Chapter 15: Sorting and Searching

Chapter Goals

• To compare the selection sort and merge sort algorithms
• To study the linear search and binary search algorithms
• To appreciate that algorithms for the same task can differ widely in performance
• To understand the big-Oh notation
• To learn how to estimate and compare the performance of algorithms
• To learn how to measure the running time of a program.

Selection Sort

• Selection sort is based upon finding the minimum value in a range of indices and placing that value at the front of the vector.
• Example: Suppose we start with the following vector.

  11

  9

  17

  5

  12

• The smallest element is 5, so we place it in the first location in the array. However we cannot lose the 11, so these two elements are swapped.

  5

  9

  17

  11

  12

• Now take the minimum of the remaining entries a[1] ... a[4]. That value is in the correct place, so no swapping is done.

  5

  9

  17

  11

  12

• Repeat the process over the unsorted region.

  5

  9

  11

  17

  12

• Continue until the unprocessed region has length 1. You are done.

  5

  9

  11

  12

  17

Selection Sort (selsort.cpp)

Profiling the Selection Sort Algorithm

• To measure the performance of a program, one could simply run it and measure how long it takes by using a stopwatch.
• Recall that default construction for the Time class initializes the object to the current time.

  Time now

• We can use Time objects to measure the performance of the sorting algorithm.

  Time before;
  selection_sort(v);
  Time after;
  
  cout << "Elapsed time = " << after.seconds_from(before)
       << " seconds\n";
  

• sorttime.cpp
  
• Here are results obtained on a Pentium III processor with a clock speed of 750 MHz running Linux.
• Actually numbers will look different, but the relationship between the numbers will be the same. Doubling the size of the data more than doubles the time needed to sort it.

  n	seconds
  5000	2
  10000	5
  20000	21
  40000	88

Analyzing the Performance of the Selection Sort Algorithm

• To analyze the selection sort algorithm, we will count how often an array element is visited.
• For a vector with n elements, we must visit all n of them to find the smallest one. Then visit two elements to swap them.
• In the next step, we visit n - 1 elements to find the minimum, plus two elements to swap.
• In the next step, we visit (n - 2) + 2 elements.
• The total elements visited is

  

  because

  

• The formula for the number of visits in the selection sort algorithm is (1/2) n² + (5/2) n - 3.
• For large values of n, the latter terms in the formula have no significant contribution, so we just ignore them.

  n	(1/2) n2	(5/2) n - 3
  1000	500 000	2 497
  2000	2 000 000	4 997

• When comparing the ratios of counts for different values of n, the coefficient 1/2 cancels out.
• We simply say "The number of visits is of order n²" and use big-Oh notation: The number of visits is O(n²).
• For selection sort, doubling the number of elements increases the time needed for sorting fourfold.
• If a 1000 element vector takes 11 seconds, a 100 000 element vector will require over 3 hours!

Merge Sort

• Suppose you have a vector of 10 integers, with the first and second half already sorted:

  

• It is an easy matter to merge the two sorted halves by taking a new element form either the first or second subvector and choosing the smaller of the elements each time:

  

• The computer keeps dividing the vector into smaller and smaller subvectors, sorting each half and merging them back together.

  void merge_sort(vector<int>& a, int from, int to)
  {  if (from == to) return;
     int mid = (from + to) / 2;
     /* sort the first and second half */
     merge_sort(a, from, mid);
     merge_sort(a, mid + 1, to);
     merge(a, from, mid, to);
  }

• The merge procedure then merges the two sorted halves of the vector.

Merge Sort (mergsort.cpp)

Analyzing the Merge Sort Algorithm

• mergetime.cpp
  
• Despite appearing to be a much more complicated algorithm, merge sort performs much better than selection sort.

• Note that the graph does not have a parabolic shape; instead it appears as if running time grows approximately linearly with the size of the array.

  n	seconds
  160000	1
  320000	2
  640000	4
  1280000	9

• To understand why the merge sort algorithm is such a tremendous improvement, estimate the number of array element visits.
• First we consider the merge process. For a vector of n elements,
  • We must visit the first two elements of the halves and decide which is smaller (2 visits)
  • We must copy the smaller element to it's place in the sorted vector (1 visit)
  • We must copy the sort vector back into the original vector (2 visits)
• So the merge process takes 5n visits.

• To analyze the full algorithm, let T(n) denote the number of visits required to sort a range of n elements through the merge sort process.
• For each of calculation we will assume n = 2^m.
• Because sorting each half will take T(n/2) visits, we have:

  

• A similar analysis yields:

  

• Putting these two results together gives us

  

• Repeating the process gives us:

  

• This analysis generalizes to the formula

  

• Since we assumed n = 2^m we have

  

• To establish growth order, we drop the lower order term and the constant factor.
• The change of base formula for logarithms allows us to drop the base of the logarithm.

  

• Hence merge sort is an O(n log n) algorithm.

• How does O(n²) compare to O(n log n)?
• Recall that it takes 100² = 10 000 times longer to sort 1 000 000 records than it takes to sort 10 000 records with the O(n²) algorithm.
• With the O(n log n) algorithm, the ratio is

  

• If it takes 4 seconds to sort 10 000 records for both sorts (merge sort is really faster, of course) then
  • it will take 10 minutes to merge sort 1 000 000 records
  • it will take 11 hours to selection sort 1 000 000 records

Searching

• If you want to find a number in a sequence of values that occur in arbitrary order, you must look through all elements until you have found a match or until you reach the end.
• This is called a linear or sequential search.
• There is nothing you can do to speed up the search, the algorithm is O(n).
  
• The procedure returns the index of the match, or a -1 if the value cannot be found.

Searching (lsearch.cpp)

Binary Search

• Now search for an item in a data sequence that has been previously sorted.
• Rather than a linear search, we will do a binary search.
• The search is called binary because the size of the search area is cut in half in each step.
• The cutting in half works only because the sequence of values has been sorted.

• Suppose we want to search the following data for 123.

  v[0]	v[1]	v[2]	v[3]	v[4]	v[5]	v[6]	v[7]
  14	43	76	100	115	290	400	511

• The last point in the first half of the data set, v[3], is 100. That means we should look in the second half for a match.

  v[4]	v[5]	v[6]	v[7]
  115	290	400	511

• The last value in the first half of this sequence is 290; hence, the value must be located in the sequence

  v[5]

  115

• Since there is only one element, the search easily shows that there is no match.

Binary Search (bsearch.cpp)

• Binary search is a O(log n) algorithm.
• Suppose n = 100, after each search the size of the search range is cut in half to 50, 25, 12, 6, 3, and 1.
  • After seven comparisons we are done.
  • log₂100 is about 6.64386, note that 2⁷ = 128.
• Since binary search is so much faster, is it worthwhile to always sort data before searching?
  • An O(n) search is less work that a O(n log n) sort, so it's not worth it if you only have to perform one search.
  • Sorting first might be worth it if you perform a number of searches of the same vector.