My interview question was that I need to return the length of an array that removed duplicates but we can leave at most 2 duplicates.

For example, [1, 1, 1, 2, 2, 3] the new array would be [1, 1, 2, 2, 3]. So the new length would be 5. I came up with an algorithm with \$O(2n)\$ I believe.

def removeDuplicates(nums):
    if nums is None:
        return 0

    if len(nums) == 0:
        return 0

    if len(nums) == 1:
        return 1

    new_array = {}
    for num in nums:
        new_array[num] = new_array.get(num, 0) + 1

    new_length = 0
    for key in new_array:
        if new_array[key] > 2:
            new_length = new_length + 2
            new_length = new_length + new_array[key]

    return new_length

new_length = removeDuplicates([1, 1, 1, 2, 2, 3])
assert new_length == 5

Can I improve this code to be any faster?

  • \$\begingroup\$ Cross-posted from Stack Overflow \$\endgroup\$ Jul 17, 2015 at 4:30
  • 2
    \$\begingroup\$ Just for clarity, can it leave two duplicates or must it leave two in situations where there are 2 or more. It gives you two very different answers. \$\endgroup\$
    – Taekahn
    Jul 17, 2015 at 4:34
  • \$\begingroup\$ Note that O(2n) is the same as O(n); don't include constants like 2 in the big-O notation. \$\endgroup\$ Jul 17, 2015 at 13:08

2 Answers 2


I am interviewing myself these days, and I think I am developing a good sense of what interviewers are after. Regarding your code, here are a few things I would have done differently:

  1. Know the language conventions! Python's PEP8 dictates lowercase_with_underscores format for function and variable names.

  2. Avoid unnecessary explicit checks. Your function should work on any iterable, there is no point in checking for None and not for, e.g. an int or a float, which would also error out when you try to compute their len. If you want to be pedantic about it, do something like:

    from collections import Iterable
    assert isinstance(nums, Iterable)

    I would personally leave that check entirely out.

  3. Your other checks are also pretty much unnecessary, as Python has very well defined rules for iteration over empty iterables (does nothing), and adding a special path for size-one iterables could easily be dismissed as premature optimization (the root of all evil!). Even if you ignore this advice, PEP8 prefers if not nums over if len(nums) == 0.

  4. If you reorganize your code a little, you can get rid of the second loop entirely:

    counts = {}
    new_length = 0
    for num in nums:
        count = counts.get(num, 0)
        if count < 2:
            counts[num] = count + 1
            new_length += 1

If you put it altogether, your function would end up looking like:

def remove_duplicates(nums, dups_to_keep=2):
    counts = {}
    new_len = 0
    for num in nums:
        count = counts.get(num, 0)
        if count < dups_to_keep:
            counts[num] = count + 1
            new_len += 1
    return new_len

It may be marginally faster than your code, but I think the most important improvement is in compactness and readability.


You're just reimplementing collections.Counter.

def remove_duplicates(lst):
    return sum(min(x, 2) for x in collections.Counter(lst).values())

By the way, \$O(2n)\$ is equivalent \$O(n)\$. However, the bound is actually tighter: \$O(n + m)\$, where m is the number of distinct elements.

This is, however, an average bound, because it relies on a dict. If the worst case hash collisions happen, you get \$O(m(n+m))\$

  • \$\begingroup\$ I had to reimplement as I was asked during an interview. \$\endgroup\$
    – toy
    Jul 17, 2015 at 4:44
  • \$\begingroup\$ Since m <= n the proper answer is O(n), or O(n**2) if you want to be pedantic about hash table performance. \$\endgroup\$
    – Jaime
    Jul 17, 2015 at 5:50
  • \$\begingroup\$ @Jaime It is equivalent in the \$O(n)\$ case, but \$O(m(n+m))\$ is smaller than \$O(n^2)\$ \$\endgroup\$
    – o11c
    Jul 17, 2015 at 5:53
  • \$\begingroup\$ Big O notation is a worst case bound, which in this case is O(n^2) if all values are unique. \$\endgroup\$
    – Jaime
    Jul 17, 2015 at 5:57
  • \$\begingroup\$ @Jaime and likewise it's \$O(n^3)\$ but we don't say that because we can do better. \$\endgroup\$
    – o11c
    Jul 17, 2015 at 5:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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