arXiv:2610.21778
Same Spectrum, Harder First-Order Training: Phase-Locking in Single- versus Multi-Qubit Re-Uploading Circuits
Michael Poppel, David Bucher, Maximilian Zorn, Markus Baumann, Sebastian Wölckert, Claudia Linnhoff-Popien, Philipp Altmann, Jonas Stein
Abstract
Angle-encoded variational quantum circuits implement truncated Fourier series. For one-dimensional inputs, a single-qubit re-uploading (serial) circuit represents every function that a multi-qubit circuit with the same encoding budget $E$ can; we find that this exchange of width for depth is not free in first-order training. Because its state lives on the Bloch sphere, all parameters of the serial circuit move the state, at each input, along one complex direction, which we call \emph{phase-locking}. We prove that this costs no Fourier direction but leaves no direction to spare: however many parameters it has, the serial circuit has at most $2E+2$ tangent directions, one more than its $2E+1$ Fourier coefficients need, while a two-qubit control with the same encoding budget and model class generically has at least $6E+6$. The cost in training is measured against this control, as the number of runs out of 100 that reach the optimum within a fixed step budget. On an exact loss with flat-spectrum targets the gap grows with $E$; at $E=32$ the serial circuit converges in 3, 53 and 83 runs under gradient descent, Adam and Levenberg--Marquardt, the control in 99, 100 and 100. On three real-world series, at matched parameter counts, the serial circuit reaches the least-squares optimum in at most 9 runs under Adam and in 32 to 93 under Levenberg--Marquardt, the control in at least 98 and in all 100; its held-out error is up to about $1.5$ times that of the control under Adam, but no larger in the typical run under Levenberg--Marquardt.
AI Summary
For univariate angle-encoded variational circuits, does exchanging qubit width for re-uploading depth preserve trainability when expressivity is unchanged? The paper compares a single-qubit serial circuit with a two-qubit ancilla control having the same encoding budget $E$, chain length, and model class: all real trigonometric polynomials of degree at most $E$ bounded by one. A two-qubit parallel circuit additionally probes chain length. The structural analysis identifies “phase-locking”: at each input, all single-qubit state derivatives orthogonal to the state lie along one complex direction. Across inputs, the serial circuit has generically exactly $2E+2$ independent real tangent directions, versus at least $6E+6$ for the ancilla control. Both coefficient Jacobians generically have full row rank $2E+1$; the serial circuit’s sole output-preserving tangent direction is a readout rotation. These results follow from tangent-space decompositions, Fourier-degree arguments, explicit rank constructions, and genericity arguments. Noise-free simulations compare gradient descent, Adam, and Levenberg–Marquardt. For exactly representable flat-spectrum targets at $E=32$, convergence within fixed budgets occurs in 3, 53, and 83 of 100 serial runs, respectively, versus 99, 100, and 100 control runs. Matched-parameter data experiments also show lower serial convergence rates. On three real-world series, Adam’s held-out error is up to about $1.5$ times the control’s; typical Levenberg–Marquardt predictions are comparable. The paper does not establish that phase-locking causes the training gap. Its scope is univariate inputs, Haar-random initialisation, fixed optimisation budgets, and exact statevector simulation; hardware costs, shot noise, native angle parameterisations, and encoding budgets exceeding target degree are not tested.
The proofs were developed with the automated prover Aristotle and are checked by the Lean kernel.
Error found: The abstract claims no increase in typical held-out error under Levenberg–Marquardt, but the reported median serial/control error ratio is 1.01 for nottem at E=24, supporting approximate equality rather than non-increase.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.