私はこの問題が大好きです。これは典型的なインタビューの質問であり、あなたの考え次第で、あなたはどんどん良い解決策を手に入れることになります。これをO(n 2)時間よりも優れた方法で実行することは確かに可能です。ここで問題について考えることができる3つの異なる方法を挙げました。うまくいけば、これはあなたの質問に答えます!
まず、分割統治ソリューション。入力を半分に分割し、各サブアレイの問題を解決してから、2つを組み合わせることでこれを解決できるかどうかを見てみましょう。実際にこれを実行でき、効率的に実行できることがわかりました。直感は以下の通りです。1日の場合、最良のオプションは、その日に購入してから、同じ日に売り戻して利益を得ないことです。それ以外の場合は、配列を2つに分割します。最適な答えが何であるかを考える場合、それは次の3つの場所のいずれかにある必要があります。
- 正しい売買ペアは完全に前半で発生します。
- 正しい売買ペアは完全に後半に発生します。
- 正しい売買ペアは両方の半分で発生します-前半で購入し、後半で販売します。
前半と後半でアルゴリズムを再帰的に呼び出すことで、(1)と(2)の値を取得できます。オプション(3)の場合、最高の利益を上げる方法は、前半の最低ポイントで購入し、後半の最高ポイントで販売することです。入力に対して単純な線形スキャンを実行して2つの値を見つけるだけで、2つの半分の最小値と最大値を見つけることができます。これにより、次の繰り返しのアルゴリズムが得られます。
T(1) <= O(1)
T(n) <= 2T(n / 2) + O(n)
マスター定理を使用して再帰を解決すると、これがO(n lg n)時間で実行され、再帰呼び出しにO(lg n)空間を使用することがわかります。単純なO(n 2)解法を打ち負かしました!
ちょっと待って!私たちはこれよりもずっと良いことができます。繰り返しにO(n)項がある唯一の理由は、各半分の最小値と最大値を見つけるために入力全体をスキャンする必要があることです。すでに各半分を再帰的に探索しているので、再帰によって各半分に格納されている最小値と最大値も返させることで、より良い結果が得られる可能性があります。言い換えると、私たちの再帰は次の3つを返します。
- 利益を最大化するための売買時間。
- 範囲全体での最小値。
- 範囲全体の最大値。
これらの最後の2つの値は、計算する再帰(1)と同時に実行できる単純な再帰を使用して再帰的に計算できます。
- 単一要素の範囲の最大値と最小値は、その要素にすぎません。
- 複数の要素の範囲の最大値と最小値は、入力を半分に分割し、各半分の最大値と最小値を見つけて、それぞれの最大値と最小値を取得することで確認できます。
このアプローチを使用すると、反復関係は次のようになります。
T(1) <= O(1)
T(n) <= 2T(n / 2) + O(1)
ここでマスター定理を使用すると、O(lg n)スペースでO(n)のランタイムが得られます。これは、元のソリューションよりもさらに優れています!
しかし、少し待ってください-私たちはこれよりもさらに良いことができます!動的プログラミングを使用してこの問題を解決することを考えてみましょう。問題は次のように考えることです。最初のk個の要素を調べたところ、問題の答えがわかっていると仮定します。最初の(k + 1)要素の問題を解決するために、最初のソリューションと組み合わせて(k + 1)番目の要素に関する知識を使用できますか?もしそうなら、最初のn個の要素について計算するまで、最初の要素、次に最初の2つ、次に最初の3つなどの問題を解決することにより、優れたアルゴリズムを実行できます。
これを行う方法について考えてみましょう。要素が1つしかない場合は、それが最良の売買ペアでなければならないことはすでにわかっています。ここで、最初のk要素の最良の答えがわかっていて、(k + 1)番目の要素を見てみます。次に、この値が最初のk要素よりも優れたソリューションを作成できる唯一の方法は、最初のk要素の最小値とその新しい要素の差が、これまでに計算した最大の差より大きい場合です。要素を移動するときに、これまでに見た最小値と、最初のk要素だけで得られる最大利益の2つの値を追跡するとします。最初は、これまでに見た最小値が最初の要素であり、最大利益はゼロです。新しい要素を見ると、まず、これまでに見た最低価格で購入し、現在の価格で販売することでどれだけ稼ぐかを計算して、最適利益を更新します。これがこれまでに計算した最適値よりも優れている場合は、この新しい利益になるように最適解を更新します。次に、これまでに確認した最小要素を、現在の最小要素と新しい要素の最小値に更新します。
各ステップで実行するのはO(1)の作業のみであり、n個の要素のそれぞれに正確に1回アクセスするだけなので、完了までにO(n)時間かかります。さらに、O(1)補助記憶域のみを使用します。これはこれまでに得たものと同じです!
例として、入力での、このアルゴリズムの実行方法を次に示します。配列の各値の間の数値は、その時点でアルゴリズムが保持する値に対応しています。これらすべてを実際に保存することはありませんが(O(n)メモリーが必要です!)、アルゴリズムが進化するのを確認すると役立ちます。
5 10 4 6 7
min 5 5 4 4 4
best (5,5) (5,10) (5,10) (5,10) (5,10)
回答:(5、10)
5 10 4 6 12
min 5 5 4 4 4
best (5,5) (5,10) (5,10) (5,10) (4,12)
回答:(4、12)
1 2 3 4 5
min 1 1 1 1 1
best (1,1) (1,2) (1,3) (1,4) (1,5)
正解:(1、5)
今はもっと上手くできる?残念ながら、漸近的な意味ではありません。O(n)未満の時間を使用すると、大きな入力のすべての数値を調べることができないため、最適な回答を見逃さないことを保証できません(要素でそれを「隠す」ことができます)見ていない)。さらに、O(1)未満のスペースは使用できません。big-O表記に隠された定数要素に対するいくつかの最適化があるかもしれませんが、それ以外の場合、根本的に優れたオプションを見つけることは期待できません。
全体として、これは次のアルゴリズムがあることを意味します。
- 素朴:O(n 2)時間、O(1)空間。
- 分割統治:O(n lg n)時間、O(lg n)空間。
- 最適化された分割統治:O(n)時間、O(lg n)空間。
- 動的計画法:O(n)時間、O(1)空間。
お役に立てれば!
編集:興味があれば、これら4つのアルゴリズムのPythonバージョンをコード化して、それらをいじって、相対的なパフォーマンスを判断できるようにしました。これがコードです:
# Four different algorithms for solving the maximum single-sell profit problem,
# each of which have different time and space complexity. This is one of my
# all-time favorite algorithms questions, since there are so many different
# answers that you can arrive at by thinking about the problem in slightly
# different ways.
#
# The maximum single-sell profit problem is defined as follows. You are given
# an array of stock prices representing the value of some stock over time.
# Assuming that you are allowed to buy the stock exactly once and sell the
# stock exactly once, what is the maximum profit you can make? For example,
# given the prices
#
# 2, 7, 1, 8, 2, 8, 4, 5, 9, 0, 4, 5
#
# The maximum profit you can make is 8, by buying when the stock price is 1 and
# selling when the stock price is 9. Note that while the greatest difference
# in the array is 9 (by subtracting 9 - 0), we cannot actually make a profit of
# 9 here because the stock price of 0 comes after the stock price of 9 (though
# if we wanted to lose a lot of money, buying high and selling low would be a
# great idea!)
#
# In the event that there's no profit to be made at all, we can always buy and
# sell on the same date. For example, given these prices (which might
# represent a buggy-whip manufacturer:)
#
# 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
#
# The best profit we can make is 0 by buying and selling on the same day.
#
# Let's begin by writing the simplest and easiest algorithm we know of that
# can solve this problem - brute force. We will just consider all O(n^2) pairs
# of values, and then pick the one with the highest net profit. There are
# exactly n + (n - 1) + (n - 2) + ... + 1 = n(n + 1)/2 different pairs to pick
# from, so this algorithm will grow quadratically in the worst-case. However,
# it uses only O(1) memory, which is a somewhat attractive feature. Plus, if
# our first intuition for the problem gives a quadratic solution, we can be
# satisfied that if we don't come up with anything else, we can always have a
# polynomial-time solution.
def BruteForceSingleSellProfit(arr):
# Store the best possible profit we can make; initially this is 0.
bestProfit = 0;
# Iterate across all pairs and find the best out of all of them. As a
# minor optimization, we don't consider any pair consisting of a single
# element twice, since we already know that we get profit 0 from this.
for i in range(0, len(arr)):
for j in range (i + 1, len(arr)):
bestProfit = max(bestProfit, arr[j] - arr[i])
return bestProfit
# This solution is extremely inelegant, and it seems like there just *has* to
# be a better solution. In fact, there are many better solutions, and we'll
# see three of them.
#
# The first insight comes if we try to solve this problem by using a divide-
# and-conquer strategy. Let's consider what happens if we split the array into
# two (roughly equal) halves. If we do so, then there are three possible
# options about where the best buy and sell times are:
#
# 1. We should buy and sell purely in the left half of the array.
# 2. We should buy and sell purely in the right half of the array.
# 3. We should buy in the left half of the array and sell in the right half of
# the array.
#
# (Note that we don't need to consider selling in the left half of the array
# and buying in the right half of the array, since the buy time must always
# come before the sell time)
#
# If we want to solve this problem recursively, then we can get values for (1)
# and (2) by recursively invoking the algorithm on the left and right
# subarrays. But what about (3)? Well, if we want to maximize our profit, we
# should be buying at the lowest possible cost in the left half of the array
# and selling at the highest possible cost in the right half of the array.
# This gives a very elegant algorithm for solving this problem:
#
# If the array has size 0 or size 1, the maximum profit is 0.
# Otherwise:
# Split the array in half.
# Compute the maximum single-sell profit in the left array, call it L.
# Compute the maximum single-sell profit in the right array, call it R.
# Find the minimum of the first half of the array, call it Min
# Find the maximum of the second half of the array, call it Max
# Return the maximum of L, R, and Max - Min.
#
# Let's consider the time and space complexity of this algorithm. Our base
# case takes O(1) time, and in our recursive step we make two recursive calls,
# one on each half of the array, and then does O(n) work to scan the array
# elements to find the minimum and maximum values. This gives the recurrence
#
# T(1) = O(1)
# T(n / 2) = 2T(n / 2) + O(n)
#
# Using the Master Theorem, this recurrence solves to O(n log n), which is
# asymptotically faster than our original approach! However, we do pay a
# (slight) cost in memory usage. Because we need to maintain space for all of
# the stack frames we use. Since on each recursive call we cut the array size
# in half, the maximum number of recursive calls we can make is O(log n), so
# this algorithm uses O(n log n) time and O(log n) memory.
def DivideAndConquerSingleSellProfit(arr):
# Base case: If the array has zero or one elements in it, the maximum
# profit is 0.
if len(arr) <= 1:
return 0;
# Cut the array into two roughly equal pieces.
left = arr[ : len(arr) / 2]
right = arr[len(arr) / 2 : ]
# Find the values for buying and selling purely in the left or purely in
# the right.
leftBest = DivideAndConquerSingleSellProfit(left)
rightBest = DivideAndConquerSingleSellProfit(right)
# Compute the best profit for buying in the left and selling in the right.
crossBest = max(right) - min(left)
# Return the best of the three
return max(leftBest, rightBest, crossBest)
# While the above algorithm for computing the maximum single-sell profit is
# better timewise than what we started with (O(n log n) versus O(n^2)), we can
# still improve the time performance. In particular, recall our recurrence
# relation:
#
# T(1) = O(1)
# T(n) = 2T(n / 2) + O(n)
#
# Here, the O(n) term in the T(n) case comes from the work being done to find
# the maximum and minimum values in the right and left halves of the array,
# respectively. If we could find these values faster than what we're doing
# right now, we could potentially decrease the function's runtime.
#
# The key observation here is that we can compute the minimum and maximum
# values of an array using a divide-and-conquer approach. Specifically:
#
# If the array has just one element, it is the minimum and maximum value.
# Otherwise:
# Split the array in half.
# Find the minimum and maximum values from the left and right halves.
# Return the minimum and maximum of these two values.
#
# Notice that our base case does only O(1) work, and our recursive case manages
# to do only O(1) work in addition to the recursive calls. This gives us the
# recurrence relation
#
# T(1) = O(1)
# T(n) = 2T(n / 2) + O(1)
#
# Using the Master Theorem, this solves to O(n).
#
# How can we make use of this result? Well, in our current divide-and-conquer
# solution, we split the array in half anyway to find the maximum profit we
# could make in the left and right subarrays. Could we have those recursive
# calls also hand back the maximum and minimum values of the respective arrays?
# If so, we could rewrite our solution as follows:
#
# If the array has size 1, the maximum profit is zero and the maximum and
# minimum values are the single array element.
# Otherwise:
# Split the array in half.
# Compute the maximum single-sell profit in the left array, call it L.
# Compute the maximum single-sell profit in the right array, call it R.
# Let Min be the minimum value in the left array, which we got from our
# first recursive call.
# Let Max be the maximum value in the right array, which we got from our
# second recursive call.
# Return the maximum of L, R, and Max - Min for the maximum single-sell
# profit, and the appropriate maximum and minimum values found from
# the recursive calls.
#
# The correctness proof for this algorithm works just as it did before, but now
# we never actually do a scan of the array at each step. In fact, we do only
# O(1) work at each level. This gives a new recurrence
#
# T(1) = O(1)
# T(n) = 2T(n / 2) + O(1)
#
# Which solves to O(n). We're now using O(n) time and O(log n) memory, which
# is asymptotically faster than before!
#
# The code for this is given below:
def OptimizedDivideAndConquerSingleSellProfit(arr):
# If the array is empty, the maximum profit is zero.
if len(arr) == 0:
return 0
# This recursive helper function implements the above recurrence. It
# returns a triple of (max profit, min array value, max array value). For
# efficiency reasons, we always reuse the array and specify the bounds as
# [lhs, rhs]
def Recursion(arr, lhs, rhs):
# If the array has just one element, we return that the profit is zero
# but the minimum and maximum values are just that array value.
if lhs == rhs:
return (0, arr[lhs], arr[rhs])
# Recursively compute the values for the first and latter half of the
# array. To do this, we need to split the array in half. The line
# below accomplishes this in a way that, if ported to other languages,
# cannot result in an integer overflow.
mid = lhs + (rhs - lhs) / 2
# Perform the recursion.
( leftProfit, leftMin, leftMax) = Recursion(arr, lhs, mid)
(rightProfit, rightMin, rightMax) = Recursion(arr, mid + 1, rhs)
# Our result is the maximum possible profit, the minimum of the two
# minima we've found (since the minimum of these two values gives the
# minimum of the overall array), and the maximum of the two maxima.
maxProfit = max(leftProfit, rightProfit, rightMax - leftMin)
return (maxProfit, min(leftMin, rightMin), max(leftMax, rightMax))
# Using our recursive helper function, compute the resulting value.
profit, _, _ = Recursion(arr, 0, len(arr) - 1)
return profit
# At this point we've traded our O(n^2)-time, O(1)-space solution for an O(n)-
# time, O(log n) space solution. But can we do better than this?
#
# To find a better algorithm, we'll need to switch our line of reasoning.
# Rather than using divide-and-conquer, let's see what happens if we use
# dynamic programming. In particular, let's think about the following problem.
# If we knew the maximum single-sell profit that we could get in just the first
# k array elements, could we use this information to determine what the
# maximum single-sell profit would be in the first k + 1 array elements? If we
# could do this, we could use the following algorithm:
#
# Find the maximum single-sell profit to be made in the first 1 elements.
# For i = 2 to n:
# Compute the maximum single-sell profit using the first i elements.
#
# How might we do this? One intuition is as follows. Suppose that we know the
# maximum single-sell profit of the first k elements. If we look at k + 1
# elements, then either the maximum profit we could make by buying and selling
# within the first k elements (in which case nothing changes), or we're
# supposed to sell at the (k + 1)st price. If we wanted to sell at this price
# for a maximum profit, then we would want to do so by buying at the lowest of
# the first k + 1 prices, then selling at the (k + 1)st price.
#
# To accomplish this, suppose that we keep track of the minimum value in the
# first k elements, along with the maximum profit we could make in the first
# k elements. Upon seeing the (k + 1)st element, we update what the current
# minimum value is, then update what the maximum profit we can make is by
# seeing whether the difference between the (k + 1)st element and the new
# minimum value is. Note that it doesn't matter what order we do this in; if
# the (k + 1)st element is the smallest element so far, there's no possible way
# that we could increase our profit by selling at that point.
#
# To finish up this algorithm, we should note that given just the first price,
# the maximum possible profit is 0.
#
# This gives the following simple and elegant algorithm for the maximum single-
# sell profit problem:
#
# Let profit = 0.
# Let min = arr[0]
# For k = 1 to length(arr):
# If arr[k] < min, set min = arr[k]
# If profit < arr[k] - min, set profit = arr[k] - min
#
# This is short, sweet, and uses only O(n) time and O(1) memory. The beauty of
# this solution is that we are quite naturally led there by thinking about how
# to update our answer to the problem in response to seeing some new element.
# In fact, we could consider implementing this algorithm as a streaming
# algorithm, where at each point in time we maintain the maximum possible
# profit and then update our answer every time new data becomes available.
#
# The final version of this algorithm is shown here:
def DynamicProgrammingSingleSellProfit(arr):
# If the array is empty, we cannot make a profit.
if len(arr) == 0:
return 0
# Otherwise, keep track of the best possible profit and the lowest value
# seen so far.
profit = 0
cheapest = arr[0]
# Iterate across the array, updating our answer as we go according to the
# above pseudocode.
for i in range(1, len(arr)):
# Update the minimum value to be the lower of the existing minimum and
# the new minimum.
cheapest = min(cheapest, arr[i])
# Update the maximum profit to be the larger of the old profit and the
# profit made by buying at the lowest value and selling at the current
# price.
profit = max(profit, arr[i] - cheapest)
return profit
# To summarize our algorithms, we have seen
#
# Naive: O(n ^ 2) time, O(1) space
# Divide-and-conquer: O(n log n) time, O(log n) space
# Optimized divide-and-conquer: O(n) time, O(log n) space
# Dynamic programming: O(n) time, O(1) space