Signal Detection and Parameter Estimation

Signal Detection and Parameter Estimation

Statistics 101: definitions

Bayes factor

\[p(m|D) = \frac{p(D|m)\,p(m)}{p(D)}\]
  • $p(m D)$ – probability of model $m$, given data $D$ (hard)
  • $p(D m)$ – probability of data, given model $m$ (easier)
  • $p(m)$ – prior (prior probability of given model)
  • $p(D)$ – normalisation

Evidence

\[p(D|m) = \int{\rm d}\theta\,{p(D|\theta,m)\,p(\theta|m)}\]
  • Also called marginal likelihood
  • Integrates likelihood over all model parameters
  • Penalises overly complex models (Occam factor)
  • Used to compare models

Likelihood

\[\mathcal{L}(\theta) = p(D|\theta,m)\]
  • Probability of observing data $D$
  • Assuming model $m$ with specific parameters $\theta$
  • Central quantity for parameter estimation
  • Often maximise $\mathcal{L}$ or $\log\mathcal{L}$

Posterior

\[p(\theta|D,m) = \frac{p(D|\theta,m)\,p(\theta|m)}{p(D|m)}\]
  • Probability of parameters after observing the data
  • Combines likelihood and prior
  • Central quantity in Bayesian inference

Interpretation:

  • peak → most probable parameters
  • width → parameter uncertainty

Odds ratio

Compare two models $m_1$ and $m_2$:

\[\frac{p(m_1|D)}{p(m_2|D)} = \frac{p(D|m_1)}{p(D|m_2)} \times \frac{p(m_1)}{p(m_2)}\]
  • Posterior odds = Bayes factor × prior odds
  • Bayes factor:
\[B_{12} = \frac{p(D|m_1)}{p(D|m_2)}\]
  • $B_{12} > 1$ favours $m_1$
  • $B_{12} < 1$ favours $m_2$

Priors

\[p(\theta|m)\]
  • Probability distribution for parameters before seeing the data
  • Encodes prior knowledge or assumptions
  • Must be specified in Bayesian inference

Examples:

  • Uniform prior — all values equally likely
  • Log prior — equal probability per decade
  • Informative prior — based on previous measurements

Frequentist vs Bayesian (signal searches)

Frequentist

  • Parameters $\theta$ are fixed but unknown
  • Data are random realisations of noise

Parameter estimation: maximise likelihood (or related statistic)

Detection:

  • compare null hypothesis (noise) with signal model
  • compute test statistic → obtain p-value

Decision rule:

\[p < p_{\rm thresh}\]

(e.g. $5\sigma$ significance)

Bayesian

  • Parameters $\theta$ are random variables
  • incorporate prior information

Parameter estimation:

  • use posterior distribution

Model comparison:

  • compare evidences using the Bayes factor

$\chi^2$ likelihood (Gaussian noise)

For data vector $\mathbf{d}$ and model prediction $\mathbf{m}(\theta)$:

\[\chi^2 = (\mathbf{d}-\mathbf{m})^{T} C^{-1} (\mathbf{d}-\mathbf{m})\]

Likelihood:

\[p(D|\theta,m) \propto \exp\!\left(-\frac{1}{2}\chi^2\right)\]
  • $C$ - noise covariance matrix
  • Includes correlations between data points

Special case: uncorrelated Gaussian noise

\[\chi^2 = \sum_i \frac{(d_i - m_i)^2}{\sigma_i^2}\]

Finding signals

Matched filter and signal-to-noise ratio (SNR)

Method for detecting a known signal shape in noise.

Inner product (overlap function):

\[(a|b) = a^T C^{-1} b\]
  • $C$ - noise covariance matrix $\approx\sigma^2$
  • weights data by inverse noise power

For constant uncorrelated noise:

\[(a|b) = \vec{a}\cdot\vec{b} / \sigma^2\]

Matched filter output:

\[z = (d|s)\]
  • $d$ - data
    • Assumed to be signal + noise:
    • $d = s + n$
  • $s$ - template signal
  • $s = A \bar s$
    • Only is cases where signal is linear in amplitude (most cases)
    • $\bar s$ is shape of signal, $A$ is amplitude

Interpretation:

  • weighted correlation of data with expected signal
  • Overlap of signal and data

Signal-to-noise ratio:

\[\rho = \frac{(d|s)}{\sqrt{(s|s)}} = \frac{(d|\bar s)}{\sqrt{(\bar s|\bar s)}}\]
  • normalised matched-filter output
  • dimensionless detection statistic

Expected SNR for true signal:

\[\rho_{\rm exp}^2 = (s|s)\]

Optimality:

  • matched filter maximises SNR in Gaussian noise

Detection:

  • candidate signal if
\[\rho > \rho_{\rm thresh}\]

Typical thresholds:

  • $\rho \sim 5$–$10$ depending on background and number of trials

Matched filter / SNR (brief derivation)

Assume additive noise:

\[d = A h + n\]

Gaussian likelihood, noise with covariance $C$:

\[p(d|A) \propto \exp\!\left[-\frac12(d-Ah)^T C^{-1}(d-Ah)\right]\]
Maximise $\ln p(d A)$ w.r.t. $A$:
\[\ln p(d|A) = \text{const} - \frac12(d-Ah|d-Ah)\]

Expand:

\[(d-Ah|d-Ah) = (d|d) - 2A(d|h) + A^2(h|h)\]

Differentiate and set to zero:

\[\frac{\partial}{\partial A}\Big[(d|d) - 2A(d|h) + A^2(h|h)\Big]=0\] \[-2(d|h) + 2A(h|h)=0 \quad\Rightarrow\quad \hat A = \frac{(d|h)}{(h|h)}\]

Define the matched-filter (amplitude) statistic:

\[z = (d|h)\]

Normalise by its noise RMS to get SNR:

Under noise-only ($d=n$),

\[\langle z \rangle = 0, \qquad {\rm Var}(z)=\langle (n|h)^2\rangle = (h|h)\]

So

\[\rho = \frac{z}{\sqrt{(h|h)}} = \frac{(d|h)}{\sqrt{(h|h)}}\]
  • matched filter: compute $(d h)$ (noise-weighted correlation)
  • SNR: matched filter output in units of its noise standard deviation

Template searches

Signal parameters unknown → search over templates:

\[s(t;\theta_k)\]

Procedure:

  1. Generate template bank
  2. Compute matched filter for each template
  3. Find maximum SNR
  • Candidate signals correspond to large peaks in SNR
    • Either have a threshold for how large we consider as the trigger
    • Or find the largest in a data set, investigate, etc.

Parameter estimation

Signal model:

\[s(t;\theta)\]

with parameters

\[\theta = \{\theta_1,\theta_2,\dots\}\]

Goal:

  • find parameters that best match the data

Statistic to maximise:

\[\mathcal{L}(\theta) = p(D|\theta) ,\quad \rho(\theta),\quad\text{etc.}\]

Interpretation:

  • parameters that maximise the detection statistic
  • give the best-fit signal model

Special case: unknown amplitude

If the signal has the form

\[s(t) = A\,\bar s(t)\]

then the best-fit amplitude can be found analytically:

\[\hat A = \frac{(d|\bar s)}{(\bar s|\bar s)}\]

So the matched filter automatically finds the optimal amplitude.


Estimating uncertainty

Uncertainty may determined by the shape of the likelihood peak around the best-fit parameters.

  • Very sharp peak: low uncertainty in best-fit value
  • Very narrow peak: high uncertainty

Approximate rule:

\[\Delta\chi^2 \sim 1\]

gives the $1\sigma$ uncertainty on a parameter.

Key idea:

  • sharper peak in likelihood → smaller uncertainty
  • uncertainty typically scales as
\[\sigma_\theta \propto \frac{1}{\text{SNR}}\]