Advanced Linear Programming Organisation
Advanced Linear Programming Organisation
Advanced Linear Programming Organisation
Organisation
Lecturers:
Leen Stougie,
Vrije Universiteit and CWI in Amsterdam
E-mail: l.stougie@vu.nl or stougie@cwi.nl
Marjan van den Akker
Universiteit Utrecht
marjan@cs.uu.nl
Advanced Linear Programming suggests that there is also something like Basic Linear Programming. And indeed, I assume everyone has had a course or part of a course on Basic Linear Programming. Usually LP is taught in the bachelors as the first subject in a course on Optimization.
Most importantly I assume that everyone has had a course in Linear Algebra, at least covering linear (sub)spaces, linear independence, bases of linear spaces, linear transformations, matrix
theory including ranks, inverse, solutions of linear systems (the Gaussian elimination method), etc.
In this course I will refresh very briefly your memory on LP. I expect everyone to have access to the
book
D. Bertsimas and J.N. Tsitsiklis, Introduction to Linear Optimization, Athena Scientific, Belmont,
Massachusetts, 1997. (ISBN 1-886529-19-1)
The book also allows you to refresh your memory, since it is a complete course on LP, and I will
skip or go very briefly over parts that I suppose you to know already. Ill try to indicate as clearly
as possible which parts of the book I assume to be known at the end of the course. Our ambition
is to cover almost the whole book, indeed relying on your knowledge of basic LP. Please feel free
to ask me questions during the lectures if anything remains unclear.
Detailed week-to-week information can be found on the website http://personal.vu.nl/l.stougie
The website also includes the lecture notes I made for myself for preparing the classes. In the lecture notes you also find the Exercises for the week. Answers to the exercises will become available
on the website. If an answer is missing, please hand in the exercise, which I will check and correct
and make it available at the website.
Last organisational detail: The exam is a written exam. There will be a re-exam, but there will
NOT be a third exam. So plan your agendas very well and make sure you are present at the exam.
Time and place will be announced through my own website and (hopefully) through the website
of Mastermath. Examples of exams are at the website.
Introduction to LP
(1)
cT x
Ax b,
(2)
or as the standard form to be used in the simplex method for solving LPs
min cT x
s.t. Ax = b, x 0.
(3)
For reasons that are obscure to me, the authors of the book have chosen for minimization as the
standard optimization criterion. I will follow this unless it bothers me too much.
The first system is usually the one that is obtained after modelling the verbal description of the
problem. It is easy to see that any such system can be transformed into an equivalent system with
only (or only ) constraints and only non-negative decision variables, or, by using slack- or
surplus variables, into a system with only = constraints and only non-negative decision variables,
obtaining the standard LP formulation.
Chapter 1 of the book gives an abundance of examples, some of which really show how LP can
capture a large variety of optimization problems. Specifically, the example in Section 1.3 shows
how minimizing a piecewise linear convex function can be solved using LP, with problems concerning absolute values as a specific example. Some of the problems in Chapter 1 model as integer
LPs, a subject Marjan will cover in her part of this course. Please read these modelling parts of
the chapter.
A basic LP-course is heavily based on Linear Algebra. However, the insights that lead to the
solution of LP-problems and to a deeper understanding of duality are based on geometry. The
book puts a lot of emphasis on the geometrical interpretation of LP, and I will use it in this course
to show how geometric and algebraic interpretations coincide, re-deriving known theorems about
LP from another theory.
x2
10
3x1 +2x2 =18
8
x1 =4
2x2 =12
10
x1
x2
10
8
(2,6)
Z=36=3x1 +5x2
Z=30=3x1 +5x2
10
x1
1.1
Graphical solution
As a step-up let us first have a look at a simple example that allows a graphical solution. The
following problem is extracted from Hillier and Lieberman, Introduction to Operations Research,
McGraw Hill, New York. I skip the verbal description and the modelling and switch immediately
to the LP-formulation.
maximize 3x1
s.t.
x1
3x1
x1 0,
+ 5x2
2x2
+ 2x2
4
12
18
(4)
x2 0.
So for once we use a maximization problem here. We have two decision variables, which allows a
graphical description of the problem:
The feasible set of an LP in the IRn , i.e., an LP with n decision variables, is bounded by (n 1)dimensional hyperplanes. A vertex of the feasible set is the intersection of n hyperplanes of
3
1.2
Geometric interpretation
P
Definition 1.3 Let x1 , . . . , xk IRn and 1 , . . . , k IR+ {0} such that ki=1 i = 1 then the
Pk
vector i=1 i xi is called a convex combination of the vectors x1 , . . . , xk . The set of all convex
combinations of x1 , . . . , xk is called the convex hull of x1 , . . . , xk .
4
constraint is tight or active or supporting in a point of P if it is satisfied with equality in that point.
In LP a point that represents a feasible solution in which n linearly independent constraints are
tight (active or supporting) is called a basic feasible solution.
Theorem 1.2 (2.3 in [B& T]) For a non-empty polyhedron P , the following three statements
are equivalent
(a) x P is an extreme point of P ;
(b) x is a basic feasible solution;
(c) x is a vertex.
Proof. (a)(b) Suppose x is not bfs. Let I = {i|aTi x = bi } be any set of linearly independent
constraints active in x . Since |I| < n the vectors ai , i I lie in a subspace of IRn , i.e., there
is at least one vector d orthogonal aTi d = 0 to all ai , i I. Consider vectors y = x + d and
z = x d, for > 0 sufficiently small. y P because
aTi y = aTi x + aTi d = bi + 0 = bi , i I,
and since for i
/ I, aTi x > bi , whence for small enough aTi y bi . Similarly z P . But
x = 12 (y + z) hence not an extreme point.
(b)(c) One shows that the choice c =
iI
(c)(a) Take any c for which x is a unique optimal point. For any two other points y, z P we
therefore have cT y > cT x and cT z > cT x . Hence, we necessarily have cT (y + (1 )z) > cT x ,
concluding that x cannot be written as a convex combination of any two other points in P .
2
Two remarks: Firstly, two different basic feasible solutions are said to be adjacent on the polyhedron P if they share n1 out of the n hyperplanes or active constraints that define each of them.
Secondly, the last theorem implies that there are a finite number of extreme points in any polyhedron P , since P is the intersection of a finite number m of halfspaces.
One of the main results of Chapter 2 concerns the existence of an extreme point.
Theorem 1.3 (2.6 in [B& T]) Given non-empty polyhedron P = {x IRn |aTi x bi , i =
1, . . . , m}, the following statements are equivalent:
a) P has at least one extreme point;
b) P does not contain a line; i.e., x P , d 6= 0 IRn , IR: x + d P ;
c) There exist n vectors out of the family of a1 , . . . , am which are linearly independent.
Proof. a) c) is a direct consequence of the last theorem. I will prove b) a) and leave the
proof of c) b) to be read by yourself.
Let x P and I(x) = {i|aTi x = bi }. If {ai |i I(x)} contains n linearly independent ai s, then x
is a bfs and hence an extreme point. Otherwise, as before, they span a subspace of the IRn and
therefore there exists a vector d such that aTi d = 0, i I(x). Consider y = x + d. Clearly
aTi y = aTi x = bi i I(x). Thus I(x) I(y). Since P does not contain any line it must be
that if we decrease or increase long enough, then some other constraint than those in I(x) will
become active, say this happens at and aTj (x + d) = bj , for some j
/ I(x). Since aTj x > bj ,
T
T
T
T
aj d = bj aj x 6= 0. Hence, because aj d 6= 0 and ai d = 0, i I(x), we have that aj is
linearly independent of {ai |i I(x)}. This procedure is repeated until we arrive at a set of n
linearly independent active constraints, corresponding to a bfs, i.e., an extreme point of P .
2
Another result that most of you have always used but probably never seen proven rigorously.
Theorem 1.4 (2.7 in [B& T]) Given an LP-problem, the feasible polyhedron of which contains
at least one extreme point, and for which an optimal solution exists, there is always an extreme
point that is an optimal solution.
Proof. Let Q = {x IRn |Ax b, cT x = v} be the set of optimal solutions, having objective
value v. Q contains no line, since it is inside P and P has an extreme point, hence contains no
line. Thus Q also has an extreme point, say x . We show by contradiction that x is also extreme
point of P . Suppose not, then y, z P, [0, 1]: x = y + (1 )z. cT y v and cT z v
and cT x = v implies that cT y = cT z = v y, z Q and therefore x is not extreme point of
Q.
2
A slightly stronger result is proved along the same lines as the proof of Theorem 2.6 in the book.
Theorem 1.5 (2.8 in [B& T]) Given an LP-problem, the feasible polyhedron of which contains
at least one extreme point, then either the solution value is unbounded or there is an extreme
point that is optimal.
In my discussions of Chapters 1 and 2 I made a selection of what to treat here in class and what
to leave to yourself for study. For those who did not have a course on complexity theory I advise
to read Section 1.6 of [B& T], that shows you how to measure the running time of an algorithm
and that making a distinction between so-called polynomial running time and exponential running
time is relevant. Complexity theory is a beautiful theory, essential in the study of combinatorial
problems. In this book it only plays an essential role in Chapters 8 and 9.
There are intriguing open problems in the study of polyhedra in relation to complexity theory.
You may skip Sections 2.7 and 2.8 in the book, since the material is covered again in Chapter
4. I do advise you to study 2.4 on degeneracy.
Exercises of Week 1
1.5, 1.15, 2.6, 2.15.
Next
week
Chapter