Information-Theoretic Constraints on Variational Quantum Optimization: Efficiency Transitions and the Dynamical Lie Algebra
By: Jun Liang Tan
Potential Business Impact:
Makes quantum computers learn better by controlling information.
Variational quantum algorithms are the leading candidates for near-term quantum advantage, yet their scalability is limited by the ``Barren Plateau'' phenomenon. While traditionally attributed to geometric vanishing gradients, we propose an information-theoretic perspective. Using ancilla-mediated coherent feedback, we demonstrate an empirical constitutive relation $ΔE \leq ηI(S:A)$ linking work extraction to mutual information, with quantum entanglement providing a factor-of-2 advantage over classical Landauer bounds. By scaling the system size, we identify a distinct efficiency transition governed by the dimension of the Dynamical Lie Algebra. Systems with polynomial algebraic complexity exhibit sustained positive efficiency, whereas systems with exponential complexity undergo an ``efficiency collapse'' ($η\to 0$) at $N \approx 6$ qubits. These results suggest that the trainability boundary in variational algorithms correlates with information-theoretic limits of quantum feedback control.
Similar Papers
Information-Theoretic Constraints on Variational Quantum Optimization: Efficiency Transitions and the Dynamical Lie Algebra
Quantum Physics
Makes quantum computers learn better by fixing information flow.
Trainability of Quantum Models Beyond Known Classical Simulability
Quantum Physics
Keeps quantum computers trainable for harder problems.
Trainability of Quantum Models Beyond Known Classical Simulability
Quantum Physics
Makes quantum computers learn without getting stuck.