A Noise-Aware Quantum Algorithm for Credit Valuation Adjustments on Real Quantum Hardware

Technical Summary: A Noise-Aware Quantum Algorithm for Credit Valuation Adjustments on Real Quantum Hardware

Problem Statement
Credit Valuation Adjustment (CVA) is a critical component of derivative pricing that accounts for counterparty credit risk. Quantitatively, CVA requires the repeated estimation of risk-neutral expectations over complex, multi-asset portfolios, a task traditionally handled by Monte Carlo (MC) simulation. While classical MC scales as $O(N^{-1/2})$, Quantum Amplitude Estimation (QAE) offers a theoretical quadratic speedup, scaling as $O(N^{-1})$. However, translating this asymptotic advantage into a practical financial workflow on current Noisy Intermediate-Scale Quantum (NISQ) devices remains an open challenge. Previous circuit-level proposals, such as that by Alcazar et al. (2022), focused on simplified single-option instances and relied on ideal simulations, leaving unaddressed the degradation of quantum information due to hardware noise, finite discretisation, and the specific complexities of multi-asset correlated dynamics.

Methodology
The authors present an end-to-end, noise-aware quantum workflow for CVA, executed on real IBM Quantum hardware. The methodology is structured into three primary components:

1. Financial Modeling and Quantum Encoding:

   • Model: The study extends previous work from a single option to a correlated two-asset netting set. It incorporates risk-neutral multi-asset lognormal dynamics, calibrated to market data (equity closes, implied volatility surfaces, CDS spreads, and discount factors).
   • Discretisation: The continuous time-market domain is truncated and discretised onto a finite grid compatible with qubit registers.
   • Circuit Construction: The CVA functional is encoded into a quantum circuit $A$. This involves:
     • State Preparation: A Quantum Circuit Born Machine (QCBM) is trained to prepare a joint time–market probability distribution representing the correlated asset dynamics.
     • Payoff Encoding: Controlled-rotation circuits (CRCA) encode the positive exposure, discount factors, and default probabilities into three ancilla qubits.
     • Target: The CVA value is recovered from the probability of measuring the "good" state (all three ancillas in $|1\rangle$).

2. Contrast-Aware Bayesian Iterative Quantum Amplitude Estimation (CABIQAE):

   • Recognizing that Grover amplification loses contrast (signal-to-noise ratio) as circuit depth increases on real hardware, the authors introduce CABIQAE.
   • Unlike standard Bayesian Amplitude Estimation (BAE) or Iterative QAE (IQAE) which assume ideal oscillation, CABIQAE incorporates a calibrated contrast-decay model. It treats the experimentally observed loss of Grover contrast not as an error, but as a parameter in the inference problem.
   • The algorithm uses a hardware-calibrated likelihood model where the success probability decays exponentially with the Grover power $k$. It dynamically selects Grover depths that maximize information gain within the "contrast window" before noise dominates.

3. Experimental Validation:

   • The workflow is validated using a "hardware-replay" methodology. The authors first characterize the hardware response of the amplified circuits and then replay adaptive estimation trajectories (CABIQAE, BIQAE, and BAE) using this empirical data. This allows for robust statistical comparison without executing every adaptive trajectory live on the quantum processor.
   • Comparisons are made against Direct Circuit Sampling (DCS) and noise-agnostic baselines (BIQAE, BAE).

Key Results

• Algorithmic Performance: In the hardware-replay regime, noise-agnostic algorithms (specifically BIQAE) fail to improve beyond a certain query cost, collapsing to a noise floor as they blindly apply deep Grover iterations. In contrast, CABIQAE and BAE continue to reduce error over a broader range.
• Efficiency: CABIQAE achieves a median relative error of $1.82 \times 10^{-3}$ with a median query cost of $4.95 \times 10^4$. While its final error is comparable to BAE ($1.77 \times 10^{-3}$), CABIQAE operates with significantly shallower circuits (median max Grover factor $K=23$ vs. $K=33$ for BAE) and drastically lower classical post-processing runtime (1.25s vs. 45.72s).
• Error Decomposition: The study decomposes the total CVA error into statistical, encoding, discretisation, and hardware contributions. The analysis reveals that the full CVA oracle is currently limited by circuit depth and discretisation resolution rather than just statistical sampling.
• Scaling: In the noisy regime, CABIQAE maintains a convergence exponent of $\beta \approx -0.69$, which is superior to the hardware-replay DCS ($\beta \approx -0.30$) and the noise-agnostic BIQAE ($\beta \approx -0.25$), though it does not reach the ideal Heisenberg limit ($\beta = -1$).

Significance and Claims
The authors explicitly state that this work does not claim an end-to-end quantum runtime advantage over classical Monte Carlo for the full CVA calculation. Instead, the significance lies in:

1. Realism: Providing a complete, end-to-end pipeline that moves beyond single-option benchmarks to a calibrated, multi-asset netting set on real hardware.
2. Noise-Awareness: Demonstrating that incorporating experimentally calibrated contrast loss into the amplitude estimation loop (via CABIQAE) makes the estimator better suited for the specific regime of current NISQ devices, where amplification is useful only within a finite depth window.
3. Practical Scalability: Showing that a noise-aware approach can achieve lower classical post-processing overhead and utilize shallower circuits compared to noise-agnostic Bayesian baselines, thereby improving the practical scalability of quantum CVA workflows.

The paper concludes that while the full quantum advantage is currently limited by hardware constraints and discretisation, the proposed noise-aware framework offers a viable path for extracting useful information from amplified circuits in the presence of realistic hardware noise.