How Quantum Circuits Actually Learn: A Causal Identification of Genuine Quantum Contributions

1. Problem Statement

Quantum Machine Learning (QML) aims to leverage non-classical phenomena (entanglement, superposition, quantum correlations) to outperform classical models. However, a critical methodological gap exists: it is unclear whether performance gains in variational quantum circuits (VQCs) stem from genuine quantum resources or merely from classical architectural scaling (e.g., increased circuit depth, parameter count, and expressivity).

Current approaches often treat quantum circuits as "black boxes," attributing improvements to quantum effects without rigorous proof. This leads to ambiguity: when a quantum model performs well but remains classically simulable, is it due to a fundamental limitation of quantum computing or simply the under-utilization of available quantum resources? The paper addresses the need to distinguish between direct architectural effects (classical scaling) and indirect quantum-mediated effects (genuine quantum contributions).

2. Methodology: Counterfactual Causal Mediation Framework

The authors introduce a Counterfactual Causal Mediation Framework to decompose performance differences between quantum architectures. This approach shifts the question from "Does quantum help?" to "How does quantum contribute?"

A. Experimental Design

• Paired Design: For each test sample, two architectures are evaluated:
  • Baseline ($t=0$): A shallow circuit (1 layer).
  • Enhanced ($t=1$): A deeper, more entangling circuit (3 or 6 layers) constructed by repeating the exact same parameterized block.
• Variables:
  • Treatment ($t$): Architectural intervention (depth).
  • Outcome ($Y$): Directional predictive performance (probability of correct classification).
  • Mediators ($M$): Quantum information-theoretic observables computed post-hoc from the trained quantum states:
    1. Entanglement Entropy ($S_A$): Subsystem non-separability.
    2. Purity ($\gamma_A$): State coherence.
    3. Linear Entropy ($L_A$): Mixedness.
    4. Quantum Mutual Information ($I(A:B)$): Total correlations.

B. Structural Causal Model (SCM)

The authors model the data-generating process using linear structural equations:

1. Mediator Equation: $M_{s,t} = \alpha_0 + \alpha t + \epsilon^{(M)}_s$ (How architecture affects quantum properties).
2. Outcome Equation: $Y_{s,t} = \tau_0 + \tau t + \beta^\top M_{s,t} + \epsilon^{(Y)}_s$ (How architecture and quantum properties affect performance).

Causal Decomposition (Theorem 1):
The Average Total Effect (ATE) of increasing circuit depth is decomposed additively:
$$\text{ATE} = \underbrace{\tau}_{\text{Direct Effect (DE)}} + \underbrace{\sum_{k} \alpha_k \beta_k}_{\text{Indirect Effect (IDE)}}$$

• Direct Effect ($\tau$): Performance gain attributable to architectural scaling (parameter count, optimization geometry) independent of quantum properties.
• Indirect Effect ($\sum \alpha_k \beta_k$): Performance gain mediated specifically through changes in quantum resources (entanglement, purity, etc.).

C. Identification Assumptions

The framework relies on three assumptions satisfied by the experimental design:

1. Sequential Ignorability: Treatment assignment is deterministic; mediators are computed post-hoc from fixed trained models, eliminating confounding.
2. No Treatment-Mediator Interaction: The effect of quantum mediators on performance is invariant across architectural regimes (tested empirically).
3. Consistency/SUTVA: Observed outcomes match potential outcomes; no interference between samples.

D. Metrics

• Mean Absolute Mediated Contribution (MAMC): Measures the average magnitude of a mediator's causal influence ($|\alpha_k \beta_k|$), ignoring sign to capture strength.
• Relative Quantum Contribution (RQC): The proportional importance of a mediator within the total mediated effect.
• Causal Regimes: A classification system (e.g., Quantum-Advantage, Classical-Dominated, Neutral) based on the signs and magnitudes of $\tau$ and the indirect effect relative to a data-driven significance threshold.

3. Key Contributions

1. Methodological Innovation: First application of counterfactual causal mediation analysis to QML to rigorously separate classical architectural scaling from genuine quantum resource utilization.
2. Quantitative Decomposition: Provides a mathematically grounded method to attribute performance gains to specific quantum properties (entanglement vs. coherence) rather than generic circuit depth.
3. Regime Classification: Introduces a taxonomy of learning behaviors (e.g., "Masked-Quantum" where quantum benefits are hidden by poor architecture; "Classical-Scalable" where gains are purely classical).
4. Resource-Aware Design: Proposes a shift from heuristic search to mechanistically informed circuit design, where architectures are engineered to amplify task-relevant quantum pathways.

4. Results

The framework was applied to 5 circuit topologies (Deep, Full, Linear, Ring, Pairwise) across 3 benchmark datasets (Breast Cancer, Diabetes, Ionosphere), totaling 43 validated configurations.

• Dominance of Direct Effects: In all configurations, direct architectural contributions systematically exceeded quantum-mediated effects.
  • Mean Ratio: Direct : Indirect = 13.1 : 1.
  • Mean Indirect Contribution: Only 0.82% of performance variation was mediated by quantum resources.
• Regime Distribution:
  • 93% Neutral: Both direct and indirect effects were below the significance threshold (indicating under-parameterization or optimization issues rather than lack of quantum potential).
  • 7% Classical-Scalable: Significant performance gains driven entirely by direct architectural scaling, with negligible quantum mediation.
  • 0% Quantum-Advantage: No configuration achieved a regime where quantum-mediated effects were statistically significant and positive.
• Mediator Specifics:
  • Purity ($\gamma_A$) was the dominant mediator in Breast Cancer and Diabetes tasks, suggesting coherence preservation is currently more impactful than entanglement in NISQ implementations.
  • Entanglement Entropy ($S_A$) was the sole quantifiable mediator in the Ionosphere dataset but contributed minimally.
  • Pairwise Topologies consistently failed to generate significant effects, confirming that localized bipartite entanglement is insufficient for these tasks.

5. Significance and Implications

• Reframing Quantum Advantage: The paper argues that current QML models are not failing due to a lack of quantum power, but due to under-utilization of quantum resources. The "quantum advantage" is not an inherent property of the hardware but a design target achievable through systematic optimization.
• Diagnosis of Current Limitations: The prevalence of the "Neutral" and "Classical-Scalable" regimes suggests that current NISQ circuits operate far below their theoretical potential. Performance gains are currently driven by classical expressivity (more parameters), not quantum mechanics.
• Actionable Design Guidelines:
  • Researchers should move beyond "deeper is better" and focus on mediator amplification (e.g., designing ansatzes that specifically enhance purity or entanglement relevant to the task).
  • The framework allows for the identification of "Masked-Quantum" regimes, where latent quantum benefits exist but are suppressed by suboptimal circuit design or optimization dynamics.
• Future Directions: The authors call for a transition from heuristic black-box tuning to mechanistically informed design, similar to the evolution of classical deep learning (gradient flow, inductive biases). The framework is generalizable to unsupervised learning, reinforcement learning, and quantum optimization.

In conclusion, this work provides the first rigorous causal evidence that current variational quantum circuits are largely classical function approximators in disguise, and that realizing true quantum advantage requires a fundamental shift in how circuits are designed to explicitly harness and amplify specific quantum information-theoretic properties.