Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quantifying the barren plateau phenomenon for a model of unstructured variational ansätze

John Napp

TL;DR

This work addresses barren plateaus in unstructured hardware-efficient ansätze for variational quantum algorithms by mapping the typical landscape flatness to a biased random-walk process and providing a Monte Carlo estimator for the typical gradient magnitude. It derives general lower and upper bounds on the gradient magnitude, showing exponential decay in circuit depth and Hamiltonian locality and polynomial decay in local dimension, with sharper 1D bounds. A key contribution is the Monte Carlo algorithm, whose runtime is $O\left(m\,(\sum_{\mathbf{x}}|c_{\mathbf{x}}|^2)^2\log(1/\delta)\,\varepsilon^{-2}\right)$, plus a domain-wall–based analysis that clarifies the roles of combinatorial versus spatial locality. The results illuminate trainability limits of HEAs, explain why global observables are prone to barren plateaus, and provide practical guidance for designing optimization strategies and architectures for VQAs.

Abstract

Quantifying the flatness of the objective-function landscape associated with unstructured parameterized quantum circuits is important for understanding the performance of variational algorithms utilizing a "hardware-efficient ansatz", particularly for ensuring that a prohibitively flat landscape -- a so-called "barren plateau" -- is avoided. For a model of such ansätze, we relate the typical landscape flatness to a certain family of random walks, enabling us to derive a Monte Carlo algorithm for efficiently, classically estimating the landscape flatness for any architecture. The statistical picture additionally allows us to prove new analytic bounds on the barren plateau phenomenon, and more generally provides novel insights into the phenomenon's dependence on the ansatz depth, architecture, qudit dimension, and Hamiltonian combinatorial and spatial locality. Our analysis utilizes techniques originally developed by Dalzell et al. to study anti-concentration in random circuits.

Quantifying the barren plateau phenomenon for a model of unstructured variational ansätze

TL;DR

This work addresses barren plateaus in unstructured hardware-efficient ansätze for variational quantum algorithms by mapping the typical landscape flatness to a biased random-walk process and providing a Monte Carlo estimator for the typical gradient magnitude. It derives general lower and upper bounds on the gradient magnitude, showing exponential decay in circuit depth and Hamiltonian locality and polynomial decay in local dimension, with sharper 1D bounds. A key contribution is the Monte Carlo algorithm, whose runtime is , plus a domain-wall–based analysis that clarifies the roles of combinatorial versus spatial locality. The results illuminate trainability limits of HEAs, explain why global observables are prone to barren plateaus, and provide practical guidance for designing optimization strategies and architectures for VQAs.

Abstract

Quantifying the flatness of the objective-function landscape associated with unstructured parameterized quantum circuits is important for understanding the performance of variational algorithms utilizing a "hardware-efficient ansatz", particularly for ensuring that a prohibitively flat landscape -- a so-called "barren plateau" -- is avoided. For a model of such ansätze, we relate the typical landscape flatness to a certain family of random walks, enabling us to derive a Monte Carlo algorithm for efficiently, classically estimating the landscape flatness for any architecture. The statistical picture additionally allows us to prove new analytic bounds on the barren plateau phenomenon, and more generally provides novel insights into the phenomenon's dependence on the ansatz depth, architecture, qudit dimension, and Hamiltonian combinatorial and spatial locality. Our analysis utilizes techniques originally developed by Dalzell et al. to study anti-concentration in random circuits.
Paper Structure (15 sections, 10 theorems, 36 equations, 1 figure, 1 algorithm)

This paper contains 15 sections, 10 theorems, 36 equations, 1 figure, 1 algorithm.

Table of Contents

  1. Introduction
  2. Setup and notation
  3. Monte Carlo algorithm
  4. General lower bound
  5. General upper bound
  6. Improved bounds for 1D architecture
  7. Discussion
  8. Proof of \ref{['lem:Zeroth_First_Inequalities']}
  9. Relating $\mathop{\mathrm{\mathbb{E}}}\limits_V \mathop{\mathrm{\mathbb{E}}}\limits_\mathbf{\uptheta} f_V (\mathbf{\uptheta})^2$ to random walks (proofs of \ref{['lem:expansion']}, \ref{['thm:biased']}, and \ref{['alg']})
  10. Expressing $g_\mathbf{x}$ as a biased random walk
  11. Expressing $g_\mathbf{x}$ as an unbiased random walk
  12. Proofs of corollaries
  13. \ref{['cor:general_upper']}
  14. \ref{['cor:1d_lower']}
  15. \ref{['cor:1d_upper']}

Key Result

Lemma 1

With $f_V(\mathbf{\uptheta})$ as defined previously,

Figures (1)

  • Figure 1: Example of an $n=6$ qudit architecture compatible with the model. Time flows from bottom to top. Orange represents entangling gates, blue represents arbitrary parameterized gates, and pink represents the final parameterized gates to act as specified in the main text. In this example, there are $m=8$ entangling gates, there are $p=36$ variational parameters, the entangling gate depth is $d=3$, and the entangling gate regular connectivity is $r=2$. While a 1D architecture is illustrated for simplicity, no such constraint is required for the Monte Carlo algorithm or general analytic bounds, although stronger bounds are obtained for the 1D setting.

Theorems & Definitions (11)

  • Lemma 1
  • Lemma 2
  • Theorem 1
  • Corollary 1
  • Corollary 2
  • Corollary 3
  • Corollary 4
  • Corollary 5
  • Lemma 3: Follows directly from Corollary 2 of dalzell2020random
  • Lemma 4
  • ...and 1 more