Trainability of IQP Quantum Circuit Born Machines Under Gaussian Initialization

Abstract: Quantum Circuit Born Machines (QCBMs) offer a natural approach to generative machine learning by leveraging the Born rule. Recent work has provided a method to classically train QCBMs with Instantaneous Quantum Polynomial (IQP) circuits via the Maximum Mean Discrepancy (MMD) loss. Despite the assumed intractability of sampling from IQP circuits classically, their expectation values can be computed classically, enabling training of these IQP QCBMs. However, quantum machine learning (QML) models have various other challenges, including trainability issues caused by exponential concentration or barren plateaus. While these issues have been explored for parameters sampled from a uniform distribution, little work has been done to rigorously treat the use of arbitrary Gaussian initialization schemes. This work leverages Stein's lemma and Lipschitz concentration bounds for Gaussian random variables to provide an analytical lower bound of the variance of the gradient and a probabilistic concentration bound of the deviation of the gradient from its mean. It discusses strategies to either avoid or encourage exponential concentration, as well as the conditions under which barren plateaus are more likely to occur.

• The paper derives a rigorous analytical lower bound on gradient variance, revealing how Gaussian parameter initialization affects trainability.
• It establishes sufficient conditions to avoid exponential gradient decay by tuning kernel bandwidth, ansatz connectivity, and parameter variance.
• The work provides actionable guidelines to mitigate barren plateaus and enhance scalability of IQP QCBMs in quantum machine learning.

Analytical Study of Trainability in IQP Quantum Circuit Born Machines Under Gaussian Initialization

Technical Context and Motivations

Quantum Circuit Born Machines (QCBMs) represent a foundational direction in quantum machine learning for modeling probability distributions via circuit-based quantum sampling governed by the Born rule [liu2018differentiable, coyle2020born, gili2023quantum]. Among QCBM architectures, those based on Instantaneous Quantum Polynomial (IQP) circuits are notable due to their classical sampling intractability and structural relevance to quantum computational advantage scenarios [bremner2016average]. Despite sampling hardness, expectation values of IQP circuits, necessary for gradient-based training via the Maximum Mean Discrepancy (MMD) objective, can be tractably computed on classical hardware [nest2009simulating]. This duality motivates the “train-on-classical, deploy-on-quantum” paradigm [recio2025train].

QML architectures, including IQP-QCBMs, face significant challenges relating to barren plateaus and exponential gradient concentration, which severely hinder large-scale trainability [mcclean2018barren, thanasilp2024exponential]. While these issues have been predominantly explored under uniform parameter initialization, the implications of Gaussian initialization—pervasive in classical ML and increasingly used in QML [zhang2022escaping]—remain largely unaddressed. This work rigorously characterizes IQP-QCBM trainability under arbitrary Gaussian parameter distributions, providing a foundational link between initialization statistics, circuit connectivity, and resultant gradient concentration.

Mathematical Framework

The analysis builds on the IQP QCBM training formalism using the MMD loss with a Gaussian kernel. The generative circuit prepares a model distribution $q_\theta$ over $n$ qubits, parameterized by $\theta_j$ drawn IID from $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$. Important technical elements include:

• Observable basis: Multiqubit Pauli-$Z$ operators indexed by binary vectors $a \in \{0,1\}^n$, expectation values $\langle Z_a \rangle_{p}$ (target) and $\langle Z_a \rangle_{q_\theta}$ (model), and generator bitstrings $g_j$ for circuit parameterization.
• MMD loss: Reformulated as an expectation over observables with weights specified by the Gaussian kernel bandwidth $\sigma^2$ through $n$0. The loss and its gradient depend on expectation values that can be evaluated efficiently classically [recio2025train].
• Gradient structure: The gradient with respect to $n$1 is sparse, involving only observables overlapping with $n$2. This induces potential locality in gradient statistics, contingent on circuit connectivity.

Main Technical Contributions

Analytical Lower Bound on Gradient Variance

A principal result is the derivation of a lower bound on the gradient variance for MMD-trained IQP-QCBMs under Gaussian-initialized parameters. Employing a multivariate extension of Stein’s lemma, the authors show:

$n$3

Explicit expansions show that the second derivative depends on the Gaussian moments of circuit observables and kernel parameters. Importantly, the exponential decay of variance—a signature of barren plateaus—can be diagnosed analytically by tracking products like $n$4, where $n$5 is the number of generator overlaps for observable $n$6.

Sufficient Conditions and Mitigation Strategies

The analysis identifies three primary factors influencing exponential concentration:

1. Kernel Bandwidth Scaling ($n$7): If $n$8 is constant, kernel weights drive $n$9 to concentrate on global observables with high Hamming weight, causing the lower bound to vanish as $\theta_j$0. By scaling $\theta_j$1, kernel mass shifts to local observables.
2. Ansatz Connectivity and Parameter Variance ($\theta_j$2, $\theta_j$3): Exponential decay is avoided if $\theta_j$4 (i.e., local circuits) or variance is inversely scaled $\theta_j$5. Even highly connected ansätze can avoid plateaus if initialized with appropriately diminished variance.
3. Initialization Mean ($\theta_j$6): Nonzero $\theta_j$7 causes the mean of trigonometric moments to decay exponentially with $\theta_j$8. Maintaining $\theta_j$9 ensures the lower bound does not vanish in the large-$\mathcal{N}(\mu_\theta, \sigma^2_\theta)$0 limit.

Consequently, exponential concentration is avoidable even at large-scale by synergistically tuning these properties—expanding the parameter distribution, employing local circuit design, or optimizing kernel bandwidth.

Probabilistic Concentration Bound

A refined Lipschitz-based concentration analysis determines that, for deviation threshold $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$1,

$\mathcal{N}(\mu_\theta, \sigma^2_\theta)$2

with $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$3 an explicit function of kernel and circuit structure. This bound tightly quantifies the probability of large gradient fluctuations, linking trainability breakdowns to regimes of small $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$4. Notably, local circuits with tunable kernel width have $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$5 polynomial in $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$6, further motivating architectural locality and adapted kernel design.

Connection to Barren Plateaus

While exponential concentration is necessary for barren plateaus, mean gradient nullification requires $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$7. Thus, strategies mitigating variance decay (e.g., small mean, large variance, locality) may not, in isolation, prevent plateaus if the mean contracts to $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$8, but still significantly improve trainability in practice.

Numerical and Empirical Considerations

The theoretical results are robust: the derived variance lower bound, for practical parameter regimes and moderate qubit numbers (e.g., $\mathcal{N}(\mu_\theta, \sigma^2_\theta)$9--$Z$0), identifies explicit paths whereby trainability failures can be avoided by altering initialization and circuit design. This analysis thus offers actionable prescriptive guidance for large-scale QCBM training protocols under Gaussian initializations.

Implications and Prospects

Practically, these results:

• Provide rigorous criteria for initializing and constructing IQP-QCBMs so as to avoid barren plateaus, supporting their scalable training on classical hardware with deployment on quantum devices.
• Establish that exponential gradient decay is not inevitable for QML architectures, given adaptive use of circuit locality, variance scaling, and kernel-parameterized losses.
• Suggest a pathway for experimental validation, including interaction effects between kernel bandwidth and data-circuit alignment.

Theoretically, this work widens the understanding of QML trainability landscapes, highlights interplay between algorithmic hyperparameters and circuit structure, and motivates further investigation into inductive biases from data-aligned circuits, more sophisticated kernel choices, and non-Gaussian initializations.

Conclusion

This paper presents a comprehensive analytical treatment of IQP-based QCBM trainability under Gaussian parameter initialization, deriving both lower variance and high-probability concentration bounds. These results substantiate the claim that barren plateaus or exponential concentration may be systematically avoided by synergistic control of kernel bandwidth, ansatz locality, and parameter distribution statistics. This work establishes rigorous criteria for scalable QML training and lays the foundation for further empirical and theoretical exploration of trainability phenomena in large-scale quantum generative models (2606.10179).