Convex Hull of Graphs of Polynomial Functions
5.3 Convex hull of graphs of polynomial functions over a polytope
5.3.1 Preliminary definitions
Let๐, ๐โNand๐ โR๐be a polytope defined by the intersection of finitely many halfspaces, i.e.,๐ := {x = (๐ฅ1, . . . , ๐ฅ๐)๐ โ R๐ | ๐๐๐x โค ๐๐, ๐ = 1, . . . , ๐}, where๐๐ โ R๐, ๐๐ โ R for๐ = 1, . . . , ๐. Then, for a given polynomial function๐ : ๐ โ Rwith ๐ โ R[x], the image of the function๐น :๐โR๐+1,xโฆโ(x, ๐(x))defines the graph of๐. Note that some of definitions and properties are also suitable for general differentiable functions. But in this chapter we focus on polynomial functions.
In this section, we study the convex hull of the set
๐ฎ:={(x, ๐ง)|๐ง=๐(x), xโ๐} โR๐+1.
Define projection functions๐x:R๐รRโR๐,(x, ๐ง)โฆโxand๐๐ง :R๐รRโR,(x, ๐ง)โฆโ๐ง. For every point(x, ๐(x))โ ๐ฎ,x=๐x((x, ๐(x)))is the corresponding domain point.
Theorem 5.1
Let๐โR๐+1be a compact set. Then the convex hull of๐is a compact set which is the intersection of all closed halfspaces containing๐.
Proof. This can be directly derived from Corollary 5.33 and 5.83 of the book [AB06]. 2 Without loss of generality, we assume๐be to full-dimensional, otherwise we can reduce the dimension by eliminating variables until the new equivalent๐is full-dimensional inR๐for ๐โค๐. Furthermore, with suitable preprocessing, every hyperplane๐๐๐x=๐๐ in๐corresponds to a facet of๐.
Let๐๐ :={xโ ๐ | โ๐, ๐๐๐x= ๐๐}denote the boundary of๐and letint๐ := ๐โ๐๐ denote the interior of๐. For๐โR+, denote the closed ball of radius๐centered at the pointx0 by๐ต๐(x0) ={xโR๐| โxโx0โ2 โค๐}. A hyperplane inR๐+1through point(x0, ๐ง0)with ๐ง0=๐(x0)and normal vectorn0โR๐,n0ฬธ= 0can be defined as
๐ป(x0,n0) ={(x, ๐ง)|๐ง=๐ง0+n0ยท(xโx0)}.
The tangent plane to๐ at (x0, ๐(x0))is๐(x0) := ๐ป(x0,โ๐(x0)). The downward closed halfspace associated with hyperplane๐ป(x0,n0)is then
๐ป(xห 0,n0) ={(x, ๐ง)|๐งโฅ๐ง0+n0ยท(xโx0)}.
The upward closed halfspace๐ป(x^ 0,n0)is defined similarly.
Hyperplanes๐ป(x0,n0) are callednonvertical (to space R๐), since๐x(๐ป(x0,n0)) = R๐. Analogously, averticalhyperplane inR๐+1through point(x0, ๐ง0)with๐ง0 =๐(x0)and normal vectorn0 โR๐can be defined as
๐ปโฅ(x0,n0) ={(x, ๐ง)|n0ยท(xโx0) = 0}.
Then the left closed halfspace associated to๐ปโฅ(x0,n0)is defined by ๐ปยดโฅ(x0,n0) ={(x, ๐ง)|n0ยท(xโx0)โค0}
and the right closed halfspace๐ป`โฅ(x0,n0)is defined similarly. For any halfspace๐ปinR๐+1, let๐๐ป denote a map which maps a halfspace to its corresponding hyperplane.
Lemma 5.2
We discuss first halfspaces in๐โฒwith corresponding hyperplanes which are nonvertical. For every๐ปห1 there exists an affine function๐1 :R๐ โRsuch that๐ปห1 ={(x, ๐ง) |๐งโฅ ๐1(x)}.
This is also true for every upward closed halfspaces.
Now we discuss halfspaces in๐โฒwith corresponding hyperplanes which are vertical. Note that for every๐ โ {1, . . . , ๐},{(x, ๐ง)|๐๐๐xโค๐๐}has the form of๐ปยดโฅwith๐ปยดโฅโ ๐ฎ. For any
5.3 Convex hull of graphs of polynomial functions over a polytope
Together with the result for all halfspaces in๐โฒ with corresponding hyperplanes which are
nonvertical, it implies that๐โฒ=๐โฒโฒ. 2
In this section, we want to find a set of halfspaces that all contain๐ฎand the intersection of them is the convex hull of๐ฎ. Lemma 5.2 shows that we can restrict attention to nonvertical halfspaces whose corresponding hyperplane intersects the graph.
Remark 5.3
Obviously,๐ฎ is bounded because of Theorem 5.1. Due to symmetry, we only need to consider the downward closed part of the convex hull since the upward closed part is equivalent to the downward closed part of functionโ๐.
5.3.2 Locally and globally convex points Definition 5.4 (Locally convex points)
A point(x0, ๐(x0))โ ๐ฎ is alocally convex pointif there exists๐ >0,n0 โR๐such that {(x, ๐ง)|xโ๐โฉ๐ต๐(x0), ๐ง=๐(x)}
โ โ
local graph
โ๐ป(xห 0,n0). (5.1) In this case, the domain pointx0is the corresponding locally convex domain point.
Let๐ฎห๐ โ ๐ฎ denote the set of all locally convex points and๐ห๐=๐x( ห๐ฎ๐)the corresponding domain points.
Lemma 5.5
Letx0 โint๐. Ifx0is a locally convex domain point, then the gradient vectorโ๐(x0)is the unique normal vectorn0tox0fulfilling (5.1).
Proof. Consider the function๐n0(x) = ๐(x)โ(n0 ยท(xโx0) +๐(x0))forn0 โ R๐and defined on๐. The point(x0, ๐(x0))is locally convex if and only if there exists๐ >0such that ๐n0(x)โฅ0over๐โฉ๐ต๐(x0). Using โTaylorโs Formula in Several Variablesโ for๐n0(x)atx0 (see the book [Edw94]),๐n0(x)may attain a local minimum0atx0if and only ifn0 =โ๐(x0).
This implies the result. 2
Definition 5.6 (Globally convex points and valid halfspaces)
A point(x0, ๐(x0))โ ๐ฎ is aglobally convex pointif there exist ann0 โR๐such that ๐ฎ ={(x, ๐ง)|xโ๐, ๐ง=๐(x)}
โ โ
total graph
โ๐ป(xห 0,n0). (5.2) Furthermore, we call๐ป(xห 0,n0)avalidhalfspace,๐ป(x0,n0)andn0the corresponding valid hyperplane and valid normal vector, respectively. The pointx0is the corresponding globally convex domain point.
Further we define๐ฎห๐ as the set of all globally convex points and๐ห๐ =๐x( ห๐ฎ๐)the set of corresponding domain points.
Remark 5.7
The definition of globally convex points is actually equivalent to the definition of generating sets, which can be found in Chapter4of the book [LS13]. Due to the different point of view, we keep on using the name defined above in this thesis.
Theorem 5.8
Letx0โint๐. Thenx0is a globally convex domain point if and only if the tangent hyperplane ๐(x0)is valid.
Proof. It is clear that a globally convex domain pointx0must also be a locally convex domain point. The result is then followed from Lemma 5.5, sincen0 is the unique candidate normal
vector to satisfy (5.2). 2
For anyx0 โ ๐ we seek a way to determine if it is a globally convex domain point. For x0 โint๐, we need only to check if๐(x0)is valid due to Theorem 5.8. On the other hand, for a boundary domain pointx0 โ ๐๐, the situation is more complex. We require several additional concepts in this case.
5.3.3 Globally convex boundary points Remark 5.9
For polytope๐, every face๐น possesses a maximal subset๐ผ๐น โ {1, . . . , ๐}such that ๐น ={๏ธxโ๐ |๐๐๐x=๐๐, ๐โ๐ผ๐น}๏ธ.
In this section, we assume that the vertices of๐are nondegenerate which means there is no pointxโR๐which satisfies๐+ 1of the given๐inequalities with equality. This implies two properties:
โข For any face๐น it holds thatdim๐น =๐โ |๐ผ๐น|
โข For any face๐น with|๐ผ๐น|= ๐ โฅ 2and๐ผ๐น = {1, . . . , ๐}without loss of generality, it implies that
๐นโ๐ ={๏ธxโ๐|๐๐๐x=๐๐, ๐ โ๐ผ๐น โ {๐}}๏ธ
is also a face of๐with๐น (๐นโ๐anddim๐น = dim๐นโ๐โ1for any๐โ๐ผ๐น.
Note that the assumption that๐is vertex-nondegenerate is not a necessary condition for this section. However, this assumption makes the notation and description easier. For a vertex-degenerate๐, the dimension of any face can be determined by calculating the rank of an auxiliary matrix. Since there are finitely many faces, for any given face๐น withdim๐น โค๐โ2, all faces that contain๐น of dimension(dim๐น+ 1)can be determined as well.
Based on this remark, we have the following lemma. Example 5.11 following after its proof illustrates the lemma in a graphical way.
5.3 Convex hull of graphs of polynomial functions over a polytope
Lemma 5.10 (Complete Projection based on the smallest face containingx0) Letx0 โ๐๐and define the index set
๐ผx0 :={๏ธ๐โ {1, . . . , ๐} |๐๐๐x0 =๐๐}๏ธ
which is nonempty. Let๐โฒ =|๐ผx0|be the cardinality with1โค๐โฒ โค๐. Then, the set ๐x0 :={xโR๐|๐๐๐x=๐๐for all๐ โ๐ผx0}
defines an affine set inR๐ of dimension๐ := ๐โ๐โฒ. Further, the set๐นx0 := ๐x0 โฉ๐is the smallest face of๐which containsx0and has dimension๐.
By permuting variables inxif necessary, there exists a bijective linear map ๐๐:๐x0 โR๐,xโฆโx๐
wherex๐= (๐ฅ1, . . . , ๐ฅ๐)๐ such that
โข The set
๐ฎ๐:={(x๐, ๐ง)|x๐=๐๐(x), ๐ง=๐(x),xโ๐นx0} โR๐+1 is the graph of some polynomial function over a polytope.
โข Either(๐๐(x0), ๐(x0))is an interior point of๐ฎ๐or๐ฎ๐={(๐๐(x0), ๐(x0))}.
Proof. Without loss of generality, we assume๐ผx0 ={1, . . . , ๐โฒ}. It is clear that๐นx0 is a face of๐withx0 โ๐นx0. Assume that there exists another face๐น of๐withx0 โ๐น (๐นx0. Then there exists๐ โ {๐โฒ+ 1, . . . , ๐}such that๐๐๐x0 =๐๐. This is a contradiction to the definition of๐ผx0. Hence๐นx0 is the smallest face of๐containingx0. Similar to the preprocessing of half spaces which form๐, which shows that๐นx0 has dimension๐.
Furthermore, we define
๐ดx0 =
โ
โ
โ ๐๐1
...
๐๐๐โฒ
โ
โ
โ โR๐
โฒร๐,
and๐x0 = (๐1, ๐2, . . . , ๐๐โฒ)๐. Then we have
๐x0 ={xโR๐|๐ดx0x=๐x0}
and๐โฒ= rank(๐ดx0) =๐โ๐. Via permuting variables, we may assume that the last๐โฒcolumns of๐ดx0 are linearly independent and form an invertible matrix๐ด๐ต. At the same time the first๐ columns of๐ดx0 form the matrix๐ด๐. By settingx๐ต= (๐ฅ๐+1, . . . , ๐ฅ๐)๐, we have
๐ดx0x= (๐ด๐๐ด๐ต)(x๐๐ x๐๐ต)๐ =๐ด๐x๐+๐ด๐ตx๐ต =๐x0,
and so
x๐ต =๐ดโ1๐ต (๐x0โ๐ด๐x๐) =๐ดโ1๐ต ๐x0 โ๐ดโ1๐ต ๐ด๐x๐=:๐โฒโ๐ดโฒx๐ for allxโ๐x0 with
๐ฅ๐+๐ =๐โฒ๐โ๐ดโฒ๐x๐=๐โฒ๐+
๐
โ๏ธ
๐=1
๐๐+๐,๐๐ฅ๐ (5.3)
for all๐= 1, . . . , ๐โ๐where๐โฒ =๐ดโ1๐ต ๐x0,๐ดโฒ =๐ดโ1๐ต ๐ด๐and๐ดโฒ๐ is the๐-th row of๐ดโฒ,๐๐+๐,๐ are constants for๐= 1, . . . , ๐โ๐and๐ = 1, . . . , ๐. Note that the๐-dimensional affine set๐x0
andR๐are isomorphic. Consider the linear map
๐๐:๐x0 โR๐,xโฆโx๐, which is bijective since the inverse exists:
๐๐โ1:R๐โ๐x0,x๐โฆโ
(๏ธ x๐ ๐โฒโ๐ดโฒx๐
)๏ธ
.
In the following we investigate the graph of function๐ under the restrictionxโ๐นx0. The set๐นx0, function๐ and pointx0can becompletely projected to๐น๐ =๐๐(๐นx0),x0๐ = ๐๐(x0) and๐๐ = ๐(x๐, ๐โฒ1 โ๐ดโฒ1x๐, . . . , ๐โฒ๐ฅโ๐ โ๐ดโฒ๐ฅโ๐x๐) respectively. It is clear that ๐๐ โ R[x๐] is a polynomial function and๐น๐ is a polytope of dimension๐. By comparing coefficients, for everyx โ ๐นx0 it implies that๐(x) = ๐๐(๐๐(x))and for everyx๐ โ ๐น๐it implies that ๐๐(x๐) =๐(๐โ1๐ (x๐)). The graph of๐๐is thus
{(x๐, ๐ง)|๐ง=๐๐(x๐), x๐โ๐น๐}={(x๐, ๐ง)|x๐=๐๐(x), ๐ง=๐(x),xโ๐นx0}=๐ฎ๐, which is the graph of the polynomial function๐๐over polytope๐น๐.
The domain pointx0 is an extreme point of๐ if๐โฒ = |๐ผx0|= ๐. In this case๐ = 0and ๐น๐ ={x0๐}which contains a single point. It implies thatx0๐is also an extreme point of ๐น๐. Otherwise, we are in case of๐โฒ < ๐. Note that every๐น๐is isomorphic to a face of๐. Assume x0๐is not an interior point of๐น๐. Then there exists a face of๐น๐which containsx0๐. It implies there exists a face of๐which containsx0. This is a contradiction to the definition of๐ผx0. Thus
x0๐is an interior point in๐น๐. 2
Example 5.11
We look at an example with๐= 2, i.e.,x= (๐ฅ, ๐ฆ). Consider the polynomial function ๐(๐ฅ, ๐ฆ) =โ๐ฅ4+ 3๐ฅ3+ 10๐ฅ2๐ฆโ5๐ฅ2+ 3๐ฅโ2๐ฆ4โ4๐ฆ2 : [โ4,4]ร[โ4,4]โR. For boundary pointx0 = (1,4)we get๐นx0 ={(๐ฅ, ๐ฆ) | โ4โค๐ฅโค4, ๐ฆ = 4},๐๐ : (๐ฅ, ๐ฆ)โฆโ๐ฅ and๐๐โ1:๐ฅโฆโ(๐ฅ,4)and๐๐(๐ฅ) =โ๐ฅ4+ 3๐ฅ3+ 35๐ฅ2+ 3๐ฅโ576as defined in Lemma 5.10.
Point(x0, ๐(x0)), set๐ฎand a valid hyperplane{(๐ฅ, ๐ฆ, ๐ง)|๐ง= 78๐ฅ+134๐ฆโ1150}are shown as the red point, the blue surface and the yellow plane in Figure 5.4a and the corresponding (x0๐, ๐๐(x0๐))and๐ฎ๐are shown as the red point and the blue curve in Figure 5.4b, respectively.โข
5.3 Convex hull of graphs of polynomial functions over a polytope
(a) In spaceR๐+1 (b) In spaceR๐+1
Figure 5.4:Graph and hyperplane restricted on a face in Example 5.11 and the complete projections
Theorem 5.12 (Globally convex boundary points )
Letx0 โ๐๐. Thenx0is a globally convex domain point of๐ฎ if and only ifx0๐=๐๐(x0)is also a globally convex domain point of๐ฎ๐with
๐ฎ๐={(x๐, ๐ง)|x๐=๐๐(x), ๐ง=๐(x),xโ๐นx0} โR๐+1.
We go back to Example 5.11. Pointx๐0= 1is a globally convex domain point of๐ฎ๐since the tangent plane{(๐ฅ, ๐ง)|๐ง=๐๐(๐ฅ) = 78๐ฅโ614}(shown as green line in Figure 5.4b) can be verified to be valid. Using the theorem above it implies that(1,4)is also globally convex. This is verified by the valid hyperplane{(๐ฅ, ๐ฆ, ๐ง)|๐ง= 78๐ฅ+ 134๐ฆโ1150}.
We prove first the direction โโโ and postpone the direction โโโ after introducing a few further definitions and auxiliary lemmas.
Proof (of โโโ). Consider the case thatx0๐andx0are both an extreme point in ๐น๐and๐, respectively. Then the single pointx0๐โ๐น๐is surely a globally convex domain point since the single point is a valid hyperplane in that space.
Otherwisex0๐is an interior point in๐น๐. Letv= (๐ฃ1, . . . , ๐ฃ๐)๐ โR๐be an arbitrary vector.
Define a function๐๐:R๐โR๐,
where all๐๐๐ are defined as in (5.3). We first give and prove the following claim.
Claim LetnโR๐. The following two sets are affinely isomorphic
{(x, ๐ง)|๐ง=๐ง0+nยท(xโx0),xโ๐x0}โผ={(x๐, ๐ง)|๐ง=๐ง0+๐๐(n)ยท(x๐โx0๐),x๐โR๐}. (5.4)
This implies
{(x, ๐ง)|๐งโฅ๐ง0+nยท(xโx0),xโ๐x0}โผ={(x๐, ๐ง)|๐งโฅ๐ง0+๐๐(n)ยท(x๐โx0๐),x๐โR๐}. (5.5)
Moreover, the tangent plane onx๐0defined as
๐๐(x0๐) ={(x๐, ๐ง)|๐ง=โ๐๐(x0๐)ยท(x๐โx0๐) +๐ง0,x๐โR๐}
satisfies
๐๐(x๐0)โผ=๐|๐
x0(x0) ={(x, ๐ง)โ๐(x0)|xโ๐x0}. (5.6)
5.3 Convex hull of graphs of polynomial functions over a polytope
To prove (5.6), it suffices to show
๐๐(โ๐(x0)) =โ๐๐(x0๐),
Equation (5.4) says that for a given hyperplane๐ป(x0,n), the graph with restrictionxโ๐x0 inR๐+1which is{(x, ๐ง)|๐ง=n(xโx0)โ๐ง0,xโ๐x0}is affinely isomorphic to the hyperplane
๐ป(x0๐, ๐๐(n))inR๐+1. Equation (5.5) translates (5.4) to halfspaces, and (5.6) similarly for tangent hyperplanes.
Now we are ready to prove the remaining part of โโโ. Let๐ป ={(x, ๐ง)|๐ง=๐ง0+n0ยท(xโ x0)}be the hypothesized hyperplane for showing globally convexity. Similar to the proof of the claim above, it implies that
Proof. We use all definitions and results from the proof of โโโ of Theorem 5.12.
Ifx0๐andx0are both an extreme point in๐น๐and๐respectively, then we have ๐๐ผ
x0(x0) ={(x0, ๐(x0))} โ๐ป.
Otherwisex0๐is an interior point in๐น๐. Similar to the proof of Equation (5.4) in the proof of
โโโ of Theorem 5.12, it implies for๐ป =๐ป(x0,n)that
where(*)is followed from Theorem 5.8. 2
5.3 Convex hull of graphs of polynomial functions over a polytope
Recall Example 5.11 again. We have๐๐ผ
x0(x0) ={(๐ฅ, ๐ฆ, ๐ง)|๐ง= 78๐ฅโ614, ๐ฆ = 4}for the globally convex pointx0 = (1,4), see the green line in Figure 5.4. Corollary 5.13 shows that every valid hyperplane through(x0, ๐(x0))contains๐๐ผ
x0(x0). Note that๐๐ผ
x0(x0)is an affine set contained in the tangent plane๐(x0), we call it a subtangent plane. We give a general definition of subtangent planes that are contained in a valid hyperplane.
Definition 5.14 (Subtangent planes)
Letx0 โ๐๐ with corresponding๐ผx0. For every๐ผ โ ๐ผx0, a correspondingsubtangent plane ๐๐ผ(x0)is defined as
๐๐ผ(x0) =๐(x0)โฉ {(x, ๐ง)|๐๐๐ x=๐๐, ๐โ๐ผ}.
A subtangent plane๐๐ผ(x0)is valid if there exists a valid hyperplane๐ป with๐๐ผ(x0) โ ๐ป. Note that๐(x0)may not be valid and yet๐๐ผ(x0)is valid by choosing๐ป ฬธ= ๐(x0). A valid subtangent plane through(x0, ๐(x0))is said to bemaximally valid if there does not exist any other valid subtangent plane๐๐ผโฒ(x0)such that๐๐ผ(x0)(๐๐ผโฒ(x0).
We also extend the definition tox0 โint๐since๐ผx0 =โ and๐โ (x0) =๐(x0).
Note that since|๐ผx0|=๐โฒthere are2๐โฒ such subtangent planes and all of them are affine sets.
For every๐ผ1, ๐ผ2with๐ผ1, ๐ผ2 โ๐ผx0 it implies that
aff{๐๐ผ1(x0), ๐๐ผ2(x0)}=๐๐ผ1โฉ๐ผ2(x0) and it is clear that
๐๐ผ1(x0)โ๐๐ผ2(x0) for๐ผ1 โ๐ผ2 โ๐ผx0.
Corollary 5.13 shows that every valid hyperplane๐ปthrough a boundary point(x0, ๐(x0)) satisfies
๐ป โ๐๐ผ
x0(x0) =(๏ธ๐(x0)โฉ {(x, ๐ง)|xโ๐x0})๏ธโ(๏ธ๐(x0)โฉ {(x, ๐ง)|xโ๐นx0})๏ธ. Face๐นx0 of dimension less than๐โ1is contained in a face of๐of higher dimension. Exam-ple 5.15 below shows that there could exist another face๐นx10 of๐ such that๐นx10 )๐นx0 and there exists also a valid hyperplane๐ป1with๐ป1 โ(๐(x0)โฉ {(x, ๐ง)|xโ๐นx10}).
Example 5.15
Consider the following example inR3. Let๐(๐ฅ, ๐ฆ) =๐ฅ2โ5๐ฅ๐ฆ+๐ฆ2be the polynomial function over domain๐ = {(๐ฅ, ๐ฆ) โ [โ3,10]ร[โ3,10]}, shown in Figure 5.6. The inequalities for the halfspaces of ๐with ๐ฅ โฅ โ3,๐ฆ โฅ โ3,๐ฅ โค 10and๐ฆ โค 10 have corresponding index 1,2,3and4, respectively. The graph of๐ inR3is illustrated in Figure 5.5a. The domain point x0= (โ3,โ3)โ๐is a boundary point with๐ผx0 ={1,2}. Let๐น1 ={(๐ฅ, ๐ฆ)|๐ฅ=โ3,โ3โค ๐ฆ โค10}and๐น2 ={(๐ฅ, ๐ฆ) | โ3 โค๐ฅ โค 10, ๐ฆ = โ3}be two faces of๐ containingx0. The
(a) The graph and graph on face(s) (b)๐{1}(x0)and๐ป1
(c)๐{2}(x0)and๐ป2 (d)๐{1,2}(x0)and๐ป3
Figure 5.5:Example 5.15 for valid subtangent planes and corresponding valid hyperplanes graph of๐ with restrictionx โ ๐น1 is the red curve in Figure 5.5a and the graph of ๐ with restrictionxโ๐น2is the blue curve in Figure 5.5a. In addition, the22 = 4subtangent planes are
๐โ (x0) ={(๐ฅ, ๐ฆ, ๐ง)|๐ง= 9๐ฅ+ 9๐ฆ+ 27}=๐(x0),
๐{1}(x0) =๐(x0)โฉ {(๐ฅ, ๐ฆ, ๐ง)|๐ฅ=โ3}={(๐ฅ, ๐ฆ, ๐ง)|๐ง= 9๐ฆ, ๐ฅ=โ3}, ๐{2}(x0) =๐(x0)โฉ {(๐ฅ, ๐ฆ, ๐ง)|๐ฆ=โ3}={(๐ฅ, ๐ฆ, ๐ง)|๐ง= 9๐ฅ, ๐ฆ=โ3}, ๐{1,2}(x0) =๐(x0)โฉ {(๐ฅ, ๐ฆ, ๐ง)|๐ฅ=โ3, ๐ฆ=โ3}={(โ3,โ3,โ27)}.
Consider the pointx1 = (10,10)with๐(x1) =โ300. In the following, we investigate whether ๐โ (x0)is valid. It is clear that the point(10,10,9ยท10 + 9ยท10 + 27) = (10,10,207)โ๐โ (x0). Note that the point๐1= (x1, ๐(x1)) = (10,10,โ300)is below๐โ (x0)because๐1is below the point(10,10,207). Hence๐โ (x0)is not valid. The subtangent plane๐{1}(x0)(red line
5.3 Convex hull of graphs of polynomial functions over a polytope
โ4 โ2 0 2 4 6 8 10
0 5 10
Figure 5.6:Domain๐ ={(๐ฅ, ๐ฆ)โ[โ3,10]ร[โ3,10]}
in Figure 5.5b and in Figure 5.5d) is valid since the hyperplane๐ป1(green plane in Figure 5.5b) with
๐ป1 = aff{๐{1},{๐1}}={(๐ฅ, ๐ฆ, ๐ง)|๐ง=โ30๐ฅ+ 9๐ฆโ90}
can be verified to be valid and๐{1}(x0)โ๐ป1. Similarly,๐{2}(x0)(blue line in Figure 5.5c and in Figure 5.5d) is valid since the hyperplane๐ป2(green plane in Figure 5.5c) with
๐ป2= aff{๐{2},{๐1}}={(๐ฅ, ๐ฆ, ๐ง)|๐ง= 9๐ฅโ30๐ฆโ90}
can be verified to be valid and๐{2}(x0)โ๐ป2. Let
๐ป3 ={(๐ฅ, ๐ฆ, ๐ง)|๐ง=โ10.5๐ฅโ10.5๐ฆโ90}
with{(โ3,โ3,โ27)} โ๐ป3and{๐1} โ๐ป3. The affine set๐ป3(green plane in Figure 5.5d) can also be verified to be valid. The subtangent plane๐{1,2}(x0)(black point in Figure 5.5d) is valid since๐{1,2}(x0)โ๐ป3.
Note that๐{1}(x0)and๐{2}(x0), both through(x0, ๐(x0)), are maximally valid subtangent planes. Hence a maximally valid subtangent plane does not have to be unique. โข Definition 5.16
Let๐1, ๐2โR๐+1be two sets. For a given domain๐ทโR๐,๐1is said to be(strictly) below๐2 over๐ทif๐x(๐1)โฉ๐x(๐2)โฉ๐ทฬธ=โ and for everyx0โ๐x(๐1)โฉ๐x(๐2)โฉ๐ทand๐ง1, ๐ง2 โR with(x0, ๐ง1)โ๐1and(x0, ๐ง2)โ๐2, we have
๐ง1โค๐ง2(๐ง1 < ๐ง2).
In this case we also call๐2is(strictly) above๐1over๐ท. We omit โover๐ทโ if ๐ทโ(๐x(๐1)โฉ๐x(๐2)).
Figure 5.7:Example of half hyperplanes for Lemma 5.18 and Lemma 5.19
Remark 5.17
In the definition above, the comparison below or above allows the compared sets to have intersection points. Note that for the special case๐1 =๐2, we also have๐1 is above๐2as well as๐1is below๐2.
Note that a hyperplane๐ป is valid if and only if it is below๐ฎ. In the example shown in Figure 5.11a,๐ป๐is below๐ป๐over[๐, ๐ข]for all๐, ๐with1โค๐ < ๐ โค4and๐ป๐is below๐ฎfor all ๐โ {1,2,3,4}. Note that for any๐1, ๐2, ๐3 โR๐+1 with๐1 is below๐2 over๐ท1 โR๐and ๐2is below๐3over๐ท2 โR๐,๐1is below๐3over๐ท1โฉ๐ท2.
Lemma 5.18 (Half hyperplanes) Let
๐ป1=๐ป(x0,n1) ={(x, ๐ง)|๐ง=๐ง0+n1ยท(xโx0)}
with the normal vectorn1 โR๐and
๐ป2=๐ป(x0,n2) ={(x, ๐ง)|๐ง=๐ง0+n2ยท(xโx0)}
with the normal vectorn2 โR๐be two nonvertical and nonparallel hyperplanes with an intersection point(x0, ๐ง0)โ๐ป1โฉ๐ป2. Let
๐x(๐ป1โฉ๐ป2) ={xโR๐|๐ง0+n1ยท(xโx0) =๐ง0+n2ยท(xโx0)}={xโR๐|(n1โn2)ยท(xโx0) = 0}.
5.3 Convex hull of graphs of polynomial functions over a polytope
Then๐ป1is below๐ป2over๐ฟ:={xโR๐|(n1โn2)ยท(xโx0)โค0}.
In addition, for any set๐with๐โ๐ฟor๐โ๐ :={xโR๐|(n1โn2)ยท(xโx0)โฅ0}, either ๐ป1is below๐ป2over๐or๐ป2is below๐ป1over๐.
Proof. For anyxโ๐ฟwe compare(x, ๐ง1)โ๐ป1and(x, ๐ง2)โ๐ป2with
๐ง1โ๐ง2= (๐ง0+n1ยท(xโx0))โ(๐ง0+n2ยท(xโx0)) = (n1โn2)ยท(xโx0)โค0, which means๐ป1is below๐ป2 over๐ฟ.
After that, the second part is trivial to prove. 2
Note that if we write๐x(๐ป1โฉ๐ป2) ={xโR๐|(n2โn1)ยท(xโx0) = 0}, then๐ป1is above ๐ป2over{xโR๐|(n2โn1)ยท(xโx0)โค0}={xโR๐|(n1โn2)ยท(xโx0)โฅ0}=R. An example is shown in Figure 5.7.
Lemma 5.19
Consider the set ๐๐โ1 = {(x, ๐ง) | ๐ง = ๐ง0 +n1 ยท(xโx0),(n1 โn2) ยท(x โx0) = 0}
forn1,n2 โ R๐,n1 ฬธ= n2 and x0 โ ๐, ๐ง0 โ R. Then๐๐โ1 is an affine set of dimension (๐โ1). For any compact set๐ โ R๐+1 with๐x(๐) โ {x | (n1โn2)ยท(xโx0) < 0}or ๐x(๐)โ {x|(n1โn2)ยท(xโx0)>0}, there exists a hyperplane๐ปโ๐๐โ1which is below๐.
Proof. The set๐๐โ1 can be equivalently written as
๐๐โ1 ={(x, ๐ง)|๐ง=๐ง0+n1ยท(xโx0)} โฉ {(x, ๐ง)|๐ง=๐ง0+n2ยท(xโx0)}, which is an intersection of two nonparallel hyperplanes. Hence it is an affine set of dimension (๐โ1).
Note that for any(x, ๐ง) โ๐, we havex ฬธโ ๐x(๐๐โ1). The affine setaff{๐๐โ1,{(x, ๐ง)}}
is then a nonvertical hyperplane. We take an arbitrary fixed pointx๐ โ๐x(๐)as areference point, with an affine set{(x๐, ๐ง)|๐งโR}. For any(x, ๐ง)โ๐,aff{๐๐โ1,{(x, ๐ง)}}has exactly one intersecting point with the affine set{(x๐, ๐ง)|๐ง โR}, see examples in Figure 5.7 with (x, ๐ง) = (x1, ๐ง1)or(x, ๐ง) = (x2, ๐ง2). Denoting the intersection point by(x๐, ๐x๐(x))with
{(x๐, ๐x๐(x))}= aff{๐๐โ1,{(x, ๐ง)โ๐}} โฉ {(x๐, ๐ง)|๐งโR}
and๐x๐ :๐x(๐)โR. Function๐x๐ is continuous since the affine function above is continuous.
Let(x1, ๐ง1) โ ๐ and(x2, ๐ง2) โ ๐. Then the hyperplaneaff{๐๐โ1,{(x1, ๐ง1)}} is below aff{๐๐โ1,{(x2, ๐ง2)}}over๐if and only if๐x๐(x1)โค๐x๐(x2), see an example in Figure 5.7.
Due to the Weierstrass Theorem, we can find(x*, ๐ง*)โ๐such that๐x๐ attains a minimum atx*. Then we haveaff{๐๐โ1,{(x*, ๐ง*)}}is belowaff{๐๐โ1,{(x๐, ๐ง๐)}}for any(x๐, ๐ง๐)โ๐.
This implies thataff{๐๐โ1,{(x*, ๐ง*)}} is below any(x๐, ๐ง๐) โ ๐. Hence the hyperplane
๐ป := aff{๐๐โ1,{(x*, ๐ง*)}}is below๐. 2
Figure 5.8:A point sequence in the neighborhood ofx0
With the following lemma we prove that a hyperplane๐ปis valid if and only if๐ปis below๐ฎ over๐ห๐.
Lemma 5.20
The hyperplane๐ป(x0,n0)is below๐ฎover๐if and only if๐ป(x0,n0)is below๐ฎover๐ห๐, which means for everyxโ๐ห๐it satisfies
๐(x)โฅ๐(x0) +n0ยท(xโx0). (5.8) Proof. Direction โโโ simply follows since๐โ๐ห๐.
For the proof of โโโ, first consider the optimization problem
xโ๐min ๐(x)โ๐(x0)โn0ยท(xโx0), (5.9) which has a continuous objective function over a compact feasible region. Since๐is compact, an optimum always exists due to the Weierstrass Theorem. Due tox0โ๐, we have
minxโ๐ ๐(x)โ๐(x0)โn0ยท(xโx0)โค๐(x0)โ๐(x0)โn0ยท(x0โx0) = 0.
The hyperplane๐ป(x0,n0)is valid if and only if the optimization problem (5.9) has optimal value0. Note that by assumption, everyxโ๐ห๐satisfies (5.8) . The hyperplane๐ป(x0,n0)is valid if the inequality holds also for everyxโ๐โ๐ห๐. We assume there exists anxโ โ๐โ๐ห๐ with
๐(xโ)< ๐(x0) +n0ยท(xโโx0).
Then the optimization problem (5.9) possesses an optimal solution(x*, ๐ง*)with value<0, i.e., ๐(x*)< ๐(x0) +n0ยท(x*โx0).
Note thatx* is a globally convex domain point since๐ป(x*,n0)is valid. This is a contradiction to (5.8) for everyxโ๐ห๐. It implies thenxโdoes not exist and (5.8) is also satisfied for every xโ๐โ๐ห๐. Hence๐ป(x0,n0)is a valid hyperplane. 2
5.3 Convex hull of graphs of polynomial functions over a polytope
Proof idea of direction โโโ of Theorem 5.12. Recall Example 5.11. As shown in Figure 5.4b, for๐ฎ๐,(x0๐, ๐๐(x0๐)) = (1,โ536)is globally convex. To show Theorem 5.12, we need to prove that there exists a valid hyperplane through(x0, ๐(x0)) = (1,4,โ536), which is the red point in Figure 5.4a. According to Corollary 5.13, every such hyperplane should contain๐๐ผ
x0(x0) which is the green line in Figure 5.4a. Note that in this example๐๐ผ
x0(x0)is an affine set of dimension1. According to Lemma 5.19, by setting๐๐โ1 =๐๐ผ
x0(x0), for any subset๐โฒ โ๐ with๐โฒโฉ๐นx0 =โ , there exists a hyperplane๐ป โ๐๐ผ
x0(x0)such that๐ปis below๐ฎover๐โฒ. The proof idea is as follows. After finding a set๐๐โ1as described in Lemma 5.19, we split๐ into two compact sets๐๐ and๐ฮsuch that๐x(๐๐โ1)โ๐๐ and๐ฮโฉ๐x(๐๐โ1) =โ as
The following lemma helps us to find๐๐ and๐ป*as mentioned in the proof idea.
Lemma 5.21
For every x0 โ ๐ห๐, there exists an ๐ > 0 such that ๐(x0) is below ๐ฎ over ๐ห๐ โฉ ๐ต๐(x0). Furthermore, every valid hyperplane๐ปthrough(x0, ๐(x0))is below๐(x0)over๐.
Proof. Note that for anyx0โint๐this lemma is clear due to Lemma 5.5 and Theorem 5.8.
We need only to consider the casex0 โ ๐๐. For a given globally convex domain point x0โ๐๐โฉ๐ห๐, we want to capture the globally convex domain points in its neighborhood.
Let๐ด, ๐ตโR๐+1be two subsets, then
which follows directly from the definition of subtangent planes.
For any๐๐ผ(x0)which is below๐ฎ,๐(x0)is then below๐ฎ over๐ห๐โฉ๐ต๐(x0)โฉ๐x(๐๐ผ(x0)) for any๐ >0. It implies that๐(x0)is below๐ฎover
โ๏ธ
๐ผโ๐ผx0 ๐๐ผ(x0)is below๐ฎ
๐ห๐โฉ๐ต๐(x0)โฉ๐x(๏ธ๐๐ผ(x0))๏ธ.
Note that
๐ห๐โฉ๐ต๐(x0) = โ๏ธ
๐ผโ๐ผx0
๐ห๐โฉ๐ต๐(x0)โฉ๐x
(๏ธ๐๐ผ(x0))๏ธ.
In the following we need only to prove that for a sufficiently small๐ >0it holds
โ๏ธ
๐ผโ๐ผx0 ๐๐ผ(x0)is below๐ฎ
๐ห๐โฉ๐ต๐(x0)โฉ๐x(๏ธ๐๐ผ(x0))๏ธ= โ๏ธ
๐ผโ๐ผx0
๐ห๐โฉ๐ต๐(x0)โฉ๐x(๏ธ๐๐ผ(x0))๏ธ (5.11)
which implies the first consequence that๐(x0)is below๐ฎover๐ห๐โฉ๐ต๐(x0).
Before proving (5.11) we first prove that for everyx0โ๐ห๐there exists an๐๐ฝ >0such that any subtangent plane๐๐ฝ(x0), ๐ฝ โ ๐ผx0 thatdoes not fulfillthe condition๐๐ฝ(x0)is below๐ฎ satisfies
๐ห๐โฉ๐ต๐๐ฝ(x0)โฉrelint(๏ธ๐น๐ฝ(x0))๏ธ=โ . (5.12) Assume that there does not exist an๐๐ฝ for which (5.12) holds. Then there exists๐ฝ โ๐ผx0 such that๐๐ฝ(x0)does not fulfill๐๐ฝ(x0)is below๐ฎand for every๐๐ฝ >0
๐ห๐โฉ๐ต๐๐ฝ(x0)โฉrelint(๏ธ๐น๐ฝ(x0))๏ธฬธ=โ .
Recall that๐น๐ฝ(x0)is a compact set. There exists then a point sequence(x๐)๐โN โ๐ห๐, as shown in Figure 5.8, such that
๐โโlim x๐ =x0and{((x๐)๐โN)} โ(๏ธ๐ห๐โฉ๐ต๐๐ฝ(x0)โฉrelint(๏ธ๐น๐ฝ(x0))๏ธ)๏ธ
for๐๐ฝ >0. The pointx๐for any๐โNis a globally convex domain point and a relative interior point in๐น๐ฝ(x0). Denotingx๐byy, the function
๐บy(x) =๐(x)โ(โ๐(y)ยท(xโy) +๐(y))
evaluates the validity of the tangent plane๐(y) through(y, ๐(y)): Tangent plane ๐(y) is valid if and only if๐บy(x)โฅ0for allxโ๐; subtangent plane๐๐ฝ(y)is below๐ฎif and only if ๐บy(x)โฅ0for allxโ๐น๐ฝ(y). Consider the function
โ:๐น๐ฝ(x0)โR,yโฆโ min
xโ๐น๐ฝ(y)๐บy(x).
5.3 Convex hull of graphs of polynomial functions over a polytope
With โโโ of Theorem 5.12,y โ ๐ห๐ implies that๐๐ฝ(y) is below๐ฎ which is equivalent to ๐(y)is below๐ฎ over๐น๐ฝ(y). Thus,โ(x๐) โฅ0for every๐โN. For anyyfunction๐บy(x)is a continuous function, it impliesโis also continuous since it maps from a compact set to the minimum of๐บy(x). Together withlim๐โโx๐ =x0 it implies thatโ(x0)โฅ 0. This implies that๐๐ฝ(x0)is below๐ฎ. This is a contradiction to our assumption. Hence there exists an๐๐ฝ >0 such that
๐ห๐โฉ๐ต๐๐ฝ(x0)โฉrelint(๏ธ๐น๐ฝ(x0))๏ธ=โ .
Since we have finitely many subtangent planes through (x0, ๐(x0)), there exists an๐ > 0 which is the minimum of all๐๐ฝ such that (5.12) holds for every such subtangent plane through (x0, ๐(x0)).
Using (5.14) iteratively and together with (5.12), it implies that ๐ห๐โฉ๐๐ฝ๐(x0) = โ๏ธ
๐ฝโ๐ผโ๐ผx0
๐๐ผ(x0)is below๐ฎ
๐ห๐โฉ๐๐ผ๐(x0)
for every๐ฝ โ๐ผx0. This implies๐(x0)is below๐ฎover
โ๏ธ
๐ผโ๐ผx0
๐๐ผ(x0)is below๐ฎ
๐ห๐โฉ๐๐ผ๐(x0) = โ๏ธ
๐ผโ๐ผx0
๐ห๐โฉ๐๐ผ๐(x0) = ห๐๐โฉ๐ต๐(x0).
Now it remains only to prove that every valid hyperplane๐ปthrough(x0, ๐(x0))is below ๐(x0)over๐. Assume that๐ป ฬธ= ๐(x0), otherwise we are done. An example for๐ = 1is shown in Figure 5.9. Let๐ = [๐, ๐ข]and๐,๐ฅ๐and๐ขbe globally convex points. Tangent plane ๐(๐), shown as the green line, is not valid. However, the example in Figure 5.9 shows that every valid hyperplane through(๐, ๐(๐))is below๐(๐)over๐; it also shows that every hyperplane through(๐, ๐(๐))which is above๐(๐)over๐cannot be valid. For interior domain point๐ฅ๐, there does not exist another hyperplane through(๐ฅ๐, ๐(๐ฅ๐))which is below๐(๐ฅ๐), thus there exists at most one valid hyperplane through(๐ฅ๐, ๐(๐ฅ๐))which is also implied by Theorem 5.8.
For domain point๐ข,๐(๐ข)is valid. In addition, every hyperplane through(๐ข, ๐(๐ข))which is below๐(๐ข)over๐is also valid; every hyperplane through(๐ข, ๐(๐ข))which is not๐(๐ข)and above๐(๐ข)over๐is not valid.
We go back to the proof for general๐. Under the assumption๐ป ฬธ=๐(x0), hyperplanes๐ป and๐(x0)are not parallel due to the intersection point(x0, ๐(x0)). Similar to the notation in Lemma 5.18, denote
๐x(๐ปโฉ๐(x0)) ={xโR๐|n0ยท(xโx0) = 0}, and
๐ฟ:={xโR๐|n0ยท(xโx0)โค0}, ๐ :={xโR๐|n0ยท(xโx0)โฅ0}
withn0 โR๐. Note that sincex0is a boundary point of๐, we have either๐x(๐ปโฉ๐(x0))โฉ int๐ ฬธ= โ or๐x(๐ปโฉ๐(x0))โฉint๐ = โ . If๐x(๐ป โฉ๐(x0))โฉint๐ ฬธ= โ , then we have ๐ฟโฉint๐ ฬธ=โ and๐ โฉint๐ฬธ=โ . Lemma 5.18 implies๐ปis above๐(x0)over๐ฟor๐ . Without loss of generality,๐ปis above๐(x0)over๐ฟ. Consider the graph of๐over๐ฟโฉ๐ต๐(x0)โฉint๐ฬธ=โ for any๐ >0. There always exists a pointx1 โ๐ฟโฉ๐ต๐(x0)โฉint๐such that(x1, ๐(x1))is below๐ป. Thus๐ปcannot be valid if๐x(๐ปโฉ๐(x0))โฉint๐ ฬธ=โ .
Now we consider the case๐x(๐ปโฉ๐(x0))โฉint๐=โ . There are two cases, either๐ โ๐ฟor ๐โ๐ . Without loss of generality, let๐โ๐ฟ. Lemma 5.18 implies that๐ปis either below๐(x0)
Now we consider the case๐x(๐ปโฉ๐(x0))โฉint๐=โ . There are two cases, either๐ โ๐ฟor ๐โ๐ . Without loss of generality, let๐โ๐ฟ. Lemma 5.18 implies that๐ปis either below๐(x0)