<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Neural Posterior Estimation | A Quest After Perspectives</title><link>https://iphysresearch.github.io/blog/tag/neural-posterior-estimation/</link><atom:link href="https://iphysresearch.github.io/blog/tag/neural-posterior-estimation/index.xml" rel="self" type="application/rss+xml"/><description>Neural Posterior Estimation</description><generator>Hugo Blox Builder (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Thu, 12 Nov 2026 00:00:00 +0000</lastBuildDate><image><url>https://iphysresearch.github.io/blog/media/sharing.png</url><title>Neural Posterior Estimation</title><link>https://iphysresearch.github.io/blog/tag/neural-posterior-estimation/</link></image><item><title>Rapid Multiparameter Gravitational-Wave Tests of General Relativity with Neural Posterior Estimation</title><link>https://iphysresearch.github.io/blog/mypublication/2026_npe_gr_tests/</link><pubDate>Thu, 12 Nov 2026 00:00:00 +0000</pubDate><guid>https://iphysresearch.github.io/blog/mypublication/2026_npe_gr_tests/</guid><description>&lt;p>DINGO-TGR replaces sampling-based inference with a trained neural posterior that proposes samples in the six-dimensional TIGER deformation space together with the binary parameters, then reweights those samples by importance sampling so that the result carries the likelihood rather than a surrogate of it. Using IMRPhenomXPHM waveforms with spin precession and higher-order modes, the framework recovers GR-consistent posteriors on 16 zero-noise injections and the injected leading-mode shifts on 12 beyond-GR injections.&lt;/p>
&lt;p>Inference of this kind takes roughly 38 minutes for GW150914 and 4.4 hours for GW250114 after training, against 177 and 238 hours for conventional sampling, and agrees with independent &lt;code>bilby-tgr&lt;/code>/&lt;code>dynesty&lt;/code> analyses on both events. GW150914 shows a mild offset from GR in the leading principal component, which a 1000-injection background study places at an empirical upper-tail &lt;em>p&lt;/em>-value of 11.5% — consistent with noise at the 5% level.&lt;/p>
&lt;h2 id="sources">Sources&lt;/h2>
&lt;ul>
&lt;li>&lt;a href="https://arxiv.org/abs/2611.17038" target="_blank" rel="noopener">arXiv abstract and submission history&lt;/a>&lt;/li>
&lt;/ul></description></item></channel></rss>