AMS 345/CSE 355 (Spring, 2006) Joe Mitchell

COMPUTATIONAL GEOMETRY
Homework Set # 6 – Solution Notes
(1). Let S be a set of n points in the plane in general position (no three are collinear). Let h denote the
number of points of S that are vertices of the convex hull, CH(S).
(a). Consider the Delaunay triangulation, D(S), of the convex hull of S. As a function of n and h how
many triangles does D(S) have? How many Delaunay edges are there in D(S)?
Actually, the fact that we are dealing with a Delaunay triangulation is irrelevant for this question: the
combinatorics depend solely on the input, T , being a triangulation. Let t denote the number of triangles in
T . There is one face at infinity, so, when we view T as a planar graph, we know it has t + 1 faces, t of which
are 3-sided and one of which is h-sided. Thus, the sum of the degrees of the faces is 3t + h, which we know
is twice the number of edges; thus, 2e = 3t + h. We also know from Euler’s formula that (t + 1) − e + n = 2.
Putting e = (3t + h)/2 into this, we get (t + 1) − (3t + h)/2 + n = 2. We solve for t to get t = 2n − h − 2 for
the number of triangles. Then, e = (3t + h)/2 = 3n − h − 3 is the number of edges.
(b). Now we are interested in decomposing the convex hull of S into “quadrangles” (4-sided polygons
(quadrilaterals)), such that each point of S is a vertex of some quadrangle.
(i). It turns out that it is not always possible to “quadrangulate” the convex hull. Show that for any set S of
n sites, if h is odd, then there is no quadrangulation of the convex hull of S.
(ii). It turns out that if h is even, there is always a quadrangulation of S. (Extra Challenge: Think about
how you might prove this (no proof is required for the assignment, though).) Suppose now that S is such that
h is even. As a function of n and h, how many quadrangles must be in any quadrangulation of the convex
hull of S? How many edges are there in such a quadrangulation?
Let q denote the number of quadrangles in a quadrangulation of the convex hull. There is one face at
infinity, so, when we view the quadrangulation as a planar graph, we know it has q + 1 faces, q of which are
4-sided and one of which is h-sided. Thus, the sum of the degrees of the faces is 4q + h, which we know is
twice the number of edges; thus, 2e = 4q + h. We also know from Euler’s formula that (q + 1) − e + n = 2.
Putting e = (4q + h)/2 into this, we get (q + 1) − (4q + h)/2 + n = 2. We solve for q to get q = n − (h/2) − 1
for the number of quadrangles. Then, e = (4q + h)/2 = 2n − (h/2) − 2 is the number of edges.
We conclude also that if h is odd, there can be no quadrangulation, since it would have a number of edges
e and a number of quadrangles q that is not integer! (Note the “h/2” term in the formula for q and for e.)
(2). O’Rourke, problem 5, section 5.5.6, page 178. In the figure below, there are four sites, {a, b, c, d} and
two Voronoi vertices, p at the center of Circle(a, b, c) and q at the center of Circle(a, c, d). Now, f (p) is the
radius of Circle(a, b, c). The function f (·) increases monotonically along the segment pq; f (p) is therefore
not a local maximum. (And p is clearly strictly interior to the convex hull of the four sites.)
a

d p
q

c
(3). O’Rourke, problem 1, section 5.6.1, page 181. The example below shows the medial axis (in red) of a
nonconvex pentagon. The medial axis has a curved segment, representing the bisector between the vertex a
and the line through cd. The locus of points that are equidistant to a point and a line is a parabola having
focus at the point (in our case, vertex a) and directrix the line (in our case, through cd).

e b

c
d
(4). Let S be the set of points {(-1,-1), (3,-1), (4,-1), (1,1), (5,1), (6,1), (3,3), (4,3), 6,3)}. (In HW5 you
constructed the Delaunay diagram for these same points.)
(a). Draw the (directed) NNG for S. The NNG is drawn with green (directed) edges below, left. The
solid red edges are the Delaunay edges.
(b). Draw the relative neighborhood graph (RNG) for S. (See O’Rourke, problem 7, section 5.5.6, page
178, for the definition of the RNG.) The RNG is drawn with magenta edges below, middle. I also show with
dashes the pairs of circles that define the lunes for each edge of the RNG, so one can verify that the interior
of the lunes are empty.

(c). Construct the MST by Kruskal’s algorithm applied to the Delaunay diagram. I draw the two possible
MSTs in green below. Note that an MST need not be unique! Which one you get when running Kruskal’s
algorithm depends on how you break ties (two edges of the same length, either of which could be added to
the solution). Since I did not specify exactly how to break ties, either solution is fine.
(d). Construct the approximate TSP tour obtained by doubling the MST and shortcutting, as in Fig-
ure 5.18. (Begin the walk around the MST at point (4,-1) and go clockwise around the MST, as done in the
example in the text.) Also determine the optimal TSP tour (easy to do in this case – can you justify your
answer?). Depending on which of the two MST’s you use, there are two possible answers to the approxi-
mate TSP question; I show each tour below in magenta. Note that these tours are NOT optimal; they are
guaranteed to be within a factor 2 of optimal, though (i.e., each will have length at most twice that of an
optimal tour).

The optimal TSP tour is shown below. Justification: All of the points that appear on the boundary of
the convex hull must appear in any optimal tour in the same order they appear on the boundary of the
hull. (Can you prove this fact? If the points did not appear in the same order as on the hull, then there
will necessarily be a crossing in the tour (do you see why?), which cannot be optimal. (Any crossing can be
“un-crossed” and lead to a shorter tour, by the triangle inequality.)) Thus, the only issue is where to insert
the point (5,1) into the tour. It can be inserted in between any pair of consecutive points on the boundary
of the hull. Testing each possibility (there are only 8), we find that it is best to insert it between (6,1) and
(4,-1). (It is easy to test each possibility! Just look at the difference in length between the detour to (5,1)
and the original length of the edge that the detour replaces.)

(e). Construct the “furthest-point Voronoi diagram” of S. Read problem 11, section 5.5.6, and see
Figure 5.19 for the definition of the diagram. (I told you that it is OK to build the diagram for only the
first 5 points of S: {(-1,-1), (3,-1), (4,-1), (1,1), (5,1)}. I show the diagram for all 9 points as well as the
diagram for just the 5 points. (Either one is acceptable.) I label each cell with its owner, using “C1” to
denote the cell owned by site 1, etc. Note that only the points that serve as vertices of the convex hull are
sites that “own” a nonempty cell. Also, I show the circle centered at the Voronoi vertices that pass through
the corresponding 3 sites: Note that for furthest-point Voronoi, we have the “full circle” property instead of
the “empty circle” property: All sites lie inside such a circle.
 C3

C6

C1
7 8 9

4 5 6

3
1 2

C9

C7
 C3

C1

4 5

3
1 2
C5

C4