lecture ppt

advertisement
Convex Hulls (2D)
-Jarvis March
- Graham Scan
- Chan’s Algorithm
What is a convex hull?
• Wikipedia: “the minimal convex set containing
X, where X is a set of points in a real vector
space V”
• In 2D: Convex polygon of smallest area
containing all given points
• In 3D: Convex polyhedron of smallest volume
containing all given points
Output-sensitive algorithms
• Any algorithm with a runtime dependent not
only on the input size, but on the output size
as well
• Occasionally pop up in problems where you
can compute parts of the output one by one
Jarvis march
• Start at the leftmost point in the set (if there
are multiple, take the closest to the bottom)
• For all other points, compute the angle
measured ccw with down being 0o
• Take point with smallest angle
• Repeat, except for further steps, measure
angles ccw with the line segment from current
to previously selected point being 0o
Jarvis march
Can we do better?
• Jarvis performs reasonably well if output size
is small
– Calculate angles for every point each time it
progresses to a new point, and it progressed to a
new point once for every point in the output
– Runtime O(nh), where n = # of points, h = # of
output points.
• If k is large (read: O(lg n)), Jarvis is slow, can be
as bad as O(n2)
Graham scan
• Take starting point to be bottom-most point (if
tied, take left-most)
• Sort all points by angle with right (+x) being 0o
• Iterate over points in order
– If the next point forms a “left turn” with the
previous two, add it to the list and continue
– If “right turn”, remove the previous point until the
angle becomes a left turn
Graham scan
Can we do even better?
Graham scan
O(n lg n)
Jarvis march
O(nh)
O(n lg h) ?
Yup.
• The “ultimate convex hull algorithm”
published by Kirkpatrick and Seidel in 1986
achieves O(n lg h)
– Uses reversal of divide-and-conquer, referred to as
“marriage-before-conquest” by authors
– Requires median-finding in O(n)
• Chan’s algorithm, published in 1996, also
achieves O(n lg h), but more simple
– As a lazy CS major, I’ll use this
Chan’s Algorithm
• Uses both Graham scan and Jarvis march as
subroutines
• Partitions points into smaller chunks, then
derives an overall convex hull by exploiting the
smaller precomputed hulls
Chan’s Algorithm
• Suppose we know h, the size of the output
• We can partition (arbitrarily) n points into
1+n/h subsets, each with h or h-1 points
• Use Graham scan on each of subsets; must be
done 1 + n/h times, and each takes O(h lg h),
so overall runtime = (1+n/h)O(h lg h) = O(n lg h)
• Finish computing convex hull from parts…
Chan’s Algorithm
Chan’s Algorithm
• Problem: In general, we don’t know the size of
the convex hull before we compute, so h is
unknown; thus, this algorithm will not work
• Instead, iterate over the algorithm with
different guesses for h to converge to the
answer
Chan’s Algorithm
t
2
min(2 ,
• For step t, let the guess for h be m =
n)
• If hull is not completed in m steps, abort
• Thus, each step takes O(n lg m) = O(n∙2t)
• Starting t at 0, we are guaranteed to find the
solution by t =
• Thus, total runtime is
Chan’s Algorithm
• Looks complicated
• But, we know
• Thus,
• So we have a working O(n lg h) algorithm, yay
Other optimizations
• Discard points found to be inside smaller
convex hulls prior to merge step on each
iteration of algorithm (if a point lies in any
convex polygon, it cannot be in convex hull)
t
2
• Alternatives to 2
Download