Regularizing quantum loss landscapes by noise injection

Daniil S. Bagaev, Maxim A. Gavreev, Alena S. Mastiukova, Aleksey K. Fedorov, Nikita A. Nemkov · 2025

Read Regularizing quantum loss landscapes by noise injection by Daniil S. Bagaev, Maxim A. Gavreev, Alena S. Mastiukova, Aleksey K. Fedorov, Nikita A. Nemkov (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-08759.txt.

Introduction

Bagaev and four co-authors tame the jagged loss landscapes that stall quantum machine learning, by deliberately injecting noise, in a paper from twenty twenty-five. Variational quantum algorithms train by descending a landscape riddled with poor local minima, making training provably hard in general. Practical heuristics, though, may still work on typical cases.

Their protocol adds targeted noise that exponentially damps the high-frequency ripples in the loss function's wave expansion, smoothing the terrain. It runs efficiently both on real hardware and in simulation. Across problem types, solution quality improved significantly and robustly. The method pairs with existing tricks such as the quantum natural gradient, adding a fresh tool to the quantum optimization kit.

Catalog record

CREATORDaniil S. Bagaev, Maxim A. Gavreev, Alena S. Mastiukova, Aleksey K. Fedorov, Nikita A. Nemkov
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-08759

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Regularizing quantum loss landscapes by noise injection Daniil S. Bagaev, Maxim A. Gavreev, Alena S. Mastiukova, Aleksey K. Fedorov, Nikita A. Nemkov Abstract The difficulty of training variational quantum algorithms and quantum machine learning models is well established. In particular, quantum loss landscapes are often highly non-convex and dominated by poor local minima. While this renders their training NP-hard in general, efficient heuristics that work well for typical instances may still exist. Here, we propose a protocol that uses a targeted noise injection to smooth and regularize quantum loss landscapes. It works by exponentially suppressing the high-frequency components in the Fourier expansion of the quantum loss function. The protocol can be efficiently implemented both in hardware and in simulations. We observe significant and robust improvements of solution quality across various problem types. Our method can be combined with existing techniques mitigating the local minima, such as the quantum natural gradient optimizer, and adds to the toolbox of methods for optimizing quantum loss functions. I Introduction and results Variational quantum algorithms (VQA) Cerezo2021 and quantum machine learning (QML) models Biamonte2017 ; Schuld2020 are generalizations of classical optimization methods and machine learning (ML) with the potential to harness quantum effects. The range of problems that can be formulated as VQA or QML is extremely broad. However, similarly to classical deep learning, studying VQA and QML remains largely an empirical effort, while firm theoretical performance guarantees are scarce. One of the key problems facing VQA and QML is trainability. Two main issues here are the onset of barren plateaus (BPs) McClean2016 ; Larocca2024 and proliferation of poor local minima. The BPs essentially manifest the curse of dimensionality, and most often arise in sufficiently deep circuits. Local minima, on the other hand, are typically associated with shallow quantum circuits, which may be free of BPs Pesah_2021 . (Note, however, that the absence of BPs is often associated with classical simulability Cerezo2023 .) This work focuses on the problem of local minima in quantum loss landscapes. On the one hand, results such as Anschuetz2021 ; Anschuetz2022 ; Bittel2021 show that the local minima problem in VQA and QML is intractable in the general case. On the other hand, there may still exist heuristics allowing for efficient training of typical instances despite the no-go results for a general worst-case. After all, training deep neural networks can be very efficient, even though the associated loss functions are typically highly non-convex Goodfellow2016 . In this work, we propose a heuristic regularization procedure that smoothes the quantum loss landscapes and alleviates the problem of local minima. It is based on the intuition that high-frequency terms in the Fourier expansion of the loss function are primarily responsible for the superfluous local minima, see illustration in Fig. 1 a. Note that this is merely a sketch and cannot faithfully represent high-dimensional VQA landscapes. Also, Fig. 1 b depicts an opposite situation, where the high-frequency modes conspire to produce a global minimum far from the global minimum of the regularized function. Essentially, our conjecture is that the latter scenario is less likely in typical quantum landscapes, and hence regularization of the high-frequency modes is beneficial for optimization. Figure 1 : A sketch showing how high-frequency terms may affect the optimization landscape. Stars mark the position of global minima. Both in (a) and (b) the high-frequency terms produce many local minima. However, in (a) the global minimum of the regularized loss function (with high frequency modes discarded) is close to the true global minimum, and can be efficiently used to warm start the optimization. In contrast, in (b) the global minimum of the full function is far from the regularized one. One analogy to this intuitive picture the use of Gaussian filtering in image processing, which effectively removes high-frequency components from a signal denoising it. Another related concept is that of graduated optimization Blake1987 , where the loss landscape during the optimization is first highly smoothed, and then gradually restored to its original form. Graduated optimization have been related to optimization techniques such as simulated annealing Mobahi2015 and also explored in the context of classical deep learning Hazan2015 ; Li2023 ; Sato2024 . Finally, we note that the Fourier expansion is non-locally related to the original loss function, and can thus give a window into the global properties of the landscape. I.1 Results In general, the Fourier expansion of quantum loss functions contains exponentially many terms (both in the number of qubits and in the number of parameters; see e.g. Nemkov2023 ), and hence is hard to access or manipulate directly. The key technical contribution of this work is a protocol that exponentially suppresses high-frequency terms in the Fourier expansion of quantum loss functions. This regularization amounts to injecting a certain amount of noise for each parameterized gate in the quantum circuit, and can be implemented efficiently both in hardware and in simulation. In hardware, a single extra qubit is in principle sufficient to implement the necessary Pauli noise channels, though allowing for more ancillary qubits can reduce the depth of the overall circuit. In software, to simulate our protocol it is sufficient to switch to the density-matrix based simulation, though there is an associated computational cost. In addition, we show that the regularized and original loss functions are related by the heat equation, which can give both qualitative and quantitative insight to when and why our protocol should make optimization more efficient. For numerical experiments, we choose two types of models. The first is based on random Wishart fields, which serve as a statistical model for a large class of VQA with local minima Anschuetz2021 ; Anschuetz2022 . The second is the quantum convolutional neural network, which is a well-known QML model that has been shown to suffer from local minima already in small circuit sizes Anschuetz2022 . Our empirical results show a consistent and significant improvement in optimization metrics across all instances studied. For example, the probability to match or improve on the best solution found by the non-regularized optimization typically increases several-fold. I.2 Related work The problem of local minima in VQA and QML models has long been recognized empirically, see e.g. Kiani2020 ; Campos2021 ; Nemkov2023a . As the problem is generally NP-hard Bittel2021 , all existing techniques attempting to solve it are heuristics. A number of studies systematically explored how the use of different gradient-based and gradient-free optimizers affects the quality of VQA and QML solutions Lockwood2022 ; Bonet-Monroig2021 . Notably, quantum natural gradient Stokes2020 has been claimed to significantly alleviate the problem of local minima in variational quantum eigensolver Wierichs2020 , though to our knowledge it was only tested in a very limited setting and furthermore incurs significant computational overhead. Other proposals include using classical heuristics such as differential evolution Failde2023 and classical neural networks to warm-start the optimization Zhang2025 or regularize the loss landscape Rivera-Dean2021 . Furthermore, the potential benefits of noise to optimization of quantum loss functions have been noted previously. Refs Nguyen2016 ; Somogyi2024 argued that quantum noise can improve generalization of quantum neural networks, while Heyraud2022 made a similar observation for quantum kernel methods. In Campos2021 it was observed that a small amount of noise may prevent layer-wise training saturation in quantum approximate

Read the complete text