3 ms·

Yeah. The higgs evidence is pretty convincing visually. Not so sure about LIGO. There is an extraordinary claim of noise reduction that requires extraordinar

Yeah. The higgs evidence is pretty convincing visually. Not so sure about LIGO. There is an extraordinary claim of noise reduction that requires extraordinary evidence and it's all obfuscated behind adaptive machine learning based filtering, and the statistical analysis on that is unparseable to a non-expert (that is worrisome). The pulsar timing network though is easily believable. Luckily, there's pretty simple statistics that one can throw at that once the third detector comes online. Hopefully that comes in before we spend too much money on LISA. It's basically this, from the article, but from astro: > Particle physics does have situations where the hypothesis are not so data driven and they rely much more heavily on the theoretical edifice of quantum field theory and our simulation of the complicated detectors. In these cases, the statistical models are implicitly defined by simulators is actually a very hot topic that blends classical statistics with modern deep learning. We often say that the simulators don't have a tractable likelihood function. This applies to frequentist hypothesis testing, confidence intervals, and Bayesian inference. Confronting these challenging situations is what motivated simulation-based inference, which is applicable to a host of scientific disciplines.

maxnoe 2y ago
How do you explain the LIGO detection of a neutron neutron star merger that was at the same time observed as GRB by many, many other telescopes? https://en.m.wikipedia.org/wiki/GW170817 https://en.m.wikipedia.org/wiki/GW170817
- throwawaymaths 2y ago
  That's an n of one (I could be wrong but its the only multi messenger GW event we've seen, with how many events are being generated, you'd have expected more). Could be a coincidence. The angular resolution of LIGO is not exactly amazing, and we don't have a real estimate of the distance to the source of the GW? In fact, IIUC the event could be exactly in the opposite direction too. 1 for 29 by 2019: https://scholarcommons.sc.edu/phys_facpub/164/ https://scholarcommons.sc.edu/phys_facpub/164/
polyphaser 2y ago
There's no fancy machine learning needed to detect some "bright" LIGO signals (some of the first blackhole-blackhole mergers). Given a set of template signals, a matched filter tries to find the one that best matches the noisy signal your instrument recorded. In order for a MF to work, what you really need is a good understanding of the noise in your instruments observations -- and there're very few people in the world better at that than LIGO folks. LIGO spent almost 20 years in construction and R&D, and almost 30 in planning. When you hear stories about how they can detect trucks miles away, and detect the waves crashing on the coast, it's possible because of scores of PhD students' who spent years characterizing each and every component that affects LIGO's noise levels. All of this to say that it's possible to download their data online and do a quick MF analysis (I did that), and with a little bit of work you get a blindingly bright statistical significance of 20-sigma or so. The actual result quoted in the papers was a bit higher. That's a testament to how well the instrument was built and its behaviour understood.

maxnoe 2y ago

How do you explain the LIGO detection of a neutron neutron star merger that was at the same time observed as GRB by many, many other telescopes? https://en.m.wikipedia.org/wiki/GW170817 https://en.m.wikipedia.org/wiki/GW170817

throwawaymaths 2y ago
That's an n of one (I could be wrong but its the only multi messenger GW event we've seen, with how many events are being generated, you'd have expected more). Could be a coincidence. The angular resolution of LIGO is not exactly amazing, and we don't have a real estimate of the distance to the source of the GW? In fact, IIUC the event could be exactly in the opposite direction too. 1 for 29 by 2019: https://scholarcommons.sc.edu/phys_facpub/164/ https://scholarcommons.sc.edu/phys_facpub/164/

polyphaser 2y ago

There's no fancy machine learning needed to detect some "bright" LIGO signals (some of the first blackhole-blackhole mergers). Given a set of template signals, a matched filter tries to find the one that best matches the noisy signal your instrument recorded. In order for a MF to work, what you really need is a good understanding of the noise in your instruments observations -- and there're very few people in the world better at that than LIGO folks. LIGO spent almost 20 years in construction and R&D, and almost 30 in planning. When you hear stories about how they can detect trucks miles away, and detect the waves crashing on the coast, it's possible because of scores of PhD students' who spent years characterizing each and every component that affects LIGO's noise levels. All of this to say that it's possible to download their data online and do a quick MF analysis (I did that), and with a little bit of work you get a blindingly bright statistical significance of 20-sigma or so. The actual result quoted in the papers was a bit higher. That's a testament to how well the instrument was built and its behaviour understood.