================ Start Lecture #1 ================

Basic Algorithms

2002-03 Fall
Monday and Wednesday 11-12:15
Ciww 102

Chapter 0 Administrivia

I start at 0 so that when we get to chapter 1, the numbering will agree with the text.

0.1 Contact Information

0.2 Course Web Page

There is a web site for the course. You can find it from my home page.

0.3 Textbook

The course text is Goodrich and Tamassia: ``Algorithm Design: Foundations, Analysis, and Internet Examples.

0.4 Computer Accounts and Mailman Mailing List

0.5 Grades

The major components of the grade will be the midterm, the final, and problem sets. I will post (soon) the weights for each.

0.6 Midterm

We will have a midterm. As the time approaches we will vote in class for the exact date. Please do not schedule any trips during days when the class meets until the midterm date is scheduled.

0.7 Homeworks, Problem Sets, and Labs

If you had me for 202, you know that in systems courses I also assign labs. Basic algorithms is not a systems course; there are no labs. There are homeworks and problem sets, very few if any of these will require the computer. There is a distinction between homeworks and problem sets.

Problem sets are

Homeworks are

0.8 Recitation

I run a recitation session on tuesdays from 2-3:15. I believe there is another recitation section. You need attend only one.

0.9 Obtaining Help

Good methods for obtaining help include

  1. Asking me during office hours (see web page for my hours).
  2. Asking the mailing list.
  3. Asking another student, but ...
    Your homeworks and problem sets must be your own.

0.10 The Upper Left Board

I use the upper left board for homework assignments and announcements. I should never erase that board. Viewed as a file it is group readable (the group is those in the room), appendable by just me, and (re-)writable by no one. If you see me start to erase an announcement, let me know.

0.11 A Grade of ``Incomplete''

It is university policy that a student's request for an incomplete be granted only in exceptional circumstances and only if applied for in advance. Naturally, the application must be before the final exam.

Part I: Fundamental Tools

Chapter 1 Algorithm Analysis

We are interested in designing good algorithms (a step-by-step procedure for performing some task in a finite amount of time) and good data structures (a systematic way of organizing and accessing data).

Unlike v22.102, however, we wish to determine rigorously just how good our algorithms and data structures really are and whether significantly better algorithms are possible.

1.1 Methodologies for Analyzing Algorithms

We will be primarily concerned with the speed (time complexity) of algorithms.

We will emphasize instead an analytic framework that is independent of input and hardware, and does not require an implementation. The disadvantage is that we can only estimate the time required.

Homework: R-1.1 and R-1.2 (Unless otherwise stated, homework problems are from the last section in the current book chapter.)

1.1.1 Pseudo-Code

Designed for human understanding. Suppress unimportant details and describe some parts in natural language (English in this course).

1.1.2 The Random Access Machine (RAM) Model

The key difference between the RAM model and a real computer is the assumption of a very simple memory model: Accessing any memory element takes a constant amount of time. This ignores caching and paging for example. (It also assumes the word-size of a computer is large enough to hold any address, which is generally valid for modern-day computers, but was not always the case.)

The time required is simply a count of the primitive operations executed. There are several different possible sets of primitive operations. For this course we will use

  1. Assigning a value to a variable (independent of the size of the value; but the variable must be a scalar).
  2. Method invocation, i.e., calling a function or subroutine.
  3. Performing a (simple) arithmetic operation (divide is OK, logarithm is not).
  4. Indexing into an array (for now just one dimensional; scalar access is free).
  5. Following an object reference.
  6. Returning from a method.

1.1.3 Counting Primitive Operations

Let's start with a simple algorithm (the book does a different simple algorithm, maximum).

Algorithm innerProduct
    Input: Non-negative integer n and two integer arrays A and B of size n.
    Output: The inner product of the two arrays.

prod ← 0
for i ← 0 to n-1 do
    prod ← prod + A[i]*B[i]
return prod

The total is thus 1+1+5n+2n+(n+1)+1 = 8n+4.

Homework: Perform a similar analysis for the following algorithm

Algorithm tripleProduct
    Input: Non-negative integer n and three integer arrays A. B, and C
    each of size n.
    Output: The A[0]*B[0]*C[0] + ... + A[n-1]*B[n-1]*C[n-1]

prod ← 0
for i ← 0 to n-1 do
    prod ← prod + A[i]*B[i]*C[i]
return prod
End of homework

Let's speed up innerProduct (a very little bit).

Algorithm innerProductBetter
    Input: Non-negative integer n and two integer arrays A and B of size n.
    Output: The inner product of the two arrays

prod ← A[0]*B[0]
for i ← 1 to n-1 do
    prod ← prod + A[i]*B[i]
return prod

The cost is 4+1+5(n-1)+2(n-1)+n+1 = 8n-1


If n=0, we access A[0] and B[0], which do not exist. The original version returns zero as the inner product of empty arrays, which is arguably correct. The best fix is perhaps to change Non-negative to Positive in the Input specification. Let's call this algorithm innerProductBetterFixed.

What about if statements?

Algorithm countPositives
    Input: Non-negative integer n and an integer array A of size n.
    Output: The number of positive elements in A

pos ← 0
for i ← 0 to n-1 do
    if A[i] > 0 then
        pos ← pos + 1
return pos

Let U be the number of updates done.

1.1.4 Analyzing Recursive Algorithms

Consider a recursive version of innerProduct. If the arrays are of size 1, the answer is clearly A[0]B[0]. If n>1, we recursively get the inner product of the first n-1 terms and then add in the last term.

Algorithm innerProductRecursive
    Input: Positive integer n and two integer arrays A and B of size n.
    Output: The inner product of the two arrays

if n=1 then
    return A[0]B[0]
return innerProductRecursive(n-1,A,B) + A[n-1]B[n-1]

How many steps does the algorithm require? Let T(n) be the number of steps required.

Problem Set #1, Problem 1.
The problem set will be officially assigned a little later, but the first problem in the set is R-1.27.

================ Start Lecture #2 ================

1.2 Asymptotic Notation

One could easily complain about the specific primitive operations we chose and about the amount we charge for each one. For example, perhaps we should charge one unit for accessing a scalar variable. Perhaps we should charge more for division than for addition. Some computers can multiply two numbers and add it to a third in one operation. What about the cost of loading the program?

Now we are going to be less precise and worry only about approximate answers for large inputs. Thus the rather arbitrary decisions made about how many units to charge for each primitive operation will not matter since our sloppiness will cover. Please note that the sloppiness will be very precise.

1.2.1 The Big-Oh Notation

Definition: Let f(n) and g(n) be real-valued functions of a single non-negative integer argument. We write f(n) is O(g(n)) if there is a positive real number c and a positive integer n0 such that f(n)≤cg(n) for all n≥n0.

What does this mean?

For large inputs (n≥n0), f is not much bigger than g (specifically, f(n)≤cg(n)).

Examples to do on the board

  1. 3n-6 is O(n). Some less common ways of saying the same thing follow.
  2. 3x-6 is O(x).
  3. If f(y)=3y-6 and id(y)=y, then f(y) is O(id(y)).
  4. 3n-6 is O(2n)
  5. 9n4+12n2+1234 is O(n4).
  6. innerProduct is O(n)
  7. innerProductBetter is O(n)
  8. innerProductFixed is O(n)
  9. countPositives is O(n)
  10. n+log(n) is O(n).
  11. log(n)+5log(log(n)) is O(log(n)).
  12. 1234554321 is O(1).
  13. 3/n is O(1). True but not the best.
  14. 3/n is O(1/n). Much better.
  15. innerProduct is O(3n4+100n+log(n)+34.5). True, but awful.

A few theorems give us rules that make calculating big-Oh easier.

Theorem (arithmetic): Let d(n), e(n), f(n), and g(n) be nonnegative real-valued functions of a nonnegative integer argument and assume d(n) is O(f(n)) and e(n) is O(g(n)). Then

  1. ad(n) is O(f(n)) for any nonnegative a
  2. d(n)+e(n) is O(f(n)+g(n))
  3. d(n)e(n) is O(f(n)g(n))

Theorem (transitivity): Let d(n), f(n), and g(n) be nonnegative real-valued functions of a nonnegative integer argument and assume d(n) is O(f(n)) and f(n) is O(g(n)). Then d(n) is O(g(n)).

Theorem (special functions): (Only n varies)

  1. If f(n) is a polynomial of degree d, then f(n) is O(nd).
  2. nk is O(an) for any k>0 and a>1.
  3. log(nk) is O(log(n)) for any k>0
  4. (log(n))k is O(nj) for any k>0 and j>0.

Example: (log n)1000 is O(n0.001). This says raising log n to the 1000 is not (significantly) bigger than the thousandth root of n. Indeed raising log to the 1000 is actually significantly smaller than taking the thousandth root since n0.001) is not O((log n)1000).

So log is a VERY small (i.e., slow growing) function.

Homework: R-1.19 R-1.20

Example: Let's do problem R-1.10. Consider the following simple loop that computes the sum of the first n positive integers and calculate the running time using the big-Oh notation.

Algorithm Loop1(n)
   s ← 0
   for i←1 to n do
      s ← s+i
With big-Oh we don't have to worry about multiplicative or additive constants so we see right away that the running time is just the number of iterates of the loop so the answer is O(n)

Homework: R-1.11 and R-1.12

Definitions: (Common names)

  1. If a function is O(log(n)), we call it logarithmic.
  2. If a function is O(n), we call it linear.
  3. If a function is O(n2), we call it quadratic.
  4. If a function is O(nk) with k≥1, we call it polynomial.
  5. If a function is O(an) with a>1, we call it exponential.
Remark: The last definitions would be better with a relative of big-Oh, namely big-Theta, since, for example 3log(n) is O(n2), but we do not call 3log(n) quadratic.

Homework: R-1.10 and R-1.12.

Example: R-1.13. What is running time of the following loop using big-Oh notation?

Algorithm Loop4(n)
   s ← 0
   for i←1 to 2n do
      for j←1 to i do
         s ← s+1
Clearly the time is determined by the number of executions of the last statement. But this looks hard since the inner loop is executed a different number of times for each iteration of the outer loop. But it is not so bad. For iteration i of the outer loop, the inner loop has i iterations. So the total number of iterations of the last statement is 1+2+...+2n, which is 2n(2n+1)/2. So the answer is O(n2).

Homework: R-1.14 (This was assigned during the third lecture in Fall 03, but would have made more sense here).

================ Start Lecture #3 ================

1.2.2 Relatives of the Big-Oh

Big-Omega and Big-Theta

Recall that f(n) is O(g(n)) if, for large n, f is not much bigger than g. That is g is some sort of upper bound on f. How about a definition for the case when g is (in the same sense) a lower bound for f?

Definition: Let f(n) and g(n) be real valued functions of an integer value. Then f(n) is Ω(g(n)) if g(n) is O(f(n)).


  1. We pronounce f(n) is Ω(g(n)) as "f(n) is big-Omega of g(n)".
  2. What the last definition says is that we say f(n) is not much smaller than g(n) if g(n) is not much bigger than f(n), which sounds reasonable to me.
  3. What if f(n) and g(n) are about equal, i.e., neither is much bigger than the other?

Definition: We write f(n) is Θ(g(n)) if both f(n) is O(g(n)) and f(n) is Ω(g(n)).

Remarks We pronounce f(n) is Θ(g(n)) as "f(n) is big-Theta of g(n)"

Examples to do on the board.

  1. 2x2+3x is Θ(x2).
  2. 2x3+3x is not θ(x2).
  3. 2x3+3x is Ω(x2).
  4. innerProductRecursive is Θ(n).
  5. binarySearch is Θ(log(n)). Unofficial for now.
  6. If f(n) is Θ(g(n)), the f(n) is &Omega(g(n)).
  7. If f(n) is Θ(g(n)), then f(n) is O(g(n)).

Homework: R-1.6

Little-Oh and Little-Omega

Recall that big-Oh captures the idea that for large n, f(n) is not much bigger than g(n). Now we want to capture the idea that, for large n, f(n) is tiny compared to g(n).

If you remember limits from calculus, what we want is that f(n)/g(n)→0 as n→∞. However, the definition we give does not use the word limit (it essentially has the definition of a limit built in).

Definition: Let f(n) and g(n) be real valued functions of an integer variable. We say f(n) is o(g(n)) if for any c>0, there is an n0 such that f(n)≤cg(n) for all n>n0. This is pronounced as "f(n) is little-oh of g(n)".

Definition: Let f(n) and g(n) be real valued functions of an integer variable. We say f(n) is ω(g(n) if g(n) is o(f(n)). This is pronounced as "f(n) is little-omega of g(n)".

Examples: log(n) is o(n) and x2 is ω(nlog(n)).

Homework: R-1.4. R-1.22

What is "fast" or "efficient"?

If the asymptotic time complexity is bad, say Ω(n8), or horrendous, say Ω(2n), then for large n, the algorithm will definitely be slow. Indeed for exponential algorithms even modest n's (say n=50) are hopeless.

Algorithms that are o(n) (i.e., faster than linear, a.k.a. sub-linear), e.g. logarithmic algorithms, are very fast and quite rare. Note that such algorithms do not even inspect most of the input data once. Binary search has this property. When you look up a name in the phone book you do not even glance at a majority of the names present.

Linear algorithms (i.e., Θ(n)) are also fast. Indeed, if the time complexity is O(nlog(n)), we are normally quite happy.

Low degree polynomial (e.g., Θ(n2), Θ(n3), Θ(n4)) are interesting. They are certainly not fast but speeding up a computer system by a factor of 1000 (feasible today with parallelism) means that a Θ(n3) algorithm can solve a problem 10 times larger. Many science/engineering problems are in this range.

1.2.3 The Importance of Asymptotics

It really is true that if algorithm A is o(algorithm B) then for large problems A will take much less time than B.

Definition: If (the number of operations in) algorithm A is o(algorithm B), we call A asymptotically faster than B.

Example:: The following sequence of functions are ordered by growth rate, i.e., each function is little-oh of the subsequent function.
log(log(n)), log(n), (log(n))2, n1/3, n1/2, n, nlog(n), n2/(log(n)), n2, n3, 2n.

What about those constants that we have swept under the rug?

Modest multiplicative constants (as well as immodest additive constants) don't cause too much trouble. But there are algorithms (e.g. the AKS logarithmic sorting algorithm) in which the multiplicative constants are astronomical and hence, despite its wonderful asymptotic complexity, the algorithm is not used in practice.

A Great Table

See table 1.10 on page 20.

Homework: R-1.7

1.3 A Quick Mathematical Review

This is hard to type in using html. The book is fine and I will write the formulas on the board.

1.3.1 Summations

Definition: The sigma notation: ∑f(i) with i going from a to b.

Theorem: Assume 0<a≠1. Then ∑ai i from 0 to n = (an+1-1)/(a-1).

Proof: Cute trick. Multiply by a and subtract.

Theorem: ∑i from 1 to n = n(n+1)/2.

Proof: Pair the 1 with the n, the 2 with the (n-1), etc. This gives a bunch of (n+1)s. For n even it is clearly n/2 of them. For odd it is the same (look at it).

Homework: R-1.14. (It would have been more logical to assign this last time right after I did R-1.13. In the future I will do so).

1.3.2 Logarithms and Exponents

Recall that logba = c means that bc=a. b is called the base and c is called the exponent.

What is meant by log(n) when we don't specify the base?

I assume you know what ab is. (Actually this is not so obvious. Whatever 2 raised to the square root of 3 means it is not writing 2 down the square root of 3 times and multiplying.) So you also know that ax+y=axay.

Theorem: Let a, b, and c be positive real numbers. To ease writing, I will use base 2 often. This is not needed. Any base would do.

  1. log(ac) = log(a)+log(c)
  2. log(a/c) = log(a) - log(c)
  3. log(ac) = c log(a)
  4. logc(a) = (log(a))/log(c): consider a = clogca and take log of both sides.
  5. clog(a) = a log(c): take log of both sides.
  6. (ba)c = bac
  7. babc = ba+c
  8. ba/bc = ba-c


Homework: C-1.12

Floor and Ceiling

⌊x⌋ is the greatest integer not greater than x. ⌈x⌉ is the least integer not less than x.

⌊5⌋ = ⌈5⌉ = 5

⌊5.2⌋ = 5 and ⌈5.2⌉ = 6

⌊-5.2⌋ = -6 and ⌈-5.2⌉ = -5

1.3.3 Simple Justification Techniques

By example

To prove the claim that there is a positive n satisfying nn>n+n, we merely have to note that 33>3+3.

By counterexample

To refute the claim that all positive n satisfy nn>n+n, we merely have to note that 11<1+1.

By contrapositive

"P implies Q" is the same as "not Q implies not P". So to show that in the world of positive integers "a2≥b2 implies that a≥b" we can show instead that "NOT(a≥b) implies NOT(a2≥b2)", i.e., that "a<b implies a2<b2", which is clear.

By contradiction

Assume what you want to prove is false and derive a contradiction.

Theorem: There are an infinite number of primes.

Proof: Assume not. Let the primes be p1 up to pk and consider the number A=p1p2…pk+1. A has remainder 1 when divided by any pi so cannot have any pi as a factor. Factor A into primes. None can be pi (A may or may not be prime). But we assumed that all the primes were pi. Contradiction. Hence our assumption that we could list all the primes was false.

By (complete) induction

The goal is to show the truth of some statement for all integers n≥1. It is enough to show two things.

  1. The statement is true for n=1
  2. IF the statement is true for all k<n, then it is true for n.

Theorem: A complete binary tree of height h has 2h-1 nodes.

Proof: We write NN(h) to mean the number of nodes in a complete binary tree of height h. A complete binary tree of height 1 is just a root so NN(1)=1 and 21-1 = 1. Now we assume NN(k)=2k-1 nodes for all k<h and consider a complete binary tree of height h. It is just two complete binary trees of height h-1 with new root to connect them.
So NN(h) = 2NN(h-1)+1 = 2(2h-1-1)+1 = 2h-1, as desired

Homework: R-1.9

================ Start Lecture #4 ================

Loop Invariants

Very similar to induction. Assume we have a loop with controlling variable i. For example a "for i←0 to n-1". We then associate with the loop a statement S(j) depending on j such that

  1. S(0) is true (just) before the loop begins
  2. IF S(j-1) holds before iteration j begins, then S(j) will hold when iteration j ends.
By induction we see that S(n) will be true when the nth iteration ends, i.e., when the loop ends.

I favor having array and loop indexes starting at zero. However, here it causes us some grief. We must remember that iteration j occurs when i=j-1.

Example:: Recall the countPositives algorithm

Algorithm countPositives
    Input: Non-negative integer n and an integer array A of size n.
    Output: The number of positive elements in A

pos ← 0
for i ← 0 to n-1 do
    if A[i] > 0 then
        pos ← pos + 1
return pos

Let S(j) be "pos equals the number of positive values in the first j elements of A".

Just before the loop starts S(0) is true vacuously. Indeed that is the purpose of the first statement in the algorithm.

Assume S(j-1) is true before iteration j, then iteration j (i.e., i=j-1) checks A[j-1] which is the jth element and updates pos accordingly. Hence S(j) is true after iteration j finishes.

Hence we conclude that S(n) is true when iteration n concludes, i.e. when the loop terminates. Thus pos is the correct value to return.

1.3.4 Basic Probability

Skipped for now.

1.4 Case Studies in Algorithm Analysis

1.4.1 A Quadratic-Time Prefix Averages Algorithm

We trivially improved innerProduct (same asymptotic complexity before and after). Now we will see a real improvement. For simplicity I do a slightly simpler algorithm, prefix sums.

Algorithm partialSumsSlow
    Input: Positive integer n and a real array A of size n
    Output: A real array B of size n with B[i]=A[0]+…+A[i]

for i ← 0 to n-1 do
    s ← 0
    for j ← 0 to i do
        s ← s + A[j]
    B[i] ← s
return B

The update of s is performed 1+2+…+n times. Hence the running time is Ω(1+2+…+n)=&Omega(n2). In fact it is easy to see that the time is &Theta(n2).

1.4.2 A Linear-Time Prefix Averages Algorithm

Algorithm partialSumsFast
    Input: Positive integer n and a real array A of size n
    Output: A real array B of size n with B[i]=A[0]+…+A[i]

s ← 0
for i ← 0 to n-1 do
    s ← s + A[i]
    B[i] ← s
return B

We just have a single loop and each statement inside is O(1), so the algorithm is O(n) (in fact Θ(n)).

Homework: Write partialSumsFastNoTemps, which is also Θ(n) time but avoids the use of s (it still uses i so my name is not great).

1.5 Amortization

Often we have a data structure supporting a number of different operations that will each be applied many times. Sometimes the worst case time complexity (i.e., the longest amount of time) of a sequence of n operations is significantly less than n times the worst case complexity of one operations. We give an example very soon.

If we divide the running time of the sequence by the number of operations performed we get the average time for each operation in the sequence, which is called the amortized running time.

Why amortized?
Because the cost of the occasional expensive application is amortized over the numerous cheap application (I think).

Example:: (From the book.) The clearable table. This is essentially an array. The table is initially empty (i.e., has size zero). We want to support three operations.

  1. Add(e): Add a new entry to the table at the end (extending its size).
  2. Get(i): Return the ith entry in the table.
  3. Clear(): Remove all the entries by setting each entry to zero (for security) and setting the size to zero.

The obvious implementation is to use a large array A and an integer s indicating the current size of A. More precisely A is (always) of size N (large) and s indicates the extent of A that is currently in use.

We are ignoring a number of error cases. For example, it is an error to issue Get(5) if only two entries have been put into the clearable table.

We start with a size zero table and assume we perform n (legal) operations. Question: What is the worst-case running time for all n operations? Once we know the answer, the amortized time is this answer divided by n.

One possibility is that the sequence consists of n-1 add(e) operations followed by one Clear(). The Clear() takes Θ(n), which is the worst-case time for any operation (assuming n operations in total). Since there are n operations and the worst-case is Θ(n) for one of them, we might think that the worst-case sequence would take Θ(n2).

But this is wrong.

It is easy to see that Add(e) and Get(i) are Θ(1).

The total time for all the Clear() operations in any sequence of n operation is O(n) since in total O(n) entries were cleared (since at most n entries were added).

Hence, the amortized time for each operation in the clearable ADT (abstract data type) is O(1), in fact Θ(1).


Note that we first found an upper bound on the complexity (i.e., big-Oh) and then a lower bound (Ω). Together this gave Θ.

1.5.1 Amortization Techniques

The Accounting Method

Overcharge for cheap operations and undercharge expensive so that the excess charged for the cheap (the profit) covers the undercharge (the loss). This is called in accounting an amortization schedule.

Assume the get(i) and add(e) really cost one ``cyber-dollar'', i.e., there is a constant K so that they each take fewer than K primitive operations and we let a ``cyber-dollar'' be K. Similarly, assume that clear() costs P cyber-dollars when the table has P elements in it.

We charge 2 cyber-dollars for every operation. So we have a profit of 1 on each add(e) and we see that the profit is enough to cover next clear() since if we clear P entries, we had P add(e)s.

All operations cost 2 cyber-dollars so n operations cost 2n. Since we have just seen that the real cost is no more than the cyber-dollars spent, the total cost is O(n) and the amortized cost is O(1). Since every operation has cost Ω(1), the amortized cost is Θ(1).

Potential Functions

Very similar to the accounting method. Instead of banking money, you increase the potential energy. I don't believe we will use this method so we are skipping it. If you like physics more than accounting, you might prefer it.

1.5.2 Analyzing an Extendable Array Implementation

We want to let the size of an array grow dynamically (i.e., during execution). The implementation is quite simple. Copy the old array into a new one twice the size. Specifically, on an array overflow instead of signaling an error perform the following steps (assume the array is A and the current size is N)

  1. Allocate a new array B of size 2N
  2. For i←0 to N-1 do B[i]←A[i]
  3. Make A refer to B (this is A=B in C and java).
  4. Deallocate the old A (automatic in java; error prone in C)

The cost of this growing operation is Θ(N).

Theorem: Given an extendable array A that is initially empty and of size N, the amortized time to perform n add(e) operations is Θ(1).

Proof: Assume one cyber dollar is enough for an add w/o the grow and that N cyber-dollars are enough to grow from N to 2N. Charge 2 cyber dollars for each add; so a profit of 1 for each add w/o growing. When you must do a grow, you had N adds so have N dollars banked. Hence the amortized cost is O(1). Since Omega;(1) is obvious, we get Θ(1).

================ Start Lecture #5 ================

Alternate Proof that the amortized time is O(1). Note that amortized time is O(1) means that the total time is O(N). The new proof is a two step procedure

  1. The total memory used is bounded by a constant times N.
  2. The time per memory cell is bounded by a constant.
Hence the total time is bounded by a constant times the number of memory cells and hence is bounded by a constant times N.

For step one we note that when the N add operations are complete the size of the array will be FS<.2N, with FS a power of 2. Let FS=2k So the total size used is TS=1+2+4+8+...+FS=∑2i (i from 0 to k). We already proved that this is (2k+1-1)/(2-1)=2k+1-1=2FS-1<4N as desired.

An easier way to see that the sum is 2k+1-1 is to write 1+2+4+8+...2k in binary, in which case we get (for k=5)

 111111 = 1000000-1 = 25+1-1

The second part is clear. For each cell the algorithm only does a bounded number of operations. The cell is allocated, a value is copied in to the cell, and a value is copied out of the cell (and into another cell).

1.6 Experimentation

1.6.1 Experimental Setup

The book is quite clear. I have little to add.

Choosing the question

You might want to know

Deciding what to measure

1.6.2 Data Analysis and Visualization

Ratio test

Assume you believe the running time t(n) of an algorithm is Θ(nd) for some specific d and you want to both verify your assumption and find the multiplicative constant.

Make a plot of (n, t(n)/nd). If you are right the points should tend toward a horizontal line and the height of this line is the multiplicative constant.

Homework: R-1.29

What if you believe it is polynomial but don't have a guess for d?
Ans: Use ...

The power test

Plot (n, t(n)) on log log paper. If t(n) is Θ(nd), say t(n) approaches bnd, then log(t(n)) approaches log(b)+d(log(n)).

So when you plot (log(n), log(t(n)) (i.e., when you use log log paper), you will see the points approach (for large n) a straight line whose slope is the exponent d and whose y intercept is the multiplicative constant b.

Homework: R-1.30

Chapter 2 Basic Data Structures

2.1 Stacks and Queues

2.1.1 Stacks

Stacks implement a LIFO (last in first out) policy. All the action occurs at the top of the stack, primarily with the push(e) and pop operation.

The stack ADT supports

There is a simple implementation using an array A and an integer s (the current size). A[s-1] contains the TOS.

Objection (your honor). The ADT says we can always push. A simple array implementation would need to signal an error if the stack is full.

Sustained! What do you propose instead?

An extendable array.

Good idea.

Homework: Assume a software system has 100 stacks and 100,000 elements that can be on any stack. You do not know how the elements are to be distributed on the stacks. However, once an element is put on one stack, it never moves. If you used a normal array based implementation for the stacks, how much memory will you need. What if you use an extendable array based implementation? Now answer the same question, but assume you have Θ(S) stacks and Θ(E) elements.

Applications for procedure calls

Stacks work great for implementing procedure calls since procedures have stack based semantics. That is, last called is first returned and local variables allocated with a procedure are deallocated when the procedure returns.

So have a stack of "activation records" in which you keep the return address and the local variables.

Support for recursive procedures comes for free. For languages with static memory allocations (e.g., fortran) one can store the local variables with the method. Fortran forbids recursion so that memory allocation can be static. Recursion adds considerably flexibility to a language as some cost in efficiency (not part of this course).

A Review of (the Real) Mod

I am reviewing modulo since I believe it is no longer taught in high school.

The top diagram shows an almost ordinary analog clock. The major difference is that instead of 12 we have 0. The hands would be useful if this was a video, but I omitted them for the static picture. Positive numbers go clockwise (cw) and negative counter-clockwise (ccw). The numbers shown are the values mod 12. This example is good to show arithmetic. (2-5) mod 12 is obtained by starting at 2 and moving 5 hours ccw, which gives 9. (-7) mod 12 is (0-7) mod 12 is obtained by starting at 0 and going 7 hours ccw, which gives 5.

To get mod 8, divide the circle into 8 hours instead of 12.

The bottom picture shows mod 5 in a linear fashion. In pink are the 5 values one can get when doing mod 5, namely 0, 1, 2, 3, and 4. I only illustrate numbers from -3 to 11 but that is just due to space limitations. Each blue bar is 5 units long so the numbers at its endpoints are equal mod 5 (since they differ by 5). So you just lay off the blue bar until you wind up in the pink.

End of Mod Review

Homework: Using the real mod (let's call it RealMod) evaluate

  1. RealMod(8,3)
  2. RealMod(-8,3)
  3. RealMod(10000007,10)
  4. RealMod(-10000007,10)

2.1.2 Queues

Queues implement a FIFO (first in first out) policy. Elements are inserted at the rear and removed from the front using the enqueue and dequeue operations respectively.

The queue ADT supports

Simple circular-array implementation

Personal rant:
I object to programming languages using the well known and widely used function name mod and changing its meaning. My math training prevents me from accepting that mod(-3,10) is -3. The correct value is 7. The book, following java, defines mod(x,y) as x-⌊x/y⌋y. This is not mod but remainder. Think of x as the dividend, y as the divisor and then ⌊x/y⌋ is the quotient. We remember from elementary school
dividend = quotient * divisor + remainder
remainder = dividend - quotient * divisor
The last line is exactly the book's and java's definition of mod.

My favorite high level language, ada, gets it right in the obvious way: Ada defines both mod and remainder (ada extends the math definition of mod to the case where the second argument is negative).

In the familiar case when x≥0 and y>0 mod and remainder are equal. Unfortunately the book uses mod sometimes when x<0 and consequently needs to occasionally add an extra y to get the true mod.
End of personal rant

Returning to relevant issues we note that for queues we need a front and rear "pointers" f and r. Since we are using arrays f and r are actually indexes not pointers. Calling the array Q, Q[f] is the front element of the queue, i.e., the element that would be returned by dequeue(). Similarly, Q[r] is the element into which enqueue(e) would place e. There is one exception: if f=r, the queue is empty so Q[f] is not the front element.

Without writing the code, we see that f will be increased by each dequeue and r will be increased by every enqueue.

Assume Q has n slots Q[0]…Q[N-1] and the queue is initially empty with f=r=0. Now consider enqueue(1); dequeue(); enqueue(2); dequeue(); enqueue(3); dequeue(); …. There is never more than one element in the queue, but f and r keep growing so after N enqueue(e);dequeue() pairs, we cannot issue another operation.

The solution to this problem is to treat the array as circular, i.e., right after Q[N-1] we find Q[0]. The way to implement this is to arrange that when either f or r is N-1, adding 1 gives 0 not N. Similarly for r. So the increment statements become
f←(f+1) mod N
r←(r+1) mod N

Note: Recall that we had some grief due to our starting arrays and loops at 0. For example, the fifth slot of A is A[4] and the fifth iteration of "for i←0 to 30" occurs when i=4. The updates of f and r directly above show one of the advantages of starting at 0; they are less pretty if the array starts at 1.

The size() of the queue seems to be r-f, but this is not always correct since the array is circular. For example let N=10 and consider an initially empty queue with f=r=0 that has
enqueue(10)enqueue(20);dequeue();enqueue(30);dequeue();enqueue(40);dequeue() applied. The queue has one element, f=4, and r=3. Now apply 6 more enqueue(e) operations
enqueue(50);enqueue(60);enqueue(70);enqueue(80);enqueue(90);enqueue(100) At this point the array has 7 elements, f=0, and r=3. Clearly the size() of the queue is not f-r=-3. It is instead 7, the number of elements in the queue.

The problem is that f in some sense is 10 not 0 since there were 10 enqueue(e) operations. In fact if we kept 2 values for f and 2 for r, namely the value before the mod and after, then size() would be fBeforeMod-rBeforeMod. Instead we, use the following inelegant formula.
size() = (r-f+N) mod N

Remark: If java's definition of -3 mod 10 gave 7 (as it should) instead of -3, we could use the more attractive formula
size() = (r-f) mod N.

Since isEmpty() is simply an abbreviation for the test size()=0, it is just testing if r=f.

Algorithm front():
    if isEmpty() then
        signal an error // throw QueueEmptyException
    return Q[f]
Algorithm dequeue():
    if isEmpty() then
        signal an error // throw QueueEmptyException
    Q[f]←NULL      // for security or debugging
    f←(f+1) mod N
    return temp
Algorithm enqueue(e):
    if size() = N-1 then
        signal an error // throw QueueFullException
    r←(r+1) mod N

================ Start Lecture #6 ================

Examples in OS

Round Robin processor scheduling is queue based as is fifo disk arm scheduling.

More general processor or disk arm scheduling policies often use priority queues (with various definitions of priority). We will learn how to implement priority queues later this chapter (section 2.4).

Homework: (You may refer to your 202 notes if you wish; mine are on-line based on my home page). How can you interpret Round Robin processor scheduling and fifo disk scheduling as priority queues. That is what is the priority? Same question for SJF (shortest job first) and SSTF (shortest seek time first). If you have not taken an OS course (202 or equivalent at some other school), you are exempt from this question. Just write on you homework paper that you have not taken an OS course.

Problem Set #1, Problem 2: C-2.2

2.2 Vectors, Lists, and Sequences

Unlike stacks and queues, the structures in this section support operations in the middle, not just at one or both ends.

2.2.1 Vectors

The rank of an element in a sequence is the number of elements before it. So if the sequence contains n elements, 0≤rank<n.

A vector storing n elements supports:

A Simple Array-Based Implementation

Use an array A and store the element with rank r in A[r].

Algorithm insertAtRank(r,e)
   for i = n-1, n-2, ..., r do

Algorithm removeAtRank(r)
   for i = r, r+1, ..., n-2 do
   return e

The worst-case time complexity of these two algorithms is Θ(n); the remaining algorithms are all Θ(1).

Homework: When does the worst case occur for insertAtRank(r,e) and removeAtRank(r)?

By using a circular array we can achieve Θ(1) time for insertAtRank(0,e) and removeAtRank(0). Indeed, that is the third problem of the first problem set.

Problem Set #1, Problem 3:
Part 1: C-2.5 from the book
Part 2: This implementation still has worst case complexity Θ(n). When does the worst case occur?

2.2.2 Lists

So far we have been considering what Knuth refers to as sequential allocation, when the next element is stored in the next location. Now we will be considering linked allocation, where each element refers explicitly to the next and/or preceding element(s).

Positions and Nodes

We think of each element as contained in a node, which is a placeholder that also contains references to the preceding and/or following node.

But in fact we don't want to expose Nodes to user's algorithms since this would freeze the possible implementation. Instead we define the idea (i.e., ADT) of a position in a list. The only method available to users is

The List ADT

Given the position ADT, we can now define the methods for the list ADT. The first methods only query a list; the last ones actually modify it.

A Linked List Implementation

Now when we are implementing a list we can certainly use the concept of nodes. In a singly linked list each node contains a next link that references the next node. A doubly linked list contains, in addition prev link that references the previous node.

Singly linked lists work well for stacks and queues, but do not perform well for general lists. Hence we use doubly linked lists

Homework: What is the worst case time complexity of insertBefore for a singly linked list implementation and when does it occur?

================ Start Lecture #7 ================


  1. Harper has offered to have an office hour thurs from 2-3. He is also available by email appointment.
  2. I think I wrote something wrong on the upper left board. I put part 2 of problem 3 of problem set 1 as a homework. You can leave it in the homework if you like, but it must be part of the problem set (which is still not assigned, but coming soon).

It is convenient to add two special nodes, a header and trailer. The header has just a next component, which links to the first node and the trailer has just a prev component, which links to the last node. For an empty list, the header and trailer link to each other and for a list of size 1, they both link to the only normal node.

In order to proceed from the top (empty) list to the bottom list (with one element), one would need to execute one of the insert methods. Ignoring the abbreviations, this means either insertBefore(p,e) or inserAfter(p,e). But this means that header and/or trailer must be an example of a position, one for which there is no element.

This observation explains the authors' comment above that insertBefore(p,e) cannot be applied if p is the first position. What they mean is that when we permit header and trailer to be positions, then we cannot insertBefore the first position, since that position is the header and the header has no prev. Similarly we cannot insertAfter the final position since that position is the trailer and the trailer has no next. Clearly not the authors' finest hour.

A list object contains three components, the header, the trailer, and the size of the list. Note that the book forgets to update the size for inserts and deletes.

Implementation Comment I have not done the implementation. It is probably easiest to have header and trailer have the same three components as a normal node, but have the prev of header and the next of trailer be some special value (say NULL) that can be tested for.

The insertAfter(p,e) Algorithm

The position p can be header, but cannot be trailer.

Algorithm insertAfter(p,e):
   If p is trailer then signal an error
   size←size+1            // missing in book
   Create a new node v
   p.next← v
   return v

Do on the board the pointer updates for two cases: Adding a node after an ordinary node and after header. Note that they are the same. Indeed, that is what makes having the header and trailer so convenient.

Homework: Write pseudo code for insertBefore(p,e).

Note that insertAfter(header,e) and insertBefore(trailer,e) appear to be the only way to insert an element into an empty list. In particular, insertFirst(e) fails for an empty list since it performs insertBefore(first()) and first() generates an error for an empty list.

The remove(p) Algorithm

We cannot remove the header or trailer. Notice that removing the only element of a one-element list correctly produces an empty list.

Algorithm remove(p)
   if p is either header or trailer signal an error
   size←size-1             // missing in book
   p.prev←NULL             // for security or debugging
   return t

2.2.3 Sequences

size, isEmptyO(1)O(1)
atRank, rankOf, elemAtRankO(1)O(n)
first, last, before, afterO(1)O(1)
replaceElement, swapElementsO(1)O(1)
insertAtRank, removeAtRankO(n)O(n)
insertFirst, insertLastO(1)O(1)
insertAfter, insertBeforeO(n)O(1)
Asymptotic complexity of the methods for both the array and list based implementations

Define a sequence ADT that includes all the methods of both vector and list ADTs as well as

Sequences can be implemented as either circular arrays, as we did for vectors, or doubly linked lists, as we did for lists. Neither clearly dominates the other. Instead it depends on the relative frequency of the various operations. Circular arrays are faster for some and doubly liked lists are faster for others as the table to the right illustrates.


An ADT for looping through a sequence one element at a time. It has two methods.

When you create the iterator it has all the elements of the sequence. So a typical usage pattern would be

create iterator I for sequence S
while I.hasNext

2.3 Trees

The tree ADT stores elements hierarchically. There is a distinguished root node. All other nodes have a parent of which they are a child. We use nodes and positions interchangeably for trees.

The definition above precludes an empty tree. This is a matter of taste some authors permit empty trees, others do not.

Some more definitions.

We order the children of a binary tree so that the left child comes before the right child.

There are many examples of trees. You learned or will learn tree-structured file systems in 202. However, despite what the book says, for Unix/Linux at least, the file system does not form a tree (due to hard and symbolic links).

These notes can be thought of as a tree with nodes corresponding to the chapters, sections, subsections, etc.

Games like chess are analyzed in terms of trees. The root is the current position. For each node its children are the positions resulting from the possible moves. Chess playing programs often limit the depth so that the number of examined moves is not too large.

An arithmetic expression tree

The leaves are constants or variables and the internal nodes are binary arithmetic operations (+,-,*,/). The tree is a proper ordered binary tree (since we are considering binary operators). The value of a leaf is the value of the constant or variable. The value of an internal node is obtained by applying the operator to the values of the children (in order).

Evaluate an arithmetic expression tree on the board.

Homework: R-2.2, but made easier by replacing 21 by 10. If you wish you can do the problem in the book instead (I think it is harder).

2.3.1 The Tree Abstract Data Type

We have three accessor methods (i.e., methods that permit us to access the nodes of the tree.

We have four query methods that test status.

Finally generic methods that are useful but not related to the tree structure.

2.3.2 Tree Traversal

Traversing a tree is a systematic method for accessing or "visiting" each node. We will see and analyze three tree traversal algorithms, inorder, preorder, and postorder. They differ in when we visit an internal node relative to its children. In preorder we visit the node first, in postorder we visit it last, and in inorder, which is only defined for binary trees, we visit the node between visiting the left and right children.

Recursion will be a very big deal in traversing trees!!

Motivating the Recursive Routines

On the right are three trees. The left one just has a root, the right has a root with one leaf as a child, and the middle one has six nodes. For each node, the element in that node is shown inside the box. All three roots are labeled and 2 other nodes are also labeled. That is, we give a name to the position, e.g. the left most root is position v. We write the name of the position under the box. We call the left tree T0 to remind us it has height zero. Similarly the other two are labeled T2 and T1 respectively.

Our goal in this motivation is to calculate the sum the elements in all the nodes of each tree. The answers are, from left to right, 8, 28, and 9.

For a start, lets write an algorithm called treeSum0 that calculates the sum for trees of height zero. In fact the algorithm, will contain two parameters, the tree and a node (position) in that tree, and our algorithm will calculate the sum in the subtree rooted at the given position assuming the position is at height 0. Note this is trivial: since the node has height zero, it has no children and the sum desired is simply the element in this node. So legal invocations would include treeSum0(T0,s) and treeSum0(T2,t). Illegal invocations would include treeSum0(T0,t) and treeSum0(T1,r).

Algorithm treeSum0(T,v)
  Inputs: T a tree; v a height 0 node of T
  Output: The sum of the elements of the subtree routed at v

  return Sum

Now lets write treeSum1(T,v), which calculates the sum for a node at height 1. It will use treeSum0 to calculate the sum for each child.

Algorithm treeSum1(T,v)
   Inputs: T a tree; v a height 1 node of T
   Output: the sum of the elements of the subtree routed at v

   for each child c of v
   return Sum

OK. How about height 2?

Algorithm treeSum2(T,v)
   Inputs: T a tree; v a height 2 node of T
   Output: the sum of the elements of the subtree routed at v

   for each child c of v
   return Sum

So all we have to do is to write treeSum3, treSum4, ... , where treSum3 invokes treeSum2, treeSum4 invokes treeSum3, ... .

That would be, literally, an infinite amount of work.

Do a diff of treeSum1 and treeSum2.
What do you find are the differences.
In the Algorithm line and in the first comment a 1 becomes a 2.
In the subroutine call a 0 becomes a 1.

Why can't we write treeSumI and let I vary?
Because it is illegal to have a varying name for an algorithm.

The solution is to make the I a parameter and write

Algorithm treeSum(i,T,v)
   Inputs: i≥0; T a tree; v a height i node of T
   Output: the sum of the elements of the subtree routed at v

   for each child c of v
   return Sum

This is wrong, why?
Because treeSum(0,T,v) invokes treeSum(-1,c,v), which doesn't exist because i<0

But treeSum(0,T,v) doesn't have to call anything since v can't have any children (the height of v is 0). So we get

Algorithm treeSum(i,T,v)
   Inputs: i≥0; T a tree; v a height i node of T
   Output: the sum of the elements of the subtree routed at v

   if i>0 then
      for each child c of v
   return Sum

The last two algorithms are recursive; they call themselves. Note that when treeSum(3,T,v) calls treeSum(2,T,c), the new treeSum has new variables Sum and c.

We are pretty happy with our treeSum routine, but ...

The algorithm is wrong! Why?
The children of a height i node need not all be of height i-1. For example s is hight 2, but its left child w is height 0. (A corresponding error also existed in treeSum2(T,v)

But the only real use we are making of i is to prevent us from recursing when we are at a leaf (the i>0 test). But we can use isInternal instead, giving our final algorithm

Algorithm treeSum(T,v)
   Inputs: T a tree; v a node of T
   Output: the sum of the elements of the subtree routed at v

   if T.isInternal(v) then
      for each child c of v
   return Sum

================ Start Lecture #8 ================

Our medium term goal is to learn about tree traversals (how to "visit" each node of a tree once) and to analyze their complexity.

Complexity of Primitive Operations

Our complexity analysis will proceed in a somewhat unusual order. Instead of starting with the bottom or lowest level routines (the tree methods in 2.3.1, e.g., is Internal(v)) or the top level routines (the traversals themselves), we will begin by analyzing some middle level procedures assuming the complexities of the low level are as we assert them to be. Then we will analyze the traversals using the middle level routines and finally we will give data structures for trees that achieve our assumed complexity for the low level.

Let's begin!

Complexity Assumptions for the Tree ADT

These assumptions will be verified later.

Middle level routines depth and height

Definitions of depth and height.

Remark: Even our definitions are recursive!

From the recursive definition of depth, the recursive algorithm for its computation essentially writes itself.

Algorithm depth(T,v)
   if T.isRoot(v) then
      return 0
      return 1 + depth(T,T.parent(v))

The complexity is Θ(the answer), i.e. Θ(dv), where dv is the depth of v in the tree T.

Problem Set #1, Problem 4:
Rewrite depth(T,v) without using recursion.
This is quite easy. I include it in the problem set to ensure that you get practice understanding recursive definitions.
The problem set is now assigned. It is due in 3 lectures from now (i.e., about 1.5 weeks).

The following algorithm computes the height of a position in a tree.

Algorithm height(T,v):
   if T.isLeaf(v) then
      return 0
      for each w in T.children(v) do
      return h+1

Remarks on the above algorithm

  1. The loop could (perhaps should) be written in pure iterator style. Note that T.children(v) is an iterator.
  2. This algorithm is not so easy to convert to non-recursive form
    It is not tail-recursive, i.e. the recursive invocation is not just at the end.
  3. To get the height of the tree, execute height(T,T.root())
Algorithm height(T)

Let's use the "official" iterator style.

Algorithm height(T,v):
    if T.isLeaf then
       return 0
       childrenOfV←T.children(v)    // "official" iterator style
       while childrenOfV.hasNext()
       return h+1

But the children iterator is defined to return the empty set for a leaf so we don't need the special case

Algorithm height(T,v):
     childrenOfV←T.children(v)    // "official" iterator style
     while childrenOfV.hasNext()
     return h+1

Theorem: Let T be a tree with n nodes and let cv be the number of children of node v. The sum of cv over all nodes of the tree is n-1.

Proof: This is trivial! ... once you figure out what it is saying. The sum gives the total number of children in a tree. But this almost all nodes. Indeed, there is just one exception.
What is the exception?
The root.

Corollary: Computing the height of an n-node tree has time complexity Θ(n).

Proof: Look at the code of the first version.

To be more formal, we should look at the "official" iterator version. The only real difference is that in the official version, we are charged for creating the iterator. But the charge is the number of elements in the iterator, i.e., the number of children this node has. So the sum of all the charges for creating iterators will be the sum of the number of children each node has, which is the total number of children, which is n-1, which is (another) $Theta;(n) and hence doesn't change the final answer.

Do a few on the board. As mentioned above, becoming facile with recursion is vital for tree analyses.

Definition: A traversal is a systematic way of "visiting" every node in a tree.

Preorder Traversal

Visit the root and then recursively traverse each child. More formally we first give the procedure for a preorder traversal starting at any node and then define a preorder traversal of the entire tree as a preorder traversal of the root.

Algorithm preorder(T,v):
   visit node v
   for each child c of v

Algorithm preorder(T):


  1. In a preorder traversal, parents come before children (which is as it should be :-)).
  2. If you describe a book as an ordered tree, with nodes for each chapter, section, etc., then the pre-order traversal visits the nodes in the order they would appear in a table of contents.

Do a few on the board. As mentioned above, becoming facile with recursion is vital for tree analyses.

Theorem: Preorder traversal of a tree with n nodes has complexity Θ(n).

Proof: Just like height.
The nonrecursive part of each invocation takes O(1+cv)
There are n invocations and the sum of the c's is n-1.

Homework: R-2.3

================ Start Lecture #9 ================

Postorder Traversal

First recursively traverse each child then visit the root. More formerly

Algorithm postorder(T,v):
   for each child c of v
   visit node v

Algorithm postorder(T):

Theorem: Preorder traversal of a tree with n nodes has complexity Θ(n).

Proof: The same as for preorder.


  1. Postorder is how you evaluate an arithmetic expression tree.
  2. Evaluate some arithmetic expression trees.
  3. When you write out the nodes in the order visited your get what is called either "reverse polish notation" or "polish notation"; I don't remember which.
  4. If you do preorder traversal you get the other one.

Problem Set 2, Problem 1. Note that the height of a tree is the depth of a deepest node. Extend the height algorithm so that it returns in addition to the height the v.element() for some v that is of maximal depth. Note that the height algorithm is for an arbitrary (not necessarily binary) tree; your extension should also work for arbitrary trees (this is *not* harder).

2.3.3 Binary Trees

Recall that a binary tree is an ordered tree in which no node has more than two children. The left child is ordered before the right child.

The book adopts the convention that, unless otherwise mentioned, the term "binary tree" will mean "proper binary tree", i.e., all internal nodes have two children. This is a little convenient, but not a big deal. If you instead permitted non-proper binary trees, you would test if a left child existed before traversing it (similarly for right child.)

Will do binary preorder (first visit the node, then the left subtree, then the right subtree, binary postorder (left subtree, right subtree, node) and then inorder (left subtree, node, right subtree).

The Binary Tree ADT

We have three (accessor) methods in addition to the general tree methods.

  1. leftChild(v): Return the (position of the) left child; signal an error if v is a leaf.
  2. rightChild(v): Similar
  3. sibling(v): Return the (unique) sibling of v; signal an error if v is the root (and hence has no sibling).

Remark: I will not hold you responsible for the proofs of the theorems.

Theorem: Let T be a binary tree having height h and n nodes. Then

  1. The number of leaves in T is at least h+1 and at most 2h.
  2. The number of internal nodes in T is at least h and at most 2h-1.
  3. The number of nodes in T is at least 2h+1 and at most 2h+1-1.
  4. log(n+1)-1≤h≤(n-1)/2


  1. Induction on n the number of nodes in the tree.

    Base case n=1: Clearly true for all trees having only one node.

    Induction hypothesis: Assume true for all trees having at most k nodes.

    Main inductive step: prove the assertion for all trees having k+1 nodes. Let T be a tree with k nodes and let h be the height of T.

    Remove the root of T. The two subtrees produced each have no more than k nodes so satisfy the assertion. Since each has height at most h-1, each has at most 2h-1 leaves. At least one of the subtrees has height exactly h-1 and hence has at least h leaves. Put the original tree back together.
    One subtree has at least h leaves, the other has at least 1, so the original tree has at least h+1. Each subtree has at most 2h-1 leaves and the original root is not a leaf, so the original has at most 2h leaves.

  2. Same idea. Induction on the number of nodes. Clearly true if T has one node. Remove the root. At least one of the subtrees is height h-1 and hence has at least h-1 internal nodes. Each of the subtrees has height at most h-1 so has at most 2h-1-1 internal nodes. Put the original tree back together. One subtree has at least h-1 internal nodes, the other has at least 1, so the original tree has at least h. Each subtree has at most 2h-1-1 internal nodes and the original root is an internal node, so the original has at most 2(2h-1)+1 = 2h-1-1 internal nodes.
  3. Add parts 1 and 2.
  4. Apply algebra (including a log) to part 3.

Theorem:In a binary tree T, the number of leaves is 1 more than the number of internal nodes.

Proof: Again induction on the number of nodes. Clearly true for one node. Assume true for trees with up to n nodes and let T be a tree with n+1 nodes. For example T is the top tree on the right.

  1. Choose a leaf and its parent (which of course is internal). For example, the leaf t and parent s in red.
  2. Remove the leaf and its parent (middle diagram)
  3. Splice the tree back without the two nodes (bottom diagram).
  4. Since S has n-1 nodes, S satisfies the assertion.
  5. Note that T is just S + one leaf + one internal so also satisfies the assertion.

Alternate Proof (does not use the pictures):

  1. Place two tokens on each internal node.
  2. Push these tokens to the two children.
  3. Notice that now all nodes but the root have one token; the root has none.
  4. Hence 2*(number internal) = (number internal) + (number leaves) + 1
  5. Done (slick!)

Corollary: A binary tree has an odd number of nodes.

Proof: #nodes = #leaves + #internal = 2(#internal)+1.

Preorder traversal of a binary tree

Algorithm binaryPreorder(T,v)
   Visit node v
   if T.isInternal(v) then
Algorithm binaryPretorder(T)

Postorder traversal of a binary tree

Algorithm binaryPostorder(T,v)
   if T.isInternal(v) then
   Visit node v
Algorithm binaryPosttorder(T)

Inorder traversal of a binary tree

Algorithm binaryInorder(T,v)
   if T.isInternal(v) then
   Visit node v
   if T.isInternal(v) then
Algorithm binaryIntorder(T)

Definition: A binary tree is fully complete if all the leaves are at the same (maximum) depth. This is the same as saying that the sibling of a leaf is a leaf.

Euler tour traversal

Generalizes the above. Visit the node three times, first when ``going left'', then ``going right'', then ``going up''. Perhaps the words should be ``going to go left'', ``going to go right'' and ``going to go up''. These words work for internal nodes. For a leaf you just visit it three times in a row (or you could put in code to only visit a leaf once; I don't do this). It is called an Euler Tour traversal because an Euler tour of a graph is a way of drawing each edge exactly once without taking your pen off the paper. The Euler tour traversal would draw each edge twice but if you add in the parent pointers, each edge is drawn once.

The book uses ``on the left'', ``from below'', ``on the right''. I prefer my names, but you may use either.

Algorithm eulerTour(T,v):
   visit v going left
   if T.isInternal(v) then
   visit v going right
   if T.isInternal(v) then
   visit v going up

Algorithm eulerTour(T):

Pre- post- and in-order traversals are special cases where two of the three visits are dropped.

It is quite useful to have this three visits. For example here is a nifty algorithm to print and expression tree with parentheses to indicate the order of the operations. We just give the three visits.

Algorithm visitGoingLeft(v):
   if T.isInternal(v) then
      print "("

Algorithm visitGoingRight(v)
   print v.element()

Algorithm visitGoingUp(v)
   if T.isInternal(v) then
      print ")"

Homework: Plug these in to the Euler Tour and show that what you get is the same as

Algorithm printExpression(T,v):
   input: T an expression tree v a node in T.
   if T.isLeaf(v) then
      print v.element()  // for a leaf the element is a value
      print "("
      print v.element()  // for an internal node the element is an operator
      print ")"
Algorithm printExpression(T):

Problem Set 2 problem 2. We have seen that traversals have complexity Θ(N), where N is the number of nodes in the tree. But we didn't count the costs of the visit()s themselves since the user writes that code. We know that visit() will be called N times, once per node, for post-, pre-, and in-order traversals and will be called 3N times for Euler tour traversal. So if each visit costs Θ(1), the total visit cost will be Θ(N) and thus does not increase the complexity of a traversal. If each visit costs Θ(N), the total visit cost will be Θ(N2) and hence the total traversal cost will be Θ(N2). The same analysis works for any visit cost providing all the visits cost the same. For this problem we will be considering a variable cost visits. In particular, assume that the cost of visiting a node v is the height of v (so roots can be expensive to visit, but leaves are free).

Part A. How many nodes N are in a fully complete binary tree of height h?

Part B. How many nodes are at height i in a fully complete binary tree of height h? What is the total cost of visiting all the nodes at height i?

Part C. Write a formula using ∑ (sum) for the total cost of visiting all the nodes. This is very easy given B.

One point extra credit. Show that the sum you wrote in part C is Θ(N).

Part D. Continue to assume the cost of visiting a node equals its height. Describe a class of binary trees for which the total cost of visiting the nodes is θ(N2). Naturally these will not be fully complete binary trees. Hint do problem 3.

2.3.4 Data Structures for representing trees

A vector-based implementation for Binary Trees

We store each node as the element of a vector. Store the root in element 1 of the vector and the key idea is that we store the two children of the element at rank r in the elements at rank 2r and 2r+1.

Draw a fully complete binary tree of height 3 and show where each element is stored.

Draw an incomplete binary tree of height 3 and show where each element is stored and that there are gaps.

There must be a way to tell leaves from internal nodes. The book doesn't make this explicit. Here is an explicit example. Let the vector S be given. With a vector we have the current size. S[0] is not used. S[1] has a pointer to the root node (or contains the root node if you prefer). For each S[i], S[i] is null (a special value) if the corresponding node doesn't exist). Then to see if the node v at rank i is a leaf, look at 2i. If 2i exceeds S.size() then v is a leaf since it has no children. Similarly if S[2i] is null, v is a leaf. Otherwise v is external.
How do you know that if S[2i] is null, then s[2i+1] will be null?
Ans: Our binary trees are proper.

This implementation is very fast. Indeed all tree operations are O(1) except for positions() and elements(), which produce n results and take time Θ(n).

Homework: R-2.7

================ Start Lecture #10 ================

However, this implementation can waste a lot of space since many of the entries in S might be unused. That is there may be many i for which S[i] is null.

Problem Set 2 problem 3. Give a tree with fewer than 20 nodes for which S.size() exceeds 100. Give a tree with fewer than 25 nodes for which S.size() exceeds 1000. Give a tree with fewer than 100 nodes for which S.size() exceeds a million.
End of problem set 2, due ??.

A linked structure for binary trees

Represent each node by a quadruple.

  1. A reference to the parent (null if we are at the root).
  2. A reference to the left child (null if at a leaf).
  3. A reference to the right child (null if at a leaf).
  4. The element().

Once again the algorithms are all O(1) except for positions() and elements(), which are Θ(n).

The space is Θ(n) which is much better that for the vector implementation. The constant is larger however since three pointers are stored for each position rather than one index.

A linked structure for general trees

The only difference is that we don't know how many children each node has. We could store k child pointers and say that we cannot process a tree having more than k children with the same parent.

Clearly we don't like this limit. Moreover, if we choose k moderate, say k=10. We are limited to 10-ary trees and for 3-ary trees most of the space is wasted.

So instead of storing the child references in the node, we store just one reference to a container. The container has references to the children. Imagine implementing the container as an extendable array.

Since a node v contains an arbitrary number of children, say Cv, the complexity of the children(v) iterator is Θ(Cv).

2.4 Priority Queues and Heaps

2.4.1 The Priority Queue Abstract Data Type

Up to now we have not considered elements that must be retrieved in a fixed order. But often in practice we assign a priority to each item and want the most important (highest priority) item first. (For some reason that I don't know, low numbers are often used to represent high priority.)

For example consider processor scheduling from Operating Systems (202). The simplest scheduling policy is FCFS for which a queue of ready processors is appropriate. But if we want SJF (short job first) then we want to extract the ready process that has the smallest remaining time. Hence a FIFO queue is not appropriate.

For a non-computer example,consider managing your todo list. When you get another item to add, you decide on its importance (priority) and then insert the item into the todo list. When it comes time to perform an item, you want to remove the highest priority item. Again the behavior is not FIFO.

Ordering properties and comparators

To return items in order, we must know when one item is less than another. For real numbers this is of course obvious.

We assume that each item has a key on which the priority is to be based. For the SJF example given above, the key is the time remaining. For the todo example, the key is the importance.

We assume the existence of an order relation (often called a total order) written ≤ satisfying for all keys s, t, and u.

Remark: For the complex numbers no such ordering exists that extends the natural ordering on the reals and imaginaries. This is unofficial (not part of 310).

Is it OK to define s≤t for all s and t?
No. That would not be antisymmetric.

Definition: A priority queue is a container of elements each of which has an associated key supporting the following methods.


Users may choose different comparison functions for the same data. For example, if the keys are longitude,latitude pairs, one user may be interested in comparing longitudes and another latitudes. So we consider a general comparator containing methods.

2.4.2 PQ-Sort, Selection-Sort, and Insertion-Sort

Given a priority queue it is trivial to sort a collection of elements. Just insert them and then do removeMin to get them in order. Written formally this is

Algorithm PQ-Sort(C,P)
   Input: an n element sequence C and an empty priority queue P
   Output: C with the elements sorted

   while not C.isEmpty() do
      P.insertItem(e,e)  // We are sorting on the element itself.

   while not P.isEmpty()

So whenever we give an implementation of a priority queue, we are also giving a sorting algorithm. Two obvious implementations of a priority queue give well known (but slow) sorts. A non-obvious implementation gives a fast sort. We begin with the obvious.

Implementing a priority queue with an unordered sequence

So insertItem() takes Θ(1) time and hence takes Θ(N) to insert all n items of C. But remove min, requires we go through the entire list. This requires time Θ(k) when there are k items in the list. Hence to remove all the items requires Θ(n+(n-1)+...+1) = Θ(N2) time.

This sorting algorithm is normally called selection sort since the dominant step is selecting the minimum each time.

Implementing a priority queue with an ordered sequence

Now removeMin() is trivial since it is just removeFirst(). But insertItem is Θ(k) when there are k items already in the priority queue since you must step through to find the correct location to insert and then slide the remaining elements over.

This sorting algorithm is normally called insertion sort since the dominant effort is inserting each element.

2.4.3 The Heap Data Structure

We now consider the non-obvious implementation of a priority queue that gives a fast sort (and a fast priority queue). The idea is to use a tree to store the elements, with the smallest element stored in the root. We will store elements only in the internal nodes (i.e., the leaves are not used). One could imagine an implementation in which the leaves are not even implemented since they are not used. We follow the book and draw internal nodes as circles and leaves (i.e., external nodes) as squares.
Since the priority queue algorithm will perform steps with complexity Θ(height of tree), we want to keep the height small. The way to do this is to fully use each level.

Definition: A binary tree of height h is complete if the levels 0,...,h-1 contain the maximum number of elements and on level h-1 all the internal nodes are to the left of all the leaves.


  1. Level i has 2i nodes, for 0≤i≤h-1.
  2. Level h contains only leaves.
  3. Levels less than h-1 contain no leaves.
  4. So really, for all levels, internal nodes are to the left of leaves.

Definition: A tree storing a key at each internal node satisfies the heap-order property if, for every node v other than the root, the key at v is no smaller than the key at v's parent.

Definition: A heap is a complete binary tree satisfying the heap order property.

Definition: The last node of a heap is the right most internal node in level h-1. In the diagrams above the last nodes are pink.

Remark: As written the ``last node'' is really the last internal node. However, we actually don't use the leaves to store keys so in some sense ``last node'' is the last (significant) node.

Homework: R-2.11, R-2.14

Implementing a Priority Queue with a Heap

With a heap it is clear where the minimum is located, namely at the root. We will also use a reference to the last node since insertions will occur at the first node after last.

Theorem: A heap storing n keys has height ⌈log(n+1)⌉


  1. Let h be the height of the tree and note that n is the number of internal nodes.
  2. Levels 0 through h-2 are full and contain only internal nodes (which contain keys).
  3. This gives 1+2+4+...+2h-2 keys.
  4. Homework: Show this is 2h-1-1 keys.
  5. Level h-1 contains between 1 and 2h-1 internal nodes each containing a key.
  6. Level h contains no internal nodes.
  7. So (2h-1-1)+1 ≤ n ≤ (2h-1-1)+2h-1
  8. So 2h-1 ≤ n ≤ 2h-1
  9. So 2h-1 < n+1 ≤ 2h
  10. So h-1 < log(n+1) ≤ h
  11. So ⌈log(n+1)&rceil = h

Corollary: If we can implement insert and removeMin in time Θ(height), we will have implemented the priority queue operations in logarithmic time (our goal).

Illustrate the theorem with the diagrams above.

================ Start Lecture #11 ================

The Vector Representation of a Heap

Since we know that a heap is complete, it is efficient to use the vector representation of a binary tree. We can actually not bother with the leaves since we don't ever use them. We call the last node w (remember that is the last internal node). Its index in the vector representation is n, the number of keys in the heap. We call the first leaf z; its index is n+1. Node z is where we will insert a new element and is called the insertion position.


This looks trivial. Since we know n, we can find n+1 and hence the reference to node z in O(1) time. But there is a problem; the result might not be a heap since the new key inserted at z might be less than the key stored at u the parent of z. Reminiscent of bubble sort, we need to bubble the value in z up to the correct location.

Up-Heap Bubbling

We compare key(z) with key(u) and swap the items if necessary. In the diagram on the right we added 45 and then had to swap it with 70. But now 45 is still less than its parent so we need to swap again. At worst we need to go all the way up to the root. But that is only Θ(n) as desired. Let's slow down and see that this really works.

Great. It works (i.e., is a heap) and there can only be O(log(n)) swaps because that is the height of the tree.

But wait! What I showed is that it only takes O(n) steps. Is each step O(1)?

Comparing is clearly O(1) and swapping two fixed elements is also O(1). Finding the parent of a node is easy (integer divide the vector index by 2). Finally, it is trivial to find the new index for the insertion point (just increase the insertion point by 1).

Remark: It is not as trivial to find the new insertion point using a linked implementation.

Homework: Show the steps for inserting an element with key 2 in the heap of Figure 2.41.


Trivial, right? Just remove the root since that must contain an element with minimum key. Also decrease n by one.
What remains is TWO trees.

We do want the element stored at the root but we must put some other element in the root. The one we choose is our friend the last node.

But the last node is likely not to be a valid root, i.e. it will destroy the heap property since it will likely be bigger than one of its new children. So we have to bubble this one down. It is shown in pale red on the right and the procedure explained below. We also need to find a new last node, but that really is trivial: It is the node stored at the new value of n.

Down-Heap Bubbling

If the new root is the only internal node then we are done.

If only one child of the root is internal (it must be the left child) compare its key with the key of the root and swap if needed.

If both children of the root are internal, choose the child with the smaller key and swap with the root if needed.

The original last node, became the root, and now has been bubbled down to level 1. But it might still be bigger than a child so we keep bubbling. At worst we need Θ(h) bubbling steps, which is again logarithmic in n as desired.

Homework: R-2.16

size, isEmptyO(1)
minElement, minKeyO(1)
insertItemΘ(log n)
removeMinΘ(log n)


The table on the right gives the performance of the heap implementation of a priority queue. As desired, the main operations have logarithmic time complexity. It is for this reason that heap sort is fast.

Summary of heaps

2.4.4 Heap-Sort (and some extras)

The goal is to sort a sequence S. We return to the PQ-sort where we insert the elements of S into a priority queue and then use removeMin to obtain the sorted version. When we use a heap to implement the priority queue, each insertion and removal takes Θ(log(n)) so the entire algorithm takes Θ(nlog(n)). The heap implementation of PQ-sort is called heap-sort and we have shown

Theorem: The heap-sort algorithm sorts a sequence of n comparable elements in Θ(nlog(n)) time.

Implementing Heap-Sort In Place

In place means that we use the space occupied by the input. More precisely, it means that the space required is just the input + O(1) additional memory. The algorithm above required Θ(n) addition space to store the heap.

The in place heap-sort of S assumes that S is implemented as an array and proceeds as follows (This presentation, beyond the definition of ``in place'' is unofficial; i.e., it will not appear on problem sets or exams)

  1. Logically divide the array into a portion in the front that contains the growing heap and the rest that contains the elements of the array that have not yet been dealt with.
  2. Do the insertions a with a normal heap-sort but change the comparison so that a maximum element is in the root (i.e., a parent is no smaller than a child).
  3. Now do the removals from the heap, moving the blue line back up.

Bottom-Up Heap Constructor (unofficial)

If you are given at the beginning all n elements that are to be inserted, the total insertion time for all inserts can be reduced to O(n) from O(nlog(n)). The basic idea assuming n=2n-1 is

  1. Take out the first element and call it r.
  2. Divide the remaining 2n-2 into two parts each of size 2n-1-1.
  3. Heap-sort each of these two parts.
  4. Make a tree with r as root and the two heaps as children.
  5. Down-heap bubble r.

Locaters (Unofficial)

Sometimes we wish to extend the priority queue ADT to include a locater that always points to the same element even when the element moves around. So if x is in a priority queue and another item is inserted, x may move during the up-heap bubbling, but the locater of x continues to refer to x.

Comparison of the Priority Queue Implementations


size, isEmpty Θ(1) Θ(1) Θ(1)
minElement, minKey Θ(n) Θ(1) Θ(1)
insertItem Θ(1) Θ(n) Θ(log(n))
removeMin Θ(n) Θ(1) Θ(log(n))

================ Start Lecture #13 ================

2.5 Dictionaries and Hash Tables

Dictionaries, as the name implies are used to contain data that may later be retrieved. Associated with each element is the key used for retrieval.

For example consider an element to be one student's NYU transcript and the key would be the student id number. So given the key (id number) the dictionary would return the entire element (the transcript).

2.5.1 the Unordered Dictionary ADT

A dictionary stores items, which are key-element (k,e) pairs.

We will study ordered dictionaries in the next chapter when we consider searching. Here we consider unordered dictionaries. So, for example, we do not support findSmallestKey. the methods we do support are

Trivial Implementation: log files

Just store the items in a sequence.

2.5.2 Hash Tables

The idea of a hash table is simple: Store the items in an array (as done for log files) but ``somehow'' be able to figure out quickly, i.e., Θ(1), which array element contains the item (k,e).

We first describe the array, which is easy, and then the ``somehow'', which is not so easy. Indeed in some sense it is impossible. What we can do is produce an implementation that, on the average, performs operations in time Θ(1).

Bucket Arrays

Allocate an array A of size N of buckets, each able to hold an item. Assume that the keys are integers in the range [0,N-1] and that no two items have the same key. Note that N may be much bigger than n. Now simply store the item (k,e) in A[k].

Analysis of Bucket Arrays

If everything works as we assumed, we have a very fast implementation: searches, insertions, and removals are Θ(1). But there are problems, which is why section 2.5 is not finished.

  1. The keys might not be unique (although in many applications they are): This is a simple example of a collision. We discuss collisions in 2.5.5.
  2. The keys might not be integers: We can always treat any (computer stored) object as an integer. Just view the object as a bunch of bits and then consider that a base two non-negative integer.
  3. But those integers might not be ``computer integers'', that is, they might have more bits than the largest integer type in our programming language: True, we discuss methods for converting long bit strings into computer integers in 2.5.3.
  4. But on many machines the number of computer integers is huge. We can't possibly have a bucket array of size 264: True, we discuss compressing computer integers into a smaller range in 2.5.4

2.5.3 Hash Functions

We need a hash function h that maps keys to integers in the range [0,N-1]. Then we will store the item (k,e) in bucket A[h(k)] (we are for now ignoring collisions). This problem is divided into two parts. A hash code assigns to each key a computer integer and then a compression map converts any computer integer into one in the range [0,N-1]. Each of these steps can introduce collisions. So even if the keys were unique to begin with, collisions are an important topic.

Hash Codes

A hash code assigns to any key an integer value. The problem we have to solve is that the key may have more bits than are permitted in our integer values. We first view the key as bunch of integer values (to be explained) and then combine these integer values into one.

If our integer values are restricted to 32 bits and our keys are 64 bits, we simply view the high order 32 bits as one value and the low order as another. In general if
⌈numBitsInKey / numBitsInIntegerValue⌉ = k
we view the key as k integer values. How should we combine the k values into one?

Summing Components

Simply add the k values.
But, but, but what about overflows?
Ignore them (or use exclusive or instead of addition).

Polynomial Hash Codes

The summing components method gives very many collisions when used for character strings. If 4 characters fill an integer value, then `temphash' and `hashtemp' will give the same value. If one decided to use integer values just large enough to hold one (unicode) character, then there would be many, many common collisions: `t21' and `t12' for one, mite and time for another.

If we call the k integer values x0,...,xk-1, then a better scheme for combining is to choose a positive integer value a and compute ∑xiai=x0+x1a+..xn-1an-1.

Same comment about overflows applies.

The authors have found that using a = 33, 37, 39, or 41 worked well for character strings that are English words.

2.5.4 Compression Maps

The problem we wish to solve in this section is to map integers in some, possibly large range, into integers in the range [0,N-1].
This is trivial! Why not map all the integers into 0.
We want to minimize collisions.

The Division Method

This is often called the mod method, especially if you use the ``correct'' definition of mod. One simple way to turn any integer x into one in the range [0,N-1] is to compute |x| mod N. That is we define the hash function h by

            h(x) = |x| mod N
(If we used the true mod we would not need the absolute value.)

Choosing N to be prime tends to lower the collision rate, but choosing N to be a power of 2 permits a faster computation since mod with a power of two simply means taking the low order bits.

The MAD Method

MAD stands for multiply-add-divide (mod is essentially division). We still use mod N to get the numbers in the range, but we are a little fancier and try to spread the numbers out first. Specifically we define the hash function h via.

            h(x) = |ax+b| mod N

The values a and b are chosen (often at random) as positive integers not a multiple of N.

2.5.5 Collision-Handling Schemes

The question we wish to answer is what to do when two distinct keys map to the same value, i.e., when h(k)=h(k'). In this case we have two items to store in one bucket. This discussion also covers the case where we permit multiple items to have the same key.

Separate Chaining

The idea is simple, each bucket instead of holding an item holds a reference to a container of items. That is each bucket refers to the trivial log file implementation of a dictionary, but only for the keys that map to this container.

The code is simple, you just error check and pass the work off to the trivial implementation used for the individual bucket.

Algorithm findElement(k):
    if B is empty then
        return NO_SUCH_KEY
    // now just do the trivial linear search
    return B.findElement(k)

Algorithm insertItem(k,e):
    if A[h(k)] is empty then
        Create B, an empty sequence-based dictionary

Algorithm removeElement(k)
    if B is empty then
        return NO_SUCH_KEY
        return B.removeElement(k)

Homework: R-2.19

Load Factors and Rehashing

We want the number of keys hashing to a given bucket to be small since the time to find a key at the end of the list is proportional to the size of the list, i.e., to the number of keys that hash to this value.

We can't do much about items that have the same key, so lets consider the (common) case where no two items have the same key.

The average size of a list is n/N, called the load factor, where n is the number of items and N is the number of buckets. Typically, one keeps the load factor below 1.0. The text asserts that 0.75 is common.

What should we do as more items are added to the dictionary? We make an ``extendable dictionary''. That is, as with an extendable array we double N and ``fix everything up'' In the case of an extendable dictionary, the fix up consists of recalculating the hash of every element (since N has doubled). In fact no one calls this an extendable dictionary. Instead one calls this scheme rehashing since one must rehash (i.e., recompute the hash) of each element when N is changed. Also N is normally chosen to be a prime number so instead of doubling, one chooses for the new N the smallest prime number above twice the old N.

================ Start Lecture #13 ================

Open Addressing

Separate chaining involves two data structures: the buckets and the log files. An alternative is to dispense with the log files and always store items in buckets, one item per bucket. Schemes of this kind are referred to as open addressing. The problem they need to solve is where to put an item when the bucket it should go into is already full? There are several different solutions. We study three: Linear probing, quadratic probing, and double hashing.

Linear Probing

This is the simplest of the schemes. To insert a key k (really I should say ``to insert an item (k,e)'') we compute h(k) and initially assign k to A[h(k)]. If we find that A[h(k)] contains another key, we assign k to A[h(k)+1]. It that bucket is also full, we try A[h(k)+2], etc. Naturally, we do the additions mod N so that after trying A[N-1] we try A[0]. So if we insert (16,e) into the dictionary at the right, we place it into bucket 2.

How about finding a key k (again I should say an item (k,e))? We first look at A[h(k)]. If this bucket contains the key, we have found it. If not try A[h(k)+1], etc and of course do it mod N (I will stop mentioning the mod N). So if we look for 4 we find it in bucket 1 (after encountering two keys that hashed to 6).
Or perhaps I should say incomplete. What if the item is not on the list? How can we tell?
Ans: If we hit an empty bucket then the item is not present (if it were present we would have stored it in this empty bucket). So 20 is not present.
What if the dictionary is full, i.e., if there are no empty buckets.
Check to see if you have wrapped all the way around. If so, the key is not present

What about removals?
Easy, remove the item creating an empty bucket.
I'm sorry you asked. This is a bit of a mess. Assume we want to remove the (item with) key 19. If we simply remove it, and search for 4 we will incorrectly conclude that it is not there since we will find an empty slot.
OK so we slide all the items down to fill the hole.
WRONG! If we slide 6 into the whole at 5, we will never be able to find 6.
So we only slide the ones that hash to 4??
WRONG! The rule is you slide all keys that are not at their hash location until you hit an empty space.

Normally, instead of this complicated procedure for removals, we simple mark the bucket as removed by storing a special value there. When looking for keys we skip over such slots. When an insert hits such a bucket, the insert uses the bucket. (The book calls this a ``deactivated item'' object).

Homework: R-2.20

Quadratic Probing

All the open addressing schemes work roughly the same. The difference is which bucket to try if A[h(k)] is full. One extra disadvantage of linear probing is that it tends to cluster the items into contiguous runs, which slows down the algorithm.

Quadratic probing attempts to spread items out by trying buckets A[h(k)], A[h(k)+1], A[h(k)+4], A[h(k)+9], etc. One problem is that even if N is prime this scheme can fail to find an empty slot even if there are empty slots.

Homework: R-2.21

Double Hashing

In double hashing we have two hash functions h and h'. We use h as above and, if A[h(k)] is full, we try, A[h(k)+h'(k)], A[h(k)+2h'(k)], A[h(k)+3h'(k)], etc.

The book says h'(k) is often chosen to be q - (k mod q) for some prime q < N. I note again that if mod were defined correctly this would look more natural, namely (q-k) mod q. We will not consider which secondary hash function h' is good to use.

Homework: R-2.22

Open Addressing vs Separate Chaining

A hard choice. Separate chaining seems to use more space, but that is deceiving since it all depends on the loading factor. In general for each scheme the lower the loading factor, the faster scheme but the more memory it uses.

2.5.6 Universal Hashing (skipped)

2.6 Java Example: Heap (skipped)

Chapter 3 Search Trees and Skip Lists

3.1 Ordered Dictionaries and Binary Search Trees

We just studied unordered dictionaries at the end of chapter 2. Now we want to extend the study to permit us to find the "next" and "previous" items. More precisely we wish to support, in addition to findElement(k), insertItem(k,e), and removeElement(k), the new methods

We naturally signal an exception if no such item exists. For example if the only keys present are 55, 22, 77, and 88, then closestKeyAfter(90) or closestElemBefore(2) each signal an exception.

We begin with the most natural implementation.

3.1.1 Sorted Tables

We use the sorted vector implementation from chapter 2 (we used it as a simple implementation of a priority queue). Recall that this keeps the items sorted in key order. Hence it is Θ(n) for inserts and removals, which is slow; however, we shall see that it is fast for finding and element and for the four new methods closestKeyBefore(k) and friends. We call this a lookup table.

The space required is Θ(n) since we grow and shrink the array supporting the vector (see extendable arrays).

As indicated the key favorable property of a lookup table is that it is fast for (surprise) lookups using the binary search algorithm that we study next.

Binary Search

In this algorithm we are searching for the rank of the item containing a key equal to k. We are to return a special value if no such key is found.

The algorithm maintains two variables lo and hi, which are respectively lower and upper bounds on the rank where k will be found (assuming it is present).

Initially, the key could be anywhere in the vector so we start with lo=0 and hi=n-1. We write key(r) for the key at rank r and elem(r) for the element at rank r.

We then find mid, the rank (approximately) halfway between lo and hi and see how the key there compares with our desired key.

Some care is need in writing the algorithm precisely as it is easy to have an ``off by one error''. Also we must handle the case in which the desired key is not present in the vector. This occurs when the search range has been reduced to the empty set (i.e., when lo exceeds hi).

Algorithm BinarySearch(S,k,lo,hi):
    Input:  An ordered vector S containing (key(r),elem(r)) at rank r
            A search key k
            Integers lo and hi
    Output: An element of S with key k and rank between lo and hi.
            NO_SUCH_KEY if no such element exits

If lo > hi then
    return NO_SUCH_KEY                    // Not present

mid ← ⌊(lo+hi)/2⌋
if k = key(mid) then
    return elem(mid)                     // Found it
if k < key(mid) then
    return BinarySearch(S,k,lo,mid-1)    // Try bottom ``half''
if k > key(mid) then
    return BinarySearch(S,k,mid+1,hi)    // Try top ``half''

Do some examples on the board.

Analysis of Binary Search

It is easy to see that the algorithm does just a few operations per recursive call. So the complexity of Binary Search is Θ(NumberOfRecursions). So the question is "How many recursions are possible for a lookup table with n items?".

The number of eligible ranks (i.e., the size of the range we still must consider) is hi-lo+1.

The key insight is that when we recurse, we have reduced the range to at most half of what it was before. There are two possibilities, we either tried the bottom or top ``half''. Let's evaluate hi-lo+1 for the bottom and top half. Note that the only two possibilities for ⌊(lo+hi)/2⌋ are (lo+hi)/2 or (lo+hi)/2-(1/2)=(lo+hi-1)/2

(mid-1)-lo+1 = mid-lo = ⌊(lo+hi)/2⌋-lo ≤ (lo+hi)/2-lo = (hi-lo)/2 < (hi-lo+1)/2

hi-(mid+1)+1 = hi-mid = hi-⌊(lo+hi)/2⌋ ≤ hi-(lo+hi-1)/2 = (hi-lo+1)/2

So the range starts at n and is halved each time and remains an integer (i.e., if a recursive call has a range of size x, the next recursion will be at most ⌊x/2⌋).

Write on the board 10 times
(X-1)/2 ≤ ⌊X/2⌋ &le X/2
If B ≤ A, then Z-A ≤ Z-B

How many recursions are possible? If the range is ever zero, we stop (and declare the key is not present) so the longest we can have is the number of times you can divide by 2 and stay at least 1. That number is Θ(log(n)) showing that binary search is a logarithmic algorithm.

================ Start Lecture #14 ================

Problem Set 3, Problem 1 Write the algorithm closestKeyBefore. It uses the same idea as BinarySearch.

When you do question 1 you will see that the complexity is Θ(log(n)). Proving this is not hard but is not part of the problem set.

When you do question 1 you will see that closestElemBefore, closestKeyAfter, and closestElemAfter are all very similar to closestKeyBefore. Hence they are all logarithmic algorithms. Proving this is not hard but is not part of the problem set.

Performance Summary of the Array-Based Implementations

MethodLog FileLookup Table
findElementΘ(n)Θ(log n)
closestKeyBeforeΘ(n)Θ(log n)
closestElemBeforeΘ(n)Θ(log n)
closestKeyAfterΘ(n)Θ(log n)
closestElemAfterΘ(n)Θ(log n)

Our goal now is to find a better implementation so that all the complexities are logarithmic. This will require us to shift from vectors to trees.

3.1.2 Binary Search Trees

This section gives a simple tree-based implementation, which alas fails to achieve the logarithmic bounds we seek. But it is a good start and motivates the AVL trees we study in 3.2 that do achieve the desired bounds.

Definition: A binary search tree is a tree in which each internal node v stores an item such that the keys stored in every node in the left subtree of v are less than or equal to the key at v which is less than or equal to every key stored in the right subtree.

From the definition we see easily that an inorder traversal of the tree visits the internal nodes in nondecreasing order of the keys they store.

You search by starting at the root and going left or right if the desired key is smaller or larger respectively than the key at the current node. If the key at the current node is the key you seek, you are done. If you reach a leaf the desired key is not present.

Do some examples using the tree on the right. E.g. search for 17, 80, 55, and 65.

Homework: R-3.1 and R-3.2

3.1.3 Searching in a Binary Search Tree

Here is the formal algorithm described above.

Algorithm TreeSearch(k,v)
Input:  A search key k and a node v of a binary search tree.
Output: A node w in the subtree routed at v such that either
        w is internal and k is stored at w or
        w is a leaf where k would be stored if it existed

if v is a leaf then
    return v
if k=k(v) then
    return v
if k<k(v) then
    return TreeSearch(k,T.leftChild(v))
if k>k(v) then
    return TreeSearch(k,T.rightChild(v))

Draw a tree on the board and illustrate both finding a k and no such key exists.

Analysis of Binary Tree Searching

It is easy to see that only a couple of operations are done per recursive call and that each call goes down a level in the tree. Hence the complexity is O(height).

So the question becomes "How high is a tree with n nodes?". As we saw last chapter the answer is "It depends.".

Next section we will learn a technique for keeping trees low.

3.1.4 Insertion in a Binary Search Tree

To insert an item with key k, first execute w←TreeSearch(k,T.root()). Recall that if w is internal, k is already in w, and if w is a leaf, k "belongs" in w. So we proceed as follows.

Draw examples on the board showing both cases (leaf and internal returned).

Once again we perform a constant amount of work per level of the tree implying that the complexity is O(height).

3.1.5 Removal in a Binary Search Tree

This is the trickiest part, especially in one case as we describe below. The key concern is that we cannot simply remove an item from an internal node and leave a hole as this would make future searches fail. The beginning of the removal technique is familiar: w=TreeSearch(k,T.root()). If w is a leaf, k is not present, which we signal.

If w is internal, we have found k, but now the fun begins. Returning the element with key k is easy, it is the element stored in w. We need to actually remove w, but we cannot leave a hole. There are three cases.

  1. If we are lucky both of w's children are leaves. Then we can simply replace w with a leaf. (Recall that leaves do not contain items.) This is the trivial case

  2. The next case is when one child of w is a leaf and the other, call it z, is an internal node. In this case we can simply replace w by z; that is have the parent of w now point to z. This removes w as desired and also removes the leaf child of w, which is OK since leaves do not contain items. This is the easy case.

  3. Note that the above two cases can be considered the same. In both cases we notice that one child of w is a leaf and replace w by the other child (and its descendents, if any).

  4. Now we get to the difficult case: both children of w are internal nodes. What we will do is to replace the item in w with the item that has the next highest key.

It is again true that the complexity is O(height) but this is not quite as easy to see as before. We start with a TreeSearch, which is Θ(height). This gets us to w. The most difficult case is the last one where w has two internal children. We spend a non-constant time on node w because we need to find the next item. But this operation is only O(height) and we simply descend the tree deeper.

3.1.6 Performance of Binary Search Trees

Time complexity of the binary search tree ADT. We use h for the height and s for the number of elements in the tree.
size, isEmptyO(1)
findElement, insertItem, removeElementO(h)
findAllElements, removeAllElementsO(h+s)

We have seen that findElement, insertItem, and removeElement have complexity O(height). It is also true, but we will not show it, that one can implement findAllElements and removeAllElements in time O(height+numberOfElements). You might think removeAllElements should be constant time since the resulting tree is just a root so we can make it in constant time. But removeAllElements must also return an iterator that when invoked must generate each of the elements removed.

Comments on average vs worst case behavior

In a sense that we will not make precise, binary search trees have logarithmic performance since `most' trees have logarithmic height.

Nonetheless we know that there are trees with height Θ(n). You produced several of them for problem set 2. For these trees binary search takes linear time, i.e., is slow. Our goal now is to fancy up the implementation so that the trees are never very high. We can do this since the trees are not handed to us. Instead they are build up using our insertItem method.

================ Start Lecture #15 ================

Problem Set 2 is (actually, should have been) assigned. It is due tues 4 Nov.

3.2 AVL Trees

Named after its inventors Adel'son-Vel'skii and Landis.

Definition: An AVL tree is a binary search tree that satisfies the height-balance property, which means that for every internal node, the height of the two children can differ by at most 1.

Homework: Draw an AVL tree of height 3 where each left child has height one greater than its sibling.

Since the algorithms for an AVL tree require the height to be computed and this is an expensive operation, each node of an AVL tree contains its height (say as v.height()).

Remark: You can actually be fancier and store just two bits to tell whether the node has the same height as its sibling, one greater, or one smaller.

We see that the height-balance property prevents the tall skinny trees you developed in problem set 2. But is it really true that the height must be O(log(n))?

Yes it is and we shall now prove it. Actually we will prove instead that an AVL tree of height h has at least 2(h/2)-1 internal nodes from which the desired result follows easily.

Lemma: An AVL tree of height h has at least 2(h/2)-1 internal nodes.

Proof: Let n(h) be the minimum number of internal nodes in an AVL tree of height h.

n(1)=1 and n(2)=2. So the lemma holds for h=1 and h=2.

Here comes the key point.

Consider an AVL tree of height h≥3 and the minimum number of nodes. This tree is composed of a root, and two subtrees. Since the whole tree has the minimum number of nodes for its height so do the subtrees. For the big tree to be of height h, one of the subtrees must be of height h-1. To get the minimum number of nodes the other subtree is of height h-2.
Why can't the other subtree be of height h-3 or h-4?
The height of siblings can differ by at most 1!

What the last paragraph says in symbols is that for h≥3,
n(h) = 1+n(h-1)+n(h-2)

The rest is just algebra, i.e. has nothing to do with trees, heights, searches, siblings, etc.

n(h) > n(h-1) so n(h-1) > n(h-2). Hence

n(h) > n(h-1)+n(h-2) > 2n(h-2)

Really we could stop here. We have shown that n(h) at least doubles when h goes up by 2. This says that n(h) is exponential in h and hence h is logarithmic in n. But we will proceed slower. Applying the last formula i times we get

For any i>0,   n(h) > 2in(h-2i)    (*)

Let's find an i so that h-2i is guaranteed to be 1 or 2. This would guarantee that n(h-2i) ≥ 1.
I claim i = ⌈h/2⌉-1 works.

If h is even h-2i = h-(h-2) = 2

If h is odd h-2i = h - (2⌈h/2⌉-2) = h - ((h+1)-2) = 1

Now we plug this value of i into equation (*) and get for h≥3

n(h) > 2in(h-2i)
   = 2⌈h/2⌉-1n(h-2i)
   ≥ 2⌈h/2⌉-1(1)
   ≥ 2(h/2)-1

Theorem: the height of an AVL tree storing n items is O(log(n)).

Proof: From the lemma we have n(h) > 2(h/2)-1.

Taking logs gives log(n(h)) > (h/2)-1 or

h < 2log(n(h))+2

Since n(h) is the smallest number of nodes possible for an AVL tree of height h, we see that h < 2 log(n) for any AVL tree of height h.

================ Start Lecture #16 ================

Midterm Exam

================ Start Lecture #17 ================
Notes on midterm:
  1. Midterm handed back at the end of class
  2. An answer sheet was also handed out.
  3. The breakdown was 90-100: 14. 80-89: 15, 70-79: 9, 60-69: 3, 0-59: 1.
  4. We will review the answers in this week's recitations.

3.2.1 Update Operations


Begin by a standard binary search tree insertion. In the diagrams on the right, black shows the situation before the insertion; red after. The numbers are the heights. Ignore the blue markings, they are explained in the text as needed.

  1. Do a find for the key to be inserted.

  2. Presumably wind up at a leaf (meaning the key not already in the tree).

  3. Call this leaf w.

  4. Insert the item by converting w to an internal node at height one (i.e., the new internal node has two leaves as children).

  5. Call the new internal node w. We view the previous operation as expanding the leaf w into an internal node (with leaves as children).

  6. In the diagram on the right, w is converted from a black leaf into a red height 1 node.

Why aren't we finished?
Ans:The tree may no longer be in balance, i.e. it may no longer by an AVL tree.

We look for an imbalanced node, i.e., a node whose children have heights differing by more than 1. If we find such a node, the tree is no longer AVL and we must perform a re-balancing operation.

  1. We start our search for an unbalanced node at w since that is where the tree has changed. Indeed the height of w has increased from 0 to 1. This height increase of w may cause a height increase of w's parent, which in tern may cause a height increase of w's grandparent, etc. That is, we may need to search the entire path from w up to the root, i.e. the ancestors of w. But we know the key fact that no other node has had its height changed.

  2. As we traverse the path from w to the root we will be looking at the ancestors of w. For each ancestor A we consider, we will also check the sibling of A. Let h be the height of A before the insert. Since the tree was AVL before the insert, we know that the height of A's sibling was and is h-1, h, or h+1. We shall see that these three cases lead us to three different actions.

    1. h-1 will be the "bad" case and will require a re-balancing operation.

    2. h will be the "unknown" case, we will need to proceed up the tree to see if a re-balance is needed.

    3. h+1 will be the "good" case, we will be able to stop and no re-balancing will be needed.

  3. We begin at w which was at height 0 and is now at height 1. Its children are leaves so w itself is in balance.

  4. Since w was at height 0, the sibling must have had height -1, 0, or +1. The height of the sibling has not changed (it is not on the path from w to the root). Since a height of -1 is impossible, we have just two possibilities, the "unknown" and "good" cases.

  5. If the sibling was at height 1, their parent was, and still is, at height 2. Hence no heights other than w's changed and the tree remains in balance. So in this case, which is illustrated in the figure above, we are indeed done.

  6. If the sibling of w was (and hence is) at height 0, we have the "unknown" case. The siblings heights now differ by 1, which is OK, but their parent's height has changed from 1 to 2 (since w is at height 1). This situation is illustrated at the right. Note that not all of the tree is shown. Also note that the black heights are before the insert and the reds are after (ignore the blue for now).

  7. Since w's parent P has had its height changed (from 1 to 2) we need to check P's sibling. The old height is 1 so the sibling must be at height 0, 1, or 2. (It is 1 in the diagram.) If we had the "good" case (height=2) the new height of P and the height of P's sibling would be equal and the height of their parent would not have changed so we would have been done (without re-balancing).

  8. In the diagram we again have the "unknown" case and proceed up to P's parent. Another "unknown" case (node's old height and its siblings height are equal, namely 2) and we keep going up.

  9. We get a final unknown case (both heights 3) and finally reach the root, which was at height 4. Since it has no sibling there is nothing that can be imbalanced so we are done, again with no re-balancing required.

  10. What is the problem? We just proceed up as in the figure on the right and eventually we hit the root and are done.

  11. WRONG.

  12. In the figure, before the insertion, siblings had the same height, the "unknown" case. This is certainly possible but not required.

  13. If an ancestor of w had height one less than its sibling, then the insertion has made them equal. That is the "good" case: The height of the parent doesn't change and we are done as illustrated in the previous (small) figure.

  14. The "bad" case occurs when an ancestor of w had height one greater than its sibling. In this case the insertion has made it two greater, which is not permitted for an AVL tree.

  15. For example the light blue node could have originally been at height 2 instead of 3. We no longer have an AVL tree, the root is imbalanced: Its children have height 2 (the blue) and 4 (the red). We must re-balance this tree.

The top left diagram illustrates the problem case from the previous figure (k was 3 above).
Definition: The operation we will perform when x, y, and z lie in a straight line is called a single rotation. When x, y, and z form an angle, the operation is called a double rotation.

Let us consider the upper left double rotation, which is the rotation that we need to apply to the example above. It is redrawn to the right with the subtrees drawn and their heights labeled. The colors are there so that you can see where they trees go when we perform the rotation.

Recall that x is an ancestor of w and has had its height raised from k-1 to k. The sibling of x is at height k-1 and so is the sibling of y. The reason x had its height raised is that one of its siblings (say the right one) has been raised from k-2 to k-1.
How do I know the other one is k-2?
Ans: It must have been the "unknown" case or we would not have proceeded further up the tree.

The double rotation transforms the picture on top to the one on the bottom. We now actually are done with the insertion. Let's check the bottom picture and make sure.

  1. The order relation remains intact. That is, every node is greater than its entire left subtree and less than its entire right subtree.

  2. The tree (now rooted at x) is balanced.

  3. Nodes y and z are each at height k. Hence x, the root of this tree is at height k+1.

  4. The tree above, rooted at z, has height k+2.

  5. But remember that before the insert z was at height k+1.

  6. So the rotated tree (which is after the insert) has the same height as the original tree before the insert.

  7. Hence every node above z in the original tree keeps its original height so the entire tree is now balanced.

Thus, if an insertion causes an imbalance, just one rotation re-balances the tree globally. We will see that for removals it is not this simple.

Here are the three pictures for the remaining three possibilities. That is, the other double rotation and both single rotations. The original configuration is shown on top and the result after the rotation is shown immediately below.

Homework: R-3.3, R-3.4, R-3.5

What is the complexity of insertion?

  1. Let n be the number of nodes in the tree before the insertion.

  2. Finding the insertion point is Θ(log n).

  3. Expanding the leaf and inserting the item is Θ(1).

  4. Walking up the tree looking for an imbalance is Θ(1) per level, which is O(log n) since the tree has height Θ(log n).

  5. Performing the one needed rotation is Θ(1).

Hence we have the following

Theorem: The complexity of insertion in an AVL tree is Θ(log n).

================ Start Lecture #18 ================

I forgot to write homework 16 on the board so I will accept it next tuesday as well as today. Also I forgot to ask for homework 15 last time so I will accept it today.

Problem Set 3, problem 2. Please read the entire problem before beginning.

  1. Draw an avl tree containing items with integer keys. Draw the internal nodes as circles and write the key inside the circle. Draw the leaves as squares; leaves do not contain items. You will want to read the remaining parts before drawing this tree.
  2. Choose a key not present in the tree you drew in part A whose insertion will require the "other" double rotation in order to restore balance (i.e., the double rotation shown in the diagram above showing one double and two single rotations). Draw the tree after the insertion, but prior to the rotation. Then draw the tree after the rotation.
  3. Choose a key not present in the tree you drew in part A whose insertion will require a single rotation in order to restore balance. Draw the tree after the insertion, but prior to the rotation. Then draw the tree after the rotation.
  4. Choose a key not present in the tree you drew in part A whose insertion will require the "other" single rotation in order to restore balance Draw the tree after the insertion, but prior to the rotation. Then draw the tree after the rotation.


In order to remove an item with key k, we begin just as we did for an ordinary binary search tree. I repeat the procedure here.

The key concern is that we cannot simply remove an item from an internal node and leave a hole as this would make future searches fail. The beginning of the removal technique is familiar: w=TreeSearch(k,T.root()). If w is a leaf, k is not present, which we signal.

If w is internal, we have found k, but now the fun begins. Returning the element with key k is easy, it is the element stored in w. We need to actually remove w, but we cannot leave a hole. There are three cases.

  1. If we are lucky both of w's children are leaves. Then we can simply replace w with a leaf. (Recall that leaves do not contain items.) This is the trivial case

  2. The next case is when one child of w is a leaf and the other, call it z, is an internal node. In this case we can simply replace w by z; that is have the parent of w now point to z. This removes w as desired and also removes the leaf child of w, which is OK since leaves do not contain items. This is the easy case.

  3. Note that the above two cases can be considered the same. In both cases we notice that one child of w is a leaf and replace w by the other child (and its descendents, if any).

  4. Now we get to the difficult case: both children of w are internal nodes. What we will do is to replace the item in w with the item that has the next highest key.

But now we have to possibly restore balance, i.e., maintain the AVL property. The possible trouble is that the light green node on the left has been replaced by the light blue on the right, which is of height one less. This might cause a problem. The sibling of the light green (shown in purple) might have height equal to, one less than, or one greater than, the light green.

  1. If the purple height was equal to the light green, it is one greater than the light blue, which is still in balance and their parent (red) has not changed height so all is well.
  2. If the sibling's height was one less than the light green, it is equal to the light blue, which remains in balance. But red's height has dropped by one so we need to look at it to see if there is trouble.
  3. If the sibling's height was one greater than the light green, it is two greater than the light blue so we are out of balance.

To summarize the three cases either

  1. We are balanced and done.
  2. We proceed up the tree to red, the parent of the light blue, and try again.
  3. We must re-balance.

In the second case we move up the tree and again have one of the same three cases so either.

  1. We will hit a case 1 and are done.
  2. We keep hitting case 2, keep going up and eventually reach the root. Done.
  3. We hit case 3 and need to re-balance.

The re-balancing is again a single or double rotation (since the problem is the same, so is the solution).

The rotation will fix the problem but the result has a highest node whose height is one less the highest prior to the rotation (in the diagrams for single and double rotation the largest height dropped from k+2 to k+1).

Unlike the case for insertions, this height reduction does not cancel out a previous height increase. Thus the lack of balance may continue to advance up the tree and more rotations may be needed.

Homework: R-3.6

Problem Set 3 problem 3 (end of problem set 3). Please read the entire problem before beginning.

  1. Draw an avl tree containing items with integer keys. Draw the internal nodes as circles and write the key inside the circle. Draw the leaves as squares; leaves do not contain items. You will want to read the remaining parts before drawing this tree.
  2. Choose a key present in the tree you drew in part A whose removal will require a double rotation and a single rotation in order to restore balance. Draw the tree after the removal, but prior to the rotations. Then draw the tree after the double rotation, but prior to the single rotation. Finally, draw the tree after both rotations.

What is the complexity of a removal? Remember that the height of an AVL tree is Θ(log(N)), where N is the number of nodes.

  1. We must find a node with the key, which has complexity Θ(height) = Θ(log(N)).
  2. We must remove the item: Θ(1).
  3. We must re-balance the tree, which might involve Θ(height) rotations. Since each rotation is Θ(1), the complexity is Θ(log(N))*Θ(1) = Θ(log(N)).

Theorem: The complexity of removal for an AVL tree is logarithmic in the size of the tree.

3.2.2 Performance

The news is good. Search, Inserting, and Removing all have logarithmic complexity.

The three operations all involve a sweep down the tree searching for a key, and possibly an up phase where heights are adjusted and rotations are performed. Since only a constant amount of work is performed per level and the height is logarithmic, the complexity is logarithmic.

================ Start Lecture #19 ================


  1. The last parts of problem set 3 are in the web page for lecture 18. The problem set is officially assigned today and is due in 3 lectures, 13 November.
  2. I rewrote parts of lecture #17 and am now a little happier with the description of AVL insertion. The revised text is in the "Giant Page" and I also extracted a revised lecture #17 to put in "Individual Lectures"

3.3 Bounded-Depth Search Trees

Might come back to this if time permits.

3.4 Splay Trees (skipped)

3.5 Skip Lists (skipped)

3.6 Java Example: AVL and Red-Black Trees (skipped)

Chapter 4 Sorting, Sets, and Selection

We already did a sorting technique in chapter 2. Namely we inserted items into a priority queue and then removed the minimum each time. When we use a heap to implement the priority, the resulting sort is called heap-sort and is asymptotically optimal. That is, its complexity of O(Nlog(N)) is as fast as possible if we only use comparisons (proved in 4.2 below)

4.1 Merge-Sort

4.1.1 Divide-and-Conquer

The idea is that if you divide an enemy into small pieces, each piece, and hence the enemy, can be conquered. When applied to computer problems divide-and-conquer involves three steps.

  1. Divide the problem into smaller subproblems.
  2. Solve each of the subproblems, normally via a recursive call to the original procedure.
  3. Combine the subproblem solutions into a solution for the original problem.

In order to prevent an infinite sequence of recursions, we need to define a stopping condition, i.e., a predicate that informs us when to stop dividing (because the problem is small enough to solve directly).

Using Divide-and-Conquer for Sorting

This turns out to be so easy that it is perhaps surprising that it is asymptotically optimal. The key observation is that merging two sorted lists is fast (the time is linear in the size of the lists).

The steps are

  1. Divide (with stopping condition): If S has zero or one element, simply return S since it is already sorted. Otherwise S has n≥2 elements: Move the first ⌈n/2⌉ elements of S into S1 and the remaining ⌊n/2⌋ elements into S2.
  2. Solve recursively: Recursively sort each of the two subsequences.
  3. Combine: Merge the two (now sorted) subsequences back into S

Example:: Sort {22, 55, 33, 44, 11}.

  1. Divide {22, 55, 33, 44, 11} into {22, 55, 33} and {44, 11}
  2. Recursively sort {22, 55, 33} and {44, 11} getting {22, 33, 55} and {11, 44}
  3. Merge {22, 33, 55} and {11, 44} getting {11, 22, 33, 44, 55}

Expanding the recursion one level gives.

  1. Divide {22, 55, 33, 44, 11} into {22, 55, 33} and {44, 11}
  2. Recursively sort {22, 55, 33} and {44, 11} getting {22, 33, 55} and {11, 44}
    1. Divide {22, 55, 33} into {22, 55} and {33}
    2. Recursively sort {22, 55} and {33} getting {22, 55} and {33}
    3. Merge {22, 55} and {33} getting {22, 33, 55}

    1. Divide {44, 11} into {44} and {11}
    2. Recursively sort {44} and {11} getting {44} and {11}
    3. Merge {44} and {11} getting {11, 44}
  3. Merge {22, 33, 55} and {11, 44} getting {11, 22, 33, 44, 55}

Expanding again gives

  1. Divide {22, 55, 33, 44, 11} into {22, 55, 33} and {44, 11}
  2. Recursively sort {22, 55, 33} and {44, 11} getting {22, 33, 55} and {11, 44}
    1. Divide {22, 55, 33} into {22, 55} and {33}
    2. Recursively sort {22, 55} and {33} getting {22, 55} and {33}
      1. Divide {22, 55} into {22} and {55}
      2. Recursively sort {22} and {55} getting {22} and {55}
      3. Merge {22} and {55} getting {22, 55}

      1. Do NOT divide {33} since it has only one element and hence is already sorted
    3. Merge {22, 55} and {33} getting {22, 33, 55}

    1. Divide {44, 11} into {44} and {11}
    2. Recursively sort {44} and {11} getting {44} and {11}
      1. Do NOT divide {44} since it has only one element and hence is already sorted

      1. Do NOT divide {11} since it has only one element and hence is already sorted
    3. Merge {44} and {11} getting {11, 44}
  3. Merge {22, 33, 55} and {11, 44} getting {11, 22, 33, 44, 55}

Finally there still is one recursion to do so we get.

  1. Divide {22, 55, 33, 44, 11} into {22, 55, 33} and {44, 11}
  2. Recursively sort {22, 55, 33} and {44, 11} getting {22, 33, 55} and {11, 44}
    1. Divide {22, 55, 33} into {22, 55} and {33}
    2. Recursively sort {22, 55} and {33} getting {22, 55} and {33}
      1. Divide {22, 55} into {22} and {55}
      2. Recursively sort {22} and {55} getting {22} and {55}
        1. Do NOT divide {22} since it has only one element and hence is already sorted.

        1. Do NOT divide {55} since it has only one element and hence is already sorted.
      3. Merge {22} and {55} getting {22, 55}

      1. Do NOT divide {33} since it has only one element and hence is already sorted
    3. Merge {22, 55} and {33} getting {22, 33, 55}

    1. Divide {44, 11} into {44} and {11}
    2. Recursively sort {44} and {11} getting {44} and {11}
      1. Do NOT divide {44} since it has only one element and hence is already sorted

      1. Do NOT divide {11} since it has only one element and hence is already sorted
    3. Merge {44} and {11} getting {11, 44}
  3. Merge {22, 33, 55} and {11, 44} getting {11, 22, 33, 44, 55}

Hopefully there is a better way to describe this action. How about the following picture. The left tree shows the dividing. The right shows the result of the merging.

Definition: We call the above tree the merge-sort-tree.

Homework: Draw the merge sort tree for {55, 33, 11, 22, 44}.

In a merge-sort tree the left and right children of a node with A elements have ⌈A/2⌉ and ⌊A/2⌋ elements respectively.

Theorem: Let n be the size of the sequence to be sorted. If n is a power of 2, say n=2k, then the height of the merge-sort tree is log(n)=k. In general n = ⌈log(n)⌉.

Proof: The power of 2 case is part of problem set 4. I will do n=2k+1. When we divide the sequence into 2, the larger is ⌈n/2⌉=2k-1+1: To see this write 2k+1 in binary. When we keep dividing and look at the larger piece (i.e., go down the leftmost branch of the tree) we will eventually get to 3=21+1. We have divided by 2 k-1 times so are at depth k-1. But we have to go two more times (getting 2 and then 1) to get one element. So the depth is of this element is k+1 and hence the height is at least k+1. The other leaves are all at height no more than k+1 (in fact they are all k). Thus the height of the tree is exactly k+1.

The rest is part of problem set 2. Here is the idea. If you increase the number of elements in the root, you do not decrease the height. If n is not a power of 2, it is between two powers of 2 and hence the height is between the heights for these powers of two. My calculation for 1 more than a power of 2, tells you what the height must be.

Problem Set 4, problem 1. Prove the theorem when n is a power of 2. Prove the theorem for n not a power of 2 either by finishing my argument or coming up with a new one.

The reason we need the theorem is that we will show that merge-sort spends time O(n) at each level of the tree and hence spends time O(n*height) in total. The theorem shows that this is O(nlog(n)).

Merging Two Sorted Sequences

This is quite clear. View each sequence as a deck of cards face up. Look at the top of each deck and take the smaller card. Keep going. When one deck runs out take all the cards from the other. This is clearly constant time per card and hence linear in the number of cards Θ(#cards).

Unfortunately this doesn't look quite so clear in pseudo code when we write it in terms of the ADT for sequences. Here is the algorithm essentially copied from the book. At least the comments are clear.

Algorithm merge(S1, S2, S):
   Input:  Sequences S1 and S2 sorted in nondecreasing order,
           and an empty sequence S
   Output: Sequence S contains the elements previously in S1 and S2
           sorted in nondecreasing order.  S1 and S2 now empty.

   {Keep taking smaller first element until one sequence is empty}
   while (not(S1.isEmpty() or S2.isEmpty()) do
     if S1.first().element()<S2.first().element() then
        {move first element of S1 to end of S}
        {move first element of S2 to end of S}

   {Now take the rest of the nonempty sequence.}
   {We simply take the rest of each sequence.}
   {Move the remaining elements of S1 to S
   while (not S1.isEmpty()) do
   {Move the remaining elements of S2 to S
   while (not S2.isEmpty()) do

Homework: R-4.2

Examining the code we see that each iteration of each loop removes an element from either S1 or S2. Hence the total number of iterations is S1.size()+S2.size(). Since each iteration requires constant time we get the following theorem.

Theorem: Merging two sorted sequences takes time Θ(n+m), where n and m are the sizes of the two sequences.

The Running Time of Merge-Sort

We characterize the time in terms of the merge-sort tree. We assign to each node of the tree the time to do the divide and merge associated with that node and to invoke (but not to execute) the recursive calls. So we are essentially charging the node for the divide and combine steps, but not the solve recursively.

This does indeed account for all the time. Illustrate this with the example tree I drew at the beginning.

For the remainder of the analysis we assume that the size of the sequence to be sorted is n, which is a power of 2. It is easy to extend this to arbitrary n as we did for the theorem that is a part of problem set 4.

Now much time is charged to the root? The divide step takes time proportional to the number of elements in S, which is n. The combine step takes time proportional to the sum of the number of elements in S1 and the number of elements in S1. But this is again n. So the root node is charged Θ(n).

The same argument shows that any node is charged Θ(A), where A is the number of items in the node.

How much time is charged to a child of the root? Remember that we are assuming n is a power of 2. So each child has n/2 elements and is charged a constant times n/2. Since there are two children the entire level is charged Θ(n).

In this way we see that each level is Θ(n). Another way to see this is that the total number of elements in a level is always n and a constant amount of work is done on each.

Now we use the theorem saying that the height is log(n) to conclude that the total time is Θ(nlog(n)). So we have the following theorem.

Theorem: Merge-sort runs in Θ(nlog(n)).

4.1.2 Merge-Sort and Recurrence Equations

Here is yet another way to see that the complexity is Θ(nlog(n)).

Let t(n) be the worst-case running time for merge-sort on n elements.

Remark: The worst case and the best case just differ by a multiplicative constant. There is no especially easy or hard case for merge-sort.

For simplicity assume n is a power of two. Then we have for some constants B and C

t(1) = B
t(n) = 2t(n/2)+Cn    if n>1
The first line is obvious the second just notes that to solve a problem we must solve 2 half-sized problems and then do work (merge) that is proportional to the size of the original problem.

If we apply the second line to itself we get
t(n) = 2t(n/2)+Cn = 2[2t(n/4)+C(n/2)]+Cn = 22t(n/22)+2Cn

If we apply this i times we get

t(n) = 2it(n/2i)+iCn

When should we stop?
Ans: when we get to the base case t(1). This occurs when 2i=n, i.e. when i=log(n).

t(n)   =  2log(n)t(n/2log(n)) + log(n)Cn
       =  nt(n/n) + log(n)Cn               since 2log(n)=n
       =  nt(1) + log(n)Cn
       =  nB + Cnlog(n)
which again shows that t(n) is Θ(nlog(n)).

================ Start Lecture #20 ================

4.2 The Set Abstract Data Type

Skipped for now

4.2.1 A Simple Set Implementation

4.2.2 Partitions with Union-Find Operations

4.2.3 A Tree-Based Partition Implementation

4.3 Quick-Sort

It is interesting to compare quick-sort with merge-sort. Both are divide and conquer algorithms. So we divide, recursively sort each piece, and then combine the sorted pieces.

In merge-sort, the divide is trivial: throw half the elements into one pile and the other half in another pile. The combine step, while easy does do comparisons and picks an element from the correct pile.

In quick-sort, the combine is trivial: pick up one pile, then the other. The divide uses comparisons to decide which pile each element should be placed into.

As usual we assume that the sequence we wish to sort contains no duplicates. It is easy to drop this condition if desired.

The book's algorithm does quick-sort in place, that is the sorted sequence overwrites the original sequence. Mine produces a new sorted sequence and leaves the original sequence alone.

Algorithm quick-sort (S)
   Input: A sequence S (of size n).
   Output: A sorted sequence T containing
           the same elements as S.

   if n < 2
      T ← S

   Create empty sequences L and G { standing for less and greater }

   { Divide into L and G }
   Pick an element P from S { called the pivot }
   while (not S.isEmpty())
      x ← S.remove(S.first())
      if x < P then
      if x > P then

   { Recursively Sort L and G }
   LS ← quick-sort (L)    { LS stands for L sorted }
   GS ← quick-sort (G)

   { Combine LS, P, and GS }
   while (not LS.isEmpty())
   while (not GS.isEmpty())

Running Time of Quick-Sort

The running time of quick sort is highly dependent on the choice of the pivots at each stage of the recursion. A very simple method is to choose the last element as the pivot. This method is illustrated in the figure on the right. The pivot is shown in red. This tree is not surprisingly called the quick-sort tree.

The top tree shows the dividing and recursing that occurs with input {33,55,77,11,66,88,22,44}. The tree below shows the combining steps for the same input.

As with merge sort, we assign to each node of the tree the cost (i.e., running time) of the divide and combine steps. We also assign to the node the cost of the two recursive calls, but not their execution. How large are these costs?

The two recursive calls (not including the subroutine execution itself) are trivial and cost Θ(1).

The dividing phase is a simple loop whose running time is linear in the number of elements divided, i.e., in the size of the input sequence to the node. In the diagram this is the number of numbers inside the oval.

Similarly, the combining phase just does a constant amount of work per element and hence is again proportional to the number of elements in the node.

We would like to make an argument something like this. At each level of each of the trees the total number of elements is n so the cost per level is O(n). The pivot divides the list in half so the size of the largest node is divided by two each level. Hence the number of levels, i.e., the height of the tree, is O(log(n)). Hence the entire running time is O(nlog(n)).

That argument sound pretty good and perhaps we should try to make it more formal. However, I prefer to try something else since the argument is WRONG!

Homework: Draw the quick-sort tree for sorting the following sequence {222 55 88 99 77 444 11 44 22 33 66 111 333}. Assume the pivot is always the last element.

The Worst Case for Quick-Sort

The tree on the right illustrates the worst case of quick-sort, which occurs when the input is already sorted!

The height of the tree is N-1 not O(log(n)). This is because the pivot is in this case the largest element and hence does not come close to dividing the input into two pieces each about half the input size.

It is easy to see that we have the worst case. Since the pivot does not appear in the children, at least one element from level i does not appear in level i+1 so at level N-1 you can have at most 1 element left. So we have the highest tree possible. Note also that level i has at least i pivots missing so can have at most N-i elements in all the nodes. Our tree achieves this maximum. So the time needed is proportional to the total number of numbers written in the diagram which is N + N-1 + N-2 + ... + 1, which is again the one summation we know N(N+1)/2 or Θ(N2.

Hence the worst case complexity of quick-sort is quadratic! Why don't we call it slow sort?

Perhaps the problem was in choosing the last element as the pivot. Clearly choosing the first element is no better; the same example on the right again illustrates the worst case (the tree has its empty nodes on the left this time).

Since are spending linear time (as opposed to constant time) on the division step, why not count how many elements are present (say k) and choose element number k/2? This would not change the complexity (it is also linear). You could do that and now a sorted list is not the worst case. But some other list is. Just put the largest element in the middle and then put the second largest element in the middle of the node on level 1. This does have the advantage that if you mistakenly run quick-sort on a sorted list, you won't hit the worst case. But the worst case is still there and it is still Θ(N2).

Why not choose the real middle element as the pivot, i.e., the median. That would work! It would cut the sizes in half as desired. But how do we find the median? We could sort, but that is the original problem. In fact there is a (difficult) algorithm for computing the median in linear time and if this is used for the pivot, quick-sort does take O(Nlog(N)) time in the worst case. However, the difficult median algorithm is not fast in practice. That is, the constants hidden in saying it is Θ(N) are rather large.

Instead of studying the fast, difficult median algorithm, we will consider a randomized quick-sort algorithm and show that the expected running time is Θ(Nlog(N)).

Problem Set 4, Problem 2. Find a sequence of size N=12 giving the worst case for quick-sort when the pivot for sorting k elements is element number ⌊k/2⌋.

4.3.1 Randomized Quick-Sort

Consider running the following quick-sort-like experiment.

  1. Pick a positive integer N
  2. Pick a N numbers at random (say without duplicates).
  3. Choose at random one of the N numbers and call it the pivot.
  4. Split the other numbers into two piles: those bigger than the pivot and those smaller than the pivot.
  5. Call this a good split if the two piles are "nearly" equal. Specifically, the split is good if the larger of the piles has no more than 3N/4 elements (which is the same as the smaller pile having no fewer than N/4 elements)

Are good splits rare or common?

Theorem: (From probability theory). The expected number of times that a fair coin must be flipped until it shows ``heads'' k times is 2k.

We will not prove this theorem, but will apply it to analyze good splits.

Theorem: The expected running time of randomized quick-sort on N numbers is O(Nlog(N)).

Proof: First we prove a little log lemma.

Lemma: logba = (log(a)) / log(b).

Proof of Lemma: First note that alog(b) = blog(a) (take the log of both sides).
Now raise each side to the power 1/log(b).
(alog(b))(1/log(b) = (blog(a))1/log(b).
Thus a(log(b))/log(b) = b(log(a))/log(b) (since (xy)z = xyz, see homework below).
So a = b(log(a))/log(b), which says that (log(a))/log(b) is the power to which one raises b in order to get a. That is the definition of logba.
End of Proof of Lemma

Proof of Theorem: We picked the pivot at random. Therefore, if we imagine the N numbers lined up in order, the pivot is equally likely to be anywhere in this line.

Consider the picture on the right. If the pivot is anywhere in the pink, the split is good. But the pink is half the line so the probability that we get a ``pink pivot'' (i.e., a good split) is 1/2. This is the same probability that a fair coin comes up heads.

Every good split divides the size of the node by at least 4/3. Recall that if you divide N by 4/3, log4/3(N) times, you will get 1. So the maximum number of good splits possible along a path from the root to a leaf is log4/3(N).

Applying the probability theorem above we see that the expected length of a path from the root to a leaf is at most 2log4/3(N). By the lemma this is (2log(N))/log(4/3), which is Θ(log(N)). That is, the expected height is O(log(N)).

Since the time spent at each level is O(N), the expected running time of randomized quick-sort is O(Nlog(N)). End of Proof of Theorem

Homework: Show that (xy)z = xyz. Hint take the log of both sides.

4.4 A Lower Bound on Comparison-Based Sorting

Theorem: The running time of any comparison-based sorting algorithm is Ω(Nlog(N)).

Proof: We will not cover this officially. Unofficially, the idea is we form a binary tree with each node corresponding to a comparison performed by the algorithm and the two children corresponding to the two possible outcomes. This is a tree of all possible executions (with only comparisons used for decisions). There are N! permutations of N numbers and each must give a different execution pattern in order to be sorted. So there are at least N! leaves. Hence the height is at least log(N!). But N! has N/2 elements that are at least N/2 so N!≥(N/2)N/2. Hence
height ≥ log(N!) ≥ log((N/2)N/2) = (N/2)log(N/2)
So the running time, which is at least the height of this tree, is Ω(Nlog(N))

Corollary: Heap-sort, merge-sort, and quick sort (with the difficult, linear-time, not-done-in-class median algorithm) are asymptotically optimal.

================ Start Lecture #21 ================

4.5 Bucket-Sort and Radix-Sort

We have seen that the fastest comparison-based sorting algorithms run in time Θ(Nlog(N)), where N is the number of items to sort. In this section we are going to develop faster algorithms. Hence they must not be comparison-based algorithms.

We make a key assumption, we are sorting items whose keys are integers in a bounded range [0, R-1].

Question: Let's start with a special case R=N so we are sorting N items with integer keys from 0 to N-1. As usual we assume there are no duplicate keys. Also as usual we remark that it is easy to lift this restriction. How should we do this sort?

Answer: That was a trick question. You don't even have to look at the input. If you tell me to sort 10 integers in the range 0...9 and there are no duplicates, I know the answer is {0,1,2,3,4,5,6,7,8,9}. Why? Because, if there are no duplicates, the input must consist of one copy of each integer from 0 to 10.

4.5.1 Bucket-sort

OK, let's drop the assumption that R=N. So we have N items (k,e), with each k an integer, no duplicate ks, and 0≤k<R. The trick is that we can use k to decide where to (temporarily) store e.

Algorithm preBucketSort(S)
   input:  A sequence S of N items with integer keys in range [0,N)
   output: Sequence S sorted in increasing order of the keys.

   let B be a vector of R elements,
       each initially a special marker indicating empty

   while (not S.isEmpty())
      (k,e) ← S.remove(S.first())
      B[k] ← e         <==== the key idea (not a comparison)

   for i ← 0 to R-1 do
      if (B[i] not special marker) then

To convert this algorithm into bucket sort we drop the artificial assumption that there are not duplicates. Now instead of a vector of items we need a vector of buckets, where a bucket is a sequence of items.

Algorithm BucketSort(S)
   input:  A sequence S of N items with integer keys in range [0,N)
   output: Sequence S sorted in increasing order of the keys.

   let B be a vector of R sequences of items
       each initially empty

   while (not S.isEmpty())
      (k,e) ← S.remove(S.first())
      B[k].insertLast(e)      <==== the key idea

   for i ← 0 to R-1 do
      while (not B[i].isEmpty())

Complexity of Bucket-Sort

The first loop has N iterations each of which run in time Θ(1), so the loop requires time Θ(N). The for loop has R iterations so the while statement is executed R times. Each while statement requires time Θ(1) (excluding the body of the while) so all of them require time Θ(N) The total number of iterations of all the inner while loops is again N and each again requires time Θ(1), so the time for all inner iterations is Θ(N).

The previous paragraph shows that the complexity is Θ(N)+Θ(R) = Θ(N+R).

So bucket-sort is a winner if R is not too big. For example if R=O(N), then bucket-sort requires time only Θ(N). Indeed if R=o(Nlog(N)), bucket-sort is (asymptotically) faster than any comparison based sorting algorithm (using worst case analysis).

Stable Sorting

Definition: We call a sort stable if equal elements remain in the same relative position. Stated more formally: for any two items (ki,ei) and (kj,ej) such that item (ki,ei) precedes item (kj,ej) in S (i.e., i<j), then item (ki,ei) precedes item (kj,ej) after sorting as well.

Stability is often convenient as we shall see in the next section on radix-sort. We note that bucket-sort is stable since we treated each bucket in a fifo manner inserting at the rear and removing from the front.

4.5.2 Radix-Sort

Let's extend our sorting study from keys that are integers to keys that are pairs of integers. The first question to ask is, given two keys (k,m) and (k',m'), which is larger? Note that (k,m) is just the key; an item would be written ((k,m),e).

Definition: The lexicographical (dictionary) ordering on pairs of integers is defined by declaring (k,m) < (k',m') if either

Note that this really is dictionary order:
canary < eagle < egret < heron
10 < 11 < 12 < 2

Algorithm radix-sort-on-pairs
   input:  A sequence S of N items with keys
           pairs of integers in the range [0,N)
           Write elements of S as ((k,m),e)
   output: Sequence S lexicographically sorted on the keys

   bucket-sort(S) using m as the key
   bucket-sort(S) using k as the key

Do an example of radix sorting on pairs.

Do an incorrect sort but starting with the most significant element of the pair.

Do an incorrect sort by using an individual sort that is not stable.

What if the keys are triples or in general d-tuples?
The answer is ...

Homework: R-4.15

Theorem: Let S be a sequence of N items each of which has a key (k1,k2,...kd), where each ki is in integer in the range [0,R). We can sort S lexicographically in time O(n(N+R)) using radix-sort.

4.6 Comparison of Sorting Algorithms

Insertion sort or bubble sort are not suitable for general sorting of large problems because their running time is quadratic in N, the number of items. For small problems, when time is not an issue, these are attractive because they are so simple. Also if the input is almost sorted, insertion sort is fast since it can be implemented in a way that is O(N+A), where A is the number of inversions, (i.e., the number of pairs out of order).

Heap-sort is a fine general-purpose sort with complexity Θ(Nlog(N)), which is optimal for comparison-based sorting. Also heap-sort can be executed in place (i.e., without much extra memory beyond the data to be sorted). (The coverage of in-place sorting was ``unofficial'' in this course.) If the in-place version of heap-sort fits in memory (i.e., if the data is less than the size of memory), heap-sort is very good.

Merge-sort is another optimal Θ(Nlog(N)) sort. It is not easy to do in place so is inferior for problems that can fit in memory. However, it is quite good when the problem is too large to fit in memory and must be done ``out-of-core''. We didn't discuss this issue, but the merges can be done with two input and one output file (this is not trivial to do well, you want to utilize the available memory in the most efficient manner).

Quick-sort is hard to evaluate. The version with the fast median algorithm is fine theoretically (worst case again Θ(Nlog(N)) but not used because of large constant factors in the fast median. Randomized quick-sort has a low expected time but a poor worst-case time. It can be done in place and is quite fast in that case, often the fastest. But the quadratic worst case is a fear (and a non-starter for many real-time applications).

Bucket and radix sort are wonderful when they apply, i.e., when the keys are integers in a modest range (R a small multiple of N). For radix sort with d-tuples the complexity is Θ(d(N+R)) so if d(N+R) is o(Nlog(N)), radix sort is asymptotically faster than any comparison based sort (e.g., heap-, insertion-, merge-, or quick-sort).

================ Start Lecture #22 ================

4.7 Selection (officially skipped; unofficial comments follow)

Selection means the ability to find the kth smallest element. Sorting will do it, but there are faster (comparison-based) methods. One example problem is finding the median (N/2 th smallest).

It is not too hard (but not easy) to implement selection with linear expected time. The surprising and difficult result is that there is a version with linear worst-case time.

4.7.1 Prune-and-Search

The idea is to prune away parts of the set that cannot contain the desired element. This is easy to do as seen in the next algorithm. The less easy part is to show that it takes O(n) expected time. The hard part is to modify the algorithm so that it takes O(n) worst case time.

4.7.2 Randomized Quick-Select

Algorithm quickSelect(S,k)
   Input:  A sequence S of n elements and an integer k in [1,n]
   Output: The kth smallest element of S

   if n=1 the return the (only) element in S

   pick a random element x of X
   divide S into 3 sequences
      L, the elements of S that are less than x
      E, the elements of S that are equal to x
      G, the elements of S that are greater than x

   { Now we reduce the search to one of these three sets }

   if k≤|L|      then return quickSelect(L,k)
   if k>|L|+|E|  then return quickSelect(G,k-|L|+|E|
   return x  { We want an element in E; all are equal to x }

4.7.3 Analyzing Randomized Quick-Select (skipped)

4.8 Java Example: In-Place Quick-Sort (skipped)

Chapter 5 Fundamental Techniques

5.1 The Greedy Method

The greedy method is applied to maximization/minimization problems. The idea is to at each decision point choose the configuration that maximizes/minimizes the objective function so far. Clearly this does not lead to the global max/min for all problems, but it does for a number of problems.

This chapter does not make a good case for the greedy method. The method is used to solve simple variants of standard problems, but the the standard variants are not solved with the greedy method. There are better examples, for example the minimal spanning tree and shortest path graph problems. The two algorithms chosen for this section, fractional knapsack and task scheduling, were (presumably) chosen because they are simple and natural to solve with the greedy method.

5.1.1 The Fractional Knapsack Method

In the knapsack problem we have a knapsack of a fixed capacity (say W pounds) and different items i each with a given weight wi and a given benefit bi. We want to put items into the knapsack so as to maximize the benefit subject to the constraint that the sum of the weights must be less than W.

The knapsack problem is actually rather difficult in the normal case where one must either put an item in the knapsack or not. However, in this section, in order to illustrate greedy algorithms, we consider a much simpler variation in which we can take a portion, say xi≤wi, of an item and get a proportional part of the benefit. This is called the ``fractional knapsack problem'' since we can take a fraction of an item. (The more common knapsack problem is called the ``0-1 knapsack problem'' since we must either take all (1) or none (0) of an item).

More formally, for each item i we choose an amount xi (0≤xi≤wi) that we will place in the knapsack. We are subject to the constraint that the sum of the xi is no more than W since that is all the knapsack can hold.

We desire to maximize the total benefit. Since, for item i, we only put xi in the knapsack, we don't get the full benefit. Specifically we get benefit (xi/wi)bi.

But now this is easy!

Why doesn't this work for the normal knapsack problem when we must take all of an item or none of it?


algorithm FractionalKnapsack(S,W):
   Input:  Set S of items i with weight wi and benefit bi all positive.
           Knapsack capacity W>0.
   Output: Amount xi of i that maximizes the total benefit without
           exceeding the capacity.

   for each i in S do
      xi ← 0        { for items not chosen in next phase }
      vi ← bi/wi    { the value of item i "per pound" }
   w ← W            { remaining capacity in knapsack }

   while w > 0 and S is not empty
      remove from S an item of maximal value   { greedy choice }
      xi ← min(wi,w)  { can't carry more than w more }
      w ← w-xi


FractionalKnapsack has time complexity O(NlogN) where N is the number of items in S.

Homework: R-5.1

5.1.2 Task Scheduling

We again consider an easy variant of a well known, but difficult, optimization problem.

In the figure there are 6 tasks, with start times and finishing times (1,3), (2,5), (2,6), (4,5), (5,8), (5,7). They are scheduled on three machines M1, M2, M3. Clearly 3 machines are needed as can be seen by looking at time 4.

Note that a good solution to this problem has three objectives that must be met.

  1. We must generate a feasible solution. That is, all the tasks assigned to a single machine are non-conflicting. A schedule that does not satisfy this condition is invalid and an algorithm that produces invalid solutions is wrong.

  2. We want to generate a feasible solution using the minimal number of machines. A schedule that uses more than the minimal number of machines might be called non-optimal or inefficient and an algorithm that produces such schedules might be called poor.

  3. We want to generate optimal solutions in the (asymptotically) smallest amount of time possible. An algorithm that generates optimal solutions using more time might be called slow.

Let's illustrate the three objectives with following example consisting of four tasks having starting and stopping times (1,3), (6,8), (2,5), (4,7). It is easy to construct a wrong algorithm, for example

Algorithm wrongTaskSchedule(T)
   Input:  A set T of tasks, each with start time si and
           finishing time fi (si≤fi).
   Output: A schedule of the tasks.

   while T is not empty do
      remove from T the first task and call it i.
      schedule i on M1

When applied to our 4-task example, the result is all four tasks assigned to machine 1. This is clearly infeasible since the last two tasks conflict.

It is also not hard to produce a poor algorithm, one that generates feasible, but non-optimal solutions.

Algorithm poorTaskSchedule(T):
   Input:  A set T of tasks, each with start time si and
   finishing time fi (si≤fi).
   Output: A feasible schedule of the tasks.

   m ← 0                         { current number of machines }
   while T is not empty do
      remove from T a task i
      m ← m+1
      schedule i on Mm

On the 4-task example, poorTaskSchedule puts each task on a different machine. That is certainly feasible, but is not optimal since the first and second task can go on one machine.

Hence it looks as though we should not put a task on a new machine if it can fit on an existing machine. That is certainly a greedy thing to do. Remember we are minimizing so being greedy really means being stingy. We minimize the number of machines at each step hoping that will give an overall minimum. Unfortunately, while better, this idea does not give optimal schedules. Let's call it mediocre.

Algorithm mediocreTaskSchedule(T):
   Input:  A set T of tasks, each with start time si and
   finishing time fi (si≤fi).
   Output: A feasible schedule of the tasks of T

   m ← 0                         { current number of machines }
   while T is not empty do
      remove from T a task i
      if there is an Mj having all tasks non-conflicting with i then
         schedule i on Mj
         m ← m+1
         schedule i on Mm

When applied to our 4-task example, we get the first two tasks on one machine and the last two tasks on separate machines, for a total of 3 machines. However, a 2-machine schedule is possible, as we will shall soon see.

The needed new idea is to processes the tasks in order. Several orders would work, we shall order them by start time. If two tasks have the same start time, it doesn't matter which one is put ahead. For example, we could view the start and finish times as pairs and use lexicographical ordering. Alternatively, we could just sort on the first component.


Algorithm taskSchedule(T):
   Input:  A set T of tasks, each with start time si and
   finishing time fi (si≤fi).
   Output: An optimal schedule of the tasks of T

   m ← 0                         { current number of machines }
   while T is not empty do
      remove from T a task i with smallest start time
      if there is an Mj having all tasks non-conflicting with i then
         schedule i on Mj
         m ← m+1
         schedule i on Mm

When applied to our 4-task example, we do get a 2-machine solution: the middle two tasks on one machine and the others on a second. But is this the minimum? Certainly for this example it is the minimum; we already saw that a 1-machine solution is not feasible. But what about other examples?

Correctness (i.e. Minimality of m)

Assume the algorithm runs and declares m to be the minimum number of machines needed. We must show that m are really needed.

OK taskSchedule is feasible and optimal. But is it slow? That actually depends on some details of the implementation and it does require a bit of cleverness to get a fast solution.


Let N be the number of tasks in T. The book asserts that it is easy to see that the algorithm runs in time O(NlogN), but I don't think this is so easy. It is easy to see O(N2).

================ Start Lecture #23 ================
  1. I added a diagram and a number of words to the previous lecture. The online notes have been updated.
  2. The following homework and problem set logically should have been assigned last time as they concern that material.
End of Notes.

Homework: R-5.3

Problem Set 4, Problem 3.
Part A. C-5.3 (Do not argue why your algorithm is correct).
Part B. C-5.4.

Remark: Problem set 4 (the last problem set) is now complete and due in 3 lectures, thurs 2 Dec. 2003.

5.2 Divide-and-Conquer

The idea of divide and conquer is that we solve a large problem by solving a number of smaller problems and then we apply this idea recursively.

5.2.1 Divide-and-Conquer Recurrence Equations

From the description above we see that the complexity of a divide and conquer solution has three parts.

  1. The time required to split the problem into subproblems.
  2. The time required to solve all the subproblems.
  3. The time required to combine the subproblem solutions

Let T(N) be the (worst case) time required to solve an instance of the problem having time N. The time required to split the problem and combine the subproblems is also typically a function of N, say f(N).

More interesting is the time required to solve the subproblems. If the problem has been split in half then the time required for each subproblem is T(N/2).

Since the total time required includes splitting, solving both subproblems, and combining we get.

    T(N) = 2T(N/2)+f(N)

Very often the splitting and combining are fast, specifically linear in N. Then we get

    T(N) = 2T(N/2)+rN
for some constant r. (We probably should say ≤ rather than =, but we will soon be using big-Oh and friends so we can afford to be a little sloppy.)

What if N is not divisible by 2? We should be using floor or ceiling or something, but we won't. Instead we will be assuming for recurrences like this one that N is a power of two so that we can keep dividing by 2 and get an integer. (The general case is not more difficult; but is more tedious)

But that is crazy! There is no integer that can be divided by 2 forever and still give an integer! At some point we will get to 1, the so called base case. But when N=1, the problem is almost always trivial and has a O(1) solution. So we write either

           r            if N = 1
    T(N) =
           2T(N/2)+rN   if N > 1
    T(1) = r
    T(N) = 2T(N/2)+rN   if N > 1

No we will now see three techniques that, when cleverly applied, can solve a number of problems. We will also see a theorem that, when its conditions are met, gives the solution without our being clever.

The Iterative Substitution Method

Also called ``plug and chug'' since we plug the equation into itself and chug along.

    T(N) = 2T(N/2)+rN
         = 2[    ]+rN
         = 2[2T((N/2)/2)+r(N/2)]+rN
         = 4T(N/4)+2rN    now do it again
         = 8T(N/8)+3rN

A flash of inspiration is now needed. When the smoke clears we get

    T(N) = 2iT(N/2i)+irN

When i=log(N), N/2i=1 and we have the base case. This gives the final result

    T(N) = 2log(N)T(N/2log(N))+irN
         = N     T(1)       +log(N)rN
         = N     r          +log(N)rN
         = rN+rNlog(N)
Hence T(N) is O(Nlog(N))

The Recursion Tree

The idea is similar but we use a visual approach.

We already studied the recursion tree when we analyzed merge-sort. Let's look at it again.

The diagram shows the various subproblems that are executed, with their sizes. (It is not always true that we can divide the problem so evenly as shown.) Then we show that the splitting and combining that occur at each node (plus calling the recursive routine) only take time linear in the number of elements. For each level of the tree the number of elements is N so the time for all the nodes on that level is Θ(N) and we just need to find the height of the tree. When the tree is split evenly as illustrated the sizes of all the nodes on each level go down by a factor of two so we reach a node with size 1 in logN levels (assuming N is a power of 2). Thus T(N) is Θ(Nlog(N)).

The Guess and Test Method

This method is really only useful after you have practice in recurrences. The idea is that, when confronted with a new problem, you recognize that it is similar to a problem you have seen the solution to previously and you guess that the solution to the new problem is similar to the old.

You then plug your guess into the recurrence and test that it works. For example if we guessed that the solution of

    T(1) = r
    T(N) = 2T(N/2)+rN   if N > 1
    T(N) = rN+rNlog(N)
we would plug it in and check that
    T(1) = r
    T(N) = 2T(N/2)+rN   if N > 1

But we don't have enough experience for this to be very useful.

The Master Method

In this section, we apply the heavy artillery. The following theorem, which we will not prove, enables us to solve some problems by just plugging in. It essentially does the guess part of guess and test for us.

We will only be considering complexities of the form

    T(1) = c
    T(N) = aT(N/b)+f(N)   if N > 1

The idea is that we have done some sort of divide and conquer where there are a subproblems of size at most N/b. As mentioned earlier f(N) accounts for the time to divide the problem into subproblems and to combine the subproblem solutions.

Theorem [The Master Theorem]: Let f(N) and T(N) be as above.

  1. (f is small) If there is a constant ε>0 such that f(N) is O(Nlogba/nε), then T(N) is Θ(Nlogba).
  2. (f is medium) If there is a constant k≥0 such that f(N) is θ(Nlogba(logN)k), then T(N) is θ(Nlogba(logN)k+1).
  3. (f is large) If there are constants ε>0 and &delta<1 such that f(N) is Ω(NlogbaNε) and af(N/b)≤δf(N), then T(N) is Θ(f(N)).

Proof: Not given.


  1. When we say f is small/medium/large we are comparing it to the mysterious Nlogba, which is the star of the show.

  2. So when f is small, T is polynomial; when f is medium, T is polynomial times a log power; and when f is big, T is f.

  3. Nlogba/Nε can also be written as Nlogba-ε and NlogbaNε can also be written as Nlogba+ε.

Now we can solve some problems easily and will do two serious problems after that.

Example: T(N) = 4T(N/2)+N.

Example: T(N) = 2T(N/2) + Nlog(N)

================ Start Lecture #24 ================
  1. We did all the examples in recitation so will go quickly.
  2. The final will be in room 303 EDUC. Where is EDUC?

Example: T(N) = T(N/4) + 2N

Example: T(N) = 9T(N/3) + N2.5

Example: T(N) = 9T(N/3) + N2

Example: T(N) = 2T(N/2) + rN (our original problem)

Homework: R-5.4

5.2.2 Integer Multiplication

We want to multiply big integers.

When we multiply without machine help, we use as knowledge the times and addition tables for all 1 digit numbers. In some sense our ``internal computer'' has as primitive operations the multiplication and addition of two 1-digit number. To multiply larger numbers, we apply the 5th-grade algorithm that requires Θ(N2) steps to multiply two N-digit numbers.

Computers typically have a single instruction to multiply two 32-bit numbers (or 64-bit on some machines). One way to enable a computer to multiply two 32N-bit numbers would be to implement the 5-th grade algorithm in software (using 32-bit numbers as ``digits'') and again compute the result in Θ(N2) time.

To make the wording easier let's consider the ``human'' case where the primitive operations deal with 1-bit numbers and we want to multiply two N-bit numbers X and Y. We also assume that N is a power of 2. As we said above, using the 5th grade algorithm we can perform the multiplication in time Θ(N2). We want to go faster.

If we simply divide the bits in half (the high order N/2, and the low order N/2), we see that to multiply X and Y we need to compute 4 sub-products Xhi*Yhi, Xhi*Ylo, Xlo*Yhi, Xlo*Ylo. The 5th grade algorithm specifies that we must put them in the correct "column" and add.

                Xhi Xlo
             x  Yhi Ylo
        Xhi*Ylo Xlo*Ylo
Xhi*Yhi Xlo*Yhi

Since addition of K-bit numbers is Θ(K) using the 3rd grade algorithm, and our multiplications are of N/2-bit values we get

   T(N) = 4T(N/2)+cn
We now excitedly apply the master theorem (case 1). Unfortunately the result is T(N)=Θ(N2), which is no improvement. We try again and with cleverness are able to get by with three instead of four multiplications.

We compute Xhi*Yhi and Xlo*Ylo as before, but then compute the miraculous (Xhi-Xlo)*(Ylo-Yhi). The miracle multiplication produces the two remaining terms we need, but has added to this two other terms. However, however those terms are just the (negative of) the two terms we did compute so we can add them back and they cancel.

Do this on the board.

The summary is that T(N) = 3T(N/2)+cN. (This c is larger than the one we had before.) Now the master theorem (case 1) gives

T(N) = Θ(Nlog23N)

Since log23<1.585, we get the surprising

Theorem: We can multiple two N-bit numbers in o(N1.585) time.

Homework: R-5.5

================ Start Lecture #25 ================

5.2.3 Matrix Multiplication

If you thought the integer multiplication involved pulling a rabbit out of our hat, get ready.

This algorithm was a sensation when it was discovered by Strassen. The standard algorithm for multiplying two NxN matrices is Θ(N3). We want to do better.

Do a matrix multiplication on the board and show that it is Θ(N3). One way to see this is to note that each entry in the product is an inner product. There are Θ(N2) entries and computing an inner product is Θ(N).

First try. Assume N is a power of 2 and break each matrix into 4 parts as shown on the right. Then X = AE+BG and similarly for Y, Z, and W. This gives 8 multiplications of half size matrices plus Θ(N2) scalar addition so

    T(N) = 8T(N/2) + bN2
We apply the master theorem and get T(N) = Θ(N3), i.e., no improvement.

But strassen found the way! He (somehow, I don't know how) decided to consider the following 7 (not 8) multiplications of half size matrices.

S1 = A(F-H)       S2 = (A+B)H       S3 = (C+D)E       S4 = D(G-E)
S5 = (A+D)(E+H)   S6 = (B-D)(G+H)   S7 = (A-C)(E+F)

Now we can compute X, Y, Z, and W from the S's

X = S5+S6+S4-S2     Y = S1+S2    Z = S3+S4    W = S1-S7-S3+S5

This computation shows that

    T(N) = 7T(N/2) + bN2

Thus the master theorem now gives

Theorem(Strassen): We can multiply two NxN matrices in time O(nlog7).


  1. log7<2.808 so we can multiply NxN matrices in time o(N2.808), which is o(N3). Amazing.
  2. There are even better algorithms, which are more complicated. I think they break the matrix into more pieces. The current best is O(N2.376).

Homework: R-5.6

================ Start Lecture #26 ================
Lecture given by Ricard Cole

5.3 Dynamic Programming

This is a little hard to explain abstractly.

One idea is that we compute a max by maximizing over a bunch of smaller possibilities (and apply this recursively). We then compute these maxima backwards. That is, instead of breaking the big one into small ones, we start with the small ones.

Another idea is that we save many intermediate results since they will be reused. So we really should analyze the space complexity as well as the time complexity.

5.3.1 Matrix Chain-Product

In this problem we want to multiple n matrices

           A0 * A1 * ... * An-1

Unlike the matrix multiplication from last lecture, these matrices are not all of the same size. But to be able to multiply the matrices, the number of columns in Ai must be the same as the number of rows in the next matrix Ai+1. Specifically Ai is of size di by di+1.

Matrix multiplication is associative, that is

         X * (Y * Z) = (X * Y) * Z
Hence in computing the matrix chain-product above, we can do the multiplications is any order we want and get the same answer. Perhaps surprisingly, some orders are much faster than others.

Example: Consider A * B * C where
A is a 20x60 matrix
B is a 60x50 matrix
C is a 50x10 matrix

First we perform the computation as A * (B * C).

B * C requires 50 multiplications for each entry in the output. Since the output is 60x10, there are 600 entries and hence 30,000 multiplications are required to produce the 60x10 result.

We now must multiple A by this new 60x10 matrix. This matrix product requires 60 multiplications for each entry in the output. Since the output is 20x10, there are 200 entries and hence 12,000 multiplications are required. So the overall computation requires 42,000 multiplications.

Now we perform the computation as (A * B) * C.

A * B requires 60 multiplications for each entry. Since the output is 20x50, there are 1000 entries and hence 60,000 multiplications. We now multiply this result by C. This matrix product requires 50 multiplications for each entry. Since the output is 20x10, there are 200 entries and hence 10,000 multiplications are required. So the overall computation requires 70,000 multiplication.

Example: Do on the board the example on the right. Notice that the answers are the same for either order of parenthesizing but that the numbers of multiplications are different.

The matrix chain-product problem is to determine the parenthesization that minimizes the number of multiplications.

(Obvious) solution. Try every parenthesization and pick the best one.

The obvious solution, often called the brute force solution, has exponential complexity since the number of ways to parenthesize an associative expression with n terms is the nth so-called Catalan number, which is Ω(4n/n3/2).

We want to do better and will do so using dynamic programming. We need to define subproblems, subsubproblems, etc and then compute the answer backwards taking advantage of common subproblems, the solutions to which we store.

Defining Subproblems

Recall that the matrix product whose multiplications we are trying to minimize is

A0 * ... * An-1
We want a convenient notation to describe various subproblems and define Ni,j to be the minimum number of multiplications needed to compute the subexpression
Ai * ... * Aj
In particular N0,n-1 is the original problem.

Characterizing Optimal Solutions

What follows is the key point of dynamic programming.

The key observation about matrix chain-product is that we can characterize the optimal solution in terms of optimal solutions of the subproblems. (The book calls this the subproblem optimality condition.)

We note specifically that when computing Ai*...*Aj there must be a last matrix multiplication so that

Ai*...*Aj = (Ai*...*Ak)*(Ak+1*...*Aj)
and (the KEY point) to get the minimum number of multiplications for Ai*...*Aj we must have the minimum number of multiplications for Ai*...*Ak and for Ak+1*...Aj. This says that Ni,j is just the minimum over all possible k's of
Ni,k + Nk,j + the cost of the final matrix multiplication.

What is the cost of the final matrix multiplication?

Designing a Dynamic Programming Algorithm

We have just seen that

Ni,j = mini≤k<j {Ni,k + Nk+1,j + didk+1dj+1}

Let's call this the fundamental equation.

Although the fundamental equation does not look simple, let's not be discouraged and keep going. First of all if we apply this recursively we will get terms in which the subscripts of N are close together. We can stop the recursion when the subscripts are equal since Ni,i=0 because no multiplications are required since no matrix product is being computed.

Remember that the idea of dynamic programming is to run the recursion backwards and start with the smallest problems.

algorithm MatrixChain(d0,...,dn) (dynamic programming)
   Input:  Sequence d0,...,dn of positive integer
           corresponding to the dimensions of a chain of matrices
   Output: Ni,j, the minimum number of multiplications
           needed to compute Ai*...*Aj

   for i ← 0 to n-1 do           { the simple base case }
      Ni,i ← 0

   for b ← 1 to n-1 do           { keep doing larger gaps }
      for i ← 0 to n-1-b         { to keep j in bounds )
         j ← i+b
         { Calculate Ni,j from the fundamental equation }
         Ni,j ← +infinity
         for k ← i to j-1 do
            Ni,j ← min(Ni,j,Ni,k+Nk+1,j+didk+1dj+1)

This algorithm calculates the minimum number of parenthesis needed for any contiguous subset of the matrices. In particular N0,n-1 gives the minimum number of parenthesis needed for the entire matrix chain-product.

Example: Compute, on the board, the minimum number of multiplications needed for the example drawn previous. We have three matrices of shape 2x2, 2x2 and 2x1 respectively so n=3 and the d vector is 2,2,2,1.

Homework: R-5.9

Analyzing the Matrix Chain-product Algorithm

This is easy now that we know the algorithm.

The algorithm clearly takes O(n3) time since it has a triply nested loop and each loop has O(n) iterations.

But the algorithm gives the number of parentheses not where they should be placed. We can fix this problem by storing in Ni,j not just the number of parentheses but the index k that gave us the solution.

Theorem: We can compute a parenthesization that minimizes the number of multiplications in a matrix chain-product in O(n3) time, where n is the number of matrices.

The diagram on the right illustrates the way dynamic program generates the elements of N.

Remark: The space complexity is Θ(N2) since the only storage used is the array N (plus a few scalars).

5.3.2 The General Technique

There are three properties needed for successful application of dynamic programming.

  1. Splitting into Subproblems: Must be able to split the original into subproblems in a recursive manner (i.e., so that subproblems can also be split into sub-subproblems, etc.). When the problem is split enough times, easy subproblems must result.
  2. Subproblem Optimality: An optimal solution to the problem must result from optimal solutions to the subproblems via some simple combining operations.
  3. Subproblem Overlap: Many subproblems themselves contain common sub-subproblems. Thus once we have solved the sub-subproblem, we can reuse this solution for many subproblems.

(The following section 5.3.A is not from the book and was covered in gottlieb's lecture #25)

5.3.A A java program to show the importance of memoization

We actually made three advances in solving this problem.

  1. We found a fairly simple recursive formulation based on singling out the last matrix multiplication.
  2. This recursion has a number of common subproblems (i.e., there is a great deal of subproblem overlap). If we remember the answer the first time we solve the subproblem, we can look it up instead of recomputing it when we encounter this subproblem again. This is called memoization.
  3. We used a dynamic programming solution that starts from small answers to generate large one. This eliminates the recursion and makes it easy to see that the solution has cubic complexity.

How important are each of these? I wrote a program that shows that the gain from 1 to 2 is much more than the gain from 2 to 3.

import java.io.StreamTokenizer;
import java.io.InputStreamReader;

public class MCP
    static int[] d;             // dimension of matrices
    static int[][] N;           // Hold values in dynamic programming
    static boolean[][] known;   // Have we memoized this value
    static int[][] NN;          // Hold memoized values

    public static void main (String[] args) throws java.io.IOException
        StreamTokenizer st =
            new StreamTokenizer (new InputStreamReader(System.in));

        // Read in data
        System.out.println("Enter number of matrices and dimensions:");
        int n = (int)st.nval;
        d = new int[n+1];
        for (int i=0; i<=n; i++) {
            d[i] = (int)st.nval;

        // Calculate N(0,n-1) recursively
        System.out.println ("\nCalculate recursively? [yn]");
        if (st.sval.equals("y")) {
            System.out.println ("Calculating recursively ...");
            System.out.println ("... done.  The answer is " + NRec(0,n-1));

        // Calculate N(0,n-1) recursively, but with memoization
        System.out.println ("\nCalculate recursively with memoization? [yn]");
        if (st.sval.equals("y")) {
            NN = new int[n][n];
            known = new boolean[n][n];
            for (int i=0;  i< n; i++)
                for (int j=0; j < n; j++)
                    known[i][j] = false;
            System.out.println ();
            System.out.println ("Calculating N(0,n-1) with memoization ...");
            System.out.println ("... done.  The answer is " + NRecMemo(0,n-1));

        // Calculate N(0,n-1) with dynamic programming
        System.out.println ("\nCalculate with dynamic programming? [yn]");
        if (st.sval.equals("y")) {
            N = new int[n][n];
            System.out.println ();
            System.out.println ("Calculating with dynamic programming ...");
            System.out.println ("... done.  The answer is " + NDyn(n));

    private static int NRec(int i, int j)
        if (i==j)
            return 0;
        else {
            int ans = Integer.MAX_VALUE;
            int val;
            for (int k=i; k < j; k++) {
                val = NRec(i,k) + NRec(k+1,j) + d[i]*d[k+1]*d[j+1];
                if (val < ans)
                    ans = val;
            return ans;

    private static int NRecMemo(int i, int j)
        if (known[i][j])
            return NN[i][j];

        int ans;
        if (i==j)
            ans = 0;
        else {
            ans = Integer.MAX_VALUE;
            int val;
            for (int k=i; k < j; k++) {
                val = NRecMemo(i,k) + NRecMemo(k+1,j) + d[i]*d[k+1]*d[j+1];
                if (val < ans)
                    ans = val;
        NN[i][j] = ans;
        known[i][j] = true;
        return ans;

    private static int NDyn(int n)
        for (int i=0; i < n; i++)

        for (int b=1; b < n; b++)
            for (int i=0; i < n-b; i++) {
                int j = i+b;
                int nij = Integer.MAX_VALUE;
                int val;
                for (int k=i; k < j; k++) {
                    val = N[i][k] + N[k+1][j] + d[i]*d[k+1]*d[j+1];
                    if (val < nij)
                        nij = val;
                N[i][j] = nij;
        return N[0][n-1];

Let's run the program and see how long it takes.

5.3.3 The 0-1 Knapsack Problem

This is the real knapsack problem, a well-known NP-complete problem. So clearly we will not get a polynomial time solution (but we will get close).

Consider a knapsack that can hold a maximum capacity W and a set S of n items with item i weighing wi and giving benefit bi. All the w's, b's and W are positive integers.

The problem is to maximize the benefit of the items carried subject to the constraint that the total weight cannot exceed W. It is called the 0-1 knapsack problem because, for each item, you leave it behind (0) or take all of it (1). You cannot choose to take for example half of the item as you could in the fractional knapsack problem discussed at the beginning of this chapter.

Homework: R-5.12

A First Attempt at Characterizing Subproblems

Let Si consist of items 1,2,...,i. So Sn is all of S. The idea is that subproblem k will be to find the optimal way to load the knapsack using only items in Sk, i.e., items 1,...,k. Then the overall answer is obtained by solving Sn.

It is indeed possible to split the problem this way and when split enough we get S1, which is trivial to optimize since there is only one element.

There is subproblem overlap as desired.

But it is not at all clear how to extend an optimal solution found for Sk to an optimal solution for Sk+1. Indeed finding the optimal solution for S1, then S2, then S3, corresponds to deciding on each element one at a time. This would be a greedy solution and doesn't give a global optimal.

One could also define Si,j, analogous to Ni,j above. However, we again don't have a way to extend an optimal solutions to subproblems into an optimal solution for a bigger subproblem.

A Better Subproblem Characterization

The problem is that by defining the subproblem simply in terms of k, we do not get enough information to construct a solution that is helpful to obtaining a global maximum.

Instead we define subproblems in a more complicated manner. This is a stroke of inspiration and is not at all obvious.

Let B[k,w] be the maximum benefit obtainable from items in Sk having total weight exactly w. Our overall goal is to find B[n,W], the original 0-1 knapsack problem.

This does split the problems into subproblems and when we get down to B[0,w] we have an trivial problem with solution 0 since we have no items to select from. (Also B[k,0] is zero since we are not permitted to choose any items since our weight limit is 0, but we don't use this fact.)

The key that makes this solution work is the observation that

        /      B[k-1,w]                   if wk>w
B[k,w] =
        \ max {B[k-1,w], B[k-1,w-wk]+bk}   otherwise

This looks formidable, but is not. We need to choose a subset of the first k items with weight exactly w. If the kth item weighs more than w, it cannot be used so the best we can do with the first k items is the same as the best we can do with the first k-1 items. That was the easier case.

If the kth item is not heavier than w, it can be used, but need not be (that gives the two possibilities in the max). If we choose not to use the kth item, we get the previous case (the first term in the max). If we use the kth item and get total weight w, the items chosen from the first k-1 must weight w-wk. Also in this case we gain the benefit of the kth item.

This looks pretty good. We can loop on k and inside that on w. B[k,*] just depends on B[k-1,*] so it works fine.

Algorithm 01knapsack(S,W)
   Input:  A set S of n items each with weight wi and benefit bi.
           A max weight W.  All are positive; all weights integers.
   Output: B[k,w] the maximum benefit obtainable from a subset of
           the first k items in S having total weight w.

   for w ← 0 to W          { base case }
      B[0,w] ← 0

   for k ← 1 to n
      for w ← 0 to W
         if wk > w then
            B[k,w] ← B[k-1,w]
            B[k,w] =  max (B[k-1,w], B[k-1,w-wk]+bk)

Analyzing the 0-1 Knapsack Dynamic Programming Algorithm

This is easy, each loop iteration is constant time so the complexity is the number of iterations. The base case loop has Θ(W) iterations and the nested loops have Θ(nW).

Theorem: Given a positive integer W and a set S of n items each with positive benefit and positive integer weight, we can find the highest benefit subset of S with total weight at most W in time Θ(nW).


  1. The algorithm given only computes the maximum not the subset. It is fairly easy to adapt the algorithm to give the items as well (the max computation tells you weather or not to use item k).
  2. My simple program needs space Θ(nW). The book gives a more subtle argument and an algorithm that only needs space Θ(W).

End of Remarks

Pseudo-Polynomial-Time Algorithms

At first glance the theorem above seems to say that there is a polynomial time solution to the 0-1 knapsack problem. The time complexity is Θ(nW), the product of two numbers characteristic of the input. But this is wrong!

The definition of polynomial time is that the algorithm takes time that is polynomial in the size of the input. What is the size of the input?

  1. We have n items each with two values a benefit b and a weight w. We should consider the size of these items to be Θ(n) in total. That is we should consider each item to be of constant size since otherwise all the operations we have used are not constant time. So that means the factor of n in the time complexity is ok.
  2. But what about the factor of W? Well W is an input, doesn't that settle it? No! The input value W requires only log(W) bits, so the size of the input W is Θ(log(W)).
  3. Hence the input size is &Theta(n+log(W)) and the complexity is Θ(nW). Is this a polynomial relationship?
  4. Consider a big W say W=2n. Then the input size is Θ(n+n)=Θ(n) and the complexity is Θ(n2n). So the complexity is exponential in this case.
  5. Indeed for large enough W, nW exceeds the time for the brute-force solution.

Remark: It is common to refer to algorithms like ours as requiring pseudo-polynomial time. That is it is polynomial the value of a number in the input, but not on its input size (the size of its binary representation).