U.S. AIR FORCE PROJECT RAND RESEARCH MEMORANDUM A STATISTICAL THEORY OF TARGET DETECTION BY PULSED RADAR: MATHEMATICAL APPENDIX J. I. Mareun RM-753, ASTIA Document Number AD 101882 2 July 19h8 Assigned 10 This is @ working paper. It may be expanded, modi 5, and recom lect the officialCONTENTS ‘Sonmary List of Symbols... Basic Formulae Relating to Thermal Noise. . Definition and Effect of Detection. Effect of Video Amplifiers Probability of Detection With No Integration. General Case— Integration of N Independent Variates.. ‘The Square Law Detector With N Variates...... Bias Level for Square Lew Ca: Cumulative Distribution for N Variates of Signal Plus Noise~Square Law Detector.. edad +13 Expansion of Functions in Gram-Charlier Series... pond Moments of Signal Plus Noise, Square Law Detector.. nom) ‘The Gram-Charlier Series for the Square Law Ce: Sample Calculation. Integration Loss, Square Law Detector. General Curves of the Cumulative Distribution Function...... ‘The Linear Detector—N Veriati Results of the Linear Detector Calculation: Expansions in Laguerre Series Other Series Approximations. Nethods of Integration Involving Subtraction of Noise. Distribution Functions for Composite Pulses of Signal-Plus-Noise Minus Noise. oa : w39 Probability Density Functions for ¥ < 0. 243 Probability Density Functions for ¥ > 0.. MhCumulative Distribution Functions M6 Another Approach to the Detection Criteria-Probability That Signal-Plus-Noise Exceeds Noise Alone --50 Use of Cumulants in Obtaining Gram-Charlier Series Coefficient Best Possible Detector Law st 58 Signal-Plus-Noise Minus Noise— Linear Law... Use of So-Called Detection Criteria. Collapsing Loss— Integration of Greater Number of Noise Varia Than of Signal-Plus-Noise Variat: Antenna Beam Shape Loss. Limiting Los: Effect of Signal Injection on Probability of Detection... Probability of Detection With Moving Target Indication Systems.. Tables of the Derivatives of the Error Function. References REFERENCE FOOTNOTES On Q Functions ea On Bessel Functions 13 On Hypergeonetric Functions. +201. n. 12 13 ua. 4s. 16. we 18. we. LIST OF FIGURES Probability Density Functions for Envelope of Sine Wave Plus Noise, Square Law Detector, N = 1. Probability Density Functions for Envelope of Sine Wave Plus Noise, Square Law Detector, N = 3.. ocoaat Probability Density Functions for Envelope of Sine Wave Plus Noise, Square Law Detector, N= 5. ceetees eee eee Probability Density Functions for Envelope of Sine Wave Plus Noise, Square Law Detector, N= 9...... Probability Density Functions for Envelope of Sine Wave Plus Noise, Square Law Detector, x= 0.... ees - Probability Density Functions for Envelope of Sine Wave Plus Noise, Square Law Detector, x * 1. Probability Density Functions for Envelope of Sine Wave Plus Noise, Square Law Detector, x* 2 Bias Level as a Function of Number of Pulses Integrated and False Alarm Number, Square Lew Detector..... Bias Level as a Function of Number of Pulses Integrated and n', Square Law Detector. Integration Loss, Non-Coherent vs. Coherent Integration Loss, Non-Coherent vs. Coherent. panos. Probability of Detection With No Integration... Tae Incomplete Toronte Fanction teQNel N-1. a), N= de The Inconplate Toronto Function TgQNel, N=1. va), Ned The Incomplete Toronto Function Tg (QN1, N=1, va), N= 3 The Incomplete Toronto Function eee a een ‘The Incomplete Toronto Function Ty, (2N-1, N~1, 79), N= 30 The Incomplete Toronto Function 1y,(2N-1, N~1, Yq), N = 100 The Incomplete Toronto Function Te(tN— 1, N=1, 4), N= 300 = 65 66 67 ..67 68 69 10 a1 n 213 14 18 16 7 8 1920. 21. 22. 23 24. 28. 26. 21. 28. 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. 39. The Incomplete Toronto Fanction Tye(QW— 1, N=1, V4), = 1000... : a0 The Incomplete Toronto Function Tyq (QN-1, N=1, 7G), N= 3000 at The Incomplete Toronto Function Ty (2N~1, N~1, VG), 9= 0... 82 The Inconplete Toronto Function Tye QN-L, NAM, VQ)s T=N/D cee ceeeceeeeeeeeeeneene coven 83 The Inconplete Toronte Function Tz (2N-1, N-1, Ya), a=. ‘The Incomplete Toronto Function Tq (QN—1, N21, G), N= 1 The Incomplete Toronto Function Ty (QN-1, WH, 49), N= 3.. The Incomplete Toronto Function Ty, (N-1, N12, 4G), N= 10... ‘The Incomplete Toronto Function Ty; QN=1, Nm2, Va), N= 30... 7 oo 88 ‘The Incomplete Toronto Function Tq (2N~1, N-1, 7G), N= 100. 88 ‘The Incomplete Toronto Function Tq (2N—1, N=1, ¥@), N= 300 ‘The Incomplete Toronto Function Tyz (QN-1, N~1, va), = 1000. The Incomplete Toronto Function Ty (2N-1, N1, vq), N= 3000.. Characteristic Function for Envelope of Noise.. Moments for Signal Plus Noise, Linear Detector, N= les.s.eeee Bias Level as a Function of Number of Pulses Integrati and False Alarm Nunber, Linear Detector... Probability Density Functions for Envelope Plus Noise, Linear Detector, N = 1. Probability Density Functions for Envelope of Sine Wave Plus Noise, Linear Detector, N = 2 Probability Density Functions for Envelope of Sine Wave Plus Noise, Linear Detector, x = 0....- Probability Density Functions for Envelope of Sine Plus Noise, Linear Detector, x = 0..40. Probability Density Functions for Envelope of Sine Wave Plus Noise, Linear Detector, x = 0.5. oo --98 41. Probability Density Functions for Envelope of Sine Wave Plus Noise, Linear Detector, x = 1. oe 42. Comparison of Linear and Square Law Detectors, n = 10%. 2.99) 43. Comparison of Linear and Square Law Detectors, n= 108, P= 0.50. 44. First Approximations to the Probability Density for Signal Plus Noise, N= 3, x= 1, Square Law Detector... 100 45. Bias Level as a Function of Number of Pulses Integrated and False Alarm Number, Square Law Detector, Signal- Plus-Noise Minus Noise..... -101 46. Probability Density Functions for a Signal-Plus-Noise Variate Minus a Noise Variate, Square Law Detector ++ +102 47. Comparison of Probability of Detection When Noise Variate is Subtracted from the Signal-Plus-Noise Variate, Square Law Detector..... +103 48. Probability That Noise Exceeds Signal-Plus-Noise, Square Law Detector. -104, 49. Collapsing Loss for Constant n, P= 0.50, N= 1-30, n= 108 105 50. Collapsing Lo: P= 0.50, V= for Constant n, 100-1000, n = 10%... peasod LH 51. Collapsing Loss for Constant P = 0.80, N= 1~1000, n= = 106 52. Collapsing Loss for Constant P= 0.90, N= 1~1000, n= =++106 53. Collapsing Loss for Constant n, P= 0.50, n= 10%...... o --107 54. Collapsing Loss for Constant n, P= 0.50, n = 10 -107 55. Collapsing Loss for Constant n‘, Piet Os fwla: tose -108 56. Collapsing Loss for Constant n', P= 0,50, n= 103°... +108 57. Collapsing Loss for Constant n', P= 0.50, » = 108 o6 sees 109 58. Collapsing Loss for Constant n, Dependence on P.. -1ogSUMMARY In a previous report(#*) a statistical theory of radar detection was presented in outline form. The mathematical details were omitted, in order that the main ideas and results might be made available as soon as possible. ‘This report contains the mathematics that led to the results presented in Ref.28, In addition, several subjects are briefly discussed that were not covered in Ref.28. These are collapsing loss, antenna bean shape loss, the effect of signal injection, limiting loss, and moving target indication. For references sce page 111.Soe as eee e > SYMBOLS anplitude of sine wave relative to R.N.S. noise level one of the independent variables in the function Q(@,A) 2 central standard moment one of the independent variables in the function Qa, §) hal f-power antenna beamvideh coefficient in the Gran-Charlier series characteristic function delta function base of natural logerithns frequency confluent hypergeometric function Campbell and Foster notation for characteristic function Canpbell and Foster notation for anticharacteristic function the ganma function probability that the sim of N noise variates will exceed the bias level 4° Hermite polynomial index, subscript, or V=T incomplete gama function as defined by Pearson‘®? nodified Bessel function of the first kind Bessel function of the first kind 1° cumlant standard it cumulant, or sometimes a modified Bessel function of the second kind integration loss collapsing Loss generalized Laguerre polynonial number of excess noise variates integrated with W signal plus noise variates false alarm nunber nw number of variates integrated 1¢ meaning in RA-15061, cone sense in various places, but other meanings shouldiw © wT sine wave anplitude probability probability that noise will exceed the bies level at least once within false alarm tine probability that the sum of N variates of signal plus noise will exceed the bias level i derivative of the error function RM.S. noise Level nodi fied Lonmel's function envelope amplitude or radar range in R/Ry idealized radar range collapsing ratio, ratio of total nunber of variates integrated to those containing signal cathode ray writing speed standard deviation (y-¥)/o, semi-independent variable in Gran-Charlier series incomplete Toronto function it moment about the mean Lonmel's function normalized envelope amplitude £8 moment. power spectrum anf power signal-to-noise ratio normalized detector output integrator output for the sum of N variates bias level *This symbol hs ‘This symbol is used in aore than one sense in various places, but other sean different seaming in RA-15061 be obvions:A STATISTICAL THEORY OF TARGET DETECTION BY PULSED RADAR: MATHEMATICAL APPENDIX BASIC FORMULAE RELATING TO THERMAL NOISE Both the thermal noise voltage across a resistor and the noise voltage due to the shot effect in a vacuum tube approach a normal distribution when the number of electrons involved per second in the processes tends toward infinity. In prac- tice, it may usually be assuned that the total noise voltage between any two points due to any combination of thermal, shot, and cosmic noise sources can be represented by the distribution function a where Jj, is the mean aquare value of the noise voltage"). This distribution is Valid provided all elements involved in the composition of the total noise voltage hare been linear. If such noise is now passed through a linear filter of center frequency f,, having pass band which is narrow compared to f,, the output will have an enve- lope, which has a probability density function fe an 2where R is the amplitude of the envelope and Y, is the mean square noise voltage, given by ad fw @ u(f) is the so-called power spectrum of the filter and is simply the square of the absolute value of the amplitude transfer function of the filter. If the input to a filter consists of a sine wave of frequency f,, as well as noise, then the probability density function of the output envelope is* Ro @ dP =0 Reo where P is the amplitude that the sine wave would have at the output of the filter in the absence of noise, and I, is a modified Bessel function of the first kind (see footnote, page 13). The envelope of the output has a correlation tine which is approximately equal to the reciprocal of the bandwidth of the filter. In simple language, it is im- probable that the envelope will change by an appreciable percentage in times mch less than the correlation tine, but it is quite probable chat it will change by a goodly percentage in times Large compared with the correlation time. It is probably 4 good approximation to assume that values of the envelope 1/Af seconds apart are independent, where Af is the bandwidth of the filter. By assuning such a discrete process it is possible to materially simplify calculations which would be very tedious if exact integration processes were used, while at the sane time sufficient accuracy is obtained for most practical purposes. A further justification for this assumption in the pulsed case shows in the results. Changing the factor 1/Af to k/Af for the correlation time has only the effect of changing the false alarm tine by the factor &. The probability of detec- tion turns out to be a very insensitive function of the false alarn time, 20 that if k is any factor of the order of magnitude of unity, the results are affected to a negligible extent. + 10 is of some interest to note that the sane form of distribution function occurs probless, For instance, if yy tepresenta the wean square telocity of The sane density function alee represents the probability that « bomb will hit at « distance between Mand A + dit fron a given point if itis initially aimed at « point whose disterce late aiming error is represented by yg, the distris from the given point is The iteton felng asnuned Grasaien 7DEFINITION AND EFFECT OF DETECTION A detector is defined as any device whose instantaneous output is a function of the envelope of the input wave only. Thus R fe tz) = Fw) oy wherey is the output of the detector and » is the normalized amplitude of the envelope. If Py is replaced by ©, Eq. (4) may be written dP = ve? Tav)dv, v0 (6) dP =0 veo Bq.(5) solved for vis v = gy) - mM If v is eliminated from (6) and (7), an equation of the form a = flayidy (a) is obtained for the probability density for the normalized voltage at the output of the detector which has the characteristics given by Eq.(5). For example, if y 24/2, then Eq. (8) becones dP = e%* I, (2vzy)dy, yO @) 20 <0 where a¥/2 has been replaced by x. The quantity * may be signal-to-noise ratio, comonly used in radar literature. identified with the por EFFECT OF VIDEO AMPLIFIERS Since a complete radio receiver usually has one or more stages of video ampli- fication following the detector, it would seem that one would want to calculate the probability density function for signal-plus-noise at the output of the video 3amplifier. This can be done theoretically, as has been shown in an excellent paper by Kac(*!, but the mathematical labor is great. On the other hand, it has been shown experimentally(®*) that the signal threshold is practically unaffected by the video bandwidth until it becomes less than about } of the IF bandwidth. Since video bandvidths less than { cf the IF bandwidth are quite uncommon in practice, it appears best for the sake of simplicity to assune the video bandwidth infinite in all the work which follows. When the results have been conputed, assuming an infinite bandwidth, it will be possible to modify them in an approximate manner so that they becone valid for any video bandwidth. This is explained on page 59, under the title "CollepsingLoss", PROBABILITY OF DETECTION WITH NO INTEGRATION ‘The calculations necestary to determine the probability of detection when exactly ore correlation interval is available are quite simple conpared with the case where the output over many correlation intervals is available, and hence the former cass is taken wp firat. In a pulsed aysten this corresponds to using a single pulse, while in a c-w aystem it is equivalent to observing the output for a tine t= 1/Af, where Mf is the over-all effective bandwidth. In either case this anounts to observing the receiver output for one correlation interval. If the output exceeds the bias level, the signal is observed or detected (see pages 9-14 of RAA50E1, A Statistical Theory of Target Detection by Pulsed Radar (#"), hereafter referred to ts No.1, for coaplete definitions of detection and bias level). It will now be shown that the probability of detecting a given signal = is independent of the detector function, everything else being held constant and only one variate being taken from the density function of Eq.(8). The false alarm tine has been defined as the time in which the probability is % that the noise alone will not exceed the bias level (Eq. (15), No.1), but it will be best here to keep things general and denote this probability as P,, rather than as 4. Eq.(15), Nol, then becomes Pe inp at ao) here the subscript NV denotes the number of variates and [is simply an abbrevia- tion. Fron Eq. (8), re Re f £(0,y) dy ap % there the symbol y, is now used for the bias number. Then the probability of detec- re ne f Hayy) dy aay nbut since y= F(v), or v = g(y), Bq.(11) may be written 2 £0) qe f ve Fay ee as) e074) wd T shy) V2 os, - as Therefore Eq. (12) becomes BR ve? I,(av) dv us) — which is independent of the detector function. ‘The integral of Eq. (15) must be evaluated by approximate methods. This function will appear in several places subsequently, and is defined as* aA) = if ve Blav)dy as) ls Fe + Te docs not appear possible to express the Q function in terms of a finite aumber of Inown functions, The Q function in similar to Lomel’s functions and in fact can be ex- Pressed 3 2 on @ Functions ota ea 6 [w, (0% op) -u,(-i5%ica)] 12 of the Sirse kind. This identity may be proven nels functi ‘540 and S41, especially Eq.5 of in Watson), shee du ead there b Sethe ee eae et eee er ecard ES) p00 eS Cae sre rity of the first of these expansions to the series for Uy (w+) given in Eg. (1), A simple for Q(@,a} analogous to Eqs. (9) ct] tcnied on pe 3In terms of this notation, Eq.(15) may be written P= QavT TRI) an ‘This is the probability of detection only if % = 1/f, where 1y is the tine available for detection. In general the probability of detection is given by Eq.(18), No.l, RB By = 1 (LP) cas) which follows from the definition of detection given on page 9, No.1. The double subscript notation Py, is used here in place of P'. If N= 7, Py, is written simply as Py. P,,, (R/R,) can be calevlated by means of Eqs.(18), (17), the tables of Q, and the single relation as) al (see Eqs.(10), (11) and (12), No.1). Footnote on Q Functions (Cont'd) which is aseful in special cases. An nsion for Q which is of value is Biven by Rice), page 109 “ fa, woras? np) = Ha steal] + tae eae) = 0 - etal] + oh were 91/7) ia given by the error function of Eq.(100). This expre: Dag bt Mae teal ° ‘The Q function ig a special case of the incomplete Toronto function described in the Footnote en page 28. The relation is as) 1 Ty ‘The Q function is graphed in Figs.13 and 14. A table in available in Ref.47 but the intervals are too large to be of general use. Proje hand “e conpuing an extensive table of the Q function which will be published as « sep 6A very good approximation for the quantity [j, which appears in Eqs.(10) and (11), may be derived from Eq. (10) by writing pi gl ay a vee ani which is valid when n’ >>1, a condition nearly always true in practice. Eq.(10) then becomes Tye © tog, F ay T£B, = 1/2, as is essumed in all the curves in Nou1, 7 (22) Bq-(11) may consequently be written BYP 2 = oS) vow toe, ee o780| - (ay ‘As an example, let R/R, = 0.595, and n= 10%. Then P, = Q(4, 4.37), which has the value 0.410 from the table given in Ref.47. Note that this is a point on the graph of Fig.l, page 22, No.1. GENERAL CASE — INTEGRATION OF N INDEPENDENT VARIATES If the output of the receiver (or filter) can be observed for a length of time much greater than one correlation period, it is of advantage to integrate the out- put, The simplest concept of an integrator is a device which linearly adds the voltage output of N samples from the detector. The tine elapsing between samplings must be at least one correlation period, in order that the samples may be considered to be independent. If the sum of N variates* of signal-plus-noise exceeds the bia level calculated from the probability density function for N variates of noise alone, then the signal is said to be detected. 3 Sends bee : srethich of the tre felons feof ar det py be the probability & 2th tree alee, lot be clatical decision metho ou thee for any ether srthed with pearmaisio ing_ay. Suppose py SeOaEeRcd fon abich sobs td fer sich TTT, Sone for daaeple, Rendell, vol a‘The integrator may take the sum of the squares of the N variates, or, in general, the sum of N variates where each variate has been processed by some general func- tion, As long as the sane weight is applied to each variate, the integrator will be called linear. The function hich the integrator applies to each variate will be called the lav of the integrator. Any nonlinear integrator will be inferior in opera- tion toa Linear integrator with the same law and would ordinarily never be used intentionally in practice. Cathode ray tubes are nonlinear, however, and thus fall short of other types of linear integrations. ‘The lov of the integrator acts in exactly the same way as the law of the devec- tor. Thus, if the detector output is y = (F(v) as given by Eq.(5), the integrator ourput is Y= Se = Pete). en It is obvious, as far as the theoretical problem is concerned, that the only function of importance is Wo) © @lFO)]. (2s) ‘There will be an infinite number of combinations of and F functions which will produce the same function W and hence the sane theoretical results. In all the work that follows, the output of the combination of integrator and detector for one independent variate will be called y = Y(v), and the sum of N variates will be (26) me ‘The symbolic solution for the case of N variates corresponding to Eq.(15) for one variate is not too difficult to obtain. The starting point is Eq.(8) for the proba~ bility density function for one variate. The characteristic function for this distribution is 6 [liane ra ‘The characteristic function for the probability density function for the sum of NV independest variates is then simply = (G)" (a)and apy = af Cl a,ererton (29) 4B, = Gla,N,¥)d¥ . (30) Corresponding to Eq.(11) is ne [conn an and to Eq. (12), Be f GaN Yia¥ G2) % If ¥, is elininaved from Eqs.(31) and (32), there results 4 solution for Ry as function of Ty, My and a, which is the desired result. It is found in most cases that the integrations required in Eqs.(27) to (32) are not possible in terme of known functions. THE SQUARE LAW DETECTOR WITH N VARIATES It seems, by a process of trial and error, that the best possible function for ¥(v) in Eq. (25) to produce integrable functions in Eqs.(27) to (32) is Wr) = Avy (33) ‘Though this represents a square law for the combined detector and integrator law, it is usual to think of it as representing a square law detector coupled with a linear Low integrator. 9In Eq. (33), the only effect of the constant A is to multiply the bias level Y, in Egs.(31) and (32) by A. The value of Py in Eq.(32) is independent of A. Tt is convenient to let A = 1/2, or ¥ = 02/2. By direct substitution from Eys.(6) and (27), 3 -f ert, (vey) EP dy an where * * 9%/2and p = ia, ‘This integral may be obtained from pair 655.1 of Campbell and Foster(?). In all pairs taken from Canpbell and Foster it is necessary to replace p by ~p, since they use €P8 for the first integration. As long as the same notation is in bota directions, the order of signs is immaterial. In order to avoid confusion, the mins sign will be used in the exponent in the first transformation and the plus sign in the second transformation. Thus all of the characteristic functions which appear hereafter are really C(-p) rather than C(p). In this way there is direct agreement with the Campbell and Foster tables as well as with tables of the Laplace tr: form. Equation (34) becomes Eee G, ert ws) oh 1 pele ‘The characteristic function for the sum of N variates is then simply (35) Gay of pair 650.0, Campbell and Foster, the probability density function is 1, (Vay ¥>0 (a y 50 or n’ > 10°. The normal approximation to Eq.(39) is unsatisfactory for N less than several thousand because of the fact that the integral is over a region which is far out on the tail of the curve. This can be seen from the Gran-Cherlier series which will be taken up presently. ‘The integral of Eq.(39) may be evaluated directly by successive integration by parts to give (a2) 43) ain this region so that Eq.(42) becomes 44) By the use of Sterling's approximation for NM, Eq.(44) reduces to %) ny B <) : (4s) au Though the expression looks more cunbersane in this form, itis actually much simpler to use in calculations than is Bq.(44). Substituting for [y from Eq.(21) gives the expression Login = 0.24 + 2 logis N+ 108 ye (¥y-NOL) (46) % + 0.434 (YM) - N log, HF Graphs of this function are shown in Figs.8 and 9. For N = 1, the exact expression for } fron Eq. (39) is Y, + 2-3 log, + 0.37 (a7) whereas Eq.(46) for N = 1 reduces to = 2.3 logy n+ 0.45 (48) The difference is seen to be practically negligible. 12CUMULATIVE DISTRIBUTION FOR N VARIATES OF SIGNAL PLUS NOISE - SQUARE LAW DETECTOR Knowing the required bias level for e given false alarm interval, it is now necessary to integrate the density function of Eq.(37) from this value to infinity to give the probability of detection for a signal of strength x, thus Be [@® "orm nova ay « us ‘This integral is not soluble directly in terms of well-known functions. The order of the Bessel function can be reduced in steps of 1 by successive integrations by parts, so that the last remaining integral is of the type given by the Q function + Te folloving are sone of the use! firat kind: ne (ay ay 2 nos coryin SOO Dts mie Gan NE Rin eet Auyaptotic expansion: oO By 7 ete ¥ (ain y) 2 Ne antag YL (ain ¥)™ a ALG) # aG) 4 ely) * on) + 21.) Koy = 1,0 Styne Satna = eo PIC 2 ria = Tyagi) Sette = 20°12) - 1,00) Setigarde = xe ce) + 1,021] Seti cnds = Florin + 11,0] Setiinds © etlorn iin + x10] Relations heeween the Jy fonctions and the hypermeometric functions will be found in the foot= note on pl. 13of Eq. (16). An easier way to arrive at the same result is by the use of the charac- teristic function, To get the cumulative distribution from -® to Y of any density function, it is only necessary to find the anticharacteristic function of C/p, where C is the characteristic function of the given density function (see pair 210, Campbell and Foster). Thus from Bq.(36), 5 (50) The term 1/p(pt1)" may be expanded in series i Popeid® ” pip) (pany? ” (end? ~ Gaye a ‘The mate of the first termof the series, by pairs 210, and 655.1, Campbell and Foster, is % om [enema . (52) The first two terns of Bare thus % “f 71, 27) dy (33) , Beane Sn Py, (w/e) av = QUAN: v2%,) using the definition of Q from Eq.(16). All the succeeding terms may be obtained by using pair 650.0 Campbell and Foster. jace of OP ig(L) 7 e+ VR ace of 2 ia( 2)” err, (2M) wa tr “Aw in Campbell and Foster, f is here used in place of 4/29 oF p/2ri. “4s) This form of solution for Bis practical for nuserical calculation only where N is less than about 10. ‘The characteristic function in Eq. (50) can be expanded in another manner using = moe” 2 eh (56) er ‘This leads to an expr sion for R, complementary to that of Eq.(55) of the form 1, QR) « GD By equating (55) and (57), one obtains one of the know expansions for Q given in tthe footnote on page 5. Equations ($5) and (57) may also be obtained direc:ly from Eq.(49) by repeated integration by parts. Equation (57) may also be obtained di- rectly from Eq. (55) by means of the identity - 58) given in McRobert “), page 32, and one of the known series for Q. For the special c e ¥, = Nx, the function Q of Eq. (55) is simply e- Efe o™n0%)] a (see footnote, page 5), and Eq.(55) becomes a Bo eM ean) 1,02) + (2%) e+ eal] (60D 1s‘This forma is useful for checking special points for values of N around 10 or below. None of the methods developed above are suitable for calculating P, for large values of N. In the next section the general method of Gram-Charlier series is developed, which will be useful in @ number of succeeding problems concerning distribution functions over large ranges of variation of N. EXPANSION OF FUNCTIONS IN GRAM-CHARLIER SERIES The function #(y) is defined by deze er - (sn) The Hermite polynomials may be defined by the relation (62) where the superscript is derivative with respect to y. The ¢ functions and the Hermite polynomials are biorthogonal, that is f HO) #8, = 0A (a) shies Therefore it is possible to expand any reasonable function in a series of the form‘) fords Dado + (64) The coefficients a, may be evaluated in a manner analogous to the Fourier series methods by multiplying both sides of Eq. (64) by Hi(y) and integrating from -= to ©. All terns drop out but one, giving a f(ndy (6s) 16It is usual to make the substitution t = y - ¥/o before making the expansion, thus causing the second and third terms of the series to vanish. The notation is J = y= the average value of y, or the first moment the variance vy, = the nt monent of the distribution. rfondy (66) Equation (64) is replaced by fly) = g(t) (67) and Eq. (65) by (68) Te follows at once from Eq. (68) that cy = 1/o, ¢, = ¢, * 0. The moments about the mean, or the central moments, are defined by O-Fifyay (69) and the standard moments about the mean by‘The coefficients c, in Eq. (68) may be easily written in terms of the a's, The first few are Fy (247150, 20) (a,721¢,+105a, ) (4-28a,+210a,-315) (a,-36a,+3780,-12600, ) 3 (71a) (mb) (me) (ma) (le) (7g) (1g) Formulae for the 11's in tems of the v's can be obtained directly from Eq. (69), giving ay 7 = ayy, + ae? ba =m = ayy © boy - 3 18 (728) (72) (720)ig Y= 8H F HONE! = 104! + Of (72) bg 8M = bY + SW = DOr? 4 SY Sef. (Te) Continuations of this series are obvious. ‘The process of obtaining the Gran-Charlier expansion is now evident: 1. Find che moments of the distribution. 2. Obtain the central moments from Eq. (72). 3. Obtain the standard central moments from Eq. (70). 4. Obtain the coefficients from Eq.(71). 5. Write the series for f(y) from Eq. (67). Te turns out that the best grouping for the terms of the series of Eq.(67) is different from the natural sequence(®), Such a regrouped series is terned an "Edgeworth series” and is actually used in this work. The grouping used by Edgeworth is 0 (13a) 0,3 (736) 0, 3, 4 6 (30) 046579. (3a) ‘This means that if the 0 and 3 terms are used as the first approximation, the addi- tion of terms 4 and 6 gives the next order approximation, and so forth. MOMENTS OF SIGNAL PLUS NOISE, SQUARE LAW DETECTOR ‘The moments of a distribution may be obtained by using the characteristic function ‘a moment generating function*. Thus (m4) 19In the case of the distribution function for the sum of N variates of signal plus noise with a square law detector, the characteristic function is given by Eq. (36), ‘and the movents are me A ao MeP , ye cof 4) 7 (75) ‘Though the first few moments may be obtained by direct differentiation, it is better im this case to expand in a MeLaurin's series and obtain the coefficient of pi/i!. Thus a Ne, __ Ws) (6) Gay ” Gn * Gat” Geral The coefficient of pi/i! is, by direct expansion of each term in Eq.(76), “pe! [ , Wied, | versie) (ved? CON Ger h ne ernNaa) ] Gp eal “en AF (Nei NNe) (78) (w=) where ,F, is the confluent hypergeometric function? Thus the moments are + The folloving are some of the useful relations concerning the confluent. hypergeanetric functin: ey De PP poy eas stent sy Tay so Ter) ei Asyuptotic expansion: wens Tae Bee Tae Kummer's {iret tranaformation: hs zal aylavers) OF leae, 2 Kamera second transfornation again =e (th 2) Fecursion relations a Fy (atliess) + (000) fe stare) F(a) ach (athena) + (ered ayh (act 2) = clara) F (overs) (Continued on next pore.) 20| eis! © Cen! ooh ENE Ni Ne) 9) ACE MWe) (80) Hypergeometric function (Contd) a (athens) # Cre) Caerdaad © (attred J (aves3) 2) 4 2, (avetlea) = 6 Caer) DED Blache) * Ged F (aed (rads A (aera) + Ceol A (aera) © ested) (avert) # hte © £ ktasheotn elecions between hypergtometric fmctions end other functions: dlaan ae con art etter sre Male, ot) using Pearson's notation for the inconplete gamma function A Gian ey fee Teel (, o) Albee) «AG ofan Berss ahemtan = 1c) (original Laguerre polynonial) cnet 1) «MTUae 42 at Tr aetemd (generalized Laguerre polynomial) o - Hemet, (2) car a1Eq.(80) being obtained from Eq. (79) by Kummer’s first transformation. The first four monents are = Nz) (81) vy = (Wa)? + QNx(We1) + M(N#L) (arb) vy © (x)? 6 3(a)*(N62) + 3Ne(Ne1) (N42) + N(W#1)(N42)—— (Be) uy = (Nxy4 + ANE) 9(N43). + 6 (Wa)? (N02)(NE3) (aia) + Ax (#1) (142) (N43) + N(N#1) (42) (N03) The generalized Laguerre polynomial L{°(2) is defined by Parten) (Glee aay Amarin). (92) Comparing (80) and (82), it is seen that the moments Laguerre polynomials are xpressed in terms of the ys ALLL DCN). (83) Another generating function for these polynomials is available through the relation 19 2) = 2S (erent) Gs) ‘The monen:s about the mean may be expressed in terms of the moments about the origin by means of Eqs. (T2a-e), resulting i wy td (85a) w= 0 (asb) 22by + Ne +N» M(2ze1) © 0? (85) by = Wz + 2N # QN(3xe1) (95a) bg = 12(Nx)? + 12Na(a42) + 342) (85e) ug = 120(Wa)? + 2OWe(SN46) + AN(SN*6) (958) A generating function for the central moments may be obtained by multiplying the generating function of Eq.(75) by ePtgiving (see pair 207, Campbell and Foster) Be tnsenl Sar = (eierm| dt eprint 86) wt (Die tgs (pei | (i The moments of Eqs. (85a-f) are most easily obtained by logarithmic differentiation in Eq. (86). ‘The standard moments about the mean are obtained from Eq.(70), and are asl (87a) a= 0 (87) aed (ate) ee oe (eta) WY8(aa01)? 6(4xe1) 4 (87e) 4 3 Nasa? 23‘There is an approximate method of computing the significant part of a, based on the fact that ¢, of Eq.(7la) i 259, Fry(®)), Thus which is always nearly equal to ¢2/2°(see page a, © 15a, + 10a} - 30 (88) 10 (108x4+rex+13) age is + 7 (s1!) For noise alone, the monents are given by ye Ween fn 4° GD! and the central moments by eee cy 4 Ger aan Equation (90) was obtained from Eq.(86) by putting =0 and expanding in a series. THE GRAV-CHARLIER SERIES FOR THE SQUARE LAW CASE (Ta-d) since The crefficients of the series may be obtained by use of Eq: the standard central moments are now known (Eqs.87a-f). They ar ql (a) qtr ws) a (te) avi aaert 24Axel «tits sia) ae? : (seen? oe sie) ©" Ten(axesy® ‘ Fron E4.(67) the required series is A ap Zo, 9800) «alo + gett + gates (92) where ' . (93) 4 = Noes) on o = VNC) and the c's are given by Eqs.(9la-e). Note that the grouping of terms is according to the Edgeworth scheme given in Eq.(73). Note further that as N tends to infinity, all the coefficients go to zero except ¢,- Thus (95) N=, In terms of N and x eh ° * Faas a °° 25Eq. (95) is precisely a statement of the central Limit theorem, and the derivaticn given is essentially a loose proof of the theorem. ‘The cumulative distribution is easily obtained from Eq.(92) by means of the simple relation feu = ot) (97) giving R -f Andy = FU-CHN) ~ gen - ge - & % where ke ra on and un ee fF ae (200) cod . The function d"*(T) is tabulated in the W.P.A. tables(*). This differs from the definition given for ¢"*(y) in Fry, page 456 but is used here because of the W.P.A. tables. ‘The series of Eq.(98) was used to calculate all the curves of Figs.1-50, No.1, with the exception of the cases where N= 1. In most cases the first two terms of the series are sufficient, though in some regions of small P four terns areneeded. SAMPLE CALCULATION Assume N= 10, = 10° From Fig.8, of Eq.dl, = 30.0 & Lat F, * 1.0 s0 that x= 1.0 26Fron Eq. (94), Fron Eq. (99), Fron Eq. (91e), Fron Eq. (91d), Fron Eq.(9le), y+ Ns = 20.0 = VATS = 5.48 Hoy 30.0 -20.0 re Ah. RE + 1.0 G(T) = g7*(1,828) = 0.9325 }0-¢"47) = 0.0338 yt aE, + -0.081 Bn ¥%(2241)* 4x41 6° Gnaazery? * 0-0139 = UHI)? | 2 9.0033, TaN(az61y $7(1.828) = 0.174 (See p.64 for references on tables of the derivative: (1.828) = -0.470 of the error function. $°(1.828) = 0,990 @°(T) = -0,01461 @°(T) = -0.0007 ego°(T) = 40.0032 P = 0.0338 + 0.0141 + 0.0007 - 0.0032 = 0.0452 This point, P = 0.045, R/R, = 1, may be found on Fig.20, No.1. INTEGRATION LOSS, It is of interest to expre: with respect to coherent integration loss as SQUARE LAW DETECTOR the effect of noncoherent integration as a loss This may be done by defining the integration Nay L, = 10 log, (1008) oe, 2where N number of pulses integrated x, * required value of signal-to-noise ratio to produce given probability of detection for N = 1. x, * required value of signal-to-noise ratio to produce the same probabilizy of detection for N = N. Thus L is a function of P and n. However, it turns out that the dependence on P and n is very small. In the case of coherent integration, 2, is always equal to z,/N, a0 that L, = 0, With noncoherent integration, x, is alvays greater than x,/N, so that noncoherent integration is never as efficient as coherent integration. ‘The results of calcula- tions are given in Figs.10 and 11. One observes that the dependence of L,; on P and nis quite small. Thus by means of the graph of Fig.12, which gives x as a function of P and n for N= 1, and any one of the curves of Fig.10, it is possible to obtain fa fairly accurate value of = for any P, mand N. GENERAL CURVES OF THE CUMULATIVE DISTRIBUTION FUNCTION ‘The integral of Eq.(49) is @ function found in other applications than the one discussed in this paper. It is desirable to have graphs of this function available in general form rather than the specialized form of Figs.1-50, No.1. The integral is a special case of the incomplete Toronto function* described by Heatley **) and Fisher (7), which is defined as Tylasnyr) = 2reett emeOLartde . (1006) Using this notation, Eq. (49) for the cumulative distribution function becomes SL Mg (201 WoL vA) (1002) CG) the linizing fore, for large. samples, of the frequency elesert e catsiple corslactee ty Fisher (27) of the quantity Bt = nh Seefficinot of « tanden'y fe ae be obtained from the curves in Fig: 18 to 32.The function plotted in Figs. 13 to 32 is Typ(2N-1,N=1, 79) (1004) and B may be found easily from these curves for any values of ¥,, N and x. THE LINEAR DETECTOR ~ W VARIATES ‘The linear detector is usually more difficult to deal with than is the square detector. The distribution function for one variate of signal-plus-noise is dP = ve? Tcav)dv . (on) In attempting to find the distribution for the sum of N variates by the method of characteristic functions, the immediate trouble is that the characteristic function of Eq. (101) does not seem to be obtainable in closed form. To give an idea of the difficulty involved, the characteristic function for one variate of noise alone is (102) This is pair 903.3, Canpbell and Foster, and may be evaluated directly by completing the square or by forming a differential equation, giving in either case (103) or in terms of @ Sas A] : ae To raise this expression to the N'* power and then obtain the anticharacteristic function is practically hopeless. 29‘The distribution function for the sum of two variates of noise alone is ob- tainable by use of the convolution theorem, giving 1-8 s weser. Ze (aos) and the cumulative distribution is also obtainable, giving . tv". af Cove eter : (206) Y However, when N > 2 there seems to be no closed solution corresponding to Eq.(105) or (106). Since these cases are for noise alone, the signal-plus-noise situation must be attacked by other means. Te turns out that if the moments of the distribution for one variate are know, the moments of the distribution for the sum of N variates may be found directly. Formulae are given, for instance, in Craner, page 345, (1° for the first few central moments, which are Hh = Mey (07a) wy = May (07) = Mag + 3NN=1I 2 (07) al x 10N%u3 + 1SN3p3 + 15N2n,u, - 45N2u2 «ord ‘The corresponding coefficients in the Gram-Charlier series then become (108) (108) (108) 30‘The a's in Eqs.(108a-c) are the central standard monents for one variate, Note Chat, in the square law case, if N is put equal to 1 in Eqs. (B5c-f) and the resulting wis used in Eqs. (107a-d), the x's for N variates are correctly given. If a moment generating function can be found for the case of N variates, then it is immaterial which method is used; but in the case in which such a function is not available, the Eqs.(107) must be used (or some method essentially equivalent). To handle the linear detector it is now sufficient to find the monents for one variate only. Rice *), page 107, gives the required expression as “2099 y = arid) 8 C4, Rice also gives the first two moments as 4 fF Floon(2) + (3) (a9) vy + 2143) (10) To calculate % one needs to know the function, ,f(-3/2,1,-x). This may be obtained by use of the recursion relation aF(arlier2) + (are) F(anl,e,2) = (2arz-e) F(aserz) (111) by putting @ = -f, #4, 2 = ule is ¥% © ay (ez) VHF, (3) (106) leo ve = 4(2eaxez?) (110d) 31‘The corresponding central monents are My 07 = 2(142) = (12a) by = tf + ay, (1020) Ee 41,(3) cana) iy = (20403?) = auf = ayy [oro afRern(S)) ane ‘The standard central moments, and then the c's of Eys.(108a-c), are directly obtain- able fron these formilae, though the process is somewhat tedious due to the curbersone form of Eqs. (112a-c). The functions to ¥,are shown graphically as » function of x in Fig.34. To ottain the bias level ¥, for the linear detector for N > 2, one can use the G.C. series for noise alone, ‘Setting x = 0 and %= V772 in Eqs.(1l2a-c) gives a, tote 2 -Z= 0.429 (sa) firen « ore Ey Hs m2 0 Bo 0.808 case) wa oy + = 0.032 ce) 0.1083 aac au wed = 3.26 (15a) 32