14: Algorithms
Lecture Amortized Analysis [Sp'10]
The goode workes that men don whil they ben in good lif al amortised by synne folwyng. -- Geoffrey Chaucer, "The Persones [Parson's] Tale" (c.1400) I will gladly pay you Tuesday for a hamburger today. -- J. Wellington Wimpy, "Thimble Theatre" (1931) I want my two dollars! -- Johnny Gasparini [Demian Slade], "Better Off Dead" (1985)
14
14.1
Amortized Analysis
Incrementing a Binary Counter
It is a straightforward exercise in induction, which often appears on Homework 0, to prove that any non-negative integer n can be represented as the sum of distinct powers of 2. Although some students correctly use induction on the number of bits--pulling off either the least significant bit or the most significant bit in the binary representation and letting the Recursion Fairy convert the remainder--the most commonly submitted proof uses induction on the value of the integer, as follows: Proof: The base case n = 0 is trivial. For any n > 0, the inductive hypothesis implies that there is set of distinct powers of 2 whose sum is n - 1. If we add 20 to this set, we obtain a multiset of powers of two whose sum is n, which might contain two copies of 20 . Then as long as there are two copies of any 2i in the multiset, we remove them both and insert 2i+1 in their place. The sum of the elements of the multiset is unchanged by this replacement, because 2i+1 = 2i + 2i . Each iteration decreases the size of the multiset by 1, so the replacement process must eventually terminate. When it does terminate, we have a set of distinct powers of 2 whose sum is n. This proof is describing an algorithm to increment a binary counter from n - 1 to n. Here's a more formal (and shorter!) description of the algorithm to add 1 to a binary counter. The input B is an (infinite) array of bits, where B[i] = 1 if and only if 2i appears in the sum.
INCREMENT(B[0 .. ]): i0 while B[i] = 1 B[i] 0 i i+1 B[i] 1
We've already argued that INCREMENT must terminate, but how quickly? Obviously, the running time depends on the array of bits passed as input. If the first k bits are all 1s, then INCREMENT takes (k) time. The binary representation of any positive integer n is exactly lg n + 1 bits long. Thus, if B represents an integer between 0 and n, INCREMENT takes (log n) time in the worst case.
14.2
Counting from 0 to n
Now suppose we want to use INCREMENT to count from 0 to n. If we only use the worst-case running time for each INCREMENT, we get an upper bound of O(n log n) on the total running time. Although this bound is correct, we can do better. There are several general methods for proving that the total running time is actually only O(n). Many of these methods are logically equivalent, but different formulations are more natural for different problems.
c Copyright 2010 Jeff Erickson. Released under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License (http://creativecommons.org/licenses/by-nc-sa/3.0/). Free distribution is strongly encouraged; commercial distribution is expressly forbidden. See http://www.cs.uiuc.edu/~jeffe/teaching/algorithms/ for the most recent revision.
1
Algorithms 14.2.1 The Summation (Aggregate) Method
Lecture 14: Amortized Analysis [Sp'10]
The easiest way to get a tighter bound is to observe that we don't need to flip (log n) bits every time INCREMENT is called. The least significant bit B[0] does flip every time, but B[1] only flips every other time, B[2] flips every 4th time, and in general, B[i] flips every 2i th time. If we start with an array full of 0's, a sequence of n INCREMENTs flips each bit B[i] exactly n/2i times. Thus, the total number of bit-flips for the entire sequence is
lg n
n 2i
<
i=0
n 2i
= 2n.
i=0
Thus, on average, each call to INCREMENT flips only two bits, and so runs in constant time. This `on average' is quite different from the averaging we consider with randomized algorithms. There is no probability involved; we are averaging over a sequence of operations, not the possible running times of a single operation. This averaging idea is called amortization--the amortized time for each INCREMENT is O(1). Amortization is a sleazy clever trick used by accountants to average large one-time costs over long periods of time; the most common example is calculating uniform payments for a loan, even though the borrower is paying interest on less and less capital over time. For this reason, `amortized cost' is a common synonym for amortized running time. Most textbooks call this technique for bounding amortized costs the aggregate method, or aggregate analysis, but this is just a fancy name for adding up the costs of the individual operations and dividing by the number of operations. The Summation Method. Let T (n) be the worst-case running time for a sequence of n operations. The amortized time for each operation is T (n)/n. 14.2.2 The Taxation (Accounting) Method
A second method we can use to derive amortized bounds is called either the accounting method or the taxation method. Suppose it costs us a dollar to toggle a bit, so we can measure the running time of our algorithm in dollars. Time is money! Instead of paying for each bit flip when it happens, the Increment Revenue Service charges a twodollar increment tax whenever we want to set a bit from zero to one. One of those dollars is spent changing the bit from zero to one; the other is stored away as credit until we need to reset the same bit to zero. The key point here is that we always have enough credit saved up to pay for the next INCREMENT. The amortized cost of an INCREMENT is the total tax it incurs, which is exactly 2 dollars, since each INCREMENT changes just one bit from 0 to 1. It is often useful to distribute the tax income to specific pieces of the data structure. For example, for each INCREMENT, we could store one of the two dollars on the single bit that is set for 0 to 1, so that that bit can pay to reset itself back to zero later on. Taxation Method 1. Certain steps in the algorithm charge you taxes, so that the total cost incurred by the algorithm is never more than the total tax you pay. The amortized cost of an operation is the overall tax charged to you during that operation. A different way to schedule the taxes is for every bit to charge us a tax at every operation, regardless of whether the bit changes of not. Specifically, each bit B[i] charges a tax of 1/2i dollars for each INCREMENT. The total tax we are charged during each INCREMENT is i0 2-i = 2 dollars. Every time a bit B[i] actually needs to be flipped, it has collected exactly $1, which is just enough for us to pay for the flip. 2
Algorithms
Lecture 14: Amortized Analysis [Sp'10]
Taxation Method 2. Certain portions of the data structure charge you taxes at each operation, so that the total cost of maintaining the data structure is never more than the total taxes you pay. The amortized cost of an operation is the overall tax you pay during that operation. In both of the taxation methods, our task as algorithm analysts is to come up with an appropriate `tax schedule'. Different `schedules' can result in different amortized time bounds. The tightest bounds are obtained from tax schedules that just barely stay in the black. 14.2.3 The Charging Method
Another common method of amortized analysis involves charging the cost of some steps to some other, earlier steps. The method is similar to taxation, except that we focus on where each unit of tax is (or will be) spent, rather than where is it collected. By charging the cost of some operations to earlier operations, we are overestimating the total cost of any sequence of operations, since we pay for some charges from future operations that may never actually occur. For example, in our binary counter, suppose we charge the cost of clearing a bit (changing its value from 1 to 0) to the previous operation that sets that bit (changing its value from 0 to 1). If we flip k bits during an INCREMENT, we charge k - 1 of those bit-flips to earlier bit-flips. Conversely, the single operation that sets a bit receives at most one unit of charge from the next time that bit is cleared. So instead of paying for k bit-flips, we pay for at most two: one for actually setting a bit, plus at most one charge from the future for clearing that same bit. Thus, the total amortized cost of the INCREMENT is at most two bit-flips. The Charging Method. Charge the cost of some steps of the algorithm to earlier steps, or to steps in some earlier operation. The amortized cost of the algorithm is its actual running time, minus its total charges to past operations, plus its total charge from future operations. 14.2.4 The Potential Method
The most powerful method (and the hardest to use) builds on a physics metaphor of `potential energy'. Instead of associating costs or taxes with particular operations or pieces of the data structure, we represent prepaid work as potential that can be spent on later operations. The potential is a function of the entire data structure. Let Di denote our data structure after i operations, and let i denote its potential. Let ci denote the actual cost of the ith operation (which changes Di-1 into Di ). Then the amortized cost of the ith operation, denoted ai , is defined to be the actual cost plus the increase in potential: ai = ci + i - i-1 So the total amortized cost of n operations is the actual total cost plus the total increase in potential:
n n n
ai =
i=1 i=1
ci + i - i-1 =
i=1
c i + n - 0 .
A potential function is valid if i - 0 0 for all i. If the potential function is valid, then the total actual cost of any sequence of operations is always less than the total amortized cost:
n n n
ci =
i=1 i=1
ai - n
i=1
ai .
3
Algorithms
Lecture 14: Amortized Analysis [Sp'10]
For our binary counter example, we can define the potential i after the ith INCREMENT to be the number of bits with value 1. Initially, all bits are equal to zero, so 0 = 0, and clearly i > 0 for all i > 0, so this is a valid potential function. We can describe both the actual cost of an INCREMENT and the change in potential in terms of the number of bits set to 1 and reset to 0. ci = #bits changed from 0 to 1 + #bits changed from 1 to 0 i - i-1 = #bits changed from 0 to 1 - #bits changed from 1 to 0 Thus, the amortized cost of the ith INCREMENT is ai = ci + i - i-1 = 2 #bits changed from 0 to 1 Since INCREMENT changes only one bit from 0 to 1, the amortized cost INCREMENT is 2. The Potential Method. Define a potential function for the data structure that is initially equal to zero and is always nonnegative. The amortized cost of an operation is its actual cost plus the change in potential. For this particular example, the potential is precisely the total unspent taxes paid using the taxation method, so not too surprisingly, we obtain precisely the same amortized cost. In general, however, there may be no way of interpreting the change in potential as `taxes'. Taxation and charging are useful when there is a convenient way to distribute costs to specific steps in the algorithm or components of the data structure; potential arguments allow us to argue more globally when a local distribution is difficult or impossible. Different potential functions can lead to different amortized time bounds. The trick to using the potential method is to come up with the best possible potential function. A good potential function goes up a little during any cheap/fast operation, and goes down a lot during any expensive/slow operation. Unfortunately, there is no general technique for doing this other than playing around with the data structure and trying lots of different possibilities.
14.3
Incrementing and Decrementing
Now suppose we wanted a binary counter that we can both increment and decrement efficiently. A standard binary counter won't work, even in an amortized sense; if we alternate between 2k and 2k - 1, every operation costs (k) time. A nice alternative is represent a number as a pair of bit strings (P, N ), where for any bit position i, at most one of the bits P[i] and N [i] is equal to 1. The actual value of the counter is P - N . Here are algorithms to increment and decrement our double binary counter.
INCREMENT(P, N ): i0 while P[i] = 1 P[i] 0 i i+1 if N [i] = 1 N [i] 0 else P[i] 1 DECREMENT(P, N ): i0 while N [i] = 1 N [i] 0 i i+1 if P[i] = 1 P[i] 0 else N [i] 1
Here's an example of these algorithms in action. Notice that any number other than zero can be represented in multiple (in fact, infinitely many) ways. 4
Algorithms
Lecture 14: Amortized Analysis [Sp'10]
P = 10001 P = 10000 P = 10000 P = 10000 P = 10011 P = 10010 P = 10001 ++ -- -- ++ ++ ++ N = 01100 - N = 01100 - N = 01100 - N = 01000 - N = 01001 - N = 01010 - N = 01010 P-N =7 P-N =6 P-N =7 P-N =8 P-N =7 P-N =6 P-N =5
Incrementing and decrementing a double-binary counter.
Now suppose we start from (0, 0) and apply a sequence of n INCREMENTs and DECREMENTs. In the worst case, operation takes (log n) time, but what is the amortized cost? We can't use the aggregate method here, since we don't know what the sequence operations of looks like. What about the taxation method? It's not hard to prove (by induction, of course) that after either P[i] or N [i] is set to 1, there must be at least 2i operations, either INCREMENTs or DECREMENTs, before that bit is reset to 0. So if each bit P[i] and N [i] pays a tax of 2-i at each operation, we will always have enough money to pay for the next operation. Thus, the amortized cost of each operation is at most -i = 4. i0 2 2 We can get even better bounds using the potential method. Define the potential i to be the number of 1-bits in both P and N after i operations. Just as before, we have ci = #bits changed from 0 to 1 + #bits changed from 1 to 0 i - i-1 = #bits changed from 0 to 1 - #bits changed from 1 to 0 = ai = 2 #bits changed from 0 to 1 Since each operation changes at most one bit to 1, the ith operation has amortized cost ai 2.
14.4
Gray Codes
An attractive alternate solution to the increment/decrement problem was independently suggested by several students. Gray codes (named after Frank Gray, who discovered them in the 1950s) are methods for representing numbers as bit strings so that successive numbers differ by only one bit. For example, here is the four-bit binary reflected Gray code for the integers 0 through 15: 0000, 0001, 0011, 0010, 0110, 0111, 0101, 0100, 1100, 1101, 1111, 1110, 1010, 1011, 1001, 1000 The general rule for incrementing a binary reflected Gray code is to invert the bit that would be set from 0 to 1 by a normal binary counter. In terms of bit-flips, this is the perfect solution; each increment of decrement by definition changes only one bit. Unfortunately, the nave algorithm to find the single bit to flip still requires (log n) time in the worst case. Thus, so the total cost of maintaining a Gray code, using the obvious algorithm, is the same as that of maintaining a normal binary counter. Fortunately, this is only true of the nave algorithm. The following algorithm, discovered by Gideon Ehrlich1 in 1973, maintains a Gray code counter in constant worst-case time per increment! The algorithm uses a separate `focus' array F [0 .. n] in addition to a Gray-code bit array G[0 .. n - 1].
EHRLICHGRAYINCREMENT(n): j F [0] F [0] 0 if j = n G[n - 1] 1 - G[n - 1] else G[ j] = 1 - G[ j] F [ j] F [ j + 1] F [ j + 1] j + 1
EHRLICHGRAYINIT(n): for i 0 to n - 1 G[i] 0 for i 0 to n F [i] i
Gideon Ehrlich. Loopless algorithms for generating permutations, combinations, and other combinatorial configurations. J. Assoc. Comput. Mach. 20:500513, 1973.
1
5
Algorithms
Lecture 14: Amortized Analysis [Sp'10]
The EHRLICHGRAYINCREMENT algorithm obviously runs in O(1) time, even in the worst case. Here's the algorithm in action with n = 4. The first line is the Gray bit-vector G, and the second line shows the focus vector F , both in reverse order: G : 0000, 0001, 0011, 0010, 0110, 0111, 0101, 0100, 1100, 1101, 1111, 1110, 1010, 1011, 1001, 1000 F : 3210, 3211, 3220, 3212, 3310, 3311, 3230, 3213, 4210, 4211, 4220, 4212, 3410, 3411, 3240, 3214 Voodoo! I won't explain in detail how Ehrlich's algorithm works, except to point out the following invariant. Let B[i] denote the ith bit in the standard binary representation of the current number. If B[ j ] = 0 and B[ j - 1] = 1, then F [ j ] is the smallest integer k > j such that B[k] = 1; otherwise, F [ j] = j . Got that? But wait -- this algorithm only handles increments; what if we also want to decrement? Sorry, I don't have a clue. Extra credit, anyone?
14.5
Generalities and Warnings
Although computer scientists usually apply amortized analysis to understand the efficiency of maintaining and querying data structures, you should remember that amortization can be applied to any sequence of numbers. Banks have been using amortization to calculate fixed payments for interest-bearing loans for centuries. The IRS allows taxpayers to amortize business expenses or gambling losses across several years for purposes of computing income taxes. Some cell phone contracts let you to apply amortization to calling time, by rolling unused minutes from one month into the next month. It's also important to remember that amortized time bounds are not unique. For a data structure that supports multiple operations, different amortization schemes can assign different costs to exactly the same algorithms. For example, consider a generic data structure that can be modified by three algorithms: FOLD, SPINDLE, and MUTILATE. One amortization scheme might imply that FOLD and SPINDLE each run in O(log n) amortized time, while MUTILATE runs in O(n) amortized time. Another scheme might imply that FOLD runs in O( n) amortized time, while SPINDLE and MUTILATE each run in O(1) amortized time. These two results are not necessarily inconsistent! Moreover, there is no general reason to prefer one of these sets of amortized time bounds over the other; our preference may depend on the context in which the data structure is used.
Exercises
1. Suppose we are maintaining a data structure under a series of operations. Let f (n) denote the actual running time of the nth operation. For each of the following functions f , determine the resulting amortized cost of a single operation. (For practice, try all of the methods described in this note.) (a) f (n) is the largest integer i such that 2i divides n. (b) f (n) is the largest power of 2 that divides n. (c) f (n) = n if n is a power of 2, and f (n) = 1 otherwise. (d) f (n) = n2 if n is a power of 2, and f (n) = 1 otherwise. 2. A multistack consists of an infinite series of stacks S0 , S1 , S2 , . . . , where the ith stack Si can hold up to 3i elements. The user always pushes and pops elements from the smallest stack S0 . However, before any element can be pushed onto any full stack Si , we first pop all the elements off Si and 6
Algorithms
Lecture 14: Amortized Analysis [Sp'10]
push them onto stack Si+1 to make room. (Thus, if Si+1 is already full, we first recursively move all its members to Si+2 .) Similarly, before any element can be popped from any empty stack Si , we first pop 3i elements from Si+1 and push them onto Si to make room. (Thus, if Si+1 is already empty, we first recursively fill it by popping elements from Si+2 .) Moving a single element from one stack to another takes O(1) time. Here is pseudocode for the multistack operations MSPUSH and MSPOP. The internal stacks are managed with the subroutines PUSH and POP.
MPUSH(x) : i0 while Si is full i i+1 while i > 0 i i-1 for j 1 to 3i PUSH(Si+1 , POP(Si )) PUSH(S0 , x)
9
MPOP(x) : i0 while Si is empty i i+1 while i > 0 i i-1 for j 1 to 3i PUSH(Si , POP(Si+1 )) return POP(S0 )
3
Making room in a multistack, just before pushing on a new element.
(a) In the worst case, how long does it take to push one more element onto a multistack containing n elements? (b) Prove that if the user never pops anything from the multistack, the amortized cost of a push operation is O(log n), where n is the maximum number of elements in the multistack during its lifetime. (c) Prove that in any intermixed sequence of pushes and pops, each push or pop operation takes O(log n) amortized time, where n is the maximum number of elements in the multistack during its lifetime. 3. Remember the difference between stacks and queues? Good. (a) Describe how to implement a queue using two stacks and O(1) additional memory, so that the amortized time for any enqueue or dequeue operation is O(1). The only access you have to the stacks is through the standard subroutines PUSH and POP. (b) A quack is a data structure combining properties of both stacks and queues. It can be viewed as a list of elements written left to right such that three operations are possible: 7
Algorithms
Lecture 14: Amortized Analysis [Sp'10] Push: add a new item to the left end of the list; Pop: remove the item on the left end of the list; Pull: remove the item on the right end of the list.
Implement a quack using three stacks and O(1) additional memory, so that the amortized time for any push, pop, or pull operation is O(1). In particular, each element pushed onto the quack should be stored in exactly one of the three stacks. Again, you are only allowed to access the stacks through the standard functions PUSH and POP. 4. Suppose we can insert or delete an element into a hash table in O(1) time. In order to ensure that our hash table is always big enough, without wasting a lot of memory, we will use the following global rebuilding rules: After an insertion, if the table is more than 3/4 full, we allocate a new table twice as big as our current table, insert everything into the new table, and then free the old table. After a deletion, if the table is less than 1/4 full, we allocate a new table half as big as our current table, insert everything into the new table, and then free the old table. Show that for any sequence of insertions and deletions, the amortized time per operation is still O(1). [Hint: Do not even look at the potential-function argument in CLRS; there is a much easier solution!] 5. Chicago has many tall buildings, but only some of them have a clear view of Lake Michigan. Suppose we are given an array A[1 .. n] that stores the height of n buildings on a city block, indexed from west to east. Building i has a good view of Lake Michigan if and only if every building to the east of i is shorter than i. Here is an algorithm that computes which buildings have a good view of Lake Michigan. What is the running time of this algorithm?
GOODVIEW(A[1 .. n]): initialize a stack S for i 1 to n while (S not empty and A[i] > A[TOP(S)]) POP(S) PUSH(S, i) return S
6. Design and analyze a simple data structure that maintains a list of integers and supports the following operations. CREATE( ) creates and returns a new list PUSH(L, x) appends x to the end of L POP(L) deletes the last entry of L and returns it LOOKUP(L, k) returns the kth entry of L Your solution may use these primitive data structures: arrays, balanced binary search trees, heaps, queues, single or doubly linked lists, and stacks. If your algorithm uses anything fancier, you must give an explicit implementation. Your data structure must support all operations in amortized 8
Algorithms
Lecture 14: Amortized Analysis [Sp'10]
constant time. In addition, your data structure must support each LOOKUP in worst-case O(1) time. At all times, the size of your data structure must be linear in the number of objects it stores. 7. Suppose instead of powers of two, we represent integers as the sum of Fibonacci numbers. In other words, instead of an array of bits, we keep an array of fits, where the ith least significant fit indicates whether the sum includes the ith Fibonacci number Fi . For example, the fitstring 101110 F represents the number F6 + F4 + F3 + F2 = 8 + 3 + 2 + 1 = 14. Describe algorithms to increment and decrement a single fitstring in constant amortized time. [Hint: Most numbers can be represented by more than one fitstring!] 8. A doubly lazy binary counter represents any number as a weighted sum of powers of two, where each weight is one of four values: -1, 0, 1, or 2. (For succinctness, I'll write 1 instead of -1.) Every integer--positive, negative, or zero--has an infinite number of doubly lazy binary representations. For example, the number 13 can be represented as 1101 (the standard binary representation), or 2101 (because 2 23 - 22 + 20 = 13) or 10111 (because 24 - 22 + 21 - 20 = 13) or 11200010111 (because -210 + 29 + 2 28 + 24 - 22 + 21 - 20 = 13). To increment a doubly lazy binary counter, we add 1 to the least significant bit, then carry the rightmost 2 (if any). To decrement, we subtract 1 from the lest significant bit, and then borrow the rightmost 1 (if any). LAZYINCREMENT(B[0 .. n]): B[0] B[0] + 1 for i 1 to n - 1 if B[i] = 2 B[i] 0 B[i + 1] B[i + 1] + 1 return LAZYDECREMENT(B[0 .. n]): B[0] B[0] - 1 for i 1 to n - 1 if B[i] = -1 B[i] 1 B[i + 1] B[i + 1] - 1 return
For example, here is a doubly lazy binary count from zero up to twenty and then back down to zero. The bits are written with the least significant bit B[0] on the right, omitting all leading 0's
0 - 1 - 10 - 11 - 20 - 101 - 110 - 111 - 120 - 201 - 210 - 1011 - 1020 - 1101 - 1110 - 1111 - 1120 - 1201 - 1210 - 2011 - 2020 - 2011 - 2010 - 2001 - 2000 - 2011 - 2110 - 2101 - 1100 - 1111 - 1010 - 1001 - 1000 - 1011 - 1110 - 1101 - 100 - 111 - 10 - 1 - 0
-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++ ++
Prove that for any intermixed sequence of increments and decrements of a doubly lazy binary number, starting with 0, the amortized time for each operation is O(1). Do not assume, as in the example above, that all the increments come before all the decrements.
c Copyright 2010 Jeff Erickson. Released under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License (http://creativecommons.org/licenses/by-nc-sa/3.0/). Free distribution is strongly encouraged; commercial distribution is expressly forbidden. See http://www.cs.uiuc.edu/~jeffe/teaching/algorithms for the most recent revision.
9