Write a program to find 100 largest numbers out of an array of 1 billion numbers

I recently attended an interview where I was asked "write a program to find 100 largest numbers out of an array of 1 billion numbers."

I was only able to give a brute force solution which was to sort the array in O(nlogn) time complexity and take the last 100 numbers.

Arrays.sort(array);

The interviewer was looking for a better time complexity, I tried a couple of other solutions but failed to answer him. Is there a better time complexity solution?

You can keep a priority queue of the 100 biggest numbers, iterate through the 1 billion numbers. Whenever you encounter a number greater than the smallest number in the queue (the head of the queue), remove the head of the queue and add the new number to the queue.

A priority queue implemented with a heap has insert + delete complexity of O(log K). (Where K = 100, the number of elements to find. N = 1 billion, the number of total elements in the array).

In the worst case you get billion*log2(100) which is better than billion*log2(billion) for an O(N log N) comparison-based sort¹.

In general, if you need the largest K numbers from a set of N numbers, the complexity is O(N log K) rather than O(N log N), this can be very significant when K is very small comparing to N.

---

The expected time of this priority queue algorithm is pretty interesting, since in each iteration an insertion may or may not occur.

The probability of the i'th number to be inserted to the queue is the probability of a random variable being larger than at least i-K random variables from the same distribution (the first k numbers are automatically added to the queue). We can use order statistics (see link) to calculate this probability.

For example, lets assume the numbers were randomly selected uniformly from {0, 1}, the expected value of (i-K)th number (out of i numbers) is (i-k)/i, and chance of a random variable being larger than this value is 1-[(i-k)/i] = k/i.

Thus, the expected number of insertions is:

And the expected running time can be expressed as:

(k time to generate the queue with the first k elements, then n-k comparisons, and the expected number of insertions as described above, each takes an average log(k)/2 time)

Note that when N is very large comparing to K, this expression is a lot closer to n rather than N log K. This is somewhat intuitive, as in the case of the question, even after 10,000 iterations (which is very small comparing to a billion), the chance of a number to be inserted to the queue is very small.

But we don't know that the array values are uniformly distributed. They might trend towards increasing, in which case most or all numbers will be be new candidates for the set of 100 largest numbers seen. The worst case for this algorithm is O(N log K).

Or if they trend towards decreasing, most of the largest 100 numbers will be very early, and our best-case run time is essentially O(N + K log K), which is just O(N) for K much smaller than N.

---

Footnote 1: O(N) integer sorting / histogramming

Counting Sort or Radix Sort are both O(N), but often have larger constant factors that make them worse than comparison sorts in practice. In some special cases they're actually quite fast, primarily for narrow integer types.

For example, Counting Sort does well if the numbers are small. 16-bit numbers would only need an array of 2^16 counters. And instead of actually expanding back into a sorted array, you could just scan the histogram you build as part of Counting Sort.

After histogramming an array, you can quickly answer queries for any order statistic, e.g. the 99 largest numbers, the 200 to 100th largest numbers.) 32-bit numbers would scatter the counts over a much larger array or hash table of counters, potentially needing 16 GiB of memory (4 bytes for each of 2^32 counters). And on real CPUs, probably getting lots of TLB and cache misses, unlike an array of 2^16 elements where L2 cache would typically hit.

Similarly, Radix Sort could look at only the top buckets after a first pass. But the constant factors may still be larger than log K, depending on K.

Note that the size of each counter is large enough to not overflow even if all N integers are duplicates. 1 billion is somewhat below 2^30, so a 30-bit unsigned counter would be sufficient. And a 32-bit signed or unsigned integer is just fine.

If you had many more, you might need 64-bit counters, taking twice the memory footprint to initialize to zero and to randomly access. Or a sentinel value for the few counters that overflow a 16 or 32-bit integer, to indicate that the rest of the count is elsewhere (in a small dictionary such as a hash table mapping to 64-bit counters).