I have an array of floats, sorted from smallest to largest, and need to be able to pick out the nearest float greater than or less than a passed input value. This input value is not necessarily present as a value in the array.

A naive approach would be to do a simple linear search through the array. That might look like this:

void FindClosestFloatsInArray( float input, std::vector<float> array, 
                               float *min_out, float *max_out )
    assert( input >= array[0] && input < array[ array.size()-1 ] );
    for( int i = 1; i < array.size(); i++ )
        if ( array[i] >= input )
            *min = array[i-1];
            *max = array[i];

But obviously as the array gets larger, this will become slower and slower.

Does anyone have an idea about an algorithm that would let me find this data more optimally? I've already switched to a binary search, which has improved matters somewhat, but it's still a lot slower than I'd like, and since I'm not actually looking for a specific value that exists in the array, it can never terminate early.

More information: The floating point values in the array are not necessarily distributed evenly (that is, the array could consist of the values "1.f, 2.f, 3.f, 4.f, 100.f, 1200.f, 1203.f, 1400.f".

I'm doing this operation hundreds of thousands of times, but I can do any amount of pre-processing on the array of floats, if it'll improve the lookup time. I absolutely can change to use something other than a vector to store them, if that'll help.

  • What makes you think that your binary search can not terminate early ? Surely you can just test the elements at i and i + 1 to see if they bracket the target value, and terminate if they do ? – Paul R Sep 20 '11 at 6:58
  • Alternately, I could test the elements at i and i-1 to see if they bracket the target value. I would also need to test whether 'i' was >= array.size()-1 so I could avoid doing your test, and whether it was <= 0 so I could avoid doing my test... it's actually a lot of extra conditionals to perform at each step, in order to check for an early-out. I imagine they would slow down the algorithm a lot, though I'll confess that I haven't actually profiled that yet. – Trevor Powell Sep 20 '11 at 7:27
  • 3
    It doesn't need to be so complicated - if your array is of size N then you just need to treat it as if it were of size N - 1. That way there is always a valid element at i + 1. You do a binary search over N - 1 element for element i which is less than your target value, with element i + 1 being greater then the target value. – Paul R Sep 20 '11 at 7:54

The code in the question (a linear search), as you rightly point out, is going to get slow for large float arrays. Technically it's O(n) where n is the number of float values in your array.

In general, the best that you can do to find a value in an ordered array is a recursive tree search of some kind (e.g. binary search), in which case you can achieve an O(log n) lookup time in the number of elements in your array. O(log n) is much better than O(n) for large values of n.

My suggested approach would therefore be a simple binary search of the array, i.e.:

  1. Set min/max integer indexes to cover your whole float array
  2. test the value in the middle of the range at index mid=(min+max/2) against the search value x
  3. if x is lower than this value, set max to mid, else set min to mid
  4. repeat (2-4) until you have found the correct value

This is an O(log n) algorithm which should be fast enough for nearly all situations. Intuitively, it works by halving the range to be searched at each step until you find the correct value.

It's really hard to beast the simple binary search, so if you have already implemented this correctly then you may be pretty close to optimal already. However, if you know the distributions of the data and/or have a limited range of lookup values (x), there are still some other more advanced tricks you can try:

  • Bucketing - create buckets (e.g. for each interval between two integers), each of which contains a smaller sorted list of the float values between the two bounding integers plus two values immediately below and immediately above each range. You can then start your search at (trunc(x)+0.5). This should give you a good speedup if you choose appropriately sized buckets (it's effectively increasing the branching factor of the tree.....). If integers don't work for you, then you can try buckets of some other fixed-point precision (e.g. multiples of 1/16).
  • Bit-mapping - if the range of possible lookup values is small enough, you could try creating a big lookup table indexed by the bitwise value of x. This will be O(1) but you may need a lot of memory which will be very unfriendly on your cache... so use with caution. This is especially nasty because you are looking up float values, so you may well need several GBs to account for all of the less significant bits......
  • Rounding and hashing - hash tables are probably not the best data structure for this problem, but if you can survive losing a bit of accuracy they could work - simply round off the lowest bits of your lookup values and use a hashmap to directly look up the correct value. You will have to experiment on the right trade-off between hashmap size and precision, and also ensure that all possible hash values are populated so this can be a bit tricky......
  • Tree-balancing - your ideal tree should have a 50% chance of going left or right. So if you create a tree based on the distribution of lookup values (x), then you can optimise the tree to produce answers with the minimal amount of tests. This is likely to be a good solution if a lot of values in your float array are very close together, since it will enable you to avoid searching these branches too often.
  • Crit-bit trees - these are still trees (so still O(log n)...) but some cases: you would however need to convert your floats into some fixed-point format in order to make the comparisons work

However, unless you are in a very special situation I'd probably recommend sticking with the simple binary search. Reasons:

  • it's much easier to implement
  • it's very fast for most common cases
  • the extra overhead of the more complex approaches (e.g. higher memory usage / cache pressure) often outweighs the minor theoretical gains
  • it will be more robust to future changes in the data distributions....
| improve this answer | |

This seems simple enough:

Do a binary search for the float that you want to bound - O (log n) time.

Then the element to the left of it is the lower bound, and the element to the right of it is the upper bound.

| improve this answer | |

The obvious answer is to store the floats in a tree. Supporting 'previous' and 'next' operations are trivial in a tree. So just do a 'next' on your value, and then do a 'previous' on the value you find in the first step.

| improve this answer | |
  • 1
    This is essentially the same as a binary search. – kevin cline Dec 19 '14 at 17:54

This paper ("sublogarithmic searching without multiplications") might be of interest; it even contains some source code. For the purposes of comparison, you can treat a float number as an integer with the same bit-pattern; this was one of the design goals of the IEEE floating point standard.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.