A Linear-Time Algorithm
for Concave One-Dimensional Dynamic Programming
Zvi Galil
Kunsoo Park
CUCS-469-89
A Linear-Time Algorithm for Concave One-Dimensional Dynamic Programming*
Zvj Galijl.2 and Kunsoo Park 1
1 Department of Computer Science, Columbia University, New York, KY 10027
2 Department of Computer Science, Tel-Aviv University, Tel-Aviv, Israel
Keywords: Dynamic programming, quadrangle inequality, total monotonicity
The one-dimensional dynamic programming problem is defined as follows: given a real-valued
function w(i,j) for integers 0 ~ i ~ j ~ nand E[O], compute
E(j] = mjn{D[i]
O~I<)
+ u'(i,j)}, for 1 ~ j ~ n,
where D[i] is computed from E[i] in constant time. The least weight subsequence problem [4] is
a special case of the problem where D[i] = E[i]. The modified edit distance problem [3], which
arises in molecular biology. geology, and speech recognition, can be decomposed into 2n copies of
the problem.
Let ,-1 be an n X m matrix. A[i, j] denotes the element in the ith row and the jth column.
A[i: i' ,j : i'] denotes the submatrix of A that is the intersection of rows i, i + 1, ... , i' and columns
j,j + 1, ... ,i. We say that the cost function w is concave if it satisfies the quadrangle inequality [7]
w(a,c) + w(b,d) ~ w(b,c) + w(a,d), for a ~ b ~ c ~ d.
In the concave one-dimensional dynamic programming problem w is concave as defined above. A
condition closely related to the quadrangle inequality was introduced by Aggarwal et al. [1] An
n X m matrix A is totally monotone if for all a < band c < d,
A[a, c] > A[b, c] ~ A[a, d] > A[b, d].
Let rU) be the smallest row index such that A[r(j), j] is the minimum value in column j. Then
total monotonicity implies
r(1) ~ r(2) ~ ... ~ rem). (1)
That is, the minimum row indices are nondecreasing. \Ve say that an element A[i. j] is dead if
i ~ r(j). A submatrix of A. is dead if all of its elements are dead. Note that for a ~ b ~ c ~ d, the
quadrangle inequality implies total monotonicity. but the converse is not true. Aggarwal et al. [1]
show that the row maxima of a totally monotone n X m matrix A can be found in O(n + m) time
if A[i,j] for any i,j can be computed in constant time. Their algorithm is easily adapted to find
the column minima. We will refer to their algorithm as the SMAWK algorithm.
Let B[i,j] = D[i] + w(i,j) for 0 ~ i < j ~ n. We say that B[i,j] is available if D[i] is known
and therefore B[i,j] can be computed in constant time. Then the problem is to find the column
minima in the upper triangular matrix B with the restriction that B[i, j] is available only after
the column minima for columns 1. 2, ... , i have been found. It is easy to see that when w satisfies
the quadrangle inequaHty, B also satisfies the quadrangle inequaHty. For the concave problem
Hirschberg and Larmore [4] and later Galil and Giancarlo [3] gave O( n log n) algorithms using
queues. Wilber [6] proposed an O(n) time algorithm when D[i] = E[i]. However, his algorithm
does not work if the availability of matrix B must be obeyed, which happens when many copies
* Work supported in part by NSF Grants CCR-86-0.5353 and CCR-88-14977
2
1 e} p n
or-----------~--------~
Figure 1. ~Iatrix B at a typical iteration
of the problem proceed simultaneously (i.e., the computation is interleaved among many copies)
as in the modified edit distance problem [3] and the mixed convex and concave cost problem [2].
Eppstein [2] extended Wilber's algorithm for interleaved computation. Our algorithm is more
general than Eppstein's; it works for any totally monotone matrix B (we use only relation (1»,
whereas Eppstein's algorithm works only when B[i,j] = D[i] + w(i,j). Our algorithm is also
simpler than both Wilber's and Eppstein ·s. Recently, Larmore and Schieber [5] reported another
linear-time algorithm. which is quite different from ours.
The algorithm consists of a sequence of iterations. Figure 1 shows a typical iteration. \Ve use
NUl, 1 :S j :S n. to store interim column minima before row r; N[j] = B[i,j] for some i < r (the
usage will be clear shortly). At the beginning of each iteration the following invariants hold:
(a) O::S rand r < e.
(b) E[j] for all 1 :S j < e have been found.
(c) E[j] for j ~ e is min(.NU],mini~r B[i,j]).
Invariant (b) means that D[i] for all 0 :S i < c are known, and therefore B[i,j] for 0 :S i < e and
c ::S j :S n is available. Initially, r = O. e = 1, and all }VU] are +00.
Let p = rnin(2c - r, n), and let G be the union of N[c : p] and B[r : e - 1, c : pl, N[c : p]
as its first row and B[r : c - 1, e : p] as the other rows. G is a (c - r + 1) x (e - r + 1) matrix
unless 2c - r > n. Let F[jl, c:S j :S p, denote the column minima of G. Since matrix G is totally
monotone, we use the SMAWK algorithm to find the column minima of G. Once F[c : p] are found,
we compute E[j] for j = c, c + 1, ... as follows. Obviously, E[e] = F[c]. For c + 1 :S j :S p, assume
inductively that B[c : j - 2,j : p] (/3 in Figure 1) is dead and B[j - 1,j : n] is available. It is
trivially true when j = c + 1. By the assumption E[j] = min(F[j], BU - 1, j]).
(1) If E[j - l,j] < FU], then E[j] = B[j - 1,j], and by relation (1) E[r: j - 2,j: n] (a, ,8,;,
and the part of G above /3 in Figure 1) and S[j : n] are dead. We start a new iteration with
c = j + 1 and r = j - 1.
(:2) If F[j] $ B[j - I,j], then E[j] = F[j]. We compare E[j - l,p] with FlP].
(2.1) If B[j - l.p] < F[p], B[r : j - 2,p + 1 : n] (a and; in Figure 1) is dead by relation (1).
B[c : j - 2.j : p] (/3 in Figure 1) is dead by the assumption. Thus only F[j + 1 : p] among
B[O : j - 2.j + 1 : n] may become column minima in the future computation. We store
F[j + 1 : p] in N[j + 1 : p] and start a new iteration with c = j + 1 and r = j - 1.
(2.2) If F(p] :S B[j - l,p], B[j - l,j : p] (6 in Figure 1) is dead by relation (1) in submatrix
B[r : j - 1. j : p] (/3. 6, and the part of G above /3). Since B[j,j + 1 : n] is available from
3
procedure concave 1D
c +- 1;
r +- 0;
N[l: n] +- +00;
while c ~ n do
p +- min(2c - r, n):
use S)'IAWK to find column minima F[e: p] of G;
E[e] +- F[e];
for j +- e + 1 to p do
if B[j - 1,j] < F[j] then
E[j] +- B(j - 1. j];
break
else
E[j] - F(j];
if B(j - 1. p] < F[p] then
N[j + 1: p] .- F[j + l:p];
break
end if
end if
end for
if j ~ p then
e.- j + 1;
r +- j - 1
else
e - p + 1;
r +- max(r, row of F[pJ)
end if
end while
end
Figure 2. The algorithm for concave 1D dynamic programming
E(j], the assumption holds at j + 1. \Vego on to column j + 1.
If case (2.2) is repeated until j = p, we have found E[j] for c ~ j ~ p. We start a new iteration
with e = p + 1. If the row of F(P] is greater than r, it becomes the new r (it may be smaller than
r if it is the row of N(pJ). Note that the three invariants hold at the beginning of new iterations.
Figure 2 shows the algorithm, where the break statement causes the innermost encl06ing loop to
be exited immediately.
Each iteration takes time O(c - r). If either case (1) or case (2.1) happens, we charge the time
to rows r, ... , c - 1 because r is increased by (j - 1) - r ~ c - r. If case (2.2) is repeated until j = p,
there are two cases. If p < n, we charge the time to columns e .... , p because c is increased by
(p+ 1) - c ~ c - r + 1. If p = n, we have finished the whole computation, and rows r, . .. , c - 1( < n)
have not been charged yet; we charge the time to the rows. Since c and r never decrease, only
constant time is charged to each row or column. Thus the total time of the algorithm is linear in
n.
4
References
[1] Aggarwal, A., Klawe, M. M., Moran, S., ShOT, P., and Wilber, R. Geometric applications of a
matrix-searching algorithm. Algorithmica 2 (1987), 195-208.
[2] Eppstein, D. Sequence comparison with mixed convex and concave costs. J. Algorithms to
appear.
[3] Galil, Z., and Giancarlo, R. Speeding up dynamic programming with applications to molecular
biology. Theoretical Computer Science 64 (1989), 107-118.
[-1] Hirschberg, D. S., and Larmore, L. L. The least weight subsequence problem. SIAM J. Comput.
16, 4 (1987), 628-638.
[5] Larmore, L. L., and Schieber, B. On-line dynamic programming with applications to the predic-
tion of RNA secondary structure. to be presented at the First Annual ACM-SIAAf Symposium
on Discrete Algorithms.
[6] Wilber, R. The concave least-weight subsequence problem revisited. J. Algorithms 9 (1988),
418-425.
[7] Yao, F. F. Speed-up in dynamic programming. SIAM J. Alg. Disc. Meth. 3 (1982),532-540.
