Optimisation and Operations Research

Transcription

1 Optimisation and Operations Research Lecture 9: Duality and Complementary Slackness Matthew Roughan Lecture_notes/OORII/ School of Mathematical Sciences, University of Adelaide July 12, 2017

2 Section 1 Duality July 12, / 30

3 Problem Recap We will start with a problem in standard equality form. We call this the primal LP max z = c T x + z 0 such that Ax = b and x 0 July 12, / 30

4 Dual Primal (P) max z = c T x + z 0 such that Ax = b and x 0 Consider a new LP called the dual problem (D) min w = b T y + z 0 such that A T y c and y free July 12, / 30

5 Origin of the dual Suppose there is an optimal solution x, z to the primal LP max z = c T x + z 0 such that Ax = b and x 0 We might have obtained this via Simplex Initial Tableau Final Tableau A b c T z 0 Simplex = Â ˆb ĉ T ẑ 0 July 12, / 30

6 Origin of the dual The final tableau comes from pivots, so there are some numbers y 1,..., y m such that ĉ j = ẑ 0 = m m y i a ij c j, y i b i + z 0. j = 1,..., n At the end of Simplex ĉ j 0 so m yi a ij c j for j = 1,..., n. So let s consider any variables y 1,..., y m which satisfy m y i a ij c j, for j = 1,..., n. We could look for an optimisation to find the y July 12, / 30

7 Origin of the dual Let s build a new objective function with m w = y i b i + z 0. From the Primal, b i = n j=1 a ijx j so m n w = y i a ij x j + z 0 = j=1 ( n m ) y i a ij x j + z 0 j=1 n c j x j + z 0 j=1 Hence w z, but also the above is true for any feasible x, so w ẑ 0, the optimal value of the primal. But we defined w so that w(y ) = ẑ 0, so ẑ 0 is the minimum of w. July 12, / 30

8 The Dual of an LP (Summary) In the dual, variables become constraints, and visa versa (P) max z = n j=1 a ijx j = b i, n c j x j + z 0 j=1 i = 1,..., m m (D) min w = y i b i + z 0 y i free, i = 1,..., m x j 0, j = 1,..., n or max z = c T x + z 0, Ax = b, x 0. m y ia ij c j, j = 1,..., n or min w = y T b + z 0, y T A c T, y free. Note how the lines are paired up July 12, / 30

9 Dual Example Example (Dual) (P) max z = 3x 1 x 2 + x 3 s.t. x 1 + 2x 2 = 5 x 1 x 2 + x 3 = 7 x 2 + 6x 3 = 11 x 1 0, x 2 0, x 3 0. (D) min w = 5y 1 + 7y y 3 s.t. y 1 + y 2 3 2y 1 y 2 + y 3 1 y 2 + 6y 3 1 y 1, y 2 and y 3 are all free variables. July 12, / 30

10 Why do we call it the dual? Theorem The Dual of the Dual is the Primal. Proof: The dual is min w = b T y + z 0 such that A T y c and y free We convert to standard equality form by multiplying the objective by -1 (to make it a max) multiplying the constraints by -1 adding slack variables replacing free variables y i with non-negative variables y + i 0 and y i 0 such that y i = y + i y i, July 12, / 30

11 Why do we call it the dual? Proof: (continued) The dual, written in standard equality form is (D ) max w = b T y b T y + z 0 such that A T y A T y + + s = c and y +, y, s 0 Now we can take the dual of (D ) which we denote (DD) by creating a variable x i for each constraint, and a constraint for each variable y j and s j. min u = c T x z 0 such that Ax b T from y and Ax b T from y + and x 0 from s July 12, / 30

12 Why do we call it the dual? Proof: (continued) We have (DD) Note that min u = c T x z 0 such that Ax b T from y and Ax b T from y + and x 0 from s we can multiply the objective by -1 (to have a max) The constraints Ax b T and Ax b T together imply that Ax = b. So (DD) (P). Q.E.D. July 12, / 30

13 Weak Duality Primal max z = c T x + z 0 Ax = b x 0 Dual min w = b T y + z 0 A T y c y free Theorem Given primal and dual problems as above have feasible solutions x and y, then z w. July 12, / 30

14 Weak Duality Proof Proof: We essentially showed this earlier: w = = = = z m y i b i + z 0 m y i n a ij x j + z 0 because (P) requires b i = j=1 j=1 ( n m ) y i a ij x j + z 0 swapping order of summation j=1 n c j x j + z 0 j=1 Hence w z from constraints of (D) n a ij x j July 12, / 30

15 Weak Duality Consequences Corollary If we have feasible solutions x and y to the primal and dual respectively and w = z, then these are optimal solutions to their respective problems. Proof: w is an upper bound on z (and visa versa), so if z = w for a feasible solution, it has achieved its upper bound, and hence we have an optimal solution. Q.E.D. July 12, / 30

16 Boundedness v Feasibility Corollary If the primal (dual) problem is unbounded then the dual (primal) problem is infeasible. Proof: If the primal were feasible and unbounded, then that means there can be no upper bound on z, so we cannot have a feasible solution the dual. Similarly if the dual is unbounded. Q.E.D. July 12, / 30

17 Strong Duality Theorem If the primal (dual) problem has a finite optimal solution, then so does the dual (primal) problem, and these two values are equal. Proof: see later as a result of complementary slackness. July 12, / 30

18 The Dual of an LP 12 Summary of Results for Primal/Dual pair (P) and (D) 1 For a feasible solution x 1,..., x n of (P), with value z, and a feasible solution y 1,..., y m of (D), with value w, we have w z (Weak Duality) 2 If (P) has an optimal solution then (D) has an optimal solution and max z = min w (Strong Duality) 3 Because the dual of the dual is the primal, if (D) has an optimal solution then (P) has an optimal solution and max z = min w (Strong Duality) 4 If (P) has no optimal solution (z ) then (D) cannot have a feasible solution (as w max z), and if (D) has no optimal solution (w ) then (P) cannot have a feasible solution (as z min w). July 12, / 30

19 The Dual of an LP 13 Summary of Results for Primal/Dual pair (P) and (D) Dual (D) Primal (P) stop 1 stop 2 stop 3 optimal feasible sol n, no feasible solution no opt. sol n solution stop 1 optimal possible impossible impossible solution stop 2 feas. sol n, impossible impossible possible no opt. sol n stop 3 no feasible impossible possible possible solution (max z = min w) July 12, / 30

20 Section 2 Complementary Slackness July 12, / 30

21 Complementary slackness Theorem (Complementary slackness) Given primal problem P and dual problem D with feasible solutions x and y, respectively, then x is an optimal solution of (P) and y an optimal solution of (D) if and only if ( m ) x j y i a ij c j = 0, for j = 1,..., n. Implicit in this theorem is the Strong Duality Theorem, which we prove now, and the second part is called the complementary slackness relations. Definition (Complementary Slackness Relations) ( m ) The relations x j y i a ij c j = 0, for each j = 1,..., n are called the Complementary Slackness Relations (CSR). July 12, / 30

22 Complementary slackness proof Proof: We have already seen the argument: n n m z = c j x j + z 0 y i a ij x j + z 0 (as j=1 j=1 ) m y i a ij c j, x j 0 = m n a ij x j y i + z 0 j=1 m = y i b i + z 0 = w as So z w for all feasible solutions n a ij x j = b i j=1 July 12, / 30

23 Complementary slackness proof Proof : (cont) From Strong Duality, if we have an optimal solution, then z = w. Reversing the above logic, we get n c j x j + z 0 = j=1 ( n m ) y i a ij x j + z 0 j=1 which, when we group the x i terms together gives ( n m ) y i a ij c j x j = 0 j=1 July 12, / 30

24 Complementary slackness proof Proof : (cont) If we know that n j 0, and we have n j = 0 j then we must have n j = 0. This applies here because we know from the LP dual and primals that m y i a ij c j 0 and x j 0. Hence ( m ) y i a ij c j x j = 0, for each j = 1,..., n. July 12, / 30

25 Complementary Slackness: Example Primal Tableax (P) x 1 x 2 x 3 x 4 x 5 b Simplex Phase I and II Resulting solution x = (5/6, 2/3, 0, 0, 1/2), with z = 1/6. July 12, / 30

26 Complementary Slackness: Example Primal Tableax (P) x 1 x 2 x 3 x 4 x 5 b Dual Tableax (D) (note that these represent ) y 1 y 2 y 3 c July 12, / 30

27 Complementary slackness Write down the dual (D) of (P). (D) min w = y 1 + 3y 2 + 2y 3 2y 1 + 2y 2 + y 3 1 (iv) y 1 + 2y 2 + y 3 1 (v) y 1 + y 2 + y 3 1 (vi) y 1 2y 2 + y 3 1 (vii) y 3 0 (viii) y 1, y 2 and y 3 are all free variables 1. (i)-(iii) Writing down the CSR, we see that From (iv), we get (2y 1 + 2y 2 + y 3 1) x 1 = 0 and since x 1 > 0, we have 2y 1 + 2y 2 + y 3 = 1. From (v), since x 2 > 0, we have y 1 + 2y 2 + y 3 = 1. Similarly, from (viii), since x 5 > 0, we have y 3 = 0. 1 Note the last constraint tightens y 3, but this is OK. July 12, / 30

28 Complementary slackness Thus from the CSR for (iv),(v) and (viii) we see that 2y1 + 2y2 + y3 = 1 y1 + 2y2 + y3 = 1 y3 = 0 Solving these equalities for y 1, y 2, y 3, we get y 1 = 2 3, y 2 = 1 6, y 3 = 0, which need to be checked for feasibility in the other constraints: That is, (vi) > 1 (vii) > 1. Also note that z = = 1 6 = w = = = 1 6. July 12, / 30

29 Takeaways Optimisation problems have a Dual this is a very general concept, and holds beyond LPs Complementary slackness relates dual solutions to primal these give us an edge in knowing when we have found optimal solutions we ll use these later! July 12, / 30

30 Further reading I July 12, / 30