I Introduction
We generally assume that the FM noises affecting an oscillator are five and can be modeled by a PSD following 5 integer powerlaws, from to . It is sometimes necessary to consider fractional or even real exponents, whether due to the process itself or to the needs of the estimation. In our case, the millisecond pulsar timing analysis, we are faced to both cases: on the one hand, we are searching for a red noise which is expected to follow a FM slope [1, 2]
; and on the other hand, we aim to estimate the exponent of this red noise by a Bayesian statistics analysis of the PVAR estimates. We have then to consider this exponent as a real number around which we have to place a confidence interval. For these reasons, we need an expression of the PVAR response for continuous powerlaw noises as well as its uncertainty, i.e. the number of degrees of freedom of a PVAR estimate versus the powerlaw exponent.
Ii Continuous PVAR response
The response of a variance XVAR (X may be A for AVAR, M for MVAR, H for HVAR, P for PVAR, etc.) for a noise is given by the following relationship
(1) 
where is the transfer function associated to XVAR and the noise level.
Following the pioneer work of Walter [3], nothing forces to be an integer and we can calculate the response of a variance for continuous powerlaw noise. Thus, using Eq. (1) leads to the continuous response of AVAR:
(2) 
which is equal to [3, Eq. (14)] since we don’t consider a high cutoff frequency in (1).
Similarly, by using the transfer function of PVAR [4, Eq. (17)], we found
(3)  
Since PVAR converges from to FM noise, we assume that this expression is valid for .
Iii Degrees of Freedom of PVAR estimates
First, we have to find a simplified expression of the number of degrees of freedom (dof) of PVAR estimates for integer powerlaw noises. Since this equation was solved for a white PM noise (see Eq. (24) in [4]), we assume that the following expression should have the same form:
(4) 
with , i.e. the integration time divided by the sampling time , is the number of summations in the computation of the variance estimates, i.e. ( is the total number of samples in the datarun) for PVAR and are coefficients to be determined for each exponent. Thanks to [4, Eq. (24)], we already know that and .
We determined these coefficients by using massive MonteCarlo simulations and verified the results by comparing them to the dof computed for continuous powerlaw noises.
Iiia Determination of the coefficients from MonteCarlo simulations
The MonteCarlo simulation was performed by computing 10 000 sequences of frequency deviations for each and for each datarun length , i.e. 150 000 simulated sequences. For given , and , we deduced the dof from the averages and the variances of the PVAR for the corresponding set of sequences by using the following wellknown property of distributions [4]:
(5) 
where and stands respectively for the mathematical expectation and the variance of the quantity between the brackets. By using a least square fit, we obtained the following results:

, , , ,

for all .
We have then modeled by the following order polynomial and assumed that is constant:
(6) 
Thanks to Eq. (4) and (6), we are now able to assess the dof of all PVAR estimates whatever the integration time or the number of samples.
IiiB Verification for continuous powerlaw noises
The dof may be computed from Eq. (5). The mathematical expectation is the response of PVAR given above in (3) and the variance can be computed from (21) and (22) of [4]:
(7)  
where is the autocorrelation function of the phasetime samples , i.e. . We used the following continuous expression of versus the powerlaw exponent (see [5, 3]):
However, since this expression involves functions with arguments of the order of , we have limited these computations to samples (!). Thanks to this equation, we have computed the theoretical variance of PVAR versus continuous and deduced the dof from (5).
Let us call . From Eq. (4), we see that . The top plot of Fig. 3 shows computed from (7) (crosses) and approximated from (6) (solid lines) versus the noise powerlaw for (we prefer to plot than for reasons of readability of the graph). The agreement is quite good for and , but there is a notable difference for and . The lower plot of Fig. 3 shows that this discrepancy is around 20% at worst but generally remains within % in most cases (all for and all for ). This agreement is satisfactory to get an acceptable assessment of the PVAR uncertainties since the relative uncertainties are proportional to : they are therefore always below 10% and mostly within .
IiiC Case of the largest integration times
The approximation given by Eq. (4) and (6) remains enough close to the empirical dof if . Moreover, we know that for . On the other hand, we can note that the approximation diverges beyond (see dashed lines in the upper plot of Fig. 4). However, it is of importance to assess the uncertainties within this interval, particularly if is not a power of 2.
In order to fill this gap, we decided to interpolate the dof within
, i.e. between and , by using the following semilogarithmic fitFor , the dof are set to 1.
Fig. 4 (top) is an enlargement of the highest 2 decades of , i.e. fo data, to focus on the result of the semilogarithmic fit. The bottom plot shows the error between the fit and the dof computed from the MonteCarlo simulations. We can see than most of these errors are within except for the case of the white FM which ranges from to and even reaches 24% for . However, this fit is sufficient to ensure an estimation of the PVAR uncertainty for the highest value within at worst.
Iv Conclusion
From a theoretical calculation, we have determined the response of PVAR for continuous powerlaw noises. From MonteCarlo simulations, we have got a simplified expression providing the dof of the the PVAR estimates within 10 %. We have proved that this expression remains valid for noninteger powerlaw noises. Finally, we have shown that a simple interpolation is efficient to fit the dof for the highest octave of integration times. Thanks to these results, we will be able to use PVAR to analyze millisecond pulsar timings and to estimate the noninteger exponent of a red noise if it is detected. On the other hand, this paper make it possible to generalize the use of PVAR to process any signal with noninteger powerlaw noise.
Acknowledgement
This work was partially funded by the ANR Programmes d’Investissement d’Avenir (PIA) Oscillator IMP (Project 11EQPX0033) and FIRSTTF (Project 10LABX0048).
References
 [1] E. S. Phinney, “A practical theorem on gravitational wave backgrounds,” August 2001, arXiv:astroph/0108028.
 [2] S. Chen, A. Sesana, and W. Del Pozzo, “Efficient computation of the gravitational wave spectrum emitted by eccentric massive black hole binaries in stellar environments,” Monthly Notices of the Royal Astronomical Society, vol. 470, no. 2, pp. 1738–1749, 05 2017.
 [3] T. Walter, “Characterizing frequency stability: A continuous powerlaw model with discrete sampling,” IEEE Transactions on Instrumentation and Measurement, vol. 43, no. 1, pp. 69–79, February 1994.
 [4] F. Vernotte, M. Lenczner, P.Y. Bourgeois, and E. Rubiola, “The parabolic variance (PVAR), a wavelet variance based on leastsquare fit,” IEEE Transactions on Ultrasonics, Ferroelectrics and Frequency Control, vol. 63, no. 4, pp. 611–623, 2016.
 [5] N. J. Kasdin, “Discrete simulation of colored noise and stochastic processes and power law noise generation,” Proceedings of the IEEE, vol. 83, no. 5, pp. 802–827, May 1995.
Comments
There are no comments yet.