Feedback-driven Recurrent Quantum Neural Network Universality¶

Conference: ICLR2026 arXiv: 2506.16332 Code: None Area: Physics Keywords: quantum reservoir computing, recurrent quantum neural network, universal approximation, fading memory filter, NISQ

TL;DR¶

This paper establishes the first quantitative approximation error bounds and universality proofs for feedback-based recurrent quantum neural networks (RQNNs), demonstrating that RQNNs can approximate arbitrary fading memory filters with a linear readout layer while requiring only \(\lceil\log_2(\varepsilon^{-1})\rceil\) qubits — growing logarithmically with precision — and are thus free from the curse of dimensionality.

Background & Motivation¶

Quantum reservoir computing (QRC) exploits quantum system dynamics to process sequential data, making it particularly suited for NISQ devices; while empirical successes are abundant, theoretical foundations remain weak
Universal approximation theorems for classical recurrent neural networks (RNNs) are well established (Hornik 1991, Barron 1993, Grigoryeva & Ortega 2018), yet quantitative approximation bounds for quantum RNNs are absent
Prior universality proofs for QRC rely on polynomial readout layers (via the Stone–Weierstrass theorem), whereas practical systems universally employ linear readout layers due to their simplicity and efficiency in training
Feedback protocols — which feed the output state back as the next-step input — allow the system to retain input history with fewer components and support real-time computation, but their approximation capabilities have lacked theoretical guarantees

Core Problem¶

Can feedback-driven RQNNs approximate general state-space systems with controllable quantum resources (qubit count, circuit size)?
Does the RQNN family possess universal approximation properties for arbitrary causal, time-invariant, fading memory filters?
Can universality be maintained using only a linear readout layer?

Method¶

RQNN Architecture¶

The network consists of \(N\) parallel quantum circuits, each corresponding to one component of the state vector:

Quantum gate \(\mathtt{U}\): a uniformly controlled quantum gate in block-diagonal form composed of \(n\) parameterized rotation blocks \(\bar{\mathtt{U}}^{(i)}\), where each block is the tensor product of an input encoding gate \(\mathtt{U}_1^{(i)}\) (depending on state \(\bm{x}\) and input \(\bm{z}\)) and a bias gate \(\mathtt{U}_2^{(i)}\)
Initialization gate \(\mathtt{V}\): maps \(|0\rangle^{\otimes \mathfrak{n}}\) to the uniform superposition \(|\psi\rangle = \frac{1}{\sqrt{n}}\sum_{i=0}^{n-1}|i\rangle \otimes |00\rangle\)
Measurement: computes the probability \(\mathbb{P}_m^{n,\bm{\theta}}\) of the target qubit being in a specified state
State map: \(\bar{F}_{R,j}^{n,\bm{\theta}}(\bm{x},\bm{z}) = R - 2R[\mathbb{P}_1^{n,\bm{\theta}^j} + \mathbb{P}_2^{n,\bm{\theta}^j}]\), equivalent to \(\frac{1}{n}\sum_{i=1}^n R\cos(\gamma^{i,j})\cos(b^{i,j} + \bm{a}^{i,j}\cdot(\bm{x},\bm{z}))\)

The overall system forms a recurrent structure via feedback \(\hat{\bm{x}}_t = \bar{F}_R^{n,\bm{\theta}}(\hat{\bm{x}}_{t-1}, \bm{z}_t)\).

Quantum Resource Analysis¶

Each circuit acts on \(\mathfrak{n} = \lceil\log_2(2n)\rceil\) qubits
\(n\) is the precision parameter (determining the number of blocks); qubit count grows only logarithmically
Achieving approximation accuracy \(\varepsilon\) requires \(\mathcal{O}(\varepsilon^{-2})\) weights and \(\mathcal{O}(\lceil\log_2(\varepsilon^{-1})\rceil)\) qubits

Main Theoretical Results¶

Theorem 4.6 (State-space system approximation bound): For a state map \(F\) satisfying a Barron-type integrability condition with contraction coefficient \(\lambda < 1\), the uniform approximation error of the RQNN filter satisfies:

\[\sup_{\bm{z}}\sup_t \|U^F(\bm{z})_t - \bar{U}(\bm{z})_t\| \leq \frac{1}{1-\lambda}\frac{\sqrt{N}\max_j C_j^\infty}{\sqrt{n}}\]

The error decays at rate \(1/\sqrt{n}\), independent of input dimension \(d\) and state dimension \(N\) (no curse of dimensionality)
Key technique: the QNN simultaneously approximates a function and its derivatives (Proposition 4.4), enabling controllable error propagation through the feedback loop

Theorem 4.8 (Universal approximation): For any causal, time-invariant, fading memory filter \(U\), there exist RQNN parameters and a linear readout \(W\) such that:

\[\sup_{\bm{z}}\sup_t \|U(\bm{z})_t - \bar{U}_W(\bm{z})_t\| \leq \varepsilon\]

No Barron integrability condition or contractivity assumption is required
The echo state property is ensured by introducing a linear preprocessing matrix \(P_j\) and a finite-step memory partition

Proof Strategy¶

An internal approximation approach is adopted: first, approximation bounds for static functions and their derivatives by QNNs are established → derivative control is then used to bound error accumulation in the feedback loop → filter approximation bounds are derived from state-map approximation bounds.

Key Experimental Results¶

This is a purely theoretical work with no numerical experiments. The core quantitative contributions are the constants and convergence rates in the approximation error bounds.

Highlights & Insights¶

First quantitative approximation bounds for RQNNs: fills a critical gap in quantum RNN approximation theory
No curse of dimensionality: the error rate \(1/\sqrt{n}\) is independent of input and state dimensions; qubit count grows only logarithmically
Linear readout suffices for universality: polynomial readout layers are unnecessary, substantially reducing the difficulty of experimental implementation
Weaker conditions than classical RNNs: the required Sobolev smoothness condition \(s > \frac{N+d}{2}+4\) is weaker than \(s > N+d+3\) needed for classical RNNs
Hardware compatibility: circuits based on uniformly controlled gates admit efficient decompositions and Rydberg atom implementations

Limitations & Future Work¶

Theory only: the absence of numerical validation makes it difficult to assess performance on actual NISQ devices
Barron-type condition restricts scope: quantitative bounds apply only to functions with sufficiently integrable Fourier transforms; no error rates are available for rough or non-contractive dynamics
Barren plateau problem unaddressed: gradient vanishing in variational quantum circuit training may significantly affect practical trainability
Monte Carlo sampling error: finite-sample measurement error is only briefly discussed in the appendix and is not incorporated into the main approximation bounds
No comparison with random reservoirs: results apply only to the fully trainable variational setting and have not been extended to genuine QRC with partially random parameters

Method	Linear Readout Universality	Quantitative Error Bound	No Curse of Dimensionality	Feedback / Sequential
Classical ESN (Grigoryeva & Ortega 2018)	✗	✗	-	✓
Classical RNN (Gonon et al. 2023)	✓	✓	✓	✓
Feedforward QNN (Gonon & Jacquier 2025)	-	✓	✓	✗
QRC polynomial readout (Sannia et al. 2024)	✗	✗	-	✓
Ours (RQNN)	✓	✓	✓	✓

Compared to classical RNNs, RQNNs impose weaker smoothness requirements on the target function. Compared to feedforward QNNs, this work addresses the additional analytical challenges introduced by feedback loops. Compared to existing QRC universality results, this paper is the first to achieve linear readout universality with quantitative bounds.

The cosine basis expansion form of RQNNs (Proposition 4.1) suggests a deep connection to classical random feature methods. The technique of combining linear preprocessing with finite-memory partitioning to guarantee the echo state property may offer design guidance for practical QRC architectures. The trade-off between barren plateaus and expressive power remains a key bottleneck for practical deployment and warrants further investigation.

Rating¶

Novelty: ⭐⭐⭐⭐☆ — First complete approximation theory for feedback-driven RQNNs
Experimental Thoroughness: ⭐⭐☆☆☆ — Purely theoretical; no experimental validation
Writing Quality: ⭐⭐⭐⭐☆ — Clear structure with precise statement of main results
Value: ⭐⭐⭐⭐☆ — Establishes theoretical foundations for sequential tasks in quantum machine learning