6.

D YNAMIC P ROGRAMMING I

‣ weighted interval scheduling

‣ segmented least squares
‣ knapsack problem
‣ RNA secondary structure

Lecture slides by Kevin Wayne

Copyright © 2005 Pearson-Addison Wesley
http://www.cs.princeton.edu/~wayne/kleinberg-tardos

Last updated on 2/10/21 2:39 PM

Algorithmic paradigms

Greed. Process the input in some order, myopically making irrevocable

decisions.

Divide-and-conquer. Break up a problem into independent subproblems;

solve each subproblem; combine solutions to subproblems to form solution
to original problem.

Dynamic programming. Break up a problem into a series of overlapping

subproblems; combine solutions to smaller subproblems to form solution
to large subproblem.

fancy name for

caching intermediate results
in a table for later reuse

2
Dynamic programming history

Bellman. Pioneered the systematic study of dynamic programming in 1950s.

Etymology.
・Dynamic programming = planning over time.
・Secretary of Defense had pathological fear of mathematical research.
・Bellman sought a “dynamic” adjective to avoid conflict.

THE THEORY OF DYNAMIC PROGRAMMING

RICHARD BELLMAN

1. Introduction. Before turning to a discussion of some representa-

tive problems which will permit us to exhibit various mathematical
features of the theory, let us present a brief survey of the funda-
mental concepts, hopes, and aspirations of dynamic programming.
To begin with, the theory was created to treat the mathematical
problems arising from the study of various multi-stage decision
processes, which may roughly be described in the following way: We
have a physical system whose state at any time / is determined by a
set of quantities which we call state parameters, or state variables.
At certain times, which may be prescribed in advance, or which may
be determined by the process itself, we are called upon to make de-
cisions which will affect the state of the system. These decisions are
equivalent to transformations of the state variables, the choice of a
decision being identical with the choice of a transformation. The out-
come of the preceding decisions is to be used to guide the choice of
future ones, with the purpose of the whole process that of maximizing
some function of the parameters describing the final state.
Examples of processes fitting this loose description are furnished
by virtually every phase of modern life, from the planning of indus-
trial production lines to the scheduling of patients at a medical
clinic ; from the determination of long-term investment programs for
universities to the determination of a replacement policy for ma-
chinery in factories; from the programming of training policies for
skilled and unskilled labor to the choice of optimal purchasing and in-
ventory policies for department stores and military establishments.
It is abundantly clear from the very brief description of possible
applications t h a t the problems arising from the study of these
processes are problems of the future as well as of the immediate 3
present.
Dynamic programming applications

Application areas.
・Computer science: AI, compilers, systems, graphics, theory, ….
・Operations research.
・Information theory.
・Control theory.
・Bioinformatics.
Some famous dynamic programming algorithms.
・Avidan–Shamir for seam carving.
・Unix diff for comparing two files.
・Viterbi for hidden Markov models.
・De Boor for evaluating spline curves.
・Bellman–Ford–Moore for shortest path.
・Knuth–Plass for word wrapping text in T X .
・Cocke–Kasami–Younger for parsing context-free grammars.
<latexit sha1_base64="EhG7/UUHfhF6o3p74Glw1DH78ik=">AAACK3icbVBNS8NAEN34WetXW49egkXwVBIRrLeCF48VGltoQ9lspu3SzSbsTqQl9C941V/hr/GiePV/mKQ52NYHC4/3ZnZmnhcJrtGyPo2t7Z3dvf3SQfnw6PjktFKtPekwVgwcFopQ9TyqQXAJDnIU0IsU0MAT0PWm95nffQaleSg7OI/ADehY8hFnFDNp0IHesFK3GlYOc5PYBamTAu1h1agN/JDFAUhkgmrdt60I3YQq5EzAojyINUSUTekY+imVNADtJvmyC/MyVXxzFKr0STRz9W9HQgOt54GXVgYUJ3rdy8T/vH6Mo6abcBnFCJItB41iYWJoZpebPlfAUMxTQpni6a4mm1BFGab5rEzJ/46ArVySzGLJWejDmipwhoou0hTt9cw2iXPduGvYjzf1VrOIs0TOyQW5Ija5JS3yQNrEIYxMyAt5JW/Gu/FhfBnfy9Ito+g5Iyswfn4BrkOohg==</latexit>
sha1_base64="K4LIWrwvep29t51jV7N8ecuok+0=">AAACGXicbVBNT8JAEN3iFyIqePWykZh4Iq0X9WbixSMmVkigIdvtFDZst83u1EAa/oBXf4W/xpMx/huXwkHAl2zy8t7MzswLMykMuu6PU9nZ3ds/qB7Wjuq145PTRv3FpLnm4PNUproXMgNSKPBRoIRepoEloYRuOHlY+N1X0Eak6hlnGQQJGykRC87QSp1ho+W23RJ0m3gr0iIrDJvO2SBKeZ6AQi6ZMX3PzTAomEbBJcxrg9xAxviEjaBvqWIJmKAo95zTS6tENE61fQppqf7tKFhizCwJbWXCcGw2vYX4n9fPMb4NCqGyHEHx5aA4lxRTujiaRkIDRzmzhHEt7K6Uj5lmHG00a1PKvzPga5cU01wJnkawoUqcomZzG6K3Gdk28a/bd23vySVVck4uyBXxyA25J4+kQ3zCSUTeyLvz4Xw6X8usK84q9CZZg/P9C1n2pHk=</latexit>
sha1_base64="SWKETCWkcgZ0sm6ZDiXeRUTM5+0=">AAACIHicbVBNT8JAEJ36iYgKXL1sJCaeSOtFvZl48YgJCAk0ZLsMsGG7bXanBtLwF7zqr/DXeNH4YyyFg4Av2eTlvZmdmRfESlpy3S9nZ3dv/+CwcFQ8Lp2cnpUrpWcbJUZgS0QqMp2AW1RSY4skKezEBnkYKGwHk4eF335BY2WkmzSL0Q/5SMuhFJwWUq+JnX655tbdHGybeCtSgxUa/YpT7Q0ikYSoSShubddzY/JTbkgKhfNiL7EYczHhI+xmVPMQrZ/my87ZZaYM2DAy2dPEcvVvR8pDa2dhkFWGnMZ201uI/3ndhIa3fip1nBBqsRw0TBSjiC0uZwNpUJCaZYQLI7NdmRhzwwVl+axNyf+OUaxdkk4TLUU0wA1V0ZQMn2cpepuZbZPWdf2u7j25UIBzuIAr8OAG7uERGtACAWN4hTd4dz6cT+d7GfeOs8q9Cmtwfn4BcFCm+g==</latexit>
sha1_base64="HJV6YlmlVjnPgPYqeSt99ZGHmBg=">AAACK3icbVBNT8JAEN3iF+IX4NFLIzHxRFov4o3Ei0dMqJBAQ7bbATZst83u1EAa/oJX/RX+Gi8ar/4P29KDgC/Z5OW9mZ2Z50WCa7SsT6O0s7u3f1A+rBwdn5yeVWv1Jx3GioHDQhGqvkc1CC7BQY4C+pECGngCet7sPvN7z6A0D2UXFxG4AZ1IPuaMYiYNu9AfVRtW08phbhO7IA1SoDOqGfWhH7I4AIlMUK0HthWhm1CFnAlYVoaxhoiyGZ3AIKWSBqDdJF92aV6lim+OQ5U+iWau/u1IaKD1IvDSyoDiVG96mfifN4hx3HITLqMYQbLVoHEsTAzN7HLT5woYikVKKFM83dVkU6oowzSftSn53xGwtUuSeSw5C33YUAXOUdFlmqK9mdk2cW6ad0370Wq0W0WcZXJBLsk1scktaZMH0iEOYWRKXsgreTPejQ/jy/helZaMouecrMH4+QWtA6iC</latexit>

・Needleman–Wunsch/Smith–Waterman for sequence alignment.

4
Dynamic programming books
 6. D YNAMIC P ROGRAMMING I

‣ weighted interval scheduling

‣ segmented least squares
‣ knapsack problem
‣ RNA secondary structure

SECTIONS 6.1–6.2
Weighted interval scheduling

・Job j starts at sj, finishes at fj, and has weight wj > 0.

・Two jobs are compatible if they don’t overlap.
・Goal: find max-weight subset of mutually compatible jobs.

sj wj fj

h
time
0 1 2 3 4 5 6 7 8 9 10 11
7
 Earliest-finish-time first algorithm

Earliest finish-time first.

・Consider jobs in ascending order of finish time.
・Add job to subset if it is compatible with previously chosen jobs.
Recall. Greedy algorithm is correct if all weights are 1.

Observation. Greedy algorithm fails spectacularly for weighted version.

weight = 999 b
weight = 1
weight = 1 a
h
time
0 1 2 3 4 5 6 7 8 9 10 11
8
Weighted interval scheduling

Convention. Jobs are in ascending order of finish time: f1 ≤ f2 ≤ . . . ≤ fn .

Def. p( j) = largest index i < j such that job i is compatible with j.

Ex. p(8) = 1, p(7) = 3, p(2) = 0. i is rightmost interval
that ends before j begins

8
time
0 1 2 3 4 5 6 7 8 9 10 11
9
Dynamic programming: binary choice

Def. OPT( j) = max weight of any subset of mutually compatible jobs for
subproblem consisting only of jobs 1, 2, ..., j.

Goal. OPT(n) = max weight of any subset of mutually compatible jobs.

Case 1. OPT( j) does not select job j.

・Must be an optimal solution to problem consisting of remaining
jobs 1, 2, ..., j – 1.
optimal substructure property
(proof via exchange argument)
Case 2. OPT( j) selects job j.
・Collect profit wj.
・Can’t use incompatible jobs { p( j) + 1, p( j) + 2, ..., j – 1 }.
・Must include optimal solution to problem consisting of remaining
compatible jobs 1, 2, ..., p( j).

0 j=0
Bellman equation. OP T (j) =
<latexit sha1_base64="qEOy02oDztIDL8rE3HZJCxx6JUo=">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</latexit>
max { OP T (j 1), wj + OP T (p(j)) } j>0
10
Weighted interval scheduling: brute force

BRUTE-FORCE (n, s1, …, sn, f1, …, fn, w1, …, wn)

_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Sort jobs by finish time and renumber so that f1 ≤ f2 ≤ … ≤ fn.

Compute p[1], p[2], …, p[n] via binary search.
RETURN COMPUTE-OPT(n).

COMPUTE-OPT( j )
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

IF (j = 0)
RETURN 0.

ELSE
RETURN max {COMPUTE-OPT( j – 1), wj + COMPUTE-OPT( p[ j ]) }.

11
Dynamic programming: quiz 1

What is running time of COMPUTE-OPT(n) in the worst case?

A. Θ(n log n)

B. Θ(n2) (1) n=1

T (n) =
C. Θ(1.618 ) n
2T (n 1) + (1) n>1
<latexit sha1_base64="eY4cuuNdtPSoc87X0TvZnXEtJMc=">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</latexit>

D. Θ(2n)

COMPUTE-OPT( j )
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

IF (j = 0)
RETURN 0.

ELSE
RETURN max {COMPUTE-OPT( j – 1), wj + COMPUTE-OPT( p[ j ]) }.

12
Weighted interval scheduling: brute force

Observation. Recursive algorithm is spectacularly slow because of

overlapping subproblems ⇒ exponential-time algorithm.

Ex. Number of recursive calls for family of “layered” instances grows like
Fibonacci sequence.

1 4 3

2
3 2 2 1
3

4
2 1 1 0 1 0
5
1 0
p(1) = 0, p(j) = j-2
recursion tree

13
Weighted interval scheduling: memoization

Top-down dynamic programming (memoization).

・Cache result of subproblem j in M [ j].
・Use M [ j] to avoid solving subproblem j more than once.

TOP-DOWN(n, s1, …, sn, f1, …, fn, w1, …, wn)

_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Sort jobs by finish time and renumber so that f1 ≤ f2 ≤ … ≤ fn.

Compute p[1], p[2], …, p[n] via binary search.
M [ 0 ] ← 0. global array

RETURN M-COMPUTE-OPT(n).

M-COMPUTE-OPT( j )
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

IF (M [ j] is uninitialized)
M [ j] ← max { M-COMPUTE-OPT ( j – 1), wj + M-COMPUTE-OPT( p[ j]) }.
RETURN M [ j].
14
Weighted interval scheduling: running time

Claim. Memoized version of algorithm takes O(n log n) time.

Pf.
・Sort by finish time: O(n log n) via mergesort.
・Compute p[ j] for each j : O(n log n) via binary search.
・M-COMPUTE-OPT( j): each invocation takes O(1) time and either
- (1) returns an initialized value M [ j]
- (2) initializes M [ j] and makes two recursive calls

・Progress measure Φ = # initialized entries among M [1.. n].

- initially Φ = 0; throughout Φ ≤ n.
- (2) increases Φ by 1 ⇒ ≤ 2n recursive calls.

・Overall running time of M-COMPUTE-OPT(n) is O(n). ▪

15
16
Weighted interval scheduling: finding a solution

Q. DP algorithm computes optimal value. How to find optimal solution?

A. Make a second pass by calling FIND-SOLUTION(n).

FIND-SOLUTION( j)
_____________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________
____________________________________________________________________________________________________________

IF (j = 0)
RETURN ∅.
ELSE IF (wj + M [ p[ j]] > M [ j – 1])
RETURN { j } ∪ FIND-SOLUTION(p[ j]).
ELSE
RETURN FIND-SOLUTION( j – 1).

M [ j] = max { M[j – 1], wj + M[p[ j]] }.

Analysis. # of recursive calls ≤ n ⇒ O(n).

17
Weighted interval scheduling: bottom-up dynamic programming

Bottom-up dynamic programming. Unwind recursion.

BOTTOM-UP(n, s1, …, sn, f1, …, fn, w1, …, wn)

_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Sort jobs by finish time and renumber so that f1 ≤ f2 ≤ … ≤ fn.

Compute p[1], p[2], …, p[n].
M [ 0 ] ← 0. previously computed values

FOR j = 1 TO n
M [ j] ← max { M [ j – 1], wj + M [ p[ j]] }.
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Running time. The bottom-up version takes O(n log n) time.

18
MAXIMUM SUBARRAY PROBLEM

Goal. Given an array x of n integer (positive or negative),

find a contiguous subarray whose sum is maximum.

12 5 −1 31 −61 59 26 −53 58 97 −93 −23 84 −15 6

187

Applications. Computer vision, data mining,

genomic sequence analysis, technical job interviews, ….

19
MAXIMUM SUBARRAY PROBLEM

Goal. Given an array x of n integer (positive or negative),

find a contiguous subarray whose sum is maximum.
i j

12 5 −1 31 −61 59 26 −53 58 97 −93 −23 84 −15 6

187

Brute-force algorithm.
・For each i and j : computer a[i] + a[i+1] + … + a[j].
・Takes Θ(n ) time.
3

Apply “cumulative sum” trick.

・Precompute cumulative sums: S[i] = a[0] + a[1] + … + a[i].
・Now a[i] + a[i+1] + … + a[j] = S[j] − S[i−1].
20

・Improves running time Θ(n ). 2

KADANE’S ALGORITHM

Def. OPT( i) = max sum of any subarray of x whose rightmost

index is i.

Goal. max OP T (i)

<latexit sha1_base64="CZcWtyZJDhy3Lko1sMUcUMz99YY=">AAACOXicbVDLTgIxFO3gC/E14NJNIzHRDZkxJmLckLhxJyYgJEBIp1y0odOZtHcIZMKfuNWv8EfcujJu/QHLYyHoSZqcnHNfPUEshUHPe3cya+sbm1vZ7dzO7t7+gZsvPJgo0RzqPJKRbgbMgBQK6ihQQjPWwMJAQiMY3Ez9xhC0EZGq4TiGTsgelegLztBKXddth2zUFbR9Te+qtVNx1nWLXsmbgf4l/oIUyQLVbt4ptHsRT0JQyCUzpuV7MXZSplFwCZNcOzEQMz5gj9CyVLEQTCednT6hJ1bp0X6k7VNIZ+rvjpSFxozDwFaGDJ/MqjcV//NaCfbLnVSoOEFQfL6on0iKEZ3mQHtCA0c5toRxLeytlD8xzTjatJa2zGbHwJd+ko4SJXjUgxVV4gg1m9gU/dXM/pL6eemq5N9fFCvlRZxZckSOySnxySWpkFtSJXXCyZA8kxfy6rw5H86n8zUvzTiLnkOyBOf7B0NwrLA=</latexit>
i

x1 i=1
Bellman equation. OP T (i) =
<latexit sha1_base64="IzPr6CPzgLq87qXSEs8B8qojGEU=">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</latexit>
max { xi , xi + OP T (i 1) } i>1

Running time. O(n). take only take element i

element i together with best subarray
ending at index i − 1

21
MAXIMUM RECTANGLE PROBLEM

Goal. Given an n-by-n matrix A, find a rectangle whose sum is maximum.

2 5 0 5 2 2 3
4 3 1 3 2 1 1
5 6 3 5 1 4 2
A = 1 1 3 1 4 1 1
3 3 2 0 3 3 2
2 1 2 1 1 3 1
<latexit sha1_base64="7MCCgMPkRZ+McrhVixdB+YGxHPQ=">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</latexit>
2 4 0 1 0 3 1

13

Applications. Databases, image processing, maximum likelihood

estimation, technical job interviews, …

22
BENTLEY’S ALGORITHM

Assumption. Suppose you knew the left and right column indices j and jʹ.
j jʹ
2 5 0 5 2 2 3 7
4 3 1 3 2 1 1 4 0−5−2
5 6 3 5 1 4 2 3
A = 1 1 3 1 4 1 1 x = 6
3 3 2 0 3 3 2 5
2 1 2 1 1 3 1 0
<latexit sha1_base64="7MCCgMPkRZ+McrhVixdB+YGxHPQ=">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</latexit>
2 4 0 1 0 3 1 <latexit sha1_base64="KksKdRyhI0qOgtDRpYhf7ThpXdA=">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</latexit>
1

An O(n3) algorithm. j solve maximum

・Precompute cumulative row sums Sij = Aik .

subarray problem
in this array

・ For each j < jʹ : k=1 <latexit sha1_base64="hAhGMhRyRjoYbIVIbkUh6Iy0kU0=">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</latexit>

– define array x using row-sum differences: xi = Sij <latexit sha1_base64="6Iv2z52WvxSnJvX12CP4j21Z+nA=">AAACQHicbVDLSgMxFM34rPXV1oULN8EiurHMiKAuhIIblxWtFWoZMultjWYyQ3JHWob5Grf6FX6Fn+BK3LoyfSxs64HA4Zz7yD1BLIVB1/1w5uYXFpeWcyv51bX1jc1CsXRrokRzqPNIRvouYAakUFBHgRLuYg0sDCQ0gqeLgd94Bm1EpG6wH0MrZF0lOoIztJJf2O75gp7Taz8Vj/sZPRyxLO8Xym7FHYLOEm9MymSMml90SvftiCchKOSSGdP03BhbKdMouIQsf58YiBl/Yl1oWqpYCKaVDi/I6J5V2rQTafsU0qH6tyNloTH9MLCVIcMHM+0NxP+8ZoKd01YqVJwgKD5a1EkkxYgO4qBtoYGj7FvCuBb2r5Q/MM042tAmtgxnx8AnLkl7iRI8asOUKrGHmmU2RW86s1lSP6qcVbyr43L1dBxnjuyQXXJAPHJCquSS1EidcJKRF/JK3px359P5cr5HpXPOuGeLTMD5+QWvkq9d</latexit>

Sij
– run Kadane’s algorithm in array x

Open problem. O(n3−ε) for any constant ε > 0.

23
 6. D YNAMIC P ROGRAMMING I

‣ weighted interval scheduling

‣ segmented least squares
‣ knapsack problem
‣ RNA secondary structure

SECTION 6.3
Least squares

Least squares. Foundational problem in statistics.

・Given n points in the plane: (x1, y1), (x2, y2) , …, (xn, yn).
・Find a line y = ax + b that minimizes the sum of the squared error:
n
!
SSE = (yi − axi − b)2
y
<latexit sha1_base64="Zy1LiC11wpnyrhTk6pqMmahVKW8=">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</latexit>
sha1_base64="Pz37uMf9ZkSqUadkulfgYpEBW/M=">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</latexit>
sha1_base64="kNKMdq3i/R2HN6/wEHFPEtvhKsQ=">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</latexit>
i=1

Solution. Calculus ⇒ min error is achieved when

! ! ! ! !
n ix i yi − ( i xi )( i yi ) i yi − a i xi
a = ! 2 ! 2
, b =
<latexit sha1_base64="uSR6MIDaEDPJwkbPgs/ygO2X/DQ=">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</latexit>
sha1_base64="h5fKMH7NWGnBcmDeMLEiuZrLM70=">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</latexit>
sha1_base64="osgmybCFlDRwfvDA5+Vdeo/74To=">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</latexit>
n i xi − ( i x i ) n

25
Segmented least squares

Segmented least squares.

・Points lie roughly on a sequence of several line segments.
・Given n points in the plane: (x1, y1), (x2, y2) , …, (xn, yn) with
x1 < x2 < ... < xn, find a sequence of lines that minimizes f (x).

Q. What is a reasonable choice for f (x) to balance accuracy and parsimony?

goodness of fit number of lines

x 26
Segmented least squares

Segmented least squares.

・Points lie roughly on a sequence of several line segments.
・Given n points in the plane: (x1, y1), (x2, y2) , …, (xn, yn) with
x1 < x2 < ... < xn, find a sequence of lines that minimizes f (x).

Goal. Minimize f (x) = E + c L for some constant c > 0, where

・E = sum of the sums of the squared errors in each segment.
・L = number of lines.
y

x 27
Dynamic programming: multiway choice

Notation.
・OPT( j) = minimum cost for points p1, p2, …, pj.
・eij = SSE for for points pi, pi+1, …, pj.

To compute OPT( j):

・Last segment uses points pi, pi+1, …, pj for some i ≤ j.
・Cost = eij + c + OPT(i – 1). optimal substructure property
(proof via exchange argument)

Bellman equation.

0 j=0
OP T (j) =
min { eij + c + OP T (i 1) } j>0
<latexit sha1_base64="3Vqhon9XuZlMDxUy0KO6LsmTpKw=">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</latexit>
1 i j

28
Segmented least squares algorithm

SEGMENTED-LEAST-SQUARES(n, p1, …, pn, c)

__________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

FOR j = 1 TO n
FOR i = 1 TO j
Compute the SSE eij for the points pi, pi+1, …, pj.

M [ 0] ← 0.
previously computed value
FOR j = 1 TO n
M [ j] ← min 1 ≤ i ≤ j { eij + c + M [ i – 1] }.

RETURN M [ n].
__________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

29
Segmented least squares analysis

Theorem. [Bellman 1961] DP algorithm solves the segmented least squares

problem in O(n3) time and O(n2) space.

Pf.
・Bottleneck = computing SSE eij for each i and j.
! ! ! ! !
n k yk − (
k x! k xk )( k yk ) k yk − aij k xk
aij = 2
! 2
, bij =
<latexit sha1_base64="BChx1F3fNhGq6mUCFDX2bORLMoA=">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</latexit>
sha1_base64="JiXEyfCjefrho5PdKjL7JUVzDmM=">AAACyXicbVHdStxAFJ6krf/a1V72ZqgUFOySLLQqpbDgTS96YaFbhc0aTiYnOt3JJM6P7Dbkqi/iO/g0gg/Tya7W7uqBmfn4zu98JykF1yYIbj3/xctXC4tLyyura+sbr1ubWz91YRXDHitEoU4T0Ci4xJ7hRuBpqRDyROBJMjxq/CdXqDQv5A8zLnGQw7nkGWdgHBW3riGu+K+aRp/pl+aKMgWskjTSNo+HdBQPx+75QHceid0H7Dy79UzsWWcu9KxT79Ho0kJKk+caPVZyiQ+j/Et3xeu4tR20g4nRpyC8B9vd/Zu7LiHkON70tqK0YDZHaZgArfthUJpBBcpwJrBeiazGEtgQzrHvoIQc9aCaSFnT945JaVYod6ShE/b/jApyrcd54iJzMBd63teQz/n61mQHg4rL0hqUbNoos4KagjZ7oSlXyIwYOwBMcTcrZRfgNDJuezNdJrVLZDM/qUZWclakOMcKMzIKGhXDec2egl6nfdgOvzs1D8jUlshb8o7skJDsky75So5JjzBvwdvzPnqf/G++8kf+72mo793nvCEz5v/5CzSm3bw=</latexit>
sha1_base64="CRJSkf/k/PXSE8sZbaW7avsGjhU=">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</latexit>
n k x k − ( k xk ) n

・O(n) to compute eij. ▪

Remark. Can be improved to O(n2) time. i i i i

・For each i : precompute cumulative sums

k=1
<latexit sha1_base64="E9Bo6rLIFL0/cuHYqsa/2SHtfn0=">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</latexit>
xk ,
k=1
yk ,
k=1
x2k ,
k=1
xk yk .

・Using cumulative sums, can compute eij in O(1) time.

30
 6. D YNAMIC P ROGRAMMING I

‣ weighted interval scheduling

‣ segmented least squares
‣ knapsack problem
‣ RNA secondary structure

SECTION 6.4
Knapsack problem

Goal. Pack knapsack so as to maximize total value of items taken.

・There are n items: item i provides value vi > 0 and weighs wi > 0.
・Value of a subset of items = sum of values of individual items.
・Knapsack has weight limit of W.
Ex. The subset { 1, 2, 5 } has value $35 (and weight 10).
Ex. The subset { 3, 4 } has value $40 (and weight 11).

Assumption. All values and weights are integral.

i vi wi

$18
1 $1 1 kg
g weights and values
5k
$22 g
6k can be arbitrary
2 $6 2 kg
11 kg positive integers
3 $18 5 kg
$6 2 kg
4 $22 6 kg

g
5 $28 7 kg
$28
7k
$1 g
1k
knapsack instance
Creative Commons Attribution-Share Alike 2.5
(weight limit W = 11)
by Dake 32
Dynamic programming: quiz 2

Which algorithm solves knapsack problem?

A. Greedy-by-value: repeatedly add item with maximum vi.

B. Greedy-by-weight: repeatedly add item with minimum wi.

C. Greedy-by-ratio: repeatedly add item with maximum ratio vi / wi.

D. None of the above.

i vi wi

$18
1 $1 1 kg
g
5k
$22 g
6k
2 $6 2 kg
11 kg
3 $18 5 kg
$6 2 kg
4 $22 6 kg

g
5 $28 7 kg
$28
7k
$1 g
1k
knapsack instance
Creative Commons Attribution-Share Alike 2.5
(weight limit W = 11)
by Dake
33
Dynamic programming: quiz 3

Which subproblems?

A. OPT(w) = optimal value of knapsack problem with weight limit w.

B. OPT(i) = optimal value of knapsack problem with items 1, …, i.

C. OPT(i, w) = optimal value of knapsack problem with items 1, …, i

subject to weight limit w.

D. Any of the above.

34
Dynamic programming: two variables

Def. OPT(i, w) = optimal value of knapsack problem with items 1, …, i,

subject to weight limit w.
Goal. OPT(n, W).
possibly because wi > w

Case 1. OPT(i, w) does not select item i.

・OPT(i, w) selects best of { 1, 2, …, i – 1 } subject to weight limit w.
Case 2. OPT(i, w) selects item i. optimal substructure property

・Collect value vi.

(proof via exchange argument)

・New weight limit = w – wi.

・OPT(i, w) selects best of { 1, 2, …, i – 1 } subject to new weight limit.
Bellman equation.
0 i=0

OP T (i, w) = OP T (i 1, w) wi > w

<latexit sha1_base64="yGc+dBk4QwRLdPr1J0EKEi2YjzI=">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</latexit>
max { OP T (i 1, w), vi + OP T (i 1, w wi ) }
35
Knapsack problem: bottom-up dynamic programming

KNAPSACK(n, W, w1, …, wn, v1, …, vn )

__________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

FOR w = 0 TO W
M [ 0, w] ← 0.

previously computed values

FOR i = 1 TO n
FOR w = 0 TO W
IF (wi > w) M [ i, w] ← M [ i – 1, w].
ELSE M [ i, w] ← max { M [ i – 1, w], vi + M [ i – 1, w – wi] }.

RETURN M [ n, W].
__________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

0 i=0

OP T (i, w) = OP T (i 1, w) wi > w

<latexit sha1_base64="yGc+dBk4QwRLdPr1J0EKEi2YjzI=">AAADBHicbVFNb9NAEF2brxK+0nLkMiICFZFGNkKlKCoq4sKNIDW0UjaK1puxs+p67e6um0RWzvwaTogr/wPxZ1i7OTgpI6309Gbem9mZKJfC2CD44/m3bt+5e2/nfuvBw0ePn7R3976ZrNAchzyTmT6PmEEpFA6tsBLPc40sjSSeRRefqvzZFWojMnVqlzmOU5YoEQvOrKMm7b9fBqf7ogvzV0D7x7TfohEmQpXceZpVi/YDeAnU4sJCCSKGFQg4hgAoHfUOMR27itrhIKw9tmrnEwEfYN6spilbAJUYW6Cl6wkNeRdot6KunOx1I3HgfKr5gGqRzJxw1WiU2RnquTDo+lHaoqim6+kn7U7QC+qAmyBcgw5Zx2Cy6+3RacaLFJXlkhkzCoPcjkumreAS3ToKgznjFyzBkYOKpWjGZX2FFbxwzBTiTLunLNRsU1Gy1JhlGrnKlNmZ2c5V5P9yo8LGR+NSqLywqPh1o7iQYDOoTgpToZFbuXSAcS3crMBnTDNu3eE3utTeOfKNn5SLQgmeTXGLlXZhNau2GG7v7CYYvum974Vf33ZOjtbr3CHPyHOyT0LyjpyQz2RAhoR7H73Ey71L/7v/w//p/7ou9b215inZCP/3P7Vj6Os=</latexit>
max { OP T (i 1, w), vi + OP T (i 1, w wi ) }
36
 Knapsack problem: bottom-up dynamic programming demo

i vi wi
1 $1 1 kg 0 i=0
2 $6 2 kg OP T (i, w) = OP T (i 1, w) wi > w
3 $18 5 kg max {OP T (i 1, w), vi + OP T (i 1, w wi }
<latexit sha1_base64="Z2vromkJx+wQ2wpWsn/gLtchm2Q=">AAAC+nicbVFda9swFJW9ry77SrvHvVwWNjqWBnuMtSN0FPayt2XQrIXIBFm5TkRl2Uhyk2D8a/Y09ro/ssf9m8luHpx0FwSHc889VzqKcymMDYK/nn/n7r37D/Yedh49fvL0WXf/4LvJCs1xzDOZ6cuYGZRC4dgKK/Ey18jSWOJFfPW57l9cozYiU+d2nWOUsrkSieDMOmra/fN1dH4o+rB8A3R4SocdGuNcqJI7T1N16DCA10AtriyUIBKoQMApBEDpZPAB08gpGoejsPHY0S6nAj7Bsq2mKVsBlZhYoCW0ZvtA+3DtBt622KPagWoxXzh51bLP7AL1Uhh0WyjtUFSzzZ2n3V4wCJqC2yDcgB7Z1Gi67x3QWcaLFJXlkhkzCYPcRiXTVnCJLoTCYM74FZvjxEHFUjRR2WRfwSvHzCDJtDvKQsO2J0qWGrNOY6dMmV2Y3V5N/q83KWxyEpVC5YVFxW8WJYUEm0H9kTATGrmVawcY18LdFfiCacat++6tLY13jnzrJeWqUIJnM9xhpV1ZzeoUw93MboPxu8HHQfjtfe/sZBPnHnlBXpJDEpJjcka+kBEZE+4de5GXeHO/8n/4P/1fN1Lf28w8J1vl//4H1VHmOQ==</latexit>

4 $22 6 kg
5 $28 7 kg
weight limit w

0 1 2 3 4 5 6 7 8 9 10 11

{ } 0 0 0 0 0 0 0 0 0 0 0 0

{1} 0 1 1 1 1 1 1 1 1 1 1 1

subset { 1, 2 } 0 1 6 7 7 7 7 7 7 7 7 7
of items
1, …, i { 1, 2, 3 } 0 1 6 7 7 18 19 24 25 25 25 25

{ 1, 2, 3, 4 } 0 1 6 7 7 18 22 24 28 29 29 40

{ 1, 2, 3, 4, 5 } 0 1 6 7 7 18 22 28 29 34 35 40

OPT(i, w) = optimal value of knapsack problem with items 1, …, i, subject to weight limit w
37
Knapsack problem: running time

Theorem. The DP algorithm solves the knapsack problem with n items

and maximum weight W in Θ(n W) time and Θ(n W) space.
Pf. weights are integers

・Takes O(1) time per table entry. between 1 and W

・There are Θ(n W) table entries.

・After computing optimal values, can trace back to find solution:
OPT(i, w) takes item i iff M [i, w] > M [i – 1, w]. ▪

Remarks.
・Algorithm depends critically on assumption that weights are integral.
・Assumption that values are integral was not used.

38
Dynamic programming: quiz 4

Does there exist a poly-time algorithm for the knapsack problem?

A. Yes, because the DP algorithm takes Θ(n W) time. “pseudo-polynomial”

B. No, because Θ(n W) is not a polynomial function of the input size.

C. No, because the problem is NP-hard.

D. Unknown.

equivalent to P ≠ NP conjecture
because knapsack problem is NP-hard

39
COIN CHANGING

Problem. Given n coin denominations { d1, d2, …, dn } and a target value V,

find the fewest coins needed to make change for V (or report impossible).

Recall. Greedy cashier’s algorithm is optimal for U.S. coin denominations,

but not for arbitrary coin denominations.

Ex. { 1, 10, 21, 34, 70, 100, 350, 1295, 1500 }.

Optimal. 140¢ = 70 + 70.

40
COIN CHANGING

Def. OPT(v) = min number of coins to make change for v.

Goal. OPT(V).

Multiway choice. To compute OPT(v), optimal substructure property

・Select a coin of denomination ci for some i. (proof via exchange argument)

・Select fewest coins to make change for v – ci.

Bellman equation. 8
<latexit sha1_base64="6B5tGCszYw5e5nluRWvumveoBFs=">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</latexit>

>
> 1 v<0
>
>
<
OP T (v) = 0 v=0
>
>
>
>
: min { 1 + OP T (v di ) } v>0
1in

Running time. O(n V).

41
 6. D YNAMIC P ROGRAMMING I

‣ weighted interval scheduling

‣ segmented least squares
‣ knapsack problem
‣ RNA secondary structure

SECTION 6.5
RNA secondary structure

RNA. String B = b1b2…bn over alphabet { A, C, G, U }.

Secondary structure. RNA is single-stranded so it tends to loop back and

form base pairs with itself. This structure is essential for understanding
behavior of molecule.

C A

A A

A U G C
base
C G U A A G

G
U A U U A
base pair G
A C G C U
G

C G C G A G C

G
A U

RNA secondary structure for GUCGAUUGAGCGAAUGUAACAACGUGGCUACGGCGAGA

43
 RNA secondary structure

Secondary structure. A set of pairs S = { (bi, bj) } that satisfy:

・[Watson–Crick] S is a matching and each pair in S is a Watson–Crick
complement: A–U, U–A, C–G, or G–C.

G G

C U

C G
A C G U G G C C A U

base pair S is not a secondary structure

in secondary structure A C
(C-A is not a valid Watson-Crick pair)

U A

B = ACGUGGCCCAU
S = { (b1, b10), (b2, b9), (b3, b8) } 44
RNA secondary structure

Secondary structure. A set of pairs S = { (bi, bj) } that satisfy:

・[Watson–Crick] S is a matching and each pair in S is a Watson–Crick
complement: A–U, U–A, C–G, or G–C.
・[No sharp turns] The ends of each pair are separated by at least 4
intervening bases. If (bi, bj) ∈ S, then i < j – 4.

G G

C G

A U G G G G C A U
A U

S is not a secondary structure

U A (≤4 intervening bases between G and C)

B = AUGGGGCAU
S = { (b1, b9), (b2, b8), (b3, b7) } 45
RNA secondary structure

Secondary structure. A set of pairs S = { (bi, bj) } that satisfy:

・[Watson–Crick] S is a matching and each pair in S is a Watson–Crick
complement: A–U, U–A, C–G, or G–C.
・[No sharp turns] The ends of each pair are separated by at least 4
intervening bases. If (bi, bj) ∈ S, then i < j – 4.
・[Non-crossing] If (bi, bj) and (bk, bℓ) are two pairs in S, then we cannot
have i < k < j < ℓ.

G G

C U

C U

A G
A G U U G G C C A U

U A S is not a secondary structure

(G-C and U-A cross)
B = ACUUGGCCAU
S = { (b1, b10), (b2, b8), (b3, b9) } 46
RNA secondary structure

Secondary structure. A set of pairs S = { (bi, bj) } that satisfy:

・[Watson–Crick] S is a matching and each pair in S is a Watson–Crick
complement: A–U, U–A, C–G, or G–C.
・[No sharp turns] The ends of each pair are separated by at least 4
intervening bases. If (bi, bj) ∈ S, then i < j – 4.
・[Non-crossing] If (bi, bj) and (bk, bℓ) are two pairs in S, then we cannot
have i < k < j < ℓ.

G G

C U

C G

A U
A U G U G G C C A U

S is a secondary structure
U A
(with 3 base pairs)
B = AUGUGGCCAU
S = { (b1, b10), (b2, b9), (b3, b8) } 47
RNA secondary structure

Secondary structure. A set of pairs S = { (bi, bj) } that satisfy:

・[Watson–Crick] S is a matching and each pair in S is a Watson–Crick
complement: A–U, U–A, C–G, or G–C.
・[No sharp turns] The ends of each pair are separated by at least 4
intervening bases. If (bi, bj) ∈ S, then i < j – 4.
・[Non-crossing] If (bi, bj) and (bk, bℓ) are two pairs in S, then we cannot
have i < k < j < ℓ.

Free-energy hypothesis. RNA molecule will form the secondary structure

with the minimum total free energy.

approximate by number of base pairs

(more base pairs ⇒ lower free energy)

Goal. Given an RNA molecule B = b1b2…bn, find a secondary structure S

that maximizes the number of base pairs.
48
Dynamic programming: quiz 5

Is the following a secondary structure?

A. Yes.

B. No, violates Watson–Crick condition.

C. No, violates no-sharp-turns condition.

D. No, violates no-crossing condition.

G C

C G U A A G

A U U A
G
G C U
G

C G A G C

A U

G 49
Dynamic programming: quiz 6

Which subproblems?

A. OPT( j) = max number of base pairs in secondary structure

of the substring b1b2 … bj.

B. OPT( j) = max number of base pairs in secondary structure

of the substring bj bj+1 … bn.

C. Either A or B.

D. Neither A nor B.

50
RNA secondary structure: subproblems

First attempt. OPT( j) = maximum number of base pairs in a secondary

structure of the substring b1b2 … bj.

Goal. OPT(n).

match bases bt and bn

Choice. Match bases bt and bj.

1 t j last base

Difficulty. Results in two subproblems (but one of wrong form).

・Find secondary structure in b1b2 … bt–1. OPT(t–1)

・Find secondary structure in bt+1bt+2 … bj–1. need more subproblems

(first base no longer b1)

51
Dynamic programming over intervals

Def. OPT(i, j) = maximum number of base pairs in a secondary structure

of the substring bi bi+1 … bj.

Case 1. If i ≥ j – 4.
・OPT(i, j) = 0 by no-sharp-turns condition.

Case 2. Base bj is not involved in a pair.

・OPT(i, j) = OPT(i, j – 1).

Case 3. Base bj pairs with bt for some i ≤ t < j – 4.

・Non-crossing condition decouples resulting two subproblems.
・OPT(i, j) = 1 + max t { OPT(i, t – 1) + OPT(t + 1, j – 1) }.
match bases bj and bt
take max over t such that i ≤ t < j – 4 and
bt and bj are Watson–Crick complements

i t j 52
Dynamic programming: quiz 7

In which order to compute OPT(i, j) ?

A. Increasing i, then j.

B. Increasing j, then i.

C. Either A or B.

D. Neither A nor B.

match bases bj and bt

i t j

OPT(i, j) depends upon OPT(i, t-1) and OPT(t+1, j-1)

53
Bottom-up dynamic programming over intervals

Q. In which order to solve the subproblems?

A. Do shortest intervals first—increasing order of ⎜j − i⎟.

RNA-SECONDARY-STRUCTURE(n, b1, …, bn ) j
6 7 8 9 10
__________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

FOR k = 5 TO n – 1 4 0 0 0
all needed values
FOR i = 1 TO n – k are already computed 3 0 0
i
j ← i + k. 2 0

Compute M[i, j] using formula. 1

RETURN M[1, n]. order in which to solve subproblems

__________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Theorem. The DP algorithm solves the RNA secondary structure problem in

O(n3) time and O(n2) space.
54
Dynamic programming summary

typically, only a polynomial

Outline. number of subproblems

・Define a collection of subproblems.

・Solution to original problem can be computed from subproblems.
・Natural ordering of subproblems from “smallest” to “largest” that
enables determining a solution to a subproblem from solutions to
smaller subproblems.

Techniques.
・Binary choice: weighted interval scheduling.
・Multiway choice: segmented least squares.
・Adding a new variable: knapsack problem.
・Intervals: RNA secondary structure.

Top-down vs. bottom-up dynamic programming. Opinions differ.

55