I looked at LeetCode question 270. Perfext Squares:
Given an integer n, return the least number of perfect square numbers that sum to n.
A perfect square is an integer that is the square of an integer; in other words, it is the product of some integer with itself. For example, 1, 4, 9, and 16 are perfect squares while 3 and 11 are not.>
Example 1:
Input: n = 12
Output: 3
Explanation: 12 = 4 + 4 + 4.
I solved it using the following algorithm:
def numSquares(n):
squares = [i**2 for i in range(1, int(n**0.5)+1)]
step = 1
queue = {n}
while queue:
tempQueue = set()
for node in queue:
for square in squares:
if node-square == 0:
return step
if node < square:
break
tempQueue.add(node-square)
queue = tempQueue
step += 1
It basically tries to go from goal number to 0 by subtracting each possible number, which are : [1 , 4, 9, .. sqrt(n)] and then does the same work for each of the numbers obtained.
Question
What is the time complexity of this algorithm? The branching in every level is sqrt(n) times, but some branches are destined to end early... which makes me wonder how to derive the time complexity.
If you think about what you're doing, you can imagine that you're doing a breadth-first search over a graph with n + 1 nodes (all the natural numbers between 0 and n, inclusive) and some number of edges m, which we'll determine later on. Your graph is essentially represented as an adjacency list, since at each point you iterate over all the outgoing edges (squares less than or equal to your number) and stop as soon as you consider a square that's too large. As a result, the runtime will be O(n + m), and all we have to do now is work out what m is.
(There's another cost here in computing all the square roots up to and including n, but that takes time O(n1/2), which is dominated by the O(n) term.)
If you think about it, the number of outgoing edges from each number k will be given by the number of perfect squares less than or equal to k. That value is equal to ⌊√k⌋ (check this for a few examples - it works!). This means that the total number of edges is upper-bounded by
√0 + √1 + √2 + ... + √n
We can show that this sum is Θ(n3/2). First, we'll upper-bound this sum at O(n3/2), which we can do by noting that
√0 + √1 + √2 + ... + √n
≤ √n + √n + √ n + ... + √n (n+1) times
= (n + 1)√n
= O(n3/2).
To lower-bound this at Ω(n3/2), notice that
√0 + √1 + √2 + ... + √ n
≥ √(n/2) + √(n/2 + 1) + ... + √(n) (drop the first half of the terms)
≥ √(n/2) + √(n/2) + ... + √(n/2)
= (n / 2)√(n / 2)
= Ω(n3/2).
So overall, the number of edges is Θ(n3/2), so using a regular analysis of breadth-first search we can see that the runtime will be O(n3/2).
This bound is likely not tight, because this assumes that you visit every single node and every single edge, which isn't going to happen. However, I'm not sure how to tighten things much beyond this.
As a note - this would be a great place to use A* search instead of breadth-first search, since you can fairly easily come up with heuristics to underestimate the remaining total distance (say, take the number and divide it by the largest perfect square less than it). That would cause the search to focus on extremely promising paths that jump rapidly toward 0 before less-good paths, like, say, always taking steps of size one.
Hope this helps!
Some observations:
The number of squares up to n is √n (floored to the nearest integer)
After the first iteration of the while loop, tempQueue will have √n entries
tempQueue can never have more than n entries, since all these values are positive, less than n and unique.
Every natural number can be written as the sum of four integer squares. So that means your BFS algorithm's while loop will iterate at the most 4 times. If the return statement did not get executed during any of the first 3 iterations, it is guaranteed it will in the 4th.
Every statement (except for the initialisation of squares) runs in constant time, even the call to .add().
The initialisation of squares has a list comprehension loop that has √n iterations, and range runs in constant time, so that initialisation has a time complexity of O(√n).
Now we can set a ceiling to the number of times the if node-square == 0 statement is executed (or any other statement in the innermost loop's body):
1⋅√n + √n⋅√n + n⋅√n + n⋅√n
Each of the 4 terms corresponds to an iteration of the while loop. The left factor of each product corresponds to the maximum size of queue in that particular iteration, and the factor at the right corresponds to the size of squares (always the same). This simplifies to:
√n + n + 2n3⁄2
In terms of time complexity this is:
O(n3⁄2)
This is the worst case time complexity. When the while loop only has to iterate twice, it is O(n), and when only once (when n is a square), it is O(√n).
The binary-search algorithm takes log(n) time, because of the fact that the height of the tree (with n nodes) would be log(n).
How would you prove this?
Now here I am not giving mathematical proof. Try to understand the problem using log to the base 2. Log2 is the normal meaning of log in computer science.
First, understand it is binary logarithm (log2n) (logarithm to the base 2).
For example,
the binary logarithm of 1 is 0
the binary logarithm of 2 is 1
the binary logarithm of 3 is 1
the binary logarithm of 4 is 2
the binary logarithm of 5, 6, 7 is 2
the binary logarithm of 8-15 is 3
the binary logarithm of 16-31 is 4 and so on.
For each height the number of nodes in a fully balanced tree are
Height Nodes Log calculation
0 1 log21 = 0
1 3 log23 = 1
2 7 log27 = 2
3 15 log215 = 3
Consider a balanced tree with between 8 and 15 nodes (any number, let's say 10). It is always going to be height 3 because log2 of any number from 8 to 15 is 3.
In a balanced binary tree the size of the problem to be solved is halved with every iteration. Thus roughly log2n iterations are needed to obtain a problem of size 1.
I hope this helps.
Let's assume at first that the tree is complete - it has 2^N leaf nodes. We try to prove that you need N recursive steps for a binary search.
With each recursion step you cut the number of candidate leaf nodes exactly by half (because our tree is complete). This means that after N halving operations there is exactly one candidate node left.
As each recursion step in our binary search algorithm corresponds to exactly one height level the height is exactly N.
Generalization to all balanced binary trees: If the tree has less nodes than 2^N we for sure don't need more halvings. We might need less or the same amount but never more.
Assuming that we have a complete tree to work with, we can say that at depth k, there are 2k nodes. You can prove this using simple induction, based on the intuition that adding an extra level to the tree will increase the number of nodes in the entire tree by the number of nodes that were in the previous level times two.
The height k of the tree is log(N), where N is the number of nodes. This can be stated as
log2(N) = k,
and it is equivalent to
N = 2k
To understand this, here's an example:
16 = 24 => log2(16) = 4
The height of the tree and the number of nodes are related exponentially. Taking the log of the number of nodes just allows you to work backwards to find the height.
Just look up the rigorous proof in Knuth, Volume 3 - Searching and Sorting Algorithms ... He does it far more rigorously than anyone else I can think of.
http://en.wikipedia.org/wiki/The_Art_of_Computer_Programming
You can find it in any good Computer Science library and on the bookshelves of many (very) old geeks.
Why is the height of a balanced binary tree equal to ceil(log2N) for N nodes?
w = width of base (maximum number of leaves)
h = height of tree (maximum number of edges from root to leaf)
Divide w by 2 (h times) to get to 1, which counts the single root node at top.
N = w + w/2 + ... + 1
N = 2h + ... + 21 + 20
= (1-2h+1) / (1-2) = 2h+1-1
log2(N+1) = h+1
Check: if N=1, h=0. If h=1, N=3.
This formula is for if the bottom level is full. N will not always be so great, but would still have the same height, h. So we must take the log's ceiling.
My knowledge of big-O is limited, and when log terms show up in the equation it throws me off even more.
Can someone maybe explain to me in simple terms what a O(log n) algorithm is? Where does the logarithm come from?
This specifically came up when I was trying to solve this midterm practice question:
Let X(1..n) and Y(1..n) contain two lists of integers, each sorted in nondecreasing order. Give an O(log n)-time algorithm to find the median (or the nth smallest integer) of all 2n combined elements. For ex, X = (4, 5, 7, 8, 9) and Y = (3, 5, 8, 9, 10), then 7 is the median of the combined list (3, 4, 5, 5, 7, 8, 8, 9, 9, 10). [Hint: use concepts of binary search]
I have to agree that it's pretty weird the first time you see an O(log n) algorithm... where on earth does that logarithm come from? However, it turns out that there's several different ways that you can get a log term to show up in big-O notation. Here are a few:
Repeatedly dividing by a constant
Take any number n; say, 16. How many times can you divide n by two before you get a number less than or equal to one? For 16, we have that
16 / 2 = 8
8 / 2 = 4
4 / 2 = 2
2 / 2 = 1
Notice that this ends up taking four steps to complete. Interestingly, we also have that log2 16 = 4. Hmmm... what about 128?
128 / 2 = 64
64 / 2 = 32
32 / 2 = 16
16 / 2 = 8
8 / 2 = 4
4 / 2 = 2
2 / 2 = 1
This took seven steps, and log2 128 = 7. Is this a coincidence? Nope! There's a good reason for this. Suppose that we divide a number n by 2 i times. Then we get the number n / 2i. If we want to solve for the value of i where this value is at most 1, we get
n / 2i ≤ 1
n ≤ 2i
log2 n ≤ i
In other words, if we pick an integer i such that i ≥ log2 n, then after dividing n in half i times we'll have a value that is at most 1. The smallest i for which this is guaranteed is roughly log2 n, so if we have an algorithm that divides by 2 until the number gets sufficiently small, then we can say that it terminates in O(log n) steps.
An important detail is that it doesn't matter what constant you're dividing n by (as long as it's greater than one); if you divide by the constant k, it will take logk n steps to reach 1. Thus any algorithm that repeatedly divides the input size by some fraction will need O(log n) iterations to terminate. Those iterations might take a lot of time and so the net runtime needn't be O(log n), but the number of steps will be logarithmic.
So where does this come up? One classic example is binary search, a fast algorithm for searching a sorted array for a value. The algorithm works like this:
If the array is empty, return that the element isn't present in the array.
Otherwise:
Look at the middle element of the array.
If it's equal to the element we're looking for, return success.
If it's greater than the element we're looking for:
Throw away the second half of the array.
Repeat
If it's less than the the element we're looking for:
Throw away the first half of the array.
Repeat
For example, to search for 5 in the array
1 3 5 7 9 11 13
We'd first look at the middle element:
1 3 5 7 9 11 13
^
Since 7 > 5, and since the array is sorted, we know for a fact that the number 5 can't be in the back half of the array, so we can just discard it. This leaves
1 3 5
So now we look at the middle element here:
1 3 5
^
Since 3 < 5, we know that 5 can't appear in the first half of the array, so we can throw the first half array to leave
5
Again we look at the middle of this array:
5
^
Since this is exactly the number we're looking for, we can report that 5 is indeed in the array.
So how efficient is this? Well, on each iteration we're throwing away at least half of the remaining array elements. The algorithm stops as soon as the array is empty or we find the value we want. In the worst case, the element isn't there, so we keep halving the size of the array until we run out of elements. How long does this take? Well, since we keep cutting the array in half over and over again, we will be done in at most O(log n) iterations, since we can't cut the array in half more than O(log n) times before we run out of array elements.
Algorithms following the general technique of divide-and-conquer (cutting the problem into pieces, solving those pieces, then putting the problem back together) tend to have logarithmic terms in them for this same reason - you can't keep cutting some object in half more than O(log n) times. You might want to look at merge sort as a great example of this.
Processing values one digit at a time
How many digits are in the base-10 number n? Well, if there are k digits in the number, then we'd have that the biggest digit is some multiple of 10k. The largest k-digit number is 999...9, k times, and this is equal to 10k + 1 - 1. Consequently, if we know that n has k digits in it, then we know that the value of n is at most 10k + 1 - 1. If we want to solve for k in terms of n, we get
n ≤ 10k+1 - 1
n + 1 ≤ 10k+1
log10 (n + 1) ≤ k + 1
(log10 (n + 1)) - 1 ≤ k
From which we get that k is approximately the base-10 logarithm of n. In other words, the number of digits in n is O(log n).
For example, let's think about the complexity of adding two large numbers that are too big to fit into a machine word. Suppose that we have those numbers represented in base 10, and we'll call the numbers m and n. One way to add them is through the grade-school method - write the numbers out one digit at a time, then work from the right to the left. For example, to add 1337 and 2065, we'd start by writing the numbers out as
1 3 3 7
+ 2 0 6 5
==============
We add the last digit and carry the 1:
1
1 3 3 7
+ 2 0 6 5
==============
2
Then we add the second-to-last ("penultimate") digit and carry the 1:
1 1
1 3 3 7
+ 2 0 6 5
==============
0 2
Next, we add the third-to-last ("antepenultimate") digit:
1 1
1 3 3 7
+ 2 0 6 5
==============
4 0 2
Finally, we add the fourth-to-last ("preantepenultimate"... I love English) digit:
1 1
1 3 3 7
+ 2 0 6 5
==============
3 4 0 2
Now, how much work did we do? We do a total of O(1) work per digit (that is, a constant amount of work), and there are O(max{log n, log m}) total digits that need to be processed. This gives a total of O(max{log n, log m}) complexity, because we need to visit each digit in the two numbers.
Many algorithms get an O(log n) term in them from working one digit at a time in some base. A classic example is radix sort, which sorts integers one digit at a time. There are many flavors of radix sort, but they usually run in time O(n log U), where U is the largest possible integer that's being sorted. The reason for this is that each pass of the sort takes O(n) time, and there are a total of O(log U) iterations required to process each of the O(log U) digits of the largest number being sorted. Many advanced algorithms, such as Gabow's shortest-paths algorithm or the scaling version of the Ford-Fulkerson max-flow algorithm, have a log term in their complexity because they work one digit at a time.
As to your second question about how you solve that problem, you may want to look at this related question which explores a more advanced application. Given the general structure of problems that are described here, you now can have a better sense of how to think about problems when you know there's a log term in the result, so I would advise against looking at the answer until you've given it some thought.
When we talk about big-Oh descriptions, we are usually talking about the time it takes to solve problems of a given size. And usually, for simple problems, that size is just characterized by the number of input elements, and that's usually called n, or N. (Obviously that's not always true-- problems with graphs are often characterized in numbers of vertices, V, and number of edges, E; but for now, we'll talk about lists of objects, with N objects in the lists.)
We say that a problem "is big-Oh of (some function of N)" if and only if:
For all N > some arbitrary N_0, there is some constant c, such that the runtime of the algorithm is less than that constant c times (some function of N.)
In other words, don't think about small problems where the "constant overhead" of setting up the problem matters, think about big problems. And when thinking about big problems, big-Oh of (some function of N) means that the run-time is still always less than some constant times that function. Always.
In short, that function is an upper bound, up to a constant factor.
So, "big-Oh of log(n)" means the same thing that I said above, except "some function of N" is replaced with "log(n)."
So, your problem tells you to think about binary search, so let's think about that. Let's assume you have, say, a list of N elements that are sorted in increasing order. You want to find out if some given number exists in that list. One way to do that which is not a binary search is to just scan each element of the list and see if it's your target number. You might get lucky and find it on the first try. But in the worst case, you'll check N different times. This is not binary search, and it is not big-Oh of log(N) because there's no way to force it into the criteria we sketched out above.
You can pick that arbitrary constant to be c=10, and if your list has N=32 elements, you're fine: 10*log(32) = 50, which is greater than the runtime of 32. But if N=64, 10*log(64) = 60, which is less than the runtime of 64. You can pick c=100, or 1000, or a gazillion, and you'll still be able to find some N that violates that requirement. In other words, there is no N_0.
If we do a binary search, though, we pick the middle element, and make a comparison. Then we throw out half the numbers, and do it again, and again, and so on. If your N=32, you can only do that about 5 times, which is log(32). If your N=64, you can only do this about 6 times, etc. Now you can pick that arbitrary constant c, in such a way that the requirement is always met for large values of N.
With all that background, what O(log(N)) usually means is that you have some way to do a simple thing, which cuts your problem size in half. Just like the binary search is doing above. Once you cut the problem in half, you can cut it in half again, and again, and again. But, critically, what you can't do is some preprocessing step that would take longer than that O(log(N)) time. So for instance, you can't shuffle your two lists into one big list, unless you can find a way to do that in O(log(N)) time, too.
(NOTE: Nearly always, Log(N) means log-base-two, which is what I assume above.)
In the following solution, all the lines with a recursive call are done on
half of the given sizes of the sub-arrays of X and Y.
Other lines are done in a constant time.
The recursive function is T(2n)=T(2n/2)+c=T(n)+c=O(lg(2n))=O(lgn).
You start with MEDIAN(X, 1, n, Y, 1, n).
MEDIAN(X, p, r, Y, i, k)
if X[r]<Y[i]
return X[r]
if Y[k]<X[p]
return Y[k]
q=floor((p+r)/2)
j=floor((i+k)/2)
if r-p+1 is even
if X[q+1]>Y[j] and Y[j+1]>X[q]
if X[q]>Y[j]
return X[q]
else
return Y[j]
if X[q+1]<Y[j-1]
return MEDIAN(X, q+1, r, Y, i, j)
else
return MEDIAN(X, p, q, Y, j+1, k)
else
if X[q]>Y[j] and Y[j+1]>X[q-1]
return Y[j]
if Y[j]>X[q] and X[q+1]>Y[j-1]
return X[q]
if X[q+1]<Y[j-1]
return MEDIAN(X, q, r, Y, i, j)
else
return MEDIAN(X, p, q, Y, j, k)
The Log term pops up very often in algorithm complexity analysis. Here are some explanations:
1. How do you represent a number?
Lets take the number X = 245436. This notation of “245436” has implicit information in it. Making that information explicit:
X = 2 * 10 ^ 5 + 4 * 10 ^ 4 + 5 * 10 ^ 3 + 4 * 10 ^ 2 + 3 * 10 ^ 1 + 6 * 10 ^ 0
Which is the decimal expansion of the number. So, the minimum amount of information we need to represent this number is 6 digits. This is no coincidence, as any number less than 10^d can be represented in d digits.
So how many digits are required to represent X? Thats equal to the largest exponent of 10 in X plus 1.
==> 10 ^ d > X
==> log (10 ^ d) > log(X)
==> d* log(10) > log(X)
==> d > log(X) // And log appears again...
==> d = floor(log(x)) + 1
Also note that this is the most concise way to denote the number in this range. Any reduction will lead to information loss, as a missing digit can be mapped to 10 other numbers. For example: 12* can be mapped to 120, 121, 122, …, 129.
2. How do you search for a number in (0, N - 1)?
Taking N = 10^d, we use our most important observation:
The minimum amount of information to uniquely identify a value in a range between 0 to N - 1 = log(N) digits.
This implies that, when asked to search for a number on the integer line, ranging from 0 to N - 1, we need at least log(N) tries to find it. Why? Any search algorithm will need to choose one digit after another in its search for the number.
The minimum number of digits it needs to choose is log(N). Hence the minimum number of operations taken to search for a number in a space of size N is log(N).
Can you guess the order complexities of binary search, ternary search or deca search? Its O(log(N))!
3. How do you sort a set of numbers?
When asked to sort a set of numbers A into an array B, here’s what it looks like ->
Permute Elements
Every element in the original array has to be mapped to it’s corresponding index in the sorted array. So, for the first element, we have n positions. To correctly find the corresponding index in this range from 0 to n - 1, we need…log(n) operations.
The next element needs log(n-1) operations, the next log(n-2) and so on. The total comes to be:
==> log(n) + log(n - 1) + log(n - 2) + … + log(1)Using log(a) + log(b) = log(a * b), ==> log(n!)
This can be approximated to nlog(n) - n. Which is O(n*log(n))!
Hence we conclude that there can be no sorting algorithm that can do better than O(n*log(n)). And some algorithms having this complexity are the popular Merge Sort and Heap Sort!
These are some of the reasons why we see log(n) pop up so often in the complexity analysis of algorithms. The same can be extended to binary numbers. I made a video on that here.
Why does log(n) appear so often during algorithm complexity analysis?
Cheers!
We call the time complexity O(log n), when the solution is based on iterations over n, where the work done in each iteration is a fraction of the previous iteration, as the algorithm works towards the solution.
Can't comment yet... necro it is!
Avi Cohen's answer is incorrect, try:
X = 1 3 4 5 8
Y = 2 5 6 7 9
None of the conditions are true, so MEDIAN(X, p, q, Y, j, k) will cut both the fives. These are nondecreasing sequences, not all values are distinct.
Also try this even-length example with distinct values:
X = 1 3 4 7
Y = 2 5 6 8
Now MEDIAN(X, p, q, Y, j+1, k) will cut the four.
Instead I offer this algorithm, call it with MEDIAN(1,n,1,n):
MEDIAN(startx, endx, starty, endy){
if (startx == endx)
return min(X[startx], y[starty])
odd = (startx + endx) % 2 //0 if even, 1 if odd
m = (startx+endx - odd)/2
n = (starty+endy - odd)/2
x = X[m]
y = Y[n]
if x == y
//then there are n-2{+1} total elements smaller than or equal to both x and y
//so this value is the nth smallest
//we have found the median.
return x
if (x < y)
//if we remove some numbers smaller then the median,
//and remove the same amount of numbers bigger than the median,
//the median will not change
//we know the elements before x are smaller than the median,
//and the elements after y are bigger than the median,
//so we discard these and continue the search:
return MEDIAN(m, endx, starty, n + 1 - odd)
else (x > y)
return MEDIAN(startx, m + 1 - odd, n, endy)
}