=== Duncan Brown - first NINJA project ===

- numerical injection analysis: study response of GW pipelines to waveforms to foster collaborations between NR and data-analysis communities

- start in spring 2008, developed format, waveforms freely submitted
	- vacuum spacetime only
	- up to two waveforms per code
	- equal and few:one masses, spinning and nonspinning, few to several cycles
- waveforms added to simulated colored Gaussian noise
	- initial LIGO and Virgo
	- generate random distances, locations, orientations, masses
	- constraints: starting frequency below 30 Hz, SNR > 5
- nine groups analyzed data with a variety of algorithms
	- signal parameters published
	- different data-analysis algorithms
- methods applied
	- CBC matched-filter pipeline (SPA, extended SPA, EOB-IMR, phenom-IMR)
	- CBC pipeline with Neyman-Pearson detection
	- LSC-Virgo ringdown
	- Q pipeline
	- Bayesian inference pipelines
	- ...
- conclusions
	- success, but hard to draw conclusions
- follow-on
	- broader parameter space
	- PN-NR stitching
	- non-Gaussian noise transients
- questions
	- possible avenues of collaboration: eased providing NR waveforms to LVC; may imagine requests of real data from NINJA
	- how can LVC improve pipelines with these results? 3.5PN EOB with ringdown cutoff would improve sensitivity

=== Arun - burst search for BBH mergers in Virgo C7 data ===

=== Laura Cadonati - systematics of NR waveform accuracy and burst searches for binary black hole mergers ===

- uses of NR in GW data analysis
	- template banks
	- injections
- NR questions: resolution, extraction radius, starting frequency, harmonics, hybrids
- DA questions: parameter space to explore, changes in reach, parameter estimation
- more questions: choice of NR parameters
- MayaKranc PSU/Gatech code
- burst search yardsticks: radius of sphere at which uniformly distributed population has 50% detection probability
- ...

=== Jon Gair - seed black holes with ET ===

- galactic black holes grow by accretion and mergers
- first generation of black holes probably created by collapse of massive, zero-metallicity pop-III stars
- seed BH parameters poorly constrained

- Einstein telescope
	- x10 over advanced LIGO, 1-10 Hz sensitivity
	- two colocated detectors at 45 deg, but triangular topology reaches same sensitivity

- estimate ET event rate using galaxy merger trees: four different models for mass distribution and accretion history
- SNR with Ajith et al. waveforms (somewhat simple piece-wise amplitude expressions)

- several high-redshifts (> 7) events -> indicates pop-III star mergers
	- no events seen indicates seed BHs not from pop-III
	- events at z < 5 may come from IMBHs in clusters
	- LISA will see similar numbers anyway
- parameter estimation: good masses, 15-30% luminosity distance

=== Lisa Goggin - possibility of inspiral-ringdown test? ===

- NR gives information about mergers, but searches can only afford analytic waveforms based on NR results
- investigation: inject EOBNR waveforms, compare recovered inspiral and ringdown parameters
- test is viable

=== Ruslan Vaulin - in search for the CBC optimal statistic ===

- question: does data contain signal?
- formal optimal strategy: compute likelihood function, announce detection if larger than threshold
- Neyman-Pearson optimization: maximize detection probability at false-alarm probability

- goal is to move the detection statistic closer to the optimal decision curve with real data
	- SNR is only zeroth-order approximation to the true likelihood
	- other parameters (chi2, e-thinca radius) can supplement it
	- estimate p(data|signal) and p(data|noise) by injections and time slides
- search pipeline maps signal parameters to detection parameters

- brute force approach
	- p(data|signal) = number of found injections near candidate / injections
	- p(data|noise) = noise trials with triggers near candidate / noise trials
	- "near" -> same parameters with epsilon fractional error
	- multidimensional visualization: injections need to make it through "gates" of epsilon fractional width in each parameter
	- get few injection "hits", expensive computationally

- less restrictive: fixed gates in chirp mass and mu, semi-open neighborhood for effective SNR
- ranking based on marginalized likelihood

- evaluate efficiency of likelihood statistics in one month of data (w/same false alarm)
	- BNS:  11% -> 16% (rho-eff) -> 19% (epsilon box)
	- NSBH: 10% -> 14%
	- BBH:  7%  -> 8%

=== Walter del Pozzo - Bayesian model selection and tests of general relativity ===

- tests of GR with GWs: coalescing binaries with advanced ground-based, EMRIs, etc.
- search for no-GR signature in GW signal
	- use a Bayesian model selection approach
	- GR vs. massive graviton as test case
- model selection: look at Bayes factor---ratios of marginal likelihoods (integrated over all model parameters)
- graviton mass adds phasing term
- actually compute P(d|H_MG)/P(d|n) vs. P(d|H_GR)/P(d|n) and take ratio
	- find little discrimination

=== Jessica Clayton - joint-LV CBC lowmass search ===

- data between May and October 2007; five one-month periods
- 3 detectors allow source localization
- Mtot in [1,35] Msun; Virgo templates cover only BNS chirp-mass region (DQ vetoes not fully studied in high-mass region)
- range limited for Virgo data
- tune parameters by monitoring detection probability while restricting the cuts
- investigating new "bank" chi2 test, more computationally efficient, of comparable utility

=== Reinhard Prix - is the F-statistic optimal? ===

- parametrized CW signal family
	- amplitude parameters (h0,cos i,psi,phi0)
	- Doppler parameters
- strain at detector: F+(t,psi,n) A+ cos phi(t,lambda,phi0)
                    + Fx(t,psi,n) Ax sin phi(t,lambda,phi0)
- with A+x(h0,cos i)
- JKS 1998: factorize signal parameters s = sum_mu A^mu h_mu

- classical hypothesis testing: point hypothesis
	- Neyman-Pearson: likelihood ratio is "most powerful" test
	- plot ROC
- for F-statistic, hypothesis is not point-wise
	- Neyman-Pearson lemma does not apply
	- we're looking at maximum likelihood detection
	- F-statistic has nice properties, but is less powerful than the likelihood ratio
	- can it be improved?

- the F-statistic is equivalent to the Bayes factor (Bayesian detection statistic, integrated over A) with a flat prior in Amu coordinates
- redo with ignorance priors on Amu' = phi0, psi, cos i, h0 (= 1/h0)
	- in this coordinates the F-statistic prior is biased towards edge-on systems and large h0
	- define new detection statistic integrated over the Amu' priors
	- ROC curve can be above or below F-stat for specific sources: neither is uniformly more powerful
	- over isotropic random angles, B statistic is more powerful
	- but difference in sensitivity is not big, and has comparable sensitivity to B-statistic

- extend to better hypothesis-setup + priors to use odds ratio directly
- extend to marginalization over Doppler space, to eliminate problem of number of independent trials