Talk detail

Participant	Sengupta, Anand Sankar
Institution	Indian Institute of Technology Gandhinagar - Village Palaj Simkheda - Gandhinagar - Gujarat - India

Session	GW9	Accepted		Order	Time
Talk	Oral abstract	Title	Efficient gravitational wave searches from compact binaries using a random projection based template-factorization
Coauthors
Abstract	Compact-binary coalescences are an important class of sources for the advanced-LIGO detectors. Improved sensitivity at low frequencies and increased bandwidth of the advanced-LIGO and Virgo detectors has resulted in a vast increase in the number of templates to cover the deemed parameter space covering target mass and spin ranges. Further, it has also led to longer template waveforms used for matched-filtering based searches. These in-turns amplify the computational cost of the gravitational wave searches by several orders of magnitude. Random matrix factorization is an efficient technique for the low-rank approximation of a given large template matrix by projecting the template waveforms into a much lower dimensional space. We report the application of such a technique to reduce the high computational costs of adv-LIGO CBC searches. We demonstrate that one can efficiently factorize large template matrices over a distributed memory computer architecture and overcome time and space complexity problem associated with the standard techniques. We also show how this method can potentially benefit existing low-latency searches and also lead to a hierarchical search paradigm using multi-processor and multi-GPU systems. Numerical experiments indicate that the reconstructed detection statistic in the proposed new algorithm is as accurate as the standard truncated-SVD searches within a fixed probabilistic error bound, and is afforded at a fraction of the current computational cost.
Pdf file

Session	GW9	Accepted		Order	Time
Talk	Oral abstract	Title	Efficient gravitational wave searches from compact binaries using a random projection based template-factorization
Coauthors
Abstract	Compact-binary coalescences are an important class of sources for the advanced-LIGO detectors. Improved sensitivity at low frequencies and increased bandwidth of the advanced-LIGO and Virgo detectors has resulted in a vast increase in the number of templates to cover the deemed parameter space covering target mass and spin ranges. Further, it has also led to longer template waveforms used for matched-filtering based searches. These in-turns amplify the computational cost of the gravitational wave searches by several orders of magnitude. Random matrix factorization is an efficient technique for the low-rank approximation of a given large template matrix by projecting the template waveforms into a much lower dimensional space. We report the application of such a technique to reduce the high computational costs of adv-LIGO CBC searches. We demonstrate that one can efficiently factorize large template matrices over a distributed memory computer architecture and overcome time and space complexity problem associated with the standard techniques. We also show how this method can potentially benefit existing low-latency searches and also lead to a hierarchical search paradigm using multi-processor and multi-GPU systems. Numerical experiments indicate that the reconstructed detection statistic in the proposed new algorithm is as accurate as the standard truncated-SVD searches within a fixed probabilistic error bound, and is afforded at a fraction of the current computational cost.
Pdf file