Download presentation
Presentation is loading. Please wait.
Published byHelen Sundqvist Modified over 5 years ago
1
Proving that a Valid Inequality is Facet-defining
Ref: W, p ๐โ ๐ + ๐ . For simplicity, assume conv(๐) bounded and full-dimensional. Consider example, ๐= ๐ฅ,๐ฆ โ ๐
+ ๐ ร ๐ต 1 : ๐=1 ๐ ๐ฅ ๐ โค๐๐ฆ . conv(๐) is full-dimensional: Consider the ๐+2 affinely independent points (0, 0), (0, 1), ( ๐ ๐ , 1), ๐=1,โฆ,๐. Problem 1: Given ๐โ ๐ + ๐ and a valid inequality ๐๐ฅโค ๐ 0 for ๐, show that the inequality defines a facet of conv(๐). Ex: Show that ๐ฅ ๐ โค๐ฆ is facet-defining. Approach 1: (Use definition) Find ๐ points ๐ฅ 1 ,โฆ, ๐ฅ ๐ โ๐ satisfying ๐๐ฅ= ๐ 0 , and then prove that these ๐ points are affinely independent. Ex: Consider ๐+1 points: (0, 0), ( ๐ ๐ , 1), and ( ๐ ๐ + ๐ ๐ , 1) for ๐โ ๐. Integer Programming 2017
2
Approach 2: (indirect but useful way, see Thm 3.5, 3.6)
Select ๐กโฅ๐ points ๐ฅ 1 ,โฆ, ๐ฅ ๐ก โ๐ satisfying ๐๐ฅ= ๐ 0 . Suppose that all these points lie on a generic hyperplane ๐๐ฅ= ๐ 0 . Solve the linear equation system ๐=1 ๐ ๐ ๐ ๐ฅ ๐ ๐ = ๐ 0 for ๐=1,โฆ,๐ก in the ๐+1 unknowns ๐, ๐ 0 . If the only solution is ๐, ๐ 0 =๐ ๐, ๐ 0 for ๐โ 0, then the inequality ๐๐ฅโค ๐ 0 is facet-defining. Ex: Show ๐ฅ ๐ โค๐ฆ is facet-defining. Select points (0, 0), ( ๐ ๐ , 1), ( ๐ ๐ + ๐ ๐ , 1) for ๐โ ๐ that are feasible and satisfy ๐ฅ ๐ =๐ฆ. As (0, 0) lies on ๐=1 ๐ ๐ ๐ ๐ฅ ๐ + ๐ ๐+1 ๐ฆ= ๐ 0 , have ๐ 0 =0. As ( ๐ ๐ , 1) lies on the hyperplane ๐=1 ๐ ๐ ๐ ๐ฅ ๐ + ๐ ๐+1 ๐ฆ=0, have ๐ ๐ =โ ๐ ๐+1 . As ( ๐ ๐ + ๐ ๐ , 1) lies on the hyperplane ๐=1 ๐ ๐ ๐ ๐ฅ ๐ โ ๐ ๐ ๐ฆ=0, ๐ ๐ =0 for ๐โ ๐. So the hyperplane is ๐ ๐ ๐ฅ ๐ โ ๐ ๐ ๐ฆ=0, and ๐ฅ ๐ โค๐ฆ is facet-defining. Integer Programming 2017
3
If conv(๐) is not full-dimensional, we may use approach 1 (find affinely independent points) or may use Thm 3.6 in the previous slides. Refer Proposition 3.5 on p.274 for a possible application of Thm 3.6. Integer Programming 2017
4
4. Describing Polyhedra by Extreme Points and Extreme Rays
Prop 4.1: If ๐={๐ฅโ ๐
๐ :๐ด๐ฅโค๐}โ โ
and rank(๐ด) =๐โ๐, ๐ has a face of dimension ๐ and no proper face of lower dimension. Pf) For any face ๐นโ๐, rank ๐ด ๐น = , ๐ ๐น = โค๐โ๐ โน dim ๐น โฅ๐. (Prop. 2.4) Show โ ๐น with dim ๐น =๐. Let ๐น be a face of minimum dimension ( >0). (If ๐=0, nothing to prove) Let ๐ฅ โ be an inner point of ๐น, dim ๐น >0 โน โ ๐ฆโ ๐ฅ โ โ๐น. Consider ๐ง ๐ = ๐ฅ โ +๐ ๐ฆโ ๐ฅ โ , ๐โ ๐
1 . Suppose z(๐) intersects ๐ ๐ ๐ฅ= ๐ ๐ for some ๐โ ๐ ๐น โค . Choose ๐ โ = min { ๐ ๐ : ๐โ ๐ ๐น โค , ๐ง ๐ ๐ lies in ๐ ๐ ๐ฅ= ๐ ๐ }, and ๐ โ = ๐ ๐ โ . Then ๐ โ โ 0 ( ๐ฅ โ is an inner point) โน ๐น ๐ โ ={๐ฅโ๐: ๐ด ๐น = ๐ฅ= ๐ ๐น = , ๐ ๐ โ ๐ฅ= ๐ ๐ โ }โ โ
is a face of ๐ of smaller dimension than ๐น, which is a contradiction. Hence ๐ง(๐) not intersect ๐ ๐ ๐ฅ= ๐ ๐ for any ๐โ ๐ ๐น โค . โน ๐ด ๐ฅ โ +๐ด๐ ๐ฆโ ๐ฅ โ โค๐ โ ๐โ ๐
1 โน ๐ด ๐ฆโ ๐ฅ โ =0 โ ๐ฆโ๐น Thus ๐น={๐ฆ:๐ด๐ฆ=๐ด ๐ฅ โ } โน dim ๐น =๐ since rank ๐ด =๐โ๐ ๏ Integer Programming 2017
5
Pf) (โธ) Suppose ๐ฅ is zero-dimensional face โน rank ๐ด ๐ฅ = =๐. (Prop 2.4)
Frequently we assume ๐โ ๐
+ ๐ โน rank ๐ด =๐ โน ๐ has zero-dimensional faces if ๐โ โ
. Assume rank ๐ด =๐ hereafter. Def 4.1: ๐ฅโ๐ is an extreme point of ๐ if there do not exist ๐ฅ 1 , ๐ฅ 2 โ๐, ๐ฅ 1 โ ๐ฅ 2 such that ๐ฅ= ๐ฅ ๐ฅ 2 . Prop 4.2: ๐ฅ is an extreme point of ๐ โบ ๐ฅ is a zero-dimensional face of ๐. Pf) (โธ) Suppose ๐ฅ is zero-dimensional face โน rank ๐ด ๐ฅ = =๐. (Prop 2.4) Let ( ๐ด , ๐ ) be submatrix of ( ๐ด ๐ฅ = , ๐ ๐ฅ = ) with ๐ด :๐ร๐ and rank ๐ โน ๐ฅ= ๐ด โ1 ๐ . If ๐ฅ= ๐ฅ ๐ฅ 2 , ๐ฅ 1 , ๐ฅ 2 โ๐, then since ๐ด ๐ฅ ๐ โค ๐ , ๐=1,2, we have ๐ด ๐ฅ 1 = ๐ด ๐ฅ 2 = ๐ ( ๐ด ๐ฅ= ๐ด ๐ฅ ๐ด ๐ฅ 2 = ๐ , ๐ด ๐ฅ 1 โค ๐ , ๐ด ๐ฅ 2 โค ๐ ) โน ๐ฅ 1 = ๐ฅ 2 =๐ฅ, so ๐ฅ is an extreme point. (โน) If ๐ฅโ๐ is not a zero-dimensional face of ๐, then rank ๐ด ๐ฅ = <๐. (Prop 2.4) โน โ ๐ฆโ 0 such that ๐ด ๐ฅ = ๐ฆ=0. For small ๐>0, let ๐ฅ 1 =๐ฅ+๐๐ฆ, ๐ฅ 2 =๐ฅโ๐๐ฆ, ๐ฅ 1 , ๐ฅ 2 โ๐. Then ๐ฅ= ๐ฅ ๐ฅ 2 , hence ๐ฅ is not an extreme point. ๏ฟ Integer Programming 2017
6
๐โ ๐
๐ , ๐โ 0 is a ray of ๐ โบ โ ๐ฅโ๐, ๐ฆโ ๐
๐ :๐ฆ=๐ฅ+๐๐, ๐โ ๐
+ 1 โ๐.
Def 4.2: Let ๐ 0 = ๐โ ๐
๐ :๐ด๐โค0 . (recession cone, characteristic cone of ๐) If ๐= ๐ฅโ ๐
๐ :๐ด๐ฅโค๐ โ โ
, then ๐โ ๐ 0 \{0} is called a ray of ๐. ๐โ ๐
๐ , ๐โ 0 is a ray of ๐ โบ โ ๐ฅโ๐, ๐ฆโ ๐
๐ :๐ฆ=๐ฅ+๐๐, ๐โ ๐
+ 1 โ๐. Note: Cone ๐พ is called pointed if ๐พโฉ โ๐พ = 0 . ๐พโฉ(โ๐พ) is called lineality space of cone ๐พ. For ๐= ๐ฅโ ๐
๐ :๐ด๐ฅโค๐ , if rank ๐ด =๐, ๐ 0 โฉ โ ๐ 0 = ๐โ ๐
๐ :๐ด๐โค0, โ๐ด๐โค0 = 0 . Hence ๐ 0 is guaranteed to be pointed. Def 4.3: A ray ๐ of ๐ is an extreme ray if there do not exist ๐ 1 , ๐ 2 โ ๐ 0 , ๐ 1 โ ๐ ๐ 2 , ๐โ ๐
+ 1 such that ๐= ๐ ๐ 2 . Integer Programming 2017
7
If ๐= 1 2 ๐ 1 + 1 2 ๐ 2 , get contradiction as in Prop 4.2.
Prop 4.3: If ๐โ โ
, ๐ extreme ray of ๐ if and only if {๐๐: ๐โ ๐
+ 1 } is one-dimensional face of ๐ 0 . Pf) (โธ) Let ๐ด ๐ = = ๐ ๐ :๐โ๐, ๐ ๐ ๐=0 . If {๐๐: ๐โ ๐
+ 1 } is a one-dimensional face of ๐ 0 , rank ๐ด ๐ = =๐โ1 โน solutions of ๐ด ๐ = ๐ฆ=0 are ๐ฆ=๐๐, ๐โ ๐
1 . If ๐= ๐ ๐ 2 , get contradiction as in Prop 4.2. (โน) If ๐โ ๐ 0 and rank ๐ด ๐ = <๐โ1, then nullity of ๐ด ๐ = โฅ2. โน โ ๐ โ โ ๐๐, ๐โ ๐
1 such that ๐ด ๐ = ๐ โ =0. Then ๐= ๐ ๐ 2 , where ๐ 1 =๐+๐ ๐ โ , ๐ 2 =๐โ๐ ๐ โ . Hence ๐ is not an extreme ray, contradiction. ๏ Cor 4.4: A polyhedron has a finite number of extreme points and extreme rays. Question: Given ๐= ๐ฅโ ๐
๐ :๐ด๐ฅโค๐, ๐ฅโฅ0 โ โ
, how can we identify the extreme rays of ๐ 0 ? Thm 4.5: If ๐โ โ
, rank ๐ด =๐, and max {๐๐ฅ:๐ฅโ๐} is finite, then there is an optimal solution that is an extreme point. Pf) Set of optimal solution is face ๐น= ๐ฅโ๐:๐๐ฅ= ๐ By Prop. 4.1, ๐น contains ๐โrank ๐ด โdimensional face. By Prop. 4.2, ๐น contains an extreme point. ๏ฟ Integer Programming 2017
8
(Compare with earlier Proposition regarding face.)
Thm 4.6: โ extreme points ๐ฅ ๐ , โ ๐โ ๐ ๐ such that ๐ฅ ๐ is the unique optimal solution of max {๐๐ฅ:๐ฅโ๐}. Pf) Let ๐ ๐ฅ ๐ = be equality set of ๐ฅ ๐ . Let ๐ โ = ๐โ ๐ ๐ฅ ๐ = ๐ ๐ , ๐=๐ ๐ โ for some ๐>0 to get integer vector ๐. Then โ ๐ฅโ๐โ ๐ฅ ๐ , ๐๐ฅ= ๐โ ๐ ๐ฅ ๐ = ๐ ๐ ๐ ๐ฅ < ๐โ ๐ ๐ฅ ๐ = ๐ ๐ ๐ = ๐โ ๐ ๐ฅ ๐ = ๐ ๐ ๐ ๐ฅ ๐ =๐ ๐ฅ ๐ . ๏ฟ (Compare with earlier Proposition regarding face.) Thm 4.7: ๐โ โ
, rank ๐ด =๐, max{๐๐ฅ:๐ฅโ๐} unbounded, then ๐ has an extreme ray ๐ โ with ๐ ๐ โ >0. Pf) ๐ขโ ๐
+ ๐ :๐ข๐ด=๐ =โ
from duality of LP โน By Farkas, โ ๐โ ๐
๐ such that ๐ด๐โค0, ๐๐>0. Consider max ๐๐:๐ด๐โค0, ๐๐โค1 =1. By Thm 4.5, โ optimal extreme point solution ๐ โ . Equality set of ๐ โ is ๐ด ๐ โ = ๐=0 and ๐๐=1 โน rank ๐ด ๐ โ = =๐โ1. โน ๐ โ extreme ray of ๐ (Prop 4.3) Integer Programming 2017
9
๐= ๐ฅโ ๐
๐ :๐ด๐ฅโค๐ โ โ
, rank ๐ด =๐ (existence of extreme point guaranteed)
Thm 4.8: (Affine) Minkowskiโs Thm: finitely constrained ๏ finitely generated. ๐= ๐ฅโ ๐
๐ :๐ด๐ฅโค๐ โ โ
, rank ๐ด =๐ (existence of extreme point guaranteed) โน ๐={๐ฅโ ๐
๐ :๐ฅ= ๐โ๐พ ๐ ๐ ๐ฅ ๐ + ๐โ๐ฝ ๐ ๐ ๐ ๐ , ๐โ๐พ ๐ ๐ =1, ๐ ๐ โฅ0 ๐๐๐ ๐โ๐พ, ๐ ๐ โฅ0 ๐๐๐ ๐โ๐ฝ}, where ๐ฅ ๐ ๐โ๐พ : extreme points of ๐, ๐ ๐ ๐โ๐ฝ : extreme rays of ๐. Pf) Let ๐={๐ฅโ ๐
๐ :๐ฅ= ๐ ๐ ๐ฅ ๐ + ๐ ๐ ๐ ๐ , ๐ ๐ =1, ๐ ๐ โฅ0, ๐ ๐ โฅ0} . ๐โ๐ is clear. Suppose โ ๐ฆโ๐โ๐ (i.e. ๐ฆโ๐, but ๐ฆโ๐). Show contradiction. Then not exist ๐,๐ satisfying ๐โ๐พ ๐ ๐ ๐ฅ ๐ + ๐โ๐ฝ ๐ ๐ ๐ ๐ =๐ฆ โ ๐โ๐พ ๐ ๐ =โ1 ๐ ๐ โฅ0 for ๐โ๐พ, ๐ ๐ โฅ0 for ๐โ๐ฝ By Farkasโ lemma, โ (๐, ๐ 0 )โ ๐
๐+1 such that ๐ ๐ฅ ๐ โ ๐ 0 โค0 for ๐โ๐พ, ๐ ๐ ๐ โค0 for ๐โ๐ฝ and ๐๐ฆโ ๐ 0 >0. Consider LP max ๐๐ฅ:๐ฅโ๐ . Integer Programming 2017
10
Hence there does not exist such ๐ฆ, i.e. ๐=๐. ๏ฟ
(continued) If LP has a finite optimal solution, then โ an extreme point optimal solution. Have ๐ ๐ฅ ๐ โ ๐ 0 โค0, but ๐๐ฆโ ๐ 0 >0 ๐๐ฆ>๐ ๐ฅ ๐ โ ๐ , contradiction. If unbounded, โ extreme ray ๐ ๐ with ๐ ๐ ๐ >0 (Thm 4.7), contradiction. Hence there does not exist such ๐ฆ, i.e. ๐=๐. ๏ฟ Integer Programming 2017
11
Now consider projections of polyhedra and Weylโs theorem.
Consider Primal-Dual pair of LP ๐ง= max ๐๐ฅ:๐ฅโ๐ , ๐={๐ฅโ ๐
+ ๐ :๐ด๐ฅโค๐} ๐ค= min ๐ข๐:๐ขโ๐ , ๐={๐ขโ ๐
+ ๐ :๐ข๐ดโฅ๐} { ๐ฅ ๐ , ๐โ๐พ} extreme points of ๐, { ๐ ๐ , ๐โ๐ฝ} extreme rays of ๐ 0 { ๐ข ๐ , ๐โ๐ผ} extreme points of ๐, { ๐ฃ ๐ก , ๐กโ๐} extreme rays of ๐ 0 Thm of the alternatives: โ ๐ฅ such that ๐ฅโฅ0, ๐ด๐ฅโค๐ โ ๐ข such that ๐ขโฅ0, ๐ข๐ดโฅ0, ๐ข๐<0 Pf) Consider primal-dual LP pair (P) max 0๐ฅ, ๐ด๐ฅโค๐, ๐ฅโฅ0 (D) min ๐ข๐, ๐ข๐ดโฅ0, ๐ขโฅ0 Integer Programming 2017
12
Pf) I) ๐โ โ
if and only if ๐ฃ๐โฅ0 โ ๐ฃโ ๐
+ ๐ with ๐ฃ๐ดโฅ0 (from previous)
Thm 4.9: The following are equivalent: The primal problem is feasible, that is, ๐โ โ
; ๐ฃ ๐ก ๐โฅ0 for all ๐กโ๐. The following are equivalent when the primal problem is feasible: ๐ง is unbounded from above; โ ๐ ๐ of ๐ with ๐ ๐ ๐ >0; the dual problem is infeasible, that is, ๐=โ
. If the primal problem is feasible and ๐ง is bounded, then ๐ง= max ๐โ๐พ ๐ ๐ฅ ๐ =๐ค= min ๐โ๐ผ ๐ข ๐ ๐ . Pf) I) ๐โ โ
if and only if ๐ฃ๐โฅ0 โ ๐ฃโ ๐
+ ๐ with ๐ฃ๐ดโฅ0 (from previous) By Mink., ๐ 0 = ๐ฃโ ๐
+ ๐ :๐ฃ๐ดโฅ0 = ๐ฃโ ๐
+ ๐ :๐ฃ= ๐ ๐ก ๐ฃ ๐ก , ๐ ๐ก โฅ0,๐กโ๐ . Hence ๐ฃ๐โฅ0 โ ๐ฃโ ๐ 0 if and only if ๐ฃ ๐ก ๐โฅ0 โ ๐กโ๐. II) ๐= ๐ฅโ ๐
๐ :๐ฅ= ๐ ๐ ๐ฅ ๐ + ๐ ๐ ๐ ๐ , ๐ ๐ =1, ๐ ๐ โฅ0, ๐ ๐ โฅ0 โ โ
. ๐ง bounded if and only if ๐ ๐ ๐ โค0 โ ๐โ๐ฝ. b โบ c: apply (I) to dual (in negation form) III) From strong duality and Minkowskiโs theorem to ๐ and ๐. ๏ฟ Integer Programming 2017
13
Note: More general form of Minkowskiโs thm (from IE531)
Decomposition Thm: Suppose ๐={๐ฅโ ๐
๐ :๐ด๐ฅโค๐}โ โ
Then ๐=๐+๐พ+๐, where ๐+๐พ is the cone {๐ฅโ ๐
๐ :๐ด๐ฅโค0} ๐={๐ฅโ ๐
๐ :๐ด๐ฅ=0} is the lineality space of ๐+๐พ ๐พ is a pointed cone. ๐พ+๐ is a pointed polyhedron. ๐ is a polytope given by the convex hull of extreme points of ๐พ+๐. Integer Programming 2017
14
Projection of a polyhedron:
Projection of (๐ฅ,๐ฆ)โ ๐
๐ ร ๐
๐ on ๐ป={ ๐ฅ,๐ฆ :๐ฆ=0} is (๐ฅ,0). Consider projection of ๐โ ๐
๐ ร ๐
๐ onto ๐ฆ=0 as a projection from the ๐ฅ,๐ฆ โspace to the ๐ฅโspace, denoted by proj ๐ฅ (๐). (๐ฅ such that (๐ฅ,๐ฆ)โ๐ for some ๐ฆโ ๐
๐ ) Thm 4.10: Let ๐= ๐ฅ,๐ฆ โ ๐
๐ ร ๐
๐ :๐ด๐ฅ+๐บ๐ฆโค๐ , then proj ๐ฅ ๐ = ๐ฅโ ๐
๐ : ๐ฃ ๐ก ๐โ๐ด๐ฅ โฅ0 โ ๐กโ๐ , where ๐ฃ ๐ก ๐กโ๐ are extreme rays of ๐= ๐ฃโ ๐
+ ๐ :๐ฃ๐บ=0 . Pf) ๐ป={(๐ฅ,๐ฆ)โ ๐
๐ ร ๐
๐ :๐ฆ=0} โน Proj๐ป(๐) ={(๐ฅ,0)โ ๐
๐ ร ๐
๐ :(๐ฅ,๐ฆ)โ๐ for some ๐ฆโ ๐
๐ } Hence, ๐ฅโ Proj๐ป(๐) โบ ๐บ๐ฆโค(๐โ๐ด๐ฅ) feasible for given ๐ฅ โบ ๐ฃโฅ0, ๐ฃ๐บ=0, ๐ฃ ๐โ๐ด๐ฅ <0 infeasible โบ โ ๐ฃโฅ0, ๐ฃ๐บ=0, we have ๐ฃ(๐โ๐ด๐ฅ)โฅ0 โบ ๐ฃ ๐ก (๐โ๐ด๐ฅ)โฅ0 for all ๐กโ๐ ( ๐ฃ ๐ก ๐ด๐ฅโค ๐ฃ ๐ก ๐) ๏ Integer Programming 2017
15
For Thm 4.10, use the thm of the alternatives:
โ ๐ฅ such that ๐ด๐ฅโค๐ โ ๐ข such that ๐ขโฅ0, ๐ข๐ด=0, ๐ข๐<0 Pf) Consider primal-dual pair (P) max 0๐ฅ, ๐ด๐ฅโค๐ (D) min ๐ข๐, ๐ข๐ด=0, ๐ขโฅ0 Cor 4.11: Projection of a polyhedron is a polyhedron. Integer Programming 2017
16
If ๐ด: ๐ 1 ร๐, ๐ต: ๐ 2 ร๐, rational matrices and
Cor 4.12: If ๐={(๐ฅ,๐ฆ)โ ๐
๐ ร ๐
๐ :๐ด๐ฅ+๐บ๐ฆโค๐} and ๐= ๐ฅโ ๐
๐ :๐ท๐ฅโค๐ , where ๐ท is ๐ร๐, then ๐= proj๐ฅ(๐) if and only if: For ๐=1,โฆ๐, ๐ ๐ ๐ฅโค ๐ 0 ๐ is a valid inequality for ๐. For each ๐ฅ โ โ๐, โ ๐ฆ โ such that ๐ฅ โ , ๐ฆ โ โ๐. Pf) I. is equivalent to ๐โ proj๐ฅ(๐). II is equivalent to ๐โ proj๐ฅ(๐). ๏ฟ Thm 4.13: (Affine Weylโs theorem) (finitely generated โน finitely constrained) If ๐ด: ๐ 1 ร๐, ๐ต: ๐ 2 ร๐, rational matrices and ๐= ๐ฅโ ๐
๐ :๐ฅ=๐ฆ๐ด+๐ง๐ต, ๐=1 ๐ 1 ๐ฆ ๐ =1, ๐ฆโ ๐
+ ๐ 1 ,๐งโ ๐
+ ๐ 2 , Then ๐ is a rational polyhedron. Pf) ๐= proj๐ฅ(๐), where ๐={(๐ฅ,๐ฆ,๐ง)โ ๐
๐ ร ๐
+ ๐ 1 ร ๐
+ ๐ 2 :๐ฅโ๐ฆ๐ดโ๐ง๐ต=0, ๐=1 ๐ 1 ๐ฆ ๐ =1} ๏ฟ (Recall that we used Fourier-Motzkin elimination in IE531.) Integer Programming 2017
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.