# [C++ / Fast / Very clear explanation / Clean Code] Solution with Greedy Algorithm and Binary Search

• First thing first, below is the code:

``````class Solution {
private:
bool doable (const vector<int>& nums, int cuts, long long max) {
int acc = 0;
for (num : nums) {
// This step is unnecessary for this problem. I didn't discard this line because I want doable function more generalized.
if (num > max) return false;
else if (acc + num <= max) acc += num;
else {
--cuts;
acc = num;
if (cuts < 0) return false;
}
}
return true;
}

public:
int splitArray(vector<int>& nums, int m) {
long long left = 0, right = 0;
for (num : nums) {
left = max(left, (long long)num);
right += num;
}

while (left < right) {
long long mid = left + (right - left) / 2;
if (doable(nums, m - 1, mid)) right = mid;
else left = mid + 1;
}
return left;
}
};
``````

## Introduction to this problem:

We can break this problem into two smaller problems:

• Given an array (A), number of cuts (CUTS), and the Largest sum of sub-arrays (MAX). Can you use at most CUTS cuts to segment array A into CUTS + 1 sub-arrays, such that the sum of each sub-array is smaller or equal to MAX?
• Given a lower bound (left), an upper bound (right), an unknown bool array (B), and an API uses i as input and tells you whether B[i] is true. If we know there exists an index k, that B[i] is false when i < k, and B[i] is true when i >= k. What is the fastest way to find this k (the lower bound)?

## Solution to the first sub-problem (Skip this part if you already knew how to solve 1st sub-problem):

For the first question, we can follow these steps:

• For each element in the array, if its value is larger than MAX, we know it's not possible to cut this array into groups that the sum of all groups are smaller than MAX. (Reason is straightforward, if A is [10, 2, 3, 5] and MAX is 6, even you have 3 cuts by which you can cut A as [[10], [2], [3], [5]], the group containing 10 will still be larger than 6).
• Use greedy algorithm to cut A. Use an accumulator ACC to store the sum of the currently processed group, and process elements in A one by one. For each element num, if we add num with ACC and the new sum is still no larger than MAX, we update ACC to ACC + num, which means we can merge num into the current group. If not, we must use a cut before num to segment this array, then num will be the first element in the new group.
• If we didn't go through A but already used up all cuts, then it's not possible only using CUTS cuts to segment this array into groups to make sure sum of each sub-array is smaller than MAX. Otherwise, if we can reach the end of A with cuts left (or use exactly CUTS cuts). It's possible to do so.

Then the first question is solved.

## Solution to the second sub-problem(Skip this part if you already knew how to solve 2nd sub-problem):

• The array B will be something like [false, false, ..., false, true, true, ..., true]. We want to find the index of the first true.
• Use binary search to find this k. Keep a value mid, mid = (left + right) / 2. If B[mid] = false, then move the search range to the upper half of the original search range, a.k.a left = mid + 1, otherwise move search range to the lower half, a.k.a right = mid.

## Why this algorithm is correct...

• No matter how we cut the array A, the Largest sum of sub-arrays will fall into a range [left, right]. Left is the value of the largest element in this array. right is the sum of this array. (e.g., Given array [1, 2, 3, 4, 5], if we have 4 cuts and cut it as [[1], [2], [3], [4], [5]], the Largest sum of sub-arrays is 5, it cannot be smaller. And if we have 0 cut, and the only sub-array is [[1, 2, 3, 4, 5]], the Largest sum of sub-arrays is 15, it cannot be larger).
• However, we cannot decide the number of cuts (CUTS), this is an given constraint. But we know there must be a magic number k, which is the smallest value of the Largest sum of sub-arrays when given CUTS cuts. When the Largest sum of sub-arrays is larger than k, we can always find a way to cut A within CUTS cuts. When the Largest sum of sub-arrays is smaller than k, there is no way to do this.

## Example

For example, given array A [1, 2, 3, 4, 5]. We can use 2 cuts.

• No matter how many cuts are allowed, the range of the possible value of the Largest sum of sub-arrays is [5, 15].
• When given 2 cuts, we can tell the magic number k here is 6, the result of segmentation is [[1, 2, 3], [4], [5]].
• When Largest sum of sub-arrays is in range [6, 15], we can always find a way to cut this array within two cuts. You can have a try.
• However, when Largest sum of sub-arrays is in range [5, 5], there is no way to do this.
• This mapped this problem into the second sub-problem. Bool array B here is [5:false, 6:true, 7:true, 8:true, ..., 15:true]. We want to find the index i of the first true in B, which is the answer of this entire question, and by solving the first sub-problem, we have an API that can tell us given an i (Largest sum of sub-arrays), whether B[i] is true (whether we can find a way to cut A to satisfy the constraint).

Below is the code with comment, just in case you don't have time to read the explanations above.

``````class Solution {
private:
/*
Params:
nums - The input array;
cuts - How many cuts are available (cuts = #groups - 1);
max - The maximum of the (sum of elements in one group);
Rtn:
Whether we can use at most 'cuts' number of cuts to segment the entire array,
such that the sum of each group will not exceed 'max'.
*/
bool doable (const vector<int>& nums, int cuts, long long max) {

// 'acc' is the temporary accumulator for the currently processed group.

int acc = 0;
for (num : nums) {

// If the current processed element in this array is larger than 'max', we cannot segment the array.
// (Reason is straightforward, if 'nums' is [10, 2, 3, 5] and 'max' is 6, even you can have 3 cuts
// (by which you can cut array as [[10], [2], [3], [5]]), the group containing 10 will be larger than 6,
//  there is no way to do this).
// Ps: This step is unnecessary in this solution. Because 'left' in the splitArray() function can assure
// 'max' will be larger than every single element. I just want to write a generalized doable() function :)

if (num > max) return false;

// If the (sum of the currently processed group) + (current element) is smaller than max, we can add current
// element into this group.

else if (acc + num <= max) acc += num;

// If not, we will make a cut before this element, and this element will be the first element in the new group.

else {
--cuts;
acc = num;

// If we've used up all cuts, this means this 'max' is not doable.
if (cuts < 0) return false;
}
}

// If we can reach here, this means we've used at most 'cuts' cut to segment the array, and the sum of each groups is
// not larger than 'max'. Yeah!
return true;
}

public:
int splitArray(vector<int>& nums, int m) {
// Use long long to avoid overflow.
long long left = 0, right = 0;
// The smallest possible value ('left') is the the value of the largest element in this array.
// The largest possible value ('right') is the sum of all elements in this array.
for (num : nums) {
left = max(left, (long long)num);
right += num;
}

// Use binary search, find the lower bound of the possible (minimum sum of groups within m - 1 cuts).
while (left < right) {
long long mid = left + (right - left) / 2;
if (doable(nums, m - 1, mid)) right = mid;
else left = mid + 1;
}
return left;
}
};
``````

• Very good post, amazing algo

• Nice! Good explanation!

• Very instructive! Thank you so much!

• I think this post is the best explanation so far for this question and it deserves more upvotes.

Here is the java implementation derived from the brilliant idea:

``````/**
* The basic idea is based on [minMaxSumValue, maxMaxSumValue], and min max sum value of subarrays is the largest
* num in the array, and max max sum value of subarrays is the total sum of all the elements.
*
* We use binary search to find the min Max sum value of the subarrays
*/
public static int splitArray(int[] nums, int m) {
// validation
if (nums == null || nums.length < 1 || m < 1) {
return -1;
}

// get [minMaxSum, maxMaxSum]
long start = 0, end = 0;
for (int n : nums) {
start = Math.min(start, n);
end += n;
}

// binary search
while (start < end) {
long middle = start + (end - start) / 2;
if (validateMaxSum(nums, middle, m - 1)) {
// valid maxSum & cut so we need to shrink the finding (reduce max sum)
end = middle;
} else {
// invalid maxSum & cut so we need to expand the finding (increase max sum)
start = middle + 1;
}
}
return new Long(start).intValue();
}

/**
* Given a max sum, we are validating the cuts. Whenever we collect/aggregate enough sum (<= maxSum), we start a new
* group
*/
private static boolean validateMaxSum(int[] nums, long maxSum, int cuts) {
int agg = 0;
for (int n : nums) {
// validate value (not needed)
if (n > maxSum || cuts < 0) {
return false;
}

agg += n;

if (agg > maxSum) {
// find a group
agg = n;
cuts--;
if (cuts < 0) return false;
}
}
// cuts may still > 1, which means we can even have less num of cuts for grouping
return true;
}
``````

• @LoseAGirlDoesNotMeanLoseASoul

Isn't start suppose to be the largest number in the array. In your code, it's always 0. I think it's suppose to be Math.max.

• I can't seem to wrap my head around the fact that when you are doing the binary search and shrinking the range down, how can you guarantee that the mid value will be a valid sum within the array.

long long mid = left + (right - left) / 2;

The divide by 2 will at times give you a sum that isn't possible .

Ex:
[7,2,5,10,8]
2

doable (1,21) // How would you get this sum [7,2,5,10,8]
doable (1,15) // 5 + 10
doable (1,18) //10+8
doable (1,17) // 2+5+10

In this case, we end up with 18 which is a valid sum. However, is this always true that you will end up with a valid sum of a continuous subarrays.?

• @dare_wreck54-yahoo-com

Hi Dare Wreck, thanks for your question.

mid value is not about can we divide the array such that the max sum of subarray equals mid, it's about can we divide the array such that the max sum of subarray less than or equals to mid. So we cannot guarantee mid is a valid sum.

This algorithm is based on the loop invariant: The magic number k is always bounded by left and right search ranges. Which is left <= k <= right. So when we reach the point left == right, we can safely say left (or right) is the magic number k we want to find.

For your question, this algorithm didn't say the mid values are always valid sums. Actually it's not, it's just a check point to narrow down the search range. Neither left nor right is alway a valid sum, too.

But we do know, the magic k must lie between left and right, and this number must be a valid sum. Because for every number smaller than k, that solution is not doable. Every number larger than k is doable, and we know k is a valid sum that we can achieve the minimum max sum.

Hope this post can be helpful. :)

• Excellent explanation. It's hard for me to understand the solution comments until yours. Thx

• It's a good one.
Thanks for the explanation.
There might be a little confused with

``````if (doable(nums, m - 1, mid)) right = mid;
``````

would it be better if it's:

``````if (doable(nums, m , mid)) right = mid;
``````

and also change:

``````if (cuts < 0) return false;
``````

to

``````if (cuts <= 0) return false;
``````

Because I don't really understand the logic of checking if it can be done with `m-1` cuts while we can just check if we can use `m` cuts.

And I tried `m` and it works.

• This is a really nice explanation. I have one quick question: the `doable` function seems not to enforce that we have to use exactly m-1 cuts. But I guess that the upper bound `max` will push us to have exactly m-1 cuts in the end, because the more groups we have, the smaller largest subarray sum we can have. Is this right? Thanks.

• Can you explain why the greedy method works for the 1st sub problem?

For example, you try to divide with 4 cuts, and not exceed MAX = 10.
While you have successfully divide with 3 cuts, [sum < 10], [sum < 10], [sum < 10] , a+b+c > 10.

So, can we proof that you can not replace a, b or c with a smaller number in previous segments, while keep [sum-d+c < 10], [sum < 10], [sum < 10], a+b+d < 10? (d < c)

• This post is deleted!

• Nice solution! How can you come up with this??!
Binary search is really an effective algorithm

Looks like your connection to LeetCode Discuss was lost, please wait while we try to reconnect.