A Dynamic Programming Approach to Sequencing Problems Michael Held; Richard M. Karp Journal of the Society for Industrial and Applied Mathematics, Vol. 10, No. 1 (Mar. 1962), 196-210. Stable URL htp:/ links jstor-org/sii sici=0368-4245% 28196203%29 10%3A 1%3C 196%3A ADPATS%3E2.0,.CO%3B2-P Journal of the Society for Industrial and Applied Mathematics is currently published by Society for Industrial and Applied Mathematics. ‘Your use of the ISTOR archive indicates your acceptance of JSTOR’s Terms and Conditions of Use, available at hup:/www,jstororglabout/terms.hml. ISTOR’s Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www jstor.org/journals/siam.himl Each copy of any part of @ JSTOR transmission must contain the same copyright notice that appears on the sereen or printed page of such transmission. STOR is an independent not-for-profit organization dedicated to creating and preserving a digital archive of scholarly journals. For more information regarding JSTOR, please contact jstor-info@ umich edu, bupslwww jstor.org/ “Tuo Jan 27 05:05:20 20041, Sor, Ieooen, Arr Mas A DYNAMIC PROGRAMMING APPROACH TO ‘SEQUENCING PROBLEMS* MICHAEL HELDS axp RICHARD M. KARPt INTRODUCTION Many interesting and important optimization problems require the determination of a best order of performing a given set of operations. ‘This paper is concerned with the solution of three such sequencing problems: a scheduling problem involving arbitrary cost funetions, the traveling- salesman problem, and an assembly-line balancing problem. Each of these problems has a structure permitting solution by means of recursion schemes ‘of the type associated with dynamie programming. In essence, these re- ‘cursion schemes permit the problems to be treated in terms of combinations, rather than permutations, of the operations to be performed. The dynamic programming formulations are given in §1, together with a discussion of various extensions stich as the inclusion of precedence constraints. In each ease the proposed method of solution is computationally effective for problems in certain limited range. Approximate solutions to larger problems may be obtained by solving sequences of small derived problems, each having the same structure as the original one. This procedure of suc cessive approximations is developed in detail in §2 with specific reference to the traveling-salesman problem, and §3 summarizes computational ex- perience with an IBM 7090 program using the procedure. 1, PROBLEM FORMULATIO: LA A scheduling problem. Suppose we aregivena setof jobs Jo ++ Ju which are to be executed successively on a single facility. Any given job Jz is assumed to requite the services of the facility for rs units of time. With J is also associated a function e4(t), giving the cost associated with completing J, at time ¢. We assume that the facility is to be constantly in use, and that no job is to be interrupted before completion.’ With these assumptions, any given order of execution of the jobs (a schedule) may be represented by an ordering (ty is “+ #,) of the integers from one through 1, indicating that the jobs are to be executed in the order Ji, Jey *-> Ji, Given such a schedule, the termination time ti, of Ja, is Deja ry, and * Reocived by the editors August 31, 1961, + International Business Machines Corporation, New York, N. Y. "there is no advantage in violating these assumptions when the functions a(t) sre monotone nonuleereasing, representing penalties inured for deferring the eom= pletion of the jobs 196DYNAMIC PROGRAMAUNG AND SEQ! (@ PROBLEMS 197 the total cost associated with the schedule is () Dh calli) We seek an ordering for which (1) assumes its minimum value €. Several authors have considered this problem for special kinds of cost funetions. For example, McNaughton [9] considers the ease in which each x(0) is constant up to some point in time, and then increases linearly. ‘The dynamic programming algorithm which follows requires no assump- tions to be made about the cost functions. Let S = fh, kry---, kus} be a subset of {1, 2,---, n}, and denote by ty the quantity 3,.,.r,- Let C(S) be the minimum cost ineurred in executing the jobs Jays Jin 5 -** Sis)» in any order whatever, in the interval (0, ts]. For € S, let S — denote the set obtained by deleting Lfrom S. Then we may define the following recurrence relations, in which C(S), for any set 8, is expressed in terms of the values of C for subsets of Ss: (a) (W(8) = 1): C(t) = exe), for any 2 oy) (a(S) >: OCS) = mimes COS = +t) To justify this, we angue as follows: In an optimal order of execution for the jobs with indices in S(Le., one which realizes the eost C(S)) some job, call it Jz, must be executed last, and the remaining jobs (those with indices in $ — 1) must be exeeuted optimally in (0, ts Then the total cost incurred by such an ordering will be (8 — J) + ex(ts). Taking the mum over all choiees of /, we obtain (2b) Tt follows that an ordering (ij i++ 4,) is optimum if and only if, Ctl fey eos GN) = CUlity ferry teal) F eylheonnild w , ! @spsn) ‘The procedure for finding an optimum solution has two phases. In the first phase, the quantities C(S), for all $C |1, 2,--» ,n}, are computed recursively from (2); € is given by C({1, 2, --» yn}).In the second phase, (3) is used to generate an optimum ordering (iis --~ i); iy is obtained fat, and then, successively, fay fay" 9 ‘The fundamental operations required are additions and comparisons, which occur in equal numbers. ‘The number of additions in the first. phase ingiven by Etak(f) = m2 the mumber of ations inthe second phase is at most >i.k = [n(n + 1)/2] — 1. If one storage location is assigned to each number C(S), the number of locations required is 2” — 1198 MICHAEL 1ELD AND RICHARD M. KARP 12. The traveling-salesman problem. ‘The dynamic programming formulation of the scheduling problem just considered is based on the fact that the cost incurred in executing a job depends only on whieh jobs preceded it. We now diseuss the traveling-salesman problem, which ex- hibits a different kind of east dependenee. “The traveling-salesman problem is that of finding a P = (1 fy ig+++ iy) of the integers from 1 through m that mi quantity @ tig + dag Fag Foe Haan here the dag are a given set of real numbers. More accurately, since there are only (n — 1) ! possibilities to consider, the problem is to find an efficient. method for choosing a minimizing permutation. “The problem takes its name from the fact that a salesman wishing, to travel by shortest total distance from his home to each of m — 1 speci- fied cities, and then return home, could use such a method if he were given the distance aas between each pair of cities on his tour. Or, if the salesman desired the shortest total travel time, the aay would represent the indi- vidual travel times” (6) While no completely acceptable computational method exists for solving the traveling-salesman problem, several procedures have been developed for obtaining optimum or near optimum solutions. ‘These procedures, however, are usually somewhat tedious, intuitive, and difficult to program for a computer. A “state of the art” discussion of the traveling-salesman problem may be found in [1}. We proceed to give a dynamic programming formulation for this problem Given $ S {2, 3, +++, n} andl € S, let C(S, 1) denote the minimum cost of starting from city one and visiting all cities in the set S, terminating at city 1, Then (a) @(S) =: CUM, = au, for any (b) (m(8) > Dt CCS, 1) = minagas[C(S — f, m) + a. “Vo see this, suppose that, in visiting the cities in S, terminating at city city m immediately precedes city J. ‘Then, assuming that the other cities are visited in an optimum order, the eost incurred is C(8 ~ 1, m) + ant ‘Taking the minimum over all choices of m, we obtain (5b). Finally, if © denotes the minimum cost of a complete tour, including the return to city one, @ © = mimreies,..n (C(12, 3, ++ 5), + a. °Th cal to one proposed by Bellmat Tormulation, discovered independently by the authors Bh essentially MentDYNAMIC PROGRAMMING AND SEQUENCING PROBLEMS 199 A permutation (1 #, +++ i,) is optimum if, and only if, @ © = C(2,3, ++, mht) aus and, for? Sp Sn—1, (8) Clin thy o> ty total tints) = Clty tayo iy to) # aig As in the scheduling problem, a two-phase computation is used to ob- tain an optimum solution. In the first phase, the quantities C(S, 1) are ‘computed recursively from (5), and @ is computed from (6). In the second phase, (7) and (8) are used to compute an optimum permutation (iy is determined first, and then, successively, tna, trey “** 5 ts). Again, the fundamental operations employed in the computation are additions and comparisons, ‘The number of each in the first phase is given by (SE He - v( : + (x= == Yin — 292" + (m0). ‘The number of occurrences of each operation in the second phase is at most. LUE? E = [n(n — 1)/2] — 1. If one storage location is assigned to each rhumber C(S, 1), the number of locations required is Sy fent = Ee y= ee 1.3. An assembly-line balancing problem. In the scheduling and travel- ingsalesman problems, all posible orderings of the elements Gobs or eties) were allowed. Many sequencing problems involve constraints which pro- hibit certain orderings from occurring, An advantage of the dynamic programming approach is that such constraints facitate the solution, rather than hinder it, We illustrate this with reference to an assembly line balancing problem posed by Salveson (10) ‘An assembly’ Ine is required to earry out, foreach unit produced, a set of jobs denoted by iy Ja, *~+ Su. A given job J may be exceuted in re units of time, and may be assigned to any of the locations. (work stations) placed serially along the assembly line. ‘The assignment is to be made so that: (1) ta, the sum of the execution times of jobs assigned to the ith work station docs not exceed the cyte time 7"; ~ tis elled the ide time at the ith station, (2) all precedence requirements arising from the technology of the as sembly process (ive assime that these take the form "Js must precede Jw) ave satisfied, (3) the numberof work stations is minimized, The existence of precedence constraints is reflected in the notions of feasible sts and feasible assignments, A given subset © 1, 2, °" 5)200 MICHAEL HELD AND RICHARD M. KARP. is said to be feasible if there is no pair (J, Sm) such that (a) 1 ¢ 8, (b) m € S, and (c) Ji is required to precede Ju . An assignment of jobs to work stations 1, 2,-+- , q is called feasible, if, for each ¢ & q, the sct of jobs assigned to the first ¢ work station is feasible. With any feasible set S = |hi, fe, =+, Fac} 18 associated a “cost” C(S), which is the mini- mum, over all feasible assignments of the jobs with indices in S, of the quantity (r — 1) + t,, where r is the mumber of work stations needed in the assembly line and f, is the sum of the execution times of jobs as- signed to the final work stations, according to a particular assignment. ‘Then the quantity C(S) may be computed according to the following re- currence relations, which hold only for feasible sets: (a) (nS) = 1): CCM) = (9) ) (lS) >: CS) = min. (C(S—D + (C(S—D, rH], then A(t, y) = y (ii) then ate) = [ERY] ty 2 ‘The quantity A(C(S — 1), m») denotes the incremental cost associated with assigning job J, assuming that the jobs specified by the elements of S — Ihave already been assigned in an optimum manner. ‘The recurrence relations (9) are nearly identical to those used in the formulation of the scheduling problem (2), and a similar two-phase com- pputation may be used to compute C([1, 2, --- ,n}), and thenee to obtain ‘an optimum assignment of an assembly tine 1.4, Some extensions. The recurrence relations employed for the solu- tion of the assembly-line balancing problem closely resemble those de- rived in treating the scheduling and traveling,salesman problems. ‘The formulation of the Tine-balaneing problem, however, necessarily included *E [denotes “integer part of” The problem of eutting specified lengths n,m, +++ 7. of stock from a minimum number of standard reels of length 7 may be treated by precisely the same recurrence relations, with all sets S taken as feasible, For linear programming approaches to the cutting stock problem see [5] and [7DYNAMIC PROGRAMMING AND SEQUENCING PROBLEMS 201 precedence relations, which were reflected in the restricted class of subsets considered. Precedence restrictions on the order of execution of jobs to be scheduled, or on the route of a traveling salesman, ean be handled by similar means. Another variation on the three problems under consideration involves the possibility of “parallelism”: a set of cities to be visited may be dis- tributed among several salesmen; assembly operations may be divided among several assembly lines; jobs may be scheduled on any of several facilities. As an illustration of this type of situation, we shall introduce ‘some modifications into the scheduling problem, Suppose that facilities F”, F'*,---, Fare available, Any given job J, may be performed by any facility, with respective execution times a, , 7”. The cost associated with terminating J, at time ¢ on facility F is given by c'?(0). Then an optimum division of jobs among facilities, together with an optimum schedule for each facility, is obtained asiollows: (a) For each set S compute the functions C'S) by means of (2), with 2{° and c{°(2) substituted for r» and c(t), respectively, in the ith ‘computation. () Compute the minimum’, over all partitions of {1, 2, +++, n} into sets S, S®, --- , 8, of the quantity }°i_. ¢(S) ; let this minimum occur for the partition 1°, 7°, --- , 2 (©) In the optimum schedule, the jobs in 1° are performed on F. ‘The order of exceution on each facility is computed from (3), as before. ‘The numbers of operations in the algorithms under discussion can be considerably reduced when equivalences are noted among the elements to be ordered. Consider, for example, a scheduling situation in which the n jobs to be scheduled may be partitioned into q equivalence classes, such that any two jobs in the same class have the same execution time and cost function. ‘Then, any set of jobs is characterized by a vector V = (1, t2, +++, 1), specifying that the set contains » elements in the Ith equivalence class; moreover, all sets for which the vector is the same may be treated as one. Then, if each class of sets is specified by a vector V, (2) may be replaced by (10) (2 6,0, +++, 0) = 0 (b) CV) = minigteeand aise (CCV — a1) + a V-r)). Here r = (11, 72, +++, 74), where 7; is the execution time of a job in the Ith equivalence class, ey) is the cost function for jobs in the ith equivalence * 1¢is possible to give a dynamic programming algorithm for the efficient compu tation of this minimum,202 acu RDM. KARP lass, and V — ¢ is a veetor obtained by subtracting one from the ith component of V5 2. SUCCESSIVE APPROXIMATIONS: It is characteristic of the algorithms under discussion that their com- plexity, measured by numbers of arithmetic operations and storage re quirements, grows quite rapidly. They ate, however, a vast improve ment over complete enumeration, and permit the rapid direct solution of problems of moderate size.’ In this section we show how the algorithms can be combined with a method of successive approximations to provide a systematie procedure for treating large problems. This procedure yields 1 sequence of permutations, each obtained from its predecessor by the solu- tion of a derived subproblem of moderate size having the same structure as the given problem. The associated costs form a monotone nonincreasing sequence which may not converge to the optimum solution; however, com- puter experimentation has yielded excellent results in a variety of cases (cf. §8). As an illustration, we shall specify a method of successive ap- proximations for solving large traveting-salesman problems Given a permutation P = (1 é ++ t,) representing a route through cities, the cities may be partitioned into u ordered sets, each consisting of cities which occur successively in P, and maintaining the same order as in P. A.wcty traveling-salesman problem is then solved in which each ordered set is treated as.a city, and the cost of going from the set (iy ijy1 +++ fas) to (ir figs +++ fmt ta) i8 aye» If 1 i8 not too large, this derived problem may be solved by the dynamic programming algorithm of §1.2. The solu- tion implies a reordering P’ of P, with P having cost less than or equal to that of P. ‘Two types of partitions have proved especially useful. In the first type, called a local partition, each of the ordered sets but one consists of a single clement, so that the tours associated with P and P” differ locally, if at all. At the other extreme, a global partition takes the w sets as nearly equal in size as possible, so that, if changes are made, they tend to be of a global nature. Examples of these two types of partitions, and of possible reorder- ings associated with them, are shown in Figs. 1a and Ib. In these examples, n= I6and u = 8, In the solution of large problems, it has proved desirable to employ alternate phases of local and global improvement. ‘Through experience with a computer program, systematic procedures "This division into equivalence lasses is a particularly useful device in the problem of eutting stock, where i is usually necestary to cut many rolls ofthe same length. "The number of equivalence classes is then equal to the number of different Tengths to be eut. An IBM 7000 program ean solve any 18-city traveling-salesman problem in 17DYNAMIC PROGHAMDMING AND SEQUENCING PROULEMS 208 (w) LOCAL P= 0 BL BLE) 6) BN Bs 15 2 4 6 § a 4 16) we ust hy im ta ott bie us 2 40 § 0 a 4 ao (b) GLoBAr, peu 1g) 13 25) 2 4118 81 49 1) 1 TH) ‘16 sO “en Pra ort) eh 0-H 31 po“) B 7 ws estoy Fc, 1. Ezamples of local ant global improvements, for the selection of derived problems have been worked out, As part of this selection process, a technique was obtained for determining whether a given derived problem is likely to yield an improvement (i.e. cost of P< cost of P), Each derived problem is specified by aw X w cost matrix © with a row and column for each set of cities to be rearranged. We asstime that the rows and columns of this matrix are so ordered that, the slant elements C4 Cn 5 *** 5 Gut sive the cost. of the transitions which ‘occur in the current permutation P. If the cost of P' is to be strictly Tess thin the cost of P, the optimum tour for the derived traveling-salesman problem defined by C must employ transitions corresponding to off-slant clements of C. The ratio of skit elements whieh are row or eolumn minima to the total number m of slant elements has proved to be a reliable measure of the “promise” of a derived problem. At any point in the computation, matrices for which this slané ratio exceeds a critical value (which may he adjusted as the computation proceeds) are not considered further Teno such selection procedure is used, it ean be shown that the successive approximation technique yields the same final tour for any two equivalent traveling-salesman problems. Two problems with associated n Xn cost, matrices (a;,) and (bi) are said to be equivalent if there are constants ecand 8 such that, for any eyelie permutation P of the indiees 1,2, +> 5m, Dray = aD bs +B. Sinco corresponding submatrices of equivalence matrices (as,) and (bi) need not have the sume shunt ratio, the invariance of the computation tunder equivalence transformations is lost when selection is employed. However, it is possible ta derive a canonical representative for each equiva- lence class (see Appendix 1) and invariance, if desired, ean be maintained by transforming any given problem inte its eannieal form,204 ANCHARL HELD AND RICHARD M. KARP 5. THE COMPUTER PROGRAM In this section, we present some computational results obtained with an IBM 7090 program for the traveling-salesman problem. ‘This program. solves problems involving 13 or fewer cities by direct application of the dynamic programming algorithm of §1.25 larger problems are treated by means of the successive approximation technique of §2. 8.1. Specifications. In the program implementing the successive approxi- mation procedure, each derived problem is of size 13.’Phe program proceeds in alternating phases of local and global improvement. Each loeal improve- ment phase cycles through the n possible derived problems corresponding to the optimal reordering of consceutive cities, until a full eyele is con- sidered without any improvement. At this point, the slant ratio threshold is increased, and, if it doos not exceed an upper bound preseribed for the current. phase, the local improvement continues; otherwise a global im- provement phase is entered. During each global improvement. phase, at, most n + 1 derived problems are considered. The first of these is obtained by choosing a partition which maximizes the sum of the slant elements in the derived cost matrix, thus tending to make the slant ratio small. The program selects the m remaining problems by a definite but arbitrary rule which makes the sizes of the 13 sets as nearly equal as possible. ‘The derived problems are examined in turn, and for those such that the slant ratio does not exceed the current threshold, a dynamie programming caleu- lation is performed to obtain an optimum tour. As soon as a derived problem yields a solution better than that associated with its slant, the improvement is incorporated into the over-all problem, and a new phase of local improvement is initiated. If no derived problem yields an improve ‘ment, the slant-ratio threshold is increased, and, if it does not exceed an upper hound specified for the current phase, the problems are re-examined; otherwise a new phase of local improvement is initiated, and no global improvement is made at this stage of the ealculation. ‘The calculation proceeds until a point of diminishing returns is reached, fas measured by an empirieal stopping criterion involving the number of derived problems treated by the dynamic programming algorithm without sueeess in any given phase. In any event, the program is terminated after the fifth phase of global improvement. ‘The strategy of the calculation, including the choice of derived prob- the adjustment of slant-ratio thresholds, and the termination of the computation, has been suggested by computational experience. Within wide limits the successful performance of the program does not depend critically on these choices. 8.2. Results. We now present six examples which were treated by the IBM 700 computer program, In these examples, the execution time for aDYNAMIC PROGRAMMING AND SEQUENCING PRORLEMS run varied between two minutes and fifteen minutes. In each ease, the initial tour was chosen at- random, i) The 42-city problem of Dantaig, Fulk cost of an optimum tour was shown by a spe rson, and Johuson. In [4] the ial argument to be 699. Re Bo ut of tit Tear Cos of Final Po 1 325 c 2 3513, 705, 3 5207 wot 4 216 6a) 3 2m 708 (ii) A 20-city problem duc to Croes. In [3], Crocs states that the cost of fan optimum tour is 246. es Fe on of Sia Toe ot of Fat Powe 1 132 26 2 80 246 3 su 26 ii) A 48.square knight’s tour. If the distance between two squares of a. chessboard is taken as the minimum number of knight’s moves required to reach one square from the other, the problem of finding a closed knight's tour (a tour of the board by knight's moves, such that each square is visited exactly once, and terminating with a return to the initial square) may be treated as a traveling-salesman problem. Since a closed knight's tour is known to exist for an abbreviated 6 X 8 chessboard (8), the cor sponding traveling-salesman problem has an optimum solution with cost 48, Rew Bo ast of Ia Tow Cot of Hin! Pa 1 1m 56 2 1 oa a 1 5 4 12 56 (iv) A 86-square king’s tour, A closed king’s tour may be formulated as a traveling;salesman problem in the sume manner as was the Knight's tour Obviously a closed king’s tour exists for a 6 X 6 chessboard; hence, the corresponding traveling-silesman problem has an optimum solution with cost 36. Ren Cot of Pal Poe on of Fiat Powe 1 7 36 2 1 3 (v) A tour of 48 cities in the United States. Machine results coupled with visual inspection suggest: the conjecture that an optimum tour has cost 11,470. The matrix for this problem and the conjectured optimum tour are given in Fig, 2.206 MICHAEL HELD AND RICHARD M. KARPDYNAMIC PROGRAMMING AND SEQUENCING PROBLEMS uoqqoud Ono-85 “8 “ONL TILT :¥800 powe}oosey 1 wantando pasm}afv09, (2.2091 22.6 01 0G 91 1eLTSe HET EAT HIST INST ecor Fe) wow ww oo oo 8 2 9 8 F 8 et SIT oe Tae gor Ose ¥92 sor 8 80 one er MF 12z SZE_ GT 86r ay Oe oe soe 828 Soe Le One 88 ser 296 ow we 8 ss 1s 66. or 99 oe oF sue 201 Tor eee scr 18 062 sr sie ea oe ml ez oe sr ee wee sz 282 OF SET AST SOT ry OMT wzt e908 oce WI WT oz gor wat ont208 MICHABL HELD AND RICHARD M. KARP Ran Yo on of tia Tow Cot of Fiat Poe 1 0,716 11,566 2 42,288 6 3 45,400 11,084 4 51228 1168s (vi) A 25-city problem for which the cost of an optimum tour is believed. to be 1711. The cost matrix for this problem and the probable optimum tour are given in Fig, ow Cot of El Po Cont of Bat Par 1 126, m1 2 588 m. 3 5u5, mu 4. CONCLUSIONS ‘The dynamic programming approach set forth in this paper is by no means a universal key to the solution af sequencing problems. However, it can be employed in a number of important situations, and, when applic- able, offers several advantages. For small problems, optimum solutions ean be obtained with only a small, and precisely predictable, amount of computation. Larger problems lend themselves to a successive-approxima- tion technique which, although lacking rigorous justification, has worked out remarkably well in practice.’ Finally, these techniques can be mecha- nized completely by rather simple computer programs. APPENDIX 1 REDUCTION OF A MATRIX TO CANONICAL FORM In §2, two nity travelingsalesman problems defined by matrices (a,,) and (,) were said to be equivalent if there existed constants a and @ such that, for any tour Pr, Xray = abs + 8. In this appendix, we characterize this equivalence directly in terms of the defining matrices, and show how to obtain a canonical representative of each equivalence class, In the discussion that follows, it is convenient to define four special classes of matrices: (e,;) is a constant-row matrix (constant-colwmn matrix) if all its elements are zero, except for a single row (column), all of whose off-diagonal elements are equal (the diagonal element is arbitrary) ; (¢)) is a potential matric’ if for every triple i,j, ky €4) + ew = em (ey) isan * An TBM 7000 program applying the successive approximation technique to the single-facility scheduling problem of §1.1 has met with success comparable to that reported above for the traveling salesman problevn " Any potential matrix defines « traveling-salesinan problem for whieh the cost of any closed tour is zero,DYNAMIC PROGRAMM (G AND SEQUENCING PROBLEMS 209 ‘mage matrix if its diagonal elements are equal to zero, and each of its row and column sums is 2070 We now define a homomorphie mapping 1 of the additive group @ of n X n (real) matrices onto the additive group 6 of nm X n (veal) image matrices. Given (di) © @ we define the mapping in two steps: ay i, = nds — Linda — Dba des 2) des = diy + Lbs aulndudse — 8 — By] + atnss - Here, dim is the Kronecker delta and ay, a, > a4. are determined: uniquely from a system of n + 1 linear equations” expressing the condi- tions that dic = 0 for all i and Df.s Dj di; = 0. It is easily seen that (a,) belongs to G. Tt can be shown that the kernel W of the map is the normal subgroup of @ generated by the set of constant-row, constant-eolumn, and potential matrices. It follows, by the law of homomorphism for groups [11], that two matrices have the same image under the mapping 7 if and only if each can be derived from the other by the addition of coustant row, constant column, and potential matrices. ‘The canonical form of a matrix is its image under the mapping T. Two matrices are said to be equivalent if each can be derived from the other using only the following operations: (a) addition of eonstant-row and con stant-column matrices, (b) addition of potential matrices, and (c) multi- plication of a matrix by a sealar constant. ‘Thus, two matrices are equi valent if and only if their eanonieal forms differ by a constant scalar multipliative factor nally it ean be shown that two n Xn traveling-salesman problems are equivalent if and only if their defining matrices are equivalent, so that an equivalence class of traveling-salesman problems is characterized by the canonical form of any one of the associated matrices Acknowledgments. ‘The authors are particularly indebted to Mr. R. Shareshian of IBM’s Mathematics and Applications Department, who was largely responsible for the preparation of the traveling-salesman eom- puter program mentioned here. Acknowledgments are due as well to Dr. A, Bomberault of the same department for several stimulating discussion particularly with regard to the canonical representation given in Ap- pendis I. REFER] 1. R.L, Aexore (ed.), Progress In Operations Research, Vole 1061 , Wiley, New York, ‘W°This aystem will he nonsingular unless m = 2210 MICHAEL HELD AND RICHARD MM. KARP 2, R, BRuiMan, Combinatorial processes and dynamic programming, Proceedings of the ‘Tenth Symposium in Applied Mathematics of the American Mathe- ‘matical Society, 1960; also see his Inter paper Dynamic programing treat- ‘ment of the traveling saleeman problem, J. Assoe. Comput. Mach., 9 (1062). pp. 61-83. 3. G. A. Cuons, A method for solving traveling-salesman problems, Operations Res, 6 (1958), pp. 701-812. 4. G.B. Dantag, D. R, Fonenson, axp 8. M. Jounson, Solution of a large-scale traseling ealeeman problem, Operations Res., 2 (1954), pp- 398-110; also see these authors later paper On a linear programming, combinatorial ap- roach tothe traveling saleeman problem, Operations Res, 7 (1950), pp. 58-65. 5. K, E1snaann, The trim problem, Management Sei. $ (1957), pp. 279-284, 6. M.M. Foon, The traveling-salceman problem, Operations Res., 4 (1980), pp. 61-75. 7. P.-C, Gutatons, ano R. E. Gonony, A linear programming approach tothe cutting stock problem, Operations Res, 9 (1061), pp. 849-850. 8. M, Knarnestns, Mathematical recreations, Dover, New York, 1953, Chapter 1 9. R. McNavontox, Scheduling with deadlines and lose functions, Management Sei, 6 (1950), pp. 1-12. 10. M. E. SaLynson, The ascombly tine Balancing problem, Journal of Industriel ngineering, 6 (1958), pp. 18-25, 11, B.L. van baw WaRRDEs, Modern Algebra, Vol. I, Ungar, New York, 1950.