■ I LIBRARY OF THE UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN 510.84 no. 131 -140 cop. 3 The person charging this material is re- sponsible for its return to the library from which it was withdrawn on or before the Latest Date stamped below. Theft, mutilation, and underlining of books are reasons for disciplinary action and may result in dismissal from the University. To renew call Telephone Center, 333-8400 UNIVERSITY OF ILLINOIS LIBRARY AT URBANA-CHAMPAIGN NOV oi 1*2 SR> 1 9 id % L161— O-1096 Digitized by the Internet Archive in 2013 http://archive.org/details/stationarydistri136shiv "X-£*br DIGITAL COMPUTER LABORATORY 7lO>!34> UNIVERSITY OF ILLINOIS C ° F' 3 URBANA, ILLINOIS REPORT NO. 136 STATIONARY DISTRIBUTION OF PARTIAL REMAINDERS IN S-R-T DIGITAL DIVISION Richard Robert Shively May 15, 1963 (This work is being submitted in partial fulfillment of the requirements for the Degree of Doctor of Philosophy in Electrical Engineering, May 1963). DIGITAL COMPUTER LABORATORY UNIVERSITY OF ILLINOIS URBANA, ILLINOIS REPORT NO. 136 STATIONARY DISTRIBUTION OF PARTIAL REMAINDERS IN S-R-T DIGITAL DIVISION by Richard Robert Shively May 15, 1963 (This work is "being submitted in partial fulfillment of the requirements for the Degree of Doctor of Philosophy in Electrical Engineering, May 1963). ,, ACKNOWLEDGMENT The author wishes to express his gratitude to Professor J. E. Robertson for certain important observations^ and for his encouragement and guidance in pursuing this work. The author is also indebted to Professor G. A. Metze for his invaluable counsel at several points during this study. The friendly help of Mrs. Phyllis Olson who typed the thesis in its final form is sincerely appreciated. TABLE OF CONTENTS BACKGROUND Page 1.1 Introduction ...................... 1 1.2 S-R-T Division ..................... 2 1.3 Division as a Markov Process .............. 5 1.3-1 Existence and Uniqueness of a Stationary- Distribution .................. 5 1.3-2 Formulation of the Chain ............. 10 1.3.3 Examples ..................... 12 1.3-4 Limitations of the Markov Chain Approach ..... 14 2. DERIVATION OF PARTIAL REMAINDER DENSITIES .......... l6 2.1 The Defining Equation .................. l6 2.2 End Point Relations ................... 22 2.3 Symmetry ........................ 23 2.4 Corollaries to Symmetry ................. 28 2.5 Generation of Solutions ................. 31 2.5.1 Properties of Solution Intervals ......... 44 2.5.2 Comments on the Set (D„) ............. 49 E 3. SHIFT AVERAGE ........................ 52 3.1 Equation for .................... 52 3.2 Evaluation of for ^ < D < 1 ........... 54 3.3 f or I < D < I . ? . .............. . 60 4. SUMMARY AND CONCLUSIONS ................... 68 4.1 Summary ......................... 68 4.2 Conclusions and Areas for Future Investigation ..... 70 BIBLIOGRAPHY ........................... 73 ■IV- STATIONARY DISTRIBUTION OF PARTIAL REMAINDERS IN S-R-T DIGITAL DIVISION Richard Robert Shively Department of Electrical Engineering University of Illinois, 19&3 The execution time required for digital division in electronic computers can be reduced if a) more than a minimum number of divisor multiples are available, and b) provisions exist to either by-pass or shorten steps which generate zero quotient digits. A division algorithm which uses this redundancy to increase the likelihood of zero digits (i.e., when the quotient is expressed in the recoded form implied by the divisor selections) then is desirable. To provide a basis for comparing various division algorithms in this respect, C. V. Freiman observed, using a Markov chain model for the division process, that for any given divisor, D, and a randomly distributed set of dividends, the distribution of partial remainders at successive steps approaches a steady-state or stationary distribution, which is independent of the distribu- tion of dividends. In attempting to determine this stationary distribution as a function of D for an algorithm known as S-R-T division, Freiman found that the pointwise solutions provided by a Markov chain are inadequate to analyze the function for certain intervals of D. The purposes in further pursuing this work started by Freiman are: a) to obtain a more complete understanding of S-R-T division, b ) to derive methods which might be found applicable in analyzing other division algorithms, and c) to provide information which could prove useful in formulating new division algorithms. If partial remainders, R., are defined to be in one-to-one corres- pondence with quotient digits (rather than considering only those associated with nonzero quotient digits), the stationary probability density, F(x), which describes R., can be shown to be symmetric with respect to the point D, F(D), for all D in the allowed range, — < D < 1. This symmetry can be used to derive a straightforward method for evaluating the stationary density for intervals in the divisor range, rather than for individual points „ The method essentially consists of locating points at which F(x) is. discontinuous, as an ordered process, starting with points of discontinuity in the relation between R. and R. , and terminating when the first such discontinuity point falls in the interval (— , 2D - p). Then, by observing that the jump sizes at successive points in this ordered set of argument values form a geometric progression, and that the sum of these jumps is — , the solution follows directly. Solution intervals, i.e., neighborhoods in the range of D for which F(x) has similar form, may be grouped into infinite families, based on the sequence of operations used to locate discontinuity points, and when considered in the order of increasing complexity, these families form a nested pattern,, The stationary densities describing R. can be used to determine the shift average, , i.e., the average number of quotient digits formed per use of the adder. is equivalent to the mean recurrence time of the event "R. > — " in S-R-T division, and is therefore the reciprocal of the probability 1 2D that R. > p. Using properties of F(x), can be shown to be — =■ for 0.75 < D < 1, reaching a maximum of 3.? over the interval 0.6 < D < 0.75 • In the interval 0.5 < D < 0.6, requires an infinite number of sub intervals to express completely, and exhibits a nonmonotonic behavior at certain points. Analysis of other division algorithms can use the fact that a) the piecewise constant functions which describe the stationary density can be formed by locating discontinuity points as an ordered process, and b) the sizes of jumps at successive arguments in this process form a geometric progression. 1 . BACKGROUND 1.1 Introduction Numerous factors are effecting a willingness to invest in redundant equipment to improve digital computer operation times. Reliability and decreasing cost of components, modularization of circuits, and miniaturization are but a few of these. In arithmetic processes, some commonly considered applications of redundancy within the structure of a general purpose computer are: a) redundant number representation to reduce or eliminate adder carry propagation time, b) multiple shift paths to reduce the number of steps required in iterative processes such as multiplication and division, and c) generation of more than the minimum number of multiples of the operand, again to improve multiplication and division. In digital division, redundancy in the generation of divisor multiples may be used to serve either or both of two general aims: a) providing overlap in the partial remainder ranges to which the respective divisor multiples are assigned, and b) minimizing the expectation value of the partial remainder magnitude. The first of these is particularly important when signed digits or stored carries are used, since distinguishing between nonoverlapping intervals by inspection of leading digits of unassimilated numbers would be impossible. Reduction in the partial remainder magnitude can, under certain circumstances, reduce the time required for the division process. (This will be discussed in Section 1.2. ) The effectiveness of a particular digital division algorithm (and the redundant equipment it may require) in reducing the average partial remainder, R., can be measured if the probability distribution (or density) functions Hereafter, R. will be used to denote the magnitude of the partial remainder occurring on the ith step of a division process. -1- -2- which describe R. are known. In Section 1.3 it will be shown, using the theory of finite Markov chains, that for a particular divisor, D, the distribution describing R. converges with increasing i toward a unique steady-state dis- tribution. Characterizing digital division as a discrete Markov process is due to C. V. Freiman (see Ref. 4). As Freiman discovered, however, the complicated behavior of the steady-state distribution of R., for certain ranges of the divisor, is not amenable to investigation using Markov chain techniques alone „ The purpose of the dissertation is to derive a method for more readily finding the steady-state distribution of R., assuming the S-R-T division algorithm is being used, and to apply this method in examining the function's behavior in those ranges of D where detailed analysis was previously impractic- able. Other than satisfying a curiosity, the worth of this information has been 2 demonstrated by Professor Gemot Metze who, using properties indicated in previous efforts and proved herein, has proposed a modification to S-R-T divi- sion which yields, for all values of the divisor, D, the maximum shift average (i.e., maximum number of quotient digits per use of adder) obtainable under the constraint that the available divisor multiples are +D, 0, and -D. 1.2 S-R-T Division Descriptions of conventional restoring and nonrestoring division can 3 be found in various texts on digital computer arithmetic. By common usage, the phrase "S-R-T division" connotates what may be regarded as nonrestoring, binary division modified to utilize the facilities of 2 ' See Ref. No. 7. 3 J e.g., Ref. No. 9. -3- k 5 6 a floating-point computer. Sweeney, Robertson, and Tocher independently- observed that if the absolute value of the fractional portion of the divisor, 7 D, is restricted to the normalized range .5 < D < 1, then any steps in which the partial remainder magnitude is found to have leading zeros can be by-passed, and zeros inserted as the corresponding quotient digits. (After being recoded from a representation using +1, 0, and -1 as digits to binary, the quotient digits just mentioned will be all zeros or all ones, depending on whether the next lower weighted nonzero quotient digit is positive or negative.) This is equivalent to saying we will always normalize the partial remainder between additions, and is profitable if paths to perform multiple shifts exist, or if the time for a given step is measurably dependent on whether or not that step requires use of the adder. Under these circumstances the average partial remainder magnitude is of interest, since the average number of leading zeros in R. is equal to the average number of quotient digits formed per addition in the divison process. To state the S-R-T division algorithm formally, the following defini- te tions will be useful: a) P. = the partial remainder after the ith step, where i = 0, 1, . . ., m-1 b) R. = |p. I l ' i ' c) D = divisor h Ref. No. 2. ? Ref. No. 10 6 Ref. No. 11 7 Inclusion of end points is dependent on choice of number representation. 1 8 T In each case, the fraction portion of a floating-point number is implied, -k- d) q = the quotient digit generated by the ith step, with possible values of +1, 0, or -1 m-1 _• e ) Q = quot ient Z 2 q . i=0 x One way of stating the recursion relation is as follows: l) If R. < ==, then P. , = 2P. l 2' l+l i q ± = 2 ) If R . > ^ and 1 — 2 a) the signs of P. and D agree, then P. ., = 2(P. - D) b) the signs of P. and D disagree, the P. , =2(P. +D) and l i+l v l q ± = -i Since the ranges of — < P < 1 and — < D < 1 would imply — < Q < 2, corrective initial and/or terminal steps may be required to obtain a quotient for which Q is a proper fraction. It can be proved by induction that the above process yields the correct quotient, since: P. . - 2(P. - q.D) (2.1) i+l v l ^i v ' and therefore after m steps are completed m-1 P_ = 2" m P + D • Z q. • 2 _1 (2.2) m . _ i i=0 or P = QD + 2 _m P (2.3) m -5- 1.3 Division as a Markov Process The partial remainder in a division process becomes a random variable if either or both of the dividend and divisor are random variables. A useful and somewhat surprising relation is that for any fixed divisor, the distribution of R. is independent of the distribution of dividends, for i suff icientely large, provided variation in the dividend density is bounded. 1.3=1 Existence and Uniqueness of a Stationary Distribution In order to demonstrate that R. has a unique stationary distribution Q in the limit, Freiman observed that the division process may be characterized as a finite, regular (or aperiodic and irreducible) Markov chain. The states of this Markov model are intervals in the range of R. , each of which satisfies i' the following requirements: a) The probability density function describing R. is uniform within the interval for all i. That is, if x n and x are the end points of such an interval, Pr[x < R < x + dx] - K • dx, for x < x < x (3.l) where K is a constant for any particular i, but may vary with io b) One such interval in the distribution of R. will map into an integral number of intervals in the distribution of R. . . for all i. This means no new sub intervals are l+l' required in order to insure that (a) is satisfied at each step. g Reference k 10 Proof that such states can be generated is deferred to Section 1.3-2 -6- Denote these states (S.), j = 1, 2, . .., N, and let X(i), J i = 0, 1, 2, . .., represent the stochastic process which ranges over this state space, where i corresponds to the step number, and X(p) = q denotes the event "R is in state S ." The Markovian character of X(i) is a result of the two P q . properties of all states listed above. If X(i) = m is know to be true, the probability that X(i + l) = n, for example, is not altered by information regarding which state X(i - l) assumed, because, due to the uniformity of the density function within an interval, this information would tell us nothing more about the probability that R. is in the subset of S which maps into S . By l m n induction, the random variable X(i + l) is also independent of all X(p), p < (i - l), as well, or: Pr[X(i + l) = n|x(i) = m, X(i - l) = k, ..., X(o) - a] (3.2) = Pr[X(i + l) = n|x(i) = m] The rather trivial observation that the division algorithm is invariant from one step to the next enables further classifying the stochastic process as a stationary Markov chain, i.e., conditional probabilities Pr[x(i + l) = n|X(i) = m] are transition probabilities which are independent of i. The procedure for finding the distribution of R. , given that of R. (or equivalently, finding the distribution of the discrete random variable X(i + l), given the distribution of X(i)) can be formally expressed by use of a transition matrix, P, with dimen- sions NXN for N states, and which has as its row m, column n entry the transition probability Pr[X(i + l) = n|x(i) = m], i.e., rows are identified with sources, columns with destinations. Then if P is premultiplied by an N-entry row vector, Jt(i), which describes the distribution of X(i), i.e., element Jt. (i) = Pr[X(i) = m], the product is Jt(i + l). Since P is independent of i, this result can be generalized to: ir(i + k) = *(i) • P k ; k = 1, 2, ... (3-3) which expresses a k-step transition. The desired result is that for i sufficiently large, fl(i) approaches a unique final distribution which is independent of u(o). If X denotes this limit distribution, then X = lim(*(o) ■ P k ) (3.^) k-»o would be true for any 1 x N probability vector, rt(0). Equation (3-^) would obviously be true if each row of P approached X as a limit, since the sum of the elements in rt(0) equals 1. We can prove that is precisely what the limiting behavior of P is, and can state as a Theorem: If the entire interval < R. < 1 can be covered with a finite number 1 of intervals, each of which has the two requisite properties of a state listed at the start of I.3.I, then there exists a unique vector X such that equation (3»*0 is true, for any tt(0)} each row of P approaches X as a limit. Proof: 1. The conclusion holds if the chain is irreducible and aperiodic (e.g., see Feller, p. 356, or Kemeny, et . al., Chapt . 6). 2. Both properties are defined to apply if there exists a finite integer B, such that for all k > B and all state pairs S , S , there is a nonzero n 7 nr probability of the system being in state S after k steps, having started in state S . (This applies to the case n = m as well.) Equivalently, the properties apply if all elements of P are positive for k > B. 3. To show that B exists, we refer to equation (2.l), which, if attention is confined to absolute values, is represented in Figure 1. The proof is completed with four observations : a) In the range of R., any interval of length < — which does not include the points R. = .5 or R. = D ; projects onto an interval twice its length in the range of R. . The same comment applies to descendents, until one or both of .5 and D are covered. b) This doubling of the image with each step applies also to an interval which includes R. = .5, provided R. = (D - .5) is not also included (see Figure l). When both of these points are included, clearly D is included in the projected image on the range of R. , to which observation (c) applies . c) An interval in the range of R. containing D has a projected image expressible as < R. < 5, with 5 > 0; on successive iterations the image will be (0, 2 • 5), n - 1, 2, . .. until the entire range of the partial remainder magnitude, (0,l), is covered. d) The first three observations established that, having started in any state (nonzero interval), the system can reach any other state, i.e., irreducibility. The fact that the last, single image of the sequence listed in (c) covers all of (0, l) proves the existence of the integer B, since this is equivalent to all elements 11 Figure 1 derives from equation (2.l), together with R. - |P. | i+l 1/2 Ri FIGURE 1' Rj +1 vs. Rj -10- of P being positive for some k, Clearly all higher powers of P must also contain only positive elements., and hence the chain is aperiodic. The last observation leads to an immediate Corollary : All elements of A, are positive. The fact that the stationary density function is nowhere zero in the unit interval will be used in Chapter 2. 1.3.2 Formulation of the Chain First it will be shown that it is possible to generate a set of intervals covering (0, l) which meet the two requirements of states listed at the start of 1.3.1; for a large class of dividend distributions. With this established, the conditional hypothesis of the theorem in 1„3°1 can be removed, and for computational simplicity, the states may be selected on the assumption that the distribution of dividends is uniform, since the final result is independent of the actual distribution of dividends. The states of the Markov chain for a given rational divisor, D, can be selected as follows: a) While otherwise arbitrary, the density function describing dividend magnitudes is assumed to be sufficiently well behaved that it can be approximated in measure to any specified accuracy by a piecewise uniform density, with all discontinuities in the approximation occurring at rational numbers. (in particular, the density is everywhere finite; there is no single point in the range R n to which a non- zero probability is assigned. ) Denoting the point set consisting of the divisor D and the arguments at which ■ 11- the discontinuities occur as -s — r (each point represented in reduced fractional form), we find the least common denominator. Denote this number as L. b) The intervals, each L in length, defined by the integral multiples of L , are then the states . Due to the definition of L in (a), the density function for dividends is uniform within any such interval. Discon- tinuity arguments for the density functions of R., i = 1, 2 t . .., all are either images of these initial discontinuities, or of the points R = .5 or R = 1. (See Figure 1; the latter two are the points in the range of R. at which the R. n versus R. relation is to l l+l l discontinuous.) Since each such image is expressible in terms of integral multiples, and sums and differences, of elements in the set -{- — r s it is clear that density function uniformity within the specified intervals is preserved, and that the set of interval end points maps only onto itself. Other than showing that the hypothesis in the theorem in 1.3-1 is satisfied, the above formulation reveals another interesting property of the stationary density function, viz., all discontinuities are images of the points .5 or 1, since the uniqueness property implies none of the discontinuities in the R density and hence none of their images can influence the limiting density. -12- 1.3-3 Examples At this point two examples may be found helpful. We know that successive distributions of X(i) are related by the equation: *(i + 1) = K(i) • P (3.5) for all i, and that for i sufficiently large: «(i) = *(i + 1) = X (3.6) Hence ^ the solution for a given D requires finding P, then solving the equation set \ = XP (3-7) To determine state boundaries we generate the images of the points x = .5 and x = 1 by application of equation (2.l) or Figure 1, stopping when repetition begins. To determine the n,m element of the transition matrix, P, we answer the question: Given an ensemble of partial remainder uniformly distributed in state S , what fraction of them will be in S after one step? n m 7 Example 1: D = ^ a) Images of the point R. - — are: (i) 2(D- |) -| (ii) 2[D - (2D - l)] = 2 - 2D = - o (iii) 2[2 - 2D] = k - ^D = 2 (iv) 2[k - k-D - D] = 8 - 10D = | Since the next application of equation (2.l) yields k q, which is also the first image point of 1, all 7 -13- potential locations of discontinuities have been determined. The states are therefore : 2 ? D b) The transition matrix then is P = .5 .5 • 25 • 75 • 5 .66... .33... l c) Solving the equation set X = XP, we obtain: X lA2 X l3*lA X 15 : [£ g 1 _3 _1 1 L 7 7 7 lU lV Example 2: D = ^ a) Images of »5 are: (i) 2(D -|) -i (ii) 2(i') = | (in) M§) = | (iv) 2(| - D) = | It The first image of 1 is also — , 5 The states are therefore: -14- o D b) P = • 25 1 1 .25 U ll K 12 \l \k ] = [ 3 I 3 12 ] 1.3.4 Limitations of the Markov Chain Approach The Markov model has been very useful both conceptually and as a vehicle for proving the existence and uniqueness of a stationary probability distribution for partial remainders. However, when attempting to use the Markov chain to solve for this distribution, two significant obstacles are encountered : a) The size of the linear equation set to be solved is governed by the value of the denominator of the divisor, D, rather than by the complexity of the density function being sought. This is a result of the definition of states which requires the set of end points to be closed under the recursion relation. Thus, in example 1 above, we find by normalizing the probability of each interval with respect to the length of the interval, the prob- ability density function has only one discontinuity, although five equations were required (see Figure 2). -15- PR[x 1 0, x < (1.1) where ) f.(y) = Pr[R. < y J( for a given divisor D; b) fj(y) =< t ± (j), T>\ f t (f), J<\ (1.2) c) | < D < 1 This relation can be demonstrated graphically as in Figure 3> where the intervals of R. which yield an R. , less than a specified number, x, are indicated. In l J l+l * ' ' the limit, if distributions of successive partial remainders are equal, let f(x) ■ lim[f .(x)] i— »oo (1.3) Then we have equation (l.l) with subscripts deleted -16- ■17- Ri+i X/2 I /2 (D-X/2) (O+X/2) Ri 1 FIGURE 3= .«x)= f(x/2)+f(D+x) -f(D-X/2) -18- f(|) + f (D + §) - f*(D - §), < x < 1 f(x) = "S 1, x > 0, x < (1A) to which (1.2) applies, again with subscripts deleted. Because attention is restricted to piecewise uniform distributions, i.e., piecewise linear distribution functions, f(x) is differentiable at all but a finite number of points. Therefore, to find the density function, we can take the derivative of both sides of equation (l.h) (to simplify notation, let F(x) = f(x)): where r F(x) = |[F(|) + F(D +§) + F*(D - §)], 0i F*(y) = < 0, y 2 Since F(x)dx has the interpretation: Pr[x < R. < x + dx], equation (1.5) is expressing the probability of a partial remainder being in any differential interval, say dx n , as the sum of probabilities that its predecessor was in one of the (at most three) mutually exclusive differential intervals, each half as long as dx , which map into dx». Hence we have, in terms of a function whose -19- domain is a continuum, the relation that was expressed in terms of discrete state probabilities by equation 1:(3.T)» The theory of Markov chains can be applied to demonstrate that a piecewise constant density function which satisfies (1.5) for a given D is unique. As in Chapter 1, generality will be slightly compromised by assuming D is rational. The first theorem below is instrumental in the proof of the second. Theorem 2 . la ) : For D rational, and F(x) a piecewise constant density function satisfying (l.5)> all discontinuities in F(x) occur at rational values of x. Proof : 1= To construct a contradiction, assume there exists an irrational number, x~, at which the conclusion is violated. 2, Since the number of discontinuities occurring at irrational arguments must be the same for both sides of (1.5 )j, jumps in two different terms on the right of that equation cannot cancel each other for x irrational. (As demonstrated later, such mutual cancellation can occur for x rational.) Hence, (1.5) implies that a second irrational discontinuity point exists at one of 2x , 2D - 2x n , or 2x„ - 2D, depending on which of (0, — ), (— , D), or (D, l) includes x~. 3. By induction, the second such discontinuity implies the existence of a third, etc. The n'th term in the sequence is obviously expressible as: x =+2 n • x^ + C (1.6) n - where C is some rational number (because D is rational). k, Since a piecewise constant function is defined to have at most a finite number of discontinuities, the sequence described by (1.6) must cycle. In the mathematical model, representation of any partial remainder, rational or irrational, is assumed possible. -20- If x is a point in this cycle, and recurs after k steps, we could infer from (1.6): C x^ = l + 2 k which is impossible with x irrational. Theorem 2.1h ) : There is one, and only one, function which satisfies equation (1.5) for a given rational divisor, D, under the implied restrictions of a piece- wise uniform probability distribution: (i) / F(x)dx =1, (ii) F(x) is piece- wise constant, and (iii) F(x) > for all x. Proof : 1. At least one solution exists, since the vector representing state probabilities of a Markov chain (i.e., the solution to equation 1:(3.2)), can alternatively be expressed as a density which must satisfy (1.5) at each point. 2. Conversely, a solution to (1.5 ) for a given divisor must be unique if the function can alternatively be expressed in terms of state probabilities of 2 a finite Markov chain, i.e., if there exists a partitioning of the remainder range such that the distribution is uniform within each interval, and the finite partition-point set is closed under the S-R-T recursion relation. Then, for two alleged solutions to (1.5 ) with D fixed, we could select a single partitioning (viz., the union of two partition point sets which apply to the solutions individually) for both functions, and thereby construct a contradiction based on the uniqueness of regular Markov chain state probabilities. 2 While this uniqueness extends to regular Markov chains with a denumerably infinite number of states, the a priori knowledge that the number of discon- tinuities is finite, gained by restricting D to the set of rational numbers, is useful in the subsequent development (see Ref. 1, Chapt. 1 for properties of chains with an infinite number of states). -21- 3. To select the partition point set hypothesized in 2, we observe that the number of discontinuities is finite, and all occur at rational numbers due to theorem 2.1a. Hence, the set \t\} n = 0, 1, . .., L, where L is the least common denominator of the divisor and discontinuity arguments, is always sufficient for the partition point set described in step 2 above, Q.E.D. The problem of finding the stationary density of partial remainders is now reduced to solving (1.5) for F(x), as a function of D. Since no single number in the range of partial remainders has a nonzero probability, the defini- tion of F(x) for x = 0, 1, or any of the finite number of discontinuity arguments will not be a primary concern. Moreover, D will be restricted to the open interval, — < D < 1, both for expediency in derivations, and because F(x) for D = — is known; to wit: F(x)=l, < x < 1. The definition of a single step in the S-R-T division process used 3 here does not agree with that used by Freiman. Instead of defining a step to be either one binary shift (i.e., if q. = 0) or one addition and shift, Freiman defines a step to be one addition and partial remainder normalization. If g(x) denotes the steady-state distribution of the normalized partial remainder magnitudes, it is apparent from what computer normalization entails that: ' 00 E [f(— ) - f(— )]; i< x < 1 n n 2 n=0 2 2 g(x) = < (1.7) x < I from which we obtain: 3 Ref . 4. -22- *(x) = < io? F( ^ ^ for all D in the (— , 1) interval. Proof ; To summarize what follows, F is extended to a function with period 2D, and it is shown that the even portion (excluding an additive constant) of the convergent Fourier series representation of this extension must sum to zero for all x, which, due to Fourier series uniqueness, implies each cosine term is identically zero. The conclusion is an immediate sequel. 1. Define r F(x), < x < 2D H(x) = < (3.1) v F(X M0D.2D^ othe ™ ise Due to the fact that F(x) is a piecewise constant function for all D and finite for all x, F(x) has bounded variation. Therefore, the Dirichlet -Jordan theorem applies when formulating a Fourier series representation of H(x). If we define: h e.g., see Zygmund; Theorem 8.1, Chapt . II < ■2k- r2D b n = ij sin(^) - F(x)dx (3.2) n =±J cos(^) • F(x)dx (3.3) such that the series corresponding to H(x) is: S[H(x)] = -| + S K sin O + a n cos Of (3 ° U) n=l ^ -* then the Dirichlet- Jordan theorem states that S[H(x)] converges uniformly to H(x) within any closed interval in which H(x) has no discontinuities, and, moreover, if x is the argument of a discontinuity S[H(x Q )] = ||h(x +) + H(x -)| (3»5) Therefore the notational distinction between H(x) and its series representation is unnecessary if we define H(x) (and F(x)) as the average of left- and right-hand limits at each jump,, 2. Since the period is 2D, and there is unit area under F(x), we know the constant of the series : ? - | B (3-6) To evaluate a , n > 0, the expression for F(x) in (l„5) may be substituted into (3.3). ■25- a = 2D, cos(— ) 1 f2D F(|)dx + ±J cos(^)F(D + |)dx (3 = 7) 2D-1 2D, ,nnx cos(^)F*(D - f)dx The limits of integration in (3° 7) are determined by the intervals through ■which x may vary in the respective terms on the right of (l„5)° The upper limit on the second integral may he any value between 2 - 2D and 2D inclusive ^ all other limits are fixed „ The flexibility of this one limit is due to H(D + — ) being zero for that interval of x ; together with the fact that it was unnecessary to impose any nonlinearity on the argument of the F(D + — ) term in (l„5). (The equation defining F(x) could alternatively be written as ° F(x) - |[F A (|) + F(D + §) + F B (D - |)], | < D < 1 (3.8) where F(y), y < | F A (y) = i 0^ otherwise F B (y) = { F(y), y>§ 0y otherwise no other restrictions on the range of x are necessary, ) Substituting the series representation for F in the integrands of (3°7);> we obtain the following expressions: -26- .) First term on the right of (3-7 ): p 1 °° r 1 / 1 /iwxv 1 v r / V\ flx / 2D J 2D COs( — } + 2 5, l a v [c ° s(n + 2 } ~ + C03(n 2 ; D ■ + b v Ls-in(n + -) — + sin(n - -) -^-J f dx (3.9) 1 . /nit. 1 1 vi tt — sin— +y-^a r , + i — Lia 4Dnit v D ' 4D 2n 4it V=l sm(n + -)^ sin(n - ^ - V n + 2 V n " 2 ri - + b / Vxlt cos(n + -)- V n + - 1 - / VnIT cos(n - -)-• V n - g where a prime on a summation symbol means the index, v, assumes all positive integer values excluding 2n for terms in which (n - is a factor. The same notation is used below, b) Second term on the right of (3. 7) is: Vn-1 r 2D oo r 1 / 1 /nitxv 1 v / -, \VJ 2dJ 2D COs( — > + 2 ^"^r / VvItX / cos(n + ^)-^r + cos(n VvItX 2' D . / V s-KX . , V » Jtx b v | sin(n + -y~ + sin(n - -)— dx (3-10) 1 a 1 v. b fi - (-i) v . i - (-D 2 a 2n " ^ t| % D v 1 V V V=l ^ n + - n - - n is constant, V is the index of the series. 6 1 Trigonometric identities used are: cosA cosB = —[cos (A + B) + cos (A - B)] cosA sinB = -[sin(A + B) + sin(A - B)] -27- c) Third term on the right of (3. 7) r2D-l . 00 1 / ,nrtx N 1 v / . O 2D J cos(— ) + - Z(-l) cos(n + -)-g- + cos(n - -)— - b sin(n + -)-g- + sin(n - gj-jj- dx (3.11) sin(^) a gn (2D - l) 1 . kDnx + 4D + Iwt ~ fsin(n + g-J^ sin(n - ^"l . + . n + L J + b cos(n + -)- V n + 2 (-1) cos(n - -)- + ■ \V - (-l)l V n - g Summing the expressions developed in (3 = 9)^ (3<10)<, an d (3«ll)j we find that (3-7) reduces to: a = a^ n 2n (3.12) 3° As mentioned earlier, the variation per period of H(x) is known to be bounded. For variation V per period of length 2D,, it can be shown that d. v. n ' — nD (3.13) Now, -by induction, (3.12 ) implies that a can be equated to any term of the form: a, , k = 2- ° n, where j is an arbitrarily large integer . Hence , (3«12) and (3.13) together imply that every |a j, n - 1, 2, is less e.g., see Zygmund, Theorem ^„12, Chapt. 2„ He derives the bound on a for a function with period 2ft, but the modifications to his derivation are straightforward . -28- than any specified positive number. The question of summing an infinite number of arbitrarily small numbers is resolved by applying Cauchy's con- dition for convergence, i.e., given e > 0, there exists an integer N such that all partial sums of the series after the Nth are within e of the limit 00 00 sum. Because E a cosf-7— ) < E a , and a I can be shown to be less n x D — ' n" ' n 1 than e/N by use of (3.12) and (3.13), it follows that: |Ea n cos(^)| < E (3.14) Since e can be arbitrarily small, the cosine portion of the series must be zero for all x. For any integer n, sin(— — ) is symmetric with respect to D, and there- for any sum of" terms of this form is also symmetric. Q.E.D. The significance of being able to select 2D as the upper limit on the second integral in (3«7) is now apparent. If any other limit in the allowed range of [2 - 2D, 2D] were used, the expression for the unknown, a , would be in terms of an infinite number of unknowns, the a and b , which, though correct, would have presented a more formidable problem than equation (3. 12). 2.4 Corollaries to Symmetry The symmetry of F(x) with respect to D will be used frequently in the remainder of this chapter, as well as the next. The properties listed in this section are some of the more immediate consequences of symmetry. Theorem 2.4a : (i) F(D) -i (4.1) (ii) F(x) =|, 0 0, i.e.^ there is no x > such that F(x Q +) - F(x Q -} > (if .7) P roof ' (A discontinuity described by (4.7) will be termed a positive jump> if the inequality is reA/ersed,, it is a negative jump, ) 1. Considered individually^ the respective terms on the right of equation (l.5| would tend to yield a positive jump in F(x)., for x = x 3 if there were: a) a positive jump at x /2 b) a positive jump at D + X n /2 3 for x < 2 - 2D, or c) a negatiAre jump at D - x r /2 } for x_< 2D - 1„ Due to {h,2) s the last of these is impossible „ Therefore, any positive jump implies the existence of another, 2. Both la) and lb) above cannot be true for the same value of x oj) since the number of positive jumps on the right in (1.5) would not then be the sum of the numbers of such jumps in the individual terms , which would lead to an inconsistency, since (1,5) is expressing F(x) in terms of itself. 3. Assume x. > is the argument of the largest positive jump (i.e., the difference in (4.7) is maximum). The contradiction is immediate., since (1.5) together with the above comments require the existence of a positive jump twice as large. Hence,, there are none . •31- 2„5 Generation of Solutions Due to equations (^.2) and (k.6), together with the theorem just proved, we can confirm that the stationary partial remainder density functions 3 for all divisors in the range r- < D < 1 are : fl D ' < x < 2D - 1 F(x) = -J | D, 2D - 1 < x < 1 (5.1) 0,, otherwise For D < j- 3 we know that x = 2D - 1 is again the argument of a dis- continuity,, nonzero in size,, since this jump is equal to that at x - 1, and the latter is nonzero since F(x) is nowhere zero inside the unit interval. In x 3 (l,3)} this discontinuity appears in the range of the F(— ) term,, for D < r-, and cannot be cancelled Toy an equal hut opposite jump in one of the other terms in that equation because F(x) is monotonic Therefore, there will be a discon- tinuity at x = k-D - 2} if 2D - 1 < y-, then discontinuities exist at ^+D - 2 and 8D - k; etc. Similar reasoning applies to discontinuities at points in the (2D - — , l) interval. Due to symmetry, a discontinuity in the range of either F(D + — ) or F*(D - — ) is cancelled by a compensating change in the other of these terms in (1,5), provided both terms are nonzero, However, for x 1 x > 2D - 1, i.e., T) + — > 2D - — , the F* term is zero; in this case discon- tinuities in the range of the F(D + — ) term cannot be cancelled, again because F(x) is monotonic. The following summarizes these observations; Theorem 2 , 5a : A discontinuity in F(x), for x = y n , < y < - } implies the existence of a discontinuity for x = 2y n , Similarly, a discontinuity for -32- x = y , 2D - — < y < 1, implies the existence of a discontinuity at x - 2y - 2D. Another useful conclusion which follows almost directly from the above observations is: Theorem 2.^b : There is at most one pair of discontinuities in F(x) for the interval — < x < 2D - — „ Proof : This restriction is simply a consequence of the fact that the number of jumps on both sides of (1.5 ) must be equal for x > 0. A pair of jumps m the — < x < 2D - — interval cancel each other on the right of (1.5), due to symmetry. The two discontinuities in the defining equation for F(x) (see equa- tion (3°8)) are each the source of one negative jump,, and thereby compensate for the mutual cancellation, or "loss,," of two other jumps on the right side, but no more. Since the intersection of the domains of the three terms on the right of equation (3=8) is null, while the union covers the unit interval (excluding the points x = D, x - —) f each discontinuity in F(x) appears in precisely one of these individual terms. Due to the two additional discontinuities intro- duced by the definitions of F„ and F^ in (3.8), the number of discontinuities in F'(x), x > is, in general,, two less than the sum of the numbers of discon- tinuities in these individual terms. This difference can be accounted for either by the mutual cancellation of jumps in the (— , 2D - — .) interval, described in theorem 2.5b, or by two pairs of jumps in the individual terms r ) viz. : jumps at x - 2D - 1 and x = 1 due to the zero-going of F(D - — ) and F(-), respectively. -33- summing to a single pair in F(x). The stationary density function for D = -jr is an example of the latter (see Figure k). Since discontinuity arguments are all linear expressions in D (e.g.., 2D - 1, h - GV) } etc.), this effective coin- cidence of two pairs of discontinuity arguments in general occurs only for isolated values of D, which will be called the "exceptional divisors," denoted [D„}„ To be precise, the set fD^l is defined to include all divisors to which v E E (5. 2d) below applies. We are now in a position to find the arguments of all discontinuities in a very straightforward way, A sequence {x.} is formed., using theorems 2.5a and 2.5b, starting with x - 2D - 1: a) If x. < — , then x. , = 2x. . l 2. 3 l+l l (5.2) b ) If x . > 2D - \, then x . . - 2D - 2x . . l 2 J l+l l c) If — < x. < 2D - — , the sequence is complete, 2—i 2 J d) The sequence is otherwise terminated at the ith step if there exists an x . , j < i„ such that x„ = x. . J J i Elements of a set fx!}, where x! = 2D - x. , are then also discon- ± JJ l i^ tinuity arguments, because of symmetry. The fact that [x.) and {x!} include all discontinuity arguments is evident from considerations similar to those used to prove no jumps are positive: in general, any discontinuity other than the pair resulting from discontinuities in the defining equation (at x = 1; x = 2D - l) implies the existence of a predecessor (in the sense of the sequence defined by (5.2)) at which the jump is twice as large, because of the coefficient, 10 1 x i ~ p anc * x i ~ ^ are limiting cases where two otherwise distinct discontinuities become coincident. -34- F(X) 5/3 5/4 5/6 5/12 2 r 4 .8 1 FIGURE 4q: SOLUTION FOR D = 3/5 1/2 F(X/2) (X<1) 1/2 F(D+X/2) FIGURE 4b FIGURE 4c 1/2 F*(D-X/2) FIGURE 4d -35- -, on the right in (1.5). Since the two just cited are the only x. which do not require such a predecessor, all discontinuities must he descendent from these, in the sense of (5-2) With the number of discontinuities known, the size of the respective jumps can also be easily evaluated if we ignore the set of divisors, (-D^}. Because F(x) is monotonic, the sum of the jumps for x > must be — (see (k.2j). Moreover, the sizes of the jumps in the sequence {x.} form a geometric progres- sion, i.e., the jump at x = x . -.is twice as large as the jump at x = jc. because of the coefficient of — in equation (1.5 ) - Denote a. = F(x.-) - F(x.+) (5.3) then for T discontinuity pairs: £ = 2 L f 2 Z a^) 1 = 1^(1 - (|) T ) (5J0 i=l i=0 or T-2 2 1 (2 T - 1) (5.5) (The coefficient, 2, in (5-^) accounts for the fact that the {a.} are half of the jumps.) Therefore, the jump at x . is: a, = (I) 1 ' 1 ■ a = 2T " X " (5=6) 1 2 1 D(2 T - 1) 11 If the divisor is in {D }, or either of the cases x. = D or x- = p occur, the predecessor requirement applies to all except one "generation of discon- tinuities, the exception resulting from the fact that one jump is half the sum of two others. -36- jumps at corresponding symmetric discontinuities are,, of course, equal. The solutions derived using (5.2) and (5° 6) apply to intervals of the divisor range., rather man individual values of the divisor (excluding solutions for [B } ) } since there are T pairs of discontinuities in F(x) for all divisors which result in x being included in the interval {— s 2D - — ) As demonstrated below., there are an infinite number of such intervals-, eacn of which involves a distinct sequence of operations when applying (^.,2) y though several non- adjacent intervals may result In the same number of discontinuities „ Equation (5«l) gives the solution for one of these intervals,, To formulate systematically the R_. distributions for — < D < f^ 1 ^ T .3 * 2 < D < Tp consider the intervals of D for which (— ) < (2D - l) < (— ) 9 n = 1 ; 2 9 . ..., x> ts i - 1 /lvn+2 . ^ - 1 ,l\n+l is true; denote tnese intervals of D (viz. : — + {—} < D < — + {^) 3 n = 1, 2 f . . . ) as I , n = I, 2 } . ...j respectively. To express a sequence of discontinuities generated by use of (5-2)^ the following definitions will be useful : N("k.) = k consecutive normalizing operations (i.e., (5.2a)) S = one application of (5 -2b) (Listing tne operations required to form, the discontinuity arguments^ rather than listing the arguments themselves s will enable observing similarities in different intervals of D which can then be used as a basis to form infinite families of solution intervals.) Then every 1 includes a sublnterval for which the sequence of opera- tions in (5.2) is N(n)j, i.e., the first x. > - also satisfies (5. 2c). Tn general, I will also include a subinterval for which the discontinuity locations n -37- in F(x) are determined in (5.2) by the sequence N(n)S; another for which the sequence is N(n)SN; etc. Table 1 is a list of all subintervals of I , n > 2 } for which the divisor results in n + 5 discontinuity pairs or less (excluding {D ]). The following two examples are provided to show how Table 1 was E derived; the solution for I appears in the first example Example 1, Consider the sub interval of I for which the sequence in (5°2) is N(n); this means the only discontinuities in F(x), for all values of D in this subinterval are at: x = 2D - 1; x = 2 (2D - l); ...; x = 2 n (2D - l), together with the respective 2D complements. For this portion of I , the fol- lowing is true : | < 2 n (2D - 1) < 2D - \ (5.7) or n+1 n+1 J-± < D < " , n = 1, 2, ... (5.8) n+2 - n+2 ^ > > (See Figure 5«) Note that I , .625 < D < .75> is covered completely in this family of solutions. Since the number of discontinuity pairs, T, is two for I (in general, T exceeds the number of steps in (5°2) by one) ? we find by use of (5.6); -38- TABLE 1. SOLUTION INTERVALS IN THE DIVISOR RANGE 2r(5o2) Interval Bounds Examples (to eight si gnificant fi gures ) Sequence n = 2 n = 3 n = 4 n = 5 1) N(n) 2 n+1 - 1 2 n+1 - 4 .5833.... o.53571^8 .516..... . 50806452 2 n+1 + 1 n+2 »5625 .53125 .515625 .5078125 2) N(n)S 2 n+2 - 1 2 n+ 3 _ 8 0625 o5535 7 l43 .525 .51209677 2 n+2 + 1 2 n+3 - 4 060714286 °55 .52419355 . 51190476 3a) N(n)SN 2 n+3 - 1 2 n+k - 12 59613846 .54310345 .52049180 .51 2 n+3 - 1 2 n+k - 8 . 58928571 o54l6.o.o . 52016129 .50992603 3b) N(n)SS 2 n+3 - 1 2 n+k - 16 .5625 .52916.00 51411290 2 n+3 + 1 2 n+k - 12 ......... . 56034483 52868852 .514 4a) N(n)SNN 2 U+k - 1 2 n+5 - 20 ......... .52813559 .51829263 50896414 2 n+k + 1 2 n+5 - 16 .5375 51814516 .50892857 4b) NfnjSNS 2 n+k - 1 2 n+5 - 2k 060576923 o5474l379 .52254098 .511 2 n+k + 1 2 n+5 - 20 060185185 54661017 .52235772 51095618 4c) N(n)SSN 2 U+k - 1 2 n+5 - 28 .55701754 . 52685950 . 51305221 2 n+k + 1 2 n+ 5 . 24 000000.00 .55603448 .52663934 .513 4d) N(n)SSS 2 n+4 - 1 2 n+5 - 32 .... . ......... .53125 .51512097 2 n+4 + 1 2 n+5 - 28 ......... .53099174 .51506024 •39- TABLE 1. SOLUTION INTERVALS IN THE DIVISOR RANGE (CONT'D) 2:(5»2) Interval Bounds Examples (to eight si gnificant fj gures ) Sequence n = 2 n = 3 n = 4 n = 5 5a) N(n)SNNN 2 n+5 - 1 2 n+6 - 36 .51720648 1 . 50844930 2 n+5 + 1 2 n+6 - 32 ......... .51713710 .50843254 i 5b) N(n)SNNS 2 n+5 - 1 2 n+6 - ^0 . 58796296 . 54025424 . 51930894 .50946215 2 n+5 + 1 2 n+6 - 36 . 58636364 o53991597 .51923077 o 5094433^ 5c) N(n)SNSN 2 n+5 - 1 2 n+6 - 44 059905660 . 54487179 .52142857 ; o 51047904 1 2 n+5 + 1 2 n+6 - 40 .5972,. ,o „ 54449153 .52134146 j .51045817 t 5d) N(n)SNSS 2 n+5 - 1 2 n+6 - U8 . 5495689 T o52356557 | ^5115 2 n+5 + 1 2 n+6 - 44 .5491^-530 . 52346939 .51147705 5e) N(n)SSNN 2 n+5 - 1 2 n+6 - 52 „ 55434783 „525720l6 1 .51252505 1 2 n+5 + 1 2 n+6 - 48 ......... o55387931 .52561475 .5125 5f) N(n)SSNS 2 n+5 - 1 2 n+6 - 56 .55921053 .52789256 ,51355422 2 n+5 + 1 2 n+6 - 52 o55869565 .5277.o,. . 51352705 5g) N(n)SSSN 2 n+5 - 1 2 n+6 - 60 .53008299 .51^58752 2 n+5 + 1 2 n+6 - 56 1 .52995868 .51^55823 5h) N(n)SSSS Q n+5 2_-_l_ 2 n+6 - 64 — -■■ -\ ,515625 2 n+5 + 1 2 n+6 - 60 ......... ,51559356 -1+0- lO • o CM 1 q cvl V N. Q V > lO Q CVJ ID • — c CVJ lO CM (& \ ^ CM ^ (0 ^ogf \ CO CM to 9T=gg^ "" - ■X-cz: ro lO CM ro V, ro ro 0> CO CD to -la- r . 1 D > < x < 2D — , 2D - 1 < x < min(lj-D - 2 ; 2 - 2D) F(x) == ^ — ', min(^D - 2 } 2 - 2D) < x < max(^D - 2, 2 - 2D) (5.9) 7==: , max(>D - 2, 2 - 2D) < x < 1 \_ 0,, otherwise For D in any given solution subinterval,, x and x' = 2D - x (and only these two discontinuity arguments) may be related by any of: x > x', x' > x _, or x' = x = Do This is the reason for the "max" and "min" in (5°9)° Example 2 „ If the sequence in (5-2) is N(n)SS., the following relation is satisfied by D° ~ < 2{2 n (2D - l) - D] - D) < 2D - | (5«io) or 2 n+3 + 1 2 n+3 - 1 - — , — < D - 2 U+k - 12 2 n+k - 16 (5=11) There are two reasons why any given sequence of operations in (5-2) may not correspond to a solution interval in the range of D. a) Some intermediate term may necessarily satisfy — < x. < 2D - — } if the implied final term also does, b) The sequence may not be self-consistent; e.g.,, if a sequence begins with N(k)S . .,, then (2D - l) > (— ) is true, and therefore an inconsistency would exist if k + 1 (or more) successive normalizations occurred at any point subsequently, because 2D - 1 is the smallest discontinuity argument „ In the example being considered,, N(l)SS cannot correspond to a solution because all values of D for which only one step is required to normalize 2D - 1 are part of the interval in which N(l) is the complete sequence „ N(2)SS cannot yield a solution since, if the first three steps are N(2)S, it can easily be shown that x, < 2D - — 3 and therefore the sequence must either terminate after N(2)S, or be followed by a normalization as the next step. For all n > 3^ equation (5.1l) correctly defines intervals of D for which the discontinuities in F(x) are generated by N(n)SS in (5»2), since neither of the two reasons for exclusion are violated in these areas , To facilitate deciding whether a given sequence of operations when applying (5-2) corresponds to some interval of D, there are four simple rules to test for consistency. 1) In a sequence starting with N(k)S, the maximum number of successive normalizations which can occur anywhere in the sequence is k. This is because 2D - 1 is the smallest discontinuity argument 2) A sequence starting with N(k)S cannot also terminate with N(k). Consider Figure 5 for a given value of D. If 2 (2D - l) > 2D - -, then any other x. occurring in the (-) < x < (— ) interval will, after normalization, also exceed 2D - — . again because all x. > 2D - 1„ 2' 1 — The other two restrictions are simply duals of the above „ Recalling that (5-2) is generating half of the total set of discontinuity arguments, it may be noted the sequence of operations to generate the 2D complement set, -43- {x! ), starting with x! = 1, is the dual (i.e., interchange N's and S's) of the sequence to generate (x.}. To determine the bound on the number of successive S's which may occur for D in a given interval, I , we assume D is such that some x. is almost equal to one. Rather than examining bounds on x . ., . etc.. $ J+l ' attention may, at that point, be shifted to the 2D complement sequence x! , J x'. , i etc., and the limits on the number of successive normalizations in the J+l complement sequence determined. From these considerations one obtains: 3) A sequence starting with N(k)S can have at most k successive S's. k) A sequence starting with N(k)S cannot terminate with a subsequence of k successive S's. After rejecting by inspection those sequences which are not consistent, the final test for the validity of a proposed sequence of operations in (5»2) is comparison of the interval of D implied with intervals corresponding to all subsequences (formed by truncation) to determine whether an intermediate x. was in the (— , 2D - — ) interval as well as the implied final one. It is therefore only necessary to find the smallest value of n, say n , for which a given sequence in Table 1 correctly describes a solution interval, and the expression then applies for all n > n . Solutions in Table 1 are arranged in the order of increasing complexity, such that all subintervals of I for which F(x) has k discontinuity pairs are listed before those with k + 1, Moreover, by ordering those which have the same number of discontinuities for I according to a binary weighting, such that S corresponds to 1 and N to zero, it happens that constants in the denominators of successive upper bounds (as well as successive lower bounds) form a readily discernible arithmetic progression. If desired, this progression may be used to extend the table indefinitely. -hk- 2.5.1 Properties of Solution Intervals Though by no means unique, the ordering of solution intervals given by Table 1 displays one of several interesting characteristics of the solutions Aside from the finite number of I for which a given form of sequence does not apply,, all I include : a) one subinterval which yields n + 1 discontinuity pairs in F(x),° b) one for which there are n + 2 pairs.; c) two for which there are n + 3 pairs ; d) four for which there are n + k pairs; and. in general,, e ) 2 for which there are n + j + 2 pairs . I, represented by the column for n = 5j is an example of this,, up to ten discontinuity pairs „ A second property which is apparent from Table 1 is that solution intervals are of the form &J and l6 discontinuity pairs, respectively. The discontinuity arguments which satisfy the — < x . < 2D - - relation for the last three examples are shown in Figure 6, which illustrates the rapid con- vergence of such a sequence,, It happens that all the solution intervals appearing in Table 1, except those described by N(2)S i N(3)SS^ N(^)SSS. and N(5)SSSS, are examples of first intervals in Infinite sequences. Due to tne above theorem, it is only necessary to observe whether any of the intervals resulting in fewer discontinuitites than the interval in question is adjacent, to decide whether a given interval is first in such a sequence . Parenthetically, it may be noted that equation (5ol2) provides a very straightforward method for computing boundaries on successive adjacent intervals in these infinite sequences . As an example, -^k = — ^o — is of the form of a lower bound in (5.12). To express the same number in the form of an upper 17 2^ - 1 2^-1 bound. 7-5 ° — j — = — rT^r~° Then the lower bound on the same interval is 20 2^" - 1 42U 25^ + 2 Tpk — ~T} again using (5. 12). In addition, the list of operations when applying (5«2) for each interval of D in such a sequence may be generated by application of the following rule : -kQ- \ \\ \ CM 1 S CM CO CM CO CD ro CO % or LU co 8 3 I- o o u. o Q. 3 O (T CD UJ Z o «— 1 CO a: o U- Q if>\ >l h- X CD o . . CO CO UJ CO a: ID CM O IT) O CO 0) 00 CO -49- If u(N,S) denotes the N-S sequence corresponding to any given solution interval, and u(N, S) is defined to be the same sequence with N's and S's interchanged, then the (5-l4) solution interval to the left of the one in question is characterized by u(N, S)Nu(N, S). This rule is merely a restatement of an observation which led to Theorem 2.5c, v where it was noted that with K discontinuity pairs, the second — terms generated by use of (5.2) closely approximate the 2D complements of the first V — . Applying it to the four intervals mentioned above, one obtains: |„}, defined to be those values of D for which (5.2d) is satisfied, may be recognized as one of three distinctly different types of limit points (or accumulation points) which would exist if Table 1 were -50- extended indefinitely. One type of limit point is the right edge of certain solution intervals; for D arbitrarily close but greater than such a boundary, the sequence of operations in (5.2) appears to cycle as discussed in the previous section. However, the {D } are accumulation points of solution intervals on either side, unlike these interval boundary points. The third type of limit point is that to which an infinite sequence of contiguous intervals converges (see Theorem 2.5c). This type of accumulation point of solution intervals is distinguished from the other two by the fact that the sequence of operations when applying (5 = 2) does not cycle. See (5.1^-) and the examples which follow it. To generate a few subsets of [D }, consider the following examples of cycling which could arise when applying (5.2): a) x = 2D - 1; x - ^D - 2; x - 8D - k; x, = ikl) - 8 = kD - 2, i.e., D = |. b) x = 2D - 1; x = kD - 2; x = 8D - k} x^ = l6D - 8; x - 30D - l6 = kO - 2, i.e., D = yL The first of these is an example of the subset of {D } for which the sequence ill of operations is N(n)SNSN ...; n = 2, 3, k, .... The second is an example of N(n)S MSM . ,.} n = 3, k, .... The values of D in the first of these subsets satisfy the equation 2 n (2D - 1) = 2[2 n+1 (2D - l) - D], n = 1, 2, . . . (5.1^) and therefore the subset is : D- 3 -^— , n-1,2, ... (5.0.}) -51- The values of D in the second subset are described by: 2 n (2D - 1) = 2[2 n+2 (2D - l) .- D] (5.1$) or D = Li^J , n = 1, 2, ... (5.17) 7 • 2 - 2 Both (5.16) and (51-7) can be shown to be proper subsets of the set in {D } 2 n - 1 where n assumes all integer values excluding powers of two. UNIVERSITY Of ILLINOIS UBR^ 3. SHIFT AVERAGE 3.1 Equation for While the density of partial remainders, as a function of D, is interesting in itself, its practical value lies in evaluating what fraction of the total number of quotient digits produced are nonzero. This figure of merit is frequently expressed in reciprocal form as the shift average, i.e., the average number of binary shifts in the division process between successive uses (addition or subtraction) of the divisor. The shift occurring as part of a step producing a nonzero quotient digit is counted when computing the shift average . Define : S(D) = S = random variable denoting the shift count between successive uses of divisor D; = expectation value of S. Since the distribution of R is independent of the dividend (for i sufficiently large), the distribution of S is also. This independence of R. was based on the assumption that the probability density describing dividends is piecewise constant, and finite (Section 1.3.2). Therefore, for a divisor D , = C means that the shift average for any ensemble of dividends which satisfies the above assumption is C; it does not mean that with divisor equal to D , the shift average is C for each individual dividend. If the distribution of the shift count, S, were known, could be evaluated by use of the relation which defines expectation: = E m • Pr[S = m] (l.l) m=l -52- -53- Pr[S = m] can be determined by converting f(x), the distribution of R. to distribution of normalized partial remainders, g(x), using equation 2:(l.7)j then: Pr[S = m] = g(D + --) - g(D + — 7=-) + g(D - — ~) - g(D - — ) B v m p m+l ' B v Q m+1 to v m ' (1-2) Alternatively, can be determined without knowledge of the dis- tribution of S if we make use of a rather basic property of ergodic, aperiodic Markov processes: if M. = mean recurrence time of state j, J and n = absolute probability of state j, (l-3) J then M. = (n . )" . J J Since S can be regarded as the recurrence time of the set of states for which x > — (where steps of the division process correspond to units of time in the terminology of a Markov process), equation (1.3) implies: " 1 = Pr[R. > |] or - - 1 - 1<£) Therefore, to find the shift average, it is only necessary to find the area under the density function, F(x), for x > — (or x < -, whichever is 1 e.g., see Feller, Chapt. XV, Section 5. -54- more convenient). This simplification in evaluating is a by-product of considering the distribution of R., rather than that of normalized partial remainders . 3 • 2 Evaluation of for j < D < 1 F(x), for r- < D < 1, is given by equation 2:(5.l), and we can easily confirm Freiman's result for this region by use of (1.4): <£> = — lj- = g^ I < S < 1 (£.1) 2D For ■£ < D < p, F(x) is given by equation 2:(5.9)> from which we obtain: f(|) = (2D - 1)| + ^(| - 2D + 1) = |, or (2.2) = 3, f or | < D < | 17 5 The next adjacent interval in the range of D is — o ^ D < tt, which appears in Table 1 as the n = 2 entry for sequence N(n)S. Since the N(2)S sequence involves three steps ;, there are four discontinuity pairs in F(x): 2D - 1, 4D - 2, 8D - 4, and Ikh - 8, together with the respective 2D complements. The area under F(x), x < — } can be determined by use of equation 2: (5.6). First: F(|) = F(D) + a T + -L(. 1 ) = i ( (2.3) 2D 2D V 2 T _ ' 2D V 2 T _ 1 ' -55- where, as defined in Chapter 2, T is the number of discontinuity pairs, and a T is the jump in F(x) at one of the two discontinuities in the (p, 2D - — ) interval, provided D is not in the set {D ). The area in question is then expressible as : f(|) = | • P(|) + (2D - 1) &1 + (4D - 2)a 2 + (k - 6D)a 3 [1 + BD . 1 + iS^ + i4iiB] {2 .k) D(2 T - 1) _ 2 3 (Note that 2D - 1, hL - 2, and k - 6D are the only discontinuity arguments less than - for D in this interval.) Therefore from (1.4): = 3, for U < D <| (2.5) 17 Thus far, the shift averages for all divisors in the range -~k < D < 1 have been found. In addition, it can be readily ascertained from Figure h, using equation (1.4), that the shift average for the point D = .6 is also 3. 17 The conjecture that = 3 holds for the interval .6 < D < ^ is a natural consequence of these observations. However, the number of solution intervals in this small interval of D is infinite, with the number of discontinuities in F(x) becoming arbitrarily large in an infinite number of distinct neighborhoods, The complex behavior of F(x) for ^ < D < 4 notwithstanding, the con- jecture raised regarding in this region can be proved by use of the symmetry property, derived in Section 2.3, as demonstrated by the following two theorems . -56- Theorem 3«2a : There are no discontinuities in F(x) in the interval h - 6D < x < 4D - 2, for | < D < |. 5 o N N N N N N I 1 I 2 I 3 I I 3 2 I 1 2D - 1 1-D to-2 J 2D-J 3D - 1 2 2 2 - 2D FIGURE 7 Proof : 1. Let N, , N p , and N^ denote the numbers of discontinuity arguments appearing in the intervals 2D-l - 3 for | < D < |. 5 8 Proof ; (The proof consists of relating areas under F(x) for certain intervals of x. Recognizing expressions of the form: f(z ) - f(z ), as the area under F(x), for z < x < z may be found helpful.) 1. Using equation 2: (1.4) for D < ■?: <, f(^D - 2) - f(2D - l) = f(3D - l) - f(2D - |) + f(2D - l) - f(D - |) (2.8) 2. Since F(x) = |, for < x < 2D - 1, f(2D - 1) - f(D - |) = 1 - ^ (2.9a) and -58- f(2D - 1) - 2 - | (2.9b) 3- Substituting (2.9) into (2.8): f(4D - 2) - f(3D - 1) - f(2D - |) + 3 - ^ (2.10) U. Due to symmetry, and the fact that F(D) = — : F(D + C) + F(D - C) - i, for any C < D (2.1l) Therefore, the sum of areas under F(x) for two symmetrically located W intervals, each of width W, is — , i.e., /Z Z "" z 2 {F(D + y) + F(D - y)}dy = 2 p X Z l provided all arguments are in the (0, 2D) interval. An application of this relation is : ] l D " I 1 f (±) - f(l - D) + f (3D - 1) - f(2D - i) = — ^ -I-535 (2 * 12) Substituting (2.12) into (2.10): fj(to - 2) = 4 - § - [f(|) - f(i - d)] or (2.13) *(§) -> -|- [f(^D - 2) - f(l - D)] -59- 5. From Theorem 3«2a: F(4D - 2-) = F(l - D) (2.l4) Using 2: (1.5) for D < i and x = 4D - 2- : F(4D - 2-) = |[F(2D - 1-) + F(3D - 1-)] = ^ + |F(3D- 1-) (2.15) From equation (2.11):: F(3D - 1) + F(l - D) = i (2.16) After substitutions, the last three equations yield: F(4D - 2-) = ~ (2.17) 6. The area under F(x), l-D f or \ < D < | For D > .6, has been found to be a relatively straightforward function. However, in the interval — < D < .6, the complex behavior of the partial remainder density is reflected in as well. To evaluate for the families of solution intervals for which F(x) was determined in Chapter 2, equation 2: (5.6) will be used to find f(^). Due to symmetry, together with Theorem 2.5b, one of each pair of discontinuity arguments in less than — , with the exception of the one pair in (— , 2D - — ). I f x * denotes which of the pair of symmetric discontinuity arguments, x^ x ^t is less than — , then: -6l- f(|) ■ \ F(|) + Z a.x* (3.1) i=l where F(p) is given by equation (2.3). As an example, consider the infinite set of intervals in the range of D for which the sequence of operations in 2: (5.2) was N(n), n = 1, 2, . ..; viz.: n+1 n+1 _ 5-= < D < - — = = 2 n+2 2 n+2 _ k Using (3«l) and 2: (5.6), and observing that T = n + 1 in this case, we find that: 1 P n_1 f (o) " £\ [1 + n (2D - 1)], n = 1, 2, ... (3.2) 2 D(2 n+1 - 1) for this family of intervals. As a second example, consider the family of intervals corresponding to the operations N(n)S in 2: (5-2); T = n + 2, and the expression is : ' - n + f [1 + "< 2D " ^ + 2D ~ ^^ " ^ (3.3a) 2 D(2 n+2 - 1) 2 n = —TZr [1 + (n - 1)(2D - 1) + -2-] (3.3b) D(2 n+2 - 1) 2 n 1 In (3.3a), the term n(2D - l) results from the n multiples of 2D - 1 (i.e., 2 J (2D - l), j = 0, 1, ..., n - l) which are less than -r. Since the sequence terminates after one subtraction, the discontinuity argument 2 (2D - l) must exceed 2D - — , and therefore its 2D complement must be less than — , and is represented by the last term in (3. 3a). Table 2 is a list of the expressions -62- TABLE 2. EXPRESSIONS FOR f(|) 2: (5.2) Sequence f(|) 1) N(n) n = 2, 3, ... f(|) " —|li [1 + n(2D - 1)] 2 D(2 n+1 - 1) 2) N(n)S n = 2, 3, ... *& = ~ nC El + (n - D(2D - 1) + -2r] 2 D(2 n+2 - 1) 2 n a 3a) N(n)SN n - 2, 3, ... 1 P n+1 T) f 2 = -7w5 [1 + n(2D " 1} + \ ] 2 D(2 n+3 - 1) 2 n 3b) N(n)SS n - 3, ^ ... t( & = — C [1 +■ (n - 2)(2D - 1) + --?-] 2 D(2 n+3 - 1) 2 n 2 4a) N(n)SNN n * 3, fc, *(i) ■ ,n+2 >2' ^n + "5 " [1 + (n + 1)(2D " 1)] D(2 ur " - 1) 4b) N(n)SNS n - 2, 3, n+2 f r i) = 2 — [1 + (n - 1)(2D - l)+—i-« D] D(2 X ^ - 1) 2 n+1 4c) N(n)SSN n - 3, fc, '(J) ■ ,n+2 N n+ D(2 U " - 1) [1 + (n - 1)(2D - 1) + -^J «D] 4d) N(n)SSS n - *, 5, n+2 - f (|) = ha [I + (» - 3) (2D - i) + -=s -D] 2 D(2 n+4 - 1) 2 n X -63- for f(p) for all subintervals of I , n = 2, 3; . .., in which there are n + h discontinuity pairs or less; the interval of D to which each applies is given by Table 1. In Table 3> expressions derived for Table 2 are used to evaluate , using equation (1.4), and with n = 2, 3> 4, 5- One interesting property of in this interval is the appearance of "plateaus" for certain intervals of D (e.g., N(2)S, N(3)SS, N(4)SSS, ...). The characteristic which identifies plateau regions is evident from examination of expressions for f (p); any sequence which involves a total of m normalizations and m subtractions yields an expression of the form: T-2 f(|) = % [1 + (m. - m )(2D - l) + Q • D] (3-4) 2 D(2 T - 1) 1 2 where 9 is a constant. (Equation (3«l) can be used to confirm readily that (3.4) is true for the general case: a) If the step required to generate x , given x , is a subtraction, then x > 2D - — and the a * x* product K 2 K K in (3.l) is of the form JC-1 2D - 2 (2D - 1) + ... 2 K-1 2 T-2 D(2 T - 1) b) Analogously, if the kth step is a normalization, then x must be less than — , and the a * x.* product K 2 K K contributes a positive multiple of 2D - 1. ) Therefore, when m exceeds m by (exactly) one, (3-^) yields a constant. Conversely, f (—), and therefore , can be constant over an interval of D only when m - m = 1. Since nothing has been assumed regarding the value of ■6k- TABLE 3- EXAMPLE EXPRESSIONS FOR , USING TABLE 2 AND EQUATION 3: (1.4) Sequence in 2:(5.2) n = 2 n = 3 n = k n = 5 l) N(n) 7D 2 - D 15D 8 - 9D 31D 2k - 33D 6 3D 6k - 97D 2) N(n)S 3 31D 8 - 3D 63D 32 - 35D 127D 96 - 131D 3a) N(n)SN 31D 8 - 3D 63D 32 - 35D 127D 96 - 131D 255D 256 - 387D 3b) N(n)SS 63 23 12 7D 32 - 9D 255D 128 - 137D 4a) N(n)SNN 12 7D 96 - 129D 255D 511D 256 - 385D 6UO - 1025D 4b) N(n)SNS 3 127D 32 - IID 255D 511D 128 - 139D 384 - 523D kc) N(n)SSN 127D 32 - IID 255D 511D 128 - 139D 384 - 523D 4d) N(n)SSS 255 103 511D 128 - 25D -65- 3 3 D, all solution subintervals in the — < D < t- region must correspond to sequences for which this difference is unity. A second consequence is that F(x) must have an even number of discontinuity pairs for intervals of D in which is constant, since the number of operations in 2: (5 .2) is odd if m i " m 2 = lo In principle, Table 3 could be extended indefinitely to obtain algebraic expressions for for as great a fraction of the „5 < D < .6 interval as we desire, although the size of solution intervals decreases rapidly with larger numbers of discontinuities and are particularly minute in the lower portion of this interval of D„ (See Table 1 for n = 3, ^ and %) Alternatively, by evaluating numerically, for a sufficiently fine grid of points in this interval, a graphical representation can be obtained to within any desired accuracy,, Figure 8 was obtained with the aid of ILLIAC II, using increments of .001 for .5 < D < 060 The discontinuity arguments, x*, were determined for i < 30o By rewriting equation (3»l) as: 1 T-2 r T-2 r t . 2 D(2 T - 1) L i=0 L 2 1+1 we observe that the maximum error in f (^) which can occur as a result of -30 T < 30, where, in fact, T may be infinite, is of the order of 2 Using increments of 10 and the trapezoidal rule, a numerical approximation to the integral / dD (3.5) ■66- 1 — 1 1 1 > o CO 0> ID in to to in in ro in CM m in m o o> CO h- CD lO *- ro CM *H O ro CM CVJ OJ CM CM CM CM CM CM CM -67- was found to be 2.61917 + 5 x 10 . This corresponds to the average value of for D in this range, if all values of D are assumed equally likely. Using this result, together with 3:(lA) and theorem 3-2b, the average for the entire divisor range is 2.6170 for S-R-T division. Because of the short computation time required per point (less than a millisecond), it was found practical to investigate in detail the behavior of for certain portions of the divisor range. The following are listed as empirical results : a) The left limit of the plateau region in which has 9 ikk D = -rz as the right end point is D = 37^. b) The above plateau is a local minimum in . Moving Ikk to the left of D = 7^j } has numerous local maxima, each of which is a plateau region. The limiting 2k 96 maxima is a point, viz., D = r-r-, at which = ^=-. While the difference of (approximately) 0.01 in the 2k ikk 9 values of at D = r-r compared to the -^p= < D < y^ plateau may have little practical significance, this behavior is sufficiently contrary to intuition to warrant future investigation. Noteworthy is the much larger peak which Freiman's simulation of the so-called 9 . 75-1 • 0-1. 5 method yields for D near y£. c) The nonmonotomic behavior of described in (b) 17 occurs again near D = — , and a plausible conjecture is that it will be found near all D = — + (p) , n = k, 5> ..., due to the similarities in these neighborhoods noted in Chapter 2. h. SUMMARY AND CONCLUSIONS 4.1 Summary Treating digital division as a Markov process in order to provide a basis for evaluating and comparing division algorithms was the contribution of Dr. C. V. Freiman. If the set of dividends encountered is assumed to have some degree of randomness, so that it may be approximated by a piecewise uniform probability distribution, then, as demonstrated in Chapter 1, the distribution function describing partial remainder magnitudes approaches a limit which is determined only by the divisor, D, and in particular, is independent of both the dividend distribution and the step numbers in the decision process. This stationary distribution is itself piecewise uniform, and the intervals of uniformity are the states of the Markov chain. However, a straightforward application of the theory of finite Markov chains in the case of S-R-T division was found imprac- ticable, for the following reasons : a) For S-R-T division, the number of discontinuities in the partial remainder stationary density function becomes indefinitely large in an infinite number of distinct neighborhoods in the divisor range, as demonstrated in Chapter 2. b) When solving for the probabilities of the individual states, or intervals, the number of unknowns in the linear equation set must at least equal, and, in general, will exceed the number of discontinuities in the density function. c ) Solving one set of linear equations yields the partial remainder distribution for only one value of the divisor. -68- -6 9 - As an alternative to the finite Markov chain approach, a functional relation was derived which was shown to be a necessary and sufficient condition for the stationary density function (2:(l.5)). By considering the Fourier series representation of a periodic extension of the density function, and applying the functional relation just mentioned together with the knowledge that the total variation (per period) is bounded, it is proved that the density functions for all values of D must satisfy an important symmetry relation (Section 2.3). This symmetry is used to derive a number of other properties which lead to a straightforward algorithm (2: (5.2)) for determining all discon- tinuity arguments in the density function, and an equation for the size of the respective jumps (2: (5.6)). The limiting distributions being sought may be expressed in terms of these jumps. The solutions thereby obtained apply for intervals in the range of D, and by grouping these intervals on the basis of the sequence of steps used in 2: (5.2), infinite families of intervals are formed. Properties of the array of solution intervals are derived by consid- ering, among other things, the relations which must exist between adjacent intervals . While the partial remainder stationary density function is interesting in itself, a practical application lies in evaluating the shift average, , which is the average number of quotient bits generated per iteration; this problem is considered in Chapter 3. Since the shift average is equal to the mean recurrence time of partial remainders which exceed 0.5 in magnitude, is found to be simply related to the partial remainder distribution function (3:(l.*0). Properties derived in Chapter 2 enable proving that = 3 for the entire interval, 0.6 < D < 0.75^ a result which was previously not obtainable due to the complex behavior of the density function for D near 0.6. For 0.5 < D < 0.6, the complicated behavior of the partial remainder density -70- function is reflected in (Table 3), and relations derived in Chapter 3 are used to evaluate with aid of ILLIAC II, for D in that interval (see Figure 8). These results, together with equation 3:(2.l), the expression for 3 when r- < D < 1, satisfy a goal defined at the outset: determination of the shift average for the entire range of divisors. h.2 Conclusions and Areas for Future Investigation In recent years there have been a number of efforts to better under- stand the digital division process when more than a minimum number of divisor multiples are available, and to apply this insight in formulating algorithms which reduce the number of nonzero quotient digits . Usually the avenue of approach is to consider bit patterns which can occur, then to select a set of rules which provides best performance in each example. This is often an iterative process. Knowledge of stationary distributions of partial remainders for the various classes of division algorithms provides a valuable theoretical foundation in this search for optimum algorithms. Thus, for example, the shift average obtained for S-R-T division is applicable to an entire class of what might be called single-threshold division processes, which can be described as follows : a) the divisor multiples available are: mD, 0, and -mD, where m > 1, b) there is a positive number, K, such that K < |mD| < 2K is true, and c) K, which in the general case is a function of D, is used as the threshold level with which each R. is i ■71- compared to decide whether or not the divisor is to be used in forming R. & i+l The abscissas in the versus D relation derived in Chapter 3 may be regarded as — D rather than D. Metze's modification of S-R-T division uses CIS. this change of scale effected by varying K to take advantage of the maximum which was found to assume when 0.6 < D < 0.75 is true. An important question then is: can any similar generalizations be made to optimize multiple -threshold algorithm, perhaps using a property analogous to the relative independence of m and K in the single-threshold case? (Only the inequalities in (b ) above must be satisfied.) In other words, when more than one (nonzero) multiple of the divisor magnitude is assumed to exist, to what extent can the solution for a specific algorithm be generalized in order to aid in formulating an optimum one? Derivation of partial remainder distributions for any useful two- threshold algorithms (MacSorley 1 s paper includes examples of such algorithms) is complicated by the fact that a number of ranges of the divisor must be analyzed individually, since the thresholds vary with D. However, the observa- tions that a) the piecewise constant functions which describe the stationary density can be formed by locating discontinuity arguments as an ordered process, starting with discontinuities in the defining equation, and b) the size of jumps at successive arguments in this process form a geometric progression, should expedite these derivations. In addition to considering binary division with further redundancy added, there is, of course, the problem of considering higher radices. A second area where further work is needed is the modification of the versus D relation to account for the fact that, in parallel computers, shifting R. an arbitrary number of places in one step is not possible. One -72- approach to this problem would be that outlined by equations 3:(l.l), (1.2), and (1.3)? with the index, m, in (l.l) confined to a finite range corresponding to the shift paths available. Finally, a more thorough understanding of some peculiarities in the stationary distribution of R. for S-R-T division would be interesting if not useful. A few of the questions which remain are: a) Can the nonmonotonic behavior of for 0.5 < D < 0,6 be related to divisor bit configurations? The fact that "bad examples," i.e., cases for which S-R-T type division is clearly nonoptimal, often involve divisors of the 1 In form — + (— ) , and that the local minima in , which are plateau regions, include these points, indicates an intuitive explanation for these minima and what determined their end points may not be difficult. b) What significance can be attached to the accumulation points of solution intervals, i.e., neighborhoods of D which lead to distributions of R. with arbitrarily large numbers of discontinuities? BIBLIOGRAPHY 1, A. T. Bharuca-Reid, Elements of the Theory of Markov Processes and Their Applicati ons ;, Chaptc P McGraw-Hill; i960 2. J. Cocke and D, W„ Sweeney, ''High Speed Arithmetic in a Parallel Device," IBM internal report, February 1957 3„ W. Feller, An Introduction to Probability Theory and Its Applications , second edition, Chapt „ Ik, 15 ? John Wiley and Sons; 196c h. C. V. Freiman, "Statistical Analysis of Certain Binary Division Algorithms," Proceedings of the IR E, vol. k$, No, 1, pp. 91-103; January 1961 5. J. Go Kemeny, H, Mirkhil, J, L, Snell, G. L. Thompson, Finite Mathematical Structure s, Chaptc 6, Prentice-Hall, Inc; 1959 6. 0, L. MacSorley, "High Speed Arithmetic in Binary Computers," Proceedings of the IRE , vol, 49 ; No* 1, pp, 67-91,. January 1961 7. G. Ac Metze, "A Class of Binary Divisions Yielding Minimally Represented Quotients," IRE Trans a ctions on Electronic Computers , vol, EC-11, ppc j6l-hj December 1962. 8. Penhollow, J, 0,, ;i A Study of Arithmetic Recoding with Applications in Multiplication and Division, : ' University of Illinois Digital Computer Laboratory, (Doctoral Dissertation ;_, 1962 9c R, K. Richards, Arithmetic Operations in Digital Computers , D, Van Nostrand Co.; 1955 10c J, E, Robertson, ! 'A New Class of Digital Division Methods," IRE Trans - actions on Electronic Computer s.- vol, EC- 7, PP^ 218-222,; September 1958 11= T t D, Tocher, '''Techniques of Multiplication and Division for Automatic Binary Computers, ,: Quart, J, Mechu Applied Math ,, vol, XI, pt, 3^ pp, 364-38^ 1958 12, J, B. Wilson and P., S. Ledley, "An Algorithm for Rapid Binary Division," IRE Transactions on Elect r onic Compute rs, vole EC-10, pp, 662-670/ December 1961 13. A Zygmund, T rigonometric Serie s, vol., I, Chapt s. 1, 2, Cambridge j; 1959 -73-