I. INTRODUCTION Closed-Form Expression for the Poisson-Binomial Probability Density Function The binomial pdf describes the numbers of successes in N independent trials when the individual probabilities of success are constant across trials. If these probabilities are allowed to vary, as is the case in most practical applications, the resulting pdf is known as a Poisson-binomial. It is an extraordinarily useful model that can be encountered in all kinds of applications across vastly differing fields; here are some examples, together with references showing where to find related cases. MANUEL FERNÁNDEZ, Member, IEEE STUART WILLIAMS Lockheed Martin The Poisson-binomial probability density function (pdf) describes the numbers of successes in N independent trials, when the individual probabilities of success vary across trials. Its use is pervasive in applications, such as fault tolerance, signal detection, target tracking, object classification/identification, multi-sensor data fusion, system management, and performance characterization, among others. We present a closed-form expression for this pdf, and we discuss several of its advantages regarding computing speed and implementation and in simplifying analysis, with examples of the latter including the computation of moments and the development of new trigonometric identities for the binomial coefficient and the binomial cumulative distribution function (cdf). Finally we also pose and address the inverse Poisson-binomial problem; that is, given such pdf, how to find (within a permutation) the probabilities of success of the individual trials. Manuscript received December 1, 2006; revised September 27, 2007, May 29, 2008, and November 18, 2008; released for publication January 13, 2009. IEEE Log No. T-AES/46/2/936818. Refereeing of this contribution was handled by W. Koch. Authors’ current addresses: M. Fernandez, Lockheed Martin, MS2, 497 Electronics Pkwy., Liverpool, NY 13088, E-mail: (manuel.f.fernandez@Imco.com); S. Williams, Sensis Corporation, East Syracuse, NY 13057. c 2010 IEEE 0018-9251/10/$26.00 ° 1) Reliability Theory/Fault Tolerance [1]: If a manufacturing process fails when at least M out of N subprocesses fail, find the probability of failure when the individual probability of failure of the nth subprocess is pn . 2) Target Tracking [2—4]: A target track is initiated when at least M detections are declared, by the given sensor, in N consecutive, independent “look opportunities.” Given pn , the (varying) probability of detection per look, determine the probability of track initiation. 3) Pattern Identification/Decision Theory [5]: If the nth expert, diagnosing whether a particular condition is present or not, does so correctly in pn percent of the cases, how many such independent experts should coincide in their diagnosis to achieve an overall success rate that exceeds some desired percentage value. 4) Educational Examination Design: Given the percentages of students answering correctly on at least n out of the N equally-weighted questions of an standardized test, n = 0, 1, : : : , N, it may be of interest to determine (under assumptions of test and test-taker independence) the percentage of students that correctly answered each of the individual questions so as to establish whether the questions exhibit a desired “spread of difficulty.” (This is an inverse Poisson-binomial problem in that we are trying to obtain the probabilities of success (the percentages in this case) of the individual trials. Since, as is seen, the Poisson-binomial model is independent of the order in which the trials take place, this can only be done up to a permutation; in other words one can determine that x percent of the students got a question right, but one may not be able to determine which question it was.) 5) Multi-Sensor Fusion [5]: Given a net of N sensors whose detection/no-detection outputs are to be combined through a voting scheme, what should each sensor’s individual probability of false alarm be so as to achieve a specified M-out-of-N “fused” false alarm probability. IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 2 APRIL 2010 803 6) Project Management/Resource Allocation [1]: Given that rk resources have already been allocated to each of K workstations, the probability of each reaching its respective production quota is pk = f(rk ). Given the availability of L additional workstations, and assuming the invertibility of function f, what resources should be allocated to the new workstations so that at least M-out-of-the-(K + L) workstations achieve their production quota with a specified probability P. A methodology for obtaining closed-form representations for Poisson-binomial expressions of common use, such as the probability density function (pdf) and the cumulative distribution function (cdf) is presented. These closed-form expressions can be applied directly to solve most of the examples just presented. We believe, however, that the main contribution of this paper is not necessarily the presentation of the closed-form expressions, but rather the description of the methodology itself since it enables the development of algorithms for addressing other applications of the Poisson-binomial model. The paper is thus intended to be tutorial in tone, striving to derive, from basic principles, nearly all the results invoked, so as to be understandable to people from across as many disciplines as possible. The breakdown of the paper is as follows. Section I introduces the relationship between the probabilities of success of individual trials and the Poisson-binomial pdf. Sections III and IV use numerical techniques (polynomial interpolation and discrete Fourier transform (DFT) methods) to derive closed-form formulas for the Poisson-binomial pdf and cdf, and Sections V and VI then demonstrate the use of these expressions by obtaining new representations of the binomial coefficient, the binomial cdf, and the Poisson-binomial moments. Section VII illustrates the use of the various techniques presented in this paper by applying them to solve the problem in example 6 of this Introduction. Section VIII again uses polynomial methods, together with matrix-theoretic techniques, to shed some light on the inverse Poisson-binomial problem (namely, given the Poisson-binomial pdf or cdf, how to obtain, up to a permutation, the probabilities of success of the individual trials); this section also sounds a cautionary note regarding numerical stability when using polynomials. Finally Section IX presents some comments and a summary of the paper. 804 II. GROUNDWORK Consider N independent trials with probabilities of success and failure, for the kth trial, equal to pk and 1 ¡ pk , respectively (“Poisson trials”). The number Y of successes can be written as the sum Y = X1 + X2 + ¢ ¢ ¢ + XN of N mutually independent random variables Xk with the distribution vectors [PrfXk = 0g PrfXk = 1g] = [1 ¡ pk pk ], where Prfug denotes “probability of u” (see Feller [6] for example). The distribution of the sum Y of these random variables, the Poisson-binomial pdf (a.k.a. Bernoulli-Poisson pdf), is then given by the linear convolution of the distributions of the Xk s; that is [PrfY = 0g PrfY = 1g ¢ ¢ ¢ PrfY = Ng] = [1 ¡ p1 p1 ] ¤ [1 ¡ p2 p2 ] ¤ ¢ ¢ ¢ ¤ [1 ¡ pN pN ]: (1) 1 Taking the Z-transform of each side of (1) yields two versions of the generating function of the Poisson-binomial pdf, P0 + P1 z + P2 z 2 + ¢ ¢ ¢ + PN z N = (1 ¡ p1 + p1 z)(1 ¡ p2 + p2 z) ¢ ¢ ¢ (1 ¡ pN + pN z) (2) where Pn was used instead of PrfY = ng to simplify the notation. A minor manipulation of the right-hand side of (2) yields P0 + P1 z + P2 z 2 + ¢ ¢ ¢ + PN z N = ®(z ¡ s1 )(z ¡ s2 ) ¢ ¢ ¢ (z ¡ sN ) where ®= N Y pk (3a) (3b) k=1 and sk = ¡(1 ¡ pk )=pk : (3c) Notice, thus, that the values comprising the Poisson-binomial pdf and the pdfs of the individual trials are merely the parameters of two different representations of the same polynomial. The former is in expanded form, while the latter is factored; hence, given one we may extract the other (i.e., it is an invertible mapping). Notice also that, since the 1 In the modern engineering literature, the Z-transform is usually defined as the representation of a sequence of numbers as coefficients of a polynomial in z ¡1 (i.e., the nth element of the sequence, n = 0, 1, : : : , N ¡ 1, becomes the coefficient of the polynomial’s (z ¡1 )n term); however, since for our purposes such representation is far simpler when defined in terms of z rather than z ¡1 , we have done so. (Some authors try to avert this definitional conflict by using x instead of z ¡1 and by dubbing it the “X-transform” (see [7], for example), but we decided to avoid introducing more esoterica.) IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 2 APRIL 2010 Fig. 1. MATLAB code for mapping (a) from individual probabilities of success in N independent trials to Poisson-binomial pdf and (b) vice versa. Q polynomial N k=1 (z ¡ sk ) on the right-hand side of (3) is monic, the coefficient of the largest order term of the polynomial in expanded form PN satisfies PN = ®. As shown in Fig. 1, using a computing language such as MATLAB, with its built-in functions “poly” and “roots,” it is very easy to quickly write a script for, given the probabilities of the individual trials (i.e., the pk ), obtaining the entries of the Poisson-binomial pdf (i.e., the Pn ), and vice versa. All these two routines produce, however, are merely lists of numbers which, albeit extremely useful in many applications, do not easily lead to new insight into further uses or manipulations of the Poisson-binomial pdf. (From the perspective of implementation complexity, using the routine of Fig. 1(a) to obtain the values of the Poisson-binomial pdf by direct expansion of the right-hand side of (3a) via the fast method implemented in MATLAB’s “poly” function is a very efficient procedure, requiring only on the order of N log2 N operations.2 ) III. A CLOSED-FORM EXPRESSION FOR THE POISSON-BINOMIAL PDF Returning to (2) let us recall that our objective is to, given the pk , obtain an expression yielding the Pn . One such expression can be obtained by using the Vandermonde polynomial interpolation method; namely by evaluating the right-hand side of (2) at N + 1 different values of z and then by finding the coefficients of the Nth-order polynomial 2 Other methods for obtaining the Poisson-binomial pdf include the use of “fast convolution” techniques to obtain the right-hand side of (1) (see [1] and [8]), as well as a variety of recursive approaches (see, for example, [1], [9], and [10], all of which also provide interesting examples of the applicability of the Poisson-binomial density across seemingly dissimilar fields). FERNÁNDEZ & WILLIAMS: CLOSED-FORM EXPRESSION FOR THE POISSON-BINOMIAL PDF 805 exactly traversing the results. In matrix-vector form, this method involves solving, for vector P, the linear system of equations 2 3 P0 21 a 2 N3 a0 ¢ ¢ ¢ a0 6 7 0 7 6 P1 7 61 a 7 a21 ¢ ¢ ¢ aN 1 6 1 76 7 6 P2 7 6 7 6 .. .. .. .. 7 6 6 7 4. . . . 5 6 .. 7 4 . 5 1 aN a2N ¢ ¢ ¢ aN N | {z } PN | {z } V 2 P N Y 3 fpk a0 + (1 ¡ pk )g 7 6 7 6 k=1 7 6 7 6 N 7 6Y 6 fpk a1 + (1 ¡ pk )g 7 7 6 7 =6 7 6 k=1 7 6 .. 7 6 7 6 . 7 6 7 6Y N 5 4 fpk aN + (1 ¡ pk )g or | k=1 {z r (4a) e¡j2¼nm=(N+1) P(m) = n=0 k=1 (N + 1), fpk ej2¼n=(N+1) + (1 ¡ pk )g m = 0, 1, : : : , N (5) a closed-form expression for the Poisson-binomial pdf. As already stated many algorithms exist for efficiently computing ), (N Y j2¼n=(N+1) fpk e + (1 ¡ pk )g (N + 1) (6) DFT k=1 which is what (5) entails. Further simplifications can be achieved by considering instead ), (N Y (ej2¼n=(N+1) ¡ sk ) (N + 1) (7) ®DFT } k=1 VP = r: (4b) Matrices with a structure as that of V are known as “Vandermonde” matrices. Theoretically speaking square Vandermonde matrices are nonsingular; hence we should be able to obtain P = V¡1 r. In practice, however, these matrices are particularly unstable numerically in the sense that, in general, as N increases, the unbalancing in the structure of the matrix (i.e., relatively large-magnitude numbers concentrating on certain zones of the matrix), coupled to computer round-off, cause enough error to perturb the matrix into singularity. And even if this were not so, the problem of having to compute the actual inverse of V still would remain since otherwise the closed-form expression for P would not provide much insight. Fortunately both issues (stability and inversion) can be easily solved by choosing the an as roots of unity; that is, as an = ej2¼n=(N+1) . Such a choice clearly thwarts magnitude growth with N, and it also solves the inversion problem by reason of the fact that the resulting Vandermonde matrix is unitary; that is, VH V = VVH = (N + 1)I, where I is an (N + 1) £ (N + 1) identity matrix and where the superscripted H is used to denote the complex-conjugate transpose. Hence, P = VH r=(N + 1), or, taking advantage of the fact that in this particular case matrix V happens to be symmetric, P = V¤ r=(N + 1), where the asterisk represents the operation of conjugating the entries of V. The matrix V¤ , for our particular choice of an (let’s call it matrix F so as to distinguish it from 806 other possibilities), is known as the DFT matrix, and countless fast Fourier transform algorithms exist for computing products of the form Fr or, as more commonly expressed, for computing DFTfr(n)g. More importantly for us the entries of P can now be expressed in terms of a summation; namely as ( ), N N X Y with ® and sk as defined in (3b) and (3c) so as to reduce the number of multiplications. Even further one may exploit the fact that, since the Poisson-binomial pdf is real-valued, the argument Q j2¼n=(N+1) (e ¡ sk ), of the DFT, that is, the values N k=1 n = 0, 1, : : : , N, must posses an “even” real part and an “odd” imaginary part (see Bracewell [11] for example), which means that we only need to compute half of the entries. Leaving to others the details of efficiently implementing the Poisson-binomial pdf formula, let’s now entertain some of the possibilities that are opening before us by possessing such a closed-form expression. We consider the formula for the Poisson-binomial cdf, the derivation of a new identity for the binomial coefficient, and new expressions for the moments of the Poisson-binomial distribution. IV. THE POISSON-BINOMIAL CUMULATIVE DISTRIBUTION FUNCTION The Poisson-binomial pdf provides the probabilities of exactly m successes out of N Poisson trials (m = 0, 1, : : : , N). Perhaps of more practical use, however, is the complement of its cdf (its “survival function”–call it Q(m)), which provides the probabilities of at least m successes in N Poisson trials; that is Q(m) = N X P(t), m = 0, 1, : : : , N (8) t=m IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 2 APRIL 2010 to maximize the likelihood of success in at least M future contests. or equivalently 8 m¡1 X > <1 ¡ P(t), Q(m) = t=0 > : 1, m = 1, 2, : : : , N : V. m=0 (9) Substituting (5) in for P(t) in (9), and exchanging the summations, yields Q(0) = 1 (10a) Q(m) = 1 ¡ ("m¡1 N X X n=0 e¡j2¼nt=(N+1) t=0 # N Y fpk ej2¼n=(N+1) k=1 ), + (1 ¡ pk )g (10b) We can now simplify the expression in brackets in (10b) to obtain a closed-form expression for the complement of the Poisson-binomial cdf; namely ( N X (1 ¡ e¡j2¼nm=(N+1) )=(1 ¡ e¡j2¼n=(N+1) ) n=0 ¢ ), N Y k=1 fpk ej2¼n=(N+1) + (1 ¡ pk )g (N + 1) = 1 ¡ m=(N + 1)+ ¡ N X n=1 ( (1 ¡ e¡j2¼nm=(N+1) )=(1 ¡ e¡j2¼n=(N+1) ) ¢ N Y k=1 ), fpk ej2¼n=(N+1) + (1 ¡ pk )g In this section we illustrate the use of the expressions and techniques presented so far by using them to derive new representations of the binomial coefficient and the binomial cdf. These new expressions are based solely on sums of sinusoidals and completely avoid the use of factorials. Consider the well-known expression for the binomial pdf μ ¶ B(k j N, p) = (N + 1), m = 1, : : : , N: Q(m) = 1 ¡ A NEW FORMULA FOR THE BINOMIAL COEFFICIENT (N + 1) N k pk (1 ¡ p)N¡k (11) Clearly, since the relationship between a cdf (call it C(m)) and its complement is C(m) = 1 ¡ Q(m), one can also extract from (11) a closed-form expression for the Poisson-binomial cdf. The outcomes of such expression (i.e., the C(m), m = 0, 1, : : : , N) would represent the probability of success in (strictly) less than m trials out of N. Applications of the complement of the Poisson-binomial cdf are countless, from determining the probability of overall system failure due to faults in at least M components [1] and that of initiating a target track after N sensor scans [2] to inferring the probability of correctly classifying a pattern after fusing (e.g., via “majority voting”) the declarations of multiple independent experts [5], or estimating what resources to allocate to a current action so as k = 0, 1, : : : , N (12) where (N k) = N!=((N ¡ k)!k!) is the “binomial coefficient.” This coefficient, also known as the “given N, choose k” function, tells us the number of ways of choosing k out of N objects without regard of their order, i.e., the number of combinations. The binomial pdf provides the probability of having k successes out of N independent trials, when the probability of success of each individual trial has the same value p. It is thus a special case of the Poisson-binomial pdf, and we can equate (5) and (12) in such instances. We do so for the special case p = 1=2, and we change the notation slightly (using “m” instead of “k” in (12) above) so as to match the notation of (5). Setting p = 1=2 in both (12) and (5) and simplifying and equating both expressions yields μ ¶ N X N ¡N 2 fe¡j2¼nm=(N+1) (ej2¼n=(N+1) + 1)N g= = 2¡N m n=0 (N + 1), for m = 0, 1, : : : , N: for m = 0, 1, : : : , N: (13) (Remark: Notice that the left-hand side of (13) is merely the binomial coefficient function scaled by the constant 2¡N . In other words one can think of the binomial coefficient, when considered a function of m, as a pdf scaled by 2N . This fact can be used to simplify its computation.) Factoring the term in parentheses in (13) as ej2¼n=(N+1) + 1 = ej¼n=(N+1) (ej¼n=(N+1) + e¡j¼n=(N+1) ), and using the identity ejÁ + e¡jÁ = 2 cos(Á), results in μ ¶ N X N = 2N m n=0 fej¼n(N¡2m)=(N+1) cosN (¼n=(N + 1))g=(N + 1), m = 0, 1, : : : , N (14) or equivalently μ ¶ N m = 2N DFTfej¼nN=(N+1) cosN (¼n=(N + 1))g=(N + 1), FERNÁNDEZ & WILLIAMS: CLOSED-FORM EXPRESSION FOR THE POISSON-BINOMIAL PDF m = 0, 1, : : : , N: (15) 807 Finally exploiting the fact that the left-hand side is real (meaning that the imaginary part of the right-hand side must vanish), we can further simplify things to obtain μ ¶ N X N = 2N fcos(¼n(N ¡ 2m)=(N + 1)) m n=0 ¢ cosN (¼n=(N + 1))g=(N + 1), m = 0, 1, : : : , N: (16) This new identity3 expresses the binomial coefficient in terms of a weighted sum of cosines rather than as a ratio of factorials, with the weights being the factors 2N cosN (¼n=(N + 1))= (N + 1). Since these weights solely depend on N, they need to be computed only once, and they can be prestored.4 However, although (16) is perhaps a “cleaner” expression, writing the identity as in (15) may be of more general utility. For example consider obtaining an identity for the power of a cosine: Taking the inverse DFT (the “IDFT”) of both sides of (15) results in μ ¶¾ Á N ½ X N ej2¼nm=(N+1) (N + 1) m m=0 = 2N fej¼nN=(N+1) cosN (¼n=(N + 1))g=(N + 1): (17) Solving for cosN (¼n=(N + 1)) yields μ ¶¾ N ½ X N cosN (¼n=(N + 1)) = 2¡N e¡j¼n(N¡2m)=(N+1) m m=0 yielding the well-known trigonometric identity for the power of a cosine. Expression (15) can also be used to obtain a new formulation for the binomial cdf. Roughly speaking this cdf depicts the probability of success in at most M out of N trials, when each trial has an equal probability p of success; it can be portrayed as μ ¶¾ M ½ X N m (N¡m) , p (1 ¡ p) m m=0 M = 0, 1, : : : , N: (20) Assuming that p is strictly less than 1 (i.e., 0 · p < 1), we can perform the manipulation where pm (1 ¡ p)(N¡m) = (1 ¡ p)N tm (21a) t = p=(1 ¡ p): (21b) Substituting this in (20), together with (15) for the binomial coefficient, yields μ ¶¾ M ½ X N m (N¡m) p (1 ¡ p) m m=0 = 2N (1 ¡ p)N =(N + 1) ( N M X X (ej¼nN=(N+1) tm ¢ m=0 n=0 ¡j2¼nm=(N+1) N ¢ cos (¼n=(N + 1))e (18) and, realizing that both sides of the equation must be real, μ ¶¾ N ½ X N N ¡N cos (Á) = 2 cos(Á(N ¡ 2m)) : m m=0 (19) ) ) : (22) By exchanging summations the right-hand side of (22) becomes ( N X N N ej¼nN=(N+1) cosN (¼n=(N + 1)) 2 (1 ¡ p) =(N + 1) n=0 At this particular stage Á = ¼n=(N + 1); however, realizing that Á doesn’t depend on m allows generalizing it to represent any desired angle,5 thus ¢ M X (te¡j2¼n=(N+1) )m ) (23a) m=0 or, after simplifying, 3 An anonymous reviewer suggests [12] for a current compilation of binomial identities. 4 Several symmetries can be exploited to reduce the number of operations required to compute (16), the most obvious being that of the binomial coefficient about m = N=2 and that of the portion in braces, for the values of n > 0, about n = (N + 1)=2 (when n = 0 the portion in braces is always equal to 1, regardless of m). 5 Another way to visualize this is to consider the convolution of both sides of (19) with a shift-inducing delta function of the form ±(n ¡ '), with ' chosen such that ¼'=(N + 1) equals the desired Á. 808 2N (1 ¡ p)N =(N + 1) N X fej¼nN=(N+1) cosN (¼n=(N + 1)) n=0 (1 ¡ tM+1 e¡j2¼n(M+1)=(N+1)) )= (1 ¡ te¡j2¼n=(N+1) )g: (23b) By realizing that (23b), being a cumulative distribution function, it must be a real number (that is, its imaginary part must vanish), we obtain IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 2 APRIL 2010 M ½ X m p (1 ¡ p) m=0 (N¡m) μ ¶¾ N m N N = 2 (1 ¡ p) =(N + 1) ( T+ N X n=1 fcosN (Ãn)[(1 ¡ t) cos(ÃnN) ¡ tM+1 (cos(Ãn(2(M + 1) + 1)N) 2 ) ¡t cos(Ãn(2(M + 1) ¡ 1)N))]=(1 ¡ 2t cos(2Ãn) + t )g with à = ¼=(N + 1) (24b) and 8 (1 ¡ tM+1 )=(1 ¡ t) > < for 0 · p < 1, p 6= 0:5 (i.e., t 6= 1) T= > : M +1 for p = 0:5 (i.e., t = 1) (24c) where L’Hopital’s rule is used to obtain the value of T for the case where p = 0:5. (Note that, for p = 1, the binomial cdf equals zero, except when M = N, in which case it equals 1. Using (24a) directly would be cumbersome for such case, which would again require the use of L’Hopital’s rule as t would be undefined.) Expressions (24a)—(24c) provide a new representation of the binomial cdf, which is solely based on trigonometric functions. These, together with (16)–the new representation for the binomial coefficient–are, as a reviewer points out, maybe somewhat surprising, before one realizes that they are just simple consequences of the binomial theorem and the properties of the direct and inverse DFT. The very surprise that the binomial coefficient (a ratio of factorials) can be formulated as a sum of sinusoidals, the fact that this formulation is nothing more–and nothing less!–than a result of the properties of the DFT, and the fact that different formulations could have been obtained by selecting other entries for the Vandermonde matrix in (4) are perhaps the main contribution of this section as they open opportunities for obtaining new representations for existing tools and for recognizing them when they appear under a different guise. VI. POISSON-BINOMIAL MOMENTS The Lth Poisson-binomial moment is the scalar EfmL P(m)g = N X m=0 = fmL P(m)g N X m=0 ( L m DFT (N Y k=1 j2¼n=(N+1) fpk e + (1 ¡ pk )g ))Á (N + 1) Using the well-known identity (e.g., see Bracewell [11]) that the Lth moment of the DFT of an N + 1 element sequence r(n) satisfies ¯ N L X ¯ L L d m DFTfr(n)g = ((N + 1)=(2¼j)) (r(n))¯¯ L dn n=0 m=0 yields the following expression for the Lth Poisson-binomial moment: dL EfmL P(m)g = (N + 1)L¡1 =(2¼j)L L dn ÃN !¯ ¯ Y ¯ ¢ fpk ej2¼n=(N+1) + (1 ¡ pk )g ¯ ¯ k=1 (26) n=0 (27) which mainly involves taking the Lth derivative of the generating function. An alternate expression can be achieved, however, that obviates the need for taking such derivatives. It is more easily obtained if we represent (25) in matrix-vector form: N X L Efm P(m)g = mL DFTfr(n)g=(N + 1) = gTL Fr m=0 (28) QN j2¼n=(N+1) + (1 ¡ pk )g, F is where r(n) = k=1 fpk e the DFT matrix, r = [r(0) r(1) ¢ ¢ ¢ r(N)]T with the superscripted T denoting transposition, and where gTL = [0 1 2L ¢ ¢ ¢ N L ]=(N + 1). By taking advantage of the symmetry of the DFT matrix F, one can express the product gTL F as the transpose (without conjugation) of the vector that results from taking the DFT of gL ; that is, EfmL P(m)g = GTL r, where vector GL = DFTfgL g. Since gL is independent of the probabilities of success of the independent trials, GL can be precomputed for the desired values of N and L. Calculation of the corresponding moment thus corresponds to a weighted sum of the results of evaluating the Poisson-binomial generating function at the values ej2¼n=(N+1) , n = 0, 1, : : : , N. We now rederive this result for those who prefer summations to matrices. We return to (28) and express the DFT in all its glory; ( )Á N N X X EfmL P(m)g = (25) (24a) mL m=0 n=0 fe¡j2¼nm=(N+1) r(n)g where use is made of (5) and (6) for P(m). FERNÁNDEZ & WILLIAMS: CLOSED-FORM EXPRESSION FOR THE POISSON-BINOMIAL PDF (N + 1): (29) 809 TABLE I Values of the Arithmetic Series Divided by N + 1 and Functions Resulting after Evaluating Lth Derivative of u(x) at x = n, for L = 1, 2, 3, and 4 N X L ¯ ¯ dL u(x)¯ ¯ dxL mL =(N + 1) m=1 1 N=2 2¼j=(1 ¡ e¡j2¼n=(N+1) ) 2 N(2N + 1)=6 (4¼ 2 =(1 ¡ e¡j2¼n=(N+1) ))f1 + 2e¡j2¼n=(N+1) =[(1 ¡ e¡j2¼n=(N+1) )(N + 1)]g 3 N 2 (N + 1)=4 (¡8¼ 3 j=(1 ¡ e¡j2¼n=(N+1) ))f1 + 3(N + 2)e¡j2¼n=(N+1) =[(1 ¡ e¡j2¼n=(N+1) )(N + 1)2 ] + 6e¡j4¼n=(N+1) =[(1 ¡ e¡j2¼n=(N+1) )2 (N + 1)2 ]g 4 N(6N 3 + 9N 2 + N ¡ 1)=30 (¡16¼ 4 =(1 ¡ e¡j2¼n=(N+1) ))f1 + (2e¡j2¼n=(N+1) =[(1 ¡ e¡j2¼n=(N+1) )(N + 1)3 ])f2N 2 + 7N + 7 + 6(N + 3)e¡j2¼n=(N+1) =(1 ¡ e¡j2¼n=(N+1) ) + 12e¡j4¼n=(N+1) =(1 ¡ e¡j2¼n=(N+1) )2 gg By noticing that r(0) = 1 and by exchanging summations, ( N ) X L L Efm P(m)g = m =(N + 1) m=0 + N N X X [mL e¡j2¼nm=(N+1) ]r(n)=(N + 1): (30) n=1 m=0 The quantity in brackets canP be expressed as ¡j2¼xm=(N+1) )jx=n [(N + 1)=(¡2¼j)]L (dL =dxL )( N m=0 e or equivalently as [(N + 1)=(¡2¼j)]L (dL =dxL ) f(1 ¡ e¡j2¼x )=(1 ¡ e¡j2¼x=(N+1) )gjx=n ; hence ( N ) X EfmL P(m)g = mL =(N + 1) m=1 + (N + 1) L¡1 =(¡2¼j) L N ½ X n=1 ¯ ¯ dL r(n) L u(x)¯¯ dx x=n ¾ (31) where u(x) = (1 ¡ e¡j2¼x )=(1 ¡ e¡j2¼x=(N+1) ): Since u(x) does not depend on the probabilities of success of the individual trials, the function resulting after evaluating its Lth derivative at x = n may be worked out and tabulated beforehand (likewise with the first term of (31)). Table I shows such functions of n for the values of L from 1 to 4. One can obtain further simplifications by comparing the expressions for the first and second moments obtained using (31) to those that would have been obtained using basic principles (i.e., EfmP(m)g = 1T p, and Efm2 P(m)g = pT (1 ¡ p) +(1T p)2 , where 1 is a vector of ones and p is a vector comprised of the individual probabilities of success–see [6]). VII. EXAMPLE For purposes of illustrating the techniques of this paper, in this section we address one of 810 x=n the application examples mentioned in Section I, namely, case 6. In summarized form: given that each of K existing workstations have a probability pk of achieving their production goals, what probability of success do we need to attain in each of L additional workstations if we are to ensure a specified probability Ps that at least M-of-the-(K + L) workstations achieve their production quotas. (We assume here, for purposes of simplification, that the probability of success for each of the new workstations is equal–call it pc .) First let us briefly further abstract and simplify this example in an attempt to obtain a more intuitive understanding of the underlying problem and its solution (and hopefully open the door to more ideas about its applicability). Namely let us consider the problem where we are given the not-necessarily-equal probabilities of success of each of N individual trials, and we are then told that one additional trial is allowed. What probability of success pc would we need to assign to that extra trial so as to attain a desired probability Ps of “at least M successes out of the N + 1 trials”? Clearly in such a case, the desired probability Ps may be expressed as Ps = Pr(“at least M out of N”) + pc ¢ Pr(“exactly M ¡ 1 out of N”): (32) Notice that this is a first-order polynomial in pc . In solving for pc we obtain the value pc = (Ps ¡ Pr(“at least M out of N”))= Pr(“exactly M ¡ 1 out of N”) (33) as the estimate of the probability of success needed to achieve exactly the desired Ps at the fatidic (N + 1)th try. (Notice, however, that the above expression for Ps has no safeguards to ensure that pc is a valid probability; hence for it to be of practical use, and depending on the application, the resulting value is usually either discarded or at least clipped, e.g., set to zero if negative or truncated when exceeding a user-selected threshold.) IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 2 APRIL 2010 By extending this example to the case where two new trials are allowed, we would obtain Ps = Pr(“at least M out of N + 1”) + pc ¢ Pr(“exactly M ¡ 1 out of N + 1”) = (Pr(“at least M out of N”) + pc ¢ Pr(“exactly M ¡ 1 out of N”)) + pc ¢ (pc ¢ Pr(“exactly M ¡ 2 out of N”) + (1 ¡ pc ) ¢ Pr(“exactly M ¡ 1 out of N”)) = qp2c + 2pc ¢ Pr(“exactly M ¡ 1 out of N”) + Pr(“at least M out of N”), (34a) where q = Pr(“exactly M ¡ 2 out of N”) ¡ Pr(“exactly M ¡ 1 out of N”): (34b) The solution to this quadratic polynomial in pc is, of course, of the form pc = [¡Pe (M ¡ 1) § fPe (M ¡ 1)2 + (Ps ¡ Pr(“at least M out of N”))qg1=2 ]=q (35) with the notation Pe (L) meaning Pr(“exactly L out of N”). Again neither pc may be a valid probability, and, depending on the application, it may also have to be discarded or clipped. Continuing the drill any further simply complicates the formulation of the problem and its solution, hence the advantage of using the methodology of this paper. Nonetheless hopefully some insight may be obtained out of the exercise and maybe even some inspiration for new application ideas. By returning now to the example and appropriately tailoring (3a) to it, we can see that the Poisson-binomial pdf for this particular case corresponds to the coefficients of the polynomial This provides, as one of its outputs, the probability that at least M-of-the-(K + L) workstations will achieve their production quotas. By exploiting the monotonicity of the Poisson-binomial cdf and, hence, of its complement (the probability of at least M successes out of N trials), one can then use a method such as bisection, or equivalently MATLAB’s “fzero” function, to iteratively converge to the solution. (Fig. 2 presents a MATLAB version of this approach.) Notice that one only needs to perform this guessing game if, when setting pc = 0, one doesn’t meet or exceed the desired Ps , but, when setting pc = 1, one does. In the first case (pc = 0), the production quota is achieved without the need of any additional workstations, while, in the second (pc = 1), the quota is not reached with the addition of only L of them. Even though, numerically speaking, the function of Fig. 2 is more than sufficient to address our example, it may still be of interest to go through the development of the expressions implicitly contained therein. Letting N = K + L and using (11) yields, for m = M, Q(M) = Ps = 1 ¡ M=(N + 1) ¡ N X ( (1 ¡ e¡j2¼nM=(N+1) )=(1 ¡ e¡j2¼n=(N+1) ) n=1 N Y ¢ k=1 j2¼n=(N+1) fpk e )Á + (1 ¡ pk )g (N + 1) = 1 ¡ M=(N + 1) ¡ N X n=1 ( (1 ¡ e¡j2¼nM=(N+1) )=(1 ¡ e¡j2¼n=(N+1) ) ¢ (pc ej2¼n=(N+1) + (1 ¡ pc ))L ¢ p(z) = ®(z ¡ r)L (z ¡ s1 )(z ¡ s2 ) ¢ ¢ ¢ (z ¡ sK ) K Y k=1 j2¼n=(N+1) fpk e + (1 ¡ pk )g )Á (N + 1) (37) (36a) where ® = pLc N Y pk (36b) k=1 r = ¡(1 ¡ pc )=pc (36c) sk = ¡(1 ¡ pk )=pk : (36d) and An anonymous reviewer suggested the following very simple, practical, and stable numerical approach for solving the example problem using these results. Guess a value for pc , and feed the vector pT = [pc 1TL p1 p2 ¢ ¢ ¢ pK ], where 1L denotes a vector of L “ones,” to function “ProbMofN” of Fig. 1(a). where use is made of the assumption that pc , the probability of reaching the production quota, is the same for all the new workstations. Alternatively (N + 1)(1 ¡ Ps ) ¡ M ( N X (1 ¡ e¡j2¼nM=(N+1) )=(1 ¡ e¡j2¼n=(N+1) ) n=1 ¢ (pc ej2¼n=(N+1) + (1 ¡ pc ))L ) K Y j2¼n=(N+1) ¢ fpk e + (1 ¡ pk )g (38) k=1 FERNÁNDEZ & WILLIAMS: CLOSED-FORM EXPRESSION FOR THE POISSON-BINOMIAL PDF 811 Fig. 2. Given vector with non-zero probabilities of success of each of K individual, independent trials, this MATLAB code computes probability of success that would be needed in each of L additional trials to obtain desired probability of success in at least M out of K + L trials. (This function can be used directly to solve the problem of Section VII.) where we have collected the terms outside the summation, and the scale factor (N + 1), on the left-hand side of the expression. Since the right-hand side of (38) is a polynomial in pc , we may follow again the strategy of expanding this polynomial by using the approach of Section III “correcting” the zeroth-order term by subtracting from it the left-hand side of (38), and then solving for its zeros by picking as pc any such zero that is also a valid probability, if it exists. Solving for the coefficients of the right-hand size of (38) by using the techniques of Section III involves substituting for pc values of the form ej2¼l=(N+1) , l = 0, 1, : : : , L, computing the L + 1 element sequence of values resulting from such substitutions, and then taking the DFT of the sequence and scaling it by 1=(L + 1). In other words the coefficients of the C(u) = N X ( n=1 ¢ 812 (1 ¡ e " L X `=0 ¡j2¼nM=(N+1) polynomial of the right-hand side of (38) are given by C(u), u = 0, 1, : : : , L, where C(u) = e¡j2¼lu=(N+1) `=0 à N ( X ¢ (1 ¡ e¡j2¼nM=(N+1) )=(1 ¡ e¡j2¼n=(N+1) ) n=1 ¢ (ej2¼(n=(N+1)+l=(L+1)) + (1 ¡ ej2¼l=(N+1) ))L )! Á K Y j2¼n=(N+1) ¢ fpk e + (1 ¡ pk )g (L + 1), k=1 u = 0, 1, : : : , L: (39) Exchanging summations and simplifying yields ¡j2¼n=(N+1) )=(1 ¡ e L X K Y ) (pk (ej2¼n=(N+1) ¡ 1) + 1) k=1 e¡j2¼lu=(N+1) (ej2¼l=(L+1)) (ej2¼n=(N+1) ¡ 1) + 1)L #Á (L + 1), ) u = 0, 1, : : : , L: (40) IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 2 APRIL 2010 By using matrix notation one may condense (40): observe that the factor in brackets is equivalent to taking the DFTs of the columns of an (L + 1) £ N matrix (call it “E”) with element E(l + 1, n) defined by E(l + 1, n) = (ej2¼l=(L+1) (ej2¼n=(N+1) ¡ 1) + 1)L , l = 0, 1, : : : , L, n = 1, 2, : : : , N: (41) Also notice that the portion in braces of (40) can be expressed as the element-by-element (“Hadamard”) product of N-element vectors d and r, with elements given by d(n) = (1 ¡ e¡j2¼nM=(N+1) )=(1 ¡ e¡j2¼n=(N+1) ) (42a) and r(n) = K Y (pk (ej2¼n=(N+1) ¡ 1) + 1), n = 1, 2, : : : , N k=1 (42b) or, alternatively, letting the d(n) represent the nth diagonal element of a diagonal matrix D, said product may also be represented as the matrix-vector product Dr. In combining these results (40) can be expressed in matrix/vector form as C = FEDr=(L + 1) (43) where F is the DFT matrix. As an aside notice that evaluating (43) by proceeding from right-to-left is efficient in that it involves a Hadamard product, followed by matrix-vector product and an “FFT”. However notice, also, that since the product FED does not depend on the pk , it can be precomputed for a given (K, L, M) triad and then used to obtain C for multiple sets of pk . Having obtained the coefficients of the polynomial that comprise the right-hand side of (38), we may now subtract the left-hand side from the zeroth-order term (i.e., from C(0) or, equivalently, from the first element of vector C; that is q = C ¡ ((N + 1)(1 ¡ Ps ) ¡ M)e1 (44) where e1 denotes the (L + 1)-element vector [1 0 ¢ ¢ ¢ 0]T . The zeros of the polynomial q(L + 1)xL + q(L)xL¡1 + ¢ ¢ ¢ + q(2)x + q(1) should include the solution to our example, if such solution exists. These zeros can be obtained by using the companion matrix approach described in Section VIII; however, as discussed in all numerical analysis textbooks, it is well known that the roots of a polynomial are not easily found.6 It is, thus, more sensible to exploit the fact that since the entries of q are all real numbers and 6 Reference [13], suggested by an anonymous reviewer, presents an excellent, concise description of the difficulties inherent to obtaining polynomial roots. the zero of interest (if it exists) must lie between 0 and 1, we can find it using an iterative zero-finding approach, such as a combination of a Newton-type method and a bisection algorithm (or equivalently MATLAB’s “fzero” function). Some comments regarding this section now follow. First notice that by setting the pk k = 1, 2, : : : , K equal to zero in the expressions above, our formulation also solves example 5 of Section I; in other words it also addresses a particular class of “inverse” binomial problem. Also (even though we do not show it here), by treading through the developments of this section for the case where the pk are zero, one can derive an alternative expression for the probability of “at least M successes out of N Bernoulli (i.e., equal-probability) trials” which is already in expanded polynomial form; namely for M > 0, μ ¶ N X N n N¡n p (1 ¡ p) n n=M " ¶!# μ ¶ ÃX μ N M X N n n n m = (¡1) p (¡1) n m¡1 n=M m=1 (45) where the left-hand side denotes the classical form of expressing such probability and where the right-hand side represents it in expanded polynomial form. Notice that, since all the coefficients can be precomputed, solving a bounded single-variable problem of the form " ¶!# μ ¶ ÃX μ N M X N n n n m Ps = (¡1) pc (¡1) n m¡1 n=M m=1 (46) can again be done using an iterative zero-finding algorithm. VIII. THE INVERSE POISSON-BINOMIAL PROBLEM The practical definition of the “inverse” of a Poisson-binomial pdf usually lies in the eye of the practitioner, which means that there are as many variants and complications as there are applications for this problem. Nonetheless one such definition may be simple enough: given an (N + 1)-element Poisson-binomial pdf, find (within a permutation) the probabilities of success of the N independent Poisson trials that gave rise to it. Theoretically these individual probabilities should be obtainable using either the MATLAB routine of Fig. 1(b) or, adding more detail, by using the companion matrix-based approach described shortly (both are equivalent). Unfortunately as will be seen, numerical stability issues can render the algorithm of Fig. 1(b) useless even for moderately-sized problems (e.g., N > 30). This section FERNÁNDEZ & WILLIAMS: CLOSED-FORM EXPRESSION FOR THE POISSON-BINOMIAL PDF 813 thus merely sketches some general hurdles, describes what may be meant by “inverse” under different circumstances, and references applications where such inverses and hurdles may be encountered. A more detailed explanation of some of these problems and their solutions will be the subject of a future paper. As can be ascertained from the expressions in Section II it is always possible to associate any N positive numbers that do not exceed unity to a Poisson-binomial pdf. Practical issues regarding the sensitivity of this pdf’s entries to errors in the estimates of the probabilities of success of the individual trials, whether the trials are or not independent or whether the pdf matches or not our particular expectations, are immaterial from the perspective that the result exhibits, under all viewpoints, the characteristics of a Poisson-binomial pdf. The converse, however, is not so clear-cut. Of course any N + 1 nonnegative numbers not exceeding unity, and adding up to 1 (i.e., a “pdf”), do not, in general, represent a Poisson-binomial pdf and, as such, we shouldn’t expect to be able to factorize its generating function into terms of the form of (3). But the problem is a little more insidious than that: even when a pdf’s probabilities indeed proceed from a Poisson-binomial process, we may still not be able to properly factorize the pdf’s generating function due to errors in the estimates (or “measurements”) of those probabilities as well as to inherent numerical instabilities. An approach for handling the case where there are errors in the pdf entries is to attempt to solve, instead, for the roots of the polynomial that qualifies as a generating function for a Poisson-binomial pdf and is closest, per some norm, to the generating function of the given, erroneous pdf data. (Experimental Presults seem to indicate that the l1 norm (i.e., kxk1 = k jxk j) may be generally better suited than the quadratic for addressing the inverse Poisson-binomial problem.) Such a polynomial should satisfy the constraints that its coefficients should constitute a valid pdf and that all of its roots (i.e., the sk of (3)) should be nonpositive real numbers. Another perspective can be gained by formulating the problem as that of finding the eigenvalues of a matrix that is a companion to the pdf’s generating function in the sense that the latter’s generating function equals, within a scale factor, the characteristic polynomial of the former. For example monic polynomials xN + aN¡1 xN¡1 + ¢ ¢ ¢ + a1 x + a0 can be associated to “companion matrices” of the form ¸ · T ¡a ¡a0 (47) Ac = I 0 where aT = [aN¡1 aN¡2 ¢ ¢ ¢ a1 ], I is an (N ¡ 1) £ (N ¡ 1) identity matrix, and 0 is an (N ¡ 1) element zero vector. It can be easily verified that the polynomial above is also the characteristic polynomial 814 of this matrix; hence the roots of the polynomial are the eigenvalues of the companion matrix. (This is the approach followed in the MATLAB function “roots”.) Since, theoretically, a polynomial’s roots are not affected by factoring out the coefficient of the polynomial’s highest order term, one can, in principle, use this eigenvalue method to solve for the roots of any polynomial. One can conceptually exploit such a fact to, under ideal circumstances, solve for the sk on the right-hand side of (3); then, since sk = ¡(1 ¡ pk )=pk , one can obtain the probabilities of success in the individual, independent Poisson trials as pk = 1=(1 ¡ sk ): (48) Unfortunately in practice this is an extremely numerically-unstable problem [13]. Depending on the values of the Poisson-binomial pdf and of the individual-trial probabilities that implicitly spawned it, even minute errors introduced by computer round-off may result in large eigenvalue displacements, and hence, in estimates for individual-trial probabilities that do not meet requirements. For example Fig. 3 shows the outcome of the simple test of running, in succession, the two routines of Fig. 1. First a list of N = 50 uniform random numbers between 0 and 1 is generated, and the corresponding Poisson-binomial pdf is computed by using the first routine. The second routine is then run using the output of the first as the input. Fig. 3 shows the results of this final step. The circles represent the randomly-generated individual-trial probabilities, while the Xs represent the output of the inverse Poisson-binomial process (i.e., the estimates of those individual-trial probabilities when given their Poisson-binomial pdf). As can be seen even with MATLAB’s fantastic precision, about half of those estimates are complex numbers. Reprising: perturbations of the elements of the Poisson-binomial pdf (i.e., of the coefficients of its generating function) can be viewed as perturbations of selected elements of its companion matrix which, even when infinitesimally small, can cause large eigenvalue excursions. The problem is, thus, how to solve for the probabilities of success of the individual trials when, due to measurement or experimental error, the perturbations of the pdf elements are anything but infinitesimal. Experiments suggest that a way of approaching this problem may be by determining and operating on a “more appropriate” companion matrix than the one defined by (47). Such a matrix could be constrained to correspond to a pdf’s generating function, have nonpositive real eigenvalues (e.g., it could be designed as Ac = ¡AP , where AP is a real, positive semi-definite symmetric matrix), and ensue from a “minimal” perturbation of the original data. That is, the inverse Poisson-binomial problem, as stated, would involve, in practice, the formulation (and solution) of a highly structured (and constrained) problem. IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 2 APRIL 2010 Fig. 3. Complex plot resulting from running in succession routines of Fig. 1 on an input consisting of list of N = 50 uniform random numbers between 0 and 1. But, albeit very interesting, the “inverse problem” just presented is not the one of most practical concern now, as one can seldom determine or specify all the entries of a particular pdf. Namely estimating these entries through experimentation may not be viable and, on the other hand, specifying them most certainly produce a pdf which is not of the persuasion desired. Thus its use in estimating/evaluating the probability of success of individual trials and/or in determining whether the underlying process is indeed Poisson-binomial may be mainly limited to experimental settings with tightly controlled conditions (e.g., see [5]). Most inverse problems of current interest involve the determination of some of the individual-trial probabilities of success that would be needed to achieve specific entries of either the Poisson-binomial pdf or of its cdf’s complement. For example in many applications, the “inverse problem” of interest is the one of Section VII: some of the individual-event probabilities are assumed known, and the objective is to estimate the probabilities (usually assumed constant) that should be achieved in the future so as to obtain a desired probability of “at least M successes out of N trials” for specified values of M and N (see [1], for example). As we see the techniques sketched in this paper can be used to attack this problem. Another inverse problem, very popular in reliability modeling circles (see [1]) and in radar engineering applications (e.g., for the binary integration of received pulses [3—4]), involves the determination of the optimal values of M and, at times, N. For example an important problem in reliability modeling involves solving for the M and N that, given the “reliabilities” (akin to the probabilities of success) of individual components (or “events”), minimize an objective function that accounts for the cost of each component and the cost of system failure. Regardless of the application of this particular “inverse” problem, however, and again for purposes of simplification, the individual probabilities of success are usually assumed to be equal across events. Finally since as part of an “inversion” operation one may need to obtain a pdf when given a cdf or its complement, we now present, for completeness’ sake, such relationship in simple matrix form. Beginning with the mapping from the pdf into/onto the cdf’s complement, we obtain HP = Q P = H ¡1 Q and/or (49) where (N + 1)-element vectors P and Q contain, respectively, the entries of the pdf and of the complement of the cdf. (For the Poisson-binomial case, the entries of these vectors would be arranged as follows. The nth entry of P represents the probability of exactly n ¡ 1 successes in N trials, while the nth entry of Q represents the probability of at least n ¡ 1 successes in N trials.) H and its inverse are (N + 1) £ (N + 1) matrices of the form 21 1 1 ¢¢¢ 1 13 60 1 1 ¢¢¢ 1 17 6 7 6 7 6 7 0 0 1 ¢ ¢ ¢ 1 1 H=6 7 6 .. .. .. .. .. 7 4 . . . ¢¢¢ . . 5 0 0 0 ¢¢¢ 0 1 2 1 ¡1 0 ¢ ¢ ¢ 0 0 3 6 0 1 ¡1 ¢ ¢ ¢ 0 0 7 6 7 6. . .. .. .. 7 ¡1 6 7 . . H =6. . . ¢¢¢ . . 7: 6 7 40 0 0 ¢ ¢ ¢ 1 ¡1 5 0 0 FERNÁNDEZ & WILLIAMS: CLOSED-FORM EXPRESSION FOR THE POISSON-BINOMIAL PDF 0 ¢¢¢ 0 (50) 1 815 Considering now the into/onto mapping between the pdf and its cdf, we have that since C = 1 ¡ Q, where 1 represents an (N + 1)-element vector of ones, and the entries of C, the (N + 1)-element cdf vector, are ordered (in the Poisson-binomial case) so that the nth entry corresponds to the probability of strictly less than n ¡ 1 successes in N trials, the relationship between P and C becomes 1 ¡ HP = C and/or P = H ¡1 (1 ¡ C): (51) REFERENCES [1] IX. SUMMARY This paper presents a methodology for obtaining closed-form representations for a number of Poisson-binomial expressions of common practical use. In a nutshell since these expressions can all be traced back to the sum of random variables drawn from binary distributions (Bernoulli-Poisson trials), they can be formulated in terms of the linear convolution of N two-element sequences. In turn the output of such convolution can be related to the coefficients of the polynomial that results from the product of monomials associated to the two-element sequences being convolved. By exploiting the Vandermonde polynomial interpolation approach, all these operations can then be written as part of a single expression that consists of a finite sum of trigonometric products. Particular expressions derived in this paper using this methodology are the closed-form formulations for the Poisson-binomial pdf and for the complement of its cdf ((5) and (11), respectively); the new trigonometric identities for the binomial coefficient and the binomial cdf ((16) and (24a)—(24c), respectively); and (31), which provides the Poisson-binomial moments. Furthermore since going from the binary distributions to the Poisson-binomial expressions can be conceptualized as the product of monomials, it is shown that the binary distributions of the individual trials can be obtained from the Poisson-binomial expressions by solving for the monomials’ coefficients or, equivalently, for the roots of the polynomial that correspond to the Poisson-binomial expression of interest. The MATLAB routine of Fig. 1(b) is provided to solve this inverse Poisson-binomial problem for relatively small-sized cases (less than 25 individual trials), and stability issues are discussed, which suggests that such an inverse problem could be made more robust by formulating it in terms of finding the eigenvalues of a real, symmetric, positive semi-definite companion matrix. 816 The just-described procedures, which implementation-wise are very attractive, can be used as tools for signal detection and tracking [2—4] and for evaluating the fusion of the outcomes of different processing strings [5]. And even a cursory look suffices to show their all-encompassing applicability to the various fields of reliability optimization theory [1]. [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] Kuo, W., and Zuo, M. Optimal Reliability Modeling. Hoboken, NJ: Wiley, 2003. Blackman, S. Multiple-Target Tracking with Radar Application. Norwood, MA: Artech House, 1986. Shnidman, D. Binary integration for Swerling target fluctuations. IEEE Transactions on Aerospace and Electronics Systems, 34, 3 (1998), 1043—1053. Frey, T. An approximation for the optimum binary integration threshold for Swerling II targets. IEEE Transactions on Aerospace and Electronics Systems, 32, 3 (1996), 1181—1185. Fernández, M., and Aridgides, A. Measures for evaluating sea-mine identification processing performance and the enhancements provided by fusing multisensor/multiprocess data via an M-out-of-N voting scheme. Proceedings of SPIE, vol. 5089, 2003. Feller, W. An Introduction to Probability Theory and Its Applications, vol. I. Wiley Series in Probability and Mathematical Statistics, New York: Wiley, 1968. McClellan, J., and Rader, C. Number Theory in Digital Signal Processing. Upper Saddle River, NJ: Prentice-Hall, 1979. Belfore, L. An O(n(log2 (n))2 ) algorithm for computing the reliability of k-out-of-n : G & k-to-l-out-of-n : G systems. IEEE Transactions on Reliability, 44, 1 (1995), 132—136. Radke, G., and Evanoff, J. A fast recursive algorithm to compute the probability of M-out-of-N events. In Proceedings of the Annual Reliability and Maintainability Symposium, 1994, 114—117. Chen, S., and Liu, J. Statistical applications of the Poisson-binomial and conditional Bernoulli distributions. Statistica Sinica, 7 (1997), 875—892. Bracewell, R. The Fourier Transform and Its Applications. Columbus, OH: McGraw-Hill, 1999. http://binomial.csuhayward.edu/Identities.html. http://en.wikipedia.org/wiki/Wilkinson’s polynomial. IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 2 APRIL 2010 Manuel Fernández (M’98) obtained his B.S.E.E. from the Georgia Institute of Technology, Atlanta, in 1979 and his M.S.E.E. and Eng.E.E. degrees from Stanford University, Stanford, CA, in 1980 and 1982. In 1983 he joined General Electric’s Electronic Laboratory in Syracuse, NY, as a research engineer, and in 1990 he became part of the GE Aerospace Advanced Engineering department (currently Lockheed Martin—MS2 Strategic Research and Technology Development group). His areas of interest include, among others, linear and nonlinear adaptive processing, signal detection and classification, and mission and sensor management. Stuart Williams received his BSEE from John Brown University, Siloam Springs, AR, in 1981 and his M.S.E.E. from Syracuse University, Syracuse, NY, in 1984. He held a variety of radar and sonar System Engineering positions for General Electric Company Aerospace from 1981—1992, primarily pursuing new business opportunities for the Advanced Development Engineering group. He worked as a Senior System Engineer for Lockheed Martin Corporation from 1992—2007, with primary emphasis on ground based radar applications, including the development of sensor management techniques. Since 2007 he has worked for Sensis Corporation in the area of long range expeditionary radar. FERNÁNDEZ & WILLIAMS: CLOSED-FORM EXPRESSION FOR THE POISSON-BINOMIAL PDF 817