Algorithm to calculate the number of 1s for a range of numbers in binary

Question

So I just got back for the ACM Programing competition and did pretty well but there was one problem that not one team got.

The Problem.

Answer 1

I think a key is first understanding the pattern of K values and how rapidly it grows. Basically, you have:

K(1) = 0
K(X) = K(bitcount(X))+1 for X > 1

So finding the smallest X values for a given K we see

K(1) = 0
K(2) = 1
K(3) = 2
K(7) = 3
K(127) = 4
K(170141183460469231731687303715884105727) = 5

So for an example like 48238 10^18 9 the answer is trivially 0. K=0 only for 1, and K=1 only for powers of 2, so in the range of interest, we'll pretty much only see K values of 2, 3 or 4, and never see K >= 5

edit

Ok, so we're looking for an algorithm to count the number of values with K=2,3,4 in a range of value LO..HI without iterating over the entire range. So the first step is to find the number of values in the range with bitcount(x)==i for i = 1..59 (since we only care about values up to 10^18 and 10^18 < 2^60). So break down the range lo..hi into subranges that are a power of 2 size and differ only in their lower n bits -- a range of the form x*(2^n)..(x+1)*(2^n)-1. We can break down the arbitray lo..hi range into such subranges easily. For each such subrange there will be choose(n, i) values with i+bitcount(x) set bits. So we just add all the subranges together to get a vector of counts for 1..59, which we then iterate over, adding together those elements with the same K value to get our answer.

edit (fixed again to be be C89 compatible and work for lo=1/k=0)

Here's a C program to do what I previously described:

#include 
#include 
#include 

int bitcount(long long x) {
    int rv = 0;
    while(x) { rv++; x &= x-1; }
    return rv; }

long long choose(long long m, long long n) {
    long long rv = 1;
    int i;
    for (i = 0; i < n; i++) {
        rv *= m-i;
        rv /= i+1; }
    return rv; }

void bitcounts_p2range(long long *counts, long long base, int l2range) {
    int i;
    assert((base & ((1LL << l2range) - 1)) == 0);
    counts += bitcount(base);
    for (i = 0; i <= l2range; i++)
        counts[i] += choose(l2range, i); }

void bitcounts_range(long long *counts, long long lo, long long hi) {
    int l2range = 0;
    while (lo + (1LL << l2range) - 1 <= hi) {
        if (lo & (1LL << l2range)) {
            bitcounts_p2range(counts, lo, l2range);
            lo += 1LL << l2range; }
        l2range++; }
    while (l2range >= 0) {
        if (lo + (1LL << l2range) - 1 <= hi) {
            bitcounts_p2range(counts, lo, l2range);
            lo += 1LL << l2range; }
        l2range--; }
    assert(lo == hi+1); }

int K(int x) {
    int rv = 0;
    while(x > 1) {
        x = bitcount(x);
        rv++; }
    return rv; }

int main() {
    long long counts[64];
    long long lo, hi, total;
    int i, k;
    while (scanf("%lld%lld%d", &lo, &hi, &k) == 3) {
        if (lo < 1 || lo > hi || k < 0) break;
        if (lo == 0 || hi == 0 || k == 0) break;
        total = 0;
        if (lo == 1) {
            lo++;
            if (k == 0) total++; }
        memset(counts, 0, sizeof(counts));
        bitcounts_range(counts, lo, hi);
        for (i = 1; i < 64; i++)
            if (K(i)+1 == k)
                total += counts[i];
        printf("%lld\n", total); }
    return 0; }

which runs just fine for values up to 2^63-1 (LONGLONG_MAX). For 48238 1000000000000000000 3 it gives 513162479025364957, which certainly seems plausible

edit

giving the inputs of

48238 1000000000000000000 1
48238 1000000000000000000 2
48238 1000000000000000000 3
48238 1000000000000000000 4

gives outputs of

44
87878254941659920
513162479025364957
398959266032926842

Those add up to 999999999999951763 which is correct. The value for k=1 is correct (there are 44 powers of two in that range 2^16 up to 2^59). So while I'm not sure the other 3 values are correct, they're certainly plausible.