VROOM-SBI: A Fast Simulation-Based Bayesian Inference Methodology for QU-Fitting

Technical Summary: VROOM-SBI: A Fast Simulation-Based Bayesian Inference Methodology for QU-Fitting

Problem Statement
Bayesian QU-fitting is recognized as a highly accurate method for inferring line-of-sight Faraday rotation measures and characterizing depolarization physics. However, its application to survey-scale data (e.g., from SKA precursors like VLA, MeerKAT, or uGMRT) has historically been infeasible due to prohibitive computational costs. Traditional implementations, such as those in RMTools using nested samplers, require minutes per pixel. For a typical image cube containing tens of thousands of active polarized pixels, a full analysis would demand thousands of CPU-hours, rendering full Bayesian inference impractical for large catalogs. While Faraday synthesis offers speed, it produces "dirty" spectra convolved with the Rotation Measure Spread Function (RMSF) and lacks the rigorous parameter posteriors of model fitting.

Methodology: VROOM-SBI
The authors introduce VROOM-SBI, a framework utilizing Simulation-Based Inference (SBI), specifically Neural Posterior Estimation (NPE), to accelerate Bayesian inference. The core approach involves training neural networks to learn a direct mapping from observed Stokes Q and U spectra to the posterior distributions of physical model parameters, bypassing the need for repeated likelihood evaluations during inference.

• Forward Simulation: The authors generate a massive training dataset by simulating Stokes spectra based on four standard depolarization models: Faraday-thin, Burn slab, External dispersion, and Internal dispersion. A two-component Faraday-thin model is also included. The simulations cover VLA L-band frequency coverage (1–2 GHz) and Faraday depths ranging from $-100$ to $100$ rad m$^{-2}$.
• Noise and Data Augmentation: The training data incorporates realistic noise models, including per-channel inverse-variance weights and additive radiometer-like noise. To ensure robustness, the dataset is augmented with Radio Frequency Interference (RFI) patterns, such as scattered missing channels, contiguous gaps, and large RFI blocks.
• Neural Architecture: The inference pipeline employs a two-stage process:
  1. Spectral Shape Estimation: A dedicated estimator fits a log-log cubic polynomial to the total intensity (Stokes I) spectrum. This provides a smooth, noise-marginalized model of $I(\nu)$, which is used to normalize Q and U spectra into fractional polarization ($q=Q/I, u=U/I$), preventing noise amplification from raw division.
  2. Polarimetric Inference: The normalized fractional polarization spectra, along with channel weights, are fed into a Neural Spline Flow (a type of normalizing flow). This network, consisting of a fully connected embedding layer followed by coupling transforms, learns the conditional density estimator $q_w(\theta|x)$ to approximate the true posterior $p(\theta|x)$.
• Inference Speed: Once trained, the network performs inference in milliseconds per pixel on a GPU, drawing thousands of posterior samples in a single forward pass.

Key Contributions

1. Speedup: VROOM-SBI achieves a computational speedup of approximately $\sim500\times$ compared to classical QU-fitting implementations (e.g., RMTools with dynesty).
2. Open Resources: The authors provide trained models for the four standard depolarization models and the spectral shape estimator via HuggingFace, along with open-source code.
3. Amortized Inference: The methodology shifts the computational burden to the training phase. Once trained for a specific frequency configuration and noise profile, the network can process millions of sources at a fraction of the cost of traditional methods.
4. Robustness: The models are trained with aggressive RFI augmentation, demonstrating resilience to channel excision and varying noise floors.

Results

• Parameter Recovery: Validation on held-out test simulations shows that VROOM-SBI accurately recovers rotation measures ($\phi$) and intrinsic polarization angles ($\chi_0$) with high precision. For the Faraday-thin model, the median absolute error (MedAE) for $\phi$ is $0.26$ rad m$^{-2}$ with 93% coverage at the 68% credible level.
• Degeneracy Handling: The method successfully resolves multi-component structures (e.g., two Faraday screens) without significant bias, provided the frequency separation is sufficient.
• Real Data Application: Applied to VLA L-band observations of the galaxy cluster MACS J1752+4440, VROOM-SBI processed 35,700 active pixels in approximately 125 minutes on a single GPU.
  • The recovered spectral index maps align with physical expectations for radio relics (steepening from shock fronts).
  • Comparison with RM synthesis and a subset of 100 pixels analyzed via traditional QU-fitting shows that VROOM-SBI recovers consistent large-scale RM structures.
  • VROOM-SBI yields slightly broader posteriors (median $\sigma_{RM} \approx 3.8$ rad m$^{-2}$ vs. $0.67$ rad m$^{-2}$ for QU-fitting) and a systematic offset of $\sim2.3$ rad m$^{-2}$ relative to QU-fitting. The authors attribute this to the amortized nature of the inference, which integrates over the range of noise realizations seen during training rather than conditioning on the exact per-observation noise level.
• Coverage: Empirical coverage for the 68% credible intervals generally falls between 64% and 93%, with the exception of the fractional polarization ($p_0$) in the Faraday-thin model (46% coverage), which the authors attribute to the model's lack of structural information regarding amplitude scaling.

Significance and Claims
The paper claims that VROOM-SBI makes full Bayesian polarimetric analysis computationally tractable for modern wide-field surveys. By reducing the inference time for a million-source catalog to approximately one day on a single GPU, the method bridges the gap between the statistical rigor of model fitting and the speed required for next-generation radio telescopes (SKA, ngVLA). The authors position this as a shift from the historical trade-off between accuracy and speed, enabling survey-scale studies that were previously limited to fast but less rigorous methods like RM synthesis. They note that while the current models are specific to certain frequency configurations, the approach is generalizable, and future work could extend this to "dirty" Faraday spectrum domains to further integrate with imaging pipelines.