MPSi fixes multipath extraction at a set depth: one pilot sounding rebuilds the channel, and refinement steps no longer track signal strength

Multipath parameter extraction moves from deciding the model order and iterating to convergence on every observation to a fixed-depth forward computation that still returns interpretable path-level parameters: complex gain, angle of arrival and delay for each path. The classical parametric channel-sounding algorithm RIMAX relies on data-dependent path-count selection, diffuse-multipath component modeling and cost-checked iterative nonlinear refinement, so its control flow varies and its number of repeated optimization steps is unknown, which makes it unsuitable for low-delay systems. In the Bahen and Dundas ray-traced scenarios, MPSi's initial-access and full-channel-estimation performance is comparable to the enhanced RIMAX baseline, while its refinement depth is fixed and substantially smaller than RIMAX's iteration count, which varies with transmit power.

The approach keeps the physical channel model in the loop and uses learned components to replace the two costliest stages of RIMAX: order selection and iterative refinement. A path-count network predicts the effective number of paths from an antenna-averaged log-power profile, a differentiable soft-top-k gate assigns weights to a fixed set of candidate paths, and a fixed number of unfolded Levenberg–Marquardt layers then jointly update the geometric parameters and compute ridge-regularized gain estimates. The training objective combines the final reconstruction normalized mean-squared error, a layer-wise deep-supervision loss and a path-count parsimony penalty, with ray tracing used solely to generate pilot observations; test results come from realizations generated independently of the training data.

The conclusions are limited to two site-specific ray-traced environments representing a small cell (Bahen) and a dense urban area (Dundas), and both training and test data are synthesized with the Sionna ray tracer over OpenStreetMap scenes rather than measured base-station data. The authors propose evaluating a broader range of environments and larger test sets, measuring platform-specific runtime, investigating MPSi-generated labels in downstream learning tasks, and extending the path model with Doppler for mobile scenarios and channel prediction.

Iteration count serves only as a control-flow complexity indicator, and deployment-level latency claims would require operation counts and timing on a specified hardware and software platform, which the work places out of scope. The residual quality-control threshold is described as a practical initial screening gate rather than a universal statistical limit, and that residual analysis relies on the known noiseless simulation signal, so a practical deployment would need a calibrated or independently estimated noise level.