o Store: & Mar 170 — PRoBLEM Ser 3 (Due TUESDAY FEBRUARY 7) SOLUTIONS BY FREDERICK LAW 1, (BUT 2.18) Consider a polyhedron P = {x : Ax > b}. Given any e > 0, show that there exists some b with the following two properties: (a) The absolute value of every component b — b is hounded by ¢. (b) Every baste feasible solution in the polyhedron P = {x : Ax > b} is nondegenerate. Solution: Intuitively, we shall find such aB by fist identifying excessive constraints and perturb ing the necessary constraints by some c. That is, if P = HT,7+--7Hy, where H, ae the halé-spaces whose intersection gives us P, then we identify those Hy which are unnecessary in defining P. 16H, is excessive, then P is well defined apart from Hy, which means that (gz; Me = P. Let 1M be the set of all indices of excessive halsjmces. ‘Then P= (leg, Ha, that we can remove all the halespaces of Mf and still get the same polyhedron. This is true just because of our construction, since all the half-spaces removed are excessive. If Hy = oc €R” : ajx > by, then let Hy = (X ER" : ahx > 0, ~c}. Let B be defined by: by jEM bal & { b-e jeM ‘Then our new polyhedron is P! = geny He Meas H}- Note that in P, the degenerate basic feasible solutions can be interpreted as having too maniy hyperplanes, boundary of hall-spaces, touching at that point. For example, on a cube in three dimensions, all the points are non- degenerate as they all have 3 hyperplanes touching them, the three facets that connect at a corner. Moreover, every facet of a polyhedron is not excessive, since it serves as 8 boundary of the polyhedron. Thus, we can interpret the excessive constraints as half-spaces whose hyperplane boundary cither touches the polyhedron only at some degenerate solution or at not degenerate solution... Therefore, by perturbing these outward by e, we remove all the degeneracy from our basic feasible solutions. Thus all the basic feasible solutions in P’ are nondegenerate. Also, since ‘we have only perturbed the excessive half-spaces by «, it follows that by construction |B; ~ bi] <= for alli. 2, (BLT 2.21) Suppose that Fourier-Motzkin elimination is used in the manner described at the fend of Section 2.8 to find the optimal cost in a linear programming problem. Show how this approach can be augmented to obtain an optimal solution as well Solution: To get an optimal solution using the Fourier-Motzkin elimination, we first find the ‘optimal solution. ‘This is done by extending our LP by one variable xo, 50 we get a new polyhedron in R**+ defined by {(¢o,x) : x € Pyelx = 20}. Then we use the Fourier Motakin elimination to project onto the fist variable, which gives us (zy ¢ R : x € Ps.t.e'x = za). Then we nninimize over this subset of R to find our optimal cost, call this c*. To get an optimal solution, ‘we really just project back upwards on n dimensions, by inverting the Fourier Motrkin algorithm. Moreover, to save time in our algorithm, we really only need to project upwards starting at c* ‘That is, if we imagine our entire process as a mapping &: P+ R where x € P gets sent to ex, then we take the preimage over c* which is just the level set @—(c*) and see what points lie on the level set. ‘This is also the same as taking the plane {x €R" : ex = c* and finding where this hyperplane intersects P. 3. (B&T 2.22) Let P and @ be polyhedra in R", Let P+Q= {x+y :x€ Py €Q) (a) Show that P +Q is @ polyhedron, Solution: Let us define M as M = {(a.x,y) i x € Py € Q2=x+y}. If Pis constructed with n; linear constraints and Q is constructed with na linear constraints, then M is constructed with n; +z +n linear constraints, where n of thei come fromx+y component wise, Therefore since M is constructed using linear constraints, A is a polyhedron in R°", ‘Then we use Fourier-Motzkin eliminstion to reduce to the first n coordinates: Thy(M) = (2 € R" : 3x € Py € Qs.x+y =2}. This can be rewritten us n(M) = (24 ¥ CRY : xe Py €Q) = P+. By the Foorier-Movzkin elimination algorithin, we know that Il,(M) is « polyhedron, and thus P+@ is a polyhedron. Cl (b) Show that every extreme point of P+@ is the sum of an extreme point of P and an extreme point of @. Solution: Suppose not, ‘Then there exists x+y which is extreme in P-+@ but elther x is not extreme in P of y is not extreme in Q or both, WLOG, suppose that x is not an cextroise in P, y may or may not be extreme in Q. Since x not extreme in P then that, means there exists #2’ € P, X€ [0,1 such that x 4 and x #2! and x= Az + (I~ A)u! ‘Then we have dat (1 Ael + Ma ty) +=! ty) But now we have written 2-+y as a convex combination of z+-y anda +y, where xty £aty and a! +y, since x / x and x # 2!. Therefore x + y is nol an extreme point, ‘This is a contradiction, so we are done, o 4, (B&L3.2) (Optimality conditions) Consider the problem of minimizing e' over a polyhedron P. Prove the following (a) A feasible solution x is optimal if and only if e'd > 0 for every feasible direction d at x. Solution: First we prove the forward direction. Suppose that x, a feasible solution, is ‘optimal. ‘Then it follows that Vy € P, e'x < e'y. Suppose d is an arbitrary feasible direction at x. Then there exists @ > 0 such that x-+0d € P. Then e'x <¢(x +6d) =c'x +0e'd. ‘Thus @e'd > 0, and since 9 > 0, we divide by 9 aud get ed > 0. Now we prove the backward direction. Suppose that e'd > 0 for every feasible direction d at x. Let y € P be arbitrary. Then let d= y ~ x. Then if we let @ = 1, then x+8d =y € P, 50 d is a feasible direction at x and so e'd > 0, Thus it follows that e'(y — x) > 0, and so cy ~e'x > O and o'x < cly. Since y was an arbitrary point in P, it follows that 2 is an optimal solution. This proves the equality a (b) A feasible solution x is the unique optimal solution if and feasible direction d at x. Solution: Our argument will be similar to that in part (a). First we prove the forward directiow.. Suppose that x, a feasible solution, is unique optimal. ‘Then this means that for any y € P such that y # x, then e'x < e'y. Let be any nonzero feasible direct at x. ‘Then there exists @ > O such that x-4 0d € P. Since @ > 0 and d 0, then it follows that x~X+6d. Therolore e'x < c'(x + 0d) = e'x + e'd. Subtracting c'x from both sides and dividing by 6 > 0, it follows that ed > 0. Now we prove the backward direction. Suppose that c’d > 0 for any nonzero feasible direction at x. Let y € P be an arbitrary point such that y # x. Since y # x, the d=y—x #0. Using #=1, then x+0d=y € P, sod is a feasible, uonzoro, direction at x. Therefore it follows that c'd > 0. Thus e'(y ~ x) = ely ~e'x > 0. And thus it follows that e'y > e'x. Since y was any point in P not equal to x, it follows that x is the unique ‘optimal solution. This proves the equality a ly iFe'd > 0 for every nonzero _)5. (3wP39) Lex ean lament ofthe standard form poyhedion P= (2 ER" + Ax= b,x > 0). Prove that a vector d'¢ Risa feasible direction at xifand ony if Ad = 0 and d > 0 fr every i such that y= 0 Solution: We fist prove the forward direction, Suppose that dis fesibe at x, ‘Then there exists 0'> O sh that x-+ 0d ¢ P. This menos A(K-+04) = Ax+0Ad-b, Since xe Py thenAx = b, thus subteacting b from both sides, and dividing by @ > 0, it follows that Ad = 0. If i is an index such that 2; =O, and since x + 8d > 0, then 2, + 6d; = @d; > 0. Dividing by 8 > 0, it follows that dy > 0 for any index ¢ such that 2 > 0. Now we prove the backward direction. Suppose dé RM such that Ad ~ 0 and dy > 0 for any index i such that 2, =0, Let us choose 0° satisfying veer cnt{B 2,904 <0} Sinco x € P, then 2; > 0 for all i. If-z; = 0, then If ay > 0, there are two cases. If dy > 0, then xy + Od; > 0. Id, < 0, then O* < —~%, and multiplying through by d; < 0, we get 0d, > —z, and thus 2, + 6"d, > 0. Thus x +6°d > 0. ‘Also, A(x-+0°d) = Ax 4-0°Ad = Ax =b since x € P. Thos it follows that x + 6"d P. This proves the equality, and we are dos a + Ody = Ody > O, since dy > 0 and 0 > 0. (BUT 25) Lot P= Gc ERY: my +2042) = 1.x 2 OF and consider the vector x = (0,0,1) Find the st of feasible directions at Solution: We find the set of feasible directions by working with the definition of feasible di- rections. We say that d is feasible atx if there exists some @ > 0 such that x-+ 8d € P. Let = (diydaydy). Then x + 04 = (Ot, Ody, + 84s). ‘To require this to be in P, we neod that (dy + da + dg) +1 = 1, which means dy + dz + dy = 0, and thos dy = —dy ~ da. We also require dy, dy t0 be non-negative, since then for aay 8 > 0, 8dr ,2da > 0. Lastly, we shall require 120ds > 0. To do this, we shall chooge 9 so that 1+ 8dg = 0, Then 1 ~ 8(4) ds) and § = 74. Note this 0 will always be positive, unless di = dz = 0. But in that ease, dg = 0, and d= whichis always a feasible dicection. ‘Ths out st of feasible directions is P= {(dssday—dh ~ da) ¢ dy 2 Oda 2 OF [As a safe mental check, if dy = dp = 0, then d = 0, and for 9 = 1,x+d=x6P. If either di or dz > 0, then let 0 = gta; > 0, and then x + 6d = (ziy,, gZtz;,0) € P, since aay + aby = BSG = Land gihay 2 O or i= 12 9 (BET 3.6) (Conditions for a unique optimum) Let x be a asic feasible solution associated, ‘with some basis matrix B. Prove the following: (a) If the reduced cost of every nonbasie variable is postive, then x is the unique optimal solution. Solution: Suppose that 2; > 0 for every noubase variable j. Let y € P be arbitrary such that y 4x. Thon Ax = Ay = b, Thus it follows that A(y ~ x) =0. Let d= y—X. Then using @ = 1, x+-d=y € P, so dis a feasible ditection at x and Ad = 0. We can rewrite this as Bdg + Soyer Apts =O where I is the set of nonbasic indice. Since B is invertible, then dy = Tye B'Ayd. Then d= cyda + Nod =D (o-egBA,) 4 = DH ya Sek ia Since y # x, there must exist some j € J such that yj # xy = 0, where 2y = 0 for all 3 € I since x is a basic feasible solution. If y € P such that yj = 0 for all j ¢ J, then ‘Ay = Byg =b. Then yg = ~B-*b = xp, and so y = x. Therefore there must exist some noubasie index on which y and x disagree. Let & € I be such an index. Since y € P, then ye 2 0, and since yp. # 2x = 0, then yx > 0. Since dj = yj ~ xy and since xj = 0 for all =Dad= Daw 2am>oce 8. (BUT 3.10) Show that if n—m where we use the fact that tj 2 0 for all j € [ and yy > 0 for all j € I since y € P. Thus wwe have c'(y — x) > 0 and thus e'x < Cy. Since this holds for ali y # x, it follows that x. js the unique optimal solution, a (b) If x is the unique optimal solution and is nondegenerate, then the reduced cost of every nonbasic variable is positive. Solution: Suppose not. Then there exists an index m € I, borrowing notation from part {@), such that gm <0. Let d be the mth basie direction.” We know ftom our results in ‘lass and in Bertsimas and ‘Tsitsiklis that d is always a feasible direction at x since x is nondegenerate, Then there exists # > 0 such that x+@d € P. Then dae Da, ie <0 since dis the hth basic direction, so dk = 1 and dj = 0 for j € 1,7 #k. Therefore e'd <0. Choosing > 0 sufficiently small so that x +d © P, we get @e/d <0 which implies that ex + 0d) = e'x + 0e'd < e'x. But x+0d © P, which means that x is not the unique ‘optimal solution. ‘This is @ contradiction, which completes the proof. o 2, then the simplex method will not eycle, no matter which pivoting rule is used. Solution: Note that there is a lot of possible choices in the simplex method, thus if this fact is indeed true it, then the face that n—m = 2 must do something eliminate our possible cholees in directions to move when executing the simplex method. One possible explanation could be th ‘when n—m = 2, then either the polyhedron is unbounded, in which ease simplex terminates at ‘optimal cost being —00 or the polyhedron is bounded and it happens that in the polytope all the basic feasible solutions are nondegenerate, and then the simplex method on this polytope will have to terminate,