CSE332: Data Abstractions Lecture 3: Asymptotic Analysis Dan Grossman Spring 2012 Gauging performance Uh, why not just run the program and time it Too much variability, not reliable or portable: Hardware: processor(s), memory, etc. OS, Java version, libraries, drivers Other programs running Implementation dependent Choice of input Testing (inexhaustive) may miss worst-case input Timing does not explain relative timing among inputs (what happens when n doubles in size) Often want to evaluate an algorithm, not an implementation Even before creating the implementation (coding it up) Spring 2012 CSE332: Data Abstractions

2 Comparing algorithms When is one algorithm (not implementation) better than another? Various possible answers (clarity, security, ) But a big one is performance: for sufficiently large inputs, runs in less time (our focus) or less space Large inputs because probably any algorithm is plenty good for small inputs (if n is 10, probably anything is fast) Answer will be independent of CPU speed, programming language, coding tricks, etc. Answer is general and rigorous, complementary to coding it up and timing it on some test cases Spring 2012 CSE332: Data Abstractions 3

Analyzing code (worst case) Basic operations take some amount of constant time Arithmetic (fixed-width) Assignment Access one Java field or array index Etc. (This is an approximation of reality: a very useful lie.) Consecutive statements Sum of times Conditionals Time of test plus slower branch Loops Sum of iterations Calls Time of calls body Recursion Solve recurrence equation Spring 2012 CSE332: Data Abstractions

4 Example 2 3 5 16 37 50 73 75 126 Find an integer in a sorted array // requires array is sorted

// returns whether k is in array boolean find(int[]arr, int k){ ??? } Spring 2012 CSE332: Data Abstractions 5 Linear search 2 3 5 16 37

50 73 75 126 Find an integer in a sorted array // requires array is sorted // returns whether k is in array boolean find(int[]arr, int k){ for(int i=0; i < arr.length; ++i) if(arr[i] == k) return true; Best case: 6ish steps = O(1) return false; Worst case: 6ish*(arr.length) } = O(arr.length) Spring 2012 CSE332: Data Abstractions

6 Binary search 2 3 5 16 37 50 73 75 126 Find an integer in a sorted array Can also be done non-recursively but doesnt matter here // requires array is sorted

// returns whether k is in array boolean find(int[]arr, int k){ return help(arr,k,0,arr.length); } boolean help(int[]arr, int k, int lo, int hi) { int mid = (hi+lo)/2; // i.e., lo+(hi-lo)/2 if(lo==hi) return false; if(arr[mid]==k) return true; if(arr[mid]< k) return help(arr,k,mid+1,hi); else return help(arr,k,lo,mid); } Spring 2012 CSE332: Data Abstractions 7 Binary search Best case: 8ish steps = O(1)

Worst case: T(n) = 10ish + T(n/2) where n is hi-lo O(log n) where n is array.length Solve recurrence equation to know that // requires array is sorted // returns whether k is in array boolean find(int[]arr, int k){ return help(arr,k,0,arr.length); } boolean help(int[]arr, int k, int lo, int hi) { int mid = (hi+lo)/2; if(lo==hi) return false; if(arr[mid]==k) return true; if(arr[mid]< k) return help(arr,k,mid+1,hi); else return help(arr,k,lo,mid); } Spring 2012 CSE332: Data Abstractions

8 Solving Recurrence Relations 1. Determine the recurrence relation. What is the base case? T(n) = 10 + T(n/2) T(1) = 8 2. Expand the original relation to find an equivalent general expression in terms of the number of expansions. T(n) = 10 + 10 + T(n/4) = 10 + 10 + 10 + T(n/8) = = 10k + T(n/(2k)) 3. Find a closed-form expression by setting the number of expansions to a value which reduces the problem to a base case n/(2k) = 1 means n = 2k means k = log2 n

So T(n) = 10 log2 n + 8 (get to base case and do it) So T(n) is O(log n) Spring 2012 CSE332: Data Abstractions 9 Ignoring constant factors So binary search is O(log n) and linear is O(n) But which is faster Could depend on constant factors How many assignments, additions, etc. for each n And could depend on size of n But there exists some n0 such that for all n > n0 binary search wins Lets play with a couple plots to get some intuition Spring 2012 CSE332: Data Abstractions

10 Example Lets try to help linear search Run it on a computer 100x as fast (say 2010 model vs. 1990) Use a new compiler/language that is 3x as fast Be a clever programmer to eliminate half the work So doing each iteration is 600x as fast as in binary search Note: 600x still helpful for problems without logarithmic algorithms! Spring 2012 CSE332: Data Abstractions 11 Another example: sum array Two obviously linear algorithms: T(n) = O(1) + T(n-1) Iterative:

Recursive: Recurrence is k+k ++k for n times Spring 2012 int sum(int[] arr){ int ans = 0; for(int i=0; i

CSE332: Data Abstractions 12 What about a binary version? int sum(int[] arr){ return help(arr,0,arr.length); } int help(int[] arr, int lo, int hi) { if(lo==hi) return 0; if(lo==hi-1) return arr[lo]; int mid = (hi+lo)/2; return help(arr,lo,mid) + help(arr,mid,hi); } Recurrence is T(n) = O(1) + 2T(n/2) 1 + 2 + 4 + 8 + for log n times 2(log n) 1 which is proportional to n (definition of logarithm) Easier explanation: it adds each number once while doing little else Obvious: You cant do better than O(n) have to read whole array Spring 2012

CSE332: Data Abstractions 13 Parallelism teaser But suppose we could do two recursive calls at the same time Like having a friend do half the work for you! int sum(int[]arr){ return help(arr,0,arr.length); } int help(int[]arr, int lo, int hi) { if(lo==hi) return 0; if(lo==hi-1) return arr[lo]; int mid = (hi+lo)/2; return help(arr,lo,mid) + help(arr,mid,hi); } If you have as many friends of friends as needed the recurrence is now T(n) = O(1) + 1T(n/2) O(log n) : same recurrence as for find

Spring 2012 CSE332: Data Abstractions 14 Really common recurrences Should know how to solve recurrences but also recognize some really common ones: T(n) = O(1) + T(n-1) T(n) = O(1) + 2T(n/2) T(n) = O(1) + T(n/2) T(n) = O(1) + 2T(n-1) T(n) = O(n) + T(n-1) T(n) = O(n) + T(n/2) T(n) = O(n) + 2T(n/2) linear linear logarithmic exponential

quadratic (see previous lecture) linear O(n log n) Note big-Oh can also use more than one variable Example: can sum all elements of an n-by-m matrix in O(nm) Spring 2012 CSE332: Data Abstractions 15 Asymptotic notation About to show formal definition, which amounts to saying: 1. Eliminate low-order terms 2. Eliminate coefficients Examples: 4n + 5 0.5n log n + 2n + 7 n3 + 2n + 3n n log (10n2 )

Spring 2012 CSE332: Data Abstractions 16 Big-Oh relates functions We use O on a function f(n) (for example n2) to mean the set of functions with asymptotic behavior less than or equal to f(n) So (3n2+17) is in O(n2) 3n2+17 and n2 have the same asymptotic behavior Confusingly, we also say/write: (3n2+17) is O(n2) (3n2+17) = O(n2) But we would never say O(n2) = (3n2+17) Spring 2012 CSE332: Data Abstractions

17 Formally Big-Oh (Dr? Ms? Mr? ) Definition: g(n) is in O( f(n) ) if there exist constants c and n0 such that g(n) c f(n) for all n n0 To show g(n) is in O( f(n) ), pick a c large enough to cover the constant factors and n0 large enough to cover the lower-order terms Example: Let g(n) = 3n2+17 and f(n) = n2 c=5 and n0 =10 is more than good enough This is less than or equal to So 3n2+17 is also O(n5) and O(2n) etc. Spring 2012 CSE332: Data Abstractions 18 More examples, using formal definition

Let g(n) = 4n and f(n) = n2 A valid proof is to find valid c and n0 The cross-over point is n=4 So we can choose n0=4 and c=1 Many other possible choices, e.g., larger n0 and/or c Definition: g(n) is in O( f(n) ) if there exist constants c and n0 such that g(n) c f(n) for all n n0 Spring 2012 CSE332: Data Abstractions 19 More examples, using formal definition Let g(n) = n4 and f(n) = 2n A valid proof is to find valid c and n0 We can choose n0=20 and c=1

Definition: g(n) is in O( f(n) ) if there exist constants c and n0 such that g(n) c f(n) for all n n0 Spring 2012 CSE332: Data Abstractions 20 Whats with the c The constant multiplier c is what allows functions that differ only in their largest coefficient to have the same asymptotic complexity Example: g(n) = 7n+5 and f(n) = n For any choice of n0, need a c > 7 (or more) to show g(n) is in O( f(n) ) Definition: g(n) is in O( f(n) ) if there exist constants c and n0 such that g(n) c f(n) for all n n0

Spring 2012 CSE332: Data Abstractions 21 What you can drop Eliminate coefficients because we dont have units anyway 3n2 versus 5n2 doesnt mean anything when we have not specified the cost of constant-time operations (can re-scale) Eliminate low-order terms because they have vanishingly small impact as n grows Do NOT ignore constants that are not multipliers n3 is not O(n2) 3n is not O(2n) (This all follows from the formal definition) Spring 2012 CSE332: Data Abstractions

22 More Asymptotic Notation Upper bound: O( f(n) ) is the set of all functions asymptotically less than or equal to f(n) g(n) is in O( f(n) ) if there exist constants c and n0 such that g(n) c f(n) for all n n0 Lower bound: ( f(n) ) is the set of all functions asymptotically greater than or equal to f(n) g(n) is in ( f(n) ) if there exist constants c and n0 such that g(n) c f(n) for all n n0 Tight bound: ( f(n) ) is the set of all functions asymptotically equal to f(n) Intersection of O( f(n) ) and ( f(n) ) (use different c values) Spring 2012 CSE332: Data Abstractions 23 Correct terms, in theory

A common error is to say O( f(n) ) when you mean ( f(n) ) Since a linear algorithm is also O(n5), its tempting to say this algorithm is exactly O(n) That doesnt mean anything, say it is (n) That means that it is not, for example O(log n) Less common notation: little-oh: intersection of big-Oh and not big-Theta For all c, there exists an n0 such that Example: array sum is o(n2) but not o(n) little-omega: intersection of big-Omega and not big-Theta For all c, there exists an n0 such that Example: array sum is (log n) but not (n) Spring 2012 CSE332: Data Abstractions 24 What we are analyzing The most common thing to do is give an O or bound to the worst-case running time of an algorithm

Example: binary-search algorithm Common: (log n) running-time in the worst-case Less common: (1) in the best-case (item is in the middle) Less common: Algorithm is (log log n) in the worst-case (it is not really, really, really fast asymptotically) Less common (but very good to know): the find-in-sortedarray problem is (log n) in the worst-case No algorithm can do better (without parallelism) A problem cannot be O(f(n)) since you can always find a slower algorithm, but can mean there exists an algorithm Spring 2012 CSE332: Data Abstractions 25 Other things to analyze Space instead of time Remember we can often use space to gain time Average case Sometimes only if you assume something about the distribution of inputs

See CSE312 and STAT391 Sometimes uses randomization in the algorithm Will see an example with sorting; also see CSE312 Sometimes an amortized guarantee Will discuss in a later lecture Spring 2012 CSE332: Data Abstractions 26 Summary Analysis can be about: The problem or the algorithm (usually algorithm) Time or space (usually time) Or power or dollars or Best-, worst-, or average-case (usually worst) Upper-, lower-, or tight-bound (usually upper or tight) Spring 2012

CSE332: Data Abstractions 27 Usually asymptotic is valuable Asymptotic complexity focuses on behavior for large n and is independent of any computer / coding trick But you can abuse it to be misled about trade-offs Example: n1/10 vs. log n Asymptotically n1/10 grows more quickly But the cross-over point is around 5 * 1017 So if you have input size less than 258, prefer n1/10 For small n, an algorithm with worse asymptotic complexity might be faster Here the constant factors can matter, if you care about performance for small n Spring 2012 CSE332: Data Abstractions

28 Timing vs. Big-Oh Summary Big-oh is an essential part of computer sciences mathematical foundation Examine the algorithm itself, not the implementation Reason about (even prove) performance as a function of n Timing also has its place Compare implementations Focus on data sets you care about (versus worst case) Determine what the constant factors really are Will do some timing on the projects too Spring 2012 CSE332: Data Abstractions 29