Optimization Methods Term Paper

000
001

054
055

002
003
004

056
057
058

CS481 Optimization Methods

005
006
007

2012

059
060
061

Topic 0: Ellipsoid Method

008
009
010

062
063
064

Instructor: Prof. C. V. Jawahar Team: Sumit Sidana, Amit Rai, Mohit Mundhra, Deepti Singhal, Avinash S

011
012
013

065
066
067

014
015
016

068
069
070

017
018
019
020
021
022
023
024
025
026
027

1. Introduction
The Ellipsoid Method, proposed by Khachiyan [4, 3], proves that the Linear Programming (LP) can be solved in polynomial time. The most prominent algorithm discussed so far, the Simplex Method, is not polynomial time and has an exponential
running time in the worst case. Ellipsoid method is theoretically better than simplex method, but very slow practically and not
competitive with Simplex. Nevertheless, it is a very important theoretical tool for developing polynomial time algorithms for
a large class of convex optimization problems, which are much more general than linear programming. The ellipsoid method
is an iterative algorithm used for minimizing convex optimization. The basic idea of the ellipsoid method is to convert the

071
072
073
074
075
076
077
078
079
080
081

028
029
030

optimization problem into the feasibility problem, and solve the feasibility problem by generating a sequence of ellipsoids

031
032
033

existing solutions.

085
086
087

034
035
036

1.1. Challenges in other methods

088
089
090

037
038
039
040
041
042
043
044

whose volume uniformly decreases at every step. Before going into the details of method, let us see the challenges in other

1.1.1

Simplex method is non-polynomial time algorithm

The simplex algorithm moves from one basic feasible solution to an adjacent one, each time improving the value of the objective function. However, it can take an exponentially large number of steps to terminate. To prove that the simplex algorithm is
not polynomial time, we need to come up with a class of instances which are unfortunate and exhibit an exponential sequence
of pivots for which the value of objective function is decreasing for minimization problem and increasing for maximization

045
046
047

problem.

048
049
050

Klee-Minty cube [5]. The Klee-Minty cube (named after Victor Klee and George J. Minty) is a unit cube whose corners have

051
052
053

Klee and Minty demonstrated that simplex algorithm has poor worst-case performance when initialized at one corner of

been slightly perturbed. In general, ddimensional cube can be represented as:

0 xj 1, j = 1, 2, ..., d.
1

(1)

082
083
084

091
092
093
094
095
096
097
098
099
100
101
102
103
104
105
106
107

Optimization Methods Term Paper

108
109

This d dimensional cube has 2d faces, one for each inequality, and 2d vertices, one for setting each subset of x1 , x2 , , xd to

162
163

110
111
112

one and rest to zero. The perturbed Klee-Minty cube is defined by inequalities, for some 0 < < 1/2:

164
165
166

1 x1

113
114
115
116
117
118

(2)

1 xj1 xj xj1 , j = 2, 3, , d.
Let us take the example of 2-dimensional perturbed Klee-Minty cube, as shown in figure 1. It can be verified that the

119
120
121

176
177
178

(1, 1e)

D
(e, 1 e2 )

179
180

125
126
127
128
129

B
A
(e,e 2)

130
131
132
133
134
135
136
137
138
139
140
141
142
143
144

184
185
186

along the route A B C D. So if the pivoting rule is always to move to the adjacent basic feasible solution for which
the entering variable has the least index (Blands rule), then the simplex algorithm will require 2d 1 pivoting steps before
terminating. With this pivoting rule the simplex algorithm has exponential worst-case time complexity.
Now let us see how it can be formalized. In order to put 2 in standard form, we add d slack and d surplus variables. The
LP problem is defined as

x1 r1 =
(3)

xj xj1 rj = 0, j = 2, 3, , d
xj + xj1 + sj = 1, j = 2, 3, , d

To prove the complexity of simplex method we use the following Lemmas for the proofs of the lemmas refer [6, 2].
Lemma 1.1 Every set of feasible bases of problem defined above will have all the xj s present in it and only one of rj or sj .

Lemma 1.2 If S and S are two subsets of {1, 2, , d} such that d S and d
/ S , then xSd > xSd . Furthermore, if S = S d,
=1

193
194
195

xSd .

201
202
203
204
205
206

xj , rj , sj 0, j = 1, , d

then

190
191
192

199
200

x1 + s1 = 1

xSd

187
188
189

196
197
198

max xd

150
151
152

159
160
161

x1

cost function increases strictly with each move along the path. Let the objective is to maximize x2 , note that x2 increases

147
148
149

156
157
158

181
182
183

(1, e)

Figure 1. Example for exponential time simplex algorithm

145
146

153
154
155

170
171
172
173
174
175

x2

122
123
124

167
168
169

207
208
209
210
211
212
213
214
215

Optimization Methods Term Paper

216
217

Here xS denote the bases feasible set where rj s are non zero. The value of xj in xS will be denoted by xSj . This lemma is

270
271

218
219
220

valid when S is non empty.

272
273
274

221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237

Lemma 1.3 Let the subset of {1, 2, , d} be enumerated in such a way that xSd 1 xsd2 ... xd 2d . Then the inequalities are
strict, and the bases feasible sets xSj and xSj+1 are adjacent for j = 1, 2, , 2d 1.
With this we have exhibited an exponential sequence of adjacent vertices for which the values of objective function is constantly increasing.
Theorem 1.4 For every d > 1 there is an LP with 2d equations, 3d variables, and inter coefficient with absolute value
bounded by 4, such that simplex may take 2d 1 iterations to find the optimum solution.
Proof Take = 1/4 and multiply all equations of 3 by 4, so that all coefficients are integers. Since the objective is to maximize xd , the exponentially long chain of adjacent bases feasible sets whose existence is established by 1.3 has decreasing

275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291

cost. Thus the theorem follows.

238
239
240

Thus the simplex method is not polynomial time algorithms.

292
293
294

241
242
243

1.1.2

295
296
297

244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269

Cutting plane method

In cutting plane method, it can be difficult to compute appropriate next query point and localization polyhedron grows in
complexity as algorithm progresses. These issues can be addressed in ellipsoid method.

298
299
300
301
302
303

2. Ellipsoid Method
Let us discuss some of the definitions related to the method, before discussing the actual algorithm:
Definition An n n symmetric matrix D is called positive definite if xT Dx > 0 fall nonzero vectors x Rn .
Definition A set E of vectors in Rn of the form E = E(z, D) = {x Rn |(x z)T D1 (x z) 1}, where D in an n n
n

positive definite symmetric matrix, is called an ellipsoid with center z R .

Definition A polyhedron P is full-dimensional if it has positive volume as shown in figure 2.
First we will cover the ground for ellipsoid method. The ellipsoid method can be used to decide whether a polyhedron
P = {x Rn |Ax b} is empty or not. And then we extend this concept to solve the LP problem (LPP). Following are the
assumptions made to simplify the understanding of ellipsoid method:
The polyhedron P is bounded.

304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323

P is either empty or full-dimensional.

3

Optimization Methods Term Paper

378
379

324
325
326
327
328
329
330
331
332
333
334
335
336
337

380
381
382

Figure 2. Bounded and full-dimensional polyhedron contains a ball and is contained in a ball

We can make our calculations in infinite precision (i.e. the square roots can be computed exactly in unit time).
The first assumption implies that there exists a ball E0 = E(x0 , r2 I), with volume V , that contains P . The second assumption
requires that either P is empty, or P has positive volume (i.e. V ol(P ) > v for some v > 0). Initially we assume that the

383
384
385
386
387
388
389
390
391

338
339
340

ellipsoid E0 , v, V are a priori known.

392
393
394

341
342

2.1. The Algorithm

395
396

343
344
345

This subsection discusses the algorithm for ellipsoid method. Fist we will describe intuitively how the method works: The

346
347
348

method generates a sequence Et of ellipsoids with centers xt , such that P is contained in Et . If xt P , then P is nonempty

349
350
351

one of the rows of A, and b is the corresponding entry of b). In this case P is contained in the intersection of the ellipsoid

352
353
354
355
356
357

and the method terminates. If xt

/ P , then there exists a constraint which is violated (i.e. xt satisfies a xt < b , where a is
Et and a half-space that passes through the center of ellipsoid as shown in figure 3. The new ellipsoid Et+1 is generated
which covers the half ellipsoid and whose volume is only a fraction of the volume of the previous ellipsoid Et . Repeating
this process, we either find a point in P , or we conclude that the volume of P is very small and, therefore P is empty.

397
398
399
400
401
402
403
404
405
406
407
408
409
410
411

358
359
360

412
413
414

361
362

415
416

363
364
365

417
418
419

366
367
368

420
421
422

369
370
371

423
424
425

372
373
374
375
376
377

Figure 3. An Iteration of Ellipsoid Method [1]

The formal algorithm for the ellipsoid method is as follows:

426
427
428
429
430
431

Optimization Methods Term Paper

432
433

Algorithm 1: Ellipsoid Method:

486
487

434
435
436

Input:

488
489
490

(a) A matrix A and a vector b that define the polyhedron P = {x Rn |ai x bi , i = 1, ..., m}.

437
438
439

(b) A number v such that either P is empty or V ol(P ) > v.

491
492
493

440
441
442

(c) A ball E0 = E(x0 , r2 I) with volume at most V , such that P E0 .

494
495
496

443
444
445
446
447
448
449
450

A feasible point x P if P is nonempty, or a statement that P is empty.

503
504

1. (Initialization)
2

508
509
510

2. (Main iteration)
(a) If t = t stop; P is empty.

457
458
459

511
512
513

(b) If xt P stop; P is nonempty.

460
461
462

514
515
516

(c) If xt
/ P find a violated constraint, that is, find an i such that ai xi < bi
(d) Let Ht = {x Rn |ai x ai xt } . Find an ellipsoid Et+1 = E(xt+1 , Dt+1 with

463
464
465

xt+1 = xt +

466
467
468
469
470

Dt+1

471
472
473

480
481
482

505
506
507

Let t = 2(n + 1) log(V /v); E0 = E(x0 , r I); D0 = r I; t = 0

454
455
456

477
478
479

500
501
502

Algorithm:

451
452
453

474
475
476

497
498
499

Output:

n2
= 2
n 1

1
Da
p t i ,
n + 1 ai Dt ai

2 Dt ai ai Dt
Dt
n + 1 ai Dt ai

517
518
519

(4)

(5)

(e) t = t + 1
The volume of the new ellipsoid formed is actually less than the volume of the previous ellipsoid. This fact is used by the
equations 4 and 5 while actually calculating the new ellipsoid. This proof has been covered in section 3.
(Note: the fractions
=

1
n+1 ,

2
n+1 ,

2
1
n+1 , n+1

and =

and

n2
n2 1

n2
n2 1

are called step, dilation and expansion terms respectively. They are denoted as

in rest of the document.)

520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539

483
484
485

Optimization Methods Term Paper

540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556

This section shows that in dimension n = 1, the ellipsoid method closely resembles binary search, a technique to decide
if several intervals in the real time have a non-empty intersection. In one dimension, ellipsoids can be seen as intervals.
Consider the polyhedron:
P = {x R | x 0 , x 1 , x 2 , x 3}

constructs the ellipsoid E1 that contains the interval E0 {x|x 2.5} = [0, 2.5]. The ellipsoid E1 is in the interval [0, 2.5]
with center x1 = 1.25. This ellipsoid E1 contained in P as shown in figure 4. Notice that this is same as the binary search
algorithm, that it always reduce the search interval to half.
0

559
560
561

This subsection describes how the ellipsoid method is used to solve the Linear Programming Problems. In general, the LP
problem and its dual is defined as:
P : min cT x

s.t. Ax b

D : max bT y

s.t. AT y = c, y 0.

(7)

582
583
584

By strong duality of linear programming cT x = bT y. Thus the optimal solution of LP and its dual exists if the solution

585
586
587

c T x = bT y

588
589
590
591
592
593

608
609
610

622
623
624

2.3. Ellipsoid Method for LPP

579
580
581

605
606
607

619
620
621

Figure 4. Ellipsoid Method in one dimension

577
578

602
603
604

616
617
618

E1

568
569
570

599
600
601

613
614
615

E0

565
566
567

596
597
598

611
612

562
563
564

574
575
576

(6)

Let E0 be the interval [0, 5], centered at x0 = 2.5. Since x0

/ P the algorithm chooses the violated inequality x 2 and

557
558

571
572
573

594
595

2.2. Bisection Method Analogy

satisfy following inequalities:

Ax b
AT y = c
y 0.

(8)

625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647

Optimization Methods Term Paper

648
649

Thus an LP problem can be formulated as an feasibility problem of equation 8, and can be solved using ellipsoid method for

702
703

650
651
652

deciding whether the polyhedron P = {x Rn : Ax b} is non-empty.

704
705
706

653
654
655
656
657
658
659
660
661
662
663
664

2.4. Variations of Ellipsoid Method

2.4.1

707
708
709

Sliding Objective Ellipsoid Method

Suppose we wish to the solve an optimization problem {cT x : Ax > b, x > 0}. Now we can apply the ellipsoid method
again (with a strict inequality) to the new polyhedron given by P {x Rn : cT x < cT x0 }. If this is empty, then x0 is
T

optimal. Otherwise, we have a new solution x1 P , say, with strict smaller objective function than c x0 . Now we can
reapply the ellipsoid method to the new polyhedron.

710
711
712
713
714
715
716
717
718

665
666

719
720

667
668
669

721
722
723

670
671
672

724
725
726

673
674
675

727
728
729

676
677
678

730
731
732

679
680
681

733
734
735

682
683
684

736
737
738

Figure 5. Sliding Objective Ellipsoid Method

739
740

685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701

At each iteration we add a new constraint in the direction of the vector c. All the constraints cT x < cT xt are parallel to
one another. One can show that by this procedure we can reach the optimum in polynomial running time.

2.4.2

741
742
743
744
745
746

Deepest Cut Ellipsoid Method

Although ellipsoid is a very robust method which guarantees convergence with any random starting criterion, it is quite slow
in going past each of its iterations considering the reasonable amount of computation that requires to be done. One way to
speed it up is by being greedy and cutting the volume of the ellipsoid by more than half in each iteration. The most efficient
cut(deepest) would be along the that hyperplane ax 6 b of the Polyhedron P whose condition is violated by the center of the
given ellipsoid. Such a method is called the Deepest Cut Ellipsoid Method. This is facilitated by the choosing the appropriate

747
748
749
750
751
752
753
754
755

Optimization Methods Term Paper

756
757
758
759
760

1 + n
,
:=
n+1
2 + 2n
:=
,
(n + 1)(1 + )
n2
:= 2
,
(n 1)(1 2 )
a b
:=
a Da

761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783

795
796
797
798
799
800
801
802
803
804
805
806
807
808
809

(10)
(11)
(12)

815
816
817
818
819
820

824
825
826
827
828

3.1. Starting Ellipsoid:

As a starting ellipsoid, we can use a ball centered at the vector
{0, 1}n vectors). This ball has a volume:

( 12 , 21 , ...., 21 )

and of radius

1
2 n

(which goes through all

1
V ol(E0 ) = n ( n)n V ol(Bn ),
2

(13)

2
V ol(Bn ) =
( n2 + 1)
n
2

829
830
831
832
833
834
835
836
837

where Bn denotes the ball of unit radius in n-space. Therefore, the volume of Bn is:

This can be approximated to

812
813
814

821
822
823

This section discusses all the mathematical arguments and proofs required to support ellipsoid method.

790
791
792
793
794

(9)

3. Mathematical Base of Ellipsoid Method

784
785
786
787
788
789

810
811

step, dilation and expansion parameters.

(14)

838
839
840
841
842
843

(or even 2n ). Thus

log(V ol(E0 )) = O(nlogn)

(15)

844
845
846
847
848

3.2. Termination Criterion:

We will now argue that if P is non-empty, then its volume may be determined. Assume that P is non-empty, say v0
P {0, 1}n . Our assumption that P is full dimensional implies that there exists v1 , v2 , ..., vn P {0, 1}n = S such that
the simplex v1 , v2 , ..., vn is full dimensional. The vi s may not be in P . Instead, define wi for i = 1, ..., n by:

vi
wi =

v0 + (vi v0 )

if cT vi 6 x +
otherwise

1
2

(16)

849
850
851
852
853
854
855
856
857
858
859
860
861
862
863

Optimization Methods Term Paper

864
865

where =

1
2ncm .This

866
867
868
869
870
871
872
873
874

918
919

implies that wi P as
1
1
c wi = c v0 + c (vi v0 ) 6 x +
= x +
2ncm
2
T

We have that P contains the convex set C = conv({v0 , w1 , w2 , ..., wn }) and V ol(C) is

(17)

1
n!

times the volume of the paral-

lelopiped spanned by wi v0 = i (vi v0 ) (with i {, 1} for i = 1, ..., n. This parallelopiped has the volume equal to
the product of the i (which is at least n ) times the volume of a parallelopiped with integer vertices, which is atleast 1.Thus,

920
921
922
923
924
925
926
927
928

(18)

929
930
931

878
879
880

Taking logs, we see that the number of iterations of the ellipsoid algorithm before either discovering that P is empty or that

932
933
934

881
882

a feasible point exists is at most log(V ol(E0 )) log(V ol(P )) = O(nlogn + nlogcm ). This is indeed polynomial. We thus

935
936

875
876
877

V ol(P ) > V ol(C) =

1
n!

1
2ncm

n

883
884
885

ascertain that Ellipsoid Method does indeed converge in polynomial-time.

937
938
939

886
887
888

3.3. Proof that E would encapsulate the Half Ellipsoid E H

940
941
942

889
890
891

Theorem 3.1 Let E = E(z, D) be an ellipsoid in Rn , and let a be a nonzero vector. Consider the halfspace H = {x
Rn |a x a z} and let

892
893
894
895
896
897
898
899
900
901
902

=
D

909
910
911
912
913
914
915
916
917

n2
1

n2

2 Daa D
n + 1 a Da

946
947
948

949
950
951

is symmetric and positive definite and thus E = E(

is an ellipsoid. Moreover, E T H E .
The matrix D
z , D)
T

Proof First, consider the case where z = 0,D = I and a = e1 = (1, 0, ..., 0) . So,E0 = {x R : x x 6 1} and
H0 = {x Rn : x1 > 1} as shown in figure 6. Hence,

903
904
905
906
907
908

Da
1

n + 1 a Da

z = z +

E = E(

943
944
945

952
953
954
955
956

n2
2
e1
, 2
(I
e1 eT ))
n+1 n 1
n+1 1

(19)

957
958
959
960
961
962

Re-writing this, we shall have

)
2
n
2
X
1
n

1
E0 = x Rn :
x1
+
x2 6 1
n+1
n2 i=2 i
)
(
2 


n
2
X
2x1
1
n+1
2(n + 1) 2
n n 1
2
61
x1 +
+
= xR :
x +
n2 i=1 i
n2
n
n + 1 (n + 1)2
(

n+1
n

2 

(20)

963
964
965

(21)

966
967
968
969
970
971

Optimization Methods Term Paper

972
973

1026
1027

974
975
976

1028
1029
1030

977
978
979

1031
1032
1033

980
981
982

1034
1035
1036

983
984
985

1037
1038
1039

986
987
988

1040
1041
1042

Figure 6. Case where E0 = {x Rn : xT x 6 1}

989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016

Therefore,
E0 =
Now suppose x E0

x Rn :

2(n + 1)
1
n2 1 X 2
x1 (x1 1) 6 1
xi + 2 +
2
n
n
n2
i=1

H0 .Then 0 6 x1 6 1, and so x1 (x1 1) 6 0. Also,

n2 1
n2

n
X

x21 +

i=1

Pn

i=1

(22)

x21 6 1. Hence

2(n + 1)
n2 1
1
1
+
x1 (x1 1) 6
+ 2 =1
2
2
2
n
n
n
n

(23)

1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056

which verifies that x E0 , proving that E0 H0 E0 .

Now, consider the general case and construct an affine transformation T (.) such that T (E) = E0 , T (H) = H0 and
T (E ) = E0 . The result then follows that as affine transformation preserve set inclusion, i.e. if A B Rn and T (.) is an
affine transformation, then T (A) T (B).
Given E = E(Z, D), introduce the affine transformation

1057
1058
1059
1060
1061
1062
1063
1064

T (x) = RD

1
2

(x z)

1
2

(24)

where R is a rotation matrix which rotates the unit ball so that D a is aligned with the unit vector e1 = (1, 0, ..., 0) i.e.

1017
1018
1019
1020
1021
1022

1043
1044

R R=I
1

RD 2 a =k D 2 k e1

(25)
(26)

1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079

1023
1024
1025

10

Optimization Methods Term Paper

1080
1081

1134
1135

So now, T (E) = E0 since

1082
1083
1084

x E (x z) D 1(x z) 6 1

1085
1086
1087

(x z)D
RD

1088
1089
1090
1091
1092
1093

1
2

RT D

1
2

1
(x z) E0
2

(27)

(x z) 6 1

(28)
(29)

T (x) E0 .

(30)

x E a (x z) > 0
1

1
2

(31)
(32)

1151
1152
1153
1154
1155

aT D 2 RT RD

1099
1100
1101

eT1 T (x) > 0

(33)

T (x) H0

(34)

1108
1109
1110
1111
1112
1113

which implies E H E .
1

3.4. Proof that Vol(E) < e (2(n+1)) Vol(E)

Theorem 3.2 Let E = E(z, D) be an ellipsoid in Rn , and let a be a nonzero vector. Consider the halfspace H = {x

R |a x a z} and let

1114
1115
1116

=
D

1117
1118
1119
1120
1121

1128
1129
1130
1131
1132
1133

z = z +

Da
1

n + 1 a Da

n2
2
n 1

2 Daa D
n + 1 a Da

V ol(E )
V ol(T (E ))
V ol(E0 )
=
=
V ol(E)
V ol(T (E))
V ol(E0 )

(35)

Now,
E0 = E

e1
n2
, 2
n+1 n 1

2
e1 eT1
n+1



(36)

n2
2
n 1

11

2
I
e1 eT
n+1 1

1162
1163
1164
1165
1166
1167

 12

1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184

So, introduce the affine transformation

e1
+
F (x) =
n+1

1159
1160
1161

1171
1172

Proof We have that

1156
1157
1158

1168
1169
1170

is symmetric and positive definite and thus E = E(

is an ellipsoid. Moreover, V ol(E ) < e1/(2(n+1))V ol(E) .
The matrix D
z , D)

1122
1123
1124
1125
1126
1127

(x z) > 0

Similarly, one can show T (E ) = E0 . Above, we proved that E0 H0 E0 , which is equivalent to T (E) T (H) T (E ),

1142
1143
1144

1148
1149
1150

1097
1098

1105
1106
1107

1139
1140
1141

1145
1146
1147

Similarly, T (H) = H0 since

1094
1095
1096

1102
1103
1104

1136
1137
1138

x.

(37)

1185
1186
1187

Optimization Methods Term Paper

1188
1189
1190
1191
1192
1193
1194
1195

1242
1243

It can be verified that E0 = F (E0 ). So

V ol(E0 ) =

V ol(E0 )
=
E0

1199
1200
1201
1202
1203
1204

1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241

n2
n2 1

2
I
e1 eT
n+1 1



V ol(E0 )

(38)

1244
1245
1246
1247
1248
1249

 n2 

 12
2
1
n+1

 (n1)
2
n2
n
=
2
n+1 n 1

 (n1)
1
1
2
< e (n+1) e (n2 1)

(39)
(40)
(41)

(42)

1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263

We have used the inequality 1 + a < ea (a 6= 0). Therefore,

(43)

1264
1265
1266

Thus, we have successfully proven that the volume of E will always be considerably lower than E and E will cover the

1267
1268
1269

1210
1211
1212

1216
1217
1218

n2
n2 1

= e (2(n+1))

1205
1206

1213
1214
1215

det

Hence,

1196
1197
1198

1207
1208
1209

1
V ol(E )
< e (2(n+1))
V ol(E)

half ellipsoid formed by E H.

1270
1271
1272

3.5. Complexity in Ellipsoid Method

This section shows that ellipsoidal method solves the linear programming feasibility problem with integer data in polynomial time. Consider now a polyhedron P = {x Rn |Ax b}, where A, b have integer entries with magnitude bounded by
some U , and assume that rows of A span Rn . In a further section on proofs, we show that if the polyhedron P is bounded and
either empty or full dimensional, then the ellipsoid method correctly decides whether P is empty or not in O(nlog(V /v))

(n+1)

, V = (2n)n (nU )n .

1276
1277
1278
1279
1280
1281
1282
1283

iterations. We also show in subsequent section on proofs that v and V can be chosen in terms of n and U as follows:
2

1273
1274
1275

(44)

1284
1285
1286

These estimates lead to an upper bound on the number of iterations of the ellipsoid method, which is O(n4 log(nU )). If

1287
1288
1289

v = nn (nU )n

we choose an arbitrary polyhedron P , we have assumed that such a polyhedron P is bounded and is either empty or full
dimensional. But, in practice Polyhedron can be unbounded or it also may not be full dimensional. But it is possible to
modify the inputs of the ellipsoid method, if the polyhedron P is unbounded or not full dimensional.
Let A be an m n integer matrix and let b be a vector in Rm . Let U be the largest absolute value of the entries in A and b.

12

1290
1291
1292
1293
1294
1295

Optimization Methods Term Paper

1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326

1350
1351

(a) Every extreme point of the polyhedron P = {x Rn |Ax b} satisfies

n

(nU ) xj (nU ) , j = 1, 2, ..., n.

(45)

1355
1356
1357

(b) Every extreme point of the standard form polyhedron P = {x Rn |Ax = b, x 0} satisfies
(mU )m xj (mU )m , j = 1, 2, ..., m.

(46)

Let P = {x Rn |Ax b}. We assume that A and b have integer entries, which are bounded in absolute value by U .
Let = 1/2n + 1((n + 1)u)(n+1) and P = {x Rn |Ax b e}, where e = (1, 1, ..., 1). Here if P is empty, then P
is empty and if P is non empty, then P is full dimensional. If the polyhedron P is arbitrary,then we first form the bounded
polyhedron PB , where PB is defined as :
PB = x P | p xj p (nU )n , j = 1, ..., n.

(47)

We can then perturb PB to form a new polyhedron PB, . As already noted, P is non empty if and only if it has an extreme
point, in which case PB is non empty. PB is nonempty if and only if PB, is nonempty. We can therefore apply the
ellipsoid algorithm to PB, , and decide whether P is empty or not. It is not hard to check that the number of iterations
6

is O(n log(nU )). We also need to ensure the fact that the number of arithmetic operations per iteration is polynomially

1327
1328
1329

bounded in n and logU . There are two difficulties, however. First,the computation of the new ellipsoid involves taking a

1330
1331
1332

integer can be an irrational number). Therefore, we need that if we only perform calculations in finite precision, the error

1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343

1352
1353
1354

square root. Although this might seem easy, taking square roots in a computer cannot be done exactly (the square root of an

1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383

we can make at each step of the computation will not lead to large inaccuracies in later stages of computation. Second, we

1384
1385
1386

need to show that the numbers we generate at each step of computation have polynomial size. A potential difficulty is that as

1387
1388

the numbers get multiplied, we might create numbers as large as 2U , which is exponential in logU . We can overcome these
difficulties. It has been shown that if we only use O(n3 logU ) binary digits of precision, the numbers computed during the
algorithm have polynomially bounded size and the algorithm still correctly decides whether P is empty in O(n6 log(nU ))
iterations. However, we do not cover the proofs of these claims here.

1389
1390
1391
1392
1393
1394
1395
1396
1397

1344
1345
1346

1398
1399
1400

1347
1348
1349

1401
1402
1403

13

Optimization Methods Term Paper

1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422

1458
1459

4. Problems in Ellipsoid Method

4.1. Performance of the ellipsoid method in practice
The ellipsoid method solves linear programming problems in O(n6 log(nU )) iterations. Since the number of arithmetic
operations per iteration is polynomial function of n and log U , this result in a polynomial number of arithmetic operations.
This running time compares favorably with the worst case running time of simplex method,which is exponential. However,the
ellipsoid method has not been practically successful, because it needs a small number of iterations on most practical linear
programming problems.
The behaviour of the ellipsoid method emphasizes the pitfalls in identifying polynomial time algorithms with efficient
algorithms. The difficulty arises because we insist on a universal polynomial bound for all instances of the problem. The
ellipsoid method achieves a polynomial bound for all instances. However, it typically exhibits slow convergence. There have
been several proposed improvements to accelerate the convergence of the ellipsoid method, such as the idea of deep cuts.

1423
1424
1425

However it does not seem that these modifications can fundamentally affect the speed of convergence.

1426
1427
1428

solvable from a theoretical point of view. In this sense, the ellipsoid method can be seen as a tool for classifying the

1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457

Rather than revolutionizing linear programming, the ellipsoid method has shown that linear programming is efficiently

complexity of linear programming problems. This is important, because a theoretically efficient algorithm is usually followed
by the development of practical methods.

1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488

4.2. Exponential number of constraints

For discussing the problems consider the linear programming problem of minimizing c x subject to constraints Ax b.
We assume that A is m n matrix and that the entries of A and b are integer. As we know that the number of iterations in

1489
1490
1491

entries of A and b. What is remarkable about this result is that the number of iterations is independent of the number m of

1492
1493
1494

constraints. This suggests that we may be able to solve, in time polynomial in n and log U , problems in which the number

1495
1496

the ellipsoid method is polynomial in n and log U , where n is the dimension of x, and U is a bound on the magnitude of the

problem in n would then appear to be impossible.

1497
1498
1499

5. Conclusion

1500
1501
1502

m of constraints is very large, e.g., exponential in n. If m is, for example, equal to 2n , we need (2n ) time just to input

We conclude our discussion on Ellipsoid Method with the remark that this method is not very practical but theoretically
shows that LP is a polynomial time algorithm, which was not possible to prove before this method. This material covers the
following details about ellipsoid method:

1503
1504
1505
1506
1507
1508
1509
1510
1511

The intuitive idea and the algorithm,

14

Optimization Methods Term Paper

1512
1513

The key geometric result behind the ellipsoid method,

1566
1567

1514
1515
1516

Mathematical proves required to support the method,

1568
1569
1570

1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542

Complexity Analysis,

1571
1572
1573

The ellipsoid method for the feasibility problem,

1574
1575
1576

The ellipsoid method for optimization (LP) problems.

1577
1578
1579

References

1580

[1] Mit lecture notes for introduction to mathematical programming fall 2009. http://ocw.mit.edu/courses/electrical-engineering-and
1581
1582

4
[2] Vedio

lecture

of

course

linear

programming

and

extensions

by

prof.

prabha

sharma,

iit

kanpur.

http://www.nptel.iitm.ac.in/courses/111104027/39. 2
[3] B. Aspvall and R. E. Stone. Khachiyans linear programming algorithm. Journal of Algorithms, 1(1):1 13, 1980. 1
[4] L. G. Khachiyan. A Polynomial Algorithm in Linear Programming. Soviet Mathematics Doklady, 20:191194, 1979. 1
[5] V. Klee and G. J. Minty. How good is the simplex algorithm? In O. Shisha, editor, Inequalities, volume III, pages 159175. Academic
Press, New York, 1972. 1
[6] C. H. Papadimitriou and K. Steiglitz. Combinatorial optimization: algorithms and complexity. Prentice-Hall, Inc., Upper Saddle
River, NJ, USA, 1982. 2

1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596

1543
1544
1545

1597
1598
1599

1546
1547
1548

1600
1601
1602

1549
1550

1603
1604

1551
1552
1553

1605
1606
1607

1554
1555
1556

1608
1609
1610

1557
1558
1559

1611
1612
1613

1560
1561
1562

1614
1615
1616

1563
1564
1565

1617
1618
1619

15