HD28 .M414 V DEWEY ZoSo- ALFRED P. WORKING PAPER SLOAN SCHOOL OF MANAGEMENT Projective Transformations for Interior Point Methods, Part II: Analysis of for finding the An Algorithm Weighted Center of a Polyhedral System Robert M. Freund Sloan W.P. No. 2050-88 October 1988 MASSACHUSETTS INSTITUTE OF TECHNOLOGY 50 MEMORIAL DRIVE CAMBRIDGE, MASSACHUSETTS 02139 MASSACHUSETTS INSTITUTE OF TECHNOLOGY NOV 1 3 2000 LIBRARIES Projective Transformations for Interior Point Methods, Part II: Analysis of for finding the An Algorithm Weighted Center of a Polyhedral System Robert M. Freund Sloan W.P. No. 205U-88 October 1988 Projective Transformations for Interior Point Methods, Part Analysis of An II: Algorithm for finding the Weighted Center of a Polyhedral System Abstract In Part II of this study, the basic theory of Part I applied to the is X We present a problem of finding the w-center of a polyhedral system projective transformation algorithm, analagous but more general than . X Karmarkar's algorithm, for finding the w-center of The algorithm At each iteration, the algorithm either improves the objective function (the weighted logarithmic barrier function) by a fixed amount, or at a linear rate of improvement. This linear rate of improvement increases to unity, and so the algorithm is superlinearly convergent. The algorithm also updates an upper bound on the optimal objective value of the weighted logarithmic barrier function at each iteration. The direction chosen at each iteration is shown to be positively proportional to the projected Newton direction. This has two consequences. On the theoretical side, this broadens a result of Bayer and Lagarias regarding the connection between projective transformation methods and Newton's method. In terms of algorithms it means that our algorithm specializes to Vaidya's algorithm if it is used with a line search, and so we see that Vaidya's algorithm is superlinearly convergent as well. Finally, we show how to use the algorithm to construct well-scaled containing and contained ellipsoids centered at near-optimal solutions to the w-center problem. After a fixed number of iterations, the current iterate of the algorithm can be used as an approximate w-center, and one can easily construct well-scaled containing and contained ellipsoids centered at the current iterate, whose scale factor is of the same order as for the w-center itself. . exhibits superlinear convergence. Keywords: analytic center, w-center, projective transformation, Newton method, program. Robert M. Freund, Sloan School of Management, M.I.T., 50 Memorial Drive, Cambridge, Mass. 02139. Author: Abbreviated ellipsoid, linear Title: Projective Transformations, Part II. Introduction L Part of this study uses the projective-centering II methodologies developed in Part I [ 3 , ] to and the improvement local develop an algorithm for finding the A w-center of a polyhedral system, which is to the the solution x problem m Pw : maximize V w = F(x) - AjX) In (bj 4 i=l Ax subject to + s =b (1.1) s>0 Mx = g m w where = (wj, . . . , wm are positive weights that are normalized to ) V w = 1, 4 i=l and F(- is ) a (weighted) logarithmic barrier function. generalization of the analytic center problem defined nonuniform positive weights on all by Sonnevend If and X of . is bounded To be more since the solution to X . However, where 8 ], [ 9 ], there exists a point X if [ it is , precise, as in Part is I understood that hyperplanes. x e then Pw among Pw [ is 11 a ], [ 12 ] , for This problem has had numerous constraints. applications in mathematical programming, see Renegar Monteiro and Adler Problem P w [ 10 ], Gonzaga [ 5 ], and others. X = {x e Rn Ax < b, Mx = g) for which Ax < b, A will have a unique solution x , called the w-center we should | say that A x is the w-center of X (A, b, M, g), dependent on the particular polyhedral representation of of this study, X we will refer to A x as the w-center of X represents a specific intersection of half-space and , For the case algorithms known problem. Vaidya [ when all are identical, there are 4 two other author that have been developed for the w-center to this 14 w weights has developed an algorithm that constructs the ] Newton-direction from the current iterate and then performs an inexact He shows in this direction. improvement that at each iteration, there improvement in F(x) or a linear rate of exhibits linear convergence. for finding the center, that is Censor and Lent convergent to [ 2 A x , is line search either constant in F(x), and so his algorithm present a primal-dual algorithm ] but not necessarily in any strong sense. The algorithm developed in this paper is based on the use of projective transformation methods to w-center a given point, as in Section rV of Part iteration the to a polyhedron polyhedron Z direction d is or unbounded. X = (x e Rn so that the current point If it is . A bounded, will then it steplength a is The algorithm can be run with by the use of iteration, there is either a constant one Mx = x in F(x). is g] the w-center of is used If A Pw search is a determined The new point x NEW x + a d back from Z improvement in F(x) , the linear rate of then determined by is to the steplength is is analytically at each X . At each or a linear rate of improvement approaches The search positively proportional to chosen at each iteration by a line search, the algorithm specializes to Vaidya's algorithm (with equal weights), shows (obliquely) bounded then computed. However, because direction. . produce an (updated) upper bound on direction determined at each iteration of the algorithm Newton Z first to test if in the limit, the algorithm exhibits superlinear convergence. the projected At each projectively transformed is a steplength a line search. projectively transforming the point improvement < b, then determined. This direction A the value of F(x) iteration, or Ax | I. and this that Vaidya's algorithm exhibits superlinear convergence, verifying a conjecture of Vaidya 15 [ might exhibit stronger convergence that his algorithm ] properties. As was shown in Section c X c E OUT — ((1 at the I, ellipsoid A w-center x of E OUT X one can construct with property that A A x , of Part and an outer an inner ellipsoid E 1N E IN II the center of is and (E OUT - x , = ) — A — - w)/ w)(E IN E IN and E OUT w = min x), where [wj] This ratio . (m - is 1) when all i A weights are identical. Thus x we algorithm is number construct ellipsoids F IN and FOUT (FOUT - x is (1.75m + the same = (1.75/w + ) 5) A x which as for is is of iterations it , it may ) . When all x In Section IV, we prove Section VI [ 14 ] , we show we show by showing II and unboundedness that the objective value that the algorithm the relationship that the direction proportional to a projected fixed number of iterations, , x "close one can easily weights are identical, then 0(1/ w) this ratio is , , and this ratio which is presents the projective . In Section III, we algorithm are valid. tests in the bounds produced by the algorithm each iteration are valid, and prove that the algorithm In Section V, never reach the c X c F OUT transformation algorithm for solving the w-center problem P w tests Although the • organized as follows: Section prove that the optimality . will exhibit a point with the property that F IN O(m). In general, the order of E IN and E Q ut The paper , - x (F IN 5) X of following sense: at the point to the w-center, in the X w-balanced point of present converges to the w-center w-center. However, after a fixed enough" in a sense a is at linearly convergent in F(x). exhibits superlinear convergence. In between the algorithm and Vaidya's algorithm d determined at each step Newton direction. one can easily construct ellipsoids In SectionVH, is positively we show that after a F IN and F OUT about (1.75/w + 5) (F IN - x A EL ) Section . VEQ contains study [ 3 ] . We We conventions. assume the reader also will cite Let the given data (A, b, X = {x e w> X c F OUT , and (F OUT Rn | Ax < b, Mx = g] are exactly the same familiar with these notation of the results presented in Part M, define the polyhedral system w e Rm Let . solving the w-center problem g) w= 1, in ) = . I of and I. be a given vector of weights satisfying where e = P w given X as in Part many that e T and normalized so is - x closing remarks. Projective Transformation Algorithm for Finding the w-center of The notation and conventions used here this c x with the property that F 1N the current iterate (1.1). (1, . . 1) , . T We make . Our interest lies in the following assumptions regarding the data: (2.1a) The matrix A mxn is and has rank n. (2.1) The matrix (2.1b) M is k x n These assumptions are and has rank for convenience, k. and if A or M lacks full rank, then one can either eliminate variables or constraints, or one can replace certain matrix inverse operations with pseudoinverse operations in the analysis. Exactly in the spirit of Karmarkar's algorithm and consistent with the local improvement algorithm of Section solving Pw . polyhedral system x = g , and of Part I, we have Let the data for the problem be [w, A, b, vector of weights satisfying M V e > X , is x is w>0 and eT w= the starting point, 1, M, (A, b, the following algorithm for g, M, which must the optimality tolerance. x, e ] Here . w is the g) are the data for the satisfy s = b-A x>0 and w = min/w Let 1 we will Because J. w/(l - w) extensively use the quantity i we throughout the description of the algorithm and the subsequent analysis, w/(l - w) the constant k = bound on for convenience. Pw the optimal objective value of Also, in the algorithm, F* is define an upper . Algorithm for the W-Center Problem StepO w=min{Wi }. Set Set k = ( w/(l - w)) Set F* = +°° . . i Step Set 1 b- A x y = , AT _1 S w . (Projective Transformation of constraints) Step 2 Step 3 s~= (Compute direction in d be the solution Let Z A = A- sy T space) to the problem (P3) : - y Td maximize d subject to dT A S^W S' 1 Ad < k Md = If this If y problem T d = 0, is unbounded, then stop, x solves Pw stop. Pw is unbounded. . Step 4 (Perform Boundedness Tests and Update Upper Bound). Set y If = (-y T d) /k y £ l/k,stop. Problem P w is unbounded, and d then set F* = min {F*, F( x) + y + 2 If Y < If 2 Y < .08567 then set F* = min {f*, F( x) + .669kY } 1 Step 5 (Compute Steplength) Set cc= 1 ; Vl+2 Y is Y /(2(l-Y))} a ray of X. ) Set z NEW = a x + Step 6 (Transform back to original XNEW = d. X space ) z new ~ x X + 1 y^NEW" + Step 7 (Stopping Criterion) Otherwise go x <- x NEW Set Step to *> F* - F( x) If . < e , stop. 1. Note This algorithm has the following straightforward explanation. Pw is an instance of the canonical optimization problem (5.1) of Part I, first that namely m P minimize F : x,s q p ln(q-p Tx) - = (x) £w ln(bj - AjX t ) i=1 subject Ax to: + s =b s > (2.2) Mx = g T p x <q where q =1 and p = , (0, . . . , 0) T , and hence F(x) = -F (x) = -Fj (x) can proceed with the local improvement algorithm presented in Section Let x e int projectively Z X be given, transform = {z e Rn with the function | let X s = b- = is 1 , and (A - sy T )z < b - Ty T x z = g(x) = P x y = let AT S _1 w . Thus we of Part I. Then we can to X— X ~ x + 1 Problem P A V . -y l transformed to , Mz = =- (x- x , g} as in (3.2), (3.3) and (3.4) of Part I. [l_ y T x (A - sy T )z + subject to z ) _ £ w t = b- sy T x t > ]_[_ y T] Uz equivalent to is Parti. Note sy T Z w-center of A < , p replaced -y , A and p replaced by (2.3) - yT x 1 problem q replaced by 1 (5.1) of - yT x , and -y) has the property that Lemma Part I x + 3.2 to replaced x the is of Part (5.3) of (ii) A with Because etc. d given by the solution then the direction , . (and, of course (1.1)) according to (2.2) that (2.3) is also an instance of A= A- by lnt = g -y Tz which . d maximizes I (with -y T z over ze {zg Rn |Mz = g,(z- x) T E IN = Let y = -y T d w - w)/ (1 precisely the quantity ) a = 1 - 1/\ g(- ), 1 + 2y namely x = . ) and z NEW = < G( x — h(z) = ^NEW — x + 1 + y ( F(x NEW ) - F( x) = F from Lemma 3.2 (ii) I W S" 1 A(z- x)< w/(l (note that ) - ( x + a d. x + — x ( x) + z— x = =- yUz- x =- z new ~ x 1# q - pT x = I. This 1). a= 1 X V^+2y ) . is -1/\1 + 2y 5.1 if space using the inverse of (see (3.4) of Part I) , we obtain ) as in Step 6 of the algorithm, (x NEW ) > ( w/(l - w )) and (l + y - , of Part ' - F^ of Part - w)} Then by Corollary w/(l - w)) (l + y- Projectively transforming back to 1 x new = ! defined in Step 4 of the algorithm. Let y = G( x + a d S- as in (5.6) of Part as in Step 5 of the algorithm, G(z NEW AT Vl+2y ) , (2.4) I, In particular, will use y whenever y ^ -08567, then F(x NEW ) - F( x) > is and we greater than or equal to a given constant, we have ( w w/(l - )) (.0033) (2.5) we make Before proceeding with the analysis and verification of the algorithm, the following remarks. Remark Use of 2.1. Because the projective transformation line search. directions relative to x, one needs 8 > x + 8 d feasible. and there will be at ) (3.3) of ( Part Because F(x) most one maximizer of start the line search I) X preserves directly. value of 8 that nearly maximizes F( x + 5 d to find a F( x 8 = with corresponds to a step of size a = 1 + ay - l/*\ 1 1 concave over is strictly + 8 d zr-= 1 Z g(- the line search can be performed in the space Specifically, could Steps 5 and 6 of the algorithm can be replaced by a a line search. x e X , then over the feasible range of 8 ) over ) One . where a = l-l/"Vl+2v which , d + 2y in the projectively transformed space . Remark 2.2. Efficient algorithm of Part dense matrices procedure projected IIL is A I, one can compute = A- sy T or d in Step Newton As 3. in the linear programming d without working with the possible very Q=A S'^W deferred to Section VI, where S" 1 A we show . The discussion that d is of this proportional to the direction. Optimality Test and Unboundedness Test In this section, valid, Computation of and that the we show that the optimality test of Step 3 of the algorithm unboundedness tests of Steps 3 and 4 are valid. is Proposition (P3) unbounded, then P w is Suppose Proof: r e the algorithm terminates in Step 3 because optimization If 3.1. Rn is unbounded. that at Step 3, that (P3) such that T r = -1 y Mr , problem = is unbounded. That means there and Ar = 0, 0. Ar = But exists a ray Ar = sy T r = implies m - s < 0. Thus X a ray of r is , and F( x + 6r) = V Wj In ( s, (1 + 6) ) -> +°° as i=l -»oo. Proposition 3.2. the algorithm terminates at Step 3 with y T If d =0, then x solves IV Proof: The solution -y = 2 pA T 0. 0, S" 1 W S" We also obtain P = so that Thus from Proposition 3.3. 1 (3 or (2.1) of d A will satisfy the following optimality conditions: d - cFa 7 A t^M where , S' 1 d = Part I, W solves the optimization If 1 A d = (1/2) (-y T d + Pw Then If not, then let r straightforward to it is then Ar = y = d1 - d2 problem sy T r , and y T r * T r = y^d 1 - T d 2 = y 0, because is (P3) that S' 1 A T d) = (1/2) (-y d) = wT S _1 A t^M = in Step 3 has a solution with a Ar = for otherwise , d2 solve the optimization problem. and Mr = A 0. Because would not have rank d 1 and d 2 are both optimal A=An. solutions, sy T However, which is a contradiction. Proposition ray of X . 3.4. If Algorithm 3 stops = d) unique. where d 1 show t^M i.e. , cFa 1 S^W . nonzero optimal value, then that solution Proof: j^M y = In either case, . x S" p(k - and p> in Step 4, then Pw is unbounded, and d is a , 10 From Theorem Proof: Part I, and Algorithm 3 stops in Step Z Because z e However, we Az < s (y Furthermore Proposition has and full Pw is Section IV. , 1 b - sy T x have y z = y 1 T x- 1) < . d * , ) 3.2). rank a < a e [-1,1]. and a > 0, Z as defined in (3.2) of by Theorem as given In particular, let whereby z = x + ct = l/(ky). 2.1 of Then if adeZ. , also Thus A(z - x 4, for all Z the w-center of is (A - sy T )z < b - sy T x so that , syT z + Az < x I, the inner ellipsoid E IN for lies in x+adeZ Thus Parti. d x + 4.1 of Part . 1 ay x + is d=y =b-s + b - sy T x But z - x i 1 = (see (2.1) ), Ad<0 — y M and d = A 0, d* x -1. d , Thus so that would have stopped so that 0. 1 Ax. a positive scalar multiple of for otherwise the algorithm Next, observe that vT d = x + *ky Thus d is F( x a ray of + d < 0. in Step 3 (see X d) -> A . +~ Because as 6 -> A <« unbounded. Linear Convergence and Improved Optimal Objective Value Bounds The purpose of this section is to establish the the algorithm for the w-center problem: following three results regarding 1 1 Lemma 4.1. (i) (Optimal Objective Value Bounds) if y<1, then P w has F( x ) < F( xj a 2 Note that Lemma 4.1 validates the A x unique optimal solution + y + A < then F( x .08567, Y At Step 4 of the algorithm (1 , , and - y) -> ) < F( xj + .669 kY^ upper bounding procedure presented in Step 4 of the algorithm. Lemma (i) if (ii) if Lemma Improvement). At Step 6 of the algorithm, 4.2 (Local y > .08567, F(x NEW ) y < .08567, 4.3 (Linear > F( F(x NEW ) - xj + (.0033) k > .4612 kY2 F( x) Convergence). At each . . iteration, at least one of the following is true: (i) F(xNEW ) (ii) F( x > F( x) + (.0033) k _ A A ) In SectionV, - F(x NEW ) < we will show .32 (F( x ) - F( x converge to A x to zero, thus establishing superlinear , x) ), where A x is the w-center of stronger than Lemma 4.3, that the constant .32 in Lemma 4.3 a result that _ that as the iterates . is convergence. X . namely (ii) will go 1 Note Lemma Lemma that 4.3(i) is a F(x NEW ) - Lemma F( x KJF 4.1 Lemma and (ii) Lemma To prove Lemma 4.2(i). 4.1 4.3(11), and 4.2. note that if 4.2(H), .4612 ) > = - F( ) an immediate consequence of is restatement of < .08567, then from F( x 4.3 -T7gbby > -68, - x) and so F( x We - F(x NEW ) ) A F( x — Z - F( x ) = F(x NEW ) - - 1 F( x ) Lemmas thus need to prove F( x ~ * ) 4.1 - F( and ) ^ - 1 .68 = .32 . x) 4.2. We start by asserting some elementary inequalities. Proposition (Inequalities) 4.1. a) In (1 + x) < x b) In (1 + x) c) 1 + y - d) 1 + y - f(h - ln(l + »2 < x - .378 x 2 V1 + 2y -^1 +2y h)h > ^ whenever -1 < x < whenever -1 < x < VI + 8 - I x2 -4612 y2 whenever (a) follows from the fact that [h - ln(l + h) h > (b) follows from \1 + - Vl +26 substituting 2 ) / 6 9 = .08567 We now will (a) is by substituting h = decreasing in 6 . Vl +29 )/Q 2 \y2 Proof: -1. .5 for .5. > ] / whenever 0<y^ h2 h. is 0<y<9. .08567. decreasing in h for (c) follows from the fact that 0. (d) follows from (c) by . prove Lemma 4.2, followed by Lemma 4.1(i) and Lemma 4.1 (11). y 2 13 Proof of Lemma F(x NEW ) - (2.4), Proof of X x e and , let is (i) a restatement of inequality "^ > k (l + y - F( xj l+y~Vl+2y Proposition 4.1(d), statement Statement 4.2.: + 2y ) 6 = .08567. Let . > .4612 y2 Then according < y < .08567 for According (2.5). to to which proves , (ii). Lemma let z = g(x) Z be given in x be the current point, and Let 4.1. (i): where g(- ) (3.2) of y = let AT S _1 w For any . the projective transformation given in (3.3) of Part is Part Then P w I. equivalent to the problem is I, (2.3): m minimize G(z) = In (1 - yT x + y Tz V w - ) In tj t i=l (A - sy T )z + subject to = b - sy T x t > t Mz = g -yTz < and by the remarks following -F} (x) G( x It . - y - y2 / ) x that = -G(z) is (2.2) and thus suffices to (2 (1 - y)) (2.3), show - yT x and lemma that y< if By construction of . 1 y, 3.2(h) of Part I, F(x) = then G(z) > 1, from Theorem 4.1 of Part I we know m Z the w-center of . Thus for any z € Z , - Y w In > tj 4 i=l m - V Wj In = s G( x i ) , where t is the slack corresponding to z in Z . i=l We now must show that -y T z because over -y T x + y . x + - yT x + ln(l d maximizes ze E OUT Thus Proposition 2.5 of Part , and so ln(l I. for y Tz -y T z ) over any ze - yT x + > y Tz) -y- y2 / (2 (1 z € E 1N Z c E OUT , (i) of then y)) . x + To > Lemma -y- y2 4.1 . see this, note d/k maximizes T ,-y z < y T x > ln(l -y) This proves statement - -y T d/k / (2 (1 - y)) = , from 14 The proof statement of (ii) Lemma 4.1 very involved, and follows from the is following sequence of lemmas: Lemma _ s =b- Let h > 4.4. be a given parameter. A _ A x, and suppose X x e X , let satisfies x- x) T A T S^W S^AC x- ( x be the w-center of Let x = p2 ) . Then m V Wj In ^ - A, x (b, h-ln(l + hL,2 ± m V - ) Wj In h-ln(l + Proof: First we In (b, - AjX ^ - ) i=l r some k = R k (n/k / p ) We now Casel. T , W(rVk - Sj ( ) ,- (2. wTr that Id) of Part I. -w T = Also, r = k and by Proposition / p) ^ n i ^ + r wnere ' i^ i=l Then note from w X = , S _1 = n TM(x- x) = A(x- x) T Wr = 2 p 2.2 of Part (p<hVk). Then |rj|<h. Then . I, ~2 2 rf , i = 1, . . . m From Proposition ™ . Thus w 2^ i=l T w'r p> hVk | r 1 4 < p / Vk i , = prove the two cases of the Lemma. (h-ln(l+h)) Tj Wj In i=l = -S" 1 A(x- x). for if r- h^kp m m ^ Wj (3 observe that m V h) -2 i=l i=l <hVk if * < Sj r- i> (h-lnO +h)) n h 2 Tt = r'Wr r»i r « (h-lnO+h)) r? h2 - P^ . , ^ i n M ^ + 4.1(a), ln(l r i) - + ^ ) < 1, . . . , m. 1 Case 2. (P > hp/k ) Because . 1, . , . . < P | r— Vk h/p ^ w m. Thus i < ) then y[k, r { mm r, ™ = r ; Proposition 4.1(a), ln(l + i | h/p < and so again by h, (h - ln(l + h)) r- ^ Vk h/p r, rIn (1 + ryvk Vk h/p kh 2 /p 2 i (h-ln(l+h» < ) - ,„ n , r , , kh^ r^ , because i=l w Tr and r T Wr = p 2 = X (Vk h/p) w. + In (1 r t ) However by . = (Vkh/P) the concavity of the log function, X m w. £ In (1) m £ WjlnaVkh/pXl+rj) w X (1-Vkh/P))= + + rjVkh/P) In (1 4 i=l i=l (h-ln(l + h)) -5 r, , Wj In (1 +t ) ^ Thus - , kh 2 h £ ^ Vk h/p) i=l m < (1 i=l i=l < + In (1 + q) - —— (h-ln(l + - < . h)) _ r- Vk hp -2 { . i=l Lemma 4.5. Q = AT S _1 x be the current point in the w-center algorithm, Let W S" 1 A A where , = sy T A- is A in Step 4. Suppose x is the optimal solution to projective transformation given If Y< 1, and h > is ) - F( x ) by (3.3) of ( h" ^ + h)) P I, and defined as is A z=g(x), where g(- is the a _ _ _ A (z - x Q(z - x) = P 2 ) that ) . 2 + pVky + P 2 ky2 /(2(1- Y)) p < if hVk <<\ ' h2 Proof: Part P w and and y 2, a given parameter, then 'F(x defined in Step A let According to Lemma — hVkp 3.2(h) of Part A F(x I, 2 2 p ky /(2(l-Y)) pVky + + _ _ ) - F( x ) = G( x m G(z) = ln(l-y T x + y Tz) - j w i=l - In (b ; i A z) ; if P > hVk A ) - G(z ) , where 5 1 A where b = b - sy T x defined in Step 2 of the algorithm, and is Noting . that m G( x = -F( ) = - x) Wj In Sj, then j£ i=l m _ a £ G( x)- G(z) = -ln(l+y T (z- x)) + Wj In (bj - m _ a z) t £ A i=l However, because Lemma x Z the w-center of is m ~ A -Ajz) Wj ln(bj X w - (defined in (3.2) of Part thus remains then by ln i s s i to show A _ A d= (z - (h - ln(l + h)) j- h _ yUz- x))< p\ky + The vector ). maximizes -y T ( x + g, -y T ( y= -y T d/k e we must have over z) E IN = {ze R n |Mz = Z c E om ={z ^Pn if (3 > r- hVk p2 ky2 T a -y) _ x -y T ( x + d) > MhVk ' ^ that -ln(l + p- if P 2(1 because I), , ^ _ i=l i=l Let .(4.1) 4.4, m It s, i=l ( --(h-ln(l +h)) £ In v?. P 2 <l/k , = , g, (z ) < k) > -pyVk . 1. ) I, where Thus Also, - x ) T Q(z - x and so pVk < . T T y d > py /p ),andso T y d that vector that is as discussed in SectionV of Part (z- x) T Q(z- x then | Step 3 of the algorithm in z e E IN all x+ d Vk Rn Mz d d/Vk However, . because < 1/k) Thus yp from Theorem Vk < y <1 . 2.1 of Part I, This then implies -ln(l ta + y d p- 1 ) < -ln(l - Pyvk j) < instance of Proposition 2.5 of Part +py\k + p..*>i —p-^k , the last inequality being an 2(1 - y) I, with e = -yp Vk and a = y . 6 17 Lemma Y< then (3 Under 4.6. < 1 the hypothesis of ln(l + h) - 2h - 1/2 V ln(l Suppose p > hyfk F(x)-F( x) < f(y,p) f(Y,P) = - +hV 12 , hVkp in y for P>0 4.5, + pVkY + and 0<y<1- -y) Straightforward calculation if V(ln(l + h)/h) 2 F Y= + 2pVk (1 - ln(l + h)/h) (4-3) 2-pVk y is l ess A optimality of x . man m e above quantity, then f( y the expression in (4.3) is 2 - ln(l + h)/h - , which equals 1 - , p) < Thus y must be greater than or equal Next, borrowing the observation in the proof of Thus,if hh (4.2) 2(1 2-ln(l+h)/h- , • where h2 reveals that f(y, p) = , ln(l1 + In r Then from Lemma . (h-ln(l +h)) Note f(Y,P) increases if 4.5, if hVk Proof: Thus Lemma — lnd + ~r rp>hVk,then Lemma , contradicting the to the expression in < pVk < 4.5 that (4.3). 1, then greater than or equal to Vdnd +h)/h) 2 + h) 1/2 V ln(l 2(1 - ln(l + h)/h) +h)1 ~¥^] /rin(l + ln(l+h) - 1/2 -^ yil— "\[ h , h)"|2 + J f I 1 " lnd + KT h~ 1 Proof of Statement RHS of (4.2) Lemma 4.5, (ii) < 5^jk Let us set h = : Thus = -.3781 p 2 + and because k < 1 y< if Then .5. on the the expression .08567, then from Lemma 4.6 and and , F(x)-F(x) <-.3781p 2 + f(p) 4.1. then greater than .08567. is (3 Lemma of pVky + P 2 ky2 Vkyp and y < 2 P k*r / (2(1 - + - / (2(1 y)) ky2 .08567, then y)) The function . - / (2(1 y)) — concave. Thus the largest value of f(p) is Let us define . given by P f(P) < .3781 , is quadratic in so that f(p) p, is \ky = ky2 ' .7562-' (l-y) with < f(p) f( p) ky2 = 2 ky 5- — 2ky2 1.5124- V. Lemma .669ky2 for y<. 08567. 2-f 1.5124- d-y) (l-y) This completes the proof of < 4.1 Superlinear Convergence In the previous section, a linear convergence rate of Lemmas 4.4, 4.5, and 4.6. arbitrarily close to zero, we showed .32, first Proposition present convergence of the algorithm, with by choosing the value In this section we show h = that as .5 we and applying choose h > and then the linear convergence rate goes to zero in the thus showing that the algorithm We linear is limit, superlinearly convergent. some elementary facts about three particular functions. 5.1. Let f(h) = 1 -i^S Let j(0) = Ll + 6 - ln(l - 1/2 Vl+26 J /6 2 +h)l 3" for £6 > ln(l + hV for h > 8 0. (5.1) (5.2) 19 , , Let p(h) = lim h->0 Then We f(h) (h-ln(l +h)) = lim , h>0 for ^2 = j(0) .5 , ( lim p(h) = h->0 and 0->O .5 5 3) - . then prove the following three propositions, after which the proof of superlinear convergence easily follows. Proposition 5.2. For any h > "1 f ( x nfw) - F( X ) + f(h) - Vl + 2f(W Proposition y < f(h) , then (f(h))^ k\l + y- in Proposition 4.1(c), 6 f(h) for if ky2 > F(x NEW )-F( x) > Proof: Step 6 of the algorithm, at 0, 5.3. we Suppose h > as in Proposition 5.1. V1 + 2Y / from and sufficiently small _ ) Now substituting obtain the desired result. At Step 6 of the algorithm, a F(x (2.4). - F( x ) if and y< f(h) f(h), and p(h) are defined then ky2 < /f/v 2k(f(h))-2 4 P (h) , "(l-f(h)) A where x Proof: and is the optimal solution to Because Lemma 4.5 f(h) is just the that if F(x)-F( x) < If h is y < ky2 /(2(1 - y)) < (f(h)) 2 expression of the f(h), -p(h)[3 2 sufficiently small p(h) /(2(1 - Pw y)) of (4.2), we have by Lemma 4.6 then + (3VkY + is RHS 2 (3 kY2 /(2(l-7)). approximately < (f(h)) 2 /(2(1 .5 - from Proposition f(h)) is (5.4) 5.1 and approximately zero. Thus the RHS of (5.4) is quadratic and concave in p. Its maximal value occurs at Vk' P= kT2 2p(h) - (1-Y) and the maximum value of the RHS in (5.4) is therefore k^ 2k?2 4p(h) - (1-7) However y < f(h) , so we have F(x ) - F( x ) < A 4 Proposition and For h > 5.4. fM kY2 2k (« h » 2 P (h) " (l-f(h)) sufficiently small, if x the optimal solution to is P w then , i) if y ^ f(h) ii) if y < then F(x NEW ) - F( x , > (l f(h) f(h), 1- , (ii) then f(h) + f(h) - >/l + 2f(h) ) k ) k (2.4), since we know 1 + 6 - 2k(f(h)) 2 - Vl + 2f(h) 4 P (h) " (f(h))^ then from + f(h) - Vl + 2f(h) Statement < y> + ' 1 F(x) -F( x) If (l then f(h), F(x) - F(x NEW ) Proof: > ) that F(x NEW ) ^l +29 is - F( x ) (1 - f(h)) > \1 + y-V 1 + 2y )k an increasing function of 9 > follows directly by combining Propositions 5.2 and 5.3. We have that 0. if y 21 ( F(x)- F(x NEW ) F(x) 1- = F(x NEW ) - F( x + f(h)- Vl + 2f(h) 1 (f(h))' ) 1- < F(x) - F( x) -F( x) \ 1 ky 4p(h) - 2k(f(h)) (l-f(h)) \ Cancelling out ky2 and rearranging yields the desired result. Lemma 5.1. Proof: It h -» , The algorithm suffices to show that as from Propostion f(h) -» Pw for solving h -» 0, 5.1. Then note + i with 6 = 9 -> f(h), and h -> as exhibits superlinear convergence in F(x). then the e-Vi 5 RHS As of (5.5) goes to zero. that + 29 = j(9) -> as .5 9 -» , by 9 Proposition p(h) -> .5 and approaches VI. The 5.1. 1 f(h) - last — term of (5.5) by Proposition > (.5) (4(.5) - 0) = 4p(h) - 2k(f(h)) 2 /(l - is 5.1. Thus the f(h)) . As h -» entire expression (5.5) . Analysis and Computation of the Improving Direction In this section, we show positively scaled projected that the direction Newton direction. As d of Step 3 of the algorithm a byproduct of is a this result, the computation of d can be carried out without solving equations involving the matrix Q = A T S^W S" 1 A, which will typically be extremely dense. Vaidya's algorithm for the center problem direction [ 14 and performing an inexact Vaidya's algorithm when our ] corresponds line search. algorithm is to computing the Newton Thus, our algorithm specializes to implemented with a line search. ? Furthermore, this establishes that ") Vaidya's algorithm then will exhibit superlinear convergence. x be the current iterate of the algorithm, Let A and and = A- Q = AT S sy T as in Steps _1 W S'^A of F(- Then ) at From the gradient of x is and 2 (2.1a), s of the algorithm, A has full A = b- and let x, and y = Q = AT rank, so that Q is S" T F(- ) given by -Q, at i.e., x V2 is given by -y F( x , i.e., VF( x ) W AT S" T S'Hv A , nonsingular and Let F(x) be the weighted logarithmic barrier function of positive definite. (1.1). . 1 let P w given in = -y, and the Hessian = -Q. ) Thus the projected Newton direction dN is the optimal solution to maximize -y T d - (1/2) d T Qd Md subject to and the Newton direction = dN together with Lagrange multipliers Jt N is the unique solution to Qd N Mtk n - Md N Because Q = -y (6.1) = has rank n and dN = M -Q _1 y has rank + k, we can write the solution to (6.1) as Q_1 m t k n (6.2) It is where jc our aim to n = (MQ^M^MQ^y show the following 23 Lemma Then 1 Let d N be the 6.1. + yT d N > by the solution direction given (6.1) or (6.2). and + y Td N > 1 if (i) , Newton 0, and d N * , d = d N Vk (^/d N Qd N ) / the is direction of Step 3 of the algorithm. If (ii) 1 + yT d N > and d N = 0, , d = dN = then Step 3 of the algorithm, and the current iterate (iii) if 1 + yT d N = 0, unbounded, and P w Remark positive scale of the Newton is Remark dN d. . Lemma Thus shows 6.1 (i) d that is is just a positive scale of the implemented with a 2.1. Then because (3.4) of Part will be d N the Newton Wj are I . Newton and 6 the projective transformations g(x) preserve directions from direction. identical. direction d N line search replacing Steps 5 Therefore, And when This is Q Lemma x, and / y d N Qd N 6.1(i) as suggested in h(z) given by using a line search, the algorithm through with or without a line search, we one need . shows Remark [ 14 is ], and (3.3) the algorithm direction in the space precisely Vaidya's algorithm a is just Suppose the algorithm . , d, X just searching in when all because the complexity analysis of Sections rv and superlinear convergence. is Rather one need . and then compute d = d N Vk dN d that in order to solve for Relation of Algorithm to Vaidya's algorithm. 6.2. . unbounded. direction (6.1) for Pw x solves not solve a system involving the possibly-very-dense matrix only solve the equations the direction of then the optimization problem (P3) of Step 3 Simplified Computation of 6.1. is V weights carries see that Vaidya's algorithm exhibits 24 Remark An 6.3. Extension of a Theorem of Bayer and Lagarias. shown Lagarias have X sends the set of optimal solutions infinity. Then the transformed space that if one is X image of the is , to First under , performing a Lemma . one can program to the linear to the line search in the Newton hyperplane direction in the 6.1 is in fact a generalization of this result. trying to find the center of any polyhedron the algorithm of Section II) is one determines steplengths by a a positive scale of the X (bounded or Newton It states not), then method Thus, direction. if line search of the objective function, then the method corresponds projective tranformation at line search) in the space the direction generated at any iteration of the projective transformation (i.e., Bayer and a projective transformation that image of Karmarkar's algorithm (with a Z ] (unbounded) center of an unbounded to finding the Z where Z Z corresponds 1 problem of minimizing Karmarkar's potential function projectively transform the polyhedron [ the following structural equivalence between Karmarkar's algorithm for linear programming and Newton's method: over a polyhedron In to Newton's method with a line search. Another important relationship between directions generated by projective transformation methods and Newton's method can be found in Gill et We now Proposition Proof: 1 Note + y Td N = We i 6.1. that _ thus must semi-definite, y T Q -1 y(l Lemma prove If (dN from yTQ^y show and - yT Q -1 y)- rc N (6.2), by solve ) a [ 4 ]. sequence of three propositions. (6.1), then 1 + y Td N > 0. we have + y TQ- 1 that that , 6.1 al. MT (MQ- M Tr yTQ 1 _1 y< 1. Q = Q - yy T Therefore y T Q . -1 Note Thus y < 1, 1 MQ- that 1 y > 1 Q = AT < yT Q _1 - yT Q_1 y S _1 QQ-1 y W S" 1 . A is positive = yTQ~ 1 (Q-yy T )Cr 1 y = completing the proof. 25 Proposition problem Proof: 6.2. (d N If (P3) of Step 3 From Also (6.3) , = -Q^y r T Qr = whereby - From . r T yTQ-V Pw Proposition 6.3. If (6.4) is if solution. 1 + yT d N = 0, then (6.3) y = T 1 Q_1 y > 0. (6.4) ^M^MQ^M 1 )'^: = we have (1 -y T Q_1 y) Thus 0. = ° satisfies r is 0, from Mr = 0, whereby Mr = (6.3). Finally, T Qr = r 0. from and -yTr > 0, unbounded. From unbounded. (d N , 7i N Q = Q - yyT d N Qd N = d NTQd N (6.2) we ) solve (6.1), and 1 + yT d N > 0, then either d N = 0, or 0. T y dN = If , (y Td 2 N) (6.5) , obtain d N T Qd N = - y T d N d N Qd N = - y Td N - If see that then the optimization 0, 0. Because d N Qd N > we the optimization problem (P3) of Step 3 3.1, and from T y dN = 1+ and = (Q-yy T )r= Proposition Proof: (6.1) unbounded, and P w has no we have -y T r = y T Q _1 y d N Qd N > solve ) y^^M^MQ^M^MQ-V = and r is N the proof of Proposition 6.1, 1 Let 7t , d N Qd N = (y , Td 2 N) then . Since . Substituting in (6.5) yields Q we must have is positive semi-definite, T y d N = -1, or -1, then this contradicts the hypothesis that T y dN = 0, Proof of Lemma which implies from 6.1.: l+y T d N >0and (6.1) that From Proposition 1 T y dN = + yTd N > 0. 0. Thus d N =0. 6.1, d N *0. Then d = d N Vk we have / 1 V d NQ d N + yT d N > ' n = :r . N Suppose / (1 + yTd N ) , and 26 P= '\/d N Qd N / (2>[k (1 + yT d N ) ) problem well-defined (by Proposition 6.3) and all M dQ d = k, satisfy the optimality conditions for the optimization are , (P3) of Step 3. d = Thus -y = 2 0, d is MT k d - (JQ > (5 , 0, the direction of Step 3 of the algorithm. Next suppose d = 0, 71 = kn + yT d N > suppose Finally, the optimization problem 3.2, 1 + yT d N = c "X, finite c E OUT and points near c , is at x solves -M T N is , x with the property that F IN c "X problem is we , see at unbounded. x A x , may a result of this section X Theorem 2.1 of Part , , F ]N and and FOUT = answers this the I. there may involve irrational data. F OUT is such that A will converge to x II (P3) conclude that of a polyhedral system with center algorithm of Section A The main 6.2, an approximate w-center point E IN and E OUT , . unbounded and P w A x = -y and 7C Pw whether one can construct good ellipsoids = 0(1 / w). affirmative: Then Then from Proposition termination, and in fact the solution A x 0. E IN =( w/(l - w)) E OUT iterates of the natural question where dN = of the special features of the w-center Although the be 0. (P3) of Step 3 fact that there exist ellipsoids E IN that the current iterate Inner and Outer Ellipsoids One and satisfy the optimality conditions for the optimization By Proposition in Step 3. VII. 1 A FOUT c F IN not about , question in the 27 Theorem (Inner and Outer Ellipsoids at an approximate w-center point.) Let 7.1. be the current iterate of the algorithm, Then Step 4 of the algorithm. F IN = {x € Rn Mx | = if let s A x and , y be as defined in let y < .08567, the ellipsoids _1 (x- x ) T A T S g, = b - x W S _1 A(x- x ) wj < and F OUT = Sxe Rn satisfy F IN Remark 7.1. Mx | c = X c Note 1/m, and Remark 7.2. OUT FOUT to F 1N this ratio is less The number x ) is 1.75-V/-^- = less 1.75m + than A x which Let 5. is (F IN - I If x ) Thus the . w = (l/m)e, O(m). to produce y < .08567 be the optimal solution to Pw the initial value of x in the algorithm, bounded. 5 needed is is 5, of iterations of the algorithm bounded Pw + than 1.75/w + is if 5^ w F, that (FOUT - ratio of the scale of w= ] g, 12 W >f- w - _1 _1 (x- x) TaT A' c-lw S W c-l S A(x- x) < we must have . Then if x° y < .08567 after at most (1-w] / f(x)-F(x A (w) \ -0033 I iterations . This follows from Lemma 4.2(i). If F* is any A finite upper bound on the value of y < .08567 after at F( x ) produced at Step 4 of the algorithm, then most (l- w)/F*-F(x w The proof of Theorem c .0033 7.1 propositions and lemmas. is iterations. a consequence of the following sequence of 28 Proposition Mx satisfy Then 52 x e int which X, and w imply Mx e x Wj i=l have \J£ < = W = b- A 52 w s x - x) < S^AC xjWWS'U x- ( let i 1, so that s Ax = b- We first . By supposition above . i-si> V. 5 < Let g. intX ; V so that , and given, is s=b-Ax, if m < 62 X w < 52 s ( , w<w Because . Sj > , i = 1, . . . we , i > 1-5 — 1 i , so that < 1-8 sj i Next, note that m ( AT S- WS- 1 A( x- x)T 1 x) = £ i ~ ,2 j m a SJ- Si v Wi V "i J - s) , i = that S^W 1, . 4 This shows that m. s -= obtain s x- x - x) S . . s S"^ , > s 0, - s) m, we J s Furthermore, where 6<1. must show u v< ( , Rn Let x e x. . By supposition, will x € int and (x - x ) T A T S _1 =g w/(l-5) 2 Proof: Suppose 7.1. rsjY I S: ~ 1 S- V 1 W; =l 62 (1-6) 2 ^ w s iy . v. i y x e int X. 29 _ Lemma suppose x e int Let 7.1. Rn F IN = {xe xe R n F OUT = be given , and X be the w-center of x let and , that (x- x) T A T Let A X Mx | = Mx | S- = ! W g, (x (x- x) T A T g, S _1 A(x- x ) - x) T A T S _1 S^W S _1 < 52 W S w _1 A(x- x) some 6<1. for A(x- x) < wj and < + (1 5) Then F IN c X Lemma Proof of c 7.1.: in Proposition 7.1, w-center of X , F OUT w \ ^ I ~- W + A/ 2 W^~ -8 1 ^ . x be an element of F IN Let we have x e X X c F IN so that , Then by . the A x Because . same argument is as the then - x ) T AT S' l VJ S- ! A(x - x ) < (1 - w)/w for any A A where from Theorem 2.1 of Part I. s = b - Ax A _ Also, by Proposition 7.1, with x = x , we have X xe (x (7.1) , , _ A A A A x-x) TAT S- WS- A( x-x) < 6 2 w/(l-8) 2 1 1 ( Taking the square-roots of (7.1) and (7.2) < - and noting the (x- x) TA T S- 1 WS- 1 A(x- x) < N w 1 =— w - + 5*V triangle inequality for ^ w 2 for 1-8 Next, note that from the hypothesis of the lemma, every x e A s -= - 1 <8 , and so , . . . , m. Now let x e X be given, and let s -= i s = b - Ax. Then (7.3) . A i 1 X si s = - we have norms, i 7 2 >- < < 1 + 8 30 (x (l - x ) TA T + 6) S- ] rt.*VV £ w m W S^ACx - x ) = s i=l V °\) x) < (1+5) 2 A A T 1 1 T A (x- x) S- WS- A(x- 2 : (s,) ( w — <- w + - 8 -= 4w 1-8 completing the proof. Lemma 7.2. and a x if Proof: Let x If is the current iterate of the algorithm the w-center of is A A z = g(x where ) P < a hVk = a _ d = z - x .5^Jk , we x± d = g, 2 Lemmas 4.5 and 4.6 that if x) Step 3 of the algorithm and in d maximizes -yT z over T Q (z . x) < k} then -y T ( x + ((3/Vk) d) > But y = -yT d/k, so ±yT d < pyVk . Therefore, (yTd 2 ) p^k < (7.4) • Next, note that since Q = Q - yy T , where Q = A T S'^W S" ] A , we have „A a A _ a a _T _ _ _ . f a z - x ) TQ ( z - x ) = (z - x ) T Q ( z - x ) + L (z - x ) T yJ2 = p 2 + (y T d ) 2 < 1 ( A p 2 + P 2 kr from Part A (x ^ I, (7.4) Also, . _ h 2k + h2 (1 _ A _ (x- x)= (z- x)/(l+y T (z- x)) from (3.4) so that - x ) T Q (x - x kV - hky) : ) < ( p 2 + p 2 k72)/(l + y T d k(h 2 + h 2 -/2 ) (1 - hy) : 2 2 ) < (p 2 + p 2 k^)/(l - pVky) 2 w(h 2 + h 2 (l-hy) : -/ 2 ) .5 y < .08567, then _ Thus ±y T d < (-p/Vk)yT d ). .55 4, _ w. . x + (z- y < .08567 at Step a _ S"'A(x- x ) < _ given in (3.3) of Part I. Let Q be as given in A — t ~ A — = ( z - x ) T Q ( z - x ) Then substituting h = d be the direction defined Then because . | ( P if — is _ ze E 1N = (ze R n Mz -yT let and W then (x- x ) T A T S _1 obtain from Let . , g(x) Step 2 of the algorithm, and in expression (4.2) X — _ a < .55 w. of let 3 Proof of Theorem s = b- Ax . i=l ^ 1- -== - We first show : suffices to It m -= \ Wj fti 7.1. T 1 - < w. show that > s cX F IN that x e F IN Because 0. Therefore, because w w ^ i x e F IN and Let . i , = then , 1, let . . , . m, s i < = l,...,m l,i , so that s, > 0, i F 1N cX which shows that = l,...,m. Thus S: i To show X that c F OUT we , A < x of w-center .08567, then the Y obtain the conclusion that Lemma VIII. A(x - x 1 k/^-=^-+ X c FOUT satisfy must we in that X S' - x ) T A T S _1 (1+6) 2 Lemma W (x 7.1, apply ) < 52 w , 8 = V^55 where n1 / - ^f^~ < 7.2, - if Next, applying 6 = V^55 . w ; — f 1.75A/^-=^ +5"V , thus showing . Concluding Remarks Alternative Convergence Constants. the algorithm that convergence to the we Lemma 4.3 asserts that at each iterate of obtain a constant improvement of at least .0033k or a linear optimal objective value, with convergence constant .32. constants .0033 and .32 are derived in Section IV by using the value h= Lemmas h = 4.4, 4.5, and 4.6. If instead of choosing h = .5, one chooses example, then by parallelling the methodology in Section with a constant improvement of convergence constant .64. at least The choice 4, one obtains .5 The in 2, for Lemma 4.3 .0133k or a linear convergence rate with of h = .5 was fairly arbitrary. Similar results can be had by choosing a different value of h to obtain different constants for the threshold value of y and the relative sizes of FOUT and F IN in Theorem 7.1. 1 32 Stronger Convergence. Vaidya algorithm is Here, linearly convergent. that the algorithm of Section A convergence. II [ 14 has ] we have shown that his Newton direction extended his result and have shown (and Vaidya's algorithm) exhibit superlinear natural question for future study whether one can show the is algorithm here to be quadratically convergent. The behavior computed. In Step write y = and a yi x) of 4, the parameter y The value . of y( x) is _ y( x) goes to Concerning the behavior of convex, y(x) if Thus No y( x) , it is y(x) is is A x , natural to ask decreases at each iteration, a level set of y(x) iteration. x approaches zero as computed, and we can Lemma 4.3, it is the optimal solution to if the level sets of yix) Pw . are The author has demonstrated examples etc. not convex, and where y{x) increases at a particular not as well-behaved as one would hope Finite Termination. is then used to derive upper bounds, a step length, is _ where x as a function of , guaranteed improvement in the objective value. From obvious that d In Step 3 of the algorithm, the search direction y. As pointed out for. in Section VII, the solution A x to the w-center problem can have irrational components and so the algorithm will not stop after finitely algorithm many [ 1 ] . may many iterations. Even This is shown by example the In that example, the iterates of a for the w-center problem, where X € {x € 2 -1, | the problem Pw may never detect unboundedness, and so iterations. R2 if Xj x1 < 1, (Newton) direction defined by each algorithm will never stop. , iterate is , the not stop after finitely Newton method with and the unbounded, of Section 4 of Bayer w = (1/3, 1/3, 1/3) x 2 > oj is and Lagarias a line search are traced and x = starting point is never a ray of X , (1 /3, 2/3). and so the The References [ 1 Bayer, D. A., and ] J.C. Lagarias. 1987. algorithm and Newton's algorithm, Karmarkar's linear programming AT&T Bell Laboratories, Murray Hill, New Jersey. [2] Censor, Y. , and A. Lent. linear equality constraints, Optimization of 'log x' entropy over Journal of Control and Optimization 25, 1987. SIAM 921-933. Freund, R. 1988. Projective transformations for interior point methods, [ 3 ] part I: basic theory and linear programming, M.I.T. Operations Research Center working paper OR 179-88. On 4 P. Gill, W. Murray, M. Saunders, J. Tomlin, and M. Wright. 1986. projected Newton barrier methods for linear programming and an equivalence to Karmarkar's projective method, Mathematical Programming 36 183-209. [ ] Gonzaga, C.C. 1987. An algorithm for solving linear programming problems in 0(n 3 L) operations. Memorandum UCB/ERL M87/10. [ 5 ] Electronics Research Laboratory, University of California, Berkeley, California. [ 6 N. Karmarkar. 1984. ] A new programming, Combinatorica [ 7 Lagarias, J.C. 1987. ] 4, polynomial time algorithm for linear 373-395. The nonlinear geometry of linear programming ITJ. and Hilbert Geometry. AT&T Bell Projective Legendre transform coordinates Laboratories, Murray Hill, N.J. Monteiro, R.C., and I. Adler. 1987. An 0(n 3 L) primal-dual interior point algorithm for linear programming, Dept. of Industrial Engineering and Operations Research, University of California, Berkeley. [ 8 ] Monteiro, R.C., and I. Adler. 1987. An 0(n 3 L) interior point algorithm for convex quadratic programming, Dept. of Industrial Engineering and Operations Research, University of California, Berkeley. [ 9 [ 10 ] ] Renegar, method, J. for linear A polynomial time algorithm, based on Newton's programming. Mathematical Programming 40 59-94. 1988. , G. Sonnevend. 1985. An 'analytical centre' for polyhedrons and new classes of global algorithms for linear (smooth, convex) programming, preprint, Dept. of Numerical Analysis, Institute of Mathematics Eotvos University, 1088, Budapest, Muzeum Korut, 6-8. [ 11 ] [ 12 ] G. Sonnevend. 1985. A new method for solving a set of linear (convex) and its applications for identification and optimization, of Numerical Analysis, Institute of Mathematics Eotvos Dept. preprint, University, 1088, Budapest, Muzeum Korut, 6-8. inequalitites [ 13 ] M.J. Todd, and B. Burrell. 1986. An extension of Karmarkar's algorithm for linear programming using dual variables, Algorithmica 1 409-424. [ 14 ] Vaidya, P. 1987. Newton-type method Laboratories, Murray [ 15 ] Vaidya, P. A locally well-behaved potential function and for finding the center of a polytope, Hill, N.J. 1988. private communication. AT&T Bell a simple MAR 200? Date Due Lib-26-67 MIT LIBRARIES 3 9080 02237 3671