You are on page 1of 5

Approximate Subdifferential of PLQ Functions

in O(log n) time
Deepak Kumar
December 16, 2016

1
1 Algorithmic Computation of -subdifferential
In this article, we present an algorithm, that we named plq epssub, to com-
pute the approximate subdifferential of an univariate convex Piecewise Linear
Quadratic(PLQ) Function at a particular x dom(f ) and  > 0. The idea
here is to use the Binary search algorithm to reduce time complexity for do-
ing so. For a PLQ function f : R R{, +} with N number of pieces,
we start with locating the index of x in f ,let say i, using binary search. After
computation of index i, we compute equation l(s) as l(s) = sx +  f (x). In-
voking binary search to Sandwich the piece of f where intersection is with l(s)

2
2 Solution for f (s) and l(s)
Let f (x) be a proper one dimensional convex PLQ function represented as
f (x) = ax2 + bx + c, where a, b, c R, f (s) be the convex conjugate of f (x)
defined in terms of supremum by f (s) = sup{< s, x > f (x) | x dom(f )}
,and l(s) =  f (x)+ < s, x >.
Since,
f (s) = sup{< s, x > f (x)}
for one dimensional case,

f (s) = sup{sx f (x)}


= sup{sx (ax2 + bx + c)}

f (s) = {sx (ax2 + bx + c)} (taking on both sides)
= s 2ax b

by equating f (s) = 0, and solving for x, we get

s 2ax b = 0
2ax = s b
sb
x =
2a

sb
= x = (1)
2a

3
by using the value of x from equation 1, we get
   
sb sb
f (s) = s f
2a 2a
  "  2   #
sb sb sb
= s a +b +c
2a 2a 2a
s2 bs
 2
s 2bs + b2 bs b2

= + +c
2a 4a 2a
s2 bs
 2
s 2bs + b2 + 2bs 2b2 + 4ac

=
2a 4a
2 2 2
2s 2bs (s b + 4ac)
=
4a
2s 2bs s2 + b2 4ac
2
=
4a
s2 + b2 2bs 4ac
=
4a

s2 + b2 2bs 4ac
f (s) = (2)
4a
To find solution between f (s) and l(s), we equate eq. 2 with l(s)

s2 + b2 2bs 4ac
= sx +  f (x)
4a
s2 + b2 2bs 4ac = 4a(sx +  f (x))
s2 s(2b + 4ax) + b2 + 4af (x) 4ac 4a = 0

by solving the above equation using quadratic formula, we get


p
(2a + 4ax) (2b + 4ax)2 4(b2 + 4af (x) 4ac 4a)
s=
2
This expression holds true when intersecting piece in f ,let say is , is
Quadratic, while if it is linear, intersection occurs at subdifferentials of the
end points of that piece. This is due to the fact that linear region in pri-
mal domain maps to a point in dual domain, hence for every x f (x) such

4
that f (x)|x[xis 1 ,xis ] is affine. Hence, to compute solution for linear case, we
equate f (s) to l(s),

sx f (x) = sx +  f (x)
s(x x) =  + f (x) f (x)
 + f (x) f (x)
s =
x x
The case x = x is not possible as it implies that  = 0. Now since we have
expression for computing the solution for linear case, we can compute the
differential value at xis 1 and xis and check for feasibility among the two
solutions.

You might also like