Prototyping a Global-fit Pipeline for LISA

Senwen Deng (APC)

with

​

S. Babak, M. Le Jeune, E. Plagnol, A. Sartirana (APC)

S. Marsat (L2IT)

Septième assemblée générale du GdR Ondes Gravitationnelles

16–17 Oct 2023

LUTH, Observatoire de Paris, Meudon

Profile of LISA Data

• Dominated by GW signals, all-sky all-time
• Many signals are long-lived (EMRI, GB) and overlapping in F&T

Simulated LISA Data: Sangria

Simulated LISA Data: Sangria

Simulated LISA Data: Sangria

Simulated LISA Data: Sangria

Simulated LISA Data: Sangria

Simulated LISA Data: Sangria

Simulated LISA Data: Sangria

Profile of LISA Data

• Dominated by GW signals, all-sky all-time
• Many signals are long-lived (EMRI, GB) and overlapping in F&T
• Unresolved GW signals contribute to the noise budget
• Non-stationary noise: gaps, glitches
• Pioneer’s problem: unknown event rate, unknown parameter distribution

Challenging!

Strategy of Global-fit

Keywords: kick-in, subtraction, iteration

Strategy of Global-fit: prototype architecture

* Preliminary reconstruction if the PE live catalogue is not available yet

Kick in

Iterations take place in each block

(Effective) subtraction in each iteration

​

The live catalogues and the noise model are updated as iterations go on

Strategy of Global-fit: prototype architecture

Each block iteration

1. subtracts unattended live catalogue signals from data
2. refines (MCMC)
3. updates the live catalogue/model

​

Specifics:

• There are up to thousands of jobs on an HTC cluster.
• The blocks can be async.
• MCMC Chains, plots, logs and debug information are stored at each iteration.

​

More blocks for other sources: SMBH, EMRI, …

* Preliminary reconstruction if the PE live catalogue is not available yet

The live catalogues and the noise model are updated as iterations go on

The kick-in step

We neglect the LISA motion and we assume the long-wavelength regime.

• Detect and reconstruct with a time-sliding F-statistic: log-likelihood ratio maximised over time of arrival (merger), distance, inclination, sky, initial phase, polarisation

How to find the maximum fast? One possible way: mesh-refinement driven by Vegas

• Adapt the meshgrid by doing Monte-Carlo integrations
• Embarrassingly parallelizable

Kick in

Preliminary MBHB reconstruction

The kick-in step: results

Subtract with the reconstructed signals

Dealing with the Galactic binaries

Dealing with the Galactic binaries

Dealing with the Galactic binaries: results

Preliminary results @4mHz

Submission

Refine MBHB PE

We start from the preliminary results of the kick-in step

• Reconstructed signal for heterodyning (Cornish & Shuman 2020)
• Initial points for the sampling

Parameter mapping is helpful (Marsat et al. 2021)

Refine MBHB PE

Get the noise level

• Starting from the PSD estimation
• Plug in a parametric model and fit

Get the noise model

Simultaneously fit the parametric model using a dozen of bins

Summary

With the Sangria analysis, we have demonstrated all the four components of the global-fit prototype

• Fast and preliminary detection/removal of MBHBs
• Search for Galactic binaries in small overlapped frequency bands
• Fast PE for MBHBs
• Noise model fitting

We built a modular architecture combining the components in concert

Next steps

Short term

• Robust stopping criterion for new source discovery (GB Block)
• Time iteration: Data accumulates with time. Each type of source has its own good cadence for data analysis.
• Dealing with gaps and glitches

Long term

• Add the modules (blocks) for other sources (SMBH, EMRI, …)

Backup slides

The parameteric noise model

For TDI A channel,

The parameters are

Sangria in FD

Sangria in FD

Sangria in FD

Sangria in FD

Low frequency sampling of MBHB signals

Low frequency sampling of MBHB signals

Low frequency sampling of MBHB signals