Estimates of loss function concentration in noisy parametrized quantum circuits

TL;DR

This work tackles the problem of loss function concentration in noisy parametrized quantum circuits by introducing a non-negative matrix theory framework based on locality transfer matrices (LTMs). It derives an exact variance expression for layered circuits with arbitrary noise, provides a deep-circuit limit that reveals an absorption mechanism, and establishes lower bounds for shallow circuits, linking noise resilience to circuit expressivity and initialization. The authors also connect these insights to initialization strategies, proposing QResNets that can mitigate barren plateaus even in non-unitary settings. The results offer practical guidance for designing near-term quantum devices and fault-tolerant approaches by clarifying how noise and unitary layers interact to shape variational optimization landscapes.

Abstract

Variational quantum computing offers a powerful framework with applications across diverse fields such as quantum chemistry, machine learning, and optimization. However, its scalability is hindered by the exponential concentration of the loss function, known as the barren plateau problem. While significant progress has been made and prior work has separately analyzed barren plateaus in unitary and noisy settings, their combined impact remains poorly understood, largely due to limitations in conventional Lie-algebraic approaches. In this work, we introduce a novel analytical framework based on non-negative matrix theory that enables the description of the variance in layered noisy quantum circuits with arbitrary noise channels. This approach enables the derivation of exact expressions in the deep-circuit regime, uncovering the complex interplay between unitary layers and noise. Notably, we identify a noise-induced absorption mechanism-a phenomenon absent in purely unitary dynamics-which provides new insight into how noise shapes circuit behavior. We further present a controlled convergence analysis, establishing general lower bounds on the variance of both deep and shallow circuits. This leads to a principled connection between noise resilience and the expressive capacity of parameterized quantum circuits, particularly under smart initialization strategies. Our theoretical results are supported by numerical simulations and illustrative applications.

Figures (8)

• Figure 1: Loss concentration in variational quantum computing. The analytical formulation proposed here employs non-negative matrix theory to describe the interplay between local 2-designs and noise. This allows for precise calculation of loss variances for generic noise maps, from strictly contractive to unitary channels, as illustrated by the coloured band in the Figure. The upper part considers deep circuits, where the loss variance $\mathbb{V}_{\rho,H}^L$ reaches its asymptotic limit. While loss concentration is well-understood for strictly contractive Schumann23Wang21 and unitary Ragone24Fontana24Diaz23 channels, \ref{['thm:main']} provides an analytical solution for the intermediate case, revealing a noise-induced absorption mechanism unique to this regime. Consistency with known limiting cases is verified in \ref{['sec:methods:recover_limits']}. The lower part focuses on "shallow" circuits, where a lower bound on $\mathbb{V}_{\rho,H}^L$ is established using \ref{['thm:lower_bound']}. This extends previous works on brickwork circuits Cerezo21_2Mele24, enabling initialization strategies such as small angle initializations Zhang22Wang23 to be represented as stochastic unravellings of noise maps. In \ref{['subsect:small_angles']} we expand upon their applicability, showing how both unitary and non-unitary QResNets can be derived thanks to \ref{['prop:noise_small_angle_equiv']}.
• Figure 2: Graphical representations of the stochastic process describing $\mathbb{V}_{\rho,H}^L$. On the left, we show the structure of the general locality transfer matrix (LTM), highlighting the decomposition into irreducible components Seneta06. Light blue blocks represent irreducible, essential components of $T$, while red blocks are related to inessential ones. In particular, $Q$ represents the collection of all irreducible, inessential components of $T$ and $R$ their relation with the essential components $T_z$. On the right, the same process is represented graphically, in terms of the local subspaces ${\mathcal{B}}_\kappa$. Here, each dot represents a single subspace, while the arrows represent the adjoint action of the channel $\mathcal{E}$. Essential and inessential components share the same colour code of $T$.
• Figure 3: Single layer structure used in the numerical examples. The circuit acts on a two-qubit register and is divided into three blocks, each representing either unitary gates ($\mathcal{U}$) or noise channels ($\mathcal{E}$ and $\mathcal{F}$).
• Figure 4: Example of NIBP-free unital, non-unitary channel. Dotted lines represent the theoretical deep-circuit limiting values for the variance in the case of layers defined by $\mathcal{U}\circ\mathcal{E}$ (blue) and $\mathcal{U}'\circ\mathcal{E}$ (pink), where $\mathcal{U}'$ is obtained from $\mathcal{U}$ in \ref{['fig:circ_example']} by substituiting the first $RX$ gate with an $RY$ gate. The solid pink line is an exponential fit of the convergence to the limiting value.
• Figure 5: Variance improvment with unital channels. Dotted lines represent the theoretical prediction given by \ref{['eq:general_formula']}, while points indicate the numerically estimated circuit variances.

Theorems & Definitions (68)

• Definition 3.1: Locality transfer matrix
• Proposition 4.1: General formula
• Theorem 4.1: Deep circuit variance
• Corollary 4.1
• Theorem 4.2: General lower bound
• Corollary 4.2: Lower bound examples
• Proposition 4.2: QResNet
• Proposition 4.3: Noise map and QResNet
• Corollary A.1: Deep, unitary circuits
• Corollary A.2: Deep, noisy circuits