Weighted Quantum Signal Processing: Low-Depth Polynomial Approximation with Applications to Kolmogorov-Arnold Networks

AI summary

Overview

Research area: Quantum algorithms, specifically Quantum Signal Processing (QSP) and its use for polynomial and continuous-function approximation, with an application to quantum-enhanced machine learning via Kolmogorov–Arnold Networks (KANs).

Technical level: Advanced. The paper is written in the language of SU(2) circuits, Chebyshev/Laurent polynomial algebra, and phase-factor synthesis; the machine-learning portion is accessible to intermediate readers, but the core theory assumes comfort with quantum circuit formalisms.

Scope (one sentence): The paper introduces Weighted Quantum Signal Processing (WQSP), a modification of QSP in which a classical weight is attached to the central rotation, and shows how suitable weights compress the parameter count and circuit depth of polynomial approximation while retaining expressive power, then uses the construction to build quantum-native KAN activation functions.

What This Paper Is About

QSP is a powerful primitive for turning a simple single-qubit rotation into a chosen polynomial of the input, but realizing high-degree polynomials demands long circuits and many trainable phase parameters, and standard QSP is restricted to polynomials of fixed parity, forcing extra decomposition machinery for general targets. The authors propose WQSP, which assigns weights to the central rotation operator so that the same circuit template can trade classical weight design for quantum resources, and they show that QSP is just the unit-weight special case. They then use WQSP-generated polynomials as the learnable univariate activation functions inside a Kolmogorov–Arnold Network, producing a compact quantum-native architecture for multivariate function approximation.

Key Contributions

1. A weighted generalization of QSP. WQSP attaches a weight function to the central rotation operator of the QSP circuit. The standard QSP circuit is recovered exactly when all weights equal one, so WQSP is a strict generalization rather than a replacement.

2. Integer weights as a pruned QSP. When the weights are natural numbers greater than one, the WQSP circuit reduces to a pruned version of QSP, exposing parameter redundancies in the standard framework. With appropriate integer weight choices, the authors report linear-to-exponential reductions in the number of parameters needed to realize arbitrary bounded univariate polynomials, while preserving approximation quality.

3. Error bounds for generic weights, and two solution strategies. For weights that are not integers, the paper derives approximation error bounds and shows that in many cases the approximation is exact. WQSP is analyzed both deterministically (polynomial generation cast as the solution of a linear system) and as a structured quantum machine-learning model with trainable parameters.

4. A WQSP-based Kolmogorov–Arnold Network. The authors use WQSP to parameterize the learnable edge activation functions of a KAN for multivariate function approximation, avoiding the block-encoding overheads of the existing quantum KAN architecture (cited as Ivashkov et al., 2024). They also emphasize that WQSP uses R_x(θ) rotations implementable as a single modulated pulse, whereas GQSP's general SU(2) rotations typically require 2 to 3 physical pulses under Euler decomposition.

Main Findings

• QSP is a special case, not a competitor. Setting all weights to one reproduces the standard QSP unitary and its polynomial structure, so the weighted formulation is a genuine superset of the existing framework rather than an alternative ansatz.

• Weights control circuit structure and expressivity. The authors state that the choice of weights "fundamentally determines the structure and expressive capabilities" of WQSP circuits — weights are the knob that governs depth, parameter count, and parity behavior.

• Redundancy in standard QSP is real and removable. Because integer weights greater than one make WQSP a pruned QSP, the paper argues that standard QSP carries removable redundancy that structured weight design can eliminate without losing expressive power.

• Parameter compression without loss of accuracy. The abstract reports linear-to-exponential reductions in parameter count for arbitrary bounded univariate polynomials while preserving approximation quality — a claim about asymptotic scaling, not a measured benchmark.

• Controlled error for non-integer weights. For generic weights the paper gives explicit approximation bounds characterizing the trade-off: heavier compression lowers circuit depth at the cost of a controlled increase in approximation error. Many cases are reported to be exact.

• Two ways to find the phases. Deterministic recovery frames polynomial generation as solving a linear system; the learning-based route trains WQSP as a variational model. The learning formulation is described as extending beyond polynomial approximation to general continuous functions.

• Hardware cost is lower than GQSP. GQSP's generic SU(2) rotations decompose into 2 to 3 physical pulses via Euler decomposition, while WQSP's R_x(θ) rotations need a single modulated pulse, reducing physical gate depth and accumulated error on NISQ hardware.

• Broader phase-synthesis context. The paper situates itself against prior phase-recovery work: an O(d²) algorithm for a degree-d polynomial, and a Riemann–Hilbert (Weiss) algorithm costing O(d³) per phase factor. These are cited from prior literature, not new results of this paper.

• No numerical experiments are reported in the provided content. The text supplied here contains no benchmark names, dataset sizes, model sizes, training results, or error tables; the WQSP-KAN claims are presented as framework-level arguments rather than measured outcomes in the accessible sections.

Methodology in Plain English

The starting point is the standard QSP circuit: a signal rotation U(x) = exp(i arccos(x) σ_x) applied repeatedly, interleaved with Z-rotations whose angles are the trainable phase parameters. The output is a 2×2 unitary whose top-left entry is a polynomial P(x) in the input, with the well-known QSP constraints — degree at most d, bounded by 1 on [−1, 1], a companion polynomial Q of degree at most d−1 with opposite parity, and the normalization |P(x)|² + (1−x²)|Q(x)|² = 1.

WQSP modifies this template by placing a weight on the central signal rotation. Choose the weights as integers greater than one and consecutive signal steps effectively coincide, so the circuit collapses to a shorter sequence — the authors describe this as a pruned QSP. Choose arbitrary weights and a different polynomial family results, for which the paper derives error bounds.

Two routes are then offered for actually finding the angles that produce a desired function. The deterministic route asks which phase vector makes the circuit's top-left entry equal the target polynomial, and solves the resulting linear system. The learning route treats the phase angles (and the weights) as variational parameters and optimizes them against the target function, which allows the same machinery to approximate continuous functions rather than only exact polynomials.

For the machine-learning application, the Kolmogorov–Arnold representation theorem expresses a multivariate continuous function as a sum of outer univariate functions composed with sums of inner univariate functions. KANs make those univariate functions learnable on the network edges. The authors replace each such edge function with a WQSP-generated polynomial, so each activation is produced by a shallow variational quantum circuit instead of a deep block-encoded one, and the whole network is trained as a hybrid quantum-classical model.

Why This Matters

Research impact. QSP underpins asymptotically optimal Hamiltonian simulation and is a core primitive for quantum algorithms, but its depth and parameter costs scale with polynomial degree, and its parity constraints force decomposition into parity sectors plus linear-combination-of-unitaries and amplitude-amplification recombinations for complex-valued targets. GQSP removes the parity restriction at the price of general SU(2) rotations that decompose into multiple pulses and require ancilla-based control. WQSP offers a third position: keep the simple single-qubit structure, and buy circuit efficiency with classical weight design. Framing QSP as a unit-weight special case of a larger family also reframes the phase-synthesis problem as a design problem over weights, which is a different research direction from faster phase-recovery algorithms.

Real-world applications (as motivated in the paper):

• Hamiltonian simulation and spectral transformation — the paper's opening motivation is that implementing f(H) underlies simulation, search, linear-system solvers, and optimization routines.
• Multivariate function approximation and regression — the WQSP-KAN framework targets approximating continuous multivariate functions with fewer trainable parameters.
• Computer vision and time-series analysis — cited among the domains where KANs have already found application.
• Physics-informed learning and graph representation learning — also listed among existing KAN application areas.

Industry relevance. The single-modulated-pulse implementation of R_x(θ), versus 2 to 3 pulses for general SU(2) rotations, speaks directly to near-term hardware budgets where gate count drives fidelity loss through cumulative noise and decoherence. Lower parameter counts also mean shorter training loops and less classical optimization overhead. For organizations evaluating whether QSP-style primitives are worth compiling onto current devices, the "fewer pulses, fewer parameters, same approximation quality" argument is the practical pitch.

Future Directions

• Empirical validation. The provided content reports no experiments; the natural next step is to benchmark WQSP against QSP and GQSP on concrete target polynomials and continuous functions, and to run WQSP-KAN on standard multivariate approximation tasks. None of these results appear in the accessible text.

• Systematic weight design. Since the weights determine structure and expressivity, the open question is how to choose them: what rule maps a target polynomial's degree and parity structure to a weight vector that minimizes depth while staying inside the error bounds?

• Scaling the quantum KAN. The paper positions WQSP-KAN as a compact alternative to block-encoded quantum KANs; whether the compression survives at network scale, how training behaves, and how expressivity compares with classical KAN edge functions are unresolved in the supplied text.

• Composition with the wider toolbox. Combining WQSP with quantum singular value transformation, with existing phase-synthesis algorithms (including the O(d²) and O(d³)-per-factor methods cited), and with error-correction-era circuit compilation are all natural extensions the framework invites.

Target Audience

Quantum algorithms researchers working on QSP, QSVT, and Hamiltonian simulation who want a lower-depth variant of a standard primitive; quantum machine-learning practitioners interested in parameter-efficient variational models; classical KAN and interpretable-ML researchers curious about learnable quantum activation functions; and graduate students in quantum information who already know the basic QSP circuit and want to see how its constraints can be relaxed. Readers without a background in SU(2) circuit algebra or Chebyshev/Laurent polynomial manipulations will find the theoretical sections demanding, but the motivation and the KAN application are readable at an intermediate level.

Authors’ abstract

Quantum Signal Processing is a powerful quantum framework for generating and approximating univariate polynomials. However, QSP is often limited by circuit-depth bottlenecks and parity constraints on the class of realizable polynomials. In this work, we introduce Weighted Quantum Signal Processing, an extension of QSP in which a weight function is assigned to the central rotation operator. This formulation provides a deeper understanding of QSP, which emerges as the special case of WQSP with unit weights. The choice of weights determines the structure and expressive capabilities of WQSP circuits. When the weights are natural numbers greater than one, WQSP reduces to a pruned version of QSP, revealing parameter redundancies in the standard framework. Through appropriate selection of integer weights, WQSP achieves linear-to-exponential reductions in the number of parameters required to realize arbitrary bounded univariate polynomials while preserving approximation quality. For generic weights, we establish corresponding approximation error bounds and show that, in many cases, the approximation is exact. We analyze WQSP from both a deterministic perspective, where polynomial generation is formulated as the solution of a linear system, and a quantum machine learning perspective, where WQSP serves as a structured and expressive quantum learning model. We further employ this learning framework to parameterize learnable activation functions in Kolmogorov--Arnold Networks for multivariate function approximation. Our results show that WQSP provides a compact, flexible, and theoretically grounded framework for realizing arbitrary univariate polynomials while requiring significantly fewer trainable parameters than conventional QSP. This yields expressive and parameter-efficient neural architectures, highlighting the potential of WQSP as a scalable primitive for quantum-enhanced machine learning.