Loss functions — overview¶
sweep_loss groups the FWI misfit literature into the following families:
| Family | Examples implemented |
|---|---|
| A. Data-domain Lp norms / robust M-estimators | L2, L1, Huber, pseudo-Huber, hybrid L1/L2, Cauchy, Tukey, Geman–McClure |
| B. Correlation / amplitude-normalised | trace-normalised L2, global correlation (NCC), local similarity (planned) |
| C. Envelope / instantaneous-phase (Hilbert) | envelope, log-envelope ratio, instantaneous phase, combined envelope+phase (planned) |
| D. Frequency / Laplace domain | Pratt frequency-domain L2, Shin-Min log misfit, phase-only & amplitude-only, Laplace, Laplace–Fourier (planned) |
| E. Travel-time / picking | cross-correlation travel-time, instantaneous travel-time, dynamic image warping (planned) |
| F. Convolution / matching-filter | AWI, deconvolution-based, matching-filter penalty (planned) |
| G. Normalized integration method (NIM) | NIM / CDF L2 (planned) |
| H. Optimal transport | W1 / KR-norm, W2 via CDF, Sinkhorn, GSOT (planned) |
Each subsection on the left-hand sidebar gives the formula, the canonical reference (with DOI) and the tests we ship.
If you came here looking for a one-page summary with every formula, see the Report.