MATHEMATIC O OPERATION RESEARC Vo No Novembe Prtnted m U - Description

SA INTEGE PROGRAMMIN WIT A FIXE NUMBE VARIABLES W LENSTRA JR Universiteit van Amsterdam i show tha th intege linea programmin proble wit a fixe numbe o variable i polynomiall solvable Th proo depend o method fro geometr o numbers Th integer linear pr ID: 28926 Download Pdf

91K - views


SA INTEGE PROGRAMMIN WIT A FIXE NUMBE VARIABLES W LENSTRA JR Universiteit van Amsterdam i show tha th intege linea programmin proble wit a fixe numbe o variable i polynomiall solvable Th proo depend o method fro geometr o numbers Th integer linear pr

Similar presentations

Download Pdf


Download Pdf - The PPT/PDF document "MATHEMATIC O OPERATION RESEARC Vo No N..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.

Presentation on theme: "MATHEMATIC O OPERATION RESEARC Vo No Novembe Prtnted m U"— Presentation transcript:

Page 1
MATHEMATIC O OPERATION RESEARC Vo 8 No 4 Novembe 198 Prtnted m U.S.A. INTEGE PROGRAMMIN WIT A FIXE NUMBE VARIABLES W LENSTRA JR Universiteit van Amsterdam i show tha th intege linea programmin proble wit a fixe numbe o variable i polynomiall solvable Th proo depend o method fro geometr o numbers Th integer linear programming problem i formulate ä follows Le n an m b positiv integers A a m X «-matri wit integra coefficients an b e T". Th questio i t decid whethe ther exist a vecto e / satisfyin th Syste o m inequalitie Ax < b. N algorith fo th solutio o thi proble i know whic

ha runnin tim tha i bounde b a polynomia functio o th length o th data Thi lengt may fo ou purposes b define t b n · m · log(i + 2) wher a denote th maximu o th absolut value o th coefficient o A an b. Indeed n suc polynomial algorithm i likel t exist sinc th proble i questio i NP-complete [3] [12] thi pape w conside th intege linea programmin proble wit a fixe valu n. I th cas n = l i i trivia t desig a polynomia algorith fo th solutio o th problem Fo n = 2 Hirschber an Won [5 an Kanna [6 hav give polynomia algorithm i specia cases A complet treatmen o th cas n = 2 wa give b Scar [10] I wa

conjecture [5] [10 tha fo an fixe valu o « ther exist polynomia algorith fo th solutio o th intege linea programmin problem I th presen pape w prov thi conjectur b exhibitin suc a algorithm. Th degre th polynomia b whic th runnin tim o ou algorith ca b bounde i a exponentia functio o n. Ou algorith i describe i §1 Usin tool fro geometr o number [1 w sho tha th proble ca b transforme int a equivalen on havin th followin additiona property eithe th existenc o a vecto Eil" satisfyin Ax < b i obvious o i i know tha th las coordinat o an suc belong t a interva whos lengt i bounde b a constan onl

dependin o n. I th latte case th proble i reduce t a bounde numbe o lowe dimensiona problems i th origina proble eac coordinat o i require t b i {0,1} n transformaüo o th proble i neede t achiev th conditio jus stated Thi suggest tha i thi cas ou algorith i equivalen t complet enumeration W remar tha th {0,1 linea programmin proble i TV/'-complete th genera cas w nee tw auxiliar algorithm fo th constructio o th require transformation Th firs o these whic "remodels th conve se (x e R" Ax < b}, i give i §2 L Loväs observe tha m origina algorith fo thi coul mad polynomia eve fo varyin n, b

employin th polynomia solvabilit o th linea programmin proble [8] [4] I a indebte t Loväs fo permissio t describ th improve algorith i §2 *Receive Novembe 13 1981 revise M 2 1982 AMS 1980 subject classification. Primary 68C25 Secondary 90C10 OR/MS Index 1978 subject classification. Primary 62 Programming/integer/algorithms Key words. Intege programming polynomia algorithm geometr o numbers. 53 0364-765X/83/0804/0538S01.2 Copyiigh 1983 Th Institut o Managemen Science
Page 2
INTEGC PROGRAMMIN WIT I 1XE NUMBE O VARIABLE 53 Th secon auxihar algonth i a leductio proces fo w-dimensiona

lattices Suc a algonthm als du t Loväsz appeare m [9 §1] an a bne sketc i give § o th presen paper Thi algonth i polynomia eve fo varym n. I supersede th muc inferio algonth tha wa describe m a earhe Versio o thi paper § w prove, followm a Suggestio o P va Emd Boas tha th intege linea programmm proble wit a fixe valu o m is als polynomiall solvabl Thi i a immediat consequenc o ou ma result i devote t th mixed integer linear programmmg problem. Combinin ou method wit Khachiyan' result [8] [4 w sho tha thi proble i polynomiall solvabl fo an fixe valu o th numbe o intege variable Thi generahze

bot ou ma lesul an Khachiyan' theore Th algonthm presente i thi pape wer designe fo theoretica purpose only an ther ar severa modification tha migh improv thei practica performance I t b expecte tha th practica valu o ou algonthm i restncte t smal value n. i a pleasur t acknowledg m mdebtednes t P va Emd Boas no onl fo permissio t mclud §4 bu als fo suggestm th proble solve m thi pape an fo severa inspmn an stimulatin discussions Descriptio o th algorithm Le K denot th close conve se = [xfER" · Ax < b] Th questio t b decide i whethe K Z = 0 I th descriptio o th algorith tha follows w mak th

followm tw simphfym assumption abou Ä' (1 K i bounded; (2 K ha positive volume. Th firs assumpüo i justifie b th followm result whic i obtame b combmm a theore o Vo zu Gathe an Sievekm [12 wit Hadamard' determman mequaht (cf (6 below) th se K T" i nonempt i an onl i n 2 contam a vecto whos coefficient ar bounde b (n + \)n" /2 a" m absolut value wher a is ä i th mtroducüon Addin thes mequahtie t th Syste make bounded Fo th justificatio o conditio (2 w refe t §2 Unde th assumption (1 an (2) § describe ho t construc a nonsmgula endomorphis o th vecto spac R" suc tha ha a "spherical appearance Mor

precisely le | denot th Euchdea lengt m R", an pu B(p,z)= (x<=W .\x-p\ < z] fo p e R" zGR >0 th close bal wit cente p an radiu z Wit thi notation th constructe wil satisf B(p, )CrKCB(p,R) (3 fo som p e r K, wit r an R satisfym 7< ( wher c i a constan onl dependm o n. Le suc a b fixed an pu L = τϊ". Thi i a lattice i R" i e , ther exist a basi bj,b . , b o R suc tha (5
Page 3
54 H W LENSTR J ca take fo example b = r(e wit e, denotm th U Standar basi vecto o R" W cal b ,b . . , b a basis fo L i (5 holds I b\,b' . . , b' i anothe basi fo L, the ^, = 2"=i ;/ f° som « X «-matri

M = (m !/ 1<(>/< wit integra coefficient an det(Af)=±l I follow tha th positiv rea numbe |det(£, . , b„)\ (th ft, bein wntte ä colum vectors onl depend o L, an no o th choic o th basis i i calle th determmant o L, notation d(L). W ca Interpre d(L) ä th volum o th parallelepipe Σ"=ι[° 0 ' *, wher [0 1 = {zElR-0 Thi Interpretatio lead t th mequahty of Hadamard d(L) < \b\. (6 = l Th equaht sig hold i an onl i th basi b\,b . . . , b i orthogonal I i a classica theore tha L ha a basi b ,b . . . , b tha i nearl orthogona m th sens tha th followm mequaht holds fl\b,\ -d(L) (7 /= wher c i a

constan onl dependm o n, cf [l Chapte VIII] [11] I § w shal mdicat a reduction process, i.e. a algonth tha change a give basi fo L mt on satisfym (7) LFMMA Let b\,b . . . ,b be any basis for L. Then VxeR :3 eL - y\ <\(\b,\ + ··· + \bf). (8 PROO W us mductio o n, th cas n=\ (o n = 0 bein obvious Le = Σ7~ Zfe, thi i a lattic m th (n - l)-dimensiona hyperplan H = ^"=1 R*, Denot b A th distanc o b t // Clearl w hav h<\b (9 No t prov (8) le G R" W ca fin m G Z suc tha th distanc o - mfc t //i < i/z Wnt - mb = x + x wit x G / an x perpendicula t H. The \x < < \b . B th mductio hypothesi ther

exists/ L suc tha -_y,| 2 + · · · + |6„„_,| Sinc x i orthogona to/ th elemen y =y + mb o L nowsaüsfies|x-j| x -J,| + |x + ·· +|*„-i| ^„| Thi prove th lemma Notic tha th proo give a effectiv constructio o th elemen y E. L tha i asserte t exist w numbe th b, suc tha \b„\ = max{|^ :\ < < n}, the (8 imphe VxeW:3yeL:\x-y\<±Tln\b„\. (10 No assum tha b ,b . . . , b i a reduced basi fo L m th sens tha (7 holds an le L' an A hav th sam meamn ä m th proo o th lemma I i easil see tha d(L) = h · d(L'). (11 Fro (7) ( 1 an (6) apphe t L' w ge fl \b,\ < c · d(L} = c · h · d(L') -h- \b,\
Page 4

INTEGE PROGRAMMIN WIT FIXE NUMBE VARIABLE ^4 an therefore wit (9) \b„\ (12 Afte thes preparation w descnb th procedur b whic w decid whethe n Z 0 01 equivalently rK n L = 0 W assum tha b ,b . . · , b i a basi fo fo whic (7 holds numbere suc tha \b max{|Z> . l Applym (10 wit = p we fin a vecto j L wit p - y\ <,\^n\b \. I 7 the TA n L i= 0, an w ar done Suppos therefor tha y & τΚ. Then & B(p,r), (3) s \p - y\> r, an thi implie tha r <±Jn\b \. Le no H, L', h hav th sam meanm ä m th proo o th lemma W hav = L' + U C H + U = U (/ + fci ). Henc L i contame m th unio o countabl man paralle

hyperplanes whic hav successiv distance h fro eac other W ar onl mtereste m thos hyper plane tha hav a nonempt mtersectio wit τΚ, thes have b (3) als a nonempt mtersecüo wit B (p, R). Suppos tha precisel t o th hyperplane H + kb mter sec B(p,R). The w hav cleail / - l < IR/h. B (4 an (12 w hav t - l < c\c {n . Henc th numbe o value fo k tha hav t b considere i bounde b a constan onl dependm o n Whic value o k nee b considere ca easil b deduced fro a representaüo o p ä a linea combmatio o Z?, Z> ...,& w fi th valu o k the w restnc attentio t thos = 2" y b, fo whic k; an thi lead t a

intege programmm pioble wit n l variable y\->y-i-> · · · >y -i- I 1 stiaightforwar t sho tha th lengt o th dat o thi ne proble i bounde b a polynomia functio o th lengt o th origina data i th direction o § hav bee followe fo th constructio o r. Eac o th lowe dimensiona problem i treate recursively Th cas o dimen sio n l (o eve n = 0 ma serv ä a basi fo th recursion Thi fmishe ou descnptio o th algonthm observ tha m th cas tha K Z i nonempty ou algonth actuall produce a elemen K Z" 2. Th conve se K. Le K = (x e R : Ax < b], an assum tha K i bounded thi sectio w descnb a algonth tha ca b use t

verif tha K satisfie conditio (2 o §1 t leduc th numbe o variable i tha conditio i foun no t satisfied an t fin th ma use i §1 Th algonth i bette tha wha i strictl neede m §1 i th sens tha i i polynomia eve fo varym n. l a indebte t L Loväs fo pointm out t ni ho thi ca b achieved th firs stag o th algonth on attempt t construc vertice Ü ,ü . . . , v o whos conve hül i a «-Simple o positive volume B maximizm a arbitrar hnea functio o K, employin Khachiyan' algonth [8] [4] on fmd a verte u o K, unles K is empty Suppose mductively tha vertice u u , . . . , v o K hav bee foun fo whic o u . . . , v

o ar hnearl independent wit d < n. The w ca construc n d lineari independen linea functions/, . . ,/„_ o R suc tha th of-dimensiona subspac V = 2 = i R( ~ % 1 give b K={ R ./,(*) -· =/„_,(* = 0}
Page 5
54 H W LENSTRA J Agai employin Khachiyan' algorithm w maximiz eac o th linea function /, /|,/ —/2 · · · ,f -d> ~J„-d o K, unti vertc ü d+ o K is foun fo whic f](v d+ \) ^ fj(v fo some e {1,2 ...,« i/} I thi occurs the - u . . . , v v d+ ü ar linearl independent an th inductiv ste o th constructio i complcted If o th othe band n suc v d+ i foun afte eac o th 2(n d) function /, —/, . . .

,/„_ —/„_< ha bee maximized the w mus hav ^(x) ^(VQ) fo al G K an al j = l 2 . . . , n d, an therefor K C u + F I thi cas w reduc th proble t a intege programmin proble wit onl d variables follows Choose for = l 2 . . . , d, a nonzer scala multipl Wj o v - v suc tha w e Z" an denot b W th ( X ä?)-matri whos column ar th w . Notic tha W ha ran d. Employin th Hermite normal form algorith o Kanna an Bacher [7 w ca find i polynomia time a integra n X «-matri U wit det(f/ = ± l suc tha UW = (%) 1<1 i wit = i i>j, (\3) k ^ 0 fo l { ' Denot b M,, . . . , u th column o th integra matri U~ Thes for a

basi R", an als o th lattic Z" Z = Σ"= 2 M, Th subspac F o R i generate b th column o W' = U~ · (k,j), s (13 implie tha κ= (14 = i Defin r, r . . . , r e R b = Σ" \^> s (Γ$ , = [/« No suppos tha j» G Ä Z" The = 2" i J,^ w it G Z an e K implie tha ü G F B (14 thi mean tha j = r fo d < n. S i a leas on o O+ > · · · ' r IB no a integer the K n Z = 0 Suppose therefore tha r ,,...,/· ar al integral Substitutin = Σ^= Τ/" + Σ"=^+ O m ou or ig ma Syste Ax < b the se tha th proble i equivalen t a intege programmin proble wit d variable y ,, y . . . , y ä required Th

vertice v v ,,..., v o K giv ris t d + l vertice v' ,v\, . . . ,v' of th conve se i R belongin t th ne problem an O' ,V'I, . · . , v' spa a i/-dimensiona simple o positiv volume Thi mean tha fo th new d-dimensiona proble th firs stag o th algorith tha w ar describin ca b bypassed conclud th firs stag o th algorithm w ma no suppos tha fo eac = 0 l,...,« l th constructio o v d+ i successful The afte n Step w hav + l vertice u ,«,,... v o K fo whic u - u . . . , u - i> ar linearl indepen dent Th «-simple spanne b v ,v . . . , v i containe i K, an it volum equal |detM|/n wher M i th matri wit

colum vector v - v . . . , v - v Thi i positive s conditio (2 o § l i satisfied J th secon stag o th algorith w construc th coordinat transformatio neede i §1 T thi en w firs tr t fin a simple o "large volum i K. Thi i , ] don b a iterativ applicatio o th followin procedure startin fro th simple spanne b D ,Ü, . . . , v„. Th volum o tha simple i denote b vol(« «, . . ,,ü„) Construc n + l linea function g g ,,..., g : R." -> suc tha g, i constan on [v :0 < j < n, i =£ i], (15 »* fo Q<
Page 6
INTEGE PROGRAMMIN WIT FIXE NUMBE O VARIABLE 54 fo / = 0 l . . . , n Maximizin th function g -g

g -g, . . . , g„, -g o K b Khachiyan' algorith w ca decid whethe ther exis / G {0 l ...,« an a verte suc tha &(*-°,)l>! &(»,-»,) j = i (th choic o j i immaterial b (15)) Suppos tha suc a pai i, i found The w replac u b x. Thi replacemen enlarge vol(ü ,ü, . . . , v„) b a facto \g,(x - ü,)|/|g,(u - u (for ^ /) whic i mor tha 3/2 W no retur t th beginnin o th procedur ("Construc n + l linea function . . . ") ever Iteratio ste vol(t> ,«,,... u increase b a facto > 3/2 O th othe hand thi volum i bounde b th volum o K. Henc afte a polynomiall bounde nurabe o iteration w reac a Situatio i whic th

abov procedur discover tha l&(*-°,)l,) (16 fo al G K an al i,j G {0 l ...,« wit i^=j. I tha cas w le b a nonsingula endomorphis o R wit th propert tha T(U τ(υ,) . . . , T(Ü spa a regulär n-simplex Wit p = (n + 1)~'Σ/= w no clai tha B(p,r)CrK CB(p,R) fo certai positiv rea number r,R satisfyin R/r<2n 3/2 i.e. tha condition (3 an (4 o § ar satisfied wit c = 2« 3/2 Thi finishe th descriptio ou algorithm prov ou claim w writ z = r(Vj), fo 0 < j < « w writ S fo th regulä «-simple spanne b z z,,... z an w define ior c > 1 (xeR" :vol(z . . .,*,_,,*,£ + , ...,z C-VO\(Z . . . , z„)forall

G (0,1 ...,«}} Conditio (16 (fo al e K an al / =£j) mean precisel tha C. T Further i clea tha S C τΚ. Ou clai no follow fro th followin lemma LEMMA Let c > 1 With the above noiation we have B(p,r) C S C T C B(p,R)for two positive real numbers r, R satisfying \ \cV + (c + l)n if n is even, >' * \cV + (2c - 2 + l)n + (c - 2c) / n i orfd PROOF Usin a similarit transformatio w ca identif W wit th hyperplan [(>})" G R" +1 Σ"= = 1 i K" suc tha z„,z,,... z i th Standar basi o +l .Thenwehav l l n+\ f^ an " } = (r)" G M" : |r, < c fo 0 < 7 < n, an r = l . y= J a straightforwar analysi on

prove tha T i th conve hül o th se o point
Page 7
54 H W LENSTRA JR obtaine b permutin th coordinate o th poin n ~ c z + c z if n = 2m, j=\ j = m+ l n - c)z - c 2 Zj + c 2 Zj i n = 2m+l. j= l }-m+ I follow tha T C B(p,R), wher R i th distanc o p t th abov point n i even «isodd + l Further B(p,r) C S, wher i th distanc o p t (0 l/n, l/n, . . . , l/n): ~ n(n+ 1 ' Thi prove th lemma REMARKS (a T th constructio o i th abov algorith on migh rais th objectio tha nee no b give b a matri wit rational coefficients Indeed fo = 2,4,5,6,10,.. ther exist n regulä n-simple al o whos vertice hav

rationa coordinates Thi objectio ca b answere i severa ways On migh replac th regulä simple b a rationa approximatio o it o indee b an fixe «-simple wit rationa vertice an positiv volume a th cos o gettin a large valu fo c, Alternatively on migh etnbe R i M" +1 ä wa don i th proo o th lemma Finally i ca b argue tha i i no necessar tha th matri M definin b rational bu onl th Symmetri matri M definin th quadrati for (rx, rx); an thi ca easil b achieve i th abov constructio o Ί. (b Th proo tha th algorith describe i thi sectio i polynomial eve fo varyin n, i entirel straightforward W indicat

th mai points Th constructio o · · · 'fn-d m tn fr rst Sta an o g gi, . . . , g i th secon stage ca b don Gaussia elimination whic i wel know t b a polynomia algorithm cf [2 §7] follow tha Khachiyan' algorith i onl applie t problem whos length ar bounde b a polynomia functio o th lengt o th origina data Th sam applie th intege programmin proble constructe i th firs stage Furthe detail ar lef t th reader (c W discus t whic exten th valu 2« fo c i (4 i bes possible Replacin th coefficien 3/ i (16 b othe constant c > l w find usin th lemma tha fo an fixe E > 0 w ca tak + e)(« + 2« 1/ i «iseven Cl

1/ [( + e)(« n n) i n i odd on i satisfie wit a algorith tha i onl polynomia fo fixe n on ca als tak e = 0 i thi formula T achiev this on use a lis o al vertice o K t fin th simple o maxima volum insid K, an transform thi simple int a regulä one Th followin resul show tha ther i stil roo fo improvement i K C R i an close conve se satisfyin (1 an (2 the ther exist a nonsingula endomorphis r R suc tha (3 an (4 hol wit c = n. T prov this on choose a ellipsoid E insid K wit maxima volume an on choose suc tha E i a sphere. Th cas tha K i a simple show tha th valu c = n i bes possible Fo fixe n an e

> 0 ther i a polynomia algorith tha achieve c = ( + ε)« I d no kno ho wel th bes possibl valu c\ n ca b approximate b a algorith tha i polynomia fo varyin n.
Page 8
INTEGE PROGRAMMIN WIT F1XE NUMBE O VARIABLE 54 (d Th algorith describe i thi sectio applie equall wel t an clas W o compac conve bodie i R" fo whic ther exist a polynomia algorith tha maximize linea function o member K o Cff . Thi remar wil pla a importan rol i §5 I particular w ca tak fo Ctf a "solvable clas o conve bodies i th terminolog o [4 §§ an 3] Th sam remar ca b mad fo th algorith presente i §1 Th reductio

process Le n b a positiv integer an le b ,b . . . , b G R" n linearl independen vectors Pu L = 2" 1 U, ; thi i a lattic i R" I thi sectio w indicat a algorith tha transform th basi b ,b . . . , b fo L int on satisfyin (7 wit c = 2" "~ 1)/ Th algorith i take fro [9 §1] t whic w refe fo a mor detaile description recal th Gram-Schmid orthogonalizatio process Th vector b* (\ < i < n) an th rea number µ, ( < j < i < n) ar inductivel define b b* = b, - µ,/ , µ = (b, , bf}/(b* , b*), = ' wher ( , ) denote th ordinar inne produc o R" Notic tha b* i th projectio b, o th orthogona complemen o ^~ R£> an

tha 2 = Rf = Σ, R*/ fo < i < n. It follow tha &*,&* - . . , b* i a orthogona basi o R". Th followin resul i take fro [9] PROPOSITION Suppose that k,l< 07 for l < j < i < n, and for ! Then i.e., (7 holds with c = 2" (w -"/ PROOF Se [9 Propositio 1.6] explai conditio (18 w remar tha th vector b* + µ, _,£>*_ an b*_ ar th projection o b, an Z>,_ o th orthogona complemen o 2;=^fy Henc i (18 doe not hol fo som / the i doe hol fo th basi obtaine fro b lt 2i . . . , b interchangin b , an b, . chang a give basi b ,b . . · , b fo L int on satisfyin (7 w ma no iterativel appl th followin

transformations First transformation: selec / ! suc tha (18 doe no hold an inter chang ft,_, an b ; Second transformation: selec / y l < j < i < « suc tha (17 doe no hold an replac b, b b, rb , wher r i th intege nearest t µ . ca b show that independentl o th orde i whic thes transformation ar applie an independentl o th choice o / an o / andy tha ar made thi lead afte a finit numbe o Step t a basi b\,b . . . ,b satisfyin (17 an (18) The (7 satisfie ä well b th proposition Thi finishe ou sketc o th algorithm particularl efficien strateg fo choosin whic transformatio t apply an fo whic / o i

andy i describe i [9 (1.15)] I w assum th b, t hav integer coordinate the th resultin algorith i polynomial eve fo varyin n, b [9 Propositio 1.26] I follow tha th sam resul i tru i w allo th coordinate o th b, t b rational.
Page 9
54 W LENSTRA J REMARKS (a Th algorith sketche abov ca b use t fin th shortes nonzer vecto i L, i th followin way Suppos tha b ,b . . . , b i a basi fo L satisfyin (7) an le G L. The w ca writ = 2? i m,b, wit m, an fro Cramer' rul i i eas t deriv tha \m,\ < c · \x\/\b,\, fo l < i < n. I i th shortes nonzer vecto i L the |jc < \b\ fo al / s \m\ < c S b

searchin th se (27=i : fo ! w ca fin th shortes nonzer vecto i L i polynomia time fo fixe n. Fo variabl n thi proble i likel t b NP-ha.rd. (b W discus t whic exten ou valu fo c i bes possible Th abov algorith yield c = 2" "~ l)//4 W indicat a algorith tha lead t a muc bette valu fo bu th algorith i onl polynomia fo fixe n. (a w showe ho t fin th shortes nonzer vecto i L b a searc procedure B a analogou bu somewha mor complicate searc procedur w ca determin th successive minima \b\\,\b'^, . . . ,\b' o L (se [l Chapte VIII fo th definition) Her b\,b' . . . , b' L ar linearl independent an b [l

Chapte VIII Theore I p 20 an Chapte IV Theore VII p 120 the satisf ι*; < · d(L) 1=\ wher y denote Hermite' constan [l §IX.7 p 247] fo whic i i know tha y„/n fo n->oo Usin a sligh improvemen o [l Chapte V Lemm 8 p 135 w ca chang b\,b' . . . , b' int a basis b" ,b' . . . , b% fo L satisfyin \b','\ ( ',\ (K i < n) d(L) (fo n > 3) conclud that fo fixe n, th basi b ,b . . . ,b produce b th algorith indicate i thi sectio ca b use t find i polynomia time a ne basi satisfyin (7) bu no wit c = (c · n)". Her c denote som absolut positiv constant th othe hand th definitio o implie tha ther exist a

n-dimensiona lattic L suc tha \x\ > y 1/ · d(L)^ /n fo al L, = 0 cf [l Chapte I Lemm 4 21] An basi b\,b · · · , b fo suc a lattic clearl satisfie Therefor th bes possibl valu fo c satisfie c > (c' · n)"/ fo som absolut positiv constan c'. 4. A fixe numbe o constraints I thi sectio w sho tha th intege linea programmin proble wit a fixe valu o m i polynomiall solvable I wa note P va Emd Boa tha thi i a immediat consequenc o ou mai result Le n,m,A,b b ä i th introduction W hav t decid whethe ther exist e Z fo whic Ax < b. Applyin th algorithm o Kanna an Bacher [7 w ca fin a (n X «)-matri U wit

integra coefficient an determinan ± l suc tha th matri ^i = (a;) 1<1 K/< satisfie = 0 fo
Page 10
INTEGE PROGRAMMIN WIT FIXE NUMBE O VARIABLE 54 Putting = U~ w se tha th existenc o Z wit Ax < b i equivalen t th existenc o y G Z wit (AU)y < b.If n > m, the th coordinate J OT+ . . . ,y o / not occu i thes inequalities sinc (19 implie tha ^ = 0 fo 7 > m. W conclud tha th origina proble ca b reduce t a proble wit onl min{«,m variables Th latte proble is fo fixe m, polynomiall solvable b th mai resul thi paper 5. Mixe intege linea programming Th mixed integer linear programming prob- lem i

formulate ä follows Le k an m b positiv integers an n a intege satisfyin 0 < n < k. Le furthe A b a m X Ä>matri wit integra coefficients an e T". Th questio i t decid whethe ther exist a vecto = ( , . . . , x wit x, e Z fo l < / < n, x, e K fo n + l < / < k satisfyin th syste o m inequalitie Ax < b. thi sectio w indicat a algorith fo th solutio o thi proble tha i polynomia fo an fixe valu o « th numbe o intege variables Thi generalize bot th resul o § l (n = k) an th resul o Khachiya [8] [4 (n = 0) Le K' = (xGR :Ax < b}, = { (x,, . . . , x„) R : ther exis x n+ ,,..., x e M suc tha (* ,x . . .

, x ~J e K'}. Th questio i whethe K Z = 0 Makin us o th argument o Vo zu Gathe an Sievekin [12 w ma agai assum tha K', an henc K, i bounded Nex w appl th algorith o § t th compac conve se K c W. T se tha thi ca b don i suffice t sho tha w ca maximiz linea function o K, se §2 Remar (d) Bu maximizin linea function o K i equivalen l maximizing o K', linea function tha depen onl th firs n coordinate x\,x . . . , x„; an thi ca b don b Khachiyan' algorithm Th res o th algorith proceed ä before A a certai poin i th algorith w hav t decid whethe a give vectc y G R belong t τΚ. Thi ca b don b

solvin a linea programmin proble wit k n variables Thi finishe th descrip tio o th algorithm i § i ca b prove tha th mixe intege linea programmin proble i als polynomiall solvabl i th numbe o inequalitie tha involv on o mor intege variable i fixed or mor generally i th ran o th matri forme b th firs n column o A i bounded Reference [1 Cassels i. W S (1959) An Introduction to the Geometry of Numbers. Springer Berlin Secon printmg 1971 [2 Edmonds 3. (1967) System o Distinc Representative an Lmea Algebra J. Res. Nat. Bur. Standards Sect. B 71 241-245 [3 Garey M R an Johnson D S (1979) Computers

and Intractabihty, A Guide to the Theory of NP-Completeness. Freema & Co. Sa Francisco [4 Grotschel M. Loväsz L an Schnjver A (1981) Th Elhpsoi Metho an It Consequence i Combmatona Optimizatio Combmatonca l 169-197 [5 Hirschberg D S an Wong C K (1976) A PolynomiaJ-tim Algorith fo th Knapsac Proble wit Tw Variables / Assoc. Comput. Mach. 2 147-15
Page 11
54 H W LENSTRA J [6 Kannan R (1980) A Polynomia Algonth fo th Two-vanabl Intege Programmin Proble J. Assoc. Comput Mach 27 118-12 [7 _ an Bachern A (1979 Polynomia Algonthm fo Computin th Smit an Heimit Norma Form o a Intege Matrix

SIAM J. Comput. 8 499-50 [8 Khachryan L G (1979) A Polynomia Algonth m Linea Programmin Dokl. Akad Nauk SSSR 24 1093-109 (Englis translation Soviel Math Dokl. 2 (1979) 191-194 [9 Lenstra A K. Lenstra H W. Jr an Loväsz L (1982 Factorin Polynomial wit Rationa Coefficients Math. Ann. 26 515-534 [10 Scarf H E (1981) Producüo Set wit Indivisibihües—Par I Generahtie Econometnca 4 1-3 Par I Th Gas o Tw Activities ibi , 395-423 [11 Va de Waerden B L (1956 Di Reduktionstheori vo positive quadratische Forme Acta Math 9 265-309 [12 Vo zu Gathen J an Sievekmg M (1978) A Boun o Solution o Linea Intege