing of momentum between the valence quark and antiquark of equal mass. For the PDA the same follows from charge conjugation and G-parity, which force all odd Gegenbauer (equivalently, odd-about-x=12x=\frac{1}{2} Mellin) moments to vanish [2, 16]. Notably, the above symmetry prescription, together with the approximately monopole behavior of the pion electromagnetic form factor, is sufficient to generate a smooth endpoint behavior for the pion valence PDF [45]. The symmetry is broken for the kaon by SU(3) flavor breaking: the heavier strange quark carries more momentum on average, producing a skewed distribution with a nonzero first Gegenbauer moment a1K≠0a_1^K\ne0 [16, 47]. Taken together, these first three constraints are generically justified for PDAs (the exchange symmetry, of course, only for the pion and quarkonia), whereas for PDFs some of them—particularly the vanishing at x→0x\to0 and the exchange symmetry—should be regarded as a consequence of evaluating the distribution at the hadron scale. In the pion case, an exact isospin-symmetric limit is also required.

A further constraint, not fixed by symmetry or endpoint behavior but no less physically motivated for mesonic ground states, concerns the overall shape of q(x)q(x) in between: we expect it to rise monotonically from the origin to a single maximum and fall monotonically thereafter, without additional bumps, and without a reversal of curvature (an inflection point) in the region around that maximum; see, e.g., Refs. [4, 7, 48]. Unlike the constraints above, this one is not a matter of exact kinematics or an exact symmetry, and it is worth spelling out on what grounds it can be expected. For the leading-twist PDA the argument is close to a theorem. The scale dependence of a PDA is governed by the ERBL evolution equation [49, 50, 51], whose eigenfunctions are the Gegenbauer polynomials Cn3/2(1−2x)C_n^{3/2}(1-2x) built on the same x(1−x)x(1-x) weight as the asymptotic form φas(x)\varphi_{\rm as}(x) above, with eigenvalues (one-loop anomalous dimensions) that grow monotonically with nn [2, 3, 47]. Every Gegenbauer moment beyond the lowest is therefore driven toward zero under evolution, increasingly so for higher nn, so that the PDA is pulled toward the single-humped asymptotic shape and any admixture capable of generating a secondary maximum or a central inflection is perturbatively suppressed—consistent with the small, rapidly decreasing Gegenbauer coefficients found in modern lattice and continuum extractions [5, 16]. Early QCD sum-rule analyses pointed in the opposite direction: Chernyak and Zhitnitsky proposed a double-humped pion PDA, φCZ(x)=30x(1−x)(1−2x)2\varphi_{\rm CZ}(x)=30x(1-x)(1-2x)^2, corresponding to a large second Gegenbauer coefficient a2=2/3a_2=2/3 [47]; this form is disfavored by the later lattice and continuum results just cited. A second, more general argument covers the valence-PDF interpretation as well, for which no analogous evolution equation singles out a preferred basis: a smooth, single-humped shape is the generic, least-structured solution to a moment-constrained reconstruction problem, which is the content of the maximum-entropy principle [52]. Under that principle, the function compatible with a finite set of moments that adds no structure beyond what the data require is the maximum-entropy density, which for a low-order moment constraint is itself unimodal—matching only a mean and a variance, for instance, already selects a Gaussian-like single-humped profile. The same logic underlies the maximum-entropy method used to reconstruct hadronic spectral functions from a finite, noisy set of Euclidean-correlator moments in lattice QCD [53], where spurious multi-peak structures are likewise recognized as reconstruction artifacts rather than genuine features. We stress that, away from the PDA case, this is a physically motivated modeling choice rather than a derived result, consistent with— but not proved by—the phenomenology of single-flavor valence distributions; § 4 turns this expectation into two explicit, differentiable penalties.

Network architectures

We construct q(x)q(x) from a univariate network core f(x;θ)f(x;\theta), evaluated pointwise on x∈[0,1]x \in[0,1], and then wrap it with the algebraic operations of § 3.3 that enforce the hard constraints of § 2 exactly. Two functionally interchangeable cores are used and compared: a multilayer perceptron (MLP) and a Kolmogorov–Arnold network (KAN). As discussed in § 1, both are, in the appropriate limit, universal representations of a one-dimensional continuous function; they differ in where their learnable flexibility lives, on the nodes for the MLP and on the edges for the KAN, and this difference is what § 3.2 argues is responsible for the KAN’s better behavior under moment matching.

Multilayer perceptron

A feedforward MLP builds f(x;θ)f(x;\theta) by alternating affine maps W(z)=Az+bW(z)=Az+b with a fixed, elementwise nonlinearity σ\sigma: the input is combined linearly into a layer of units, each unit is passed through σ\sigma, and the resulting vector becomes the input to the next layer. With LL layers,

fMLP(x;θ)=(W(L)∘σ∘W(L−1)∘⋯∘σ∘W(1))(x),(6)f_{\mathrm{MLP}}(x;\theta)=\left(W^{(L)}\circ\sigma\circ W^{(L-1)}\circ\cdots\circ\sigma\circ W^{(1)}\right)(x), \tag*{(6)}

which for a single hidden layer of width HH reduces to the familiar one-layer form f=b(2)+∑j=1Hwj(2)σELU(wj(1)x+bj(1))f=b^{(2)}+\sum_{j=1}^{H}w_j^{(2)}\sigma_{\mathrm{ELU}}(w_j^{(1)}x+b_j^{(1)}). We use the ELU activation,

σELU(z)={z,z≥0,λ(ez−1),z<0,(7)\sigma_{\mathrm{ELU}}(z)= \begin{cases} z, & z\ge0,\\ \lambda\left(e^z-1\right), & z<0, \end{cases} \tag*{(7)}

with fixed λ>0\lambda>0, and do not impose positivity on the output. The parameters θ\theta are the collection of all weights A(ℓ)A^{(\ell)} and biases b(ℓ)b^{(\ell)}; the nonlinearity σ\sigma itself is never trained, so all of the network’s flexibility resides in how these linear weights combine and recombine a fixed nonlinear building block.

Kolmogorov–Arnold network

The KAN [54] departs from this picture by placing a trainable nonlinearity on every edge of the network, in the form of a spline, and summing only at the nodes. We use here only the property of splines that matters for the construction and for the argument below—that a spline can be written as a linear combination of compactly supported basis functions, so that a single coefficient only ever affects the curve locally—and refer to Appendix § 7 for the explicit definitions.

Formally, a KAN layer with ninn_{\mathrm{in}} inputs and noutn_{\mathrm{out}} outputs is a matrix of learnable univariate functions Φℓ={ϕℓ,j,i}\Phi_{\ell}=\{\phi_{\ell,j,i}\}, one per edge, and each node value is simply the sum of its already-transformed incoming signals,

xℓ+1,j=∑i=1ninϕℓ,j,i(xℓ,i),fKAN(x;θ)=(ΦL−1∘ΦL−2∘⋯∘Φ0)(x).(8)x_{\ell+1,j}=\sum_{i=1}^{n_{\mathrm{in}}}\phi_{\ell,j,i}(x_{\ell,i}),\qquad f_{\mathrm{KAN}}(x;\theta)=\left(\Phi_{L-1}\circ\Phi_{L-2}\circ\cdots\circ\Phi_{0}\right)(x). \tag*{(8)}

This construction is motivated by the Kolmogorov–Arnold representation theorem [27], which guarantees that any continuous f:[0,1]n→Rf:[0,1]^n\to\mathbb{R} can be written through univariate functions and addition alone,

f(x1,…,xn)=∑q=12n+1Φq(∑p=1nϕq,p(xp)),(9)f(x_1,\ldots,x_n)=\sum_{q=1}^{2n+1}\Phi_q\left(\sum_{p=1}^{n}\phi_{q,p}(x_p)\right), \tag*{(9)}

in the same way that the universal approximation theorem motivates Eq. (6): a KAN of depth two and the prescribed width could in principle realize Eq. (9) exactly, and deeper, wider KANs are used in practice for greater flexibility. Each edge function ϕ\phi combines a fixed residual activation with a trainable spline built from the B-spline basis {Bi(z)}\{B_i(z)\} of Appendix § 7,

ϕ(z)=wbSiLU⁡(z)+ws∑iciBi(z),SiLU⁡(z)=z1+e−z,(10)\phi(z) = w_b \operatorname{SiLU}(z) + w_s \sum_i c_i B_i(z), \qquad\operatorname{SiLU}(z) = \frac{z}{1 + e^{-z}}, \tag*{(10)}

where SiLU⁡\operatorname{SiLU} is a single fixed function, shared by every edge and nonzero across the whole domain, while each BiB_i is one element of a fixed local basis, nonzero only over a small window of zz (Appendix § 7); the two terms of Eq. (10) therefore play different roles even though both come with a trainable weight. The spline order and knot grid underlying {Bi}\{B_i\} are fixed, and the trainable parameters θ\theta on every edge are the coefficients cic_i together with the two scalars wbw_b, wsw_s. The architectural contrast with the MLP is now precise: the MLP fixes a nonlinearity on its nodes and learns only linear weights on its edges, while the KAN learns the nonlinearity itself, edge by edge, and only sums on its nodes; for the present one-dimensional problem the network reduces, layer by layer, to a composition of learnable spline curves.

The compact support of the B-spline basis means that fitting a KAN to the moments adjusts f(x;θ)f(x;\theta) locally in xx: a gradient step that improves the fit in one region of the domain leaves the spline unchanged in regions far from it. The MLP has no analogous structure—every weight enters activations with global support in xx—so a gradient step aimed at correcting the fit in one region can perturb it everywhere, which is precisely the mechanism that seeds the domain-spanning spurious oscillations characteristic of ill-posed moment inversion. This picture is consistent with our empirical finding that the KAN’s advantage over the MLP is most visible for the asymmetric kaon distribution, whose peak is displaced from x=12x=\frac{1}{2} and whose reconstruction therefore benefits most from being able to sculpt one side of the domain without disturbing the other.

Hard-constraint wrappers

The constraints of § 2 are imposed on the network output, independently of the core. For the symmetric case (in the following, we consider qπq_\pi for a pion PDF as a paradigmatic symmetric case, although the same applies generically to any quarkonium distribution), we symmetrize and subtract the endpoint value,

fsym(x)=12(f(x)+f(1−x)),q~π(x)=fsym(x)−fsym(0),(11)f_{\mathrm{sym}}(x) = \frac{1}{2}\bigl(f(x) + f(1-x)\bigr), \qquad\widetilde{q}_\pi(x) = f_{\mathrm{sym}}(x) - f_{\mathrm{sym}}(0), \tag*{(11)}

which guarantees q~π(0)=q~π(1)=0\widetilde{q}_\pi(0) = \widetilde{q}_\pi(1) = 0 and exact q~π(x)=q~π(1−x)\widetilde{q}_\pi(x) = \widetilde{q}_\pi(1-x) without an explicit prefactor. For the asymmetric case (here specialized to the kaon), we multiply by a fixed endpoint prefactor,

q~K(x)=x(1−x)f(x),(12)\widetilde{q}_K(x) = x(1-x)f(x), \tag*{(12)}

matching the linear PDA endpoint suppression; the prefactor can be generalized to a (learnable) xα(1−x)βx^\alpha(1-x)^\beta to accommodate the Brodsky–Farrar exponent β=2\beta= 2 of valence PDFs, in the same spirit as the NNPDF preprocessing factor mentioned in § 1. In both cases the final distribution is normalized by construction,

q(x)=q~(x)∫01q~(t) dt.(13)q(x) = \frac{\widetilde{q}(x)}{\int_0^1 \widetilde{q}(t)\,\mathrm{d}t}. \tag*{(13)}

Loss function and shape regularization

Neither core, as defined above, is ever asked to match the training data directly. The data are the moments {Mntr}\{M_n^{\mathrm{tr}}\}—integrals of xnq(x)x^n q(x) over [0,1][0,1]—while the network only ever produces a pointwise value f(x;θ)f(x;\theta), and hence q(x;θ)q(x;\theta) after the wrapping of § 3.3, at whatever xx it is evaluated. Turning these pointwise values into moments therefore requires a numerical integration step, and this step has to happen before every gradient-descent update, not just once at the end of training: at each iteration we evaluate q(xi;θ)q(x_i;\theta) on a fixed mesh {xi}i=0M\{x_i\}_{i=0}^{M} covering [0,1][0,1] and combine these samples with fixed quadrature weights into estimates Mnpred(θ)M_n^{\mathrm{pred}}(\theta) of the moments (and, by the same construction, of the normalization integral of Eq. (13)); only these estimates, not the pointwise values themselves, enter the loss below5. Because the mesh and quadrature weights are fixed, this step is simply a fixed linear functional of the sampled network outputs, and therefore inherits the differentiability of q(x;θ)q(x;\theta) in θ\theta: automatic differentiation propagates gradients through the quadrature exactly as through any other layer of the computation, so ∂θMnpred\partial_\theta M_n^{\mathrm{pred}} is available directly from the same graph used for f(x;θ)f(x;\theta), and no special-purpose implicit-differentiation machinery is needed. In practice we use the composite Boole’s rule, the five-point closed Newton–Cotes quadrature [55], on a uniform mesh fine enough that its truncation error is negligible against the resolution already lost by working with a finite moment set; we do not reproduce the quadrature formula here, as it plays a purely technical role.

Given target moments {Mntar}\{M_n^{\mathrm{tar}}\}, we match the moment estimates just described with a weighted least-squares term,

Lmom(θ)=∑nλn(Mnpred(θ)−Mntar)2,(14)\mathcal{L}_{\mathrm{mom}}(\theta) = \sum_n \lambda_n \left(M_n^{\mathrm{pred}}(\theta)-M_n^{\mathrm{tar}}\right)^2, \tag*{(14)}

where the weights λn\lambda_n are unity except for the first one or two orders, which are enhanced (to ∼10\sim10) because they fix the global location and width of the profile. The n=0n=0 moment is already enforced by Eq. (13) and is used only as a consistency check.

Because the inversion is ill posed, moment matching alone does not guarantee the single-maximum, no-central-inflection behavior motivated in § 2: in exploratory runs it produced reconstructions with a second, spurious maximum, or with a reversal of curvature near the dominant peak. We enforce the preference argued for in § 2 with two differentiable penalties on the sampled values qi=q(xi)q_i=q(x_i), evaluated on the same mesh used for quadrature. The first enforces an “increase-then-decrease” (unimodal) profile by penalizing the wrong sign of the discrete slope Δqi=qi+1−qi\Delta q_i=q_{i+1}-q_i on each side of the global-maximum index kk,

Luni=1max⁡(1,k)∑i<kReLU⁡(−Δqi)2+1max⁡(1,M−k)∑i≥kReLU⁡(Δqi)2.(15)\mathcal{L}_{\mathrm{uni}} = \frac{1}{\max(1,k)}\sum_{i<k}\operatorname{ReLU}(-\Delta q_i)^2 + \frac{1}{\max(1,M-k)}\sum_{i\geq k}\operatorname{ReLU}(\Delta q_i)^2. \tag*{(15)}

The second forbids inflection points in a central window around the peak (while leaving the tails free) by penalizing positive curvature there,

Lconc=∑i:qi≥τqmax⁡ReLU⁡(q′′(xi))2,(16)\mathcal{L}_{\mathrm{conc}} = \sum_{i:q_i\geq\tau q_{\max}}\operatorname{ReLU}(q''(x_i))^2, \tag*{(16)}

with τ∈(0,1)\tau\in(0,1) defining the central region by height. The full objective is

L(θ)=Lmom+λ2L2+λuniLuni+λconcLconc,(17)\mathcal{L}(\theta)=\mathcal{L}_{\mathrm{mom}}+\lambda_2\mathcal{L}_2+\lambda_{\mathrm{uni}}\mathcal{L}_{\mathrm{uni}}+\lambda_{\mathrm{conc}}\mathcal{L}_{\mathrm{conc}}, \tag*{(17)}

where L2=∑ℓ∥A(ℓ)∥F2+∑ℓ∥b(ℓ)∥22\mathcal{L}_2=\sum_\ell\|A^{(\ell)}\|_F^2+\sum_\ell\|b^{(\ell)}\|_2^2 is an L2L_2 penalty on every trainable parameter of the core, which stabilizes training; for the KAN, the analogous sum runs over the base weights, spline coefficients, and spline scalers of every edge, in place of A(ℓ)A^{(\ell)} and b(ℓ)b^{(\ell)}.

For the KAN alone we add one further term, absent for the MLP, which has no counterpart in its parameterization. Writing cji,k(ℓ)c_{ji,k}^{(\ell)} for the kk-th B-spline coefficient (k=1,…,Kk = 1, \ldots, K) of the edge from input ii to output jj in layer ℓ\ell (§ 3.2), define its mean absolute size cˉji(ℓ)=1K∑k∣cji,k(ℓ)∣\bar{c}_{ji}^{(\ell)} = \frac{1}{K} \sum_k \lvert c_{ji,k}^{(\ell)} \rvert and its normalized weight within the layer, pji(ℓ)=cˉji(ℓ)/∑i′,j′cˉj′i′(ℓ)p_{ji}^{(\ell)} = \bar{c}_{ji}^{(\ell)} / \sum_{i',j'} \bar{c}_{j'i'}^{(\ell)}. The KAN loss includes an additional activation-sparsification penalty [54],

Lspline=∑ℓ[∑i,jcˉji(ℓ)−∑i,jpji(ℓ)log⁡pji(ℓ)],(18)\mathcal{L}_{\mathrm{spline}} = \sum_{\ell} \left[ \sum_{i,j} \bar{c}_{ji}^{(\ell)} - \sum_{i,j} p_{ji}^{(\ell)} \log p_{ji}^{(\ell)} \right], \tag*{(18)}

added to the total loss of Eq. (17) with the same coefficient λ2\lambda_2 as L2L_2. The first (entropy-free) term is an L1L_1 penalty that discourages large spline coefficients unless they materially improve the fit; the second is the Shannon entropy of the normalized edge magnitudes, which discourages spreading activation evenly across many edges. Together they push the trained KAN toward a sparse, concentrated set of active edges rather than a diffuse one—a standard KAN regularizer that has no analogue for the MLP, whose parameters carry no such edge-wise structure to sparsify.

A useful heuristic for the hyperparameters λ2\lambda_2, λuni\lambda_{\mathrm{uni}}, λconc\lambda_{\mathrm{conc}} is to choose them so that the loss contributions are of comparable magnitude in the first training epochs.

Results

We benchmark the method on synthetic targets for which the exact moments and the pointwise function are known [12], so that both moment reproduction and pointwise fidelity can be assessed. To quantify the variability induced by random initialization, each reconstruction is reported as the mean over nensn_{\mathrm{ens}} independent trainings, with a band given by the pointwise standard deviation.

Symmetric benchmark (pion-like)

As a symmetric target we use the smooth function

qtest(x)=Nlog⁡(1+x2(1−x)2ρ2),ρ=0.3,(19)q_{\mathrm{test}}(x) = \mathcal{N} \log\left(1 + \frac{x^2(1-x)^2}{\rho^2}\right), \qquad\rho= 0.3, \tag*{(19)}

with N\mathcal{N} fixed by unit normalization. The architecture enforces q(x)=q(1−x)q(x) = q(1-x), endpoint vanishing, and normalization, and is trained on the first N=10N = 10 moments. We compare the two cores of § 3 under identical constraints and loss: the MLP has a single hidden layer of 150 neurons, while the KAN has two hidden layers of 16 nodes each, with spline order 3 and grid size 10; both are evaluated on a mesh of 301 points. Each core is trained independently nens=30n_{\mathrm{ens}} = 30 times, and the reconstruction is reported as the ensemble mean, with a band (figures) or parenthetical uncertainty (tables) given by the pointwise standard deviation. Table 1 compares the exact and reconstructed moments for both cores, and Figure 1 shows the two pointwise reconstructions side by side. We note a practical training detail: the Adam learning rate had to be initialized at 10−510^{-5} for the first few hundred epochs and then increased toward 10−310^{-3} to obtain stable convergence.

Table 1: First ten Mellin moments for the symmetric benchmark, Eq. (19). MntrueM_n^{\mathrm{true}} are the exact moments; M‾npred\overline{M}_n^{\mathrm{pred}} and εn=∣M‾npred−Mntrue∣/∣Mntrue∣\varepsilon_n = \left|\overline{M}_n^{\mathrm{pred}} - M_n^{\mathrm{true}}\right|/\left|M_n^{\mathrm{true}}\right| are the ensemble-averaged prediction and relative error, for the MLP and the KAN (both nens=30n_{\mathrm{ens}} = 30).

nnMntrueM_n^{\mathrm{true}}MLPKAN
M‾npred\overline{M}_n^{\mathrm{pred}}εn\varepsilon_nM‾npred\overline{M}_n^{\mathrm{pred}}εn\varepsilon_n
10.50.49999977×10−77 \times10^{-7}0.50000000(1)5×10−95 \times10^{-9}
20.28840.2883(1)3×10−43 \times10^{-4}0.28838(6)5×10−55 \times10^{-5}
30.18260.1825(2)7×10−47 \times10^{-4}0.18257(8)1×10−41 \times10^{-4}
40.12340.1233(1)7×10−47 \times10^{-4}0.12341(6)2×10−42 \times10^{-4}
50.08760.08756(2)2×10−42 \times10^{-4}0.08756(3)2×10−42 \times10^{-4}
60.06450.06457(9)9×10−49 \times10^{-4}0.06450(5)1×10−41 \times10^{-4}
70.04890.0491(2)3×10−33 \times10^{-3}0.04895(9)4×10−54 \times10^{-5}
80.03810.0383(3)5×10−35 \times10^{-3}0.0381(1)5×10−45 \times10^{-4}
90.03020.0305(3)9×10−39 \times10^{-3}0.0302(2)1×10−31 \times10^{-3}
100.02440.0247(4)1×10−21 \times10^{-2}0.0244(2)2×10−32 \times10^{-3}

Table 1.

Symmetric benchmark reconstructions by MLP and KAN

Figure 1. Symmetric benchmark, Eq. (19) with ρ=0.3\rho= 0.3, trained on the first ten moments: MLP (left) and KAN (right) reconstructions, each shown as the ensemble mean (nens=30n_{\mathrm{ens}} = 30) with a one-standard-deviation band.

Asymmetric benchmark (kaon-like): MLP vs. KAN

For the asymmetric target we use a Gegenbauer expansion built on the x2(1−x)2x^2(1-x)^2 weight, which is the natural orthogonality weight of the Cj5/2C_j^{5/2} polynomials and encodes the Brodsky–Farrar endpoint behavior,

qtest(x)=Nx2(1−x)2(1+∑j=1JajCj5/2(1−2x)),(20)q_{\mathrm{test}}(x) = \mathcal{N} x^2(1-x)^2\left(1+\sum_{j=1}^{J}a_j C_j^{5/2}(1-2x)\right), \tag*{(20)}

where odd aja_j generate the asymmetry expected from SU(3) breaking. (This is a valence-PDF-motivated construction, chosen for its endpoint behavior and convenient orthogonal basis; it should not be confused with the Cn3/2C_n^{3/2} ERBL eigenbasis of Sec. 2, which applies to the PDA.) We consider two instances of Eq. (20), both fit from the first N=10N = 10 moments: a two-term expansion with (a1,a2)=(0.1,0.02)(a_1,a_2) = (0.1,0.02), and a four-term expansion with (a1,a2,a3,a4)=(−0.08,0.1,−0.01,0.013)(a_1,a_2,a_3,a_4) = (-0.08,0.1,-0.01,0.013), the latter of similar magnitude to (and rounded from) the Gegenbauer-moment values reported for the kaon distribution amplitude in Table I of Ref. [48]. We emphasize that these coefficients are used here only as a representative, realistically sized asymmetric input for the Cj5/2C_j^{5/2}, valence-PDF-style test function of Eq. (20), not as a literal instance of the kaon PDA itself, whose natural expansion is the distinct Cn3/2C_n^{3/2} basis discussed above. Both instances are fit with the same architectures and training setup: the MLP has three hidden layers of

8 neurons each, and the KAN has the same architecture as in § 5.1 (two hidden layers of 16 nodes, spline order 3, grid size 10); each core is trained independently nens=30n_{\mathrm{ens}}=30 times per instance.

Two-term expansion, (a1,a2)=(0.1,0.02)(a_1,a_2)=(0.1,0.02)

Table 2 compares the exact and reconstructed moments for both cores, and the top row of Figure 2 shows the pointwise reconstructions.

Table 2: Asymmetric benchmark, Eq. (20) with (a1,a2)=(0.1,0.02)(a_1,a_2)=(0.1,0.02). M‾ntrue\overline{M}^{\mathrm{true}}_n, M‾npred\overline{M}^{\mathrm{pred}}_n, and ϵn=∣M‾npred−Mntrue∣/∣Mntrue∣\epsilon_n=\left|\overline{M}^{\mathrm{pred}}_n-M^{\mathrm{true}}_n\right|/\left|M^{\mathrm{true}}_n\right| are defined as in Table 1, for the MLP and the KAN (nens=30n_{\mathrm{ens}}=30 each).

MLPKAN
nnMntrueM^{\mathrm{true}}_nM‾npred\overline{M}^{\mathrm{pred}}_nϵn\epsilon_nM‾npred\overline{M}^{\mathrm{pred}}_nϵn\epsilon_n
10.46430.4640(1)5×10−45 \times10^{-4}0.4643(3)5×10−55 \times10^{-5}
20.25240.2527(1)1×10−31 \times10^{-3}0.2525(4)4×10−44 \times10^{-4}
30.15240.1525(1)8×10−48 \times10^{-4}0.1523(3)6×10−46 \times10^{-4}
40.09910.0991(1)8×10−48 \times10^{-4}0.0989(3)2×10−32 \times10^{-3}
50.06820.06803(7)2×10−32 \times10^{-3}0.0680(3)3×10−33 \times10^{-3}
60.04900.04881(5)3×10−33 \times10^{-3}0.0488(2)3×10−33 \times10^{-3}
70.03640.03628(5)2×10−32 \times10^{-3}0.0363(2)1×10−31 \times10^{-3}
80.02780.02775(7)7×10−47 \times10^{-4}0.0278(2)2×10−32 \times10^{-3}
90.02170.0218(1)2×10−32 \times10^{-3}0.0218(1)6×10−36 \times10^{-3}
100.01730.0174(1)6×10−36 \times10^{-3}0.0175(1)1×10−21 \times10^{-2}

Table 2.

Asymmetric benchmark reconstructions comparing MLP and KAN with exact distributions

Figure 2. Asymmetric (kaon-like) benchmark, Eq. (20): MLP (left column) and KAN (right column) reconstructions against the exact distribution, for the two-term expansion (a1,a2)=(0.1,0.02)(a_1, a_2) = (0.1, 0.02) (top row) and the four-term expansion (a1,a2,a3,a4)=(−0.08,0.1,−0.01,0.013)(a_1, a_2, a_3, a_4) = (-0.08, 0.1, -0.01, 0.013) adapted from Ref. [48] (bottom row). Each panel shows the ensemble mean (nens=30)(n_{\mathrm{ens}} = 30) with a one-standard-deviation band.

Four-term expansion, (a1,a2,a3,a4)=(−0.08,0.1,−0.01,0.013)(a_1,a_2,a_3,a_4)=(-0.08,0.1,-0.01,0.013)

Table 3 compares the exact and reconstructed moments for both cores, and the bottom row of Figure 2 shows the pointwise reconstructions. This instance is more structured than the two-term one, and moment reproduction degrades faster with nn for both cores, but markedly less so for the KAN.

Table 3: Asymmetric benchmark, Eq. (20) with (a1,a2,a3,a4)=(−0.08,0.1,−0.01,0.013)(a_1,a_2,a_3,a_4)=(-0.08,0.1,-0.01,0.013), adapted from Ref. [48]. MntrueM^{\mathrm{true}}_n, M‾npred\overline{M}^{\mathrm{pred}}_n, and ϵn\epsilon_n as in Table 2.

MLPKAN
nnMntrueM^{\mathrm{true}}_nM‾npred\overline{M}^{\mathrm{pred}}_nϵn\epsilon_nM‾npred\overline{M}^{\mathrm{pred}}_nϵn\epsilon_n
10.52860.5288(1)4×10−44 \times10^{-4}0.5286(2)2×10−52 \times10^{-5}
20.32620.3259(1)9×10−49 \times10^{-4}0.3262(2)3×10−53 \times10^{-5}
30.22060.2200(1)3×10−33 \times10^{-3}0.2204(2)9×10−49 \times10^{-4}
40.15850.15799(5)3×10−33 \times10^{-3}0.1582(2)2×10−32 \times10^{-3}
50.11880.11861(8)2×10−32 \times10^{-3}0.1186(1)2×10−32 \times10^{-3}
60.09200.0921(2)2×10−32 \times10^{-3}0.0919(1)1×10−31 \times10^{-3}
70.07300.0735(2)7×10−37 \times10^{-3}0.0730(1)4×10−44 \times10^{-4}
80.05910.0599(2)1×10−21 \times10^{-2}0.0593(2)3×10−33 \times10^{-3}
90.04860.0497(3)2×10−22 \times10^{-2}0.0489(3)6×10−36 \times10^{-3}
100.04060.0419(3)3×10−23 \times10^{-2}0.0410(3)1×10−21 \times10^{-2}

Table 3.

Across both instances the KAN reconstructs the pointwise profile of the target distribution more faithfully than the MLP, consistent with the symmetric benchmark of § 5.1. We stress that this advantage refers to the xx-profile, not to each individual moment: both cores reproduce the input moments at a comparable level (in Table 2 the MLP is even slightly more precise for several nn), which is precisely the ill-posedness discussed in § 2—nearly identical moment sets are compatible with visibly different profiles, as the MLP panels of Figure 2 show. What does distinguish the KAN systematically is the lowest orders: in all three benchmarks it reproduces the moments n=1n = 1–33 more precisely than the MLP, by up to an order of magnitude (Table 1, Table 2, and Table 3). These are the moments that fix the location, width, and skewness of the profile, and which the loss already weights most heavily, so their more accurate reproduction is consistent with the KAN’s more faithful pointwise reconstruction—although, the inversion being ill posed, low-order moment accuracy alone cannot guarantee it. At the level of moments, the difference is most visible in the higher moments of the four-term, more structured instance, where the MLP’s relative error grows to 3×10−23 \times10^{-2} at n=10n = 10 against 1×10−21 \times10^{-2} for the KAN.

Effect of the number of input moments (KAN)

Having established that the KAN outperforms the MLP at fixed moment count, we now assess, for the KAN alone, how the reconstruction degrades as fewer moments are supplied. (The MLP was also trained on N=5N = 5 moments; its reconstructions deviated from the true profile considerably more than the KAN ones, and are not shown.) We repeat both asymmetric instances of the previous section, training the same KAN architecture on only the first N=5N = 5 moments, again with nens=30n_{\mathrm{ens}} = 30 independent trainings, and compare the resulting reconstruction to the N=10N = 10 case already reported, both pointwise and for all ten moments—including the five beyond what the N=5N = 5 fit was ever asked to match.

Table 4 and 5 compare the N=10N = 10 and N=5N = 5 KAN moments for the two-term and four-term expansions, respectively, and Figure 3 shows the corresponding N=5N = 5 pointwise reconstructions. In both instances the moments that entered the N=5N = 5 training (n≤5n \le5, above the rule in each table) remain about as well reproduced as in the N=10N = 10 fit, but the higher moments (n>5n > 5, below the rule), which the N=5N = 5 network never saw during training, show visibly larger deviations from the true values than their N=10N = 10 counterparts—most markedly at n=9,10n = 9, 10, where the relative error grows by roughly a factor of four to five. This is the expected signature of a more strongly ill-posed inversion at smaller NN: with less information about the tail behavior supplied during training, the reconstruction is less constrained away from the region directly probed by the low moments, and the pointwise bands in Figure 3 are correspondingly wider than the N=10N = 10 KAN panels (right column) of Figure 2.

nnM‾ntrue\overline{M}^{\mathrm{true}}_nM‾npred\overline{M}^{\mathrm{pred}}_nεn\varepsilon_nM‾npred\overline{M}^{\mathrm{pred}}_nεn\varepsilon_n
10.46430.4643(3)5×10−55 \times10^{-5}0.4642(2)2×10−42 \times10^{-4}
20.25240.2525(4)4×10−44 \times10^{-4}0.2523(3)2×10−42 \times10^{-4}
30.15240.1523(3)6×10−46 \times10^{-4}0.1522(2)1×10−31 \times10^{-3}
40.09910.0989(3)2×10−32 \times10^{-3}0.0989(2)2×10−32 \times10^{-3}
50.06820.0680(3)3×10−33 \times10^{-3}0.0681(2)7×10−47 \times10^{-4}
60.04900.0488(2)3×10−33 \times10^{-3}0.0491(2)3×10−33 \times10^{-3}
70.03640.0363(2)1×10−31 \times10^{-3}0.0367(3)9×10−39 \times10^{-3}
80.02780.0278(2)2×10−32 \times10^{-3}0.0283(3)2×10−22 \times10^{-2}
90.02170.0218(1)6×10−36 \times10^{-3}0.0224(4)3×10−23 \times10^{-2}
100.01730.0175(1)1×10−21 \times10^{-2}0.0181(4)5×10−25 \times10^{-2}

Table 4. KAN reconstruction of the two-term asymmetric benchmark, Eq. (20) with (a1,a2)=(0.1,0.02)(a_1,a_2) = (0.1, 0.02), trained on the first N=10N = 10 vs. the first N=5N = 5 moments (nens=30n_{\mathrm{ens}} = 30 each). M‾ntrue\overline{M}^{\mathrm{true}}_n, M‾npred\overline{M}^{\mathrm{pred}}_n, and εn\varepsilon_n as in Table 1. The rule after n=5n = 5 separates the moments used to train the N=5N = 5 model from those extrapolated beyond it.

KAN reconstructions using only the first N = 5 moments for the two-term and four-term expansions

Figure 3. KAN reconstructions using only the first N=5N=5 moments, for the two-term expansion (a1,a2)=(0.1,0.02)(a_1,a_2)=(0.1,0.02) (left) and the four-term expansion (a1,a2,a3,a4)=(−0.08,0.1,−0.01,0.013)(a_1,a_2,a_3,a_4)=(-0.08,0.1,-0.01,0.013) adapted from Ref. [48] (right). Each panel shows the ensemble mean (nens=30n_{\mathrm{ens}}=30) with a one-standard-deviation band; compare to the N=10N=10 KAN panels (right column) of Figure 2.

Table 5. KAN reconstruction of the four-term asymmetric benchmark, Eq. (20) with (a1,a2,a3,a4)=(−0.08,0.1,−0.01,0.013)(a_1,a_2,a_3,a_4)=(-0.08,0.1,-0.01,0.013), adapted from Ref. [48], trained on the first N=10N=10 vs. the first N=5N=5 moments (nens=30n_{\mathrm{ens}}=30 each). M‾ntrue\overline{M}_n^{\mathrm{true}}, M‾npred\overline{M}_n^{\mathrm{pred}}, and ϵn\epsilon_n as in Table 1; the rule after n=5n=5 is as in Table 4.

nnMntrueM_n^{\mathrm{true}}\multicolumn{2}{c}{KAN, N=10N=10}\multicolumn{2}{c}{KAN, N=5N=5}
M‾npred\overline{M}_n^{\mathrm{pred}}ϵn\epsilon_nM‾npred\overline{M}_n^{\mathrm{pred}}ϵn\epsilon_n
10.52860.5286(2)2×10−52 \times10^{-5}0.52860(6)5×10−55 \times10^{-5}
20.32620.3262(2)3×10−53 \times10^{-5}0.32616(8)9×10−59 \times10^{-5}
30.22060.2204(2)9×10−49 \times10^{-4}0.22050(5)5×10−45 \times10^{-4}
40.15850.1582(2)2×10−32 \times10^{-3}0.1585(1)3×10−53 \times10^{-5}
50.11880.1186(1)2×10−32 \times10^{-3}0.1191(2)2×10−32 \times10^{-3}
60.09200.0919(1)1×10−31 \times10^{-3}0.0926(3)6×10−36 \times10^{-3}
70.07300.0730(1)4×10−44 \times10^{-4}0.0739(4)1×10−21 \times10^{-2}
80.05910.0593(2)3×10−33 \times10^{-3}0.0603(4)2×10−22 \times10^{-2}
90.04860.0489(3)6×10−36 \times10^{-3}0.0501(4)3×10−23 \times10^{-2}
100.04060.0410(3)1×10−21 \times10^{-2}0.0422(4)4×10−24 \times10^{-2}

Table 5 (PDF p. 13).

Discussion and outlook

Our aim has been to reconstruct meson distributions from finite moment data with minimal functional bias: a flexible neural core is shown to accurately capture the shape, while only first-principles properties—support, endpoint behavior, normalization, and (for the pion and quarkonia) the x↔1−xx \leftrightarrow1-x symmetry—are hard-wired, and soft penalties suppress the multiple-maxima and central-concavity artifacts of ill-posed inversion. Replacing the MLP core by a KAN improves the accuracy with which the pointwise profile is reconstructed, most visibly for the asymmetric kaon and at the smallest moment counts; we attribute this to the compact support of the B-spline activations, which localizes moment-driven corrections and damps the global oscillations to which the MLP is susceptible.

This comparison is not matched in parameter count: for the architectures of § 5, the KAN carries substantially more trainable parameters than the MLP (of order 4×1034 \times10^{3} for the KAN cores used here, against a few hundred for the MLP, i.e. roughly an order of magnitude more in the symmetric benchmark and over twenty-fold more in the asymmetric one), so capacity alone could in principle account for part of the KAN’s advantage. We do not believe this is the explanation, however: enlarging the MLP core, by widening or adding hidden layers, did not improve its reconstructions in our trials and in fact made them less stable, consistent with the picture of § 3.2—additional capacity gives an architecture with globally supported activations more ways to fit the target moments with a spurious oscillation, not fewer. A controlled comparison at matched parameter count, isolating the effect of the spline basis from raw capacity, would nonetheless make this conclusion more rigorous and is a natural next step.

Furthermore, having demonstrated that the neural cores described here can reproduce the shape of a distribution from a limited set of moments, imposing only a few well-grounded physical constraints and no particular functional form, we plan to follow this feasibility study with a reconstruction of meson PDFs and PDAs from the moments available in the literature, as obtained within the DSE/BSE framework.

On the other hand, a limitation of this “fit-per-dataset” strategy is that a full optimization is required for every new moment sequence, so it is not an instantaneous emulator. A natural complement, borrowed from the variational- and simulation-based-inference literature [56, 57], is amortized inference: rather than solving a fresh optimization problem for every new instance, as we do here, a single network is trained once on a large collection of instances, so that the cost of inference is paid up front during training and “amortized” over all subsequent uses—at test time a new instance requires only a forward pass through the trained network, not a new optimization. In our setting this means training a single model on a large synthetic library of moment sequences drawn from many functional families, mapping the (sequential) input {Mn}n=1N\{M_n\}_{n=1}^{N} directly to the reconstructed mesh values {q(xi)}\{q(x_i)\}. The sequential structure is naturally suited to a Transformer. Note that this supervised setting does not admit the hard/soft constraints of § 2–§ 4 in the same form: the network regresses directly on mesh values rather than on an unconstrained core that is subsequently wrapped or penalized, so there is neither an architectural wrapper to apply nor a moment-matching loss to augment with shape penalties. The constraints are instead retained only implicitly, through the choice of training data: the functional families used to generate the synthetic library are themselves built to satisfy support, endpoint, symmetry, and unimodality properties by construction, so the network only ever learns from (and hence only ever outputs) constraint-satisfying profiles. We have already obtained encouraging results along these lines with an encoder-only Transformer architecture, which we plan to present in a forthcoming paper. Such an approach trades some functional bias—bounded by the diversity of the training library—for immediate inference once trained, enabling fast scans and uncertainty propagation. The two strategies are thus complementary: the present constrained fit minimizes bias per dataset, while the amortized model prioritizes rapid inference after an up-front training cost.

B-splines

This appendix collects the definitions needed to make the KAN construction of § 3.2 self-contained; see [58] for a complete treatment. Fix a non-decreasing sequence of knots t0≤t1≤⋯≤tmt_0 \le t_1 \le\cdots\le t_m on the domain of interest. The B-splines of degree 00 associated with this knot sequence are the piecewise-constant indicator functions of the knot intervals,

Bi,0(x)={1,ti≤x<ti+1,0,otherwise,(21)B_{i,0}(x) = \begin{cases} 1, & t_i \le x < t_{i+1},\\ 0, & \text{otherwise}, \end{cases} \tag*{(21)}

and the B-splines of degree p≥1p \ge1 are defined from those of degree p−1p-1 by the Cox–de Boor recursion,

Bi,p(x)=x−titi+p−tiBi,p−1(x)+ti+p+1−xti+p+1−ti+1Bi+1,p−1(x),(22)B_{i,p}(x) = \frac{x-t_i}{t_{i+p}-t_i} B_{i,p-1}(x) + \frac{t_{i+p+1}-x}{t_{i+p+1}-t_{i+1}} B_{i+1,p-1}(x), \tag*{(22)}

with the convention that any term with a vanishing denominator is set to zero. Each Bi,pB_{i,p} is itself a spline of degree pp: a polynomial of degree pp on every sub-interval [tj,tj+1)[t_j,t_{j+1}), joined p−1p-1 times continuously differentiably across knots. Iterating Eq. (22) shows that Bi,pB_{i,p} is supported on [ti,ti+p+1][t_i,t_{i+p+1}] and vanishes identically outside this window of p+1p+1 consecutive knot intervals—the compact-support property used in § 3.2. Any spline of degree pp on the given knot sequence can be written uniquely as a linear combination

s(x)=∑iciBi,p(x),(23)s(x) = \sum_i c_i B_{i,p}(x), \tag*{(23)}

and, because of the support property above, changing a single coefficient cic_i leaves s(x)s(x) unchanged everywhere outside [ti,ti+p+1][t_i,t_{i+p+1}]. This is the property Eq. (10) relies on, with pp fixed by the spline order of the KAN and {ti}\{t_i\} a fixed, uniform knot grid on the edge’s input range.

References

  1. [1]R. Keith Ellis, W. James Stirling, and B. R. Webber. QCD and collider physics, volume 8. Cambridge University Press, 2 2011.
  2. [2]G. P. Lepage and S. J. Brodsky. Exclusive processes in perturbative quantum chromodynamics. Phys. Rev. D, 22:2157–2198, 1980. Related earlier work: Phys. Rev. Lett. 43, 545 (1979); Erratum: Phys. Rev. Lett. 43, 1625 (1979).
  3. [3]A. V. Efremov and A. V. Radyushkin. Factorization and asymptotic behavior of pion form factor in qcd. Phys. Lett. B, 94:245–250, 1980. Related work by A. V. Radyushkin, Phys. Lett. B 131, 179 (1983).
  4. [4]Khépani Raya, Adnan Bashir, Daniele Binosi, Craig D. Roberts, and José Rodríguez-Quintero. Pseudoscalar Mesons and Emergent Mass. Few Body Syst., 65(2):60, 2024. arXiv:2403.00629.arxiv.org/abs/2403.00629
  5. [5]Lei Chang, I. C. Cloet, J. J. Cobos-Martinez, C. D. Roberts, S. M. Schmidt, and P. C. Tandy. Imaging dynamical chiral symmetry breaking: pion wave function on the light front. Phys. Rev. Lett., 110(13):132001, 2013. arXiv:1301.0324.DOI
  6. [6]Yin-Zhen Xu, Khépani Raya, Zhu-Fang Cui, Craig D. Roberts, and J. Rodríguez-Quintero. Empirical Determination of the Pion Mass Distribution. Chin. Phys. Lett., 40(4):041201, 2023. arXiv:2302.07361.DOI
  7. [7]Zhen-Ni Xu, Zhao-Qian Yao, Cédric Mezrag, Khépani Raya, and José Rodríguez-Quintero. Pseudoscalar charmonium and bottomonium: light-front wave functions, distribution amplitudes and distribution functions. 6 2026. arXiv:2606.28442.arxiv.org/abs/2606.28442
  8. [8]Z. F. Cui, Minghui Ding, J. M. Morgado, K. Raya, D. Binosi, L. Chang, J. Papavassiliou, C. D. Roberts, J. Rodríguez-Quintero, and S. M. Schmidt. Concerning pion parton distributions. Eur. Phys. J. A, 58(1):10, 2022. arXiv:2112.09210.DOI
  9. [9]Z. F. Cui, Minghui Ding, J. M. Morgado, K. Raya, D. Binosi, L. Chang, F. De Soto, C. D. Roberts, J. Rodríguez-Quintero, and S. M. Schmidt. Emergence of pion parton distributions. Phys. Rev. D, 105(9):L091502, 2022. arXiv:2201.00884.DOI
  10. [10]Pei-Lin Yin, Yin-Zhen Xu, Zhu-Fang Cui, Craig D. Roberts, and José Rodríguez-Quintero. All-Orders Evolution of Parton Distributions: Principle, Practice, and Predictions. Chin. Phys. Lett., 40(9):091201, 2023. arXiv:2306.03274.arxiv.org/abs/2306.03274
  11. [11]Ya Lu, Yin-Zhen Xu, Khépani Raya, Craig D. Roberts, and José Rodríguez-Quintero. Pion distribution functions from low-order Mellin moments. Phys. Lett. B, 850:138534, 2024. arXiv:2311.08565.arxiv.org/abs/2311.08565
  12. [12]Zhen-Ni Xu, Daniele Binosi, Chen Chen, Khépani Raya, Craig D. Roberts, and José Rodríguez-Quintero. Kaon distribution functions from empirical information. Phys. Lett. B, 865:139451, 2025. arXiv:2411.15376.arxiv.org/abs/2411.15376
  13. [13]Minghui Ding, Khépani Raya, Daniele Binosi, Lei Chang, Craig D Roberts, and Sebastian M Schmidt. Drawing insights from pion parton distributions. Chin. Phys. C, 44(3):031002, 2020. arXiv:1912.07529.DOI
  14. [14]Zhu-Fang Cui, Minghui Ding, Fei Gao, Khépani Raya, Daniele Binosi, Lei Chang, Craig D Roberts, Jose Rodríguez-Quintero, and Sebastian M Schmidt. Kaon and pion parton distributions. Eur. Phys. J. C, 80(11):1064, 2020.
  15. [15]Zhen-Ni Xu, Zhao-Qian Yao, Peng Cheng, and Khépani Raya. Distribution amplitudes of vector and axial-vector mesons in a nonperturbatively improved symmetry-preserving framework. 9 2026. arXiv:2609.02329.arxiv.org/abs/2609.02329
  16. [16]Gunnar S. Bali, Vladimir M. Braun, Simon Bürger, Meinulf Göckeler, Michael Gruber, Fabian Hutzler, Piotr Korcyl, Andreas Schäfer, André Sternbeck, and Philipp Wein. Light-cone distribution amplitudes of pseudoscalar mesons from lattice QCD. JHEP, 08:065, 2019. [Addendum: JHEP 11, 037 (2020)].arxiv.org/abs/1903.08038
  17. [17]Bálint Joó, Joseph Karpie, Kostas Orginos, Anatoly V. Radyushkin, David G. Richards, Raza Sabbir Sufian, and Savvas Zafeiropoulos. Pion valence structure from Ioffe-time parton pseudodistribution functions. Phys. Rev. D, 100(11):114512, 2019. arXiv:1909.08517.
  18. [18]Constantia Alexandrou, Simone Bacchio, Ian Cloët, Martha Constantinou, Kyriakos Hadjiyiannakou, Giannis Koutsou, and Colin Lauer. Pion and kaon ⟨x3⟩ from lattice QCD and PDF reconstruction from Mellin moments. Phys. Rev. D, 104(5):054504, 2021. arXiv:2104.02247.DOI
  19. [19]Xiang Gao, Andrew D. Hanlon, Nikhil Karthik, Swagato Mukherjee, Peter Petreczky, Philipp Scior, Shuzhe Shi, Sergey Syritsyn, Yong Zhao, and Kai Zhou. Continuum-extrapolated NNLO valence PDF of the pion at the physical point. Phys. Rev. D, 106(11):114510, 2022. arXiv:2208.02297.DOI
  20. [20]J. Hadamard. Lectures on Cauchy's Problem in Linear Partial Differential Equations. Yale University Press, 1923.
  21. [21]A. N. Tikhonov and V. Y. Arsenin. Solutions of Ill-Posed Problems. Winston, 1977.DOI
  22. [22]H. W. Engl, M. Hanke, and A. Neubauer. Regularization of Inverse Problems. Springer, 1996.
  23. [23]J. A. Shohat and J. D. Tamarkin. The Problem of Moments. American Mathematical Society, 1943.DOI
  24. [24]Jingxuan Chen, Xiaopeng Wang, Yanbing Cai, Xurong Chen, and Qian Wang. Valence quark distributions of pions: insights from Tsallis entropy*. Chin. Phys. C, 50(1):013103, 2026. arXiv:2408.03068.arxiv.org/abs/2408.03068
  25. [25]K. Hornik. Multilayer feedforward networks are universal approximators. Neural Networks, 2:359–366, 1989.DOI
  26. [26]G. Cybenko. Approximation by superpositions of a sigmoidal function. Math. Control Signals Systems, 2:303–314, 1989.
  27. [27]A. N. Kolmogorov. On the representation of continuous functions of several variables by superpositions of continuous functions of one variable and addition. Dokl. Akad. Nauk SSSR, 114:953, 1957. See also V. I. Arnold, Dokl. Akad. Nauk SSSR 114, 679 (1957).
  28. [28]M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys., 378:686–707, 2019.DOI
  29. [29]S. Forte, L. Garrido, J. I. Latorre, and A. Piccione. Neural network parametrization of deep-inelastic structure functions. JHEP, 0205:062, 2002. arXiv:hep-ph/0204232.DOI
  30. [30]NNPDF Collaboration, R. D. Ball, et al. Parton distributions for the lhc run ii. JHEP, 1504:040, 2015. arXiv:1410.8849.
  31. [31]H. Dutriex, O. Grocholski, H. Moutarde, and P. Sznajder. Artificial neural network modelling of generalised parton distributions. Eur. Phys. J. C, 82:252, 2022. Erratum: Eur. Phys. J. C 82, 389 (2022); arXiv:2112.10528.DOI
  32. [32]P. Dall’Olio, F. De Soto, C. Mezrag, J. M. Morgado Chávez, H. Moutarde, J. Rodríguez-Quintero, P. Sznajder, and J. Segovia. Unraveling generalized parton distributions through lorentz symmetry and partial dglap knowledge. Phys. Rev. D, 109:096013, 2024. arXiv:2401.12013.arxiv.org/abs/2401.12013
  33. [33]M.-H. Chu, K. Cichy, M. Constantinou, P. Sznajder, and J. Wagner. A unified neural-network framework for nucleon imaging from numerical simulations of qcd. JHEP, 05:210, 2026. arXiv:2509.15931.DOI
  34. [34]S. J. Brodsky and G. R. Farrar. Scaling laws for large-momentum-transfer processes. Phys. Rev. Lett., 31:1153–1156, 1973. See also Phys. Rev. D 11, 1309 (1975).
  35. [35]G. R. Farrar and D. R. Jackson. Pion and nucleon structure functions near the physical end point. Phys. Rev. Lett., 35:1416–1419, 1975.
  36. [36]Edmond L. Berger and Stanley J. Brodsky. Quark Structure Functions of Mesons and the Drell-Yan Process. Phys. Rev. Lett., 42:940–944, 1979.DOI
  37. [37]Yuri L. Dokshitzer. Calculation of the Structure Functions for Deep Inelastic Scattering and e+ e- Annihilation by Perturbation Theory in Quantum Chromodynamics. Sov. Phys. JETP, 46:641–653, 1977.
  38. [38]V. N. Gribov and L. N. Lipatov. Deep inelastic e p scattering in perturbation theory. Sov. J. Nucl. Phys., 15:438–450, 1972.
  39. [39]L. N. Lipatov. The parton model and perturbation theory. Yad. Fiz., 20:181–198, 1974.
  40. [40]Guido Altarelli and G. Parisi. Asymptotic Freedom in Parton Language. Nucl. Phys. B, 126:298–318, 1977.DOI
  41. [41]Raza Sabbir Sufian, Colin Egerer, Joseph Karpie, Robert G. Edwards, Bálint Joó, Yan-Qing Ma, Kostas Orginos, Jian-Wei Qiu, and David G. Richards. Pion Valence Quark Distribution from Current-Current Correlation in Lattice QCD. Phys. Rev. D, 102(5):054508, 2020. arXiv:2001.04960.DOI
  42. [42]Lei Chang, Khépani Raya, and Xiaobin Wang. Pion Parton Distribution Function in Light-Front Holographic QCD. Chin. Phys. C, 44(11):114105, 2020. arXiv:2001.07352.DOI
  43. [43]M. B. Hecht, Craig D. Roberts, and S. M. Schmidt. Valence quark distributions in the pion. Phys. Rev. C, 63:025213, 2001. arXiv:nucl-th/0008049.DOI
  44. [44]Roy J. Holt and Craig D. Roberts. Distribution Functions of the Nucleon and Pion in the Valence Region. Rev. Mod. Phys., 82:2991–3044, 2010. arXiv:1002.4666.DOI
  45. [45]Xiaobin Wang, Lei Chang, Minghui Ding, Khepani Raya, and Craig D. Roberts. Symmetry Constraints on Pion Valence Structure. Chin. Phys. Lett., 43(7):070203, 2026. arXiv:2510.23950.arxiv.org/abs/2510.23950
  46. [46]R. D. Ball, E. R. Nocera, and J. Rojo. The proton spin and its decomposition. Eur. Phys. J. C, 76:383, 2016. arXiv:1604.00024.
  47. [47]V. L. Chernyak and A. R. Zhitnitsky. Asymptotic behavior of exclusive processes in qcd. Phys. Rept., 112:173–318, 1984.DOI
  48. [48]K. Raya, A. Bashir, and J. Rodríguez-Quintero. Mapping spatial distributions within pseudoscalar mesons. Chin. Phys. Lett., 42:020201, 2025. arXiv:2412.06025.arxiv.org/abs/2412.06025
  49. [49]G. Peter Lepage and Stanley J. Brodsky. Exclusive Processes in Quantum Chromodynamics: Evolution Equations for Hadronic Wave Functions and the Form-Factors of Mesons. Phys. Lett. B, 87:359–365, 1979.
  50. [50]A. V. Efremov and A. V. Radyushkin. Factorization and Asymptotical Behavior of Pion Form-Factor in QCD. Phys. Lett. B, 94:245–250, 1980.
  51. [51]G. Peter Lepage and Stanley J. Brodsky. Exclusive Processes in Perturbative Quantum Chromodynamics. Phys. Rev. D, 22:2157, 1980.
  52. [52]E. T. Jaynes. Information theory and statistical mechanics. Phys. Rev., 106:620–630, 1957.DOI
  53. [53]M. Asakawa, T. Hatsuda, and Y. Nakahara. Maximum entropy analysis of the spectral functions in lattice qcd. Prog. Part. Nucl. Phys., 46:459–508, 2001. arXiv:hep-lat/0011040.DOI
  54. [54]Z. Liu et al. Kan: Kolmogorov–arnold networks. In International Conference on Learning Representations (ICLR), 2025. arXiv:2404.19756.arxiv.org/abs/2404.19756
  55. [55]W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. Numerical Recipes: The Art of Scientific Computing. Cambridge University Press, 3 edition, 2007.
  56. [56]D. P. Kingma and M. Welling. Auto-encoding variational bayes. In International Conference on Learning Representations (ICLR), 2014. arXiv:1312.6114.arxiv.org/abs/1312.6114
  57. [57]K. Crammer, J. Brehmer, and G. Louppe. The frontier of simulation-based inference. Proc. Natl. Acad. Sci., 117:30055–30062, 2020. arXiv:1911.01429.
  58. [58]C. de Boor. A Practical Guide to Splines. Springer, revised edition, 2001.

Paper details

vlnc:W_6chu6xfe
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents