1: \documentclass[12pt,preprint]{aastex}
2: %\documentclass{emulateapj}
3:
4:
5: \begin{document}
6:
7: \title{Detecting the stochastic gravitational wave background using pulsar timing}
8:
9: \author{Fredrick A. Jenet\altaffilmark{1}, George B. Hobbs\altaffilmark{2}, K.J. Lee\altaffilmark{3}, Richard N. Manchester\altaffilmark{2}}
10: %\altaffiltext{1}{California Institute of Technology, Jet Propulsion Laboratory, 4800 Oak Grove Drive, Pasadena, CA 91109}
11: \altaffiltext{1}{Center for Gravitational Wave Astronomy, University of Texas at Brownsville, TX 78520 (merlyn@alum.mit.edu)}
12: \altaffiltext{2}{Australia Telescope National Facility, CSIRO, P.O.~Box~76, Epping, NSW~1710, Australia}
13: \altaffiltext{3}{Department of Astronomy, Peking University, Beijing 100871, China}
14:
15: \begin{abstract}
16:
17: The direct detection of gravitational waves is a major goal of current
18: astrophysics. We provide details of a new method for detecting a
19: stochastic background of gravitational waves using pulsar timing
20: data. Our results show that regular timing observations of 40 pulsars
21: each with a timing accuracy of 100\,ns will be able to make a direct
22: detection of the predicted stochastic background from coalescing black
23: holes within five years. With an improved pre-whitening algorithm, or
24: if the background is at the upper end of the predicted range, a
25: significant detection should be possible with only 20 pulsars.
26: \end{abstract}
27:
28: \keywords{pulsars:general --- gravitational waves}
29:
30: \section{Introduction}
31:
32: Analysis of pulsar pulse time-of-arrival (TOA) data shows that
33: pulsars, especially millisecond pulsars (MSPs), are very stable
34: clocks. Measurement of timing residuals, that is, the differences
35: between observed and predicted TOAs, enables the direct detection of
36: gravitational waves (GWs) \citep{ew75,saz78,det79}. The fluctuating
37: TOAs induced by a GW will be correlated between widely-spaced
38: pulsars. \citet{hd83} attempted to detect this correlation by
39: cross-correlating the time derivative of the timing residuals for
40: multiple pulsars. In our work, we have developed a similar
41: cross-correlation technique and have, for the first time, a fully
42: analyzed method for combining multiple pulsar observations in order to
43: make an unambiguous detection of a GW background. We emphasize that,
44: in contrast to \citet{hd83}, our method is based entirely on the
45: measured residuals.
46:
47: Only the effects of a stochastic background of GWs are considered.
48: Astrophysical sources of such a background include cosmological
49: processes \citep[e.g.][]{mag00} and coalescing massive black hole
50: binary systems \citep{jb03,wl03,ein+04}. We show that a direct
51: detection of a stochastic GW background is possible using pulsar
52: timing observations and that the significance of the detection depends
53: upon the number of pulsars observed, the root-mean-square (RMS) timing
54: noise achieved, the number of observations, and the power spectrum of
55: the measured timing residuals. The results are applied to the case of
56: the Parkes pulsar timing array
57: (PPTA\footnote{http://www.atnf.csiro.au/research/pulsar/psrtime}).
58:
59:
60: In the next section, the analysis technique is described. In \S 3
61: the significance of detecting a given stochastic background using this
62: method is estimated. The effects of pre-filtering the residual time
63: series are also discussed. The results are summarized in
64: \S 4.
65:
66: \section{Detection Technique}
67: As a first step, the power spectra of the pulsar timing residuals are
68: analyzed. If they all show a very red power-law spectrum, the
69: residuals may be dominated by a GW background. However, such red
70: spectra can also be due to period noise intrinsic to the pulsar,
71: uncorrected interstellar delays, inaccuracies in the Solar-System
72: ephemeris, or variations in terrestrial time standards
73: \citep[e.g.][]{fb90}. A GW background produces a unique signature in
74: the timing residuals which can only be confirmed by observing
75: correlated signals between multiple pulsars widely distributed on the
76: sky.
77:
78: The presence of a stochastic GW background will cause the pulse TOAs
79: to fluctuate randomly, but these fluctuations will be correlated between
80: different pulsars. In order to detect the presence of a GW background,
81: one needs to first calculate the correlation coefficient
82: between the observed timing residuals of each pair of observed
83: pulsars:
84: \begin{equation}
85: r(\theta) = \frac{1}{N}\sum_{i=0}^{N-1} R(t_i,\hat{k}_1) R(t_i,\hat{k}_2) \label{r_stat}
86: \end{equation}
87: where $R(t_i,\hat{k})$ is the time series of $N$ pulsar residuals
88: sampled regularly in time, $\hat{k}_1$ and $\hat{k}_2$ are the
89: directions to the two pulsars, and $\cos(\theta) = \hat{k}_1 \cdot
90: \hat{k}_2$. It will be assumed that $R$ has zero mean and that each
91: pulsar pair has a unique angular separation. $r(\theta)$ is written
92: only as a function of the angular separation since the GW background
93: is expected to be isotropic. In the presence of an isotropic GW
94: background, the ensemble-averaged value of $r(\theta)$ is given
95: by\footnote{For an outline of the calculation of $\zeta$ see
96: \citet{hd83}.}:
97: \begin{eqnarray}
98: \langle r(\theta)\rangle &=& \sigma_g^2 \zeta(\theta) \\
99: \zeta(\theta) &=& \frac{3}{2} x\log(x) -\frac{x}{4} + \frac{1}{2} +
100: \frac{1}{2}\delta(x)
101: %x &=& \frac{1 - \cos(\theta)}{2}
102: \end{eqnarray}
103: where $x = (1 - \cos(\theta))/2$, $\sigma_g$ is the RMS of the timing residuals
104: induced by the stochastic GW background, and $\delta(x)$ equals 1 for
105: $x=0$ and 0 otherwise. The detection technique proposed here simply
106: looks for the presence of the function $\zeta(\theta)$ in the measured
107: correlation coefficients $r(\theta)$.
108:
109: Since one cannot perform the ensemble average in practice, the measured
110: statistic, $r(\theta)$, will be of the form $r(\theta) = \langle r(\theta)\rangle + \Delta r(\theta)$,
111: %\begin{equation}
112: %r(\theta) = \langle r(\theta)\rangle + \Delta r(\theta) \label{measured_r}
113: %\end{equation}
114: where $\Delta r(\theta)$ is a ``noise term''. Since $r(\theta)$ is
115: calculated by summing over a large ($\geq 20$) number of data points,
116: $\Delta r(\theta)$ will be a Gaussian random variable for practical
117: purposes. The optimal way to detect the presence of a known functional
118: form within random data is to calculate the correlation between the
119: data and the known function. Hence, to detect the presence of
120: the GW background one needs to calculate
121: \begin{equation}
122: \rho = \frac{\frac{1}{N_p}\sum_{i=0}^{N_p-1} (r(\theta_i)-\bar{r})(\zeta(\theta_i)-\bar{\zeta})}{\sigma_r \sigma_\zeta} \label{rho_def}
123: \end{equation}
124: where $\theta_i$ is the angle between the $i$th pair of pulsars and
125: $N_p$ is the number of distinct pairs of pulsars. $\bar{r}$ and
126: $\bar{\zeta}$ indicate the mean values over all pairs of pulsars and
127: $\sigma_r^2$ and $\sigma_\zeta^2$ are the variances of $r$ and $\zeta$
128: respectively. For $M$ pulsars, $N_p = M(M-1)/2$.
129:
130: From the definition of $r(\theta)$ and Eqn. \ref{rho_def}, one can show that the
131: expected value of $\rho$ is approximately:
132: \begin{eqnarray}
133: \rho &\approx& \frac{\sigma_g^2 \sigma_{\zeta}}{\sqrt{\sigma_g^4 \sigma_\zeta^2 + \sigma_{\Delta r}^2}}\\
134: \sigma_{\Delta r}^2 &=& \frac{1}{N_p} \sum_{i=0}^{Np-1} \langle (r(\theta_i) - \langle r(\theta_i)\rangle )^2\rangle. \label{sDr}
135: \end{eqnarray}
136: For the case where there is no correlation in the data, the statistics
137: of $\rho$ will be Gaussian with zero mean and variance given by $\sigma_\rho^2 = 1/N_p = 2 / (M^2-M)$.
138: %\begin{eqnarray}
139: %\sigma_\rho^2 &=& \frac{1}{N_p} = \frac{2}{M(M-1)} \label{srho}
140: %\end{eqnarray}
141: Hence, the significance of a measured value of $\rho$ may be defined as $S = \rho/\sigma_\rho$.
142: %\begin{eqnarray}
143: %S &=& \frac{\rho}{\sigma_\rho} \label{sig1}
144: % &=& \frac{\sqrt{\frac{M(M-1)}{2}}}{\sqrt{1 + \frac{\sigma_r^2}{\sigma_g^4 \sigma_\zeta^2}}} \label{sig1}
145: %\end{eqnarray}
146: The probability of measuring a correlation greater than or equal to
147: $\rho$ when no actual correlation is present is given by
148: $\mbox{erf}(S/\sqrt{2})/2$.
149:
150: \section{Estimating the Detection Significance}
151:
152: In order to estimate the expected detection significance, $S$, one
153: needs to estimate $\sigma_g$ and $\sigma_{\Delta r}$. It is
154: assumed that the timing residuals, $R(t,\hat{k})$, are stationary
155: Gaussian random variables that are sampled at regular intervals
156: denoted by $\Delta t$. It is also assumed that terms proportional to $t$ and
157: $t^2$ (i.e., the period and period-derivative terms) have been
158: subtracted from $R(t,\hat{k})$.
159:
160: The space-time fluctuations induced by a stochastic GW background are
161: described by a quantity known as the characteristic strain spectrum
162: denoted by $h_c$ \citep[e.g.][]{mag00}. Models of the GW background propose a
163: power-law dependence between $h_c$ and the GW frequency, $f$: $h_c(f) = A f^{\alpha}$ \citep{jb03,wl03,mag00,ein+04}.
164: %\begin{equation}
165: %h_c(f) = A f^{\alpha}
166: %\end{equation}
167: Using this form of the characteristic strain spectrum, the power
168: spectrum of the induced residuals is given by $P_R(f) = \langle |\tilde{R}(f)|^2\rangle = \frac{A^2}{4 \pi^2} f^{2\alpha -3}$,
169: %\begin{equation}
170: %P_R(f) = \langle |\tilde{R}(f)|^2\rangle = \frac{A^2}{4 \pi^2} f^{2\alpha -3}, \label{P_R}
171: %\end{equation}
172: where $\tilde{R}(f)$ is the Fourier transform of $R(t)$.
173: Given $P_R(f)$, the total RMS fluctuation induced by the
174: stochastic GW background is given by
175: \begin{eqnarray}
176: \sigma_g^2 &=& \int_{f_l}^{f_h} P_R(f) df \\
177: &=& \frac{A^2}{2 \pi^2 (2-2\alpha)} \left(f_l^{2\alpha -2} - f_h^{2\alpha -2}\right) \label{sigmag2}
178: \end{eqnarray}
179: where $f_l$ is the lowest detectable frequency given by $1/T$ and
180: $f_h$ is the highest detectable frequency typically given by
181: $1/2\Delta t$. $T$ is the total time span of the data set. Since $\alpha <0$ for backgrounds of interest \citep{mag00}, the term containing $f_h$ is negligible.
182:
183: Estimating $\sigma_{\Delta r}$ is slightly more complicated. To take
184: into account the effects of subtracting linear and quadratic terms
185: from the residuals, a semi-analytic approach was adopted. As outlined
186: below, an estimate for $\sigma_{\Delta r}$ is made analytically but
187: with one free parameter $\beta$. For a given value of $\beta$, $S$ is
188: calculated as a function of $A$ for a given set of pulsars and timing
189: parameters. $S(A)$ is compared to Monte-Carlo simulations in order to
190: determine the correct value of $\beta$. This showed that the value of
191: $\beta$ is insensitive to the values $\alpha$, N, M, $\sigma_g$ and
192: the RMS residual noise level.
193:
194: Using Equation~\ref{r_stat} together with the assumption that
195: $R(t,\hat{k})$ is a Gaussian random variable, one can show that
196: \begin{eqnarray}
197: \sigma_{\Delta r}^2 &\approx& \overline{\frac{1}{N^2}\sum_{i=0}^{N-1}\sum_{j=0}^{N-1} c_{ij}(\hat{k}_1) c_{ij}(\hat{k}_2)}, \label{eqnsd}
198: %c_{ij}(\hat{k}) &=& \langle R(t + i \Delta t,\hat{k})R(t+ j \Delta t,\hat{k})\rangle.
199: \end{eqnarray}
200: where $c_{ij}(\hat{k}) =\langle R(t + i \Delta t,\hat{k})R(t+ j \Delta t,\hat{k})\rangle$. The bar above Equation~\ref{eqnsd} represents an average over all
201: pairs of pulsars. As the autocorrelation function and the power
202: spectrum are Fourier transforms of one another, one can estimate
203: $\sigma_{\Delta r}^2$ from the expected power spectrum of the
204: residuals. The statistics of the residuals are assumed to be
205: stationary so that $c_{ij}(\hat{k})$ depends only on $i - j$. The
206: expected discrete power spectrum of $R(t,\hat{k})$, which includes both a GW component and a white noise component, is given by
207: \begin{equation}
208: P_d(i,\hat{k}) = \left\{ \begin{array}{lc}
209: P_g(i) + \frac{2 \sigma_n(\hat{k})^2}{N} \hspace{20pt} & \mbox{for $i > 0$} \\
210: 0 & \mbox{for $i = 0$}
211: \end{array}
212: \right.. \label{p_d}
213: \end{equation}
214: $P_g(i)$ is the discrete power spectrum of the GW-induced timing
215: residuals, $i$ is the discrete frequency bin number corresponding to
216: frequency $i/T$, $\sigma_n(\hat{k})$ is the RMS value of the
217: residual fluctuations caused by all non-GW sources for the pulsar in
218: the $\hat{k}$ direction. It is assumed that all noise sources have a
219: flat spectrum. This assumption is consistent with most observations of
220: MSPs. $P_g(i)$ is given by
221: \begin{equation}
222: P_g(i) = \frac{A^2 T^{2-2\alpha}}{(2 \pi)^2 (2-2\alpha)}m(i)
223: \end{equation}\label{pg}
224: where
225: $$\begin{array}{ll}
226: m=0 & {\rm for~} i = 0\\
227: m= \beta^{2\alpha -2} - (1.5)^{2\alpha-2} & {\rm for~} i = 1 \\
228: m=(i-0.5)^{2\alpha-2} - (i+0.5)^{2\alpha-2} & {\rm for~} i > 1.
229: \end{array}$$
230: %
231: Effectively, $\beta$ is the lowest frequency used to calculate the
232: correlation function $c_{ij}$. Monte-Carlo simulations show that
233: $\beta = 0.97$.
234:
235: For the case where all pulsars have the same noise level, the detection significance becomes
236: \begin{eqnarray}
237: S &=& \sqrt{\frac{M(M-1)/2}{1 + \frac{\chi + 2(\sigma_n/\sigma_g)^2 + (\sigma_n/\sigma_g)^4}{N \sigma_\zeta^2}}}\label{sig_nolambda}
238: %\chi &=& \frac{1}{\sigma_g^4 N}\sum_{i=0}^{N-1} \sum_{j=0}^{N-1}c_{g_{ij}}^2.
239: \end{eqnarray}
240: where $\chi = \frac{1}{\sigma_g^4 N}\sum_{i=0}^{N-1} \sum_{j=0}^{N-1}c_{g_{ij}}^2$, and $c_{g_{ij}}$ is the correlation function for the GW-induced
241: component of the timing residuals. $\chi$ is a measure of the
242: ``whiteness'' of the residuals.
243:
244: The solid curve in Figure~\ref{fig1} panel A) plots the detection
245: significance versus power-law amplitude for $\alpha = -2/3$, the
246: expected value for a background generated by an ensemble of
247: super-massive black hole binaries \citep{jb03}. This spectral index
248: together with the removal of the linear and quadratic terms from $R$
249: effectively makes $\chi = 0.6 N$. The parameters are set as follows:
250: $N=250$, $M=20$, $\sigma_n=100$ ns and $T = 5$ years. These values are
251: the target values for the PPTA \citep{hob04}. Note that the
252: significance saturates for high values of $A$. This effect can easily
253: be seen in Equation~\ref{sig_nolambda} since all terms of the form
254: $\sigma_n/\sigma_g$ go to zero as $\sigma_g$ gets very large. This
255: saturation is due to the ``self-noise'' associated with the stochastic
256: nature of the background and its asymptotic value is independent of
257: $\sigma_n$. The roll-off at low values of $A$ occurs at $\sigma_g =
258: \sigma_n$.
259:
260: %\section{Effects of pre-filtering}
261:
262: Since the power spectrum of the GW-induced timing residuals will be
263: dominated by low frequencies, one can apply a low-pass filter to each
264: of the residual time series before correlating. This is similar to
265: fitting a low-order polynomial to the data and then correlating the
266: resulting fits. To estimate the significance for this technique, one
267: evaluates $\sigma_g^2$ and $\sigma_{\Delta r}^2$ using
268: Equations~\ref{sigmag2} and \ref{p_d} but with a high frequency
269: cut-off $f_{hc}$. For purposes of this discussion, $f_{hc}$ was set to
270: $4/T$. The dashed line in Figure~\ref{fig1} panel A) shows the effect
271: of using a low-pass filter on the residuals. All the other parameters
272: are the same as those used to generate the solid line. Low-pass
273: filtering effectively reduces $\sigma_n$ while keeping $\sigma_g$
274: relatively unchanged. It also has the effect of increasing $\chi/N$
275: when $P_g$ is a red power-law spectrum. Hence, low-pass filtering will
276: not increase the maximum attainable significance, but it will lower
277: the value of $\sigma_g$ where the roll-off starts to occur.
278:
279: We next try to increase the maximum achievable significance. This
280: method involves both low-pass filtering and a technique called
281: ``whitening''. When correlating two time series that each have a steep
282: power-law spectrum, an optimal signal-noise ratio is obtained if filters
283: are applied to give each time series a flat spectrum before
284: correlation. This will act to reduce $\chi$ in
285: Equation~\ref{sig_nolambda}. In practice, starting from the lowest
286: non-zero frequency bin, we give each Fourier component with
287: significant power equal amplitude and set higher components to zero.
288: In this way, we are correlating only that part of the signal which has
289: a high signal-to-noise ratio and adjusting the power spectrum to
290: optimize the measurement of the correlation function.
291:
292: $P_d$ and $\sigma_g$ need to be calculated in order to estimate
293: $S$ using the whitening method. After whitening $P_d(i,\hat{k}) = 2
294: \sigma_d(\hat{k})^2 / N$, where $\sigma_d(\hat{k})$ is the
295: RMS of the residual data from the $\hat{k}$th
296: pulsar. The whitening also affects $\sigma_g$. In
297: the general case where the pulsars have different noise levels,
298: $\sigma_g$ will depend on the pulsar. The expression for $\rho$ then
299: becomes:
300: \begin{equation}
301: \rho \approx \frac{\left(\overline{\sigma_g^2\zeta^2} - \overline{\sigma_g^2\zeta} \mbox{ } \overline{\zeta} \right)/\sigma_\zeta}{\sqrt{\left(\overline{ \sigma_g^4 \zeta^2} - \left(\overline{\sigma_g^2\zeta}\right)^2\right) + \sigma_{\Delta r}^2}} \label{rho}\\
302: \end{equation}
303: with $\sigma_g(\theta)^2$ given by
304: \begin{eqnarray}
305: \sigma_{g}(\theta)^2 &=& \frac{2}{N} \sigma_d(\hat{k}_1)\sigma_d(\hat{k}_2) \sqrt{\left(\sum_{i=0}^{N_{max}}P_g(i)/P_d(i,\hat{k}_1)\right)\left(\sum_{i=0}^{N_{max}}P_g(i)/P_d(i,\hat{k}_1)\right)}
306: \end{eqnarray}
307: where $N_{max}$ is the largest
308: frequency bin used based on the criterion discussed above. The
309: solid line in Figure~\ref{fig1} panel B) plots the significance using the
310: whitening versus $A$. The same parameters were used for this case as
311: in the previous cases.
312:
313: The above discussion assumes that the noise levels were the same for
314: all pulsars. Next, the case where the pulsars have different noise
315: levels will be considered. All curves in Figure~\ref{fig1} panel B)
316: were generated using the whitening technique. Unless specified, 250
317: observations were taken on each pulsar over 5 years. The dashed line
318: corresponds to 20 pulsars, 10 with $\sigma_n=100$~ns and 10 with
319: $\sigma_n=500$~ns. The dashed-dot line has 10 pulsars each with
320: $\sigma_n=100$~ns and 500 observations. The dashed-triple-dot line has
321: 20 pulsars with $\sigma_n=100$~ns and 500 observations over ten years.
322:
323: When given a choice between observing a large sample of pulsars with
324: different noise levels and observing only those pulsars with the
325: lowest noise levels but for a longer time, the above curves
326: demonstrate that one should actually observe the larger sample of
327: pulsars. This is not a general statement, but rather it depends on the
328: level of the GW background and the noise level. However, the levels
329: chosen above are relevant to the PPTA
330: \citep{jb03,wl03,hob04}. Note that for large $M$, the
331: significance scales as $M$. Hence, doubling the number of pulsars will
332: double the expected significance.
333:
334:
335: \section{Summary}
336:
337: The main goal of this work is to determine the effectiveness of an
338: array of pulsars, such as the PPTA, for detecting a stochastic
339: background of GWs. Using a simple correlation technique, the detection
340: significance was calculated given the number of pulsars, the location
341: of each pulsar, the TOA precision, the number of observations, the
342: total time span of the data, and the amplitude and power-law index of
343: the GW background. For the case where all pulsars have the same
344: white-noise spectrum, Equation~\ref{sig_nolambda}, may be used to
345: calculate the detection significance. For the case of the PPTA, it was
346: found that the maximum achievable significance will be about 3 for a
347: background with spectral index $\alpha = -2/3$ and $A \sim 10^{-15}$
348: which is the expected level of the GW background from an
349: ensemble of super-massive binary black holes in galaxies
350: \citep{jb03,wl03,ein+04}. Note that lowering the RMS noise level will only
351: decrease the minimum detectable value of $A$ and not increase the
352: maximum attainable significance.
353:
354: Low-pass filtering the timing residuals, or equivalently, fitting
355: low-order polynomials (i.e. cubic terms) to the residuals and
356: correlating the coefficients, does not increase the maximum attainable
357: significance. The significance level is increased by pre-whitening of
358: the timing residuals. Using whitening, it is estimated that the PPTA
359: could obtain a detection significance greater then 4 for $A \geq 3
360: \times 10^{-15} \mbox{yr}^{-2/3}$ provided that efficient whitening
361: filters can be designed and implemented. This is an area of further
362: study and will be addressed in a later paper. With the same
363: qualifiers, increasing the total time span of the PPTA to 10 years
364: would yield a significance greater then 4 for $A \geq 10^{-15}
365: \mbox{yr}^{-2/3}$. Since the significance scales as the number of
366: pulsars, doubling that number will double the expected
367: significance. Hence, using the simple correlation technique described
368: here without any pre-filtering, a stochastic background with $A \geq
369: 10^{-15} \mbox{yr}^{-2/3}$ will be detectable at a significance of about 5.5
370: using 40 pulsars observed 250 times over 5 years and each having
371: 100~ns timing precision.
372:
373: Part of this research was carried out at the Jet Propulsion
374: Laboratory, California Institute of Technology, under a contract with
375: the National Aeronautics and Space Administration and funded through
376: the internal Research and Technology Development program. The authors
377: wish to thank Russell Edwards for useful discussions.
378:
379: \begin{figure}
380:
381: \plotone{f1.ps}
382: \caption{\label{fig1} The detection significance, S, versus the
383: logarithm of the amplitude $A$ of the characteristic strain
384: amplitude $h_c(f)$. The strain spectral index $\alpha = -2/3$,
385: corresponding to an astrophysical background of GWs generated by
386: super-massive binary black holes. The vertical lines bound the
387: values of $A$ expected by models of the GW background
388: \citep{jb03,wl03,ein+04}. In panel A), the curves were calculated
389: with 20 pulsars each with RMS residual noise fluctuations of 100
390: ns. The solid line corresponds to the simple correlation
391: technique. The dashed line includes the effect of low-pass
392: filtering. Panel B) shows the effects of the whitening
393: technique. The solid line was calculated with the same parameters as
394: in A). The remaining curves were generated using different noise
395: levels and number of pulsars. See text for further details.}
396:
397: \end{figure}
398:
399:
400: %\bibliographystyle{apj}
401: %\bibliography{journals,modrefs,psrrefs,crossrefs}
402:
403: \begin{thebibliography}{9}
404: \expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
405:
406: \bibitem[{Detweiler(1979)}]{det79}
407: Detweiler, S. 1979, ApJ, 234, 1100
408:
409: \bibitem[{{Estabrook} \& {Wahlquist}(1975)}]{ew75}
410: {Estabrook}, F.~B. \& {Wahlquist}, H.~D. 1975, General Relativity and
411: Gravitation, 6, 439
412:
413: \bibitem[Enoki et al.(2004)]{ein+04} Enoki, M., Inoue, K.~T.,
414: Nagashima, M., \& Sugiyama, N.\ 2004, \apj, 615, 19
415:
416: \bibitem[{Foster \& Backer(1990)}]{fb90}
417: Foster, R.~S. \& Backer, D.~C. 1990, ApJ, 361, 300
418:
419: \bibitem[{Hellings \& Downs(1983)}]{hd83}
420: Hellings, R.~W. \& Downs, G.~S. 1983, ApJ, 265, L39
421:
422: \bibitem[{Hobbs (2004)}]{hob04}
423: Hobbs, G., PASA, in press (astro-ph/0412153)
424:
425: \bibitem[{Jaffe \& Backer(2003)}]{jb03}
426: Jaffe, A.~H. \& Backer, D.~C. 2003, ApJ, 583, 616
427:
428: \bibitem[{Maggiore(2000)}]{mag00}
429: Maggiore, M. 2000, Phys. Rep., 331, 283
430:
431: \bibitem[{Sazhin(1978)}]{saz78}
432: Sazhin, M.~V. 1978, Sov. Astron., 22, 36
433:
434: \bibitem[{{Wyithe} \& {Loeb}(2003)}]{wl03}
435: {Wyithe}, J.~S.~B. \& {Loeb}, A. 2003, ApJ, 590, 691
436:
437: \end{thebibliography}
438:
439:
440: %\clearpage
441:
442: %\clearpage
443: %\begin{figure}
444: %\plotone{f2.ps}
445: %\caption{\label{fig2} The detection significance, S, versus the
446: % logarithm of the amplitude $A$ of the characteristic strain
447: % amplitude $h_c(f)$. $\alpha = -2/3$. The curves were generated using
448: % different noise levels and number of pulsars. See text for further
449: % details. }
450: %\end{figure}
451:
452:
453:
454:
455:
456:
457: \end{document}
458:
459: