A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University

Size: px
Start display at page:

Download "A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University"

Transcription

1 A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University Lecture 18 Op0mal localiza0on Modeling Topics plan: Modeling (linear/non- linear least squares) Bayesian inference Bayesian approaches to spectral es0ma0on; also prewhitening methods Op0miza0on methods (needed for posterior PDFs, Bayes factors) Reading: Ch 10, 11 Gregory Assignment 3: now on web page References: Webpage: 1 Arrival Time Estimation with Matched Filtering Preliminary approaches 1

2 Localization Using Matched Filtering This handout describes localization of an object in a parameter space. For simplicity we consider localization of a pulse in time. The same formalism applies to localization of a spectral feature in frequency or to an image feature in a 2D image. The results can be extrapolated to a space of arbitrary dimensionality. I. First consider finding the amplitude of a pulse when the shape and location are known. Let the data be I(t) =aa(t)+n(t), where a = the unknown amplitude and n(t) is zero mean noise. The known pulse shape is A(t). Let the model be Î(t) =âa(t). Define the cost function to be the integrated squared error, 2 Q = dt I(t) Î(t). 20 Taking a derivative, we can solve for the estimate of the amplitude, â: âq = 2 dt I(t) Î(t) âî(t) =0 dt I(t) Î(t) A(t) =0 dtî(t)a(t) = dti(t)a(t) â dta 2 (t) = dti(t)a(t) Note that: â = dti(t)a(t) dta 2 (t). a. The model is linear in the sole parameter, â b. The numerator is the zero lag of the crosscorrelation function (CCF) of I(t) and A(t). c. The denominator is the zero lag of the autocorrelation function (ACF) of A(t). 21 2

3 II. Now consider the case where we don t know the location of the pulse in time (the time of arrival, TOA) and that it is the TOA we wish to estimate. We still know the pulse shape, a priori. Let the data, model and cost function be I(t) = aa(t t 0 )+n(t) Î(t) = âa(t ˆt 0 ). Q = dt I(t) Î(t) 2. Note that the model is linear in â but is nonlinear in ˆt 0. A nonlinear model requires an iterative solution, generally. 22 Minimizing Q with respect to â, we have âq = 2 dt dt â I(t) I(t) Î(t) dtî(t)a(t ˆt 0 )= dta 2 (t ˆt 0 )= â = dti(t)a(t ˆt 0 ) dta 2 (t ˆt 0 ). Î(t) âî(t) =0 A(t ˆt 0 )=0 dti(t)a(t ˆt 0 ) dti(t)a(t ˆt 0 ) (4) This last equation has the same form as in I. except that the estimate for the arrival time ˆt 0 is involved. 23 3

4 Now, minimizing Q with respect to ˆt 0, we have ˆt0 Q = 2 dt I(t) Î(t) ˆt0 Î(t) =0 â dt Î(t) A (t ˆt 0 )= â dt I(t)A (t ˆt 0 ) âa(t ˆt 0 ) â dt A(t ˆt 0 )A (t ˆt 0 )= dt I(t)A (t ˆt 0 ). (5) Grid Search: One approach to finding the arrival time is to search over a 2D grid of â, ˆt 0 to find the values that satisfy equations 4 and 5. This approach is inefficient. Instead, one can search over a 1D space for the single nonlinear parameter, ˆt 0, and then solve for â using either equation 4 or 5. Linearization + Iteration: Another method is to find solutions for â and ˆt 0, we can linearize the equations in ˆt 0 t 0 by using Taylor-series expansions for A(t ˆt 0 ) and A (t ˆt 0 ). Let ˆt 0 = t 0 + ˆt 0. Then, to first order in ˆt 0 : A(t ˆt 0 ) A(t t 0 ) A (t t 0 ) ˆt 0 A (t ˆt 0 ) A (t t 0 ) A (t t 0 ) ˆt 0 A 2 (t ˆt 0 ) A 2 (t t 0 ) 2A (t t 0 )A(t t 0 ) ˆt Now equations (4) and (5) become â â Consider the integral dt [A 2 (t t 0 ) 2 ˆt 0 A(t t 0 )A (t t 0 )] = dt I(t)[A(t t 0 ) ˆt 0 A (t t 0 )] dt [A(t t 0 )A (t t 0 ) ˆt 0 A(t t 0 )A (t t 0 ) ˆt 0 A 2 (t t 0 )] = dt I(t)[A (t t 0 ) ˆt 0 A (t t 0 )]. The integrand may be written as and so the integral equals dt A(t t 0 )A (t t 0 ). A(t t 0 )A (t t 0 )= 1 d 2 dt A2 (t t 0 ) 1 2 A2 (t t 0 ) t 2 t 1 0 in the limit of (e.g.) t 1,2 = T/2 with T pulse width. 25 4

5 We then have ˆt 0 â â dt A 2 (t t 0 ) = dt I(t)[A(t t 0 ) ˆt 0 A (t t 0 )] dt [A(t t 0 )A (t t 0 ) + A 2 (t t 0 )] = dt I(t)[A (t t 0 ) ˆt 0 A (t t 0 )]. Solving for â in both cases we have dt [I(t)A(t t0 ) ˆt 0 I(t)A (t t 0 )] â = dt A2 (t t 0 ) dt I(t) A (t t 0 )+ ˆt 0 A (t t 0 ) â = ˆt 0 dt A(t t0 )A (t t 0 )+A 2 (t t 0 ). (6) (7) 26 Using the notation we have Solving for i 0 dt I(t)A(t t 0 ) i 1 dt I(t)A (t t 0 ) i 2 dt I(t)A 2 (t t 0 ) i 3 dt I(t)A (t t 0 ) i 4 dt I(t) A(t t 0 )A (t t 0 )+A 2 (t t 0 ). ˆt 0 (to first order) we have â = i 0 ˆt 0 i 1 i 2 â = i 1 + ˆt 0 i 3 ˆt 0 i 4. ˆt 0 = i 1 i 2 i 0 i 4 + i 2 i (8) 5

6 Iterative Solution for ˆt 0 This equation can be solved iteratively for ˆt 0 : 0. choose a starting value for ˆt calculate ˆt 0 using the linearized equations. 2. is ˆt 0 =0? 3a. if yes, stop. 3b. if no, update ˆt 0 ˆt 0 + ˆt 0 and go back to step 1. For the best fit value for ˆt 0, the change is zero, ˆt 0 =0(top of the hill) and â can be calculated using one of the equations 6 or 7. Correlation Function Approach The iterative solution for ˆt 0 is similar to the following procedure that uses a crosscorrelation approach more directly: 1. cross correlate the template A(t) with I(t) to get a CCF. 2. find the lag of peak correlation as an estimate for the arrival time, ˆt 0 = max. 3. calculate â if needed. Subtleties of the Cross Correlation Method The CCF is calculated using sampled data and therefore is itself a discrete quantity. Often one wants greater precision on the arrival time than is given by the sample interval. I.e. we want a 28 floating point number for ˆt 0, not an integer index. Therefore we want to calculate the peak of the CCF by interpolating near its peak. The interpolation should be done properly by using the appropriate interpolation formula for sampled data (using the sinc function). Using parabolic interpolation yields excessive errors for the arrival time. In practice, the proper interpolation is effectively done in the frequency domain by calculating the phase shift of the Fourier transform of the CCF, which is the product of the Fourier transform of the template and the Fourier transform of the data. Arrival Time Errors 6

7 7

8 Arrival Time Estimation with Matched Filtering Least-squares approach 8

9 Arrival Time Estimation from Matched Filtering Least-squares solution and localization error Consider discrete sampling of a template and measured profile from which we want to determine the amplitude b and location through matched filtering (MF). The time and frequency domain quantities are: s t = template s k m t = model = a + bs t m k p t = measured profile = m t + n t p k n t = noise ñ k, N 1 where the Fourier transform is defined as s k = s t e 2itk/N. Frequency-domain approach: Using weights w k, minimize 2 = = t=0 N 1 w k p k m k 2 k=0 N 1 w k p k an k0 b s k e 2ik /N 2 k=0 1 We ignore the k =0term and consider only the first half of the array because time-domain quantities are real. So the cost function is (rewritten as Q) Q = w k p k 2 + b 2 s k 2 2bR e { p k s ke +2 ik/n }. (1) Taking derivatives we can solve for b and find an implicit equation for. b Q = 2b b = w k s k 2 2 w k R e p k s ke +2 ik/n w k s k 2 w k R e p k s ke +2 ik/n =0 2 9

10 is the solution of Q = 4( ib/n)b w k I m p k s k e +2 ik/n =0 w k I m p k s k e +2 ik/n =0 (2) Another approach: The same implicit equation for (Equation 2) can be gotten by multiplying the weighted DFT of the cross correlation, w k s k p k by the phasor e 2 ikˆ/n and finding the best value of ˆ: Q(ˆ) = = b When ˆ = the sum is maximized and real, w k s k p k e 2 ikˆ/n Q max = Q() =b w k s k 2 e 2 ikˆ/n (3) w k s k 2 3 Another solution for the scale factor is b = Q max w k s k 2 ˆ is the solution of ˆ R e {Q(ˆ )} =0or the solution of I m {Q(ˆ )} =0. Either one gives the same equation as Equation 2. Parameter errors: Expand Q from Equation 1 to second order, Q(b, ) Q min + b Q b + Q + 2 b Q ( b) Q ( ) b Q b = 2 b Q ( b) Q ( ) b Q b. Scale factor: Defining b as the error along the b axis in the b- plane and using b 2 Q =2 w k s k 2, 4 10

11 we have Q = Q min + 2 b Q ( b) 2 yielding Q = Q min + 2 b Q ( b) 2 Q min + 2 b Q 2 b b 2 = Q Q min b 2Q = Q Q min = 2 w k s k w k s k 2, where the rightmost equation results from defining the 1 error as the contour of Q that is one unit above the minimum. It is reasonable to define the weights in terms of the additive noise, so w k =1/k 2 and if they are all the same (as for white noise), then 2 b = 2 2 s k 2 = Nt 2 2 s k 2, where all k have been set equal to and then related to the rms noise level t in the time series. 5 Arrival time: Using the same approach we have From earlier expressions we get 2 = 2 2 = Q Q min 2 Q = 1 2 Q b w k k 2 R e { p k s N ke +2 ik /N } Note that the units of are in sample numbers. To get time units, we need to multiply by the sample interval, t. 6 11

12 Template Fitting of Millisecond Pulsars Pulse widths range from ~25 to 1000 µs In the best cases, timing precision is ~ 50 ns (a factor of 1/500 of the pulse width) Templates (average pulse shapes) appear to be stable over decades so pulsars can be used as precise clocks Departures from the templates occur because of fluctuations at the single pulse level Templates for MSPs 12

13 13

14 14

15 Pulsar P (ms) 1.56 B J B J J J J J J J B J J J J J J J J J J J J J J J J J J J J J J J B J J J Frequency Band = 1400 MHz Characterization of all MSPs used by NANOGrav for precision timing. Most MSPs are limited by template fitting errors (finite S/ N) but a few are limited by departures of pulse shapes from the template by pulsar-intrinsic effects min [µs] J,1/FWHM Implementation The method is easily implemented in python. Specifically, a nonlinear solver is used to minimize the cost function written here as Q = p k m k 2 = p k b s k e 2 ik/n 2, where we ignore the k =0term so that any mean offset of the template is irrelevant. We also use only the first half of the array because data and template are real. The two parameter fit yields the scale factor b and time offset. Errors are calculated as above. import scipy.optimize as spo tfft = fft(template) pfft = fft(profile) bhat0 = bccf tauhat0 = tauccf+ishift paramvec0 = array((bhat0, tauhat0)) paramvec = spo.minpack.leastsq(tfresids, paramvec0, args=(tfft, pfft)) bhat = paramvec[0][0] tauhat = paramvec[0][1] The module leastsq in scipy.optimize.minpack starts with an initial guess of the parameters paramvec0. As such, the returned result can be at a local minimum rather than the global minimum. How to check: try different starting values for paramvec

16 def tfresids(params, tfft, pfft): """ Calculates residuals between scaled and rotated template and data. """ b=params[0] tau=params[1] Nfft = size(pfft) Nsum = Nfft/2 arg=(2.*pi*tau/float(nfft)) * arange(0., Nfft, 1.) phasevec = cos(arg) - 1j*sin(arg) resids = abs(pfft[1:nsum] - b*tfft[1:nsum]*phasevec[1:nsum]) return resids def toa_errors_additive(tfft, b, sigma_t): """ Calculates error in b = scale factor and tau = TOA due to additive noise. input: fft of template b = fit value for scale factor sigma_t = rms additive noise in time domain output: sigma_b sigma_tau """ Nfft = size(tfft) Nsum = Nfft / 2 kvec = arange(1,nsum) sigma_b = sigma_t*sqrt(float(nfft) / (2.*sum(abs(tfft[1:Nsum])**2))) sigma_tau = (sigma_t*nfft/(2.*pi*abs(b))) \ * sqrt(float(nfft) \ / (2.*sum(kvec**2*abs(tfft[1:Nsum])**2))) return sigma_tau, sigma_b 8 About 700 lines of code. 16

17 17

18 Analysis of template fitting errors for simulated Gaussian profile Fourier RMS Fourier Total CCF RMS CCF Total MF Prediction Template fitting errors for Gaussian pulse with width of 10 bins vs. S/N TOA (bins) realizations S/N Parabolic interpolation of the CCF gives the same standard deviation as the Fourier approach and the theoretical MF error. But there is a systematic deviation from the true TOA that causes the large total error for parabolic interpolation RMS = std(best fit) Total = total mean-square difference from true TOA MF prediction = predicted RMS error for matched filtering. 18

19 Pulse timing Taylor 1992 Pulsar timing and relativistic gravity BY J. H. TAYLOR Joseph Henry Laboratories and Physics Department, Princeton University, Princeton, New Jersey 08544, U.S.A. In addition to being fascinating objects to study in their own right, pulsars are exquisite tools for probing a variety of issues in basic physics. Recycled pulsars, thought to have been spun up in previous episodes of mass accretion from orbiting companion stars, are especially well suited for such applications. They are extraordinarily stable clocks, approaching and perhaps exceeding the long-term stabilities of the best terrestrial time standards. Most of them are found in binary systems, with orbital velocities as large as 10-3 c. They provide unique opportunities for measuring neutron star masses, thereby yielding fundamental astrophysical data difficult to acquire by any other means. And they open the way for high precision tests of the nature of gravity under conditions much more 'relativistic' than found anywhere within the Solar System. Among other results, pulsar timing observations have convincingly established the existence of quadrupolar gravitational waves propagating at the speed of light. They have also placed interesting limits on possible departures of the strong-field nature of gravity from general relativity, on the rate of change of Newton's constant, G, and on the energy density of low-frequency gravitational waves in the universe. extending over several decades could lead to important new results in cosmology and Phil. Trans R. Soc. Lond. A (1992) 341, ? 1992 The Royal Society 1. Introduction and 19

20 Appendix A. Determining pulse times of arrival The applications of pulsar timing discussed in this paper depend crucially on the precision of measuring pulse times of arrival, so it is important to make the best possible use of all information in the data. Let us assume that observed and standard profiles, p(t) and s(t), have been obtained as described in?2a: the profiles are sampled and recorded at equally spaced intervals of time, tj = jat, j = 0,1,...,N- 1, with P = NAt. Before sampling, the detected signals will have been low-pass filtered at a cutoff frequency fc < (2At)-1. To avoid filtering out useful data, At is chosen small enough that fc can exceed the highest frequencies significantly present in the data. (Obviously what is 'significant' will depend on the available signal-to-noise ratio.) If the foregoing criteria are met, the finite sampling theorem (Bracewell 1965) ensures that all potentially available information is fully and unambiguously contained in the discretely sampled values pj = p(tj). If p(t) is equal to a shifted and scaled replica of s(t) plus random noise, as defined in (1), their Fourier transforms are also related in a simple way. Discrete Fourier transforms of the two profiles can be written as.n-1 Pkexp (iok,) = E pjei2zjk/n, (A 1) j=o N-i S exp (ik) = E Sjei2jk/N, (A 2) j=o where the frequency index k runs from 0 to N-1. Thus, the real quantities Pk and Sk are the amplitudes of the complex Fourier coefficients, and 0k and 5k are the phases. Linearity of the transform relationship implies that Pkexp (i)k) = an+bskexp [i( +? kr)] + Gk, k = 0,..., (N- 1), (A 3) where Gk represents random noise equal to the Fourier transform of the sampled noise in the time-domain profile, g(tj). Note that the bias a and scale factor b assume similar roles in the time and frequency domain (cf. equation (1)). As a consequence of the 'shift theorem' (Bracewell 1965), the time offset T appears in the frequency domain as the slope of a linear ramp, kr, added to the phases of the standard profile's Fourier coefficients. After the transforms have been computed, the value of a can be obtained immediately from the relation a = (P-bSo)/N. (A 4) The desired pulse time of arrival r, as well as the gain factor b, can be obtained by minimising the goodness-of-fit statistic Phil. Trans. R. Soc. Lond. A (1992) pk bsk exp [i(sx(b,t ) T= E. k + kt)] 2(A 5) (A 5) k=l k Pulsar timing and relativistic gravity 133 In this equation o(j is the root-to-mean-square amplitude of the noise at frequency k, and presumably the anti-aliasing low-pass filter will make the ok,s fall off somewhat at larger values of k. In practice, however, this subtlety is usually unimportant since the amplitudes P, and Sk decrease even faster than S0. Owing to inherent symmetries in the transforms, the limits of summation in (A 5) can be taken as 1 to VN, rather than 0 to N-1. For convenience of notation, in the remaining equations the summation limits have been omitted and the 0oks treated as constant. By replacing the complex exponential in (A 5) with trigonometric equivalents and expanding the indicated squared modulus, one obtains a more convenient expression for X2, namely X2 = k-2 (P+ b2s )-2b -~2 P Sk cos (k- Ok + kt). (A 6) At the global minimum of the two-dimensional function X2(T, b), its derivatives with respect to r and b must vanish. This requirement yields two equations in the two unknowns, namely ax2 2b = 02 E kpk Sk sin (k - 0k + k) = (A 7) ax2 2b - 2 ab =2 -S -2 E Pk Sk cos = 0. (A 8) (Okk-kr) Equation (A 7) can be solved for T by a straightforward iterative procedure such as Brent's method (see, for example, Press et al. 1986). Equation (A 8) then yields b = PSk COS (k- ok + kt)/ Sk. (A 9) Uncertainties in the estimated values of T and b may be found by approximating X2 near its minimum by the leading terms of a Taylor series, and determining the excursions of b or T required to increase the value of X2 by 1. This procedure leads to 2 22 = = - (A10) (J ( at2) 2b E k2p Sk cos (O+ - 1k + k7()' -1 _(02,yX2 2'3 (- ab2)= 2 2 (A 11) Notice that the data sampling interval, At = P/N, appears nowhere in (A 1)-(A 10). Our formalism for fitting TOAS in the Fourier transform domain places no limits on accuracy expressed as a fraction of At. In contrast, experience has shown that timedomain methods widely in use for determining TOAS do not readily produce arrivaltime accuracies smaller than about O.1At (see, for example, Rawley 1986). 20

21 Arrival Time Errors Here we wish to localize the occurrence of a function A(t). We will consider this to be a pulse whose arrival time t 0 we want to estimate, along with its expected error. Let the signal be I(t) =a 0 A(t where a 0 is the amplitude and n(t) is zero mean noise. t 0 )+n(t), We will find the time of arrival (TOA) by cross correlating the presumed known pulse shape A(t) with the signal: C AI ( )= dt I(t)A(t ). First, assume that the signal has been coarsely shifted so that the template is already aligned with the signal and that template is centered on t =0. This way we can assume that the arrival time estimate is a small correction to the coarse estimate. 29 We can expand the signal as I(t) a 0 A(t) We also expand the template, but to second order in find the lag of maximum correlation: Then we can write a 0 t 0 A (t)+n(t). A(t ) A(t) A (t)+ 2 A (t) A (t ) A (t) A (t). C IA( ) = d d C AI( )=0 = dt I(t)A (t ) C IA (0) because we will be taking a derivative to 2 C IA (0) An estimator for is then = C IA (0) C IA (0) = a 0C AA (0) a 0 t 0 C A A (0) + C na (0) a 0 C AA a 0 t 0 C A A (0) + C na (0). Using previous approximations we encounter the terms C AA and C A A (0) that vanish for pulses that are zero at ±. Also the C na term in the denominator yields a second-order term that can be ignored. Then the TOA estimator becomes = a 0t 0 C A A (0) + C na (0). a 0 C AA (0) 30 21

22 Noiseless case: When there is no noise we have = t 0 C A A (0) C AA (0). Using a trial function, such as a Gaussian shape, it can be shown that expect! Even better, the denominator can be integrated by parts to show that C AA (0) = pulses that vanish at ±, so the equality is general. With noise: We can now write = t 0 + C na (0) a 0 C AA (0). = t 0, as we would C A A (0) for 31 Then the mean-square TOA error is, using = t 0, ( ) 2 = C2 na (0) a 2 0 C2 AA (0) = C2 na (0) a 2 0 C2 A A (0) dtdt n(t)n(t ) A(t)A(t ) = a 2. 0 C2 A A (0) 32 22

23 Now assume white noise (for specificity) so that n(t)n(t ) = 2 nw n (t t ) where w n is a short characteristic time scale (such as an inverse bandwidth) to keep the units correct. Then ( ) 2 = 2 n w n dta 2 (t) a 0 CA 2 A (0) = 2 n w n a 0 C A A (0) We can then write this out as a TOA error = n a 0 = 1 SNR 1/2 w n dt [A (t)] 2 1/2 w n dt [A (t)] 2. We see that the error scales as the inverse of the signal to noise ratio (SNR). The denominator also involves the integral of the squared derivative of the pulse shape, suggesting that sharper pulses with larger derivatives will produce smaller arrival time errors. 33 Localization using CCFs :time and frequency domains 23

24 Localization: Time vs. Frequency Domains For a Nyquist sampled, bandlimited process with bandwidth B the sampling theorem implies that the continuous-time signal can be reconstructed from the sampled data as x(t) = where t =1/B and, as usual, sinc x (sin x)/x. Consider a model appropriate for matched filtering, x(t) =aa(t x n sinc(t n t)/ t t 0 )+n(t), where a is the amplitude, A(t) is the known template, and n is noise. Suppose we have sampled versions of the data and template, x n,a n. One approach is to calculate the discrete CCF C xa ( ) = 1 N n x n A n and find the maximum to determine an esimate ˆt 0 = max. However, we really want the CCF of the continuous time quantities, x(t) and A(t), C xt () = dt x(t)a(t ). 34 We cannot get the true lag of maximum correlation, max, by interpolating the sampled correlation function unless we interpolate according to the sampling theorem. If we interpolate differently, we will get biased results. Another approach: the frequency domain Take the FT of the model equation to get X(f) =aã(f)e 2 ift 0 + ñ(f). No noise: Write the FT in terms of its real and imaginary parts and find the phase: See example. X(f) = X r (f)+i X i (f) = X(f) e i (f) Xi (f) = tan 1 (f) sin 2 =tan X 1 ft0 = 2 ft 0. r (f) cos 2 ft 0 With noise: The phase will have a noise-like error that is nonlinearly related to ñ(f). In the limit of large SNR, the rms phase error will scale as 1/SNR. Working directly with the phase to determine t 0, however, is numerically problematic because the phase will wrap for large offsets and for low SNR

25

26

27

28 Best approach: Fit to the complex FFT rather than to the phase. Use as a model Then define the product M(f) = Ã(f)e 2 ifˆt 0 J(f) = X(f) M (f) =a Ã(f) 2 e 2 if(t ˆt 0) + ñ(f)ã (f)e 2 ifˆt 0 and integrate over frequency: S df J(f) f = a df Ã(f) 2 e 2 if(t 0 ˆt 0) + f df ñ(f)ã (f)e 2 ifˆt 0. f This quantity can be maximized vs. ˆt 0. For no noise, we expect ˆt 0 = t 0. The maximum can be found using standard search methods over the nonlinear parameter ˆt 0 (grid search, linearization, etc.). Note that the integrand is naturally weighted by the actual signal. An equivalent approach is to use a different test statistic, S 2 = df X(f) M(f), f that we would minimize to find ˆt Note that we have used continuous notation here. With sampled data we can reconstruct the continuous FT as X(f) = (f/b) x n e 2 ifn t, which can be implemented using the DFT. n 43 28

29 Modeling Criteria for Modeling See slides 58 29

30 Suppose we have the following y = data vector x = independent variable Criteria for Modeling ŷ( ) = model for data with parameters f y (y; ) = multivariate PDF for the data. ˆ = vector of parameters that yield the best fit of the model to the data according to some criterion or ˆ = parameters of a probability density function for the data. Suppose we have a model for the data that consists of parameters. These might be parameters of a time series or parameters of a PDF for the data. 1 (1) Least squares minimize with respect to : Q( ) j [y i ŷ i ( )] 2 (2) Maximum likelihood: For the data {y i,i =1,N} suppose the joint PDFis modeled as or known to be f y (y; ) After obtaining the N data points, we view the PDF as the likelihood of getting those actual data points given a choice of the parameters,. i.e. L ( )=f y (y; ) and those values of that maximize L ( ) are maximum likelihood estimators for. 2 30

31 Measures of Optimality For a given parameter and data set, there may be several or many possible estimators. We need quantitative guidelines on how to choose the best one. The criteria often used include: Consistency: (convergence) Let n = sample size lim ˆ =0 n Unbiased: ˆ = or B = ˆ ; consistency unbiased because for finite n, the estimator can be biased even though the bias 0 as n. (e.g. maximum likelihood estimators) Minimum variance: Var (ˆ) ˆ2 ˆ 2 if Var ˆ is minimized, the resultant estimator yields the least variation of estimates (note, however, that the MV estimator may be biased) 3 Mean-square error: MSE = ˆ 2 = [ˆ ˆ + ˆ ] 2 =Var(ˆ)+B 2 Efficiency: A well-designed experiment yields data that are all used in an estimate. An inefficient experiment yields superfluous data. Thus, experiments should be designed vis á vis an estimator. As an example, an estimator is said to be mean-square efficient if no other estimator has a smaller cost function: (ˆ ) 2 < (ˆ ) 2 where is any other estimator of. 4 31

32 Sufficiency: if ˆ is a sufficient statistic it contains all information obtainable from a sample on. Formally, consider an estimator ˆ and another ˆ and the conditional distribution F (ˆ ˆ) P {ˆ < some number ˆ} If ˆ is not a function of ˆ and if F (ˆ ˆ) is independent of (actual value) then ˆ is a sufficient statistic. More transparent is an example! Suppose you have N data points, {x n,n=1,n} that you know are distributed as N(µ, ). You calculate the sample mean and standard deviation, ˆµ = N N 1 n=1 x n ˆ 2 = 1 N (x n N n=1 ˆµ) 2. 5 In calculating the likelihood function L (µ, ) for µ and it can be shown that L depends only on ˆµ and ˆ 2. Thus ˆµ and ˆ 2 contain all the information needed for likelihood inference and are thus sufficient statistics. 6 32

33 Least squares fitting Linear Minimum chi^2 (or ML) solu0on is easy and unique Nonlinear Solving the NL problem (a plethora of approaches, including MCMC) can be challenging with large dimensionality General Want solu0ons for parameter vectors Want errors and covariances of parameters à covariance matrix for the parameters to be derived from (e.g.) covariance matrix of the data Confidence intervals on parameter values Comparison of models Hypothesis tes0ng 65 33

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2013

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2013 A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2013 Lecture 26 Localization/Matched Filtering (continued) Prewhitening Lectures next week: Reading Bases, principal

More information

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011 A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011 Reading: Chapter 10 = linear LSQ with Gaussian errors Chapter 11 = Nonlinear fitting Chapter 12 = Markov Chain Monte

More information

F & B Approaches to a simple model

F & B Approaches to a simple model A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 215 http://www.astro.cornell.edu/~cordes/a6523 Lecture 11 Applications: Model comparison Challenges in large-scale surveys

More information

A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University

A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University Lecture 19 Modeling Topics plan: Modeling (linear/non- linear least squares) Bayesian inference Bayesian approaches to spectral esbmabon;

More information

imin...

imin... Pulsar Timing For a detailed look at pulsar timing and other pulsar observing techniques, see the Handbook of Pulsar Astronomy by Duncan Lorimer and Michael Kramer. Pulsars are intrinsically interesting

More information

A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University. False Positives in Fourier Spectra. For N = DFT length: Lecture 5 Reading

A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University. False Positives in Fourier Spectra. For N = DFT length: Lecture 5 Reading A6523 Modeling, Inference, and Mining Jim Cordes, Cornell University Lecture 5 Reading Notes on web page Stochas

More information

A523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011

A523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011 A523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011 Lecture 6 PDFs for Lecture 1-5 are on the web page Problem set 2 is on the web page Article on web page A Guided

More information

Statistics. Lent Term 2015 Prof. Mark Thomson. 2: The Gaussian Limit

Statistics. Lent Term 2015 Prof. Mark Thomson. 2: The Gaussian Limit Statistics Lent Term 2015 Prof. Mark Thomson Lecture 2 : The Gaussian Limit Prof. M.A. Thomson Lent Term 2015 29 Lecture Lecture Lecture Lecture 1: Back to basics Introduction, Probability distribution

More information

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2015 http://www.astro.cornell.edu/~cordes/a6523 Lecture 4 See web page later tomorrow Searching for Monochromatic Signals

More information

Lecture 3. G. Cowan. Lecture 3 page 1. Lectures on Statistical Data Analysis

Lecture 3. G. Cowan. Lecture 3 page 1. Lectures on Statistical Data Analysis Lecture 3 1 Probability (90 min.) Definition, Bayes theorem, probability densities and their properties, catalogue of pdfs, Monte Carlo 2 Statistical tests (90 min.) general concepts, test statistics,

More information

Primer on statistics:

Primer on statistics: Primer on statistics: MLE, Confidence Intervals, and Hypothesis Testing ryan.reece@gmail.com http://rreece.github.io/ Insight Data Science - AI Fellows Workshop Feb 16, 018 Outline 1. Maximum likelihood

More information

Stochastic Processes: I. consider bowl of worms model for oscilloscope experiment:

Stochastic Processes: I. consider bowl of worms model for oscilloscope experiment: Stochastic Processes: I consider bowl of worms model for oscilloscope experiment: SAPAscope 2.0 / 0 1 RESET SAPA2e 22, 23 II 1 stochastic process is: Stochastic Processes: II informally: bowl + drawing

More information

Signal Modeling, Statistical Inference and Data Mining in Astrophysics

Signal Modeling, Statistical Inference and Data Mining in Astrophysics ASTRONOMY 6523 Spring 2013 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Course Approach The philosophy of the course reflects that of the instructor, who takes a dualistic view

More information

Regression. Oscar García

Regression. Oscar García Regression Oscar García Regression methods are fundamental in Forest Mensuration For a more concise and general presentation, we shall first review some matrix concepts 1 Matrices An order n m matrix is

More information

Continuous Wave Data Analysis: Fully Coherent Methods

Continuous Wave Data Analysis: Fully Coherent Methods Continuous Wave Data Analysis: Fully Coherent Methods John T. Whelan School of Gravitational Waves, Warsaw, 3 July 5 Contents Signal Model. GWs from rotating neutron star............ Exercise: JKS decomposition............

More information

Statistical Data Analysis Stat 3: p-values, parameter estimation

Statistical Data Analysis Stat 3: p-values, parameter estimation Statistical Data Analysis Stat 3: p-values, parameter estimation London Postgraduate Lectures on Particle Physics; University of London MSci course PH4515 Glen Cowan Physics Department Royal Holloway,

More information

Gravity and the Unseen Sky. Sydney J. Chamberlin

Gravity and the Unseen Sky. Sydney J. Chamberlin Gravity and the Unseen Sky Sydney J. Chamberlin Outline Gravitational-wave detectors The road to detecting stochastic backgrounds with pulsar timing arrays Testing general relativity with pulsar timing

More information

Modern Methods of Data Analysis - WS 07/08

Modern Methods of Data Analysis - WS 07/08 Modern Methods of Data Analysis Lecture VIc (19.11.07) Contents: Maximum Likelihood Fit Maximum Likelihood (I) Assume N measurements of a random variable Assume them to be independent and distributed according

More information

ESS Finite Impulse Response Filters and the Z-transform

ESS Finite Impulse Response Filters and the Z-transform 9. Finite Impulse Response Filters and the Z-transform We are going to have two lectures on filters you can find much more material in Bob Crosson s notes. In the first lecture we will focus on some of

More information

Gaussian processes. Chuong B. Do (updated by Honglak Lee) November 22, 2008

Gaussian processes. Chuong B. Do (updated by Honglak Lee) November 22, 2008 Gaussian processes Chuong B Do (updated by Honglak Lee) November 22, 2008 Many of the classical machine learning algorithms that we talked about during the first half of this course fit the following pattern:

More information

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring Lecture 8 A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2015 http://www.astro.cornell.edu/~cordes/a6523 Applications: Bayesian inference: overview and examples Introduction

More information

Statistical techniques for data analysis in Cosmology

Statistical techniques for data analysis in Cosmology Statistical techniques for data analysis in Cosmology arxiv:0712.3028; arxiv:0911.3105 Numerical recipes (the bible ) Licia Verde ICREA & ICC UB-IEEC http://icc.ub.edu/~liciaverde outline Lecture 1: Introduction

More information

SGN Advanced Signal Processing: Lecture 8 Parameter estimation for AR and MA models. Model order selection

SGN Advanced Signal Processing: Lecture 8 Parameter estimation for AR and MA models. Model order selection SG 21006 Advanced Signal Processing: Lecture 8 Parameter estimation for AR and MA models. Model order selection Ioan Tabus Department of Signal Processing Tampere University of Technology Finland 1 / 28

More information

Physics 6720 Introduction to Statistics April 4, 2017

Physics 6720 Introduction to Statistics April 4, 2017 Physics 6720 Introduction to Statistics April 4, 2017 1 Statistics of Counting Often an experiment yields a result that can be classified according to a set of discrete events, giving rise to an integer

More information

ELEG 3143 Probability & Stochastic Process Ch. 6 Stochastic Process

ELEG 3143 Probability & Stochastic Process Ch. 6 Stochastic Process Department of Electrical Engineering University of Arkansas ELEG 3143 Probability & Stochastic Process Ch. 6 Stochastic Process Dr. Jingxian Wu wuj@uark.edu OUTLINE 2 Definition of stochastic process (random

More information

Statistics: Learning models from data

Statistics: Learning models from data DS-GA 1002 Lecture notes 5 October 19, 2015 Statistics: Learning models from data Learning models from data that are assumed to be generated probabilistically from a certain unknown distribution is a crucial

More information

Advanced Statistical Methods. Lecture 6

Advanced Statistical Methods. Lecture 6 Advanced Statistical Methods Lecture 6 Convergence distribution of M.-H. MCMC We denote the PDF estimated by the MCMC as. It has the property Convergence distribution After some time, the distribution

More information

Lecture 23. Lidar Error and Sensitivity Analysis (2)

Lecture 23. Lidar Error and Sensitivity Analysis (2) Lecture 3. Lidar Error and Sensitivity Analysis ) q Derivation of Errors q Background vs. Noise q Sensitivity Analysis q Summary 1 Accuracy vs. Precision in Lidar Measurements q The precision errors caused

More information

Why is the field of statistics still an active one?

Why is the field of statistics still an active one? Why is the field of statistics still an active one? It s obvious that one needs statistics: to describe experimental data in a compact way, to compare datasets, to ask whether data are consistent with

More information

Lecture 5. G. Cowan Lectures on Statistical Data Analysis Lecture 5 page 1

Lecture 5. G. Cowan Lectures on Statistical Data Analysis Lecture 5 page 1 Lecture 5 1 Probability (90 min.) Definition, Bayes theorem, probability densities and their properties, catalogue of pdfs, Monte Carlo 2 Statistical tests (90 min.) general concepts, test statistics,

More information

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011 A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011 Reading Chapter 5 of Gregory (Frequentist Statistical Inference) Lecture 7 Examples of FT applications Simulating

More information

that efficiently utilizes the total available channel bandwidth W.

that efficiently utilizes the total available channel bandwidth W. Signal Design for Band-Limited Channels Wireless Information Transmission System Lab. Institute of Communications Engineering g National Sun Yat-sen University Introduction We consider the problem of signal

More information

EE 574 Detection and Estimation Theory Lecture Presentation 8

EE 574 Detection and Estimation Theory Lecture Presentation 8 Lecture Presentation 8 Aykut HOCANIN Dept. of Electrical and Electronic Engineering 1/14 Chapter 3: Representation of Random Processes 3.2 Deterministic Functions:Orthogonal Representations For a finite-energy

More information

Estimation of Parameters

Estimation of Parameters CHAPTER Probability, Statistics, and Reliability for Engineers and Scientists FUNDAMENTALS OF STATISTICAL ANALYSIS Second Edition A. J. Clark School of Engineering Department of Civil and Environmental

More information

A523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011

A523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011 A523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2011 Lecture 1 Organization:» Syllabus (text, requirements, topics)» Course approach (goals, themes) Book: Gregory, Bayesian

More information

Signal Design for Band-Limited Channels

Signal Design for Band-Limited Channels Wireless Information Transmission System Lab. Signal Design for Band-Limited Channels Institute of Communications Engineering National Sun Yat-sen University Introduction We consider the problem of signal

More information

Lecture 3. Linear Regression II Bastian Leibe RWTH Aachen

Lecture 3. Linear Regression II Bastian Leibe RWTH Aachen Advanced Machine Learning Lecture 3 Linear Regression II 02.11.2015 Bastian Leibe RWTH Aachen http://www.vision.rwth-aachen.de/ leibe@vision.rwth-aachen.de This Lecture: Advanced Machine Learning Regression

More information

Gravitational-Wave Data Analysis: Lecture 2

Gravitational-Wave Data Analysis: Lecture 2 Gravitational-Wave Data Analysis: Lecture 2 Peter S. Shawhan Gravitational Wave Astronomy Summer School May 29, 2012 Outline for Today Matched filtering in the time domain Matched filtering in the frequency

More information

Digital Baseband Systems. Reference: Digital Communications John G. Proakis

Digital Baseband Systems. Reference: Digital Communications John G. Proakis Digital Baseband Systems Reference: Digital Communications John G. Proais Baseband Pulse Transmission Baseband digital signals - signals whose spectrum extend down to or near zero frequency. Model of the

More information

Pulsars: Observation & Timing. Stephen Eikenberry 28 Jan 2014

Pulsars: Observation & Timing. Stephen Eikenberry 28 Jan 2014 Pulsars: Observation & Timing Stephen Eikenberry 28 Jan 2014 Timing Pulsars Figure: D. Nice Pulsars send out pulsed signals We receive them Goal is to measure a Time of Arrival (TOA); this provides great

More information

Problem Sheet 1 Examples of Random Processes

Problem Sheet 1 Examples of Random Processes RANDOM'PROCESSES'AND'TIME'SERIES'ANALYSIS.'PART'II:'RANDOM'PROCESSES' '''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''Problem'Sheets' Problem Sheet 1 Examples of Random Processes 1. Give

More information

Probing Relativistic Gravity with the Double Pulsar

Probing Relativistic Gravity with the Double Pulsar Probing Relativistic Gravity with the Double Pulsar Marta Burgay INAF Osservatorio Astronomico di Cagliari The spin period of the original millisecond pulsar PSR B1937+21: P = 0.0015578064924327 ± 0.0000000000000004

More information

The Expectation-Maximization Algorithm

The Expectation-Maximization Algorithm 1/29 EM & Latent Variable Models Gaussian Mixture Models EM Theory The Expectation-Maximization Algorithm Mihaela van der Schaar Department of Engineering Science University of Oxford MLE for Latent Variable

More information

Stochastic Processes. M. Sami Fadali Professor of Electrical Engineering University of Nevada, Reno

Stochastic Processes. M. Sami Fadali Professor of Electrical Engineering University of Nevada, Reno Stochastic Processes M. Sami Fadali Professor of Electrical Engineering University of Nevada, Reno 1 Outline Stochastic (random) processes. Autocorrelation. Crosscorrelation. Spectral density function.

More information

Stochastic Processes. A stochastic process is a function of two variables:

Stochastic Processes. A stochastic process is a function of two variables: Stochastic Processes Stochastic: from Greek stochastikos, proceeding by guesswork, literally, skillful in aiming. A stochastic process is simply a collection of random variables labelled by some parameter:

More information

Machine Learning 4771

Machine Learning 4771 Machine Learning 4771 Instructor: Tony Jebara Topic 11 Maximum Likelihood as Bayesian Inference Maximum A Posteriori Bayesian Gaussian Estimation Why Maximum Likelihood? So far, assumed max (log) likelihood

More information

Statistics. Lecture 2 August 7, 2000 Frank Porter Caltech. The Fundamentals; Point Estimation. Maximum Likelihood, Least Squares and All That

Statistics. Lecture 2 August 7, 2000 Frank Porter Caltech. The Fundamentals; Point Estimation. Maximum Likelihood, Least Squares and All That Statistics Lecture 2 August 7, 2000 Frank Porter Caltech The plan for these lectures: The Fundamentals; Point Estimation Maximum Likelihood, Least Squares and All That What is a Confidence Interval? Interval

More information

Lecture 7. Fourier Analysis

Lecture 7. Fourier Analysis Lecture 7 Fourier Analysis Summary Lecture 6 Minima and maxima 1 dimension : Bracket minima : 3 values of f(x) : f(2) < f(1) and f(2)

More information

Signal Processing - Lecture 7

Signal Processing - Lecture 7 1 Introduction Signal Processing - Lecture 7 Fitting a function to a set of data gathered in time sequence can be viewed as signal processing or learning, and is an important topic in information theory.

More information

Design Criteria for the Quadratically Interpolated FFT Method (I): Bias due to Interpolation

Design Criteria for the Quadratically Interpolated FFT Method (I): Bias due to Interpolation CENTER FOR COMPUTER RESEARCH IN MUSIC AND ACOUSTICS DEPARTMENT OF MUSIC, STANFORD UNIVERSITY REPORT NO. STAN-M-4 Design Criteria for the Quadratically Interpolated FFT Method (I): Bias due to Interpolation

More information

Correlator I. Basics. Chapter Introduction. 8.2 Digitization Sampling. D. Anish Roshi

Correlator I. Basics. Chapter Introduction. 8.2 Digitization Sampling. D. Anish Roshi Chapter 8 Correlator I. Basics D. Anish Roshi 8.1 Introduction A radio interferometer measures the mutual coherence function of the electric field due to a given source brightness distribution in the sky.

More information

Part 4: Multi-parameter and normal models

Part 4: Multi-parameter and normal models Part 4: Multi-parameter and normal models 1 The normal model Perhaps the most useful (or utilized) probability model for data analysis is the normal distribution There are several reasons for this, e.g.,

More information

Lecture 4 - Spectral Estimation

Lecture 4 - Spectral Estimation Lecture 4 - Spectral Estimation The Discrete Fourier Transform The Discrete Fourier Transform (DFT) is the equivalent of the continuous Fourier Transform for signals known only at N instants separated

More information

Frequentist-Bayesian Model Comparisons: A Simple Example

Frequentist-Bayesian Model Comparisons: A Simple Example Frequentist-Bayesian Model Comparisons: A Simple Example Consider data that consist of a signal y with additive noise: Data vector (N elements): D = y + n The additive noise n has zero mean and diagonal

More information

Bayesian Regression Linear and Logistic Regression

Bayesian Regression Linear and Logistic Regression When we want more than point estimates Bayesian Regression Linear and Logistic Regression Nicole Beckage Ordinary Least Squares Regression and Lasso Regression return only point estimates But what if we

More information

Statistics and Data Analysis

Statistics and Data Analysis Statistics and Data Analysis The Crash Course Physics 226, Fall 2013 "There are three kinds of lies: lies, damned lies, and statistics. Mark Twain, allegedly after Benjamin Disraeli Statistics and Data

More information

Neural Network Training

Neural Network Training Neural Network Training Sargur Srihari Topics in Network Training 0. Neural network parameters Probabilistic problem formulation Specifying the activation and error functions for Regression Binary classification

More information

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2013 Lecture 4 For next week be sure to have read Chapter 5 of

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2013 Lecture 4 For next week be sure to have read Chapter 5 of A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2013 Lecture 4 For next week be sure to have read Chapter 5 of Gregory (Frequentist Statistical Inference) Today: DFT

More information

An Introduction to Expectation-Maximization

An Introduction to Expectation-Maximization An Introduction to Expectation-Maximization Dahua Lin Abstract This notes reviews the basics about the Expectation-Maximization EM) algorithm, a popular approach to perform model estimation of the generative

More information

Delphine Perrodin INAF-Osservatorio Astronomico di Cagliari (Italy) IPTA Student Week 2017, Sèvres, France

Delphine Perrodin INAF-Osservatorio Astronomico di Cagliari (Italy) IPTA Student Week 2017, Sèvres, France Delphine Perrodin INAF-Osservatorio Astronomico di Cagliari (Italy) IPTA Student Week 2017, Sèvres, France Outline! Introduction to Gravitational Wave detectors! Introduction to Pulsar Timing Arrays! Stochastic

More information

POLYNOMIAL-BASED INTERPOLATION FILTERS PART I: FILTER SYNTHESIS*

POLYNOMIAL-BASED INTERPOLATION FILTERS PART I: FILTER SYNTHESIS* CIRCUITS SYSTEMS SIGNAL PROCESSING c Birkhäuser Boston (27) VOL. 26, NO. 2, 27, PP. 115 146 DOI: 1.17/s34-5-74-8 POLYNOMIAL-BASED INTERPOLATION FILTERS PART I: FILTER SYNTHESIS* Jussi Vesma 1 and Tapio

More information

Chapter 3: Maximum-Likelihood & Bayesian Parameter Estimation (part 1)

Chapter 3: Maximum-Likelihood & Bayesian Parameter Estimation (part 1) HW 1 due today Parameter Estimation Biometrics CSE 190 Lecture 7 Today s lecture was on the blackboard. These slides are an alternative presentation of the material. CSE190, Winter10 CSE190, Winter10 Chapter

More information

An Introduction to Bayesian Linear Regression

An Introduction to Bayesian Linear Regression An Introduction to Bayesian Linear Regression APPM 5720: Bayesian Computation Fall 2018 A SIMPLE LINEAR MODEL Suppose that we observe explanatory variables x 1, x 2,..., x n and dependent variables y 1,

More information

Module 2. Random Processes. Version 2, ECE IIT, Kharagpur

Module 2. Random Processes. Version 2, ECE IIT, Kharagpur Module Random Processes Version, ECE IIT, Kharagpur Lesson 9 Introduction to Statistical Signal Processing Version, ECE IIT, Kharagpur After reading this lesson, you will learn about Hypotheses testing

More information

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring A653 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 15 http://www.astro.cornell.edu/~cordes/a653 Lecture 3 Power spectrum issues Frequentist approach Bayesian approach (some

More information

Estimators as Random Variables

Estimators as Random Variables Estimation Theory Overview Properties Bias, Variance, and Mean Square Error Cramér-Rao lower bound Maimum likelihood Consistency Confidence intervals Properties of the mean estimator Introduction Up until

More information

Data Converter Fundamentals

Data Converter Fundamentals Data Converter Fundamentals David Johns and Ken Martin (johns@eecg.toronto.edu) (martin@eecg.toronto.edu) slide 1 of 33 Introduction Two main types of converters Nyquist-Rate Converters Generate output

More information

Practice Problems Section Problems

Practice Problems Section Problems Practice Problems Section 4-4-3 4-4 4-5 4-6 4-7 4-8 4-10 Supplemental Problems 4-1 to 4-9 4-13, 14, 15, 17, 19, 0 4-3, 34, 36, 38 4-47, 49, 5, 54, 55 4-59, 60, 63 4-66, 68, 69, 70, 74 4-79, 81, 84 4-85,

More information

Variable and Periodic Signals in Astronomy

Variable and Periodic Signals in Astronomy Lecture 14: Variability and Periodicity Outline 1 Variable and Periodic Signals in Astronomy 2 Lomb-Scarle diagrams 3 Phase dispersion minimisation 4 Kolmogorov-Smirnov tests 5 Fourier Analysis Christoph

More information

Tracking of Spread Spectrum Signals

Tracking of Spread Spectrum Signals Chapter 7 Tracking of Spread Spectrum Signals 7. Introduction As discussed in the last chapter, there are two parts to the synchronization process. The first stage is often termed acquisition and typically

More information

Additional Keplerian Signals in the HARPS data for Gliese 667C from a Bayesian re-analysis

Additional Keplerian Signals in the HARPS data for Gliese 667C from a Bayesian re-analysis Additional Keplerian Signals in the HARPS data for Gliese 667C from a Bayesian re-analysis Phil Gregory, Samantha Lawler, Brett Gladman Physics and Astronomy Univ. of British Columbia Abstract A re-analysis

More information

p(z)

p(z) Chapter Statistics. Introduction This lecture is a quick review of basic statistical concepts; probabilities, mean, variance, covariance, correlation, linear regression, probability density functions and

More information

DETECTION theory deals primarily with techniques for

DETECTION theory deals primarily with techniques for ADVANCED SIGNAL PROCESSING SE Optimum Detection of Deterministic and Random Signals Stefan Tertinek Graz University of Technology turtle@sbox.tugraz.at Abstract This paper introduces various methods for

More information

On 1.9, you will need to use the facts that, for any x and y, sin(x+y) = sin(x) cos(y) + cos(x) sin(y). cos(x+y) = cos(x) cos(y) - sin(x) sin(y).

On 1.9, you will need to use the facts that, for any x and y, sin(x+y) = sin(x) cos(y) + cos(x) sin(y). cos(x+y) = cos(x) cos(y) - sin(x) sin(y). On 1.9, you will need to use the facts that, for any x and y, sin(x+y) = sin(x) cos(y) + cos(x) sin(y). cos(x+y) = cos(x) cos(y) - sin(x) sin(y). (sin(x)) 2 + (cos(x)) 2 = 1. 28 1 Characteristics of Time

More information

Sensor Tasking and Control

Sensor Tasking and Control Sensor Tasking and Control Sensing Networking Leonidas Guibas Stanford University Computation CS428 Sensor systems are about sensing, after all... System State Continuous and Discrete Variables The quantities

More information

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2015 http://www.astro.cornell.edu/~cordes/a6523 Lecture 23:! Nonlinear least squares!! Notes Modeling2015.pdf on course

More information

Universität Potsdam Institut für Informatik Lehrstuhl Maschinelles Lernen. Bayesian Learning. Tobias Scheffer, Niels Landwehr

Universität Potsdam Institut für Informatik Lehrstuhl Maschinelles Lernen. Bayesian Learning. Tobias Scheffer, Niels Landwehr Universität Potsdam Institut für Informatik Lehrstuhl Maschinelles Lernen Bayesian Learning Tobias Scheffer, Niels Landwehr Remember: Normal Distribution Distribution over x. Density function with parameters

More information

Covariance function estimation in Gaussian process regression

Covariance function estimation in Gaussian process regression Covariance function estimation in Gaussian process regression François Bachoc Department of Statistics and Operations Research, University of Vienna WU Research Seminar - May 2015 François Bachoc Gaussian

More information

2. SPECTRAL ANALYSIS APPLIED TO STOCHASTIC PROCESSES

2. SPECTRAL ANALYSIS APPLIED TO STOCHASTIC PROCESSES 2. SPECTRAL ANALYSIS APPLIED TO STOCHASTIC PROCESSES 2.0 THEOREM OF WIENER- KHINTCHINE An important technique in the study of deterministic signals consists in using harmonic functions to gain the spectral

More information

Discrete-Time Signals and Systems. Efficient Computation of the DFT: FFT Algorithms. Analog-to-Digital Conversion. Sampling Process.

Discrete-Time Signals and Systems. Efficient Computation of the DFT: FFT Algorithms. Analog-to-Digital Conversion. Sampling Process. iscrete-time Signals and Systems Efficient Computation of the FT: FFT Algorithms r. eepa Kundur University of Toronto Reference: Sections 6.1, 6., 6.4, 6.5 of John G. Proakis and imitris G. Manolakis,

More information

Metropolis Algorithm

Metropolis Algorithm //7 A Modeling, Inference, and Mining Jim Cordes, Cornell University Lecture MCMC example Reading: Ch, and in Gregory (from before) Chapter 9 of Mackay (Monte Carlo Methods) hip://www.inference.phy.cam.ac.uk/itprnn/

More information

Likelihood-Based Methods

Likelihood-Based Methods Likelihood-Based Methods Handbook of Spatial Statistics, Chapter 4 Susheela Singh September 22, 2016 OVERVIEW INTRODUCTION MAXIMUM LIKELIHOOD ESTIMATION (ML) RESTRICTED MAXIMUM LIKELIHOOD ESTIMATION (REML)

More information

sine wave fit algorithm

sine wave fit algorithm TECHNICAL REPORT IR-S3-SB-9 1 Properties of the IEEE-STD-57 four parameter sine wave fit algorithm Peter Händel, Senior Member, IEEE Abstract The IEEE Standard 57 (IEEE-STD-57) provides algorithms for

More information

If we want to analyze experimental or simulated data we might encounter the following tasks:

If we want to analyze experimental or simulated data we might encounter the following tasks: Chapter 1 Introduction If we want to analyze experimental or simulated data we might encounter the following tasks: Characterization of the source of the signal and diagnosis Studying dependencies Prediction

More information

Parameter estimation! and! forecasting! Cristiano Porciani! AIfA, Uni-Bonn!

Parameter estimation! and! forecasting! Cristiano Porciani! AIfA, Uni-Bonn! Parameter estimation! and! forecasting! Cristiano Porciani! AIfA, Uni-Bonn! Questions?! C. Porciani! Estimation & forecasting! 2! Cosmological parameters! A branch of modern cosmological research focuses

More information

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring

A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring Lecture 9 A6523 Signal Modeling, Statistical Inference and Data Mining in Astrophysics Spring 2015 http://www.astro.cornell.edu/~cordes/a6523 Applications: Comparison of Frequentist and Bayesian inference

More information

SEISMIC WAVE PROPAGATION. Lecture 2: Fourier Analysis

SEISMIC WAVE PROPAGATION. Lecture 2: Fourier Analysis SEISMIC WAVE PROPAGATION Lecture 2: Fourier Analysis Fourier Series & Fourier Transforms Fourier Series Review of trigonometric identities Analysing the square wave Fourier Transform Transforms of some

More information

(Gaussian) Random Fields

(Gaussian) Random Fields 23/01/2017 (Gaussian) Random Fields Echo of the Big Bang: Cosmic Microwave Background Planck (2013) Earliest view of the Universe: 379000 yrs. after Big Bang, 13.8 Gyr ago. 1 CMB Temperature Perturbations

More information

Data Detection for Controlled ISI. h(nt) = 1 for n=0,1 and zero otherwise.

Data Detection for Controlled ISI. h(nt) = 1 for n=0,1 and zero otherwise. Data Detection for Controlled ISI *Symbol by symbol suboptimum detection For the duobinary signal pulse h(nt) = 1 for n=0,1 and zero otherwise. The samples at the output of the receiving filter(demodulator)

More information

Advanced Introduction to Machine Learning CMU-10715

Advanced Introduction to Machine Learning CMU-10715 Advanced Introduction to Machine Learning CMU-10715 Gaussian Processes Barnabás Póczos http://www.gaussianprocess.org/ 2 Some of these slides in the intro are taken from D. Lizotte, R. Parr, C. Guesterin

More information

STATISTICS/ECONOMETRICS PREP COURSE PROF. MASSIMO GUIDOLIN

STATISTICS/ECONOMETRICS PREP COURSE PROF. MASSIMO GUIDOLIN Massimo Guidolin Massimo.Guidolin@unibocconi.it Dept. of Finance STATISTICS/ECONOMETRICS PREP COURSE PROF. MASSIMO GUIDOLIN SECOND PART, LECTURE 2: MODES OF CONVERGENCE AND POINT ESTIMATION Lecture 2:

More information

Multivariate Distribution Models

Multivariate Distribution Models Multivariate Distribution Models Model Description While the probability distribution for an individual random variable is called marginal, the probability distribution for multiple random variables is

More information

Parametric Techniques Lecture 3

Parametric Techniques Lecture 3 Parametric Techniques Lecture 3 Jason Corso SUNY at Buffalo 22 January 2009 J. Corso (SUNY at Buffalo) Parametric Techniques Lecture 3 22 January 2009 1 / 39 Introduction In Lecture 2, we learned how to

More information

Kalman filtering and friends: Inference in time series models. Herke van Hoof slides mostly by Michael Rubinstein

Kalman filtering and friends: Inference in time series models. Herke van Hoof slides mostly by Michael Rubinstein Kalman filtering and friends: Inference in time series models Herke van Hoof slides mostly by Michael Rubinstein Problem overview Goal Estimate most probable state at time k using measurement up to time

More information

Physics 403. Segev BenZvi. Parameter Estimation, Correlations, and Error Bars. Department of Physics and Astronomy University of Rochester

Physics 403. Segev BenZvi. Parameter Estimation, Correlations, and Error Bars. Department of Physics and Astronomy University of Rochester Physics 403 Parameter Estimation, Correlations, and Error Bars Segev BenZvi Department of Physics and Astronomy University of Rochester Table of Contents 1 Review of Last Class Best Estimates and Reliability

More information

6. Advanced Numerical Methods. Monte Carlo Methods

6. Advanced Numerical Methods. Monte Carlo Methods 6. Advanced Numerical Methods Part 1: Part : Monte Carlo Methods Fourier Methods Part 1: Monte Carlo Methods 1. Uniform random numbers Generating uniform random numbers, drawn from the pdf U[0,1], is fairly

More information

Sound & Vibration Magazine March, Fundamentals of the Discrete Fourier Transform

Sound & Vibration Magazine March, Fundamentals of the Discrete Fourier Transform Fundamentals of the Discrete Fourier Transform Mark H. Richardson Hewlett Packard Corporation Santa Clara, California The Fourier transform is a mathematical procedure that was discovered by a French mathematician

More information

Detecting Gravitational Waves with a pulsar timing array

Detecting Gravitational Waves with a pulsar timing array Detecting Gravitational Waves with a pulsar timing array Key Concept in General Relativity: Mass curves space-time Picture of curved space time First observed during solar eclipse Light from stars behind

More information

Introduction to Machine Learning. Maximum Likelihood and Bayesian Inference. Lecturers: Eran Halperin, Lior Wolf

Introduction to Machine Learning. Maximum Likelihood and Bayesian Inference. Lecturers: Eran Halperin, Lior Wolf 1 Introduction to Machine Learning Maximum Likelihood and Bayesian Inference Lecturers: Eran Halperin, Lior Wolf 2014-15 We know that X ~ B(n,p), but we do not know p. We get a random sample from X, a

More information

Using training sets and SVD to separate global 21-cm signal from foreground and instrument systematics

Using training sets and SVD to separate global 21-cm signal from foreground and instrument systematics Using training sets and SVD to separate global 21-cm signal from foreground and instrument systematics KEITH TAUSCHER*, DAVID RAPETTI, JACK O. BURNS, ERIC SWITZER Aspen, CO Cosmological Signals from Cosmic

More information