sample function properties of Gaussian processes carry over di- rectly to those of SIRPs, and it is impossible to tell the difference between a Gaussian process and a SIRP on the basis of one sample function. Sample function properties of Gaussian pro- cesses have been widely studied (see, for example, [5, chapters g-131). We will consider a few special cases of interest, and we will mation theory, IEEE Trans. Inform. Theory, vol. IT-14, pp. 12-16, Jan. 1968. I31 K. Yao, A representation theorem and its applications to spherically-invariant random processes, IEEE Trans. Inform. Theory, vol. IT-19, pp. 600-608, Sept. 1973. [41 A. N. Kolmogorov, Foundations of the Theory of Probability. New York: Chelsea, 1956, pp. 27-33. [51 H. Cramer and M. R. Leadbetter, Stationary and Related Stochastic Pro- cesses. New York: Wiley, 1967. state the results in terms of SIRPs. In the sequel, the zero-mean SIRP X(t) will be assumed to be separable, stationary, second- order, and mean-square continuous. Also, we will assume that P(X(t) = 0) = 0. The next theorem follows from the work of Do- brushin [7]. [61 I. S. Gradshteyn and I. W. Ryzhik, Table oflntegrals, Series, and Products. New York: Academic Press, 1965, p. 341, Sect. 3.472-l. [71 R. L. Dobrushin, Properties of sample functions of a stationary Gaussian process, Theory of Prob. and Its Appl., vol. 5, pp. 120-122,196O. [81 Yu. K. Belyaev, Local properties of the sample functions of stationary Gaussian processes, Theory of Prob. and Its Appl., vol. 5, pp. 117-120, 1960. Theorem 4: Either 1) the sample functions of X(t) are con- 191 G. A. Hunt, Random Fourier transforms, Trans. Amer. Math. Sot., vol. 71, tinuous with probability one or 2) with probability one, the pp. 38-69,195l. sample functions of X(t) have discontinuities of the second kind 1101 N. D. Ylvisaker, The expected number of zeros of a stationary Gaussian at every point. process, Ann. Math. Statist., vol. 36, pp. 1043-1046,1965. 1111 B. Picinbono, Spherically invariant and compound Gaussian stochastic The following corollary is a result of the further work of Be- processes, IEEE Trans. Inform. Theory, vol. IT-16, pp. 77-79, Jan. 1970. 1121 B. Picinbono and G. Vezzosi, Detection dun signal certain dan un bruit non lyaev [8]. stationaire et non Gaussien, Ann. des Telecommunications. vol. 25, pp. __ Corollary 1: If the sample functions of X(t) are not almost 433-439,197o. surely continuous, then they are almost surely unbounded in any 1131 G. Vezzosi and B. Picinbono, Detection dun signal certain dans un bruit interval. spheriquement invariant, structure et caracteriques des receptions, Ann. The next theorem follows from the work of Hunt [9]. des Telecommunications, vol. 27, pp. 95-110,1972. Theorem 5: Let S(w) denote the spectral distribution function of x(t). If S o [log (I+ w,] dS(w) < m On the Entropy of Continuous Probability Distributions for some a > 1, then X(t) has continuous sample functions with probability one. Also, if AIDA C. G. VERDUGO LAZO AND PUSHPA N. RATHIE S 0 m d [log (1 + ,)]a dS(w) < co Abstract-A table is given of differential entropies for various continuous probability distributions. The formulas, some of which for some a > 1, then X(t) has continuous sample function de- are new, are of use in the calculation of rate-distortion functions rivatives with probability one. and in some statistical applications. The expected rate of zero crossings of a Gaussian random process has been studied by many investigators. The following theorem is based on the work of Ylvisaker [lo]. Shannons differential entropy of an absolutely continuous random variable X with probability density function f(x) is de- Theorem 6: Let R(T) denote the autocorrelation function of fined by X(t). The expected number of zero crossings of X(t) in an in- terval of length T is given by h(X) = -s-f(x) logf(x) dx. -m T -R(O) 112 - All logarithms are taken to the base e, and-the entropy h(X) is a [ 1 R(O) thus measured in nats thoughout this paper. if R (7) has a finite second derivative at the origin. If R (7) does not The proplem of finding which probability density function, have a finite second derivative at the origin, then the expected among those satisfying certain conditions, has maximum entropy number of zero crossings in any interval is infinite. has been considered extensively in the literature. The solutions of some basic maximum entropy problems can be found in Kagan, III. CONCLUSION Linnik, and Lao [8] and in Lisman and van Zuylen [9]. These are all univariate density results; some results for multivariate In this correspondence, we have established the useful repre- densities can be found in the paper of Gokhale [4]. sentation for SIRPs given in Theorem 3. This theorem was then In the course of the analysis of the maximum entropy problem, employed to establish several properties of SIRPs. use is being made of the following obvious fact. If Consider a random process of the form AY(t), where Y(t) is a zero-mean Gaussian process and A is an independent random f(x) = exp (UO + a&l(x) + - -- + anfnb)l (2) variable. It is easy to show that such a random process is spheri- then cally invariant. This approach was explored by Picinbono and Vezzosi [ll]-[13] (see [3, p. 6041 for comments regarding [ll]). h(X) = -(a0 + alEhI +a.. + u,Eh,(X)} (3) However, the essential result of Theorem 3 is that any SIRP can always be represented by an equivalent random process having where this form. This result illustrates the essential role played by Gaussian processes in the representation of SIRPs. Ehi(X) 0 J+- hi(x)f(x) dx. -cm REFERENCES Manuscript received March 29,1976, revised May 3,1977. This work was sup- ported by FAPESP, under Grant 76/1111 and CNP,, under Grant 2222.0042/77. [l] A. M. Vershik, Some characteristic properties of Gaussian stochastic pro- The authors are with the Departamento de Estatistica, Institute de Matemltica, cesses, Theory of Prob. and Its Appl., vol. 9, pp. 353-356,1964. Estat;stica e Ciencias da Computacao, Universidade Estadual de Campinas, [z] I. F. Blake and J. B. Thomas, On a class of processes arising in linear esti- Campinas, SBo Paulo, Brasil. 00189448/78/0100-0120$00.75 0 1978 IEEE 120 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. IT-24, NO. 1, JANUARY 1978 ONDENCE 121 TABLE I TABLE OFDIFFERENTIALENTROPIES t 7 DISTRIBUTIONS E"tP.?py Name Density Function (in nats) f(x) = xp-I(1 - x)q-l/B(P.q) i 0 5 x 5 1 log E3(p,q) - (p-l) IIJI(P)-*(P+q)l Beta Where B(p,q) = r(p)r(q)/r(P+q) : P. 4 ' 0 - (q-l) [$ (q)-a(p+q)] Cauchy f(X) = (i/n) (h2+x2)-l i -m < x < m, i > 0 log(4nh) Chi f(x) ={2/[2"'2 0" r(n/2)]Ix"-l e-x2'(202);x>0 log[or(n/2)/JZ] _ y $(n/2) + n/2 and n is a positive integer Chi-square f(X) = .(42)-l .-x/(202),[2n/2 ?r(n/Z)]; x,0 loq[202 r(n/2)] + (l-n/2) yw2) + n/2 and n is a positive integer Erlang f(x) = [p/(*-l,!] 2-l e-BX ; x, 6 > 0 (1 - n) $'(n) + loqCr(n,L3] + n and n is a positive integer Exponential f(X) = o-1 e-x/a : x, '3 > 0 .J /2 " z/2 F f(X) = "I ' 2 ,(v,/2)-1 log[(",/V*,B(V,/2.V2/2~l+(l-~~/2)*~~,/2) B(V,/2.Vz/2) (" +v x) ("1+"2)'2 ;x>o --==-A and ", , "2 are positi& integer -(l+"~/2)$(v,/2)+[(v,+v*~/~~[~vlfvz)/2] Gamma f(x) = xa-l e-X'@/[fi"r(a)] i x, a, B ) 0 log@r(u)] + (1-a) $(a) + a Laplace f(x) = (l/2) q-1 .-lx-ei'!J ; -m < x < m, @.o 1 + log(2 v ) Logistic f(x) = e-x (1 + e-X)-2 ; -m < x < - 2 Lognormal f(x) = [,xsrzll)] -1 e-uogx-m)2/(202) : x > 0 m + (l/2) log(2neo2) I Maxwell-Bolt7.ma ftx) = [4n-1/283/2] x2 e-flX2 i x. 6 > 0 (l/2) log (n/B) + y - l/2 / Normal f(x) = Lomi,,,-l e-x2/(2o2) -m < x < ml a>0 (1/2)loq(2neo2) I Gsnerallzed-Noml,f(x) = [2&2/r(a/2,] xa-l e-6x2: x,a, 6>0 log~(u/2)/(261'2 )] - [(a-1)/2]$W2) +a/2 Pare to f(x) = a ka/xa+' ; x z k > 0, a > 0 loq(k/a) + 1 +1/a Rayleigh f(x) = (x/b') e-x2'(2b2' ; x, b > 0 1 + log(B/J2) + y/2 f(x) = (1 + x2/")-("Cl)'2 ;--m<X<m Student-t J(v) 8(1/2,v/2) [(v + 1)/2]I$[(V + 1)/2] - $(V/2)1 and v is a positive integer + log [J(V)B(1/2,V/2)] Triangular f(x) = 2x/a ; ojxca = 2(1 - X)/(1 - a) ;- l/2 - log2 a<x<1 Uniform f(X) = l/(6 - a) : a~x<6 log(B - cl) Weibull f(x) = (c/a) x c-1 e -xc/a i x, c, a ) 0 cc - l)y/c + lclg(P/c) + 1 122 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. IT-24, NO. 1, JANUARY 1978 This reduces the problem of calculating h(X) to that of calcu- lating Ehi (X). Shannon [ll] was the first to compute the entropy of the uni- form, the exponential and the normal distributions. Golomb [5] computed the entropy for the Pareto distribution; the entropy for the squared Cauchy distribution appeared in Bergers book [l]. For the Laplace, logistic, lognormal, and Weibull distribu- tions, the Eh;(X)s can be found in the books of Johnson and Kotz [6], [7] expressed in terms of elementary functions. We add to this list the following distributions: Cauchy, Maxwell- Boltzmann, Rayleigh, Triangular. For each of the remaining distributions listed in this paper (beta, chi, &i-square, F, gamma, generalized-normal, student-t), at least one of the terms Ehi(X) is not expressible in terms of elementary functions. The main object of this correspondence is to point out that a single special function, the logarithmic de- rivative # of the gamma function, may be used to specify h(X) for all these distributions. The definitions of the #-function and of Eulers constant y (to enough precision for numerical calculations) are as follows: Hz) = $log I%) = -y + (2 - l)$(k + l)(z + iz)]-1 (4) y = 0.5772156649 -. - . (5) For more details on the #-function, see [lo]. The entropy h(X), given by (3),can be easily obtained by using a) b) 4 4 [2, p. 314, (S)] and [2, p. 316, (22)] for the beta distribu- tion, [3, p. 233, (8)], [2, p. 314, (8)], and [2,316, (23)] for the F- distribution, [2, p. 315, (9)] and [2, p. 312, (l)] for the gamma distribu- tion, [2, p. 312, (l)], [2, p. 315, (9)], and [2, p. 133, (3)] for the generalized normal distribution. The results are summarized in Table I. ACKNOWLEDGMENT The authors would like to thank Prof. T. Berger and the ref- erees for their helpful comments and suggestions. Ill 121 [31 141 [51 [3 [71 is1 PI IlO1 Iill REFERENCES T. Berger, Rate Distortion Theory. Englewood Cliffs, NJ, Prentice-Hall, 1971. A. ErdBlyi et al., Tables of Integral Transforms, vol. I. New York: McGraw-Hill, 1954. A. Erdblyi et al., Tables of Integral Transforms, vol. II. New York: McGraw-Hill. 1954. D. V. Gokhale, Maximum entropy characterization of some distributions, in Statistical Distributions in Scientific Work, vol. 3, Patil, Katz, and Ord, Eds. Boston, MA: Reidel, 1975, pp. 299-304. S. W. Golomb, The information generating function of a probability distri- bution, IEEE Trans. Inform. Theory, vol. IT-12, pp. 75-77, Jan. 1966. N. L. Johnson and S. Katz, Distributions in Statistics: Continuous Uniuariate Distributions--I. New York: Wiley, 1970. N. L. Johnson and S. Katz, Distributions in Statistics: Continuous Uniuariate Distributions-2. New York: Wiley, 1970. A. M. Kagan, Ju. V. Linnik, and C. R. Rao, Characterization Problems in Mathematical Statistics. New York: Wilev. 1973. J. H. C. Lisman and M. C. A. van Zuylen, %Jote on the generation of most probable frequency distributions, Statistica Neerlandica, vol. 26, pp. 19-23, 1972. W. Magnus, F. Oberhettinger, and R. P. Soni, Formulas and Theorems for the Special Functions of Mathematical Physics. New York: Springer-Verlag, 1966. C. E. Shannon, A mathematical theory of comunication (concluded), Bell Syst. Tech. J., vol. 27, pp. 629-631,1948. Advent of Nonregularity in Photon-Pulse Delay Estimation ISRAELBAR-DAVID, MEMBER,IEEE,AND MOSHELEVY Abstract-The mean-square error of the delay estimate of a photon pulse is known to decrease as Q- if the pulse envelope is smooth, but as Qe2 if it has sharp edges, thereby belonging to nonregular estimation cases (Q denotes the expected photon count). The transition from the Q-l to the Qm2 law is investigated for trapezoidal pulse models, and is found to occur in the region where the standard deviation of the error is on the order of the width of the pulse slopes. Thus for values of Q below this region, practical pulses are effectively rectangular, and their estimation problem may be qualified as nonregular. This note relates to one facet of the problem of estimation of photon-pulse delay with direct-detection receivers, which has been exposed in a previous contribution [l] to which we refer the reader for appropriate background. It has been shown that, asymptotically with the expected photon count Q in the pulse, the mean-square error (mse) of the delay estimate decreases only as Q-l with pulses that have smooth envelopes, i.e., regular estimation cases [2], but as Q-z with pulses that have sharp edges. The latter envelopes pertain to nonregular estimation problems and, apparently, the nonregularity changes the dependence from the Q-l to the Qe2 law. As many practical pulse envelopes are more readily associated with a rectangular model than with a smooth one, the way in which the signal shape affects the tran- sition between these laws has been suggested for further inves- tigation [l, item (3) of Section II-C]. Reformulated, the question is how inclined need a slope be in order that the pulse may qualify as a rectangular one? Herein, we analyze the effect by modeling real-life pulse en- velopes by trapezoids of varying slope as illustrated in Fig. 1. Let the pulse envelope X(t) be given by -(l + a)D I t I -D -DItID [(l + a)D - t]X, D 5 t I (1 + a)D a-D 0, otherwise. (1) Then the expected photon count in the pulse is Q = I:::&:, X(t) dt = (a + 2)0X,, the received envelope is X(t - T), and the observables are the instants tk, k = 1,2, . . - , L, of the photoelectron emissions at the direct detection receivers output. The minimum m.s.e. estimate of the delay r for such an envelope is not expressible explicitly in terms of tk, as it fortunately is for rectangular ones [I]. We therefore adopt here the estimate which is optimum for the rectangular envelope (with (Y = 0 in (1)) and is given [ 11, for L I Manuscript received March 2.1977; revised May 6.1977. This correspondence incorporates results from a Master of Science theses submitted by M. Levy to the Senate of the Technion-Israel Institute of Technology, Haifa, Israel, in October 1976. The authors are with the Faculty of Electrical Engineering, Technion-Israel Institute of Technology, Haifa, Israel. 0018-9448/78/0100-0122$00.75 0 1978 IEEE