Architecture Shape Governs QNN Trainability: Jacobian Null Space Growth and Parameter Efficiency

Michael Poppel, David Bucher, Maximilian Zorn, Markus Baumann, Sebastian Wölckert, Claudia Linnhoff-Popien, Philipp Altmann, Jonas SteinarXiv 2026 · arXiv

TL;DR

Quantum Neural Networks (QNNs) sharing the exact same encoding budget can exhibit wildly different training behaviors based purely on their architectural "shape" (serial vs. parallel). This paper proves that serial (single-qubit) architectures suffer from a fundamental bottleneck called structural gradient starvation, where adding more parameters fails to improve training because they get trapped in the Jacobian's null space. Parallel architectures avoid this, offering a $1.6\times$ to $2.2\times$ boost in parameter efficiency when scaled by adding encoding layers rather than just trainable blocks.

What Problem Does This Paper Address?

When designing Quantum Neural Networks (QNNs) using angle encoding, a common strategy is data re-uploading. If you keep the total encoding budget $E = N \times L$ fixed (where $N$ is the number of qubits and $L$ is the number of encoding layers per qubit), the network's theoretical expressivity—its frequency spectrum and redundancy—remains identical. For example, 1 qubit with 12 layers covers the exact same function class as 12 qubits with 1 layer.

However, in practice, their trainability is completely different. Single-qubit or low-qubit deep architectures often fail to converge entirely, while multi-qubit parallel architectures succeed effortlessly with far fewer parameters. This paper uncovers the mathematical mechanism behind this discrepancy, ruling out expressivity or spectral bias as the culprits.

Figure 1: Trainability of data re-uploading QNNs at fixed encoding budget E = NL = 12, trained with Adam on 10 randomly drawn degree-12 Fourier targets. Left: success rate (R2 ≥0.95) vs. parameter count P. Right: mean R2 (test) vs. P. All architectures share identical frequency spectra and redundancy structure, ruling out expressivity and spectral bias as explanations.

Methods

The authors analyze the coefficient matching Jacobian ($J$), which maps changes in ansatz parameters to changes in the Fourier coefficients of the QNN's output function.

  • Rank Ceiling Proof: They mathematically prove that for serial ($N=1$) architectures, the rank of the Jacobian is strictly bounded by $2L + 1$, regardless of how many trainable parameters $P$ you add. This causes the dimension of the null space ($\dim(\ker J)$) to grow linearly with $P$, a phenomenon they term structural gradient starvation.
  • Bilinear Factorization: They show that parallel architectures break this limitation because different qubits track independent phase trajectories, preventing the "phase-locking" that cripples serial circuits.
  • Scaling Analysis: They compare two ways to add parameters to a parallel architecture:
  • The FM route: Increasing the number of feature map (encoding) layers $L$.
  • The Trainable Blocks route: Increasing the number of trainable layers ($tbl$) while keeping encoding fixed.

Datasets

  • Synthetic Fourier Targets: Randomly generated Fourier series of fixed degrees ($d \in \{6, 12, 20, 28\}$) with coefficients drawn from a normal distribution.
  • Real-World Data: The Nottingham temperature dataset, consisting of 240 monthly mean air temperatures recorded at Nottingham Castle from 1920 to 1939.

Key Insights and Ideas

  • Phase-Locking Destroys Gradients: In single-qubit (serial) architectures, a geometric constraint forces all parameter gradient directions to share a common global phase. This locks the Jacobian rank, meaning any extra parameters added to the circuit are structurally decoupled from the loss function and receive a flat zero gradient signal.
  • Not All Parameters Are Created Equal: When scaling a parallel QNN, adding feature map (FM) layers provides a quantum-specific benefit by expanding the accessible Fourier spectrum and lifting the Jacobian rank ceiling.
  • The Classical Interpolation Trap: Conversely, adding more trainable blocks without adding encoding layers keeps the rank ceiling frozen. It only improves training through classical overparameterization (the interpolation threshold where $P \ge n_{\text{train}}$), offering no unique quantum advantage.

Key Results

  • Parameter Efficiency: The FM route achieves excellent training performance ($R^2 \ge 0.95$) using $1.6\times$ to $2.2\times$ fewer parameters than the trainable blocks route across all tested architectures.
  • Null Space Growth: For serial circuits, as $P$ increases, the fraction of parameters stuck in the null space approaches 100% ($(P - \text{rank}(J))/P \to 1$), confirming severe gradient starvation.
  • Real-World Validation: On the Nottingham temperature dataset, the trainable blocks route failed completely ($R^2 < 0$) due to an expressivity ceiling, whereas the FM route succeeded by expanding both expressivity and Jacobian rank. However, for $N=6$ qubits, an inversion occurs where the circuit enters the underparameterized "Larocca regime," making trainable blocks temporarily more efficient.

Limitations and Future Work

  • Architecture Extremes: The strict theoretical proofs are derived for the extremes: $N=1$ serial circuits and completely unentangled parallel product ansätze. Extending these rigorous proofs to hybrid architectures and complex entangling layers remains an open challenge.
  • The Underparameterized Boundary: The parameter efficiency advantage of the FM route holds below the geometric threshold ($R_{\max} = 2^{N+1}-2$). The exact trade-off dynamics inside the highly underparameterized, Larocca-limited regime require a more complete characterization.

Summary updated

Abstract

Variational quantum circuits with angle encoding implement truncated Fourier series, and architectures arranging $N$ qubits with $L$ encoding layers each -- sharing encoding budget $E = NL$ -- generate identical frequency spectra, identical frequency redundancy, and require the same minimum parameter count for coefficient control. Despite this equivalence, trainability varies substantially with architecture shape $(N,L)$ at fixed $E$. We identify structural rank deficiency of the coefficient matching Jacobian $J$ as the mechanism responsible. For serial single-qubit architectures, we prove $\mathrm{rank}(J) \leq 2L+1$ regardless of parameter count $P$, with $\dim(\ker J) \geq P-(2L+1)$ growing without bound -- a phenomenon we term \emph{structural gradient starvation}: a growing fraction of parameters become structurally decoupled from the loss as $P$ increases at fixed $L$. Parallel architectures avoid this via independent phase trajectories, ensuring $σ_{\min}(J^{(\mathrm{par})}) > 0$ generically for $P \leq 2E+1$, so no parameter lies in $\ker J$. For practitioners, we further show that the two natural routes to increasing parameter count have fundamentally different effects: adding feature map (FM) layers monotonically strengthens the Jacobian QFIM eigenvalue spectrum and achieves $R^2 \geq 0.95$ with $1.6$--$2.2\times$ fewer parameters than adding trainable blocks across all tested architectures, while trainable blocks improve training only through the classical interpolation mechanism with no quantum-specific benefit.
BibTeX
@misc{poppel2026architecture,
  author = {Michael Poppel and David Bucher and Maximilian Zorn and Markus Baumann and Sebastian Wölckert and Claudia Linnhoff-Popien and Philipp Altmann and Jonas Stein},
  title = {{Architecture Shape Governs QNN Trainability: Jacobian Null Space Growth and Parameter Efficiency}},
  year = {2026},
  eprint = {2605.05942},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.05942},
}
Figure