* "L I B RARY OF THE UN IVE.RSITY Of ILLINOIS 510.84 Iflfcr 00.226-236 cop 2- The person charging this material is re- sponsible for its return to the library from which it was withdrawn on or before the Latest Date stamped below.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. UNIVERSITY OF ILLINOIS LIBRARY AT URBANA-CHAMPAIGN MAY 1 1 1979 MAY 3 JUN 7 Mm MAY 04 1938 L161 — O-1096 Digitized by the Internet Archive in 2013 http://archive.org/details/partiallyseparab228sliv S/o. Xf- -flC . <££-% Report No. 228 tOJ> . IM- PARTIALLY SEPARABLE FUNCTIONS by THOMAS ANDREW SLIVINSKI COO-IU69-OO61 y ' -z June 1 1967 ^ DEPARTMENT OF COMPUTER SCIENCE • UNIVERSITY OF ILLINOIS • URBANA, ILLINOIS Report No. 228 PARTIALLY SEPARABLE FUNCTIONS* by Thomas Andrew Slivinski June 1, 1967 Department of Computer Science University of Illinois Urbana, Illinois 6l801 *This work was submitted in partial fulfillment of the Doctor of Philosophy degree in Computer Science and supported in part by Atomic Energy Commission Contract AT(ll-l) -IU69. ■ID. S i v ACKNOWLEDGMENT The author is very grateful to his advisor, Professor S. Muroga, for his invaluable guidance, encouragement and constructive suggestions in the preparation of this thesis. The author also wishes to thank Professor D. E. Muller for his comments and suggestions. The support of the Department of Computer Science, University of Illinois is also gratefully acknowledged. 111 TABLE OF CONTENTS CHAPTER PAGE 1 INTRODUCTION 1 1.1 Introduction 1 1.2 Basic Definitions and Examples 2 1.3 Statement of the Problem 5 2 PROPERTIES OF PSEUDO -SEPARABLE FUNCTIONS 7 2.1 Introduction 7 2.2 Elementary Algebraic Properties of Pseudo-Separable Functions 7 2.3 Bounds on the Degree of Hyperplanes and Conditions for Separation 19 2.k Assummability as a Condition for Pseudo-Separability 26 2.5 Simultaneous Realizability and Pseudo-Separability 42 3 GENERAL PROPERTIES OF MARGINALLY SEPARABLE FUNCTIONS 58 3.1 Introduction 58 3.2 Basic Properties of Marginally Separated Functions 59 3.3 A Necessary and Sufficient Condition for Quasi -Separability 68 3.4 Marginal Separation and the Degree of Supersummability 76 k SYNTHESIS OF PARTIALLY SEPARABLE FUNCTIONS 8$ k.l Introduction 89 k.2 Integer Programming Algorithm for Pseudo-Separability 89 4.3 Threshold Synthesis of Pseudo-Separable Functions 96 k.h Integer Programming Algorithm for Marginal Separability 99 4.5 Threshold Synthesis of Marginally Separable Functions 104 5 SUMMARY 108 LIST OF REFERENCES 110 VITA 111 PARTIALLY SEPARABLE FUNCTIONS Thomas Andrew Slivinski, Ph.D. Department of Computer Science University of Illinois, 1967 Threshold logic has been the center of much interest in recent years. This attention has "been generated both by demands for logical elements in computer circuits which can perform more complex Boolean operations than "AND" and "OR" and by the desire to develop a single unified theory in which the mathematical models for these various logical elements appear as special cases. This dissertation presents an extension of the concept of linear separation to generate a wider class of functions. This generalization is affected by modifying the requirements for separation to allow both true and false assignments to reside on the same hyperplane (pseudo-separation) or within the same neighborhood of a hyperplane (marginal separation). Functions which can be separated in either of these ways are called' partially separable functions. Some of the theoretical algebraic properties of partially separable functions are investigated and the relationship between these and the corresponding properties for separable functions is explored. Necessary and sufficient conditions for each type of separation are also discussed. A parameter which shows whether a given function is separable, pseudo-separable, or marginally separable is introduced. Integer linear programming algorithms which are guaranteed to minimize the number of true or false assignments on the same hyper- plane for pseudo-separation or the number of assignments within the same neighborhood of a hyperplane for marginal separation are presented. A systematic procedure for synthesizing an arbitrary partially separable Boolean function in a network of threshold elements based on these algorithms is developed. CHAPTER 1 INTRODUCTION 1.1 Introduction Linear separability of Boolean functions has been widely studied in recent years. Techniques for deciding if a given function is a threshold function have been developed and procedures for determining threshold networks to realize an arbitrary function have been proposed. This interest has been generated by two considerations. The first is the demand for logical elements which express more complex Boolean functions than the simple 'AND' and 'OR'. If a logical element can perform more varied and complex logical operations, a smaller number of such elements need be used in synthesizing an arbitrary Boolean function. The second motivation has been the desire to incorporate the mathematical models for various logical elements into a single unified theory. Thus, much effort has been spent in the research of separable or threshold functions [7] « This paper represents an attempt to generalize the concepts and properties of separable functions to include a wider class of functions. As n, the number of variables, increases, the number of threshold functions on n variables becomes very small when compared to the total number of functions on n variables. Three percent of all h variable functions are separable and only .002/o of all 5 variable functions are separable [12] . Therefore, an extension of the theory and ideas of separable functions to a more general class of functions is needed. The generalization which this paper presents is a modification of the actual requirements for separation. Instead of strict separation, both true and false assignments are allowed to reside on the same hyperplane or within the same neighborhood of the hyperplane. Functions which can be non-trivially separated in either of these weaker senses will be called "partially-separable" functions. Very little appears in the literature about this type of generalization. Robinson [10] discusses some theoretical properties in devising a technique to determine if a function is separable using determinants. Otherwise the present author is unaware of any other work in this area. 1.2 Basic Definitions and Examples Before any definitions are given, a few words about the notation used in this paper are needed. Let f(x,, x p , . .., x ) be an arbitrary Boolean function. Then f(x., x 2 , . .., x ) is a mapping from (0, l) into {0,1} and will always be assumed to be dependent on all n variables, x will be used to denote (x , x p , ..., x ) and any n-tuple A = (a , a p , . .., a ) where each a. = 1 or 0, will be termed an assignment to f (x) ■ If f (A) = 1, A is called a true assignment and if f(A) = 0, a false assignment. When discussing hyperplanes, [w; T] will represent the hyperplane W«x = T, where. W* = (w.,, Wp, ..., w ). We will also refer to the hyperplane [W; T] as a "structure" for a particular function f (x) . Definition 1.2.1; A Boolean function f(x) is said to be separable if there exists some hyperplane [wj T] such that for any true assignment A*, ^-A* > T and for any false assignment B*, W^B* < T. This definition is equivalent to allowing either true or false assignments to lie on the hyperplane [w; T], i.e., alternate requirements tf't > T if t(t) = 1 and \t-t < T if f (a) = or ^-A* > T if f(t) = 1 and vf-A* < T if f(3 =0 are equivalent. Definition 1.2.2; A Boolean function f (x) is said to be pseudo- separable if there exists some hyperplane [W; T] such that for any true assignment A, W^'A > T and for any false assignment B, W*B < T. Notice that both true and false assignments are allowed to lie on the hyperplane [w; T]. Example 1.2.1; Consider the function f(x) = x x p s/ x x, and the hyper- plane [W; T] = [1,1,1,1; 2]. If A is any true assignment, then A must either have more than 2 components a. with value 1 or else must be (1,1,0, 0) or (0, 0,1,1). If A has more than 2 components with value 1, then $.t> 3. If t is either (l, 1,0,0) or (0,0,1,1) then tf-£ = 2. In either case, w*A > T. If Bis any false assignment, then F must either have less than 2 components b. -with value 1 or else must be (1,0, 1,0), (1,0,0,1), (0,1,1,0) or (0,1,0,1). In the first case W-B* < 1, while in the second, w»B = 2. Therefore, for any false assignment w «B < T. Hence x,x p ^ x x x k ^ s pseudo-separable . k Definition 1.2.3: A Boolean function f (x) is said to be marginally separable if there exists some hyperplane [w; T] and some non-negative constant, A, called the margin , such that for any true assignment A*, tf«A*> T - a/2 and for any false assignment % V?.I? < T + a/2. Notice that we allow both true and false assignments to lie in the region between the hyperplanes [% T - a/2] and [% T + a/2]. Example 1.2.2: Consider the function f (x) = x-.Xp >/ x_x^. By Example 1.2.1 this function is pseudo- separable,, Therefore it must also be marginally separable with margin a = 0. Example 1.2.3° Consider the function f (x) = x-|Xp ^ x,x. x and the structure [3,3,2,2,2; 6 l/2] with margin A = 1. If (3,3,2,2,2) -A; > 7, — ^ then A must either have a, = a. = 1 or a, = a^ = a- = 1. Therefore A must be a true assignment. If (3,3,2,2,2) °B < 6, then either b = — > or b = 0, or one of b,, b. or b,_ is 0* In either case, B is a false assignment. Therefore, for any true assignment (3,3,2,2,2) 'A > 6 and for any false assignment (3,3,2,2,2) «B< 7» Hence x-^x^ v x x, x must be marginally separable with margin A = 1° Definition 1.2.4: A Boolean function f (x) is said to be a partially separable function either if f(x) is pseudo-separable and W / (0,0,.. »,0) or if f (x) is marginally separable and n a< 2livi. i=l 5 1.3 Statement of the Problem Thus, by modifying the requirements for separation, a larger class of functions is obtained. This can easily be seen by checking the functions given in the examples in the previous section. These functions are not separable, but are partially 'separable. The natural question concerning the algebraic properties and necessary and sufficient conditions for each type of separation arises. Along -with this comes the desire to find an optimal structure, in the sense that a minimal number of true and false assignments ]le on the same hyperplane (in the case of pseudo-Separatior) or in the same neighborhood of the hyperplane (in the case of marginally separable functions) . A function which is not separable, but is partially separable -will have such assignments. If we want to synthesize a partially separable function in a network of threshold elements, then most of the assignments of f (x) can be realized by a single threshold element having the structure [w; T] where f(x) is partially separable, under [W; T], and if we use [w; T] as a threshold element, then we must correct for false (or true) assignments lying on the hyperplane by using additional threshold elements in the network. Thus, if the number of such assignments is minimal, then the number of these corrective elements in the network will also" be minimal. As a special case, if all of these assignments are given as "don't care" conditions in the actual specification of the network, then the function can be realized by a single threshold gate. Therefore this paper will be directed toward three goals: 6 (1) A general description of the algebraic properties of partially separable functions . (2) Methods for reducing or minimizing the number of assign- ments for which corrective elements are necessary in a network. (3) Techniques for the general synthesis of functions using threshold elements. CHAPTER 2 PROPERTIES OF PSEUDO- SEPARABLE FUNCTIONS 2.1 Introduction In this chapter some of the algebraic properties of pseudo- separable functions are investigated. Some elementary properties and ideas are presented in the first section. In the second part the con- nection between pseudo-separable and separable functions is explored with special emphasis on bounds for the number of points which lie on the hyperplane. In the third section necessary and sufficient conditions for a function to be pseudo-separable are developed based on the idea of asummability. The relationship between simultaneously realizable functions and pseudo-separable functions is discussed in the last section, 2.2 Elementary Algebraic Properties of Pseudo-Separable Functions .. In this section some elementary properties of pseudo-separable functions are presented. Many of these properties are analogous to properties of separable functions. However, because both true and false assignments can lie on the same hyperplane, they are in general weaker, In this section fix) is assumed dependent on exactly n-variables . Definition 2.2.1 : The hyperplane LW; TJ is said to pseudo-separate the function f (x) if for any true assignment A, W«A > T and for any false -+-»—» — > -+ -> assignment B, W«B < T. The number of assignments A. such that W°A. = T — l l is called the degree of [W; T] . 8 Lemma 2,2.1: The structure [Oj 0] pseudo-separates every function. n Proof ; For any assignment A, 0»A = L o«a. = 0. Therefore every i=l 1 assignment A must lie on the hyperplane LOj 0J. Q.E.D. Since [Oj 0] pseudo-separates every Boolean function, and since this pseudo-separation is trivial in the sense that every assign- ment must lie on [0$ 0], we would like to avoid this structure in our future discussion,, Definition 2.2.2 : A structure [Wj T] is called non-trivial if W ^ (0,a,I'oi,0). A structure [Wj T] is called zero-free if all w. / 0'U« 1,2, ...,n). Lemma 2.2.2 : The degree of any non-trivial structure is strictly less than 2 n , Proof : If the degree 6f [W; T] is 2 , then every assignment A. must be such that W°A. = T» Consider the assignments (0,0,.„.,0) and (0,0, .„. ,0,1,0, . . o ,0) where this second assignment has 1 in position xj. Because every assignment lies on the hyperplane [WJ T], W°(0,0,...,0) = = W.(0,0,.. .,0,1,0, ...,0) = w.. Hence all w. =0 and T = 0. Therefore the only structure of degree 2 n is [Oj 0). Q.E.D. 9 Definitipn 2.2. 3 l Let f (x) be an arbitrary Boolean function. Let [W; T] be a non-trivial hyperplane which pseudo-separates f(x). An assignment A is said to be segregated by LW; TJ if W-A f T. Since [W; T] pseudo-separates f(x), this means that if A is segregated by [¥; T] and true, then W-A > T and if A is false, W-A < T. Using Definition 2.2. 3> Lemma 2.2.2 can be restated as follows: Corollary 2.2.1 : Every non-trivial structure segregates at least one assignment. Notation 2.2.1 : If f (x) is any given Boolean function, the notation "f(x; x. -* a,,...,x. -* a. )" will be used to represent the function ' 1, 1 l k which results when a (another variable or value) is substituted into f (x) for x. , a^ substituted for x. and so forth. The notation l 2 l 1 2 "f(x) > g(x)" means that for any assignment A, f(A) > g(A). If for some assignment A, f(A) > g(A), and f(x) > jg'(x) we .Vill' wr'ite "f(x) > g(x)." Definition 2.2.^ : f (x) is said to be positive in some variable x. if f(x; x. -* 1) > f(xj x. ■* 0). f(x) is negative in x , if f(x 3 x. -> 0) > f(xj x. -> l) . f(x) is called unatein x. if f(x) is either positive in x. or negative in x.. f(x) is positive if it is positive in all variables and negative if it is negative in all variables. f(x) is called unate if it is unate in all variables. Th eorem 2.2.1 : Let f(x) be pseudo-separable under [Wj tj). Let x. be any variable. Let A. be any assignment on the remaining n-1 variables. 10 If f(A.: x. - l) = 1 and f(A. j x. -♦ 0) = 0, then w. > O If i' i 11 i — fit.; x. -* l) =0 and f(A. j x. - 0) = 1, then w. < 0. i' i ii i — Proof : For any variable x., it is well known that f (x) = x.f (xj x. -*• l) v x.f(x.s x. -*• 0). / i N > i i i' l If f(A. | x. -»• 1) = 1 and f(A.$ x. -* 0) = 0, then since [W; T] pseudo- separates f(x), n n Z w.a. < T and E w.a + w. > T. j/i J J ~ tfi' J J X ~ Combining these two expressions, w. > 0„ A similar.' proof is used to show that if f(A.» x. -* l) = and f(A.$ x. -* 0) = 1 then w. < 0. ii ii i — QoE„D, Corollary 2.2,2 : Let f (x) be a pseudo-separable function. If f (x) is not unate in some variable x., then the weight corresponding to that variable must be 0. Proof ; If f(x) is not unate in x., then neither f(xj x. -*• l) > f(x; x„-» 0) nor f(xj x„ -*• 0) > f(x; x. -* 1), Therefore there is a pair of assign- or -* ments A. and B. on the other n-1 variables such that l l (1) f(A. , x. «* l) = 1 and f(A. j x. -♦ 0) = ii ii (2) f(B. j x. -* 1) = and f($.', X. -♦ 0) = 1 i' l i- l By Theorem 2.2.1, (l) implies w. > and (2) implies w. < 0. Therefore v = 0. Q.E.D. 1] Corollary 2.2.3 : If f(x) is pseudo-separable under a zero-free structure, f(x) must be unate. Corollary 2,2.^4-: If f(x) is pseudo-separable under [W; T] and positive (or negative) in x., then w„ > (or w. < 0). . i l — i — Lemma 2.2.3 : Let f(x) be pseudo-separable under lW° T]. Then f(x) x. -*x.) is pseudo-separable under -*1 1 [w ± , w g , ...., * ±mml > ^ ± , \ +1 v v. M 'w n *,T - w 1 ] = [W ; T ]. Further, the degree of [W ; T ] is the same as the degree of [Wj T]„ Proof : Since [Wj T] pseudo-separates f(x), n (1) if f(x) = 1, then I w.x. > T u n (2) if f(x) = 0, then E w x. < T x. can be expressed arithmetically as l-x. « l i n Substituting x. for x. in f(x) and (l-x„) for x. in £ w.x. n (l)' if f(xs x. -* x.) = 1, then L w.x. +■ (l-x.) w. > T i i j/i J J l' l - i -* - n (2) if f(x, x. ■+ x.) =0, then L w.x. + (l-x.) w. < T. Let w. = -w, and subtract w. from both sides in (l)' and li i (2) V- then 12 n (l)" if f(x; x. ->x.) = 1, then E w x. + w. x.'>-T - w. ' i i j^ J J i -i - ;. l -♦ n 1 (2)" if f (xs x. -* x. ) =0, then E w.x. + w. x. < T - w. i V tf ± 3 J i i " Therefore, [wj T ] pseudo-separates f(x; x. =»"xT)„ — * — »• — > Let A he any assignment such that W'A = T„ Consider the assignment A where a„ = a.(j 4 i) and a. = a . Then -*1 -»1 n n W"°A = E w,a. +- (l-a 4 ) (-w ) = E w,°a. + w„°a. - w. = T - w„ . Therefore if A lies on hyperplane LW; TJ, A lies on hyperplane [Wj T ]. Similarly if A is any assignment on [fj T ], then A will lie on [W; T], Therefore there is a one to one correspondence between assignments on these two hyperplanes so that they must have the same degree. Q.E.D. The above lemma allows us to restrict our attention to positive functions and to then generalize all results to unate functions. Hence- forth, the term "positive function" will carry this extension implicity. Lemma 2 2 U : If f (x) is pseudo-separated by [W$ T], then f (x) is pseudo- separated by [-Wj -T] and the two hyperplanes have the same degree. Proof : If f(x) = 1, then W-x > T or -(W-x) < -T and f-W)«x < -T. If f(x) = 1, then f(x) = and j[-W)«x < -T. Further, if f(x) = 0, then f(x) = 1 and W-3 < T or (-W) x > .?% Therefore [-Wj -T] pseudo-separates f(x). 13 To show that [W; T] and [-W; -T] have the same degree, let A be any assignment such that W°A = T, Then (-W)°A = -T and A also lies on [-W; -T]o Similarly if (-W)°A = -T, then Wo A = T. Q.E.D. Theorem 2.2,2 : If f (x) is pseudo-separated by [Wj T], then f (x) is n pseudo-separated by [W; Z w„ -T] and the degree of [Wj Tj is the same i=l x is the degree of [Wj Z w. -T] i=l n i Proof: Since f (x) = f(x) = f(xs x. -» x.. , x. -» x . . ... x -> x ) , the ■ ' 1 12 2 n n ' result follows from Lemmas 2»2.3 and 2„2 ^. Q.E.D. We would now like to investigate the relationship between symmetric variables and their corresponding weights „ Lemma 2-2.jp : Let f(x) be pseudo-separated by [W> T]. Let f(x) be symmetric in x. and x.« Even if w. 4 w., f(x) is also pseudo-separated i j i ' y by [w^ w 2 , ..., w w w 1+1 , ..., w , w., w ..., w n ; T] = y l+i- j-i- l- j- [W ; T] and the degree of [w; T] is the same as the degree of [Wj Tj. Proof : (l) Assume that [w ; T] does not pseudo- separate f(x). Then there must exist some true assignment A such that W °A < T or some false assignment B such that W -B > T. Assume that f (A) = 1 and VT-A < To Since [W; T] pseudo-separates f(x), W°A > T. Let C be the assignment which agrees with A in all variables other than the i and j lU coordinates and c, = a., c, = a„. Then W-C = ¥ „A < T„ Therefore C i 3 J i must be a false assignment because [VT ; T] pseudo-separates f(x)„ Since x. and x. are symmetric in f(x), A must also be a false assignment. Therefore, we have a contradiction „ (2) Let A be any assignment which lies on the hyperplane [Wj T]„ Then the corresponding assignment C, which was defined in part (l) will lie on hyperplane [W; T], Similarly if C is on [VT°, T], then A must be on [w; T]„ Therefore, there is a one to one correspondence between assignments on [Wj T] and [w $ T] and these two hyperplanes must have the same degree „ QoEoDo Theorem 2 2,3' If f(x) is pseudo -separable under [W° T] and if x„ and x. are symmetric in ftx), but w„ f- w., then there exists another structure [w ; T ] with w. " tsw ; which, pseudo-separates f(x) ■■■'k^. and whose degree is no greater than the degree of [Wj Tj„ Proof ; Consider the hyperplane [VT ; T ] = [W +• w ; 2T] where [f; T] was defined in Lemma 2„2<,5° We will first show that this hyperplane pseddo-separates f(x)» Let x be any assignmento (W + w )«x = W°x +■ W °x, If f (x) = 1, then W.x > T and W^x > T and (W 4- W 1 ) >x > 2T Similarly, if f(x) = 0, then W*x < T and fa < T and (W 4- w~) -x < 2T„ Therefore [W 4- w ; 2T] pseudo-separates f(x)„ Also, because w. = w., and 1 1 1 , _. - 11 11 w, = w., w. +■ w. = w„ 4- w. = w. 4- w.o Therefore w„ = w„ iJ'iiiJOJ i z Finally, we want to show that the degree of [W 4- w°, 2T], is less than or equal to the degree of [Wj T]. Let x be segregated by [W$ T]„ If 15 W-x > T, then W~-x > T and therefore, (W + W~) °x > 2T„ Similarly, if W-x < T, then (W +■ F)»x < 2T» Thus x must also be segregated by [W f W 1 ) 2T]. We next define a partial ordering relation between different variables of a function, The following theorem shows how this relation, when it can be defined, relates the magnitudes of the corresponding weights in any pseudo-separating hyperplane. Definition 2.2,5 : Let f (x) be an arbitrary Boolean function. Let x, X. * and x, be any two variables, x„ > x. means that J i - 3 f(x» x. -* 1, x. -* 0) > f(x: x„ -> 0, x. -*■ 1) / l 3 ~ 1 and x > x. means that i 3 f(x; x„ ■* 1, x. ■* 0) > f(xs x„ -* O.'x. -* 1) ' i 3 i Theorem 2.2. U : Let f(x) be pseudo-separated by [W; T]. If x. > x., then -*- J w, > w.. If w. = w.- then all assignments A such that i - J i J f(A; x. -» 1, x. -+ 0) = 1 and f(A; x. -» 0, x . -* l) - must lie on the hyperplane [W| T], Proof: Assume x. > x . . Then there must be some assignment B. . to the i J i>0 other n-2 variables such that f(B. .; x. -* 1, x. -* 0) = 1 and f(B. .) x. -* 0, x . -* l) = 0. Since [W; T] pseudo-separates f(x), 1 > 3 i J 16 £ w °b f w. > T and Z w °b + w„ < T so that w. > w.o If fa.i p p x _ p/i,j p p ° " x " ° n n w = w , then Z w b ■*- w. = Z w b f w, = T must hold and these 1 J p^,j pp x P^i.j pp a true and false assignments must lie on the hyperplane [W| T]„ Q.E.D. The preceeding theorem is weaker than the corresponding result for separable functions, where x„ > x„ implies w„ > w. in all structures [7]. The next example shows that this weaker result is the strongest property which we can guarantee „ Example 2,2ol : Consider the function f (x) = x.x %s x x v x x, v x x „ f (xj x =*• 1, x p -* 0) = x^ v x> while f (x$ x -» 0, x -> l) = x . There- fore f (xj x.. -*• 1, x -* 0) > f (xj x„ -* 0, x -> l) and so x > x_„ Consider the hyperplane [1,1,1,1$ 2] If A is any true assignment, then A has more than 2 components of value 1 or else is one of (1,1,0,0), (1,0,1,0), (1,0,0,1) or (0,l,l,0)o In either case (l,l,l,l) »A > 2„ If B is any false assignment, then B has less than 2 components of value 1 or else is (0, 0,1,1) or (0, 1,0,1) so that (l,l,l,l) °B < 2. Therefore [1,1,1,1; 2] pseudo-separates f(x)„ Because w. = w = 1 and x > x , we may not strengthen Theorem 2 2,i+ to require that x. > x. implies w„ > w.o 1 3 1 J Another difference "between pseudo-separable and separable functions is that pseudo-separable functions need not have all variables related by this ordering, while for separable functions the relation IT "<" simply orders the set of all variables [73. Consider the function xx v xx,, which was shown in Example 1.2.1 to be pseudo-separable, f (x; x -* 1, x -» 0) - x and f(xj x. -* 0, x -* l) = x, , so that neither x l _ x -? nor x i - X V Definition 2.2.6 : Let f (x) be an arbitrary Boolean function. Let A = (a., a , ..., a ) and B = (b , b , ..., b ) be any pair of assign- > -> ments. Assume that A and B agree in position x. , x. , .... x. 12 p (i.e., a. = b. , .... a. = b. ) and disagree in x . , x. , .... x. where 11 P P 1 2 d q .i -» -» .-*. p + q = n. We will say that A and B are f(x; - comparable if either f(xj x. -* a . , x . -+ a .,..., x . -*■ a ) > J l J l J 2 J 2 J q J q f(xj x. -* b . , x. -»• b . , . . . , x. -* b . ) J l J l J 2 J 2 J q J q or f(xj x. •* b . , x. — b . , ..., x. -> b . ) > J l J l J 2 ^2 J q J q f(x; x. -* a. , x . -* a . , . . . , x . -+ a . ) . J l J l J 2 J 2 ^q J q The former case will be denoted by "A > B" and the latter by "A § B" . If A > B and f(x; x. -* a . , x . -* a. , ..., x. -* a ) ^ J l J l J 2 J 2 J q J q f (x; x. ■* b . , x -♦ b .,..., x . -* b . ) , then we will write "A > B" . J l J l J 2 J 2 J q J q f (x) is called completely monotonic if every pair of assignments is f(x) - comparable [8], [9], [11]. 18 Theorem 2.2.5 ' Let f(x) "be pseudo-separable under [Wj T]. Assume that f (x) is such that for two assignments A and B, A > B„ Then W°A > W°B„ Proof: Assume that A and B agree on variables x_ , x^, .... x and dis= 1 2 q. agree on variables x , n , x _,..., x where a , . = . . = a =1 and & q+1' q+2' ' n q+1 p a ,=.„.= a = 0(ri < p < n). Since A > B, there must exist some p-t-1 n — — ' -> -> pair of assignments C and D which also differ on x , . , x r ,. „„., x * q+1' q+2' ' n where C agrees with A on these variables and f (C) = 1, while D agrees with B and f (d) = 0, If W°A < W.B, then w ,_*■ w _-,f .. . + w < w , + ...+v v ' ' q+1 q+2 p p+1 n :\U~. and because C and D differ on exactly this set of variables, W«D > W»C„ This is impossible because f (S) = 1 and f (d) = 0. Therefore W°A > W.B. Q.E.D. Theorem 2.2.6 : Let f(x) be pseudo-separable under [W$ T]. Assume that f (x) is completely monotonic and that for any pair of assignments A and B, if A > B, then W°A > W°B„ Then f (x) is separable under [W$ T], Proof ; Let A and B be any true and false assignments respectively such that W«A = W°B = To Since f(x) is completely monotonic, A and B must be f (x) - comparable. Because f (a) = 1 and f (b) = 0, A > B. Therefore -+ -*■ -* -* W»A > W°B, a contradiction. Q.E.D. The concept of complete monotonicity also leads to an important difference between pseudo- separable and separable functions. Every , ; . separable function mast be completely monotonic, but : an arbitrary pseudo- separable function need not be so. To see this, again consider the 19 pseudo-separable function x x 4s x^k" Let assignment A "be (l, 1,0,0) and B be (1,0,1,0). These two assignments disagree on variables x and x . f(xj x 2 -* 1, x -* 0) = x and f (x; x_ -> 0,x -> l) = x, , which says that neither A > B nor B > A. 2.3 Bounds on the Degree of Hyperplanes and Conditions for Separation In this section, the relationship between the number of assign- ments on a pseudo-separating hyperplane and requirements for separability of a given function are discussed. Bounds on the degree of pseudo- separating hyperplanes are also investigated. Theorem 2.3.1 : Assume that f(x) is pseudo-separated by both [W j T ] and [W j T ]. Then f(x) is also pseudo-separated by [W + L T +■ T ] and the degree of [W + W ; T + T ] is the number of assignments which lie on both hyperplanes [W _j T J and [W ; ^ ]. Proof : We will show two things. First, if [W ; T ] and [L T ] pseudo-separate f(x), then [W. +■ W j T + IE ] also pseudo-separates f(x), and second, if A is any assignment that is segregated by either [W^ T ] or [W 2 j T ], then A is segregated by [W^ + ft ; T + T g J. (l) Let A be any assignment. If f (A) = 1, then W-A > T^and W-A > T and (% + W rt ) «A > T, + T„. If f (A) = 0, then W n -A < T n and 2-2 12' -12 ' 1-1 W 'A < T and (W + W ) 'A < T. + T . Therefore [W +■ W > T +■ T ] pseudo-separates f(x). 20 (2) Assume that A is segregated "by either [W , T ] or [W ; T ]. Then either W «A / T or W* 2 °A / T^ Therefore (W^ f W 2 )°A / ^ + T 2 and A vill he segregated by [W + Wj T + T J. 1 * Q E„D« When hyperplanes [W_ $ T_] and [W-j T ] have no assignments in common or when all common assignments have the same value, then the _— >• — ► = hyperplane [W. + W j T f T J must separate the function. 1 2 1 ^ Corollar y 2o3°l° Assume that f(x) is pseudo-separated by [W.^ °, T. ] and [W_j T ]„ Let P he the set of all assignments which lie on hyperplane [W. 5 T. ] and Q, he the set of all assignments which lie on [W p $ T ]„ If PTIQ = 0, f(x) is separated by [W^ + W g j T ± + Tg], Corollary 2o3°2: Assume that f(x) is pseudo-separated by [W j T. ] and LW j T Jo If all assignments A„ which lie on both hyperplanes have the same value, f(A„) = a, then f(x) is separable under [W- f W_^ T + T ]„ Example 2o3°l ° We can illustrate this last result hy considering the function which was defined in Example 2 2„1<, Let f (x) = xx v x„x v x p x v x.x, „ We have already seen that it is pseudo- separated hy [l,l,l,lj 2] The assignments which lie on this hyperplane are (1,1,0,0), (1,0,1,0), (1,0,0,1), (0,1,1, 0), (0,1,0,1) and (0,0,1,1). The first four are true and the last two are false „ If we can find a hyperplane which pseudo-separates f(x) and segregates (0, 1,0,1) and (.0,0,1,1), we will then he able to add the two structures to generate a 21 new hyperplane which will separate f(x). Consider the hyperplane [2,1,1,0; 2]. Let A he any assignment such that (2,1,1,0). A > 2, We want to show that f (A) = 1. A must he such that a = 1 and either a = 1 or a - 1. In either case, f(A) = 1. If (2,1,1,0) »A < 2, then a = and either a = or a = 0, therefore f(A) = 0. Therefore [2,1,1,0; 2] pseudo-separates f(x). The assignments which lie on [2,1,1,0; 2] are (1,0,0,0), (1,0,0,1), (0,1,1,0) and (0,1,1,1). The only assignments common to both [1,1,1,1; 2] and [2,1,1,0; 2j are (l, 0,0,1) and (0,1,1,0) which are both true. Therefore, [3,2,2,1; k"} must separate f(x) and the only assignments which lie on this hyperplane are (1,0,0,1) and (0,1,1,0). Definition 2.3.1 : Let f(x) be a positive function. Let A = (a ,a p ,...,a ) and B = (b ,b , . . . ,b ) be any pair of assignments, A and B are comparable (as opposed to f(x) - comparable used in Definition 2.2.6) if for all i = 1, 2, . . . , n, a, > b. or for all i, t>. > a.. We will denote the i-i ' i — i former case by A > B and the latter by A < B. If A > B and for some j, a > b_, we will write A > B. A is an extremal true assignment if f(A) = 1 and for all B such that B < A, f (B) = 0„ A is an extremal false ass ignment if f (A) = and for all B, B > A, f (b) =1. A is an extremal assignment if A is either an extremal true or an extremal false assignment. If A is not an extremal assignment, A is called non- extremal. Theorem 2.3.2 : If f (x) is pseudo-separated by a zero-free structure [W; T], then all non-extremal assignments must be segregated by [W; T], 22 Proof ; If f (x) is pseudo-separable under a zero=free structure, then f (x) must be unate "by Corollary 2<>2»3o We will assume that f (x) is positive. Therefore, all w„ > by Corollary 2o2„U and because [Wj T] is a zero-free structure, all w. > 0. l Let A be any non-extremal true assignment . (A similar proof -*■ . will hold if A is any non-extremal false assignment „ ) Then there exists -> —»-»-* another true assignment B such that B < A„ If A is not segregated by — *• _ -> — > — > -» -*■ [W$ TJ, then W-A = T„ Because B is also a true assignment, W°B > T Therefore W° (A - B) < 0„ Since A > B. there must be at least one a. > b„< — ' 11 Hence W» (A - B) > w„ „ Because all weights are positive, we have a °"* r"* -\ contradiction,, Therefore A must be segregated by LW$ TJ» Q.E.D, Theorem 2„3°3 ; let f (x) be a positive function., Assume that f (x) is pseudo-separated by a zero-free structure [Wj T] Then the degree of [W; T] is less than or equal to 2([pl) . Proof : Since [Wj T] is zero-free, f(x) is unate and 2 ([~]) is the maximum number of possible extremal assignments [7]° By the previous theorem all non-extremal assignments must be segregated by [W; T]. Therefore 2 ( [r] J is the maximum number of possible assignments which can lie on [W> T}. Q.E.D. Theorem 2.3.^ : Assume :that f(x) is pseudo-separated by [Wj T], Let -►■-+• -*•♦-+ -*■ -* _ A , A , . „„, A, , B , B , ,.., B be all assignments which lie on LW3 TJ„ J_ <— K i_ c. 6 Let [w j T ] be any hyperplane which contain^ A_ , A , „ 00 , A, and 23 segregates B , B , . . „ , B 'i: ( [W ? T ] need not pseudo -separate f(x)i ~* ~*1 -*• 1 There may exist true assignments C. such that W D C. < T and false assign- ments D. such that W °D„ > T „ However such assignments may not lie on [W> T]). Then there exists a hyperplane [W -, T ] which pseudo- .-» -» -* =-* separates f(x) and segregates B , B , „.., B together with all assign- /J. c. e ments which are segregated by [W> T]. r~*l - l-i '~*\ Proof : The problem nere is that I.W ; T J need not pseudo-separate f(x), r ~* ~*1 1- Consider the set of structures LmW + W f mT + T j where m is any positive integer. We will show that for some m, the resulting hyperplane pseudo- separates f(x), segregating B , B , „„,, B and all assignments segregated by [W; T], (1) Let D. , D , ooo. D be all false assignments such that L d. S — >~]_ "■* 1 -» -* -» W. 'D. > T . Let C, , C_, . . „ . C^ be all true assignments such that i 1' 2 t -+1 -* 1 W °C. < T o Let M be the minimum positive integer such that for all D. and C. , [^.C. - T 1 ] - M[T - W-C, ] > and [T 1 - W^D. j - MLW.D. - TJ > 0„ Because all such D, and C, were segregated by [W; T], W-C. > T and W°D> < T and so we have [T - (W»C. )] < and i l l [(W°D.) - T] < Oo Therefore, such a positive M can always be found, (2) We now show that [MW + W j MT + T j pseudo-separates f(x). -» — » — * -*j_ — * 1 Let A be any assignment,, If W°A = T and w "A = T^ then (MW + W )»A = MT + T , Therefore, if A is not segregated by either [W; T] or [W 1 ? T 1 ], then A is not segregated by [MW + W^ MT + T 1 ], "* -*• -* r^l 1- "? If W°A = T and A is segregated by LW$ T j, then A is segregated by 2k [MW + VT f MT + T ] because if A is a true ^assignment, w °A > T and (MW + W )*A > MT + T and if A is a false assignment ^T°A < T and (MW + w )»A < MT + T . If A is a true assignment "which is segregated by LWj TJ, then there are two different possibilities, ~*1 ""*• ^ 1 ~tl °* 1 -»■ either W^-A > T or W^A < T „ In the first case A must be segregated by [MW + W 1 ; MT + T 1 ] because (MW + W 1 ) -A > MT + T . If W^A < T, then A must be one of the C. m part (,!;„ (MW + W 1 )-^ - (MT + T 1 ) = [(W^A) ? T 1 ] - MfT - (W° A) ] and by the choice of M in part (l), this expression is always positive. Therefore (MW + W;°A > MT + T and A is segregated by this structure A similar proof holds if A is a false assign- ment which has been segregated by [W$ T]. Hence [MW + W; MT + T ] pseudo-separates f (x) and segregates all assignments segregated by Q.E.D. either [Wj T] or [W X j T 1 ] Corollary 2. 3. 3 « Assume that f(x) is pseudo-separable under [W$ Tj and that there exists another hyperplane [Wj T ] which segregates all assignments which lie on [W; T], then f(x) is separable. Proof : If [Wj T ] segregates all of the assignments which lie on [Wj T], hhthefaii'! [MW f W; MT + T ] will separate f(x) since all assign- ments lying on the hyperplane [Wj TJ will be segregated. Q.E.D. 25 Certain sets of assignments lying on a hyperplane are always separable. One such set consists of only one true assignment and any number of false assignments as the next theorem shows. Theorem 2. 3. 5 - Assume that f(x) is pseudo-separable under [Wj T] and that there exists only one true (false) assignment and any number of false (true) assignments which lie on [Wj Tj. Then f(x) is separable. Proof -; We need only show that there exists some hyperplane which — > segregates the one true assignment, A, and by applying Corollary 2. 3° 3, the result follows. Assume that A = (a n , a OJ .... a ). Let v 1' 2' n a n = a_ = . . . = a =1 and a , = a n = ... = a =0. Let w. = 1 if 12 p p+1 p-t-2 n l a. = 1 and w. = -1 if a. = 0. Let T. = p - 1. Clearly i l l :. -*1 -* . 1 -> r -»l i- W «A = p > T = p - 1 so that A is segregated by IVT°, T J. For any ■* -*1 -» 1 r -*l 1. other assignment B, W -B < p - 1 = T , Therefore Lw \ T J segregates the set of assignments which lie on [Wj T] and so f(x) is separable = Q.E.D. Definition 2.3.2 : Let {A 1 , A , . .., A J be a set of assignments. [A , A , ..., A ] is called consistent if the rank of the matrix! V - u s) i" . i •. ■; i , ' K, 1, f(A ) is the same as the rank of the augmented matrix! : : \, 1, f(A k ) Theorem 2.3.6 : Let (A^, A , . .., A k } be a consistent set of assignments, -* . Let A. = (a. , a. , ..., a. ). Then this set is separable if the rank 1 2 X n of the augmented matrix in Definition 2.3.2 is less than n + 1. 26 Proof : Consider a set of linear inequalities given "by W°A. - T = for f(A. ) = 1 and W-A. - T = - 1 for f(A. ) = 0. If the rank of this l i l matrix is less than n + 1, and if the assignments are consistent, then we can have a solution [W; TJ, This yields a segregating hyperplane for I A, ) A , . . „ , A j „ 1 2 k Q.E.D. Corollary 2o3.^ ' Assume that f(x) is pseudo-separable under [W; T]„ -> -» -> . Let [A , A , „ . „ , A } be the set of all assignments which lie on hyper- 1 (— K. plane [W- T]. Then if {A.*, A , . .., A } is consistent and if the rank I is less than n + 1, f (x) is separable. a ll a i2 * ' ' a ln' a kl a k2 * ' ' a kn Proof : If this set is consistent and if the rank is less than n + 1, then by the previous theorem, there exists some hyperplane [Wj T ] which segregates this set. Since this set includes all assignments which lie on [Wj T], f(x) must be separable by Corollary 2. 3° 3° Notice that the above corollary also puts a lower bound on the number of con- sistent assignments which can reside on the hyperplane „ Q.E.D. 2.M Asummability .: as, ,'a Condition for Pseudo-Separabilit y • In this section the idea of asummability is introduced and from this property, a necessary and sufficient prerequisite for a function to be pseudo-separable is developed. This condition is also used to explore other algebraic properties of pseudo-separable functions „ 27 Definition 2.4.1 : Let f (x) be an arbitrary function. f(x) is called k-summable if there exist k true assignments A , A , . .., A and k 1 <— ri false assignments B , B , ..., B , with possible repetitions, such that k k E A. = S B. . f(x) is called asummable if it is not k-summable for . _ l . . l i=l i=l any k„ (See Example 2.4.1.) The following theorem is due to Elgot [2] and Chow [l], and the proof is omitted here. Theorem 2.4.1: Let f (x) be an arbitrary Boolean function, f (x) is separable if and only if it is asummable. In developing a necessary and sufficient condition for pseudo- separation, we would like to refer to the summability or asummability of particular assignments. Therefore we introduce the next definition. Definition 2.4.2 : Let A be any true (or false) assignment to f(x). A is called summable if for some k, there exist k - 1 true (false) assignments A , A , . .., A and k false (true) assignments d 5 K k B , B , ..., B , with possible repetitions, such that A + L A. = 1 Cm K 1 . ,-. 1 1=2 k B. . Note that in such a case all of A n , A . .... A. and i=l —>—>—> — » B , B , ..., B are summable. An assignment A is called asummable if it 1 £- K. is not summable. 28 Lemma 2.4.1 : Let f(x) be pseudo-separated by [Wj T]. If A is segregated by [W; T], then A is asammable. Conversely, all summable assignments must lie on the hyperplane [W$ T]. -> , Proof : Assume that A is a true assignment. (A similar proof can be given if A is any false assignment) .Let W.A > T. If A is summable, — > — * — * then there exist true assignments A , A , . . . , K, and false assignments b\ , B , . . . , B. such that A + E A. = Z B, . Therefore, 1 ' 2' k . i . , i i=2 i=l k k W- (A + E A. ) = W° ( E B. ) . Because W-A > T and each W.A. > T, . „ l . , l l — i=2 1=1 k W« (A + Zj A. ) > kT. Because each B:. is a false assignment, W-B. < T l ■ • i i — i=2 x k so that W- ( .E.. B.) < kT. Therefore, we have a contradiction. Hence, i=l i — ' -» ^-> _ A can not be summable if it is segregated by LWj TJ. Q.E.D. Corollary 2.4.1 : Assume f(x) is pseudo-separable under a zero-free structure [W; T]. Then all non-extremal assignments must be segregated by [W; T] and hence must be asummable. Proof : If f (x) is pseudo-separated by a zero-free structure, all non- extremal assignments must be segregated (by Theorem 2.3.2) and hence must be asummable (by Lemma 2.4.1). Q.E.D. Definition 2.4.3 ' f (x) is called super summable if, for some k, there exist k true assignments A , A , ..., A and k false assignments I c. K. k k B , B c , ..., B , with possible repetitions, such that E A. < E B„ . 1=1 1=1 29 If we let each A. = (a. , a. , . .., a. ) and 1 i, i_ l 12 n B. = (b. , "b. , .... b. ), then the above condition is fulfilled when 12 n k k k k Z a. < Z b. for j = l,2,...,n and for some j = p. Z a. < Z b. . i=l j i=l j i=l p i=l p Example 2.4.1: Consider the function f(x) = x n x_ v x x, x . (1,1,0,0,0) 12 j 4 ^ and (0,0,1,1,1) are both true assignments while (1,0,0,1,0) and (0,1,1,0,1) are both false. . Because (l, 1,0,0,0) + (0,0,1,1,1) = (1,1,1,1,1) = (1,0,0,1,0) + (0,1,1,0,0), these assignments are summable. Therefore, f(x) is not separable by Theorem 2.4.1. Now consider false assignments (1,0,1,1,0) and (0,1,1,0,1). (1,0,1,1,0) + (0,1,1,0,1) = (1,1,2,1,1) > (1,1,1,1,1). Therefore, f(x) is supersummable. Definition 2.U.4 : Let f(x) be a supersummable function. f(x) is said to be supersummable in variable x if for some k, there are k true — c -P assignments A , A , ..., A and k false assignments B , B , ..., B , with _L c. K. _L c. K k k k k possible repetitions, such that Z A. < Z B. and Z a. < Z b. . .,i.,i .,i.-.i 1=1 1=1 i=l p 1=1 p In the previous example f (x) is supersummable in x and by symmetry is also supersummable in x, and x . Lemma 2.4.2 : If f(x) is a positive supersummable function, f(x) is summable. 30 Proof : Let true assignments A , A p , , . . , A and false assignments k -> k B, , B_, . ... B, be such that Z A. < Z B. . Let 12 k . , 1 . , i 1=1 1=1 A. = (a. , a. , .„., a, ) and B. = (b. , b. , .... b. ). Assume i i, i ' i i i , i ' i 12 n 12 n k k Z a. = Z b„ for all j except j = p, , p., . .., p . Consider those . , i . . , l . L d q i=l j i=l j H k k assignments B for which b = 1. If Z b. - Z a. = s of these . , l ■ i ■ l i=l p x i=l p l e assignments are changed to the corresponding assignments with b = 0, 6p ] -> k k_s s and renamed B , then .Z a. = Z b. + Z b. Since f(x) is e ' . - l . -, i . -. l i=l p x i=l Pl i=l ?1 positive and B > B , f (B ) = implies f (B ) = 0. If this procedure k is repeated for each of p n , p^, .... p , the result will be Z a. 1 2 q . . l . k k ^ k Z b q for all j so that Z A. = Z B. q . i=l X j i=l X i=l X Q.E.D. Example 2.4.2 : Consider the function x x \y x x, . (l, 1,0,0) and (0,0, 1,1) are true assignments and (0, 1,0,1) and (l, 0,1,0) are false assignments. Because (l, 1,0,0) +■ (0,0, 1,1) = (l, 1,1,1) = (0,1, 1,0) + (l, 0,1,0), the function is summable. Now assume that it is supersummable. Then there are true assignments A , A , ..., A and false assignments B , B , ..., B k -> k such that Z A. < Z B. . Therefore, if we let 1 = (l,l,l, . . . ,l) , then i=l x i=l x k - - k 1- ( Z A.) < 1*( Z B.). Each true assignment A. has at least . _ i . , i i — — ■ — — ■ i=l i=l k 2 components of value 1. Therefore* 1-A. > 2 and 1- ( Z A.) > 2k. l — . , l — i=l 31 Since each B. is false, B. has at most 2 components of value 1. Hence, k 1-B. < 2 and 1- ( E B.) < 2k. Therefore, it is impossible to have i=l k -* k 1- ( E A.) < 1- ( E B. ) and x.x. v x x, is not supersummable. 1-1 x 1=1 x 12 3 4 Lemma 2.4.3: If f(x) is positive and supersummable in x and f (x) is pseudo-separated "by LW; TJ, then w = 0. k k Proof: If f(x) is supersummable in x , then for some k, E A. < E' B. * i=l i=l k k and E a. < E b. , where A. = (a. , a. , .... a. ) are true assignments . , i . , i 1 1-, i ! i=l p i=l p 1 2 n and B. = (b. , b. , .... b. ) are false assignments. If w > 0, then we V X 2 P k _> - k have W- ( E A. ) < W. ( E B. ) because of w; > 0. Since each A. is true, .,i . , l i- l i=l i=l k k W-A. > T, so that W- ( E A. ) > kT. Similarly W- ( E B. ) < kT. This l — . , l — .,i — 1=1 1=1 k k contradicts W- ( E A. ) < W. ( E B. ) . Therefore, w =0. i=l x i=l X P Q.E.D. Theorem 2.4.2 : If f(x) is supersummable, f(x) can not be pseudo- separated by a zero-free structure. Proof : If f (x) is pseudo-separated by a zero-free structure, then it must be unate by Corollary 2.2.2. Assume it is positive. Then because f(x) is supersummable in some x , some w = by Lemma 2.4.3, a p P contradiction. Q.E.D. 32 Theorem 2.4.3 : Assume that f (x) is a positive Boolean function,, Then f(x) is supersummable if and only if some non-extremal assignment is summable. Proof : (l) Let A be a summable non-extremal true assignment. Then k - k A + Z A = Z B.. Since A is not extremal, there is some true 1=2 i=l assignment A such that A < A. Therefore A 4- Z A. < Z B. and f (x) . i . , i 1=2 i=l is supersummable, (2) If f(x) is supersummable, f(x) is summable by Lemma 2.4.2. Notice that the B used in the proof of Lemma 2.4.2 are non-extremal assignments. Therefore a non-extremal assignment must be summable. Q.E.D. The following theorem presents the major result for pseudo- separation by zero-free structures. Theorem 2.4.4 : Let f(x) be a positive function. Then f(x) is pseudo- separable under a zero-free structure if and only if f(x) is not supersummable. Proof : (l) By Theorem 2.4.2, the condition that f(x) not be super- summable is necessary for f (x) to be pseudo-separable under a zero-free structure. (2) We now want to show that this condition is sufficient. Because all segregated assignments must be asummable and all summable 33 assignments must lie on the hyperplane by Lemma 2.4.1, we can consider the following set of linear inequalities: (a) W'A. > T + 1 if A. is asummable and true i — l (b) W-U. = T if U. is summable (either true or false) J J (c) W*B, < T - 1 if 1 is asummable and false k — k — > Clearly, if we can find W and T which will satisfy the above inequalities, Vre will then have a hyperplane which pseudo-separates f (x) . We can combine inequalities of the form (a) and (b), (a) and (c), and (b) and (c) to generate another set of linear inequalities. (a) • W-(A. - U.) > 1 l j - (b) ' W-(A. - B, ) > 2 i k — (c)» W. (U. - B ) > 1 J k A necessary and sufficient condition for this set of inequalities to be inconsistent and therefore not to have any solution [Wj T] is that there exists a. . > 0, b., > and c ., > such that .E. a. . + Z 2b.. ij — lk — ik — . . ij . . lk + E c., - > and E a. . W- (A. - IJ.) + .E. b., W- (A. - B, ) •+ E c, W. (U. - % ) < i,j 1J X J i,k lk X k j,k ^ J k " (Fan, [2]). Notice that in order for the above system to be inconsistent and all w. ^ 0, this last inequality must be either identically zero or identically less than zero. Therefore, if a solution [W; T] does not exist, 3^ .£. a.. (A. -U.)+ E b..(A. -B.) + E c to -\)<0 or i,J i,k j,k E a. . A. + E b., A. + E e U. < E a. . if. +■ E b.. S. + E c ., fi . i,u 1J x i,k lk x j,k -> k = ~ i,j 1J J i,k lk k ; j, k J k k - Therefore E a. + E u < E u. + E b. (2.U.1) i=i x j=i 3 j=i J i=i i where p = E a. .+ E b ,q= E c „ and r = E a . . . i„j 1J i,k ik J,k o k i,d « We now want to show that if condition (2.U.1) is fulfilled, then fix) is supersummable. Let us divide U. into true assignments x. i i 3, s y 1 ~*1 v ~*1 ~*1 U. = Z, x. f L Y. where x. are true and 3=1 J 3=1 J 3=1 J J n s q.- s a ~*1 V ""* V ~* V ~* Y. are false* Substituting L x. + L Y. for L U. and J 3=1 J 3=1 J 3=1 J t -+1 r_t -tl r -» E x. + E Y. for E U., we can rewrite (2.4.1) as: 3=1 J 3=1 J 3=1 J E _ s _ q- s _ t _>i r -"t _*i p+q-r E A. + E x. + E Y. < E x. + E YT + E B. (2.4.2) i=l x 3=1 J 3=1 J 3=1 J j=l 3 i=l x We want all true assignments on the left side of this expression and — > false assignments on the right. Each Y. is summable, therefore g g Y + L V = L Z where Z are true assignments and V are false, jmm m m m=2 m=l g g . Therefore, we can replace each Y. in (2.4.2) by E 7? - E V and add <] , m m u m=l m=2 35 L V to both sides of this expression. Then we will have effectively m=2 eliminated one false assignment from the left side of (2.1+.2). If we repeat this process for each Y. and perform the corresponding substitution J -*1 and addition for each x., we will arrive at a summation of the following J form: p q+d ^ P" 1 " 3, ■*■*■_> r4 "d Z A, + Z T, < L B. + Z F. (2.U.3) i=l x j=l J i=l x j=l J -» -> -». -*• where all A. and T. are true assignments and B. and F. are false assign- 1 J 1 j ments. If (2.U„3) is an equality, we have a contradiction, since we — »■ -> have assumed that all of the A. and B. are asummable, Therefore, 1 1 ' p q+d ^ p+q-r r+d Z A, 4- Z T. < Z B. + Z F. and f(x) is supersummable,, i=l X j=l J i=l X j=l J Therefore, if f(x) is not supersummable, then f(x) is pseudo- separable under a zero-free structure, Q.E.D, Corollary 2.^.2: If f (x) is positive and pseudo-separated by a zero- free structure, then all asummable assignments can be segregated by some hyperplane [W; T], Proof : If f (x) is pseudo-separated by a zero-free structure, f (x) can not be supersummable by Theorem 2,^.2. If f(x) is not supersummable, the proof of Theorem 2„k„k guarantees the existence of a structure which pseudo-separates f(x) and segregates all asummable assignments- Q.E.D. 36 If we combine the results of Theorems 2.^.3 and 2.k.k, we have the following important corollary: Corollary 2.U.3 : Let f (x) be a positive function, f (x) can be pseudo- separated by a zero-free structure [Wj T] if and only if all of the non-extremal assignments are asummable. The previous results require zero-free structures. These are more restrictive than non-trivial structures. The following example exhibits a function which is supersummable, but can be pseudo-separated by a non-trivial (but not zero-free) hyperplane. Example 2.U.3: Consider the function f(x) = x n x^ v x x x v x_x x^. — 12 135 2po Since assignments (1,0,1,0,1,1) and (0,1,1,1,0,1) are false, and (1,0,1,1,0,0) and (0,1,0,0,1,1) are true and (1,0,1,0,1,1) + (0, 1,1,1,0,1): (1,1,2,1,1,2) > (1,1,1,1,1,1) = (1,0,1,1,0,0) 4- (0,1,0,0,1,1), this function is supersummable. Consider the structure [1,1,0,0,0,0) l]„ — > The only assignments segregated by this structure have x = x . Let A be any segregated assignment. If a = a - 1, then the assignment must be true and (1,1,0,0,0,0) 1. If a = a = 0, then A must be false and (1,1,0,0,0,0) °A = < 1. Therefore, [1,1,0,0,0,0; l] pseudo- separates f (x) . Necessary and sufficient conditions for pseudo-separation of a positive function by anon-trivial hyperplane appear to be much more difficult than the zero-free case. However, much information can be obtained by reducing this problem to the zero-free case by partitioning 37 the set of variables x into two subsets, x , variables whose corresponding weights are zero and x , variables whose weights are non-zero. Notation 2.4.1 : Assume that a positive function f(x) is pseudo- separated by [Wj T]. Assume that w = ... = w =0 and all w. > for i > e + 1. Let x q - (x^ Xg, . .., x g ) and x^ = (* e+1 , x g+2 , . .., x^) Then we have the following expansion because of the positiveness of the function. 1) ±u. ) = XL...X fi 1 e ^', x x — > 1, ... V x r ., ,x e _ x f (x, x 1 - ♦ 1, .. , . , Vl - 1, X e V . '• • v f(x"j X 1 - 0, , , x^O). Let f 1 (x 1 ) = f(x, X;L - l,.. 4 ,x e - 1), f 2 (x x ) = f (xj x x - 1, . . . , x e _ x -* 1, x g -» 0) , , f, (x\ ) = f (x: x n -> 0, . . . , x -* 0) where k = 2 k 1 ' 1 e e Lemma 2.4.4 ; If [Wj T] pseudo-separates f(x) where w = ... = w = — > ^ — > and f (A ) / f (A ) for some s and t where A is an assignment of (n - e) variables, instead of all variables, A., = (a ,_, . ... a ), \ / > '1 e+1 n then A must be extremal for every f.(x n ). 1 J x 1 Proof: Let A = (A , A n ) where A = (&., . . . , a ) is any assignment on ■ — o 1 o 1 e n e variables. Then W-A = Z w.a„ is well defined. Since w. = . . . = w = 0, .,11 1 e i=l 38 the first e coordinates do not matter. If f (A ) f f (A ) , then S JL Xj 1 W*A = T because lWj TJ pseudo-separates f (x) . If A is not extremal for f . (x ), then there exists B, such that f . (B ) = f . ( A ) and B > A u J J ^ 1 ± or B, < A n depending on whether A., is false or true for f ( x ) . If 11 1 J 1 f(A ) = 0, then f(B ) - and B > A so that W-B > W-A = T, a contradiction. If f(A ) = 1, then f (B ) = 1 and A > B so that W-B < W*A = T, a contradiction. Q.E.D. Many of the properties which were discussed earlier in this section are applicable to the set of functions, f , f , . .., f „ 1 c- K These will be merely stated without proof. After each the corresponding theorem or lemma for single functions is given inside parenthesis. Lemma 2.k.^ : If A is segregated by [Wj T] where w = . . . = w =0, then A must be asummable in all f . (x ) and if A is summable in any f „(x ), A must lie on the hyperplane [Wj T]„ (Lemma 2.4.1) Th eorem 2.k.5 - No f „ (x ) can be supersummable. (Theorem 2.4.2) The orem 2.4.6 : Every non-extremal assignment of each , f . (x..) must, be segregated by [Wj T] where w = ... = w =0. (Theorem 2„3.2) Theorem 2.4.7: Let f(x) be a positive function„ Let (x. , x , .... x } 12 e be the set of variables in which f (x) is supersummable. If (— > . — > A ; x , -*• 0, . . . , x -» 0) = 1 for some assignment A and if for each o' e+1 ' n ' & o 39 x.(e+l < i < n) there exists some assignment B . such that J - - oj f( V X e + 1^°^ ■•" Y 1 ^' T 1 ' X 3 + l- > °' ~" X n~* G) = °> then f (x) is not pseudo-separable . Proof : If [W; T] pseudo-separates f(x), w, = . .. . = w = by Lemma 2.4.3 .-->. -> because f(x) is supersummable in all of these variables. Let A be the -» n-tuple whose first e components are A and whose last n-e are all 0. Then f(A) = 1, but W-A = 0. For each x., (e +■ 1 < j < n), there exists J an assignment B . on fx_ , x^, .... x ] such that & oj l 1' 2' ' e f(B .; x , -> 0, .... x. , -* 0, x. -> 1, x. , -> 0, .... x -* 0) = 0. v oj' e+1 ' ' jil ' d ' j+1 ' n — » — » Let B. be the n-tuple whose first e coordinates are B . and whose last 3 oj th n-e are 0, ..., 0, 1, 0, ..., with 1 in the j position. Then W-A > W*B. "because f(A) = 1 and f(B.) = 0. Because W-A = and W-B. = w., 3 J 3 3 and since all w. > because f(x) is positive, w. = 0, Therefore all J J w. = and only the trivial hyperplane pseudo-separates f(x)„ 3 Q.E.D. Since f (x) and f (x) are pseudo-separable together by Theorem 2.2.3, Theorem 2.4.7 can be extended to the next corollary. Corollary 2.4.2 : Let f (x) be positive and supersummable in {x ,x , . . . ,x } If f(A j x , -+ 1. .... x -> 1) = for some A and if for each o e+1 n o x.(e+l < j < n) there exists some assignment B . such that J - - oj f( V Vi"* 1 ' •••/ Vi" 1 ' x j -* > Vi"* 1 ' ■■•' \^ z) =1 > then f(x) is not pseurlo- separable o Example 2.4.3 : Consider the function f (x) = x.x v x x, x . We showed in Example 2. 4.1 that this function is super summable in x , x, and x . Since (0,0,1,1,1) is a true assignment and (1,0,0,0,0) and (0,1,0,0,0) are false assignments, f(x) is not pseudo-separable. Henceforth, results under the condition that [Wj T] must be zero-free will be presentedo Theorem 2.4.8 : Let f(x) be an arbitrary positive function. Assume that f (x) is completely monotonic and pseudo-separable under a zero-free structure [W; T]. If there are only 2 true assignments A and B and two false assignments C and D which lie on [Wj T], then f(x) is separable Proof : If f(x) is asummable, then it is separable by Theorem 2„4„1. Assume that f(x) is not separable. Since A and B are the only true assignments which lie on [W; Tj, they are the only summable true assign- . -> -» ments by Lemma 2.4.1,, Similarly C and D are the only false summable .-» ->-*->-» -> -» assignments for f(x). Therefore, C + D = A + B. Assume that A and C agree in positions x n , x n , . ..< x and disagree in x ., x , _.,..., x . v 1' 2 q q+1 q+2' ' n — » — » — > — > — > -> If A t B = C + D, B and D must also disagree on x , , x _,„.., x and ' q+1 q+2' ' n agree on x., , x^, . ... x . Futher a. = d. for j = q+1, q+2, . ... n and 12 q 3 3 b. = c. on this same set of components because a. + b. = c. + d. and 3 3 3 3 3 3 kl Then we have that A = C and B = D . q q q q f(A : x , -» a ,, x _ -» a _,..., x -> a ) = 1 = f(A), v q' q-t-1 q+1' q+2 q+2' ' n n J K ' ' and f(A jx - -» c , , x ~ -» c _,..., x -> c ) = = v q' q+1 q+1' q+2 q+2' ' n n y f(C ; x , -* c _ , x -*c _,..., x -> c ) = f(C). v q' q+1 q+1' q+2 q+2' ' n n y v ' -» £ -» This means that we can not have A < C. Also, f (D S x l -* a , # x ~ ^ a ^, . „ . , x -» a ) = v q' q+1 q+1' q+2 q+2' ' n n y f(D -, x - -> d . , x -»d _,..., x -> d ) = f(D) = 0, s q' q+1 q+1' qt-2 q+2' ' n n J v ' ' f (D ix v^ c -, f x ^ "** c ^> • • ° > x -»■ c ) = v q* q+1 q+1' q+2 q+2' ' n rr and f (B j x n ^ b ., x -» b _, . . . , x - b ) = f (B) = 1. v q' q+1 q+1' q+2 q+2' ' n n y v ' This means that we can not have A > C. Therefore A and C are not f (x) - comparable and hence f(x) can not be completely monotonic. Q.E.D. Lemma 2.4.6 : If f (x) is self -dual and A is any true (or false) assign- -* ment, then A is a false (or true) assignment. Proof : f d (x) = f(x). Therefore, if f(A) = 1, then f(A) = 1 or f(A) = 0. Q.E.D. Lemma 2.U.7 : If f(x) is self -dual, then any assignment A is summable if — * and only if A is summableo h2 — » Proof : If A is true and summable, then there are true assignments — > -> — * —►->-> A , A , . .., A and false assignments B , B , . .., B such that k _> k - A + Z A. = Z B. . Because A = (l, 1, 1, . .., l) -A, i=2 x i=l x -»k^ k _, k -* k -* -* (A + Z A ) = (k, k, . . . , k) - Zl. -X Therefore, Z B. = Z A. + A. i=2 i=2 i i=l x i=2 x Since f(x) is self -dual all A, A , A , . .., A are all false assignments, ->-*-» -^ -* while B. , B_, .... B are all true. Therefore, A is summable. 1 2 v k Q.E.D. Theorem 2.k.Q : If f (x) is self -dual and pseudo-separable under some — > zero-free structure, then any assignment A is segregatable if and only if —* A is segregatable. Proof : We know that if f (x) is pseudo-separated by some zero-free structure, then all of its asummable assignments can be segregated by some hyperplane [W; T] by Theorem 2.^.4. A is summable if and only if — — > — A is summable by Lemma 2.4.7« Hence A and A must be segregated together by this structure. Q.E.D. 2.5 Simultaneous Realizability and Pseudo-Separability . In this section the concept of simultaneous realizability is examined in relationship to pseudo-separability. The major results of this section are that the upper and lower bounds (see Definition 2.5*2) for any refinement (see Definition 2.5.3) of a pseudo-separation are simultaneously realizable. This idea is also applied to the question of pseudo-separability for completely monotonic functions and to other algebraic properties. U3 Definition 2.5.1 : Let g (x) and g p (x) be two separable Boolean functions. We will say that g. (x) and g p (x) are simultaneously realizable , if there exist some set of weights W and two thresholds T and T_, such that g 1 (x) is separated by [W; T ± ] and g (x) by [W; T p ]. [7] Definition 2.5.2 : Let f (x) be a pseudo-separable Boolean function on exactly n variables. Let [W; T] be any structure which pseudo-separates f(x). Let M be the minimum T = W°A. for all assignments that are segregated by [W; T] (i.e. W.A. ^ T). Let < e < M. Let g(x") be the — » — > separable function on exactly n variables such that if W°A > T - e, then g(A) = 1 and if W°A < T - e then g(A) = 0. Let h(x) be the separable function on exactly m variables such that if W°A > T -f e, then h(A) =1 and if W-A < T + e, then h(A) = 0. g(x) is called an upper bound for [¥j T] and h(x) is called a lower bound for [Wj T], If [Wj T] is the structure which segregates all segregatable assignments of f(x), then we will call g(x) the least upper bound and h(x), the greatest lower bound for f (x) . Notice that in the above definition, € is defined small enough so that for any assignment A of f(x), h(A) = and g(A) = 1 if and only — » — > if W°A = T. To facilitate the definition of € for each distinct hyper- plane [Wj T], we will refer to the function M(-WjT). This is meant to represent the minimum T - W«A. for all A. segregated by LW$ TJ. Example 2.5.1 : Consider the function x x s/ x o x ), ^ e saw ^" n ■$^ a . m .P.^. e . 1/2..1, that this function was pseudo-separated by [1,1,1,1$ 2]. The upper bound for this structure is x x v x x v x x, v x x, , which has hk structure [1,1,1,1; 1.5]. The lower "bound for this structure is separated by the hyperplane [1,1,1,1; 2.5] and hence must be xxx v xxx, sy xxx, v xxx,. The true assignments which lie on [1,1,1,1; 2] are (l, 1,0,0) and (0,0, l,l). The false assignments which lie on [1,1,1,1, 2] are (l, 0,1,0), (0,1, 0,1), (l,0,0,l) and (0,1, 1,0). Since (1,1,0,0) + (0,0,1,1) = (l,l,l,l) = (1,0,1,0) + (0,1,0, l) = (l, 0,0,1) + (0,1, 1,0). All six of these assignments are summable. Therefore [l, 1,1,1; 2] segregates all asummable assignments (and in that sense is the optimal pseudo-separating structure for x x ^ x x, ). Hence x x v x x s/ x x, v x x \/ x x. v x x, is the least upper bound 1 X 2 X 3 V X 1 X 2 X 4 v X 1 X 3% " X 2 X 3^ and x^x^x^ s/ x-.x^x, ^ x x„x, v x.x x. is the greatest lower bound Definition 2.5-3 ' let g(x) and h(x) be two separable functions on exactly n variables. Assume g(x) > h(x) . g(x) and h(x) are called linearly distinguishable if there exists some hyperplane [W; TJ such that g(x) is the upper bound of [W; T] and h(x) is its lower bound. In such a case [W> T] is called the distinguishing hyperplane of g(x) and h(x) . Theorem 2.5.1: If f(x) is pseudo-separated by [W; T], then f(x) can be written in the form f (x) = h(x) v f (x)g(x) where h(x) is the lower bound of [W; T] and g(x) is the upper bound. Conversely, if two separable functions g(x) and'h(x) are linearly distinguishable, then for any function f (x), h(x) v/ f (x)g(x) will be pseudo-separated by the distinguishing hyperplane for g(x) and n (x), [W; T]. U5 Proof : Let f(x) be pseudo-separable under [Wj T]„ Let h(x) "be the lower bound of [W; T~J and g(x) be the upper bound. Define f (x) as follows: for any assignment A, if h(A) / g(A), then f (A) = f(A) and if h(A) = g(A), then f (A) - 1. Consider the function h(x) v f (x)g(x). For any assignment A, if f (A) = 1, then W°A > T. If W.A > T, then W°A > T + € because < € < M(W, T) and therefore h(A) = 1. Hence h'(A) s/ f 1 (A)g(A) - 1. Tf W°A = T, then because W«A > T - €, g(A) - 1. Since W-A < T - e, h(A) = Therefore, h( A) f g(A) and f X (A) = 1 = f(A). Hence h(A) v f 1 (A)g(A) = 1. Similarly, if f(A) = 0, then h(A) v/ f 1 (A)g(X)=0 Therefore f(x) = h(x) v f (x)g(x), Now assume that h(x) and g(x) are linearly distinguishable and let f (x) be any Boolean function. We Claim that h(x) v^ f (x)g(x) is pseudo-separated by LW$ T], the distinguishing hyperplane for h(x) and g(x). Let f(x) = h(x) v» ^(xJgCx). If A is any true assignment to f(x), then either h(A) = 1 or f 1 (A) = g(A) = 1. If h(A) = 1, because h(x) is the lower bound of [W; T], W.A > T f€ and A will be segregated by [Wj T] ( If f 1 ® = g(A) = 1, then W-A > T - € because g(x) is the upper bound of [Wj T°J and W°A = T because < € < M(w, T) . In any case, W=A > T. A similar proof is used to show that if f(B) = Q, then W-B'< T. Therefore, [W$ TJ pseudo- separates h(x) v f (x)g(x) for any Boolean function f (x) if [W; T] is the distinguishing hyperplane for h(x) and g(x)„ Q.E.D. Corollary 2, 5« 1 : If two separable functions, h^x) and g(x) are linearly distinguishable, then for any f (x), h(x) v f (x)g(x) is pseudo-separated U6 by the distinguishing hyperplane [Wj T] of h(x) and g(x), and for any assignment A, A is segregated "by LW) TJ if and only if h(,A) and g(A) are equal. Proof : If h(x) and g(x) are linearly distinguishable, then for any f (x), [W; T] pseudo-separates f(x) = h(x) ^ f (x)g(x) by Theorem 2.5.1. Assume that A is segregated by [W; T]. If f(A) = 1, then W-A > T and because < e < M(W, I), W°A >T+e>T-e. Therefore, h(A) = g(A) = 1. If f (1) = 0, then W«A < T and again because < e < M(W, T) W-A < T - e < T + €; arid so g(A) = h(A) = 0. Hence, if A is segregated by [W; T], then g(A) = h(A) . Now assume g(A) = h(A). If g(A) = h(A) = 0, then W'A < T - € because g(A) = 0. Therefore, A is segregated by [Wj T]. Similarly, if g(A) = h(A) = 1, then W-A > T + € and so A is again segregated by [W; T]. Q.E.D. Definition 2^.k : Let f(x) be pseudo-separable under [W, j T ]. Assume that h (x) and g (x) are the lower and upper bounds for this hyperplane. If there exists another hyperplane [W ; T ] which pseudo-separates f(x) with lower bound h p (x) and upper bound g p (x) such that h (x) > h (x) and g p (x) < g,(x), [VI ) T p ] will be called a refinement of [W. \ T ]. Example 2.5.2 : Consider the function x x v x x^ ^ x i x h v x o x v ^ e saw in Example 2.2.1 that this function is pseudo-separated by [1,1,1,1; 2]. In Example 2.3-1, we showed that [3,2,2,1; k] separates this function, and therefore also pseudo-separates it. The upper bound hi for [1,1.1,1; 2], we saw in Example 2.5.1, is x x v x x <• x x, v x x v v xx, v xx, = g (x). The upper bound (and least," . upper bound for f(x)) for [3,2,2,1; VJ is x x x 2 v x^ s/ x^ v x x = g^x) < g^x) . Further, the lower bound for [1,1,1,1; 2] is X..X-X v x x x, -v x x x, v xxx, = h (x) and the lower bound of [3,2,2,1$ k] is h 2 (x) = x^ v x.x v x g x x^„ Because g^x) > gg(x) and h^x) < h^x) , [3,2,2,1$ U] is a refinement of [1,1,1,1; 2] for x x v x x v x x, v x x Lemma 2.5»1 « Let f (x) be an arbitrary Boolean function on exactly n variables. Then f(x) is separable under [W; T] if and only if either the lower bound h(x) = f(x) or the upper bound g(x) - f(x)„ Proof : If f(x) is separable under [W$ T], then we have one of the following 3 pairs of conditions: (1) if f (A) = 1, then W.A > T and if f (A) = 0, then W«A < T. (2) if f (A) = 1, then W-A > T and if f (A) = 0, then W-A < T. (3) if f (A) = 1, then H > T and if f (A) = 0, then W-A < T. In case (l), there is no assignment which lies on [W; T], Therefore h(x) = g(x), and [Wj Tj must also separate h(x) and g(x). Since g(x), h(x) and f(x) are on exactly n-variables fi(x) = g(x) = f(x)„ In case (2), we claim that f(x) = g(x) . Clearly if f(A) = 1, then U8 W-A > T and g(A) = 1 because g(x) is separated by [Wj T - e]. If f (A) = 0, then W-A < T - e because < € < M(W, T) and therefore g(A) = 0. Therefore f(x) = g(x) . In case (3), if :•.:" f(A) = 0, then W-A < T and h(A) = 0, while if f(A) = 1, W-'A>T':+ e and h(A) = 1. Therefore f(x) =h(x).;If f(x) = h(x) or g(x), then [Wj T + e] or [Wj T - e] must separate f(x), and because < £ < M(W, T), [Wj T] also separates f(x). Q.E.D. In Example 2.3-1, we showed that x x v x x ^ x x v x x, was separable under [3,2,2,1; k] . We also noted that true assignments (l, 0,0,1) and (0, 1,1,0) resided on this hyperplane. Therefore this is a separation of f (x) in the sense of case (2) and g(x) must be equal to x l x 2 v X 1 X 3 v x l x i+ v X 2 X 3' We would now like to show that the upper and lower bounds of any refinement are both simultaneously realizable with the upper and lower bounds of the original hyperplane. Theorem 2.5.2 : If [W ; T ] is a refinement of [W j T ] for f(x), h (x) , h Q (x), g. (x) and g p (x) are all simultaneously realizable. Proof : We will demonstrate that we can always determine p, q, T , T , T and T. such that [pW +■ qW ; T ] separates h (x) , [pW + qW„; T ] separates h (x) , [pW + qW ; T_] separates g (x) and [pW + qW 2 ; T^] separates g (x) . This means that we must find p, q, T , T , T , and T. such that: U9 (1) if h (x) = 1, then (pW + qW 2 ) -x > T^ if h (x) = 0, then (pW + qW )-x < T?" (2) if h 2 (x) =1, then (p^ + qW 2 )-x > 1^ if h 2 (x) = 0, then (p^ +qW 2 )-x < T^ (3) if gg(x) - 1, then (p^ + qW 2 )-x > T^ if g 2 (x) = 0, then (pW^ +■ qW 2 )»x < T^ (h) if §1 (x) = 1, then (pW x + qW 2 )«x > tJ if g x (x) - 0, then (pW x f qW 2 )-x < tJ We will assume that h (x) < h (x) < g_(x) < g (x) . This implies that T. < T < T < T and the above problem can be reduced to finding p, q, T , T , T and T, such that: (a) if h (A) = 1, then (pW + qW* 2 ) • A > T?" (b) if h (A) = and h (A) = 1, then T^~ > (pW + qW ) "A > T 2 (c) if h (A) = and g g (2) = 1, then T 2 > (pW^ + qWg) ' A > T^ (d) if g 2 (A) = and g;L (A) = 1, then T^ > (p^ f qW 2 ) -A > tJ (e) if g x (A) = 0, then (pW^ f qW £ ) »A < tJ. We will use the following notation. Let Q.(s,t,u,v) denote the j maximum of W.-A where h (A) = s, h (A) = t, g p (,A) = u g (A) = v and r.(s,t,u,v) the minimum of W.-A where h (A) = s, h (A) = t, g p (A) = u, g (A) = v. Using this notation, the five conditions above become: (1) p0 1 (0,0,0,0) + q0 2 (0, 0,0,0) < tJ (2) P 1 (0,0,0,l) + q0 2 (0,0,0,l) < T^ (3) p0 1 (0,0,l,l)+ q0 2 (0,0,l,l) < T 2 50 (k) p0 1 (0,l,l,l) + q©-g(0,l,l,l) < T^ (5) pT^O, 0,0,1) + qp 2 (0,0,0,l) > tJ (6) pI^O, 0,1,1) + q? 2 (0, 0,1,1) > T^ (7) pr i (0,l,l,l) + qF 2 (0, 1,1,1) > T 2 (8) p^d,!,!,!) + 0F 2 (1,1,1,1) > T^ 1111 Now eliminate T , T , T and T . (i)* P [r x (o, 0,0,1) - q 1 (o,o,o,o)] + q[r 2 (o,o,o,i) - o 2 (o,o,o,o)] >0 (2)' p[r l (o, 0,1,1) - o 1 (o, o,o,i)] + q [r 2 (o, 0,1,1) - o 2 (o, 0,0,1)] >0 (3)' P [r x (o, 1,1,1) - q i (o, 0,1,1)] + q[r 2 (o,i,i,i) - e 2 (o,o,i,i)] > o (1+)' pC^d, 1,1,1) - Q 1 (0, 1,1,1)] + q[r 2 (l, 1,1,1) - Q 2 (0, 1,1,1)] > 0. From the knowledge of h (x), h (x), g„(x) and g (x) we now calculate the Q.'s and r.'s. If g. (A) =1 and h (a) = 0, then W -A = T. J 3 1 I'll because T - € < W -A < T + e , and < e < M(W, T'). Similarly, if g (A) = 1 and h (A) = 0, then W 'A = T . Hence f 1^(0,0,0,1) = 0^0,0,0,1) = ^(0,0,1,1) = Q 1 (0,0,1,1) = 1^(0,1,1,1) = 1 (0, 1,1,1) = T x and 1^(0,0,1,1) = 9^0,0,1,1) = Tg- Because g (x) is separated "by [W, j T - £ ] where < e < M(W , T ) the maximum W -A such that g,{A) = is less than T - €. . Similarly, the maximum W -A such that g p (Aj = is less than T - € where < € < M(W , T ). Because h (x) is separated by [W x ; T 1 + 6 1 ] and h 2 (x) by [W^ T g + 6 g ] if h^A) = 1, then W -A > T + € and if h (A) = 1, W -A > T + £ . Therefore: 1 (0, 0/0,0) < T x - € 1 and 1^(1,1,1,1) > ^ +- g 9 2 (0, 0,0,1) < T 2 - e 2 and r g (0,l,l,l) > T g + e 2 9 2 (0, 0,0,0) < T 2 - e 2 and r 2 (l,l,l,l) > T g + e 2 Q (0,1,1,1) * T + e o sad rjD,.0,Q,l) < T - e o . Then (a) P [r 1 (o, 0,0,1) - 9^0,0,0,0)] > p^ -i't^ r i >},^ ;■ x (b) P [r 1 (o, 0,1,1) - 9(0,0,0,1)] = p [t 1 - t 1 ] = 0. (c) p[r i (0,l,l,l) ^ 0^0,0,1,1)3 = p[T 1 - T x ] = and (d) P [r x (i, 1,1,1) - 9 L (o,],i,i)] > p [t x f e 1 - T x ] = pe r Similarly, (e) q[r 2 (0, 0,1,1) _ 9 2 (0, 0,0,1)] > $[T g - (T g - e g )!| = qc^ and (f) q[l 2 (0, 1,1,1) - 9^0,0,1,1)] > q[T 2 + € ? - T g ] = qe g . With this information, for any positive p and q conditions (2)* and (3)' are automatically satisfied. Thus, only inequalities (l)' and (h) ' need be considered. (l)' becomes (i)" P € 1 + , q [r 2 (o, 0,0,1) - 9 2 (0, 0,0,0)] > and ('4) ' becomes (h) n p€ 1 + q[r 2 (l, 1,1,1) - 9 2 (0, 1,1,1)] > 0. Since e > 0, both of these conditions can be satisfied by fixing q > and by increasing p. Let q = 1. Then choose p large enough so that 52 pe i > l^ 1 ' 1 ' 1 ' 1 ) - e 2 (o,i,i,i)l ^d pe 1 > |r 2 (o,o,o,i) - 9 2 (o,o,o,o)|. Since such a positive integer can always be found, the weights pW + W can he used to realize all four functions, h (x), h p (x), g p (x) and g (x) . Q7E.D. Example 2.5-3 : Consider the function given in Example 2.5.2, f(x) = x..x_ ss x n x s/ x^,x„ v x.x, o To calculate a common set of weights, '12132314 & ' consider r (0,0,0, l), 9 (0, 0,0,0), r (l, 1,1,1) and 9 (0, 1,1,1). Since [W.^ T 1 ] = [1,1,1,1; 2], g]L (x) = x x x 2 v x.^ v X] x^ n, x^ ^ y^a^x^ , h 1 (x) = x^x v X 1 X 2 X 1 + v X 1 X 3 X 4 v X 2 X 3 X l4° Because [Wgj. T g ] = [3,2,2,1; 4], g 2 (x) = x^ v x x x v x^ v x 2 x and h 2 (x) = xx v xx \s x x x, o The only assignments A such that g (A) = 1 and g (A) = are (0,1,0,1) and (0, 0,1,1). Therefore r (0,0,0,1) = (3,2,2,1). (0,1,0,1) = (3,2,2,1)^(0,0,1,1) = 3. If g 1 (A) = 0, then A has at most one component of value 1 and hence 9 (0,0,0,0) = 3 where the maximum occurs for A = (l, 0,0,0). Similarly r (l, 1,1,1) = 5 = 9 (0,1,1,1) . Therefore, for any positive p, pW + W will yield a common set of weights for these four functions. Let p = 1, then W x +■ W 2 = (4,3,3,2) and T* = 8, T 2 = 7, T^ = 6 and tJ = 5. Theorem 2. 5-3 - If two positive functions, h(x) and g(x) are linearly distinguishable, all assignments A such that g(A) = 1, h(A) = must be true extremal assignments for g(x) and false extremal assignments for h(x) . 53 Proof: If h(x) and g(x) are linearly distinguishable, for any k(x), f (x) = h(x) s/ k(x)g(x) is pseudo-separated by the distinguishing hyper- plane [W; T] by Theorem 2.5.1. Assume that g(A) = 1 and h(A) = 0, but ~* f~*\ ~*1 A is not extremal true for g(x). Then there is another assignment A .— ►» -»l -* /->. such that g(A) = 1 and A < A. Since k(,x) can be any function in : .. Theorem 2.5-1, let k(x) be such that k(A) = k(A ) = 1. Then A is not extremal true for f(x) and hence must be segregated by [W; T]. Therefore h(A) = g(A) = 1 by Corollary 2. 5.1> a contradiction. Therefore A must be extremal true for g(x). A similar proof shows that A must be extremal false for h(x) . Q.E.D. — > Definition 2. 5.5 - An extremal true assignment A is called extrinsic if -* -» — > there exists some other extremal true assignment B such that A and B are f(x) -comparable and A > B. An extremal false assignment C is called — » extrinsic if there exists another extremal false assignment D such that C and D are f(x) -comparable and C < D. An extremal assignment which is not an extrinsic assignment is called an intrinsic assignment. [7] We now wish to extend Theorem 2.5-3 to show that all assign- ments on which the values of g(x) and h(x) differ are intrinsic assign- ments. Theorem 2.5.k : If h(x) and g(x) are linearly distinguishable and g(A) = 1 while h(A) = 0, then A is intrinsic in both h(x) and g(x). 5** Proof : A must "be extremal in "both h(x) and g(x) by Theorem 2.5-3. If A is not intrinsic in g\x), then there is another assignment B g such that g(B) = 1 and A > Band W- A > W-B by Theorem 2.2.5. Since g(A) fi h(A), W-A = W-B by Corollary 2.5.1. Assume that A and B agree in positions x n , x^., . . . , x and disagree in x . . , x _, . . . , x with 1 2' ' p pfl' p+2' n a = a = ... = a - l and a = a o = . . . = a = 0(p < q < n). p+1 p+2 q q+1 q+2 n K * — * - ' Because W«A = W-B, w , + w _+...+■ w = w , + w _+...+ w . ' p+1 p+2 q q+1 q+2 n — » — » Therefore for any pair of assignments C and D which agree in :: — * -> x. , x_, .... x and which are such that C agrees with A on 1 2 p — * — >■ — * — > — > — > x ,, x „,..., x and D agrees with B on these variables, W«C = W°D. p+1' p+2' ' n e ' Hence it is impossible to have g(x: x ..->.£. n , x , • -* a .... x -* a ) ^ &v ' p+1 p+1' p+2 p+2' ' n n y > g(xt x ,, -» b -,,-x rt -* b _, ...,x -> b ) . Therefore A K B, a &v ' p+1 p+1' p+2 p+2' ' n n - ' contradiction. Similarly if B > A, then the contradiction A > B would arise. Therefore A must be intrinsic in both h(x) and g(x). Q.E.D. Theorem 2.5.5 - h(x) and g(x) are linearly distinguishable if and only if h(x) v k(x)g(x) is pseudo-separable for any k(x) and for some k(x) exactly those assignments A such that h(A) = g(A) are segregated by the distinguishing hyperplane [W; T]. Proof : If h(x) and g(x) are linearly distinguishable, the result follows from Theorem 2.5-1 and Corollary 2.5.1. If f(x) = h(x) ^ k(x)g(x) is pseudo-separable under [W; T] and if all assignments A such that h(A) = g(A) are segregated, then h(x) must have structure [Wj T + e] and g(x), [W; T - e] where < € < M(W, T) by Corollary 2.5-1 and Theorem 2.5.1. Therefore h(x) and g(x) must be linearly distinguishable. ' ."Q/E.D. 55 Corollary 2.5.2 : If h(x) and g(x) are linearly distinguishable then no assignment A. such that h(A. ) = 1 is summable in a sum including assignments B. such that g(B.) = 0. Proof : We know that g(x) > h(x) . Therefore if h(Aj_) = 1, then g(A. ) = 1 and if g(B. ) = 0, then h(B. ) = 0. Let k(x) he any Boolean function. If g(x) and h(x) are linearly distinguishable then f (x) = h(x) v k(x)g(x) is pseudo-separated by [Wj T], the distinguishing hyperplane for h(x) and g(x) by Theorem 2.5.1- Also, all assignments C such that g(C) = h(C) are segregated by [Wj T] (by Corollary 2.5-1) • Therefore all assign- — > _— » — » — > — > ments A. must be segregated by |_W; TJ and W-A. > T. Therefore each A. must be asummable by Lemma 2.U.I. Q.E.D. Lemma 2.5-2 : Let f (x), f (x), ..., f v (x) be simultaneously realizable functions. Then there exists a set of weights W and thresholds T. such that each f.(x) has structure [Wj T.] (i = l,2,...,k) and W.A. is distinct for each assignment A.. [2] — > Proof : Assume that we already have a set of hyperplanes of the same W but with different thresholds, T., under which each f . (x) is separable, but W-A = W-B for some A and B. Let A = (a , a., ..., a. ) and 1 2 n B = (b., , b^, .... b ). Since A 4- B there is some coordinate x., where * 1' 2 n ' j a . = 1 and b = or b . = 1 and a . = 0. Let M be the minimum difference 3 3 3 3 -» -> 1 between the distinct W-A. 's and the T. 's. Replace w. by w . = w. 4- e i i 3 3 3 56 where < € < M. Call the new structures [w~; T.], where w = w (p ^ j), w. = w. + e and T. = T . . Then W-A will be different from W-B. Since — > — » < € < M, any W°A„ ' s which were already distinct are still different under [W> T.]. Similarly the W°A 's which were previously less than T.'s will remain less than T.'s. Cbntinuing this replacement in this way -*L1 a set of weights W will be eventually constructed so that each W 1 " -A. is distinct and each f . (x) is separable under [w s T. ] (i'= ; l,2,...,k). Q.E.D. Assume that f(x) is pseudo-separable under [W .°, T ] with upper bound g, (x) and lower bound h (x) „ There exists some set of weights, W such that WT°x is distinct for all x and such that [WTj T ] separates h (x) and [W j T. ] separates g (x) by Lemma 2„5o2 If 1 l g 1 W i T p ] is a refinement of [VL $ T.], then its bounds g p (x) and h p (x) are simultaneously realizable with h_ (x) and g (x) . Hence by applying -*11 ->11 -* the same lemma, a weight vector W can be found for which all w *x are distinct and under which h (x), h (x^gJx), and g_ (x) can be represented respectively as [WTj T ], [W_; T ], [W p $ T ] and c - "-i *- o "P [WTj T ] where T > T, > T > T Any assignment which is segregated *- S-> "■•] ftp So ^>i "by CW ) T p ], but not by [W > T.. ] will then be in the range 1 -tL -* 1 1 -tL -> 1 T < W p *x < T or T > W p -x > T . Therefore any assignment in the g l 2 g 2 1 2 ""*1 ~* 1 /~*\ range T < W°x < T must be false for f (x) and any assignment in the g l d g 2 1 -+1 -* 1 ,-+\ range T, > W *x > T, must be true for f (x) „ h ! 2 h 2 57 Theorem 2.5.6 : If f (x) is pseudo-separated "by [Wj T] with upper bound g(x) and lower bound h(x) and if a refinement of [W; T] exists, -»1 ,--*l n then there must exist a set of weights W such that lW ; T J separates /~*\ •- _> 1 -i /~*\ ~* ~~*1 ~* h(x), LW °, T J separates g(x) and the assignment A such that W »A is g -* — » — * -*] -± ->2. ~* the maximal W°x < T. and B such that w -B is the minimal W -x > T h g must be true and false assignments for f(x) respectively. — > — > —> Therefore, if some set of false assignments B, , B , . . . } B — > — » -» and true assignments A , k , . . . , A is segregatable, then there must exist a set of weights such that W*A. > W«C. > W«B„ where the C„ are 111 1 all assignments which are summable and hence not segregatable. 58 CHAPTER 3 GENERAL PROPERTIES OF MARGINALLY SEPARABLE FUNCTIONS 3*1 Introduction In this chapter we will discuss a type of separation where a mixture of true and false assignments are allowed to lie between two parallel hyperplanes. This type of separation will be called marginal separation. This is more general than pseudo-separation. We saw in Corollary 2.4.1 that in order for a function f(x) to be pseudo- separable under a zero free structure, all of its non-extremal assignments must be asummable. An arbitrary positive function need not have this property. We saw in Example 2.4,1 that the function f (x) = x x sy x-XiX. has summable false assignments (1,0,1,0,1) and (0,1,0,0,1) and because this last assignment was not extremal, x,x_ v x x.x is not pseudo-separable, While all non-extremal assignments are not asummable, this function does possess some asummable assignments. We now would like to extend the theory that we have developed to include any positive function which has asummable assign- ments. This desire leads to the concept of marginal separation. The class of marginally separable functions is also a generalization of asummability as follows: If A. are true assignments for positive f(x) and B. are false assignments, and if k k k k T, t. < E!., then W*' ( Z A*. ) < W • ( £ f . ) i=l i=l i=l i=l if W is to be zero free. Thus some vf-B. must be greater than some 59 W'A. where B. is false and A. is true. The case is impossible if f(x) is a pseudo- separable function. Therefore, if we wish to extend our theory to all positive functions, we must allow w°A. < w«B. for some true assignments A. and false assignments B. . In marginal separation, we define some constant /\, called the margin, and require that all such assignments be in the: 'range T - a/2 < ^°~&- < ^'^-. < T + V 2 ° In this way we can localize all such assignments between two parallel hyperplanes, [w; T + a/2] and [W; T - a/2] and can segregate all other assignments . In the first section of this chapter we will put on a priori bound on this margin and will examine the properties of functions which can be realized under this restriction. In the second section we will develop necessary and sufficient conditions for this type of separation. In the last section we will not demand any prior bound on A, but will develop bounds which will allow the segregation of all asummable assignments. As before we will be dealing with Boolean functions on exactly n-variables. We also will assume that all w. =}= so as to eliminate the problems encountered when w. =0. 3»2 Basic Properties of Marginally Separated Functions In this section the basic idea of marginal separation is reintroduced. The concept of a quasi -separable Boolean function is defined and some of the basic algebraic properties of qua si -separable functions are discussed with particular emphasis on the relationship between the function and the weights of the quasi -separating hyperplane. 6o Definition 5«2.1: Let f(x) be an arbitrary Boolean function. Let [wj T] be an arbitrary hyperplane and A> a constant. We -will say that [w; T,A] marginally separates f(x) if for any assignment A, if t(f) = 1, then W*»A*> T - A/2 and if t(t) = 0, then W*-A* < T +A/2. The constant A will be called the margin of [w; T,a] and [W*; T,A] will be, called a marginal structure. Lemma 3»2.1: Let f(x) be an arbitrary Boolean function. Let [W*; T] be an arbitrary hyperplane. Then there exists some A > such that [w; T,a] marginally separates f (x) . n Proof: Let E |w. | = Z. i=l (1) If T > 0, let A = 2(T + Z) . Then T + a/2 = 2T + Z and T - a/2 = Z. For any assignment k, - Z < W*'A < Z. Therefore for any assignment A,T - A/2 < "•'&* < T + A/ 2 and [»j t >a] marginally separates (2) If T < 0, let A = 2(Z - T) . Then by the above reasoning we have that [w; T,A] marginally separates f (x) . 6i Lemma 3«2.2: Let A > be any constant. Then there exists some hyper- plane [w; T] such that [w; T,a] marginally separates f(x) . n Proof; Let w be any weight vector such that E |w.|< a/2. Then, because i=l X each assignment A must lie in the range - a/2 < w A < a/2, if we let T = 0, -we will have that [w; 0, a] marginally separates f (x) . Q.E.D. The two preceeding lemmas show that A and [w; T] must be related in some way for the concept of marginal separation to be meaningful. In this first section, we will use the relation a < I w . I where w . is 7 - ' mm ' mm the minimum weight of w. This means that if A is any true assignment and B any false assignment, then in order to have marginal separation of this type tf.I? - \?.t < K» . I. ° * ' min 1 Definition 3»2.2; A Boolean function f (x) is called qua si -separable under [w; T,a] if [w; T,a] marginally separates f(x) and if A < | w . \ . The pair of parallel hyperplanes [W*; T - a/2] and [% T + a/2] will be called quasi -separating hyperplanes if A < j w . |. As special cases, if a = 0, we have pseudo- separability and if A < 0, we have separability. We will now examine some algebraic properties of qua si -separable functions. Lemma $.2.3> Let f(x) be qua si -separable under [w; T,a] • Let x. be J any variable and A. any partial assignment in the other n-1 variables. J 62 If f(A*.; x. -> l) =1 and f(A*.; x^ -> 0) =0, then v. > and if •1 3 - J o, J f (A*.; x -> 0) =1 and f(A.) xj * l) #. 0; (then w. < 0. J 3 J J -, J Proof: As in Theorem 2.2.1, we can write f(x) = x.f(x; x. -> l) v 1 v/ x.f(x; x. -> 0) . J J J J n (1) If f (A*.; x. -» 1) = 1, then Z v. a. + v > T - a/2. 3 3 ■*■ -*• J n (2) If f(A*.j x. -*0) = 0, then I v. a. < T + a/2. J J l l - *-* Combining (l) and (2), w . > - A. J Now assume w. < 0. Then we have© > w. > - A. Therefore, 3 ~ ~ 3 - A > J w. | a contradiction because it was assumed that [w; T,A] quasi- J separates f (x) . A similar proof holds if f(A,; x. -* 0) =1 and f(A*.; x. -> 1) = 0. J J Q.E.D, Theorem ^.2.1; If f(x) is quasi -separable, then f(x) must be unate. Proof: If f(x) is not unate in x., then neither - — : — J f (x; x -> 1) > f (x; x -> 0) J J nor f (x; x -> 0) > f (x; x -» l) J J 63 Thus there must exist a pair of assignments A. and B. on the other J J n-1 variables such that (l) f(A.; x. -> 1) = 1 and f(A.; x. -» 0) = J J J J (2) f(H\; x. ->0) =1 and f(B > .; x. -> l) = 0* (l) implies w. > by Lemma 3»2.3 and (2) implies v. < 0. J J Therefore we have a contradiction. Hence f(xj if it is not unate, can not be quasi -separable. Q.E.D, Corollary 3«2.1: Let f(x) be positive (or negative) in some variable x. and quasi -separable under [w; T,A]. Then the corresponding weight J w. must be positive (or negative). J Proof: If f(x) is positive in x. then f(x; x. -> l) > f (3c; x. -» 0) . J J J Therefore there must exist some assignment A. on the remaining n-1 j variables such that f(A.; x. -» l) =1 and f(A.: x. -»0) =0. J J JO Therefore w . > by Lemma 3.2.3. J Q.E.D. Theorem 3*2.2: If f(xj is marginally separable under [w; T,a], then f(x; x. ->x.) is marginally separable under [w x , w , . .., ^ i _ 1 ^ -y ± , v ±+1 > •"> v n 3 T - v 1 ,a] = [^; T - v ± ,A] Proof:' ' If . [W; I:/ A] marginally separates f'(x),' then- for any x 6k (1) If f(x) = 1, then tf-x > T - a/ 2 » (2) If f (x) * 0, then ^«x < T + A/2.' As in the proof of Theorem 2.2,3, substitute x. for x. in f(x) and ( 1-x . ) for x . in vf °x. Then (l) If f(x) x. ->x.) = 1, E w.x.-:+ w.(l - x.) > T - a/2< . _ n (2) If f (x; x. -» x.) = 0, Z w jXj + w. (l - x.) < T + a/2. Subtracting v. from both sides in the linear inequalities: (l) If f(x; x. -*x.) = 1, L w.x. - w.x. > T - a/2 - w, v/ v ' 1 i' ' j j 11— ^ : -» - n (2) If f(x; x. ->x.) = 0, £ w.x. - w.x. <■ T + a/2 - w . , v/ v ' 1 i' ' j j i i — ■ Hr 1 Therefore, [W'; T - w.,a] marginally separates f(x; x. -> x^) . Q.E.D, Lemma ^.2.k: If f(x) is marginally separable under [w; T,a], then f(x) is marginally separable under [-W; -T,a] • 65 Proof: If f(x) = 1, f(x) = and V?.x > T - ^2 or (-w) -x < - T + /±/2. If f(x) = 0, f(x) = 1 and W*-x < T + a/ 2 or (-tf)«x > - T -a/2. Hence [-w; - T,a] marginally separates f (x) . Q.E.D. Theorem 3>2»3: If f(x) is quasi-separable under [w; T,a], then f (x) n is quasi -separable under [W; Z w. - T,a] . i=l d - - Proof: Since f (x) = f(x), "we can combine the results of the Theorem 3.2.2 and Lemma J>,2,h. Q.E.D, Definition 3*2. 3- Let f(x) be an arbitrary Boolean function. Let [W; T,a] be any structure "which marginally separates f (x) . Let A be any assignment to f (x) . We "will say that A is segregated by the structure [% T,a] if f(A*) = 1 and W**A* >T + a/2 or f(t) = and W C A < T - a/2. If there exists some structure which marginally separates f (x) and segregates A, "we "will call A segregatable. The concept of segregated assignment defined here, is very similar to that used in pseudo-separability. In both instances if an assignment A is segregated, we can determine f (a) by only examining W*A and T. Notice that if an assignment is not segregated, then we have either W*A = T in the case of pseudo-separation or T - a/2 < W«A < T + a/2 for marginal separation and we need additional information to determine f (a) . 66 Lemma 3*2.5: If f(x) is quasi -separable under [w; T,a], then all non- extremal assignments must be segregated by [w; T,a]« Proof: f(x) is unate by Theorem 3»2ol. Therefore, we can assume all w. > 0. Let A be any non-extremal true assignment. There must be some other true assignment R such that A > Bo Since all w. > 0, W»A > W-B. Since A* > B*, vf-A* - ^»B*> w . . If A* is not segregated by [% T,A], T - A/2 < V?.A*< T + A/2. Also ^«B*> T - a/2 since if is true. Therefore, A> w . , a contradiction. Similarly, every non-extremal false assignment must be segregated by [Wj T,a]. Q.E.D. Theorem 3»2A: Let f(x) be a positive function. If f(x) is marginally separated by [w; T,a] and if all non-extremal assignments are segregated by this structure, then f(x) is quasi -separable. Proof: Assume that A> v. for some i. There must exist a true assignment A = (a., a p , ..., a ) and a false assignment, B = (b_, b ? , .«,., b ) such that ^»A* = T - a/2 and tf-lf = T + a/2 or else A could be reduced. If a. = or b. = 1, then the corresponding assignments A with a. = a. for J4=l and a. ' = 1 or B ' with b. ' = b (j/i) and b. ' = will not be . . segregated by [^ T,A] because A> w. and ^-A* = T - a/2. This would contradict the original assumption that all non-extremal assignments are segregated by fv?; T,A], Therefore a. = 1 and b„ =0. Consider the transformation w ' = w (p=/i), w. ' = w. + €. and T' = T + e./ 2 where p p v i 11 r €. > 0. 1 67 Because we want e. to be as large as possible and still main- tain marginal separation, we can increase e. until we find either a true assignment C such that "W* « A = ~W' if € or a false assignment D such that W' *D = W' *B. An important point here is that c. = and d. = 1» If c. = 1, then the transformation w.' = w. + £. has the same effect on l ' ill C as on A so that w*C = W°A and therefore 6. can be increased without i disturbing the marginal separation. Similarly if d. =0, then W' .D = W'D = W'B = w ' °B and e. can be increased. Therefore, assume i ' that we have found a true assignment C with c. = such that i w' «A = w ' *C . Then 6. can not be increased further or else [w'; T',A] will not marginally separate f (x) . Consider A and €. at the point that V? ! •(? •= \t' 't and c . = 0. If A < e. ,'■ then because ^' «A* = T - a/2 + e. i ■ — i' ^ l and W f •]? = !?•!? = T + &J2, ^' -t > vf' «b\ Therefore, for all true assignments A. and false assignments B . y W'-A. > w ' «B. . Hence i ° li—i [w'; T - a/2 + e. ] pseudo- separates (or separates) f (x) . So assume that A > e. . Then tf» .]? > V?* -A*. Let A' = tf l ^ - tf 1 "A*. If A* > w. ' , l — l we employ the same argument used above to get a contradiction. Since C has c. =0, the corresponding assignment C' where c : . ' =: c. for j/i, c. = 1, must be true and non-extremal. Since W ' *(f = T + £./2 - A'/ 2 and A' > w i ' , T + €./2 - a/2 < tf' "(ft < T + e . /2 + a/2 . Since ^••(f 1 = tf«lf + €., T - S./2 - A'/ 2 < $'(?' < T - €./2 + A'/ 2 « Since A' =A-£.,T-a/2< tf-cf' < T + a/2 - €. or T - a/2 < tf.(f* < T + a/2 . Therefore C 1 is not segregated by [W; T,a], a contradiction since it is non-extremal. Therefore A' < w. '. 68 Consider f(x) under this structure. Clearly, [w'| T',A'] marginally separates f(x) since e. was increased only until we were in danger of destroying this property o Also A' < u : «. . If w. : is the only weight such that A> w., we are done. If not, we repeat the procedure for each such weight until a marginal structure [W M j T",A"] is found which quasi-separates fix) . Q.E.D. Example 3»2.1; We would like to illustrate a very important fact about marginal separation, namely that an assignment need not be asummable to be segregated by even a quasi -separating structure. Consider f(x) = x,Xp v x,x^x . In Example 1.2o2, we showed that this function is marginally separable under [3,3>2 5 2, 2; 6 l/2,l] Since l/2 (1,0,0,0,1) = 5 < 6 = T - A/2. Therefore (1,0,0,0, l) is segregated by [3,3,2,2,2; 6 1/2,1]. 3«3 A Necessary and Sufficient Condition for Quasi-Separability In this section we will develop a necessary and sufficient criterion for quasi -separability. The condition which is presented here is analogous to the supersummability conditions for pseudo-separable functions which were introduced in Section 2.4-. 69 Definition 3-5«l* Let f(x) be an arbitrary function, f(x) is called extremely summable if there exists k true non-extremal assignments —>—>■—> — > — > •=* A,, A p , . .., A, and k false extremal assignments EL, B p , .«.., B, such k k that Z A*. = Z st.:. 1 1 i=l i=l Lemma 3»3«1« If f (x) is a positive function, then f (x) is extremely summable if and only if there exists k true extremal assignments A-,, Ap, . .., A, and k false non-extremal assignments B, , Bp, ..., B, k k such that E A. = Z B. . i i i=l i=l Proof: If k such true extremal assignments and k such false non-extremal assignments exist and f (x) is positive, then we can, by changing some components of the B. and A. from to 1, change each A. into A'.' and B. "into BJ , where each B.' is extremal false and each A', is non- i V i i k k extremal true. We can maintain Z A.' = Z B] if we "change the same i=l i=l component in one B. and one A. at a time. Therefore f(x) must be extremely summable. Conversely, if f(x) is extremely summable, then we can find k true non-extremal assignments A,', AX, .... A,' and k false extremal 1 2' ' k 70 k k assignments B 1 , B' , ..., B' such that £ A' = Z B' . If we repeat 1 c. K 1 1 1=1 i=l the process outlined in the first half of this proof, except changing components from 1 to in both A* 1 and B*' at once, we will find k — > — * — > true extremal assignments A., A ? , ..., A, and k false non-extremal k k ;ignments B 1 ,, BJ , . .., B' such that A A„ = Z B ' . 1 ^ K 1=1 1 -i=l X Example 3«3«1' Consider the function f (x) = x* x p vioc^x-x^ v XpX,_X/- (1,0,1,1,0,0) and (0,1,0,0,1,1) are extremal true assignments. (1,0,0,0,1,1) and (0,1,1,1,0,0) are false assignments. Since (1,0,1,1,0,0) + (0,1,0,0,1,1) = (1,0,0,0,1,1) + (0,1,1,1,0,0) = (1,1,1,1,1,1), these assignments are summahle. Since (1,0,1,1,0,0) < (1,0,1,1,0,1) which is false and (0,1,0,0,1,1) < (0,1,1,0,1,1) which is also false, both (1,0,1,1,0,0) and (0,1,0,0,1,1) are non-extremal false. Therefore this function is extremely summable. We can also relate the concept of extreme summability with the idea of supersummability . Because we want to employ extreme summability for quasi -separable functions in much the same way as we used super- summability for pseudo- separable functions, and because we wish to extend the range of functions realized, we expect extreme summability to be a weaker property than supersummability. 71 Theorem 3*3-1: If f(x) is positive and extremely summable, f (x) is supersummable. Proof: Let A... A., ....A, be non-extremal true assignments and — — ■ 12k to -» -* -» B , B , . .., B be extremal false assignments such that 12 k k k A, = L B.„ Replace each A. with a corresponding extremal true i=l i=l ■• k k k assignment U. « (A< > U 4 ) . Then Z U 4 < Z A. = Z B. and f(x) is supersummable. i=l x i=l X 1=1 Q.E.D. Example 3° 3° 2: Consider the function X..X v x x, x . As was shown in Example 2.1+.2, this function is supersummable with false assignments (0,1,0,1,1) and (1,0,1,1,0) summing to (1,1,2,1,1) and true assignments (1,1,0,0,0) and (0,0,1,1,1) summing to (1,1,1,1,1). The only extremal true assignments for this function are: (1,1,0,0,0) and (0,0,1,1,1) „ The only non-extremal false assignments are: (0,0,0,0,0), (0, 0,0,0,1), (0,0,0,1,0), (0,0,1,0,0), (0,1,0,0,0), (1,0,0,0,0), (0,0,0,1,1), (0,0,1,0,1), (0,0,0,1,1), All other assignments are either false extremal assignments or true assignments „ Suppose there exists a f p non-extremal false assignments B_ , B_, . ... B_ _, such that 12 Cft-p a+p Z B. = a(l,l,0,0,0) + p(0, 0,1,1,1). Let 1 = (1,1,. ..,1). Then 1=1 X 72 OS-0 "?•( Z 1^) = ti[cc(l, 1,0,0,0) + p(0,0, 1,1,1,)] = 2a + 3p„ Since each i=l _V ->-> B. has at most two coordinates of value 1, 1*B. < 2 l ' i — ' . Of? or l'( Z B.) < 2« + 20. Therefore 2a+ 20 > 2a + 30. This implies i=l 1 a . assignments such that Z B = Ol( 1,1, 0,0,0), which JLs impossible. . There- , i=1 fore x.x v x^x.x^ is Supersummable hut not extremely summable. Theorem 3«3»2s A necessary and sufficient condition for a positive function f (x) to be quasi -separable is that f (x) not be extremely summable. Proof: (l) If f(x) is positive : and quasi -separable, then there is some marginal structure [w: T,4i with A< w „ which marginally mm separates f (x) . If f (3c) is extremely summable, there exists k — > — > — * extremal true assignments A , A ? , <.<,<>, A. and k non-extremal false k -> k assignments B,, B c , . «.,, B, such that Z A, = Z B. . 12 k 1=1 X 1=1 x Let U. be an extremal false assignment which corresponds to B. . k k k Then \? > t . The tf° ( Z A*. ) = tf ° ( Z ? . ) < tf« ' ( Z % ) • Since 1=1 2 1=1 1=1 73 if. > t . , tf-T?. > W*.B* + w . . i 1' i — min k k k Therefore \f' ( Z if.' > W*« ( E B*. ) + k v . 4- (E t) +kw . . Because j , X — . , i, min . _ i mm i=l i=l ' i=l each A. is a true assignment, w«A, > T -A/2. Since f(x) is positive, k -* all v. > 0- Therefore W*' ( £ A.) > kT - kA/2. Similarly, issl x k W*«( Z if. ) < kT + kA/2. Therefore kw . < kA and f(x) is not quasi- ._ n i — ' min — separable . (2) We will now show that if f (x) is not extremely summable, then f(x) is quasi -separable. If f(x) is quasi-separable, "we know that all non-extremal assignments must be segregated by the marginal structure by Lemma 3«2.5« Therefore only extremal assignments can reside in the range T - a/2 < W*«A*. < T + A/2. Also, if f(x) is positive and quasi- separable, then all w. > by Corollary 3.2.1. Therefore if A > B, W«A > w.B*. Conversely if all extremal assignments satisfy these conditions, then the function is quasi -separable by Theorem 3«2.4. There- fore, the problem of determining sufficient conditions for a positive function to be qua si -separable is the same as finding a marginal structure [W*; T,a1 such that: vf'tf. > T -A/2 if \f. is extremal and true. (3.3^l) 7^ V?°V^ < T + A/2 if V* k is extremal and false and (3..3-1) A < w . . man Combining these inequalities w o (u. - v. ) > - w . or ^•(lf. - ^. ) + w . > 0. x j k' mm A necessary and sufficient condition that such a set of inequalities not "be consistent and therefore not have a solution, is the existence of a., > such that Z a., > and Ea., $'(\f. - f. ) + w . ] < a [3] .. ,Jk J k min - J,k Let 2 be the assignment with zeros in all positions except the coordinate which corresponds to w . <> If more than one such position exists (i.e., min ■y . v . . . v all correspond to v . ) write 2 = — (fT.n + a © +...+ Z_ ) i ■ • ' ■? > x i mm p x X d p ' X l X 2 p Then we can rewrite the above expression as Yj a,, V?°(tf. - V*,) + . X a.,w»S<0. Since we have assumed all . , j K. j K J /K. J K w.> 0, this expression must be identically zero or identically less than 0. Therefore 75 Za (I? - f) + Eal<0 (3. 3.2) jk Jk J k j,k jk or ^k and true. Therefore pU. + pZ = E A. and (3.3.2) becomes J e=l je V a ( I A J < v ^ # J,k Jk e=l J j,k Jkk OOO; 76 We can always make the relation in (3-3»3) an equality by — -. changing components in A. from to 1 and still he guaranteed that the resulting assignment is non-extremal true. This means that L A. =■-■ T, v. for some q and f(x) is extremely summable. Therefore, i=l x i*l. x if f (x) is not quasi -separable, then (3«3»l) are inconsistent and hence (3«3«3) which implies f(x) is extremely summable. Q ;E .D . Example 3»3«3 : In Example 3«3»1> f(x) = x^-v x.. x x. w x ? x x •- was shown to be extremely summable. Therefore this function is not quasi -separable. In Example 3.3.2.x n x^v/ x X|,x c was shown to be quasi-separable, but; not i. 2 5 *r J V. .?. T..1 T ..'. ■. >G .: i q '.;.'..-,' j_~ .' ' ..3Xi } . i> :.\C ,, pseudo- separable. Therefore there exists positive functions which are quasi-separable, but not pseudo -separable. J.h Marginal Separation and the Degree of Super summability This section is devoted to an analysis of marginal separation without any a priori bound on the margin \ However, later we will obtain lower bounds on A in terms of the weight vector W for specific functions, as we did in the previous sections. The degree of super- summability tells us how -^ and W are related. In particular it gives a lower bound on A as a multiple of the minimum weight (i.e., 77 A >T w . where F is the degree of supersummability) . Before we — mm examine this property, we prove some general results about marginal separation. Theorem 3»^»1' Let f(x) be a positive function. Then there exists a marginal structure :[w; T, A] which marginally separates f (x) and which is such that for any assignment A, A is segregated by [W; T, A] . if and cl only if A is asummable in f (x) . Proof: Because we require that all assignments segregated by [W; T, a] must be asummable and all asummable assignments must be segregated by [w; T,a], the problem of finding such a structure can be reduced to determining [w; T,a] such that (1) tf-A* t > T + A/2 if A*, is asummable and f(A*. ) = 1. (2) \t-t. < T - A/2 if If. is asummable and f(lf.) = (3) T - A/2 < tf'C*. < T + A/2 if (f is summable. Combining these inequalities, (1)' If- (2^ - 1) > A (2)' tf.(t ± - C^) >0 (3)' ^.(c^ -?,) >o 78 This system is inconsistent if and only if there exists non-negative numbers a.,, h , and c. such that Z a A «* and ij IK jK » i» lj Z • w-C^ - S ) - Z b lk ».ff 1 -? k ) ♦ Z c jk (c k - B k ) 0, some a. . must be positive. As in the proof of Theorem 2»k k, we again denote the C„ which -* •>♦ -♦ are true as D. and the C. which are false as E. and write 11 1 Zc. = Z D. + L E J . Similarly, we can write Z C. = Z D ! + Z E. . . i . - i . . i "' . . i . . i . , i i=l i=l i=l i=l i=l . i=l where D. are the true C. and E. are the false. Substituting these equalities into (3A2) 79 L A. + Z D. + L E. < 2 B. + Z D. + Z I 13 4.3) . , i . , l . , i - . , i . - i . -, i i=l i=l i=l i=l i=l i=l To arrive at an expression with only true assignments on the left side and false assignments on the right, we use the fact that each E, and D. is summable. Thus E. = Z V 1 - L Z 1 and D. = Z Z 1 - Z V 1 , i l . , m m l ., m _ m ' m=l m=2 m=l m=2 -*d ~*i ' -h -*i ' where the Z and Z are false summable assignments and the V and V mm mm are true summable assignments. Substituting the corresponding equality into (3-^-3) for each E. and D. we get p q+d p+q-r r-t-d Z A. 1- Z T. < Z B. + Z F. . - 1 . , l — . , l . , i 1=1 1=1 1=1 1=1 -> -> -* — * where A. and T. are true assignments and B. and F. are false. Since 11 11 p f- and p + q - r ^ (because some a. . > 0),this implies that some -*- J of the A. and B. must be summable or supersummable. If it is super- 11 summable, it is accordingly summable by Lemma 2.4.2. This, however, -* contradicts the original assumption that the C. were the only summable assignments. Therefore, the above system of inequalities always has a solution. Q.E.D. The marginal structure which is guaranteed to exist by this theorem will not in general be optimal in the sense that a maximal number of assignments need not be segregated. We can in general find another structure which not only segregates all asummable assignments, but some of the summable assignments as Example 3-2,1 shows. 80 Definition 3.^-1 : Let f (x) be an arbitrary Boolean function on exactly n variables. Assume that [Wj T, A] marginally separates f(x). Let M(Wj T, A) be the minimum of (T - A/2) - W»A. for segregated false assignments and W'A - (T ¥ A/2) for segregated true assignments. Let g(x) be the separable function on n variables such that if ' ^ W'A > T - A/2 - e, then g(A) = 1 and if W'B < T - A/2 - e, g(B) = where < € < M(W, T, A) , Let h(x) be the separable function on n variables such that if W«A > T f A/2 f c, then h(A) = 1 and if W-B > T + A/2 + e, then h(B) = where < € < M(W, T, A). g(x) is then called the marginal upper bound for [W; T, A] and h(x) is called the marginal lower bound . Notice that because < € < M(W, T, A), if g(A) = 1 and h(A) = 0, then T - A/2 < W*A < T f A/2, Lemma 3°*+«l: Let h(x) and g(x) be any pair of simultaneously realizable functions. Assume h(x) < g(x) . Then there exists some [Wj T, A] such that h(x) is the marginal lower bound of [W; T, A] and g(x) is its marginal upper bound. Proof : Let h(x) have structure [W; T ] and g(x) [W; T ]. Since h(x) < g(x), T. > T . Let A = T,_ - T - 2c, where < e < minimum " h g h g (¥-A. - T u for h(A.) = 1 and T - W.A, for g(A.) = 0) . Let i h l h i i T = T t l/2 A + e. Then h(x) is the marginal lower bound for g [W; T, A] and g(x) is the marginal upper bound, Q.E.D. $1 Theorem 3.4.2 : If g(x) and h(x) are the marginal upper bound and lower "bound for [Wj T,A ], then for any function k(x), h(x) v g(x)k(x) is > — > marginally separated by [Wj T,A ]. Conversely, if f(x) is marginally separable under [Wj T,A ], then there exists some function k(x) such that f (x) = h(x) v g(x)k(x) where g(x) and h>(x) are the marginal bounds for [W; T, A] and any assignment A is segregated "by £W; T, a] if and only if g(A) and h(A) are equal. Proof : If f(x) = h(x) v k(x)g(x) and h(x) has structure [W; T + a/2 + e] while g(x) has structure [W; T - a/2 - e] where < e < M(W, T, a), then we claim [Wj T, A] marginally separates f (x) and segregates all assign- ments such that g(x) = h(x). Let A be any true assignment. Then either h(A) - 1 or g(A)k(A) = 1. If h(A) - 1, W-A > T + A/2 f e and A is segregated by [Wj T, A]. Also, because h(A) = 1 and g(x) > h(x), g(A) = 1. If g(A) = 1 (and h(A) - 0), then W-A > T - a/2 - e and since < e T - A/2. Since h(A) = 0, W-A < T + a/2. Therefore .. T -A/2 < W-A < T + A/2 and A is not segregated by [W; T, A]. A similar arguement shows that if f (B) = 0, then h(B) = must hold. If g(B) = 0, W-B < T -A/2. If h(B) = and g(S) = 1, then T - A/2 < W-B ^ T + A/2 and B is not segregated. Therefore f (x) is marginally separated by [Wj T, A] and all segregated assignments are such that g(A) = h(A). If f (x) is marginally separated by [WJ T, A], pick the marginal lower and upper bounds to be h(x) and g(x) respectively. Then f(x) must be of the form h(x) v k(x)g(x) where k(x) is such that f(A) = 1 and h(A) =0, k(A) = 1 and if f(S) = and g(B) = 1, k(A) = 0. Since O h(x)« Similarly if g(A) = 0, h(A) = 0. Because all assignments A such that g(A) = h(A) are segregated by [W) T, A] ("by- Theorem 3° 4. 2) and because all segregated assignments are asummable (by Theorem 3. 4,1), the result follows Q.E.D. Another important result about marginal separation is that there exists a structure [w; T,a] which segregates only one assignment, and for each distinct assignment a structure exists which segregates only that assignment. Theorem 3«^Q: Let f (x) be an arbitrary Boolean function and let A be any assignment. Then there exists some marginal structure [W; T,a] which segregates only A. 83 Proof: Assume that f(Aj = 1. Let A = (a,, a~, .... a ). Let n / ^ -. \ .. -| h(x) = IT x. i where x. = x. and x. = x. . Let g(x) be identically . -, i ii ii i=l equal to 1. Then f(x) = h(x) v f(x)g(x). Clearly h(x) and g(x) are simultaneously realizable, Therefore* by Theorem J>,k.2, we can find [W; T, A] which marginally separates f (x) and also segregates only assignments such that g(A) = h(A). Because g(x) = 1, and A is the only assignment such that h(A) = 1, it is the only assignment segregated by [Wj T, A]. Q.E.D. .-*. -> -> -> Let f(x) be a given function. Let A , A , . , . , A be any 1 £- K. -> -> -» set of k true assignments and B , B , . .., B any set of k false assign- k k ments. If these two sets are summable, i.e. if Z A. = E B., the • -i i • n i 1=1 i=l i:. :•.;.-.. ■:; i ■: W. . '.': k k function can not be separable. If Z A. = Z B. and some of the . , l . ., l i=l i=l -* -> A. or B, are non-extremal, then the function is supersummable and hence -> -> not pseudo-separable. If all of the A. or B. are non-extremal, then the function is extremely summable and therefore not quasi-separable. We would now like to develop a lower bound on A, the margin, in terms of W for specific functions. Thus, we introduce the concept of the degree of supersummability. Definition 3.^.3 ' Let f(x) be an arbitrary positive function. Let {A , A ..., A J be a set of k true assignments and {B , B , ..,, B ) k false assignments. These sets are called k-comparable if 81+ k k k k Z A. > Z B. or Z B. > Z A. . i=l 1_ i=l X i=l x ~i=l X Notice that if A. = (a. , a. ..... a. ) and 1 i_f i ' ' i 12 n B. = (b . , b. , . .., b, ) then two sets are k-comparable if J D l J 2 J n k k k k Z a. > Z b. for all j or Z a. < Z b. for all j. , i . - . , l . , , i . — . i i . ° i=l j i=l j i=l j i=l j Definition 3.^.^ s Let f (x) be an arbitrary positive function . Let k k 8, be the maximum possible It ( Z B. - Z A.) for all k-comparable _^ i=l i=l sets of assignments, where 1 = (l,l, ...,l). Let r be the maximum 8,/k for K. all k. F will be called the degree of supersummability for f(x). Lemma 3° h.2 : The degree of supersummability is attained by summing only extremal assignments. Proof : Assume F is the degree and that r is attained as ! =~ . • ' k k k r = 1: ( L B. - L A.). If some B. or A. is non-extremal, then the . , l . , l 11 i=l i=l -+ -> degree can be increased by substituting for A. or B. the corresponding extremal assignment. In such a case, T is not maximal, a contradiction. Q.E.D. Lemma 3.^.3 - Assume T > 0. Then the assignments which are summed to obtain this degree are summable. k k Proof: Let kT = Is ( Z B. - Z A.). Then since T > and . , i . , i i=l i=l k -+ k -* {a\ , a\, ..., A, } and {B n , B„, ... B. } are k-comparable, Z B. > Z A., 1' 2' k 1' 2' k . , i . , l i=l 1=1 85 k k Assume that Z bi ^ Z a. . If we change x components in the B.'s — > from 1 to and call the resulting assignments B'. , then J k _> k _ k k Z B'. rl Z A. and Z b ! = £ a . . Since f(x) is positive i=l 3 " i=l J i=l X j i=l ^ and B. > B., B. is a false assignment. We repeat this procedure for J J J k k k k ^ each component such that £ b! > £ a. . Then £ "BY = Z A, .,i.,i . , l . , l i=l p i=l p i=l i=l — > where BV are the resulting false assignments. Therefore, all true assignments are summable. Similarly all of the false assignments — »• B. must be summable. 3 Q.E.D. Example 3.^.1: Consider the function x n x_ s, x x,x_. We have already 1 c. 3^5 shown that this function is supersummable since (l,l,0,O,0)+(O,0,O,l,l) = (1,1,1,1,1) < (1,1,2,1,1) = (1,0,1,1,0) + (0,1,1,0,1). Therefore the degree is at least 1/2. We have also shown in Example 3.3-2, that this function is not extremely summable. Since all assignments used in achieving the degree must be extremal and since the only true extremal assignments are (l, 1,0,0,0) and (0,0,1,1,1), the degree must be achieved as p+q _ (p-: + q)r = l-[ Z B - p(l, 1,0,0,0) -q(0, 0,1,1, l) ] . i=l Again we note that each extremal false assignment has at most 3 ones p+q _ as components. Therefore, 1- ( Z B.) < 3p + 3q while i=l X 1- [p(l, 1,0,0,0) 4- q(0,0,l,l,l)] - 2p + 3q. Therefore (p + q)r : p or P. p+q* r <-— -. Therefore the degree of this function is 1/2 < r < 1. since q / 86 We will later show that this bound on V is sufficient to determine what type of separation is possible* Theorem 3-4.4 : If T < 0, then a positive function, f (x) is separable. Proof : If T < 0, then whenever {£. , A , . .., X) and (B , B , . .., B. ) 12 It k ^12 k ^k are comparable sets of true and false assignments, Z A. > E B. . i=l 1 i=l x Therefore, f (x) can not be summable for any k and this implies that f(x) is separable by Theorem 2.4.1. For incomparable sets of true and false assignments which are not considered in computing r, of course, equality can not hold. Q.E.D. Theorem 3«4.5 : If r =0, then a positive function f (x) is pseudo- separable. Proof : If r = 0, then whenever {A A , . . . , A.) and {B , ]L, . . . , B } k k are comparable sets of true and false assignments, Z A. > Z B. . i=l x ' i=l 1 Therefore, f (x) can not be supersummable and this implies that f (x) is pseudo-separable by Theorem 2.4.3. For incomparable sets of true k k and false assignments, the inequality Z A. < Z B. can not hold. 1=1 " i=1 ' Q.E.D. Theorem 3-4.5 : If r > and [W$ T, A] marginally separates a positive function f(x), A > Tw . . ' — mm 87 k k k k Proof: Let T = 1- ( E B. - £ A.). Then E B. > E A.. If i=l i=l 1=1 i=l k [W; T, A] marginally separates f(x), $• ( E B. ) < kT + kA/2 and i=l X k k h. k -> W- ( E A.) > kT - kA/2. Let E B. - Z A. = (c , c , . .., c ) i=l i=l i=l n n where all c. > and some c. > 0. Then clearly, E c.w. > E c.w J . =1 i i - i=1 i min k k Since W- ( E B. ) < kT + kA/2 and W- ( E A. ) > kT - kA/2, .. i=l x i=l x k k n n n W-( E B - E A. ) = Ewc .Ew.c >Ecw. . . i . t i . , l i - -*•=, i l -, n i mm i=l i=l i=l i=l i=l n c. n c .. E — rv . . Since T = E — -. i=l- kmn i=l k ' Q.E.D. or a > E -rrv . . Since T = E —r-, then A > T w . k min . , k — mm Notice that if some w. 4 w.. then E c.w. > E c.w . and l ' ,v . , i I . , i mm u i=l "■ i=l A > r w . , " min Theorem 3* ^-.6 : If T < 1, then a positive function, f (x) is quasi- separable. Proof : If f(x) is extremely summable, then there exists k true extremal assignments A , A , . . . , A and k false non-extremal assign- 12 k k k ments B , B , ..., B such that L A. = L B. . Since each B. is -Lc. K. _ i . _ i i 1=1 1=1 non-extremal and false, there is another false assignment U. > B. . ^k^k 1 ^ 1 Since U, > B. , 1- (U, - B. ) > 1. Therefore, 1- ( E U. - E B. ) l i' l l — ' . , i . , i i=l i=l k -> k = l-( E U. - E A.) >k. Hence r > 1. Thus if V < 1, f(x) is not i=l x i=l X extremely summable and by Theorem 3.3-2, f(x) is quasi-separable. Q.E.D. 83 The above theorem gives the reason that the bounds on I? which were found in Example 3«^.1> are sufficient to determine the type of separation for this function. 89 CHAPTER 4 SYNTHESIS OF PARTIALLY SEPARABLE FUNCTIONS 4.1 Introduction In this chapter methods for realizing arbitrary partially separable functions in networks of threshold elements are presented. In the first section an algorithm to minimize the number of true assignments (or false assignments) lying on a pseudo-separating hyperplane is discussed. In the second section a procedure for network synthesis for pseudo- separable functions is given. Algorithms to minimize the number of assignments not segregated by a marginally separating structure are developed in the third section. In the last part a general synthesis procedure for marginally separable functions is outlined. 4.2 Integer Programming Algorithm for Pseudo-Separability Before any discussion about an algorithm which minimizes the number of true assignments (or false assignments) lying on the pseudo-separating hyperplane is presented, we should explain why -a hyperplane which has the minimum number of true (or false) assignments will be optimal for our network for a given function. The complete procedure for network synthesis will be given in Section 4.3. (This does not mean that the network discussed in Section 4.3. has the minimum number of threshold elements. For a synthesis which guarantees the optimum number of elements see Muroga's synthesis by integer linear programming. [7] It requires in general many more inequalities 90 than the synthesis of the present paper)* • Our primary concern is to find a network of threshold elements to realize a pseudo- separable function f(x), using the pseudo-separating hyperplane as the threshold structure for one of the threshold, element s. jj The network, which we propose to use, employs the pseudo- separating hyperplane as a threshold element and then uses other threshold elements to correct for either false (or true) assignments which lie on this hyperplane. An important point is that we must correct for all such true assignments or all such false assignments, but not both. To see this, assume that f(x) is pseudo-separated by [W; T]. Let the upper bound of [w; T] be g(x) and the lower bound be h(x) . Assume that true assignments A,, A„, ..., A, and false assignments B., IL, *.., B lie on [W*; T]. In using [w; T] as a threshold structure, we have two alternatives. If k > e we can assume that all assignments which lie on [W; T] are true for f(x) and then add extra threshold elements to correct for those which are actually false. This corresponds to representing [w; T] by its upper bound g(xj and then correcting for B.. , B_, . .., B . If e > k we can assume that all assignments on [w; T] are false and then correct the network for those which are actually true. This corresponds to repre- senting [w; T] by h(x) and adding extra threshold elements for — * — i — i A,, Ap, ..., A, . Either alternative is permissable since both of the resulting networks will realize f (x) . When h(x) or g(x) is realized as a threshold element, we then need extra elements to correct for only the true or false assignments on [w; T]. Therefore, we want a pseudo-separating hyperplane where the number of such extra elements is minimal. 91 Because we have the choice of correcting for either true or false assignments, it would seem that we must devise two algorithms, one to find [w ; T ] where the number of true assignments on this hyper- plane is minimal and another to find [w ; T ] where the number of false assignments is minimal. Then we would choose the hyperplane which requires the least number of extra elements in our network. Actually, we can combine both of these algorithms into a single algorithm to determine [W; T] for which the total number of assignments on this hyperplane is minimal, as follows: Lemma 4.2.1: Let f(x) be pseudo -separated by [w ; T ] . Let A, , Ap, ..., A, be all true (or false) assignments lying on [w ; T ]. If [w ; T ] minimizes the number of true (or false) assignments lying on it, then every hyperplane which pseudo-separates f(x) must contain A* A* A* 1' 2' ' ' "' k* Proof: Assume [w ; T ] minimizes the number of true assignments lying on it. If some hyperplane [VL ', T ] segregates at least one of these A. ( [w. ; T, ] may contain true assignments not on [W ; T ]), then [W + W ; T + T ] will segregate all true assignments segregated by either [vf 1 ; T 1 ] or [#; T^] (by Theorem 2.3.1). Hence, [#" + #j T 1 + T^] will contain less true assignments than [W , T ] } a contradiction. Similarly, if false assignments A,, A~, ..., A, lie on [w ; T ], then no pseudo-separating hyperplane can segregate any A. . Q.E.D. 92 Theorem U.2.1: Let f(x) be pseudo- separated by [W% T ]•, [V?" 1 ; T 11 ] and 1 m l- [Tf;' T]. Assume that the number of true assignments on [w ; T ] is minimal for all pseudo-separating hyperplanes. Assume that the number of false assignments on [W j T ] is minimal for all pseudo-separating hyperplanes and that the total number of assignments on [w; T] is minimal. Then every assignment that is segregated by either [W ; T ] or [W ; T ] must be segregated by [Wj T], i.e.. if true assignments A, , A , ..., A lie on [W j T J and false assignments B , B c , . „ . , B r~€.l 11-, -* ~* -»-»-» -> lie on [W j T ], then only A , A , ..., A , B , B , ..., B will lie on [w$ T]. (Note the following: this theorem does not assert that the number of false assignments lying on [W ; T ] or the true assign- ments lying on [w -, T ] is minimal. ) Proof : Xf [W ; T ] minimizes the number of true assignments lying on - * ~* — * — 1 1 — *L1 11 it and if A , A , . .., A lie on [W -, T ], then neither [W~ j T ] nor [w; T] can segregate any of these A. (by Lemma U.2.1). Similarly, neither [wj T ] nor [W; T] can segregate any B. . Therefore [W; T] -> -» can not segregate any A. or B. . Consider the hyperplane [w + w ) T +■ T ]. This hyperplane pseudo-separates f(x) and - * ~* i contains all A. and B. by Lemma 4.2.1. Also it segregates all assign- "~ 1 1 " - ' , *11 11 ments that are segregated by either [W ; T ] or [W j T ] by Theorem 2.3.1. Therefore, it must segregate all assignments except A , A , ..., A , B , B , ..., B . Because these assignments must also lie on Iwj T], [w j T] must also segregate all other assignments. If 93 — > — »]_ — >n i 11 [W; T] did not segregate all other assignments, then [W +■ W ; T + T ] — > ; would contain less assignments than [Wj T] contradicting the assumption that the number of assignments on [W> T] is minimal. Q.E.D. r~tl It As stated in the theorem, [W > T J, while segregating the largest number of true assignments, need not segregate the most false assignments. In addition to B , B , . .., B , there may exist other — »}_ ]_ ~~ *11 11 false assignments on [W j T ]. Similarly, [ w \ T ] may contain more true assignments than A , A , ..., A . However, LW; TJ is guaranteed to contain both the least number of true assignments and the least number of false assignments. Therefore, when we need to solve two problems, one to minimize the number of true assignments on the pseudo-separating hyperplane and the other to minimize the number of false assignments, it is sufficient to solve one problem to minimize the total number of assignments there. We now introduce a definition to facilitate our discussion of the threshold elements used in our synthesis procedure. Definition U.2.1 : Let f(x) be a separable function. Assume that [W$ T] separates f(x). [Wj T] is called a threshold structure for f(x) if for any assignment A, f (A) = 1 implies W-A > T and f (a) = implies W«A < T - u where u, a positive constant called the gap of f(x), has been predetermined. The gap u is necessary for the reliable operation of the — > threshold element with structure [Wj T] and is predetermined. [7] 9 k By regarding w./li and T/U as a new structure, u may always be normalized to 1. In the rest of this section we will assume that the gap is 1. let us now discuss the actual algorithm to minimize the total number of assignments on a pseudo-separating hyperplane for a given function. Let f (x) be an arbitrary Boolean function. We want to develop and algorithm based on integer linear programming to find the optimal pseudo-separating structure for f(x). In other words, we want to find [W; T] such that -» — > — » (1) if f(A.) = 1, W.A. > T l l — -> -> -> (2) if f(B.) = 0, W-B. < T (3) the number of assignments which lie on [Wj T] is minimal. Consider conditions (l) and (2). These can be rewritten as constraints using slack variables a. >~ and a . > as follows: (1)' if f(A\) = 1, W«A. - a. = T l ' i i (2)' if f(B*.) = 0, WoB. +a. = T Now let us turn to the problem of choosing an objective function to minimize the number of assignments which lie on the hyper- plane [wV T]. 'To minimize the number of such assignments, -consider ,tha.t if A., or B resides on [w; T], then the corresponding a. (or a.) must be J -^ J 95 0. Thus, we want to introduce into the objective function a penalty for each assignment of CC. = or a. =0. Let J. be a variable with the l j l following properties: a. +• J. ^ .1 for all i = 1, 2, . .., 2 and J. = or 1, tvhere lisMihe) 'normalized gap). We then add a penalty to the object function if J. 4 (or a. = 0). ° i ' l n n 2 Define the objective function as S |w. |+ F( E J.) where i=l 1 i=l X n 2 n n F » E lw. J Then our procedure will be to minimize F(Z -J. )+ E |w. I i=l 1 i=l 1; i=l x under the constraints (1) if f(£.) = 1, f-1. - a. = T l ' i i (2) if f (B ) = 0, W-B* + a, = T J' 9 J (3) a. $i 0, a. > formal!, i and j (U) a. + J. > 1 for all i and j 11 — (5) j. = v or 1 for all i and j. 2 n If E J. is minimized, then the number of assignments on i=l X the hyperplane is minimized because if J. = 0, OL. must be positive — » and the corresponding assignment A. is not on the hyperplane. Therefore the resulting hyperplane [Wj'.T] must segregate the most assignments and hence must be optimal. Because we know that [0; 0] pseudo-separates every Boolean function (by Lemma 2.2.1), we can use [Oj 0] as an initial feasible 96 solution to our problem and hence we do not need artificial variables. [7]. k.2) Threshold Synthesis of Pseudo-Separable Functions In this section we will use the algorithm which was presented in the previous section to develop a synthesis procedure for pseudo- separable functions . The basic threshold configuration which we will use in this synthesis is shown in Figure 4.3.1* when the threshold element JI repre--:- sents the pseudo-separating structure, and also has additional inputs from the elements Q , fi , ..., Q , which represent corrections:' for either the true assignments on this hyperplane or the false assignments. Figure U.3.1; Configuration for Synthesis of Pseudo-Separable Function 97 Let f(x) be an arbitrary pseudo- separable function. Let [w; T] be the pseudo -separating hyperplane which segregates the greatest number of assignments. (This hyperplane can be found by using the algorithm outlined in the last section) . Let h(x) be the lover bound of [W; T] and let g(x) be its upper bound. Assume that true assignments A*, , A* 2 , . .., At and false assignments K, Bp, . .., B lie on [w; T]. Let e > k. Then in order to minimize the number of extra elements ft., ft p , . .., ft , we will represent [w; T] by its lower bound h(x) and will correct for all true assignments A.., A p , . .., A, . (Actually, we have the alternative of deleting all true assignments on [w; T] in h(x) , and using the resulting partially specified function to represent II . For notational ease, however, we will simply use h(x) .) The problem of determining an optimal threshold structure , , n [w ; T ] for h(x) (optimal, in the sense that £ | w. | is minimal for i=l 1 the gap u or by normalization for the unit gap) has been investigated previously. Muroga, Toda, and Takasu present a clear discussion of this problem. [8] Therefore, the procedure will only be outlined here. (1) Convert h(x) into a positive function p(x) (2) Solve the following linear programming problem for [ton; T ] n Minimize £ w. under the constraints i=l X (a) all w p > (^.5.1) (b) if v (t) = 1, \P-t> T p (c) if p(ft = 0, fc*.B*T h or a. > Z |w h | + T h . i=l i=l (Actually, we need only sum w. which are negative, but the Labove lower bound on a. is sufficient to guarantee the operation of the circuit) . 99 We now will show that the network we have designed realizes the function f (x) . Let B be any false assignment to f (x) » Since [W; T] pseudo- separates f (x) and h(x) is- the upper bound, h(B) =0. Since each element SI. corresponds to one true assignment on [w; T], the output from each of these elements will be (T.< Therefore, the element II will output 0. If A is any true assignment, there are two possibilities, either h(Aj = 1, or h(A) =0. If h(Aj = 1, then II must output 1 because it realizes h(x) . If h(A) = and A is a true assignment, then g(A) = 1 and A must reside in the hyperplane [W*; T] (by Corollary 2.5.1) and hence one of the ft. must have output 1. Because h n h a. > T + £ | w 4 j t then n must also have output 1. i=l Therefore, this network realizes f (x) . In most of this work we have assumed that there were more false assignments on [w; T] than true assignments. However if there are more true assignments than false, we can repeat the same procedure, with slight modification, using the upper bound instead of the lower bound and then correcting for false assignments instead of true. h-.k Integer Programming Algorithm for Marginal Separability The procedure for synthesizing a marginally separable function in a network of threshold elements which we will use in very similar to the technique discussed earlier in this chapter for pseudo -separable functions. This approach entails representing the marginal structure which segregates the largest number of true (or false) assignments by its marginal lower (or upper) bound and then adding extra elements to 100 correct for the true (or false) assignments which were not segregated by the original structure . Marginal separation differs from pseudo- separation in that it does not possess the property of Theorem ^.2.1, i.e., we have no guarantee that by finding a marginal structure which segregates the largest total number of assignments we automatically can find no other structure which segregates more true assignments but less false ones (or more false assign- ments and less true ones) . Therefore, if we want to minimize the number of extra elements we must add to the network, we must solve two problems. The first problem will be to m i ni m i ze the number of true assignments not segregated by the marginal structure and the second will be to minimize the number of false assignments not segregated. We now discuss an algorithm which will find [wl ; T-.,A,] which maximizes the number of segregated true assignments. Later we will show how this algorithm can be modified to .maximize the number of segregated false assignments. Let f (x) be an arbitrary Boolean function. We want to devise an algorithm to find [W.; T,,^] which segregates the largest number of true assignments for f (x) . If [W.; T-.,^] marginally separates f(x), then we must have: ■ (1) if t(t ± ) = 1, then \'t ± £ T - Aj/2 (i = 1, 2, ..., k) (2) if f (I? ) = 0, then W^-lt < 0^ + Aj/2 (i = k + 1, ..., 2 n ) Let W- = (w..-., ..., w.. ) and let (? = (c n , ..., c ) be any 1 11' ' lir 1' ' n assignment. Let w' = (v,,, ..., w ,&-/2) and C' = (c.., ..., c , 5) 101 where 5 = 1 if f(C^) = 1 and 5 =-1 if f(C?) = 0. Using this notation, the above requirements become: (1) if f(t.) = 1, then ft 1 'tl > T n -and l 1 i — 1 (2) if f(lf. ) = 0, then fl •#! < T n x ' i' ' 1 l — 1 As we did in the previous algorithm in Section ^.3> we now add slack variables a. > for all i. Our constraints then become: i — (1) if f(t.) = 1, tf'-A*. - a. = T for 1, 2, ..., k i ' 1 i i l ' ' (2) if f(]?.) = 0, tf'.it + a. = T n for i = k + 1, ..., 2 n (3) a. > for i = 1, 2, ..., 2 n To minimize the number of true assignments not segregated by the marginal structure, we must penalize the objective function for any true assignments such that ¥ »A. < T + A,/2. ; Such assignments must have jc a. < A* • As we did for pseudo-separable functions, we introduce a variable J. for each a.(i = 1, 2, . .., k) with the following properties: a i + ('^1 + 1 ^ J - - \ + 1 and J - = ° or 1 « An important fact about such a procedure as this is that we can predefine A, at the outset and still be guaranteed that we will find a marginal structure which segregates the most true assignments, because if [w..; T ,A, ] segregates the most true assignments and A'. r is any other number, then IF 1' A' 1' A' : 102 ■will then segregate exactly those assignments segregated by [w, ; T ,A, ] • Therefore, -we can let A, = 1. Then J. is defined so that a. + 2J. > 2 and J. = or 1. T. i 11-1 k n We then add F( E J.) to our objective function where F » E |w . | i=l x i=l - LL Because [0; 0] pseudo-separates every function (by Lemma 2.2.1), the marginal structure [0; 0,0] marginally separates every function and hence can be used as an initial feasible solution. Therefore, there is no need for artificial variables. [7] (Kenneth Breeding has formulated a similar procedure for a similar problem) . Our mixed integer programming algorithm which will determine [w j T ,l] segregating the greatest number of true assignments is K n to minimize F( E J.) + E |w n . I under the constraints . , l , ' li ' i=l n=l (1) tfl-tl -a. = T if f(t.) = 1 (i :=.1,.2, , k) (2) tfl'tl + a. = T n if f(lf.) = (i = k+1, ..., 2 n ) 1 i i 1 i (3) a. > for i = 1, 2, ..., 2 n (^.l) (k) a. + 2J. > 2 for i = 1, 2, ..., k. l — (5) J. = or 1 for i = 1, 2, ..., k. 103 k Again note that "by adding F( E J.) as a penalty to our i=l x objective function we can minimize the number of true assignments n not segregated by the .marginal structure by minimi zing E J. because i=l 1 J. = 1 if and only if A. is not segregated. As we have stressed, the algorithm outlined in (^A.l) ■will determine the marginal structure which segregates the largest number of true assignments. No claim of any sort is made for the number of false assignments segregated. Therefore, in order to find the juarginal structure [W_, T n ,A_] which segregates the most false assignments, we must solve another mixed integer programming problem which is obtained by modifying this algorithm. If we want to minim ize the number of -false assignments not segregated by the marginal structure, we must penalize the objective function for any false assignments such that « 'if. > T fi - A-,/2. -A false 'assignment B. will not be segregated if < a. < A.. Therefore, the variable J. must be introduced so that a. + ( A. + 1 and n J. = or 1 for i = k + 1, ..., 2. Hence the algorithm to minimize the number of non-segregated false assignments is to: c n d n minimize F( Z J.) + Z w . under the constraints i=k+l i+I (1) $^l - a..="T ,if i\t ± ) = 1 (i = 1, 2, _., k) (2) ^.g: + a. = T Q , if f(B\) = (i = k + 1, ..., 2 n ) 104 (3) a. > for all i = 1, 2, ■'-'..;, 2 n (UA.2) (4) a. + 2J >>2 for i * k+1, . .., 2 n (5) J. = or 1 f or i = k+1, ..., 2 n 4.5 Threshold Synthesis of Marginally Separable Functions The synthesis procedure in this section is very similar to the procedure for threshold synthesis of pseudo-separable functions outlined in Section 4.3. The basic configuration of the circuit -which we will use in our threshold network will be the same as before . The reader is directed to Figure 4.3.1 for this configuration. Once again, II represents the marginal separating structure as a threshold element and the elements Q, , Q p , . .., 0, represent corrections for either the true or false assignments not segregated by this marginal structure, depending upon which assignments not segregated have a fewer number. Let f (x) be an arbitrary marginally-separable function. Let [w.,; T, ,A, ] be the marginal structure which segregates the most true assignments and [w n ; T ,A~.] the structure which segregates the largest number of false assignments. (These structures can be found by using the algorithms (4.4.1) and (4.4.2).) Assume that true assign- ments A. ,. A~, ..., A are not segregated by [w, ', T .,&. ] and false assignments E_ , B_, ..., B are not segregated by [W ; T q ,aJ. 105 Assume that k> e. Then in order to minimize the number of extra corrective elements ft . , we use [W-j ^-i'^i) an ^ we will represent this marginal structure "by its marginal lower bound h (x) and then — > — > — > correct for A, • A„, .... A in the network, 1 d e Having determined that -we will represent [w ; T-,A,] by its marginal lower bound h (x) and add extra elements for — > — > — > A,, A p , . .., A , the rest of the procedure is exactly the same as we outlined in Section h .3 and will be omitted here. Notice that if e >> kj then we must use [^L; T n ,/V] and we must represent this marginal structure by its marginal upper bound . g n (xj and then correct for false assignments BL , B p , . .., B, by adding extra elements, Q. . There is another approach to synthesis of marginally separable functions. Assume that some procedure persists which yields the ; '■ structure [W, ; T,,,^] for which the total number of assignments of f(x), not segregated, is minimal. (Such an algorithm can be obtained by combining the mixed integer linear programming algorithms (U.^.l) and (h.k.2).) lUse two threshold elements, one corresponding to the marginal upper bound of [w ; T , A,], g, (x) , and the other corres- ponding to its marginal lower bound nix) to separate all of the assignments segregated by [\f ' ; T,,Al. Let f , (x) be the partially specified function which results from f (x) when all assignments segregated by [w ; T,,A,] are deleted, (all assignments A where g, (A) = h (a) ) . Repeat the procedure to find an optimal marginal io6 structure for f , (x) , [w p ; T , A~J by solving the mixed integer linear programming. Again construct two threshold elements corresponding to the marginal upper and lower bounds of [w p ; T p , A-,]. If all assignments for f(x) are segregated by either [W, ; T ,A.]or fw p j T p ,Ap] "we will terminate the procedure. If not, find f p (x) by deleting all assignments segregated by [w p ; T p ,/V,] from f , (x) and repeat the procedure. In this r, way a finite chain of marginal structure can be found -with the property every assignment of f (x) is segregated by at least one structure in the chain (by Theorem 3-^«3)« Figure 4.5.1; shows how the pair of threshold elements g. and h. corresponding to [ w . °, T.,A. ] will be connected to the pair of -J. elements g.,, and h. n corresponding to [w. ., ; T. ...A. -. ] . We want D i+1 l+l * ° l+l l+l' i+I to weigh the inputs of g. and h in such a way sq that if the output of g. is the same as the output of h., then both g. and h. will have the same output. Let g. have structure fw ;T. , ] and h. n have structure [w : T. n ]. Let the input weight a. corres- l+l i+I l n . ponding to g. be such that a. ' <■- £. |w. j and let the input weight 1 x 3=1 J corresponding to h. be B. = | a. |. Then if g. = h. = 1, because of the negation (expressed with a short bar in Figure 4.5.1) of output of g- }&• = and both g. , and h. n will output 1. Similarly if g. = h. = 0, •^i'^i & i+l i+I * J °i l then both g. n and h. , will output 0. i+I i+I 107 Figure U.|?.l : Configuration for Synthesis of Marginally Separable Function. i'oa CHAPTER 5 SUMMARY The major effort in this paper has been to explore the theoretical properties of pseudo and marginal separation. Most of this investigation has been directed toward the algebraic properties which a Boolean function must have if it is to be pseudo or marginally separated. The basic concepts of pseudo-separation and marginal separation were introduced in Chapter 1. In Chapter 2 some general properties of pseudo-separable functions were presented. The concept of a supersummable function was introduced and used to determine necessary and sufficient conditions for pseudo-separability. The relationship between simultaneously realizable functions and the upper and lower bounds for a hyperplane were also discussed as well as a method for determining better bounds if they exist. The correspondence between separable and pseudo- separable functions was also discussed. Chapter 3 was devoted to the algebraic properties of margin- ally separable functions. The idea of quasi-separation was introduced and a necessary and sufficient condition for this type of separation was presented in terms of extreme summability. The concept of the degree of supersummability was also discussed and the relationship between this parameter and the various types of separation was explored. Integer linear programming formulations to find optimal segregating structures for an arbitrary partially separable function 109 were discussed in Chapter k. Synthesis procedures based on these algorithms for realizing a function in a network of threshold elements were also introduced. This work represents an attempt to generalize the concept of a threshold element and the idea of total separation. It serves two purposes. First, it presents a new approach to the problem of threshold synthesis in terms of a more general type of separation. Second, it provides better insight into the properties of separable functions by allowing them to be examined as special cases of more general types of separated function. 110 LIST OF REFERENCES [I] Chow, C. K. , "On the Characterization of Threshold Functions", Switching Circuit Theory and Logical Design , pp. 31-38* Sept. 1961. [2] Elgot, C. C, "Truth Functions Realizable by Single Threshold Organs," Switching Circuit Theory and Logical Design , pp. 225-21+5, Sept. 196l. [3] Fan, K., "On Systems of Linear Inequalities," Linear. Inequalities and Related Systems, Annals of Mathematical Studies #38, Princeton University Press, 195^ [U] Hadley, G., Linear Programming , Addison-Wesley Publishing Company, 1962. [5] Hadley, G. , Non-linear Programming , Addison-Wesley Publishing Company, I96U. [6] McNaughton, R., "Unate Truth Functions", IRE Transactions on Electronic Computers , Volume EC -10, pp. 1-6, March I96I. [7] Muroga, S., "Threshold Logic", Lecture Notes, Department of Computer Science, University of Illinois, 1966. [8] Muroga, S., Toda, I., and Takasu, S., "Theory of Majority Decision Elements", Journal of the Franklin Institute , Volume 271* pp. 376-418, 196l ' [9] Paull, M. C, McCluskey, E. J., "Boolean Functions Realizable with Single Threshold Devices", Proceedings of IRE , pp. 1335-1337, July i960. [10] Robinson, A., "Some Remarks on Threshold Functions", IBM Report, 1963. [II] Winder, R. 0., "Threshold Logic", Ph„D. Dissertation, Princeton University, 1962. [12] Yen, Y. , "Multi-threshold Logic", Ph.D. Dissertation, University of Illinois, 1966. Ill VITA Thomas Andrew Slivinski was born in Evergreen Park, Illinois in 19^+2. He received a B. A. degree in Mathematics from Lehigh University Bethlehem, Pennsylvania, in June 19^3 • ^ n September 1963, he began his graduate study at the University of Illinois, Urbana, where he received a M. S. degree in Mathematics in February 1965. Since September 19^3^ he has been working as a research assistant in the Digital Computer Laboratory, the University of Illinois, Urbana. % %