Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

The time complexity for the merge sort algorithm is $T(n) = 2T(n/2)+\Theta(n)$, and the solution to this recurrence is $\Theta(n\lg n)$.

However; assume you are not dividing the array in half for sorting, but instead in, say, $a$ pieces. Then the time would be something like $T(n) = aT(n/a) + \Theta(n \log_2 a)$. (The last term being the time to perform divide-and-merge for $a$ pieces with a total size of $n$. For positive, integer $a$, what is the solution to this recurrence? (assuming $T(n) = 1$ for $n = 1$ and that $n=a^x$ where x is a positive integer as well).

I have a hunch it would probably be $\Theta(n \log_2 n)$ for any $a$, but I'm not really sure how to prove it, with the logarithm being in a different base than $2$ .

share|improve this question
2  
This is not a proof, but $n \log n$ seems to satisfy this equation quite nicely: $a (n/a) \log (n/a) + n \log a$ simplifies to the $n \log n$ we were looking for. Can we turn this into a proof by induction? –  Srivatsan Jul 28 '11 at 19:03

2 Answers 2

up vote 1 down vote accepted

Use the Master Theorem, case 2.

share|improve this answer
1  
Ahh, yes. Obviously $\Theta(n \log_2 a) = \Theta(n)$ since $\log_2 a$ doesn't change when $n$ grows, didn't think about that... Thanks! –  Johan Sannemo Jul 28 '11 at 19:17

This is the one piece of time-complexity theory that I remember from my undergrad Algorithms course =) The solution is known in [Cormen-Leiserson-Rivest] as the "master method". The text has the full formula, and I refer you there for more explanation.

The behavior of the recurrence $T(n) = aT(n/a) + \Theta(n)$ could be a bit different than that of $T(n) = aT(n/a) + \Theta(n\log_a n)$ due to the larger overhead of the latter. In the former case (including $a=2$ in your first line), the time complexity works out to $$ n^{\log_a(a)} \lg n = n \lg n,$$ as you have written. In the latter case, with overhead $\Theta(n \log_a n)$, we have to be more clever... [I'll write more later, as obligations are pulling me away from the computer at the moment].

share|improve this answer
    
Edited it a while ago: actually meant $n \log_2 a$, not $n \log_a n$, sorry! –  Johan Sannemo Jul 28 '11 at 19:12

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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