\documentstyle [twoside]{article}
%
%-----------------------------------------------------------------------
%; Copyright (C) 1998
%; Associated Universities, Inc. Washington DC, USA.
%;
%; This program is free software; you can redistribute it and/or
%; modify it under the terms of the GNU General Public License as
%; published by the Free Software Foundation; either version 2 of
%; the License, or (at your option) any later version.
%;
%; This program is distributed in the hope that it will be useful,
%; but WITHOUT ANY WARRANTY; without even the implied warranty of
%; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
%; GNU General Public License for more details.
%;
%; You should have received a copy of the GNU General Public
%; License along with this program; if not, write to the Free
%; Software Foundation, Inc., 675 Massachusetts Ave, Cambridge,
%; MA 02139, USA.
%;
%; Correspondence concerning AIPS should be addressed as follows:
%; Internet email: aipsmail@nrao.edu.
%; Postal address: AIPS Project Office
%; National Radio Astronomy Observatory
%; 520 Edgemont Road
%; Charlottesville, VA 22903-2475 USA
%-----------------------------------------------------------------------
%
\newcommand{\DoMemo}{T}
\newcommand{\doFig}{T}
\newcommand{\showwork}{T}
\newcommand{\workyes}{F}
%
\newcommand{\memnum}{101}
\newcommand{\whatmem}{\AIPS\ Memo \memnum}
\newcommand{\AIPS}{{$\cal AIPS\/$}}
\newcommand{\VPOPS}{{$\cal VPOPS\/$}}
\newcommand{\RANCID}{{$\cal RANCID\/$}}
\if T\DoMemo
\newcommand{\memtit}{The Calculation of SNR in KRING's FFT stage }
\else
\newcommand{\memtit}{KRING - FFT SNR calculation }
\fi
%
\newcommand{\POPS}{{$\cal POPS\/$}}
\newcommand{\Cookbook}{${{\cal C}ook{\cal B}ook\/}$}
\newcommand{\TEX}{\hbox{T\hskip-.1667em\lower0.424ex\hbox{E}\hskip-.125em X}}
\newcommand{\AMark}{AIPSMark}
\newcommand{\AMarks}{AIPSMarks}
\newcommand{\figyes}{T}
\newcommand{\uv}{{\it uv}}
\newcommand{\eg}{{\it e.g.},}
\newcommand{\ie}{{\it i.e.},}
\newcommand{\daemon}{d\ae mon}
\newcommand{\Aipsletter}{{${\cal AIPSL}etter\/$}}
\newcommand{\ust}{{\rm st}}
\newcommand{\uth}{{\rm th}}
\newcommand{\und}{{\rm nd}}
\newcommand{\urd}{{\rm rd}}
%
\newcommand{\boxit}[3]{\vbox{\hrule height#1\hbox{\vrule width#1\kern#2%
\vbox{\kern#2{#3}\kern#2}\kern#2\vrule width#1}\hrule height#1}}
%
\title{
\vskip -35pt
\if T\DoMemo
\fbox{{\large\whatmem}} \\
\fi
\vskip 28pt
\memtit\\}
\author{Ketan M. Desai}
%
\parskip 4mm
\linewidth 6.5in
\textwidth 6.5in % text width excluding margin
\textheight 8.81 in
\marginparsep 0in
\oddsidemargin .25in % EWG from -.25
\evensidemargin -.25in
\topmargin -.5in
\headsep 0.25in
\headheight 0.25in
\parindent 0in
\newcommand{\normalstyle}{\baselineskip 4mm \parskip 2mm \normalsize}
\newcommand{\tablestyle}{\baselineskip 2mm \parskip 1mm \small }
\input psfig
%
%
\begin{document}
\pagestyle{myheadings}
\thispagestyle{empty}
\if T\DoMemo
\newcommand{\Rheading}{\whatmem \hfill \memtit \hfill Page~~}
\newcommand{\Lheading}{~~Page \hfill \memtit \hfill \whatmem}
\else
\newcommand{\Rheading}{K. M. Desai\hfill \memtit \hfill Page~~}
\newcommand{\Lheading}{~~Page \hfill \memtit \hfill K. M. Desai}
\fi
\markboth{\Lheading}{\Rheading}
%
%
\vskip -.5cm
\pretolerance 10000
\listparindent 0cm
\labelsep 0cm
%
%
\vskip -30pt
\maketitle
%\vskip -30pt
\normalstyle
\begin{abstract}
This memo describes the SNR calculation during the FFT stage of KRING.
I describe the assumed statistical properties of the visibilities used
for the fringe search and show how KRING's SNR estimate differs from
that in FRING. I also discuss the probability of false detection, a
statistic users have often requested be reported while fringe-fitting.
\end{abstract}
\section{Introduction}
The estimation of residual phase variations in VLBI data is of great
practical importance. Residual phase variations occur for a variety
of reasons including antenna and source position errors in the
correlator model, source structure, ionospheric and atmospheric
effects as well as thermal noise. Linear residual variations of phase
with frequency and time are called residual delays and rates.
Fringe-searching or fringe-fitting is the process of estimating
residual delays and rates from time/frequency data, or dynamic
spectra. Schwab and Cotton\footnote{Schwab, F. \& Cotton, B., ``
Global Fringe-Fitting Search Techniques for VLBI'', AJ, 688, {\bf 88}, 1983.} proposed
a two stage algorithm to construct the best possible estimate of the
residual phase variations using all available information. In the
first stage, dynamic spectra are Fast Fourier Transformed one baseline
at a time to determine an initial antenna-based model of the phase
variations\footnote{See the note about baseline stacking in the next
section.}; this is the fringe detection stage. In the second stage,
a Least Squares Fit is performed using data from all baselines to
simultaneously refine all model parameters; this is the fringe
refinement stage. The fringe detection stage is the subject of this
memo.
An important component of the Schwab \& Cotton algorithm is the SNR
estimate of the fringe detection. The SNR estimate in FRING is based
upon simulations by B. Cotton. In the fringe-detection stage, after Fourier Transforming the dynamic spectra, peaks
in the resultant delay-rate spectrum are `searched' to find the one
with the greatest amplitude. If the SNR estimated using this peak
amplitude exceeds some threshold, the residual phase, rate, and delay
are accepted as part of the initial model. The correctness of the SNR
computation is important because it is used to determine whether or
not to accept the search results for a given baseline. If the
computation incorrectly reports a high SNR, erroneous model values may
be passed to the fringe refinement stage. If the computation
incorrrectly reports a low SNR, acceptable model values may be
discarded, forcing more baselines to be searched, increasing
computation time. If the computation systematically reports low SNRs,
the FFT stage may fail to pass model values for some antennas,
resulting in lost data. The common wisdom is that the present
calculation in FRING is approximately correct at low SNRs but is an
underestimate at high SNR. I present a formal analysis of SNR
estimation in section 2. I will show that the common wisdom is
incorrect at low SNR, this is illustrated by example in section 2.
There are three implementations of the Schwab and Cotton algorithm in
\AIPS, FRING, BLING/BLAPP, and KRING. The analysis in this memo has
been implemented in KRING to compute SNRs during the fringe detection
stage.
The time/frequency sampling rate of a dynamic spectrum determines the
delay/rate range of its FFT; the exact relation between the sampling
rate and the corresponding range is determined by the Nyquist theorem.
The delay-rate peak is the point with the largest amplitude inside a
search window, or range of delays and rates, specified by the user.
If too large a search window is specified, a noise spike may be found
that, by chance, has amplitude greater than the true delay-rate peak;
this is a false detection. If too small a search window is specified,
the true delay-rate peak may lie outside the search window; any point
found within the search window will be a false detection. The
probability of false detection is of interest for deciding whether or
not the search windows need to be made smaller or increased; it is not
reported by FRING, BLING, or KRING. Although only a small fraction of
the full Nyquist range may have been specified for the search window,
FRING and BLING still must calculate full-size FFTs while KRING
actually performs smaller FFTs when smaller search windows are given,
decreasing computation time\footnote{In KRING, If the delay search window is
smaller than the rate search window, the frequency axis is first FFT'd to
delay, then only those times that lie within the the delay search window are
FFT'd into rates. If the rate search window is smaller than the delay search,
the order of the FFTs is reversed appropriately. This offers tremendous
savings in execution time when the search is over a restricted range of delay/
rate space.}. I present a brief analysis of the
probability of false detection in section 3.
\subsection{Assumptions}
Signal processing of the voltages received at each antenna can, in
principle, introduce biases in the visibility amplitude and more
importantly in the visibility phase. Post-correlation processing can
correct for amplitude and phase biases but cannot recover from all
associated sensitivity losses. I assume that all amplitude and phase
biases have been removed, leaving only a bias due to an additive noise
component, before fringe-searching. This is a common assumption when
analyzing visibility data\footnote{ see {\it e.g.\/} chapter 6 of
Thompson, Moran, and Swenson, Interferometry and Synthesis in Radio
Astronomy, Krieger Publishing Co., 1991}. I also assume that any
sensitivity losses are properly incorporated into the weights so that
each weight properly reflects the noise in its associated measurement.
The dynamic spectrum produced by the correlator and presented for
fringe-searching includes amplitude and phase variations due to source
structure as well as non-linear phase variations. Schwab and Cotton
offered three alternative approaches to the nominal Least Squares
formulation of the fringe search problem in which the model and the
data are altered for reasons of efficiency or convenience. In this
memo, I compare the standard pre-Cotton and Schwab method and the
method actually implemented in FRING, both of which are described in
the next section.
Generally, analyses of fringe-searching techniques assume that the
source is a point source. This is, of course, usually not true. If
the data are divided by a good model of the source, the resultant
visibilities will reflect a source with unit strength at the phase
center, {\it i.e.} a point source. I assume that such a model is
available and has been divided out of the data. It must be remembered
however that the model division distorts the noise distribution,
magnifying the noise wherever the model is of low amplitude. If during the
model division phase, the weights are also divided by the square of the
model amplitude, this procedure will not bias the location of the peak in
delay/rate space. The weights are not
modified when model division is peformed in FRING or KRING. I do not
address the effect of this omission in this memo.
Schwab and Cotton introduced stacking, a technique for increasing
sensitivity when fringe-searching weak sources. The analysis in this
memo does not address the subtleties introduced by stacking.
Besides estimating the residual delay and rate in the data, it is of
interest to determine the signal-to-noise ratio of the final estimate and
the probability of false detection. The latter is more easily
calculated if it is assumed that the true residual delays and rates
are actually zero, {\it i.e.} there are no residual phase variations
present in the input visibilities. Note that as long as this
assumption is not used in the fringe-searching procedure, it does not
result in any loss of generality for the final results. I therefore
assume that the true residual delay and rate in the data are both
zero.
All the analyses in this memo involve averaging random quantities that
are either Gaussianly distributed or satisfy the Central Limit Theorem
criteria. I assume that the averaged quantities are Gaussianly
distributed and so can be described entirely using the first and
second moments.
In summary, I analyze the estimation of SNR and the probability of
false detection in this memo using the following assumptions:
$\bullet$ The visibilities are well behaved; besides additive noise,
the data is bias-free in amplitude and phase.
$\bullet$ The weights attached to each visibility properly reflect the
noise associated with each measurement.
$\bullet$ The data reflects a point source. Any source structure has
been pre-divided out of the visibility data.
$\bullet$ The random variables analyzed in this memo are completely
characterized by their first and second moments.
\subsection{Notation}
The complex visibility function, $V$, is sampled at discrete times and
frequencies and is modelled as a constant real signal $S$ plus
complex noise $n_r + jn_i$. The former is independent of time and
frequency while the latter is independent between one time/frequency
sample and others.
This assumes that any source structure has been divided out, as
discussed in the last section. The complex noise is distributed as a
Gaussian with half-width at half-max $\sigma\sqrt{\ln{8}}$ in its real and
imaginary parts.
The probability density distributions for the real and imaginary part
of $V$ are Gaussians centered upon $S$ and zero respectively. The
probability density distribution for the phase is obtained by
integrating over all possible signal amplitudes. Here are the
probability distributions for the real part, imaginary part, and phase
of the complex visibilities:
$$p_{V_r}(r) = {1\over\sqrt{2\pi\sigma^2}}e^{-{(r-S)^2\over2\sigma^2}} \qquad
p_{V_i}(i) = {1\over\sqrt{2\pi\sigma^2}}e^{-{i^2\over2\sigma^2}} \qquad
p_\Phi(\phi) = {e^{-{1\over2}k^2}\over2\pi}\int\limits_0^\infty xdx
e^{-{1\over2}x^2 + xkcos\phi}$$
Throughout this memo, $k=S/\sigma$ is the signal-to-noise ratio {\em
for an individual visibility}. In some places, $\gamma = k^2/4$ is
used for convenience.
I use the `symmetric' form of the Fourier transform in this memo
$$M_U(\zeta) = \langle e^{ju\zeta}\rangle =
\int_up_U(u)e^{i2\pi u\zeta}du$$
except in section 2 where, to avoid a plethora of `$2\pi$'s, I use the
`asymmetric' form:
$$M_U(\zeta) = \langle e^{ju\zeta}\rangle =
\int_up_U(u)e^{iu\zeta}du$$
The pre-Schwab and Cotton approach was to fringe-search the
visibilities directly. I denote by $Y_{\omega\tau}$ the elements of
the Fourier Transform of the visibilities.
$$Y_{\omega\tau} \equiv{1\over TF}\sum\limits_{f=1,F\atop t=1,T}
w_{tf}V_{tf}e^{2\pi jf\tau/D + 2\pi jt\omega/R}$$
The sums here are over T time and F frequency points, $\omega$ takes
rate values between $0$ and $R-1$, and $\tau$ takes delay values
between $0$ and $D-1$. The weights are denoted by $w_{tf}$. The
first and second moments, and variance of $w_{tf}$ are denoted by
$\langle w\rangle$, $\langle w^2\rangle$, and $\sigma_w^2 = \langle
w^2\rangle-\langle w\rangle^2$.
The method actually implemented by Schwab and Cotton in FRING uses a
normalized form of the visibilities. I denote by $X_{\omega\tau}$ the
elements of the Fourier Transform of the normalized visibilities
according to the following formula.
$$X_{\omega\tau} \equiv{1\over TF}\sum\limits_{f=1,F\atop t=1,T}
w_{tf}{V_{tf}\over|V_{tf}|}e^{2\pi jf\tau/D + 2\pi jt\omega/R}
= {1\over TF}\sum\limits_{f=1,F\atop t=1,T}
w_{tf}e^{j\phi_{tf} + 2\pi jf\tau/D + 2\pi jt\omega/R}$$
When $\omega = \tau = 0$, I will write
$$X \equiv X_{00} = {1\over TF}\sum\limits_{f=1,F\atop t=1,T}w_{tf}e^{j\phi_{tf}}
\qquad
Y \equiv Y_{00} = {1\over TF}\sum\limits_{f=1,F\atop t=1,T} w_{tf}V_{tf}$$
I will also write $X_r$ and $X_i$ for the real and imaginary parts of
$X$ while $X_a^2 = X_r^2 + X_i^2$ and $X_{ri} = X_rX_i$. Similar
notation applies for $Y$ and when $\omega$ and $\tau$ are non-zero, {\it e.g.}, $Y_r$, $X_{\omega\tau r}$, {\it etc}.
The rate and delay indices $\omega$ and $\tau$ are assumed to be
integers except as otherwise noted. The points in the delay-rate
plane are also sometimes called delay-rate cells or just cells.
\section{Statistics of the true delay-rate peak in the delay-rate plane}
In order to derive the statistics of the true delay-rate peak in the
delay-rate plane, I first compute the moments of $X$ and $Y$. These
moments are easily computed if I first compute the Fourier Transforms
of the probability functions of $X$ and $Y$.
The Fourier Transform of $p_{Y_r}(y_r)$, $p_{Y_i}(y_i)$ and $p_\Phi(\phi)$ are:
$$M_{Y_r}(\zeta) = e^{jk\zeta\sigma - {1\over2}\zeta^2\sigma^2} \qquad
M_{Y_i}(\zeta) = e^{-{1\over2}\zeta^2\sigma^2}$$
$$M_\Phi(\zeta) = \int\limits_{-\pi}^\pi d\phi p_\Phi(\phi)e^{j\zeta\phi}
= e^{-{1\over2}k^2}\int\limits_0^\infty xdxe^{-{1\over2}x^2}I_\zeta(kx)
= e^{-2\gamma}\delta(\zeta) +
\sqrt{\pi\gamma\over2}e^{-\gamma}
\left(I_{|\zeta-1|\over2}(\gamma) +
I_{|\zeta+1|\over2}(\gamma) \right)$$
Here $I_\zeta$ is the modified Bessel function of integer
order\footnote{as defined in Eqn. 9.6.20 of Abramowitz and Stegun,
Handbook of Mathematical Functions, Dover Publications,
1965.}\footnote{The integral was evaluated using A.\&S. Eqns. 9.6.26
and 11.4.31}. Note that $M_\Phi$ is even in $\zeta$.
Evaluating $M_{\Phi}$ at different values of $\zeta$ yields the
necessary moments of $X$ but computation of the moments of $Y$
requires the first and second derivatives of $M_{X_r}(\zeta)$ and
$M_{X_i}(\zeta)$ with respect to $\zeta$:
\vspace{-10pt}
\begin{center}
\begin{tabular}{llll}
$M_{Y_r}'(\zeta) = (jk\sigma-\zeta\sigma^2)e^{jk\zeta\sigma-{1\over2}\zeta^2\sigma^2} $&$
M_{Y_i}'(\zeta) = -\zeta\sigma^2e^{-{1\over2}\zeta^2\sigma^2}$ \\
$M_{Y_r}''(\zeta) = \sigma^2\left[\zeta^2\sigma^2 - k^2 - 1 - 2jk\zeta\sigma\right]e^{jk\zeta\sigma - {1\over2}\zeta^2\sigma^2} $&$
M_{Y_i}''(\zeta) = \sigma^2(\zeta^2\sigma^2-1)e^{-{1\over2}\zeta^2\sigma^2}$
\end{tabular}
\end{center}
\vspace{-10pt}
The moments of $X$ and $Y$ are summarized below. The analysis of
$X$'s moments is based upon the treatment found in Goodman's
Statistical Optics, Appendix B\footnote{Goodman, J., {\it Statistical Optics}, Wiley Interscience, 1985.}:
\vspace{-10pt}
\begin{center}
\begin{tabular}{llll}
$ \langle X_r\rangle $&$= \langle w\rangle M_\Phi(1)$ &
$ \langle Y_r\rangle $&$ = -j\langle w\rangle M_r'(0)$ \\
$ \langle X_i\rangle $&$= 0$ &
$ \langle Y_i\rangle $&$ = -j\langle w\rangle M_i'(0)$ \\
$ \langle X_r^2\rangle $&$= {\langle w^2\rangle \over2TF}\left[1+M_\Phi(2)\right] + \langle w\rangle ^2{TF-1\over TF}M_\Phi^2(1)$ &
$ \langle Y_r^2\rangle $&$ = -{\langle w^2\rangle \over TF}M_r''(0) - \langle w\rangle ^2{TF-1\over TF}M_r^{'2}(0)$ \\
$ \langle X_i^2\rangle $&$= {\langle w^2\rangle \over2TF}\left[1-M_\Phi(2)\right]$ &
$ \langle Y_i^2\rangle $&$ = -{\langle w^2\rangle \over TF}M_i''(0) - \langle w\rangle ^2{TF-1\over TF}M_i^{'2}(0)$ \\
$ \langle X_{ri}\rangle $&$= 0$ &
$ \langle Y_{ri}\rangle $&$ = 0$ \\
$ \langle X_a^2\rangle $&$= {\langle w^2\rangle \over TF} + \langle w\rangle ^2{TF-1\over TF}M_\Phi^2(1)$ &
$ \langle Y_a^2\rangle $&$ = -{\langle w^2\rangle \over TF}(M_r''(0)+M_i''(0)) - \langle w\rangle ^2{TF-1\over TF}(M_r^{'2}(0)+M_i^{'2}(0))$
\end{tabular}
\end{center}
\vspace{-10pt}
Plugging in the FT's, evaluated at the apropropiate values, these moments become:
\vspace{-10pt}
\begin{center}
\begin{tabular}{llll}
$ \langle X_r\rangle = \langle w\rangle G(\gamma)$ &
$ \langle Y_r\rangle = \langle w\rangle S$ \\
$ \langle X_i\rangle = 0$ &
$ \langle Y_i\rangle = 0$ \\
$ \langle X_r^2\rangle = {\langle w^2\rangle \over2TF}\left[2 - \left({\sinh\gamma\over e^{\gamma}\gamma}\right)\right] + \langle w\rangle ^2{TF-1\over TF}G^2(\gamma)$ &
$ \langle Y_r^2\rangle = {\langle w^2\rangle \over TF}(\sigma^2 + S^2) + \langle w\rangle ^2{TF-1\over TF}S^2$ \\
$ \langle X_i^2\rangle = {\langle w^2\rangle \over2TF}\left[\left({\sinh\gamma\over e^{\gamma}\gamma}\right)\right]$ &
$ \langle Y_i^2\rangle = {\langle w^2\rangle \over TF}\sigma^2$ \\
$ \langle X_{ri}\rangle = 0$ &
$ \langle Y_{ri}\rangle = 0$ \\
$ \langle X_a^2\rangle = {\langle w^2\rangle \over TF} + \langle w\rangle ^2{TF-1\over TF}G^2(\gamma)$ &
$ \langle Y_a^2\rangle = {\langle w^2\rangle \over TF}(S^2 + 2\sigma^2) + \langle w\rangle ^2{TF-1\over TF}S^2$
\end{tabular}
\end{center}
\vspace{-10pt}
where $G(\gamma) = M_\Phi(1) = \sqrt{\pi\gamma\over2}e^{-\gamma}
\left(I_0(\gamma) + I_1(\gamma) \right)$. Note that the
cross-correlations $X_{ri}$ both vanish so that $X_r$ and $X_i$ are
uncorrelated.
\subsection{The limiting cases of weak and strong signal}
The limiting forms of the moments are summarized below for the weak
and strong signal limits.
In the weak signal limit, the moments have the following functional
forms:
\vspace{-10pt}
\begin{center}
\begin{tabular}{llll}
$\langle X_r\rangle = \langle w\rangle \sqrt{\pi\over8}k $ &
$\langle Y_r\rangle = \langle w\rangle k\sigma $ \\
$\langle X_i\rangle = 0 $ &
$\langle Y_i\rangle = 0 $ \\
$\langle X_r^2\rangle = \left[\langle w^2\rangle + \langle w\rangle^2(TF-1){\pi\over4}k^2\right]{1\over2TF} $ &
$\langle Y_r^2\rangle = \left[\langle w^2\rangle + \langle w\rangle^2(TF-1)k^2\right]{1\over2TF}2\sigma^2 $ \\
$\langle X_i^2\rangle = \langle w^2\rangle{1\over2TF} $ &
$\langle Y_i^2\rangle = \langle w^2\rangle{1\over2TF}2\sigma^2 $ \\
$\langle X_{ri}\rangle = 0 $ &
$\langle Y_{ri}\rangle = 0 $ \\
$\langle X_a^2\rangle = \left[2\langle w^2\rangle + \langle w\rangle^2(TF-1){\pi\over4}k^2\right]{1\over2TF} $ &
$\langle Y_a^2\rangle = \left[2\langle w^2\rangle + \langle w\rangle^2(TF-1)k^2\right]{1\over2TF}2\sigma^2 $
\end{tabular}
\end{center}
\goodbreak
In the strong signal limit, the moments have the following functional
forms:
\vspace{-10pt}
\vspace{-10pt}
\begin{center}
\begin{tabular}{llll}
$\langle X_r\rangle = \langle w\rangle $ &
$\langle Y_r\rangle = \langle w\rangle S $ \\
$\langle X_i\rangle = 0 $ &
$\langle Y_i\rangle = 0 $ \\
$\langle X_r^2\rangle = \left[\langle w^2\rangle + \langle w\rangle^2(TF-1)\right]{1\over TF} $ &
$\langle Y_r^2\rangle = \left[\langle w^2\rangle + \langle w\rangle^2(TF-1)\right]{1\over TF}S^2 $ \\
$\langle X_i^2\rangle = \langle w^2\rangle{1\over k^2TF} $ &
$\langle Y_i^2\rangle = \langle w^2\rangle{1\over k^2TF}S^2 $ \\
$\langle X_{ri}\rangle = 0 $ &
$\langle Y_{ri}\rangle = 0 $ \\
$\langle X_a^2\rangle = \left[\langle w^2\rangle + \langle w\rangle^2(TF-1)\right]{1\over TF} $ &
$\langle Y_a^2\rangle = \left[\langle w^2\rangle + \langle w\rangle^2(TF-1)\right]{1\over TF}S^2 $
\end{tabular}
\end{center}
\vspace{-10pt}
At low SNRs, the square of the delay-rate peak's amplitude is slightly
lower when using normalized visibilities instead of the traditional
visibilities (as compared to the `off-peak' rms noise). This may
result in some small loss in sensitivity for weak sources.
\subsection{Estimating SNR from the statistics of the delay-rate peak}
Assuming that the fringe-search has found the origin as the highest
point in the delay-rate plane, we are still not assured that the
amplitude of the peak at the origin is the amplitude of the true
delay-rate peak. This is because the amplitude at the origin is one
realization of the possible values at the origin. Still, this
realization is the best estimate we have of the true delay-rate peak's
amplitude. It is easier to estimate the SNR ($k$) using the square of the
amplitude of the delay-rate peak found by the fringe search.
If the expressions for $\langle X_a^2\rangle$ and $\langle
Y_a^2\rangle$ are rearranged, we have:
$$ {\langle X_a^2\rangle TF - \langle w^2\rangle \over \langle w\rangle ^2 (TF-1)} = G^2({k^2\over4}) \qquad
{\langle Y_a^2\rangle TF/\sigma^2 - 2\langle w^2\rangle \over\langle w^2\rangle + \langle w\rangle ^2(TF-1)} = k^2$$
Note that in each expression, the left-hand sides are composed
entirely of known quantities with the possible exception of
$\sigma^2$. On the right-hand sides, $G^2({k^2\over4})$ and $k^2$ can
be inverted, numerically if not directly.
The second equation can be simply inverted to determine the SNR ($k$),
but requires estimation of the noise, $\sigma$. The first equation
involves the inverse of $G$ which must be carried out numerically, or
using a look-up table, but involves only known quantities.
When an {\it a priori} estimate of the noise level is not available,
the SNR is easier to estimate when the input visibilities are
normalized to unit amplitude. This procedure also renders the SNR
estimate less susceptible to amplitude errors in the source model.
\subsection{Comparison of the present analysis with that in FRING}
The mapping of the squared-amplitude to SNR as implemented in FRING is
here compared to that derived in the previous section.
Figure~\ref{fig:FRINGtoKRING} plots the mapping function currently
used in FRING and that derived in this memo for use in KRING. Given
an observed peak value on the y-axis, the SNR is to be read off of the
x-axis. Note that the x-axis is the SNR {\it per visibility} and
should be multiplied by the square root of the number of visibilities
to determine the SNR of the fringe search, the number that is reported
by FRING and KRING. For example, in a solution interval of 10
minutes, with 2 second integrations and 16 channels across an IF, 4800
visibilities are used in the fringe search. Suppose a delay-rate peak
squared-amplitude of 0.15 is seen. This corresponds to an SNR {\em
per visibility} of 0.65. The actual SNR reported by FRING or KRING
would be $0.65\times\sqrt{4800} = 45$. For a larger delay-rate peak,
FRING would report a larger SNR than KRING. For a smaller delay-rate
peak, FRING would report a smaller SNR than KRING.
Figure~\ref{fig:FRINGoverKRING} shows the overestimation factor for
FRING versus KRING. Continuing the example of the last paragraph, if
FRING reported an SNR of $5$, that corresponds to an squared-amplitude
of $\approx 0.002$ for which KRING would report an SNR of $7.25$.
\begin{figure}
\if\doFig\figyes
\centerline{\hss\psfig{figure=FIG/AM101A.EPS,height=8.0in}\hss}
\else
\vskip 8.0in
\fi
\caption{ The solid line shows the mapping of squared-amplitude used
by FRING to estimate SNR while the dashed line shows the mapping
derived in this memo. The squared-amplitude of the delay-rate peak on
the y-axis is to be mapped to an SNR on the x-axis. The abcissa is
the SNR {\it per visibility} and should be multiplied by the square
root of the number of visibilities used in the fringe search to obtain
the value actually reported by FRING/KRING.}
\label{fig:FRINGtoKRING}
\end{figure}
\begin{figure}
\if\doFig\figyes
\centerline{\hss\psfig{figure=FIG/AM101B.EPS,height=8.0in}\hss}
\else
\vskip 8.0in
\fi
\caption{
The factor by which the SNR reported by FRING exceeds that derived in
this memo for a given fringe-search.}
\label{fig:FRINGoverKRING}
\end{figure}
\section{Probability of false detection}
There are two ways in which a fringe-search results in a false
detection. The first is when there is actually no signal but the
fringe-search yields a `detection' with significant estimated SNR.
The second is when there is a signal present with some delay and rate
residual but the fringe-search yields a `detection' at a different
delay and rate residual.
The traditional analysis of the probability of false detection assumes
that signal is present only in one cell in delay-rate space. In this
case the two types of false detection described above are identical.
If any cell besides the one containing signal is chosen by the
fringe-search procedure, it is a false detection. Analysis of the
probability of false detection is described in many
places\footnote{see {\it e.g.} Thompson, Moran, \& Swenson}. I begin
by summarizing the traditional analysis, then present a similar
analysis using normalized visibilities, and conclude with a more
general analysis in which signal is allowed to be present in more than
one cell in the fringe-search. In the latter analysis, the root
causes of the two types of false detection can be differentiated.
\subsection{Traditional analysis of the probability of false detection (PFD)}
I begin by reviewing the traditional analysis of the PFD, in which the
input visibilities $Y$ are directly Fourier Transform'ed.
First, we need the statistics of points in the delay-rate plane when
signal is and isn't present. Since $Y$ is formed by averaging
Gaussianly distributed random variables, its real and imaginary parts
obey the central limit theorem. The statistics of the real and
imaginary parts of $Y$ are described by:
$$P_{Y_r}(y_r) = {1\over\sqrt{2\pi}\sigma_r}
e^{-{(y_r - \langle w\rangle S)^2\over2\sigma_r^2}} \qquad
P_{Y_i}(y_i) = {1\over\sqrt{2\pi}\sigma_i}
e^{-{y_i^2\over2\sigma_i^2}}$$
where $\sigma_r^2 = {\langle w^2\rangle \sigma^2 + \sigma_w^2S^2\over
TF}$ and $\sigma_i^2 = {\langle w^2\rangle \sigma^2\over TF}$. It can
be shown that $Y_r$ and $Y_i$ are un-correlated\footnote{see Appendix
B of Goodman's Optics book}.
In the absence of signal, these probability distributions become:
$$P_{Y_r}(y_r)|_{k=0} = {1\over\sqrt{2\pi}\sigma_a}
e^{-{y_r^2\over2\sigma_a^2}} \qquad
P_{Y_i}(y_i)|_{k=0} = {1\over\sqrt{2\pi}\sigma_a}
e^{-{y_i^2\over2\sigma_a^2}}$$
where $\sigma_a^2 = {\langle w^2\rangle \sigma^2\over TF}$.
What is the probability that the squared amplitude of an off-center
delay-rate cell, $y_{\omega\tau a}^2$, is less than the squared
amplitude of the delay-rate origin, $y_{00a}^2$ ? This is just:
$$P(y_{\omega\tau a}^2 < y_{00a}^2)
= \int\limits_{y_{\omega\tau a}^2 < y_{00a}^2}dy_rdy_iP_{Y_r}(y_r)P_{Y_i}(y_i)
= 1 - e^{-{y_{00a}^2\over2\sigma_a^2}}$$
Now what is the probability that jointly, for a search window in the
delay-rate plane consisting of D delays by R rates, no point exceeds
the true delay-rate peak in amplitude-squared?
$$P(y_{00a}^2 > y^2_{\omega\tau a}: \forall \omega\neq0, \tau\neq0 )
= {1\over2\pi\sigma_r\sigma_a} \int dy_rdy_i
e^{-{(y_r - \langle w\rangle S)^2\over2\sigma_r^2}-{y_i^2\over2\sigma_a^2}}
\left[1 - e^{-{y_r^2+y_i^2\over2\sigma_a^2}}\right]^{RD-1}$$
This is the probability of a true detection. Subtracting it from 1
yields the PFD:
$$PFD= \sum\limits_{n=1}^{RD-1}{\Gamma(RD+1)\over\Gamma(n+1)\Gamma(RD-n+1)}(-1)^{n+1}
{\sigma_ae^{-{\langle w\rangle^2S^2\over2}{n\over\sigma_a^2+n\sigma_r^2}}\over\sqrt{(1+n)(\sigma_a^2+n\sigma_r^2)}}$$
Setting $\langle w\rangle = \langle w^2\rangle = 1$, which corresponds
to not using any weights for the input visibilities, this is the PFD
of Eqn. 9.61 in Thompson, Moran, and Swenson.
\subsection{PFD using normalized visibilities}
I now calculate the PFD when the input visibilities are first
normalized to unit amplitude to form $X$ which is then Fourier
Transform'd.
Again, we need the statistics of points in the delay-rate plane when
signal is and isn't present. As described in Appendix B of Goodman,
the real and imaginary parts of $X$ obey the central limit theorem
even if the signal disappears and the phase of $X$ is uniformly
distributed. The statistics of the real and imaginary parts of $X$
are:
$$P_{X_r}(x_r) = {1\over\sqrt{2\pi}\sigma_r}
e^{-{(x_r - \langle w\rangle G(\gamma))^2\over2\sigma_r^2}} \qquad
P_{X_i}(x_i) = {1\over\sqrt{2\pi}\sigma_i}
e^{-{x_i^2\over2\sigma_i^2}}$$
where
$\sigma_r^2 = {\langle w^2\rangle (1-{sinh\gamma\over2\gamma e^\gamma}) - \langle w\rangle ^2G^2(\gamma)\over TF}$ and
$\sigma_i^2 = {\langle w^2\rangle \over2TF}({sinh\gamma\over2\gamma e^\gamma})$.
In the absence of signal, these probability distributions become:
$$P_{X_r}(x_r)|_{k=0} = {1\over\sqrt{2\pi}\sigma_a}
e^{-{x_r^2\over2\sigma_a^2}} \qquad
P_{X_i}(x_i)|_{k=0} = {1\over\sqrt{2\pi}\sigma_a}
e^{-{x_i^2\over2\sigma_a^2}}$$
where
$\sigma_a^2 = {\langle w^2\rangle \over2TF}$.
What is the probability that the squared amplitude of an off-center
delay-rate cell, $x_{\omega\tau a}^2$, is less than the squared
amplitude of the delay-rate origin, $x_{00a}^2$ ? This is just:
$$P(x_{\omega\tau a}^2 < x_{00a}^2) = 1 - e^{-{x_{00a}^2\over 2\sigma_a^2}}$$
Now what is the probability that jointly, for a search window in the delay
rate plane consisting of D delays by R rates, no point exceeds the
true delay-rate peak in amplitude-squared?
$$P(x_{00a}^2 > x^2_{\omega\tau a}: \forall \omega\neq0, \tau\neq0 )
= {1\over2\pi\sigma_r\sigma_i} \int dx_rdx_i
e^{-{(x_r-\langle w\rangle G(\gamma))^2\over2\sigma_r^2}-{x_i^2\over2\sigma_i^2}}
\left[1 - e^{-{x_r^2+x_i^2\over2\sigma_a^2}}\right]^{RD-1}$$
So, the probability of false detection is
$$PFD= \sum\limits_{n=1}^{RD-1}{\Gamma(RD+1)\over\Gamma(n+1)\Gamma(RD-n+1)}(-1)^{n+1}
{\sigma_a^2e^{-{\langle w\rangle^2G^2(\gamma)\over2}{n\over\sigma_a^2+n\sigma_r^2}}\over\sqrt{(\sigma_a^2+n\sigma_i^2)(\sigma_a^2+n\sigma_r^2)}}$$
\subsection{Generalized analysis of the PFD}
I now analyze the statistics for the generalized case when signal is
present in more than one cell in the delay-rate search. I also
discuss the conditions under which signal occupies more than one cell
in the delay-rate search.
First, we need the statistics of an arbitrary point in delay-rate
space in terms of the input visibilities.
$$X_{\omega\tau} = {1\over TF}\sum\limits_{t=1,T\atop f=1,F}
w_{tf}e^{j(\phi_{tf}+2\pi{t\omega/R}+2\pi{f\tau/D})} \quad
Y_{\omega\tau} = {1\over TF}\sum\limits_{t=1,T\atop f=1,F}
w_{tf}V_{tf}e^{j2\pi t\omega/R+j2\pi f\tau/D}$$
What are the moments of the real and imaginary parts of
$X_{\omega\tau}$ at an arbitrary point in delay-rate space?
$$\langle X_{\omega\tau r}\rangle
= M_\Phi(1)\langle w\rangle Q_{\omega\tau}\cos\psi$$
$$\langle X_{\omega\tau i}\rangle
= M_\Phi(1)\langle w\rangle Q_{\omega\tau}\sin\psi$$
$$\langle X_{\omega\tau ri}\rangle
= \langle X_{\omega\tau r}\rangle\langle X_{\omega\tau i}\rangle + HQ_{2\omega2\tau}sin2\psi$$
$$\langle X_{\omega\tau r}^2\rangle
= \langle X_{\omega\tau r}\rangle ^2+ {\langle w^2\rangle -\langle w\rangle ^2M^2_\phi(1)\over2TF} + HQ_{2\omega2\tau}cos2\psi$$
$$\langle X_{\omega\tau i}^2\rangle
= \langle X_{\omega\tau i}\rangle ^2+ {\langle w^2\rangle -\langle w\rangle ^2M^2_\phi(1)\over2TF} - HQ_{2\omega2\tau}cos2\psi$$
$$\langle X_{\omega\tau a}^2\rangle
= \langle X_{\omega\tau r}\rangle ^2 + \langle X_{\omega\tau i}\rangle ^2 + {\langle w^2\rangle -\langle w\rangle ^2M^2_\phi(1)\over TF}$$
where $\psi = \pi\omega{T-1\over R} + \pi\tau{F-1\over D}$,
$H={\langle w^2\rangle M_\Phi(2)-\langle w\rangle ^2M^2_\phi(1)\over2TF}$ and
$Q_{\omega\tau} =
{sinc(2\pi\omega T/R)sinc(2\pi\tau F/D)\over sinc(2\pi\omega/R)sinc(2\pi\tau/D)}$.
If the Fourier-transforms are carried out with no over-sampling, $T=R$
and $F=D$ and $Q_{\omega\tau} = \delta_{\omega0}\delta_{\tau0}$. The
moments then simplify considerably into the results of section 2.
In particular, note that the first moments become
$$\langle X_{\omega\tau r}\rangle = M_\Phi(1)\langle w\rangle
\delta_{\tau0}\delta_{\omega0} \qquad \langle X_{\omega\tau i}\rangle =
0 \qquad \langle X_{\omega\tau ri}\rangle = 0$$
Implicit here is the assumption that $\omega$ and $\tau$ are integers.
The assumption that signal is present in only one cell in delay-rate
space is equivalent to the assumption that the FFT is not oversampled,
and that true delay-rate residual lies at the center of one of the
searched FFT cells. The former is not true in practice while the
latter is not true in general. The FFT is zero-padded, in FRING and
KRING, by a factor of at least 4 and up to 16.
The statistics of $X_{\omega\tau}$ are those of two correlated Gaussian
random variables:
$$P_X(x_r,x_i) = {1\over2\pi\sigma_{\omega\tau r}\sigma_{\omega\tau i}(1-\rho^2)}
e^{-{(x_r - \langle X_{\omega\tau r}\rangle)^2\over2\sigma_{\omega\tau r}^2(1-\rho^2)}
+{(x_i - \langle X_{\omega\tau i}\rangle)(x_r - \langle X_{\omega\tau r}\rangle)\rho\over\sigma_{\omega\tau r}\sigma_{\omega\tau i}(1-\rho^2)}
-{(x_i - \langle X_{\omega\tau i}\rangle)^2\over2\sigma_{\omega\tau i}^2(1-\rho^2)}}$$
where
$\rho = {\langle X_{\omega\tau ri}\rangle - \langle X_{\omega\tau r}\rangle\langle X_{\omega\tau i}\rangle\over\sigma_{\omega\tau r}\sigma_{\omega\tau i}}$,
$\sigma_{\omega\tau r}^2 = \sigma_a^2 + HQ_{2\omega2\tau}cos2\psi$,
$\sigma_{\omega\tau i}^2 = \sigma_a^2 - HQ_{2\omega2\tau}cos2\psi$
and
$\sigma_a^2={\langle w^2\rangle -\langle w\rangle ^2M^2_\phi(1)\over2TF}$.
What is the probability that the squared amplitude of an off-center
delay-rate cell, $x_{\omega\tau a}^2$, is less than the squared
amplitude of the delay-rate origin, $x_{00a}^2$?
$$P(x_{\omega\tau a}^2 < x_{00a}^2)
= \int\limits_{x_a^2 < x_{00a}^2}dx_rdx_i
P_X(x_r,x_i) =
{1\over2\pi\sigma_{x\omega\tau}\sigma_{y\omega\tau}}
\int\limits_{x_r^2+x_i^2 < x_{00a}^2}dx_rdx_i
e^{-{(x_r-x_0)^2\over2\sigma_{x\omega\tau}^2}
-{x_i^2\over2\sigma_{y\omega\tau}^2}}$$
where $x_0 = M_\Phi(1)\langle w\rangle Q_{\omega\tau}$,
$\sigma_{x\omega\tau}^2 = \sigma_a^2 + HQ_{2\omega2\tau}$,
$\sigma_{y\omega\tau}^2 = \sigma_a^2 - HQ_{2\omega2\tau}$.
What is the probability that $x_{\omega\tau a}^2$ will jointly not exceed
$x_{00a}^2$ for all $\omega\tau$ in the search space?
$$ 1-PFD = \prod\limits_{\omega\tau}
{1\over2\pi\sigma_{x\omega\tau}\sigma_{y\omega\tau}}
\int\limits_{x_r^2+x_i^2 < x_{00a}^2}dx_rdx_i
e^{-{(x_r-x_0)^2\over2\sigma_{x\omega\tau}^2}
-{x_i^2\over2\sigma_{y\omega\tau}^2}}$$
This probability is too difficult to evaluate and is currently a
matter of active research.
\section{Conclusions}
I have presented a formal analysis of the method by which the SNR can
be estimated for the results of a fringe-search. This analysis shows
that the SNR estimate in FRING is an underestimate at low SNR and an
overestimate at high SNR. The former is contrary to the conventional
wisdom. I have also shown that, if the input visibilities are
normalized to unit amplitude before the fringe-search is carried out,
the SNR can be estimated directly from the phases of the resultant
visibilities which contain only information about the ratio of the
signal to the noise, not either one separately. The procedure
discussed in section 2 of this memo has been implemented in KRING.
Use of the normalized visibilities comes at a slight cost in sensitivity.
However, there is a cost, in that the probability of false detection is
slightly greater when using normalized visibilities .
Traditionally, the 'probability of false detection' is computed
assuming that signal is present in only one cell in delay-rate space.
When zero-padded Fourier transforms are used, this assumption is not
valid. More generally, whenever the true delay/rate residual is not
an integral number of search cells away from the starting delay/rate
point, there will be signal present in many cells in the delay/rate
plane, even if the Fourier transforms are not padded. The traditional
analysis of the probability of false detection is generalized in
section 3.3; there, signal is allowed in more than one cell in
delay-rate space. The analysis in section 3.3 can be used to formally
derive the probability of false detection when zero-padding is used
before performing the FFTs.
\section{Future Directions}
When padded Fourier transforms are used, the delay-rate peak is more
likely to lie at the center of a delay-rate cell, decreasing the
probability of false detection. On the other hand, padding the
Fourier transforms increases the probability of false detection
because a noise spike close to the true peak may be found that has
greater amplitude. A more complete analysis of the effects of
zero-padding upon the probability of false detection is needed. Also,
alternatives to zero-padding such as tapering and their effects upon
the probability of false detection should be analyzed.
The PFDs of sections 3 remain to be calculated using numerical
methods.
\section{Acknowledgements}
I would like to acknowledge Walter Brisken, Bryan Butler, and Michael
Rupen, Vivek Dhawan, Fred Schwab, and Leonid Kogan for their helpful
comments in preparing this memo.
\end{document}