Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep-Circuit QAOA

Gereon Koßmann, Lennart Binkowski, Lauritz van Luijk, Timo Ziegler, René Schwonnek

TL;DR

This paper investigates the deep-circuit regime of QAOA by shifting attention from parameter landscapes to the geometry of the accessible state space, using a Lie-group framework to analyze optimization landscapes. It proves that, in the asymptotic limit with universal generators, critical points align with eigenstates of the encoded problem and the global minimum corresponds to the optimal classical solution, while deep circuits introduce local attractors and traps that complicate optimization. The authors introduce the mu–f diagram as a practical, data-driven performance indicator that depends only on the classical objective function and its local-neighborhood structure, enabling prediction of when deep QAOA is favorable (e.g., for random QUBOs) versus when it suffers from no-free-lunch behavior. Numerical experiments with simple local-search routines confirm these insights, showing favorable performance on QUBO-like problems and obstacles on random, structureless objectives. Overall, the work provides a geometric lens and a concrete diagnostic tool to assess the potential of deep-circuit QAOA and delineates the constraints under which it may offer practical advantages.

Abstract

Despite its popularity, several empirical and theoretical studies suggest that the quantum approximate optimization algorithm (QAOA) has persistent issues in providing a substantial practical advantage. Numerical results for few qubits and shallow circuits are, at best, ambiguous, and the well-studied barren plateau phenomenon draws a rather sobering picture for deeper circuits. However, as more and more sophisticated strategies are proposed to circumvent barren plateaus, it stands to reason which issues are actually fundamental and which merely constitute - admittedly difficult - engineering tasks. By shifting the scope from the usually considered parameter landscape to the quantum state space's geometry we can distinguish between problems that are fundamentally difficult to solve, independently of the parameterization, and those for which there could at least exist a favorable parameterization. Here, we find clear evidence for a 'no free lunch'-behavior of QAOA on a general optimization task with no further structure; individual cases have, however, to be analyzed more carefully. Based on our analysis, we propose and justify a performance indicator for the deep-circuit QAOA that can be accessed by solely evaluating statistical properties of the classical objective function. We further discuss the various favorable properties a generic QAOA instance has in the asymptotic regime of infinitely many gates, and elaborate on the immanent drawbacks of finite circuits. We provide several numerical examples of a deep-circuit QAOA method based on local search strategies and find that - in alignment with our performance indicator - some special function classes, like QUBOs, indeed admit a favorable optimization landscape.

Deep-Circuit QAOA

TL;DR

This paper investigates the deep-circuit regime of QAOA by shifting attention from parameter landscapes to the geometry of the accessible state space, using a Lie-group framework to analyze optimization landscapes. It proves that, in the asymptotic limit with universal generators, critical points align with eigenstates of the encoded problem and the global minimum corresponds to the optimal classical solution, while deep circuits introduce local attractors and traps that complicate optimization. The authors introduce the mu–f diagram as a practical, data-driven performance indicator that depends only on the classical objective function and its local-neighborhood structure, enabling prediction of when deep QAOA is favorable (e.g., for random QUBOs) versus when it suffers from no-free-lunch behavior. Numerical experiments with simple local-search routines confirm these insights, showing favorable performance on QUBO-like problems and obstacles on random, structureless objectives. Overall, the work provides a geometric lens and a concrete diagnostic tool to assess the potential of deep-circuit QAOA and delineates the constraints under which it may offer practical advantages.

Abstract

Despite its popularity, several empirical and theoretical studies suggest that the quantum approximate optimization algorithm (QAOA) has persistent issues in providing a substantial practical advantage. Numerical results for few qubits and shallow circuits are, at best, ambiguous, and the well-studied barren plateau phenomenon draws a rather sobering picture for deeper circuits. However, as more and more sophisticated strategies are proposed to circumvent barren plateaus, it stands to reason which issues are actually fundamental and which merely constitute - admittedly difficult - engineering tasks. By shifting the scope from the usually considered parameter landscape to the quantum state space's geometry we can distinguish between problems that are fundamentally difficult to solve, independently of the parameterization, and those for which there could at least exist a favorable parameterization. Here, we find clear evidence for a 'no free lunch'-behavior of QAOA on a general optimization task with no further structure; individual cases have, however, to be analyzed more carefully. Based on our analysis, we propose and justify a performance indicator for the deep-circuit QAOA that can be accessed by solely evaluating statistical properties of the classical objective function. We further discuss the various favorable properties a generic QAOA instance has in the asymptotic regime of infinitely many gates, and elaborate on the immanent drawbacks of finite circuits. We provide several numerical examples of a deep-circuit QAOA method based on local search strategies and find that - in alignment with our performance indicator - some special function classes, like QUBOs, indeed admit a favorable optimization landscape.
Paper Structure (23 sections, 6 theorems, 38 equations, 11 figures)

This paper contains 23 sections, 6 theorems, 38 equations, 11 figures.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Variational Quantum Algorithms
  4. The Deep-Circuit Regime
  5. Optimization Landscapes
  6. Barren Plateaus
  7. Asymptotic Circuits
  8. The Geometry of Accessible States
  9. The Lie Algebra of the Encoded Problem
  10. Classification of Critical Points
  11. Uniqueness of Minima and Maxima
  12. Deep Circuits
  13. Local Attractors and Traps
  14. Trap Sizes
  15. A Performance Indicator
  16. ...and 8 more sections

Key Result

theorem 1

$B$ and $C$ together form universal generators of QAOA-gates if the underlying classical optimization problem given by a target function $f$ fulfills the conditions In particular, the set of optimization problems $f$ for which $\mathfrak{g}=\mathfrak{su}(2^{N})$ is open and dense (and hence the complement is a null set).

Figures (11)

  • Figure 1: The geometry of the QAOA visualized on a qubit: Here, $\Omega$ is given by the surface of the Bloch sphere (all pure qubit states). Applying the basic gates $U_C$ or $U_B$ corresponds to movements along vector fields $\Phi_C$ or $\Phi_B$ that are oriented along lines of constant latitudes (left). Performing a QAOA sequence (coral colored line) corresponds to alternatingly move along these vector fields.
  • Figure 2: Schematic visualization of different troughs without any computational state, including a local extremum $\ketbra{z_{0}}$, and the optimal state $\ketbra{z^{*}}$, respectively. Cyan lines indicate orbits of movements in $i C$ direction. They are also the level sets of the functional. Blue lines indicate orbits of movements in $i B$ direction. A component of the trough, i.e., a set of states with vanishing first and positive second derivative in $i B$ direction, is marked with red.
  • Figure 3: Schematic visualization of a valley (troughs in higher dimensions not indicated): Driving with $i C$ (cyan lines), does not change the value of the functional. Driving with $i B$ is in some sense the 'orthogonal' direction. In a valley, all $i B$ trajectories have a local minimum. We identify the size of a valley by the region in which the second derivative is positive.
  • Figure 4: $\mu$ -$f$ diagrams of some randomly sampled functions $f$ on $13$ bits. The values of $f$ are (a) uniformly randomly distributed with support on $[- 1, 1]$ and (b) distributed with respect to a bimodal distribution that favors values at the boundaries of $[- 1, 1]$. From a local search perspective these are considered unfavorable instances. (c), on the contrary, stems from a random QUBO instance. This diagram unveils an optimization landscape that is favorable for QAOA.
  • Figure 5: Minimization of a function with uniformly randomly distributed values (step size $\varepsilon = 0.1$). As the $\mu$ -$f$ diagram (\ref{['figure:MuFRandom']}(a)) suggests, layer-wise optimization struggles to find an optimum. Both, the success probability and the approximation ratio increase only very slowly, and the final state spreads over many computational basis states, with the focus on a non optimal state.
  • ...and 6 more figures

Theorems & Definitions (9)

  • theorem 1
  • proof
  • proposition 2
  • proof
  • corollary 3
  • corollary 4
  • proposition 5
  • proof
  • corollary 6