Skip to main content Link Search Menu Expand Document (external link)

Identifying the maximum in a sliding window

Given an array $x$ and an integer $k$, determine for every index $1 \leq j\leq n$ the maximum $x[i]$ among all indices $\max\{1,j-k+1 \} \leq i \leq j$.

Use a double ended queue

The high level idea is to maintain a window over $x$, ranging from index $j-k+1$ to index $j$. We say that in this range, index $b$ dominates index $a$ if $a < b$ and $x[a] \leq x[b]$.

For an intuitive explanation, imagine that on a coastline, there are $n$ building standing in a row. Each building has the same width, which we call a block. Building $i$ (numbered from left to right) consists of $x[i]$ stories. You drive a car on the coast line left to right and look in your rear mirror. Due to the fog, you can’t see buildings that are more than $k$ blocks far. In addition, any building in some block $b$ will hide all smaller or equal height buildings in blocks $a$ to the left of $b$. This is meant by index $b$ dominating index $a$.

Instead of maintaining the actual window, we maintain in a queue $S$ all non-dominated indices in the window. The queue contains indices in increasing order, and their corresponding $x$-values decrease in this order. Hence $S[0]$ is the index of the largest value in the window, and $S[-1]$ is the index of the last value which entered the window.

When sliding the window $[i,j]$, $x[j]$ enters the window, while $x[i]$ leaves the windows. If $x[i]$ was the maximum in the window, i.e. $S[0]=i$, then $i$ has to be removed from $S$. And before $j$ enters $S$, we need to remove all indices dominated by $j$.

We use a double ended queue, because we need to remove dominated elements from the right, insert new elements to the right, and remove the maximum element from the left.

Note that when $j$ ranges over the first $k$ indices, no element leaves $S$ yet, since the window does not have size $k$ yet.

For the time complexity, note that every element might enter and leave $S$ at most once, hence the work done on every index is constant in amortized time, and the overall complexity is $O(n)$, where $n$ is the size of $x$.

from collections import deque

def latest_max(x, k):
    """returns a table y of the same size as x, such that
    y[j] = max(x[max(0, j-k+1):j+1]).
    complexity: O(len(x))
    """
    n = len(x)
    y = [] 
    i = 0 # invariant: i = max(0, j - k + 1)
    S = deque() # list of undominated elements in window
    for j in range(n):
        while S and x[S[-1]] <= x[j]:
            S.pop()
        S.append(j)
        if i == j - k:
            if S[0] == i:
                S.popleft()
            i += 1
        y.append(x[S[0]])
    return y  

Comments