(a) The graph of a polynomial function and a piece of the boundary (b) A subgraph on the boundary
Figure 5.1:Investigating the boundary of the graph of a polynomial function
are presented, respectively. In this chapter, we focus on the characteristics of the convex hull of graphs of general polynomial functions over a polytope.
5.2 Basic ideas of this chapter
Consider the following example in R3. Let ๐(๐ฅ, ๐ฆ) = ๐ฅ2โ 5๐ฅ๐ฆ + ๐ฆ2be a polynomial function
over the domain ๐ = {(๐ฅ, ๐ฆ) โ [โ3, 10] ร [โ3, 10]}. For the constraint ๐ง = ๐ฅ2โ 5๐ฅ๐ฆ + ๐ฆ2,
the feasible region, denoted by ๐ฎ, is shown in Figure 5.1a which corresponds to the graph of ๐ over domain ๐. Recalling our MINLP algorithms, we add linear constraints to strengthen the LP relaxation. For any hyperplane ๐ป in R3defined in the form {(๐ฅ, ๐ฆ, ๐ง) | ๐ง = ๐๐ฅ + ๐๐ฆ + ๐}
with constants ๐, ๐, ๐ โ R, ๐ป is said to be a linear underestimator to ๐ over ๐ if ๐ฎ = {(๐ฅ, ๐ฆ, ๐ง) | ๐ง = ๐ (๐ฅ, ๐ฆ), (๐ฅ, ๐ฆ) โ ๐} โ {(๐ฅ, ๐ฆ, ๐ง) | ๐ง โฅ ๐๐ฅ + ๐๐ฆ + ๐}
โ โ
downward closed halfspace to ๐ป
.
Graphically, it means that the corresponding downward closed halfspace completely contains the graph of ๐ over ๐.
In contrast to general MINLP algorithms, we want to find such linear underestimators directly. They are expected to strengthen the LP relaxation. The intuition is that we only want to consider such hyperplanes ๐ป that support the graph, otherwise we can move it upwardly until the new generated hyperplane intersects the graph.
In other words, we say a linear underestimators ๐ป is below (see Definition 5.16) the graph ๐ฎ. Thus ๐ป is said to be valid.
(a) The view of the graph and a linear underestimator (b) The view from the other side
Figure 5.2:A linear underestimator which supports two boundary points of the graph of a polynomial function
To find linear underestimators ๐ป, we study the intersection points ๐ป โฉ๐ฎ. After a series of pre- liminary definitions in Section 5.3.1, we define locally and globally convex points in Section 5.3.2. A point (๐ฅ0, ๐ฆ0, ๐ง0)on the graph is said to be locally convex if there exists ๐ป โ (๐ฅ0, ๐ฆ0, ๐ง0)
and ๐ป is below the graph of ๐ over a neighborhood of (๐ฅ0, ๐ฆ0, ๐ง0). A point (๐ฅ1, ๐ฆ1, ๐ง1)on the
graph is said to be globally convex if there exists ๐ป โ (๐ฅ1, ๐ฆ1, ๐ง1)and ๐ป is below ๐ฎ.
The hyperplane ๐ป๐ก = {(๐ฅ, ๐ฆ, ๐ง) | ๐ง = 9๐ฅ โ 30๐ฆ โ 90}, shown as the yellow hyperplane
in Figure 5.2, can be verified to be a linear underestimator for ๐ over ๐. The hyperplane
๐ป๐กintersects ๐ฎ in two points (โ3, โ3, โ27) and (10, 10, โ300). Hence (โ3, โ3, โ27) and
(10, 10, โ300)both are globally convex points. Note that they both are boundary points of ๐ฎ. Consider further a point (๐ฅ0, ๐ฆ0, ๐ง0)such that the corresponding domain point (๐ฅ0, ๐ฆ0)is an
interior point of ๐. As we will show in Section 5.3.2, to check if (๐ฅ0, ๐ฆ0, ๐ง0)is globally convex,
we need only to check if the corresponding tangent plane is below ๐ฎ. However, in practice, it is quite hard to find those globally convex points such that the corresponding domain points are interior points of ๐. In addition, the property of global convexity usually depends on the domain. On the one hand, any locally convex point may become globally convex if the domain size is small enough. On the other hand, a globally convex point with respect to the current domain could be only locally convex for a larger domain. Note that in the example above ๐ is neither convex nor concave over ๐.
Now we move our attention to those globally convex points for which the corresponding domain points are on the boundary of ๐. Consider the subgraph with restriction ๐ฆ = โ3, which is presented as
5.2 Basic ideas of this chapter
(a) Subtangent plane (b) Globally convex boundary point in R2
Figure 5.3:Example for a globally convex boundary point
This subgraph is shown as the red curve in Figure 5.1a. Since ๐ฆ = โ3 is satisfied for any point in the red subgraph, after projecting the space {(๐ฅ, โ3, ๐ง)} โ R3to the space {(๐ฅ, ๐ง)} โ R2,
we get an isomorphic two-dimensional curve in R2
{(๐ฅ, ๐ง) | ๐ง = ๐ (๐ฅ, โ3) = ๐ฅ2+ 15๐ฅ + 9 =: ห๐ (๐ฅ), โ3 โค ๐ฅ โค 10}.
In general, we show at the beginning of Section 5.3.3 that certain subgraphs on the boundary can be projected to an isomorphic graph in a space with lower dimension. The one-dimensional curve is shown as the red curve in Figure 5.1b. Note that the corresponding function ห๐ is a
univariate function. Fortunately, the study of the convexity of univariate functions is much easier than that for bivariate functions. In the example ห๐ has domain [โ3, 10] and is a convex
function over [โ3, 10].
According to the definition of globally convex points, any point ๐ฅ* โ [โ3, 10]in Figure 5.1b
is globally convex in the projected space R2. Theorem 5.12 implies that for any such ๐ฅ*, because
๐ฅ*is globally convex in the projected space, the boundary point (๐ฅ*, โ3)in ๐ is also globally
convex in the original space. This means there exists a hyperplane ๐ป โ (๐ฅ*, โ3, ๐ (๐ฅ*, โ3))and
๐ปis below ๐ฎ. Consider the case ๐ฅ*= 0. Then (0, โ3, 9) is a globally convex point. Figure 5.3b shows that in the projected space R2, the tangent plane, shown as the green line, is the unique
underestimator. The corresponding line in R3, shown as the green line in Figure 5.3a, is then
{(๐ฅ, โ3, 15๐ฅ + 9) | ๐ฅ โ R} which is defined as subtangent plane in Section 5.3.3. Corollary 5.13 implies that every linear underestimator ๐ป with ๐ป โ (0, โ3, 9) satisfies
๐ป โ {(๐ฅ, โ3, 15๐ฅ + 9) | ๐ฅ โ R},
The blue line {(๐ฅ, โ3, 9๐ฅ) | ๐ฅ โ R} in Figure 5.2a is a subtangent plane on (โ3, โ3, โ27), as defined in Section 5.3.3. We can verify that the yellow hyperplane is the affine hull of the blue line and the point (10, 10, โ300), i.e.,
๐ป๐ก= {(๐ฅ, ๐ฆ, ๐ง) | ๐ง = 9๐ฅ โ 30๐ฆ โ 90} = aff {{(๐ฅ, โ3, 9๐ฅ) | ๐ฅ โ R}, {(10, 10, โ300)}} . For any point (10, 10, ๐ง1)with ๐ง1< โ300, we can also verify that
๐ป๐= aff{๏ธ{(๐ฅ, โ3, 9๐ฅ) | ๐ฅ โ R}, {(10, 10, ๐ง1)}}๏ธ is also a linear underestimator. By comparing ๐ป๐กand ๐ป๐we have
๐ป๐กโฉ ๐ฎ = {(โ3, โ3, โ27), (10, 10, โ300)} ) {(โ3, โ3, โ27)} = ๐ป๐โฉ ๐ฎ.
From the intuition, we prefer ๐ป๐กsince the resulting relaxation is tighter. For this purpose
we define tight and loose hyperplanes in Section 5.3.4. In general, a valid hyperplane ๐ป๐ is
definitely loose if there exists another valid hyperplane ๐ป๐กwhich preserves all intersection
points and intersects in additional point(s) with ๐ฎ, which means (๐ป๐กโฉ ๐ฎ) ) (๐ป๐โฉ ๐ฎ).
This is a sufficient but not necessary condition for loose hyperplanes. Using Lemma 5.26 in Section 5.3.4 we verify that the yellow hyperplane in Figure 5.2a is a tight hyperplane.
After that, in Section 5.3.5 we prove for every loose hyperplane ๐ป๐that there exists a tight
hyperplane ๐ป๐กthat preserves intersection points with
(๐ป๐กโฉ ๐ฎ) โ (๐ป๐โฉ ๐ฎ).
We call the corresponding halfspaces tight or loose halfspaces. Note that in the example above we have ๐ป๐กโฉ ๐ป๐ = {(๐ฅ, โ3, 9๐ฅ) | ๐ฅ โ R} which is the blue line in Figure 5.2a. Graphically,
we can rotate ๐ป๐around the blue line as axis to generate ๐ป๐ก. The rotation approach is the basic
idea of a few proofs in this section.
Finally, in Section 5.3.6, we prove that to form the convex hull of ๐ฎ using halfspaces, we only need tight hyperplanes. In other words, any loose hyperplane is proved to be redundant.
In Section 5.3 we only include theoretical results. We cannot use them to solve MINLP directly. In Section 5.4 we develop algorithms to compute tight hyperplanes for the graph of bivariate polynomial functions with degree up to 3 over a polygon in R2. Note that the
domain does not have to be box-constrained. In the algorithms, we first find all globally convex domain points on the boundary. This is very tractable since we only need to find globally convex points in the graph of univariate polynomial functions with degree 3 over a closed interval in R. Based on those globally convex domain points, the algorithms find a series of tight halfspaces. Computations in Section 5.5 show that these tight halfspaces improve the dual bounds significantly.