Approach the problem using the "trial and error" algorithm

  • 2

    Well, normally we would refrain from using the naive trial and error algorithm for solving problems since it generally leads to bad time performance. However, there are situations where this naive algorithm may outperform other more sophisticated solutions, and LeetCode does have a few such problems (listed at the end of this post -- ironically they are all "hard" problems). So I figure it might be a good idea to bring it up and describe a general procedure for applying this algorithm.

    The basic idea for the trial and error algorithm is actually very simple and summarized below:

    Step 1: Construct a candidate solution.
    Step 2: Verify if it meets our requirements.
    Step 3: If it does, accept the solution; else discard it and repeat from Step 1.

    However, to make this algorithm work efficiently, the following two conditions need to be true:

    Condition 1: We have an efficient verification algorithm in Step 2;
    Condition 2: The search space formed by all candidate solutions is small or we have efficient ways to search this space if it is large.

    The first condition ensures that each verification operation can be done quickly while the second condition limits the total number of such operations that need to be done. The two combined will guarantee that we have an efficient trial and error algorithm (which also means if any of them cannot be satisfied, you should probably not even consider this algorithm).

    Now let's look at this problem: Find The K-th Smallest Pair Distance, and see how we can apply the trial and error algorithm.

    I -- Construct a candidate solution

    To construct a candidate solution, we need to understand first what the desired solution is. The problem description requires we output the K-th smallest pair distance, which is nothing more than a non-negative integer (since the input array nums is an integer array and pair distances are absolute values). Therefore our candidate solution should also be a non-negative integer.

    II -- Search space formed by all the candidate solutions

    Let min and max be the minimum and maximum numbers in the input array nums, and d = max - min, then any pair distance from nums must lie in the range [0, d]. As such, our desired solution is also within this range, which implies the search space will be [0, d] (any number outside this range can be ruled out immediately without further verification).

    III -- Verify a given candidate solution

    This is the key part of this trial and error algorithm. So given a candidate integer, how do we determine if it is the K-th smallest pair distance?

    First, what does the K-th smallest pair distance really mean? By definition, if we compute all the pair distances and sort them in ascending order, then the K-th smallest pair distance will be the one at index K - 1. This is essentially the naive way for solving this problem (but will be rejected due to MLE, as expected).

    Apparently the above definition cannot be used to do the verification, as it requires explicit computation of the pair distance array. Fortunately there is another way to define the K-th smallest pair distance: given an integer num, let count(num) denote the number of pair distances that are no greater than num, then the K-th smallest pair distance will be the smallest integer such that count(num) >= K.

    Here is a quick justification of the alternative definition. Let num_k be the K-th pair distance in the sorted pair distance array with index K - 1, as specified in the first definition. Since all the pair distances up to index K - 1 are no greater than num_k, we have count(num_k) >= K. Now suppose num is the smallest integer such that count(num) >= K, we show num must be equal to num_k as follows:

    1. If num_k < num, since count(num_k) >= K, then num will not be the smallest integer such that count(num) >= K, which contradicts our assumption.

    2. If num_k > num, since count(num) >= K, by definition of the count function, there are at least K pair distances that are no greater than num, which implies there are at least K pair distances that are smaller than num_k. This means num_k cannot be the K-th pair distance, contradicting our assumption again.

    Taking advantage of this alternative definition of the K-th smallest pair distance, we can transform the verification process into a counting process. So how exactly do we do the counting?

    IV -- Count the number of pair distances no greater than the given integer

    As I mentioned, we cannot use the pair distance arrays, which means the only option is the input array itself. If there is no order among its elements, we got no better way other than compute and test each pair distance one by one. This leads to a O(n^2) verification algorithm, which is as bad as, if not worse than, the aforementioned naive solution. So we need to impose some order to nums, which by default means sorting.

    Now suppose nums is sorted in ascending order, how do we proceed with the counting for a given number num? Note that each pair distance d_ij is characterized by a pair of indices (i, j) with i < j, that is d_ij = nums[j] - nums[i]. If we keep the first index i fixed, then d_ij <= num is equivalent to nums[j] <= nums[i] + num. This suggests that at least we can do a binary search to find the smallest index j such that nums[j] > nums[i] + num for each index i, then the count from index i will be j - i - 1, and in total we have an O(nlogn) verification algorithm.

    It turns out the counting can be done in linear time using the classic two-pointer technique if we make use of the following property: assume we have two starting indices i1 and i2 with i1 < i2, let j1 and j2 be the smallest index such that nums[j1] > nums[i1] + num and nums[j2] > nums[i2] + num, respectively, then it must be true that j2 >= j1. The proof is straightforward: suppose j2 < j1, since j1is the smallest index such that nums[j1] > nums[i1] + num, we should have nums[j2] <= nums[i1] + num. On the other hand, nums[j2] > nums[i2] + num >= nums[i1] + num. The two inequalities contradict each other, thus validate our conclusion above.

    V -- How to walk the search space efficiently

    Up to this point, we know the search space, know how to construct the candidate solution and how to verify it by counting, we still need one last piece for the puzzle: how to walk the search space.

    Of course we can do the naive linear walk by trying each integer from 0 up to d and choose the first integer num such that count(num) >= K. The time complexity will be O(nd). However, given that d can be much larger than n, this algorithm can be much worse than the naive O(n^2) solution mentioned before.

    The key observation here is that the candidate solutions are ordered naturally in ascending order, so a binary search is possible. Another fact is the non-decreasing property of the count function: give two integers num1 and num2 such that num1 < num2, then count(num1) <= count(num2) (I will leave the verification to you). So a binary walk of the search space will look like this:

    1. Let [l, r] be the current search space, and initialize l = 0, r = d.
    2. If l < r, compute the middle point m = (l + r) / 2 and evaluate count(m).
    3. If count(m) < K, we throw away the left half of current search space and set l = m + 1; else if count(m) >= K we throw away the right half and set r = m.

    You probably will wonder why we throw away the right half of the search space even if count(m) == K. Note that the K-th smallest pair distance num_k is the minimum integer such that count(num_k) >= K. If count(m) == K, then we know num_k <= m (but not num_k == m, think about it!) so it makes no sense keeping the right half.

    VI -- Putting everything together, aka, solutions

    Don't get scared by the above analyses. The final solution is much simpler to write once you understand it. Here is the Java program for the trial and error algorithm. The time complexity is O(nlogd + nlogn) (don't forget the sorting) and space complexity is O(1).

    public int smallestDistancePair(int[] nums, int k) {
        int n = nums.length, l = 0, r = nums[n - 1] - nums[0];
        while (l < r) {
            int m = l + ((r - l) >> 1), cnt = 0;
            for (int i = 0, j = 0; i < n; i++) {
                while (j < n && nums[j] <= nums[i] + m) j++;
                cnt += j - i - 1;
            if (cnt < k) {
                l = m + 1;
            } else {
                r = m;
        return l;

    Lastly here is a list of LeetCode problems that can be solved using the trial and error algorithm (you're welcome to add more):

    1. 719. Find K-th Smallest Pair Distance

    2. 668. Kth Smallest Number in Multiplication Table

    3. 644. Maximum Average Subarray II

    Anyway, this post is just a reminder to you that the trial and error algorithm is worth trying if you find all other common solutions suffer severely from bad time or space performance. Also it's always recommended to perform a quick evaluation of the search space size and potential verification algorithm to estimate the complexity before you are fully committed to this algorithm.

  • 0

    this analysis/tutorial rocks!

Log in to reply

Looks like your connection to LeetCode Discuss was lost, please wait while we try to reconnect.