15-451/651: Design & Analysis of Algorithms

Lecture #10 (draft!!) September 26, 2013

Dynamic Programming is a powerful technique that allows one to solve many different types of
problems in time O(n2 ) or O(n3 ) for which a naive approach would take exponential time. In this
lecture, we discuss this technique, and present a few key examples. Topics in this lecture include:

• The basic idea of Dynamic Programming.

• Example: Longest Common Subsequence.

• Example: Knapsack.

• Example: Matrix-chain multiplication.

1 Introduction
Dynamic Programming is a powerful technique that can be used to solve many problems in time
O(n2 ) or O(n3 ) for which a naive approach would take exponential time. (Usually to get running
time below that—if it is possible—one would need to add other ideas as well.) Dynamic Pro-
gramming is a general approach to solving problems, much like “divide-and-conquer” is a general
method, except that unlike divide-and-conquer, the subproblems will typically overlap. This lecture
we will present two ways of thinking about Dynamic Programming as well as a few examples.
There are several ways of thinking about the basic idea.
Basic Idea (version 1): What we want to do is take our problem and somehow break it down into
a reasonable number of subproblems (where “reasonable” might be something like n2 ) in such a way
that we can use optimal solutions to the smaller subproblems to give us optimal solutions to the
larger ones. Unlike divide-and-conquer (as in mergesort or quicksort) it is OK if our subproblems
overlap, so long as there are not too many of them.

2 Example 1: Longest Common Subsequence

Definition 1 The Longest Common Subsequence (LCS) problem is as follows. We are given
two strings: string S of length n, and string T of length m. Our goal is to produce their longest
common subsequence: the longest sequence of characters that appear left-to-right (but not necessarily
in a contiguous block) in both strings.

For example, consider:

S = ABAZDC
T = BACBAD

In this case, the LCS has length 4 and is the string ABAD. Another way to look at it is we are finding
a 1-1 matching between some of the letters in S and some of the letters in T such that none of the
edges in the matching cross each other.
For instance, this type of problem comes up all the time in genomics: given two DNA fragments,
the LCS gives information about what they have in common and the best way to line them up.
Let’s now solve the LCS problem using Dynamic Programming. As subproblems we will look at
the LCS of a prefix of S and a prefix of T , running over all pairs of prefixes. For simplicity, let’s
1
worry first about finding the length of the LCS and then we can modify the algorithm to produce
the actual sequence itself.
So, here is the question: say LCS[i,j] is the length of the LCS of S[1..i] with T [1..j]. How
can we solve for LCS[i,j] in terms of the LCS’s of the smaller problems?

Case 1: what if S[i] 6= T [j]? Then, the desired subsequence has to ignore one of S[i] or T [j] so we
have:
LCS[i, j] = max(LCS[i − 1, j], LCS[i, j − 1]).

Case 2: what if S[i] = T [j]? Then the LCS of S[1..i] and T [1..j] might as well match them up.
For instance, if I gave you a common subsequence that matched S[i] to an earlier location in
T , for instance, you could always match it to T [j] instead. So, in this case we have:

LCS[i, j] = 1 + LCS[i − 1, j − 1].

So, we can just do two loops (over values of i and j) , filling in the LCS using these rules. Here’s
what it looks like pictorially for the example above, with S along the leftmost column and T along
the top row.
B A C B A D
A 0 1 1 1 1 1
B 1 1 1 2 2 2
A 1 2 2 2 3 3
Z 1 2 2 2 3 3
D 1 2 2 2 3 4
C 1 2 3 3 3 4
We just fill out this matrix row by row, doing constant amount of work per entry, so this takes
O(mn) time overall. The final answer (the length of the LCS of S and T ) is in the lower-right
corner.
How can we now find the sequence? To find the sequence, we just walk backwards through
matrix starting the lower-right corner. If either the cell directly above or directly to the right
contains a value equal to the value in the current cell, then move to that cell (if both to, then chose
either one). If both such cells have values strictly less than the value in the current cell, then move
diagonally up-left (this corresponts to applying Case 2), and output the associated character. This
will output the characters in the LCS in reverse order. For instance, running on the matrix above,
this outputs DABA.

3 More on the basic idea, and Example 1 revisited

We have been looking at what is called “bottom-up Dynamic Programming”. Here is another way
of thinking about Dynamic Programming, that also leads to basically the same algorithm, but
viewed from the other direction. Sometimes this is called “top-down Dynamic Programming”.
Basic Idea (version 2): Suppose you have a recursive algorithm for some problem that gives you
a really bad recurrence like T (n) = 2T (n − 1) + n. However, suppose that many of the subproblems
you reach as you go down the recursion tree are the same. Then you can hope to get a big savings
if you store your computations so that you only compute each different subproblem once. You can
store these solutions in an array or hash table. This view of Dynamic Programming is often called
memoizing.
For example, for the LCS problem, using our analysis we had at the beginning we might have
produced the following exponential-time recursive program (arrays start at 1):
2
LCS(S,n,T,m)
{
if (n==0 || m==0) return 0;
if (S[n] == T[m]) result = 1 + LCS(S,n-1,T,m-1); // no harm in matching up
else result = max( LCS(S,n-1,T,m), LCS(S,n,T,m-1) );
return result;
}

This algorithm runs in exponential time. In fact, if S and T use completely disjoint sets of characters
(so that we never have
1 S[n]==T[m]) then the number of times that LCS(S,1,T,1) is recursively
called equals n+m−2
m−1 . In the memoized version, we store results in a matrix so that any given
set of arguments to LCS only produces new work (new recursive calls) once. The memoized version
begins by initializing arr[i][j] to unknown for all i,j, and then proceeds as follows:

LCS(S,n,T,m)
{
if (n==0 || m==0) return 0;
if (arr[n][m] != unknown) return arr[n][m]; // <- added this line (*)
if (S[n] == T[m]) result = 1 + LCS(S,n-1,T,m-1);
else result = max( LCS(S,n-1,T,m), LCS(S,n,T,m-1) );
arr[n][m] = result; // <- and this line (**)
return result;
}

All we have done is saved our work in line (**) and made sure that we only embark on new recursive
calls if we haven’t already computed the answer in line (*).
In this memoized version, our running time is now just O(mn). One easy way to see this is as
follows. First, notice that we reach line (**) at most mn times (at most once for any given value of
the parameters). This means we make at most 2mn recursive calls total (at most two calls for each
time we reach that line). Any given call of LCS involves only O(1) work (performing some equality
checks and taking a max or adding 1), so overall the total running time is O(mn).
Comparing bottom-up and top-down dynamic programming, both do almost the same work. The
top-down (memoized) version pays a penalty in recursion overhead, but can potentially be faster
than the bottom-up version in situations where some of the subproblems never get examined at
all. These differences, however, are minor: you should use whichever version is easiest and most
intuitive for you for the given problem at hand.

More about LCS: Discussion and Extensions. An equivalent problem to LCS is the “mini-
mum edit distance” problem, where the legal operations are insert and delete. (E.g., the unix “diff”
command, where S and T are files, and the elements of S and T are lines of text). The minimum
edit distance to transform S into T is achieved by doing |S| − LCS(S, T ) deletes and |T | − LCS(S, T )
inserts.
In computational biology applications, often one has a more general notion of sequence alignment.
Many of these different problems all allow for basically the same kind of Dynamic Programming
solution.
1
This is the number of different “monotone walks” between the upper-left and lower-right corners of an n by m
grid.

3
4 Example #2: The Knapsack Problem
Imagine you have a homework assignment with different parts labeled A through G. Each part has
a “value” (in points) and a “size” (time in hours to complete). For example, say the values and
times for our assignment are:

A B C D E F G
value 7 9 5 12 14 6 12
time 3 4 2 6 7 3 5

Say you have a total of 15 hours: which parts should you do? If there was partial credit that was
proportional to the amount of work done (e.g., one hour spent on problem C earns you 2.5 points)
then the best approach is to work on problems in order of points-per-hour (a greedy strategy).
But, what if there is no partial credit? In that case, which parts should you do, and what is the
best total value possible?2

Exercise: Give an example where using the greedy strategy will get you less than 1% of the optimal
value (in the case there is no partial credit).

The above is an instance of the knapsack problem, formally defined as follows:

Definition 2 In the knapsack problem we are given a set of n items, where each item i is
specified by a size si and a value vi . We are also given a size bound S (the size of our knapsack).
The goal is to find the subset of items of maximum total value such that sum of their sizes is at
most S (they all fit into the knapsack).

We can solve the knapsack problem in exponential time by trying all possible subsets. With
Dynamic Programming, we can reduce this to time O(nS).
Let’s do this top down by starting with a simple recursive solution and then trying to memoize
it. Let’s start by just computing the best possible total value, and we afterwards can see how to
actually extract the items needed.

// Recursive algorithm: either we use the last element or we don’t.

Value(n,S) // S = space left, n = # items still to choose from
{
if (n == 0) return 0;
if (s_n > S) result = Value(n-1,S); // can’t use nth item
else result = max{v_n + Value(n-1, S-s_n), Value(n-1, S)};
return result;
}

Right now, this takes exponential time. But, notice that there are only O(nS) different pairs of
values the arguments can possibly take on, so this is perfect for memoizing. As with the LCS
problem, let us initialize a 2-d array arr[i][j] to “unknown” for all i,j.

Value(n,S)
{
if (n == 0) return 0;
if (arr[n][S] != unknown) return arr[n][S]; // <- added this
2
Answer: In this case, the optimal strategy is to do parts A, B, F, and G for a total of 34 points. Notice that this
doesn’t include doing part C which has the most points/hour!

4
 if (s_n > S) result = Value(n-1,S);
else result = max{v_n + Value(n-1, S-s_n), Value(n-1, S)};
arr[n][S] = result; // <- and this
return result;
}

Since any given pair of arguments to Value can pass through the array check only once, and in
doing so produces at most two recursive calls, we have at most 2n(S + 1) recursive calls total, and
the total time is O(nS).
So far we have only discussed computing the value of the optimal solution. How can we get
the items? As usual for Dynamic Programming, we can do this by just working backwards: if
arr[n][S] = arr[n-1][S] then we didn’t use the nth item so we just recursively work backwards
from arr[n-1][S]. Otherwise, we did use that item, so we just output the nth item and recursively
work backwards from arr[n-1][S-s n]. One can also do bottom-up Dynamic Programming.

Exercise: The fractional knapsack problem is the one where you can add δi ∈ [0, 1] fraction of task i
to the knapsack, using δi si space and getting δi vi value. The greedy algorithm adds items in decreasing
order of vi /si . Prove that this greedy algorithm produces the optimal solution for the fractional problem.

5 Example #3: Matrix product parenthesization

Our final example for Dynamic Programming is the matrix product parenthesization problem.
Say we want to multiply three matrices X, Y , and Z. We could do it like (XY )Z or like X(Y Z).
Which way we do the multiplication doesn’t affect the final outcome but it can affect the running
time to compute it. For example, say X is 100x20, Y is 20x100, and Z is 100x20. So, the end result
will be a 100x20 matrix. If we multiply using the usual algorithm, then to multiply an `xm matrix
by an mxn matrix takes time O(`mn). So in this case, which is better, doing (XY )Z or X(Y Z)?
Answer: X(Y Z) is better because computing Y Z takes 20x100x20 steps, producing a 20x20 matrix,
and then multiplying this by X takes another 20x100x20 steps, for a total of 2x20x100x20. But,
doing it the other way takes 100x20x100 steps to compute XY , and then multplying this with Z
takes another 100x20x100 steps, so overall this way takes 5 times longer. More generally, what if
we want to multiply a series of n matrices?

Definition 3 The Matrix Product Parenthesization problem is as follows. Suppose we need to

multiply a series of matrices: A1 ×A2 ×A3 ×. . .×An . Given the dimensions of these matrices, what
is the best way to parenthesize them, assuming for simplicity that standard matrix multiplication is
to be used (e.g., not Strassen)?

There are an exponential number of different possible parenthesizations, in fact 2(n−1)

n−1 /n, so we
don’t want to search through all of them. Dynamic Programming gives us a better way.
As before, let’s first think: how might we do this recursively? One way is that for each possible
split for the final multiplication, recursively solve for the optimal parenthesization of the left and
right sides, and calculate the total cost (the sum of the costs returned by the two recursive calls
plus the `mn cost of the final multiplication, where “m” depends on the location of that split).
Then take the overall best top-level split.
For Dynamic Programming, the key question is now: in the above procedure, as you go through the
recursion, what do the subproblems look like and how many are there? Answer: each subproblem
looks like “what is the best way to multiply some sub-interval of the matrices Ai × . . . × Aj ?” So,
there are only O(n2 ) different subproblems.

5
The second question is now: how long does it take to solve a given subproblem assuming you’ve
already solved all the smaller subproblems (i.e., how much time is spent inside any given recursive
call)? Answer: to figure out how to best multiply Ai × . . . × Aj , we just consider all possible middle
points k and select the one that minimizes:

optimal cost to multiply Ai . . . Ak ← already computed

+ optimal cost to multiply Ak+1 . . . Aj ← already computed
+ cost to multiply the results. ← get this from the dimensions

This just takes O(1) work for any given k, and there are at most n different values k to consider, so
overall we just spend O(n) time per subproblem. So, if we use Dynamic Programming to save our
results in a lookup table, then since there are only O(n2 ) subproblems we will spend only O(n3 )
time overall.
If you want to do this using bottom-up Dynamic Programming, you would first solve for all sub-
problems with j − i = 1, then solve for all with j − i = 2, and so on, storing your results in an n
by n matrix. The main difference between this problem and the two previous ones we have seen is
that any given subproblem takes time O(n) to solve rather than O(1), which is why we get O(n3 )
total running time. It turns out that by being very clever you can actually reduce this to O(1)
amortized time per subproblem, producing an O(n2 )-time algorithm, but we won’t get into that
here.3

6 High-level discussion of Dynamic Programming

What kinds of problems can be solved using Dynamic Programming? One property these problems
have is that if the optimal solution involves solving a subproblem, then it uses the optimal solution
to that subproblem. For instance, say we want to find the shortest path from A to B in a graph,
and say this shortest path goes through C. Then it must be using the shortest path from C to B.
Or, in the knapsack example, if the optimal solution does not use item n, then it is the optimal
solution for the problem in which item n does not exist. The other key property is that there should
be only a polynomial number of different subproblems. These two properties together allow us to
build the optimal solution to the final problem from optimal solutions to subproblems.
In the top-down view of dynamic programming, the first property above corresponds to being
able to write down a recursive procedure for the problem we want to solve. The second property
corresponds to making sure that this recursive procedure makes only a polynomial number of
different recursive calls. In particular, one can often notice this second property by examining
the arguments to the recursive procedure: e.g., if there are only two integer arguments that range
between 1 and n, then there can be at most n2 different recursive calls.
Sometimes you need to do a little work on the problem to get the optimal-subproblem-solution
property. For instance, suppose we are trying to find paths between locations in a city, and some
intersections have no-left-turn rules (this is particularly bad in San Francisco). Then, just because
the fastest way from A to B goes through intersection C, it doesn’t necessarily use the fastest way
to C because you might need to be coming into C in the correct direction. In fact, the right way
to model that problem as a graph is not to have one node per intersection, but rather to have one
node per hintersection, directioni pair. That way you recover the property you need.

6.1 Why “Dynamic Programming”?

The term dynamic programming comes from Richard Bellman, who also gave us the Bellman-Ford
algorithm from the next lecture. In his autobiography he says
3
For details, see Knuth (insert ref).

6
 “Let’s take a word that has an absolutely precise meaning, namely dynamic, in the
classical physical sense. It also has a very interesting property as an adjective, and
that is it’s impossible to use the word, dynamic, in a pejorative sense. Try thinking of
some combination that will possibly give it a pejorative meaning. [. . .] Thus, I thought
dynamic programming was a good name. It was something not even a Congressman
could object to. So I used it as an umbrella for my activities.”

Bellman clearly had a way with words, coining another super-well-known phrase, “the curse of
dimensionality”.