EMRI parameter inference no longer needs a true-value start: a neural network narrows the search, and a six-month signal yields posteriors in about ten hours
Synopsis
Bayesian parameter inference for extreme-mass-ratio inspirals (EMRIs) has relied on narrow priors centered on the true parameters or Markov chains initialized near them, information real observations do not provide. A deep-learning-assisted framework first uses a neural network to produce approximate posteriors from the data and shrink the search region, then particle swarm optimization finds a high-likelihood starting point, and staged Markov chain Monte Carlo sampling produces the final posteriors. On two noise-free six-month simulated signals, the resulting posteriors agree closely with reference analyses initialized near the true values, and once the network is trained each signal takes about ten hours.
Interpretation
EMRI parameter inference can be initialized without knowledge of the true source parameters and still yield posteriors close to a reference analysis started near the truth. Earlier EMRI analyses of simulated signals typically adopted narrow priors centered on the true parameters or initialized Markov chain Monte Carlo chains near them, a strategy that relies on knowledge unavailable in real observations. For two six-month simulated signals, Signal 1 and Signal 2 (the latter having the smallest volume contraction factor among fixed-phase test signals), the posteriors are compared with a reference analysis using the same priors and sampling settings, showing substantially overlapping credible regions and similar parameter correlations.
The framework chains learned localization to likelihood-based sampling: the neural network only narrows the range, and the final estimates remain likelihood-driven. Approximating EMRI posteriors directly with neural networks over broad priors remains difficult, while a direct global search over a large parameter space requires repeated costly waveform generation. The network combines an embedding network of a temporal convolutional network and a multilayer perceptron with a neural spline flow, trained by cross-entropy loss; its marginal credible intervals define a reduced axis-aligned search region, particle swarm optimization maximizes the joint overlap inside it, and Eryn then runs staged Markov chain Monte Carlo with the data duration increased stage by stage to six months.
The search-space contraction is pronounced for most parameters and keeps comparable coverage for random initial phases unseen in training. All three initial phases are fixed to zero during training, so signals with random initial phases form an out-of-distribution test. On an independent test set the volume contraction factor has a median of about one value and a minimum of about another, and for a stated fraction of test signals all true values lie within their search intervals; a random-phase test set gives a comparable median contraction factor and the same joint coverage fraction.
Perspective
The demonstration is limited to noise-free simulated six-month EMRI signals over a restricted source-parameter domain, with all three initial phases fixed to zero during training; the random-phase test set only checks search-region contraction and true-value coverage, and the full inference pipeline is run only on two fixed-phase signals.
The study describes itself as a proof-of-principle demonstration based on noise-free simulations and a restricted source-parameter domain; realistic EMRI observations will contain instrumental noise and an unresolved Galactic binary foreground, and may be affected by glitches, data gaps, and uncertainties or variations in the noise power spectral density. Signal 2 shows weaker search-region contraction, and some of its marginal posteriors differ from the reference analysis in width or weak secondary structure.