Fenwick tree batch updating - algorithm

Fenwick tree has an update(idx, delta) function, which may be implemented like this:
for ( ; idx < fTree.length; idx += idx & -idx ){
this._fTree[ idx ] += delta;
}
So you are updating your current idx and then pushing your change till the end.
I want to perform several update calls in a loop, like
update( 12, 100 );
update( 13, 400 );
update( 17, 200 );
If I know the range of updated indexes in advance ( 12 - 17 here ), update may be optimized. Rather than pushing till the end everytime, we can push till certain index, buffer rest change, and then, after last update, push buffered change till the end. Providing a picture to explain concept:
For example taking range 3 - 7.
updating of index 3 will lead to update of following indexes: 3 - 4 - 8 - 16.
Updating of index 4 will lead to: 4 - 8 - 16
Updating of index 5 will lead to: 5 - 6 - 8 - 16
It would be optimal to push everything till 8, buffer rest, and then push it till 16.
Question
Is there better way to find this node to push until ( 8 here ), knowing range start and end?
Non-ideal solution
const getLimitIndex = ( from, to ) => {
let lim = to;
for( let tmp = lim & -lim; lim - tmp > from; ){
lim += tmp;
tmp = lim & -lim;
}
return lim;
}

Related

Rotate Image - Java

I am doing a question where, given an n x n 2D matrix representing an image, rotate the image by 90 degrees (clockwise).You have to rotate the image in-place, which means you have to modify the input 2D matrix directly. DO NOT allocate another 2D matrix and do the rotation. This my my code:
class Solution {
public void rotate(int[][] matrix) {
int size = matrix.length;
for(int i = 0 ; i < matrix.length; i++){
for(int y = 0 ; y < matrix[0].length ; y++){
matrix[i][y] = matrix[size - y - 1][i];
System.out.println(size - y - 1);
System.out.println(i);
System.out.println("");
}
}
}
}
This is the input and output results:
input matrix: [[1,2,3],[4,5,6],[7,8,9]]
output matrix: [[7,4,7],[8,5,4],[9,4,7]]
expected matrix: [[7,4,1],[8,5,2],[9,6,3]]
I do not really understand why I am getting duplicates in my output such as the number seven 3 times. On my System.out.println statement, I am getting the correct list of indexes :
2
0
1
0
0
0
2
1
1
1
0
1
2
2
What can be wrong?
I have found a solution. I will try my best to explain it.
Let us consider an array of size 4.
1 2 3 4
5 6 7 8
9 10 11 12
13 14 15 16
Now lets look at the numbers present only on the outside of the array:
1 2 3 4
5 8
9 12
13 14 15 16
We will proceed by storing the first element 1 in a temporary variable. Next we will replace 1 by 13, 13 by 16, 16 by 4 and at last 4 by 1 (whose value we already stored in the temporary variable).
We will do the same for all the elements of the first row.
Here is a pseudocode if you just want to rotate this outer ring, lets call it an outer ring:
for i = 0 to n-1
{
temp = A[0][i];
A[0][i] = A[n-1-i][0];
A[n-1-i][0] = A[n-1-0][n-1-i];
A[n-1-0][n-1-i] = A[i][n-1-0];
A[i][n-1-0] = temp;
}
The code runs for a total of n times. Once for each element of first row. Implement this code an run it. You will see only the outer ring is rotated. Now lets look at the inner ring:
6 7
10 11
Now the loop in pseudocode only needs to run for 2 times and also our range of indexes has decreased. For outer ring, the loop started from i = 0 and ended at i = n-1. However, for the inner ring the for loop need to run from i = 1 to i = n-2.
If you had an array of size n, to rotate the xth ring of the array, the loop needs to run from i = x to i = n-1-x.
Here is the code to rotate the entire array:
x = 0;
int temp;
while (x < n/2)
{
for (int i = x;i < n-1-x;i++)
{
temp = arr[x][i];
arr[x][i] = arr[n-1-i][x];
arr[n-1-i][x] = arr[n-1-x][n-1-i];
arr[n-1-x][n-1-i] = arr[i][n-1-x];
arr[i][n-1-x] = temp;
}
x++;
}
Here each value of x denotes the xth ring.
0 <= x <= n-1
The reason why the outer loop runs only for x < n/2 times is because each array has n/2 rings when n is even and n/2 + 1 rings if n is odd.
I hope I have helped you. Do comment if face any problems with the solution or its explanation.

Maximum number achievable by converting two adjacent x to one (x+1)

Given a sequence of N integers where 1 <= N <= 500 and the numbers are between 1 and 50. In a step any two adjacent equal numbers x x can be replaced with a single x + 1. What is the maximum number achievable by such steps.
For example if given 2 3 1 1 2 2 then the maximum possible is 4:
2 3 1 1 2 2 ---> 2 3 2 2 2 ---> 2 3 3 2 ---> 2 4 2.
It is evident that I should try to do better than the maximum number available in the sequence. But I can't figure out a good algorithm.
Each substring of the input can make at most one single number (invariant: the log base two of the sum of two to the power of each entry). For every x, we can find the set of substrings that can make x. For each x, this is (1) every occurrence of x (2) the union of two contiguous substrings that can make x - 1. The resulting algorithm is O(N^2)-time.
An algorithm could work like this:
Convert the input to an array where every element has a frequency attribute, collapsing repeated consecutive values in the input into one single node. For example, this input:
1 2 2 4 3 3 3 3
Would be represented like this:
{val: 1, freq: 1} {val: 2, freq: 2} {val: 4, freq: 1} {val: 3, freq: 4}
Then find local minima nodes, like the node (3 3 3 3) in 1 (2 2) 4 (3 3 3 3) 4, i.e. nodes whose neighbours both have higher values. For those local minima that have an even frequency, "lift" those by applying the step. Repeat this until no such local minima (with even frequency) exist any more.
Start of the recursive part of the algorithm:
At both ends of the array, work inwards to "lift" values as long as the more inner neighbour has a higher value. With this rule, the following:
1 2 2 3 5 4 3 3 3 1 1
will completely resolve. First from the left side inward:
1 4 5 4 3 3 3 1 1
Then from the right side:
1 4 6 3 2
Note that when there is an odd frequency (like for the 3s above), there will be a "remainder" that cannot be incremented. The remainder should in this rule always be left on the outward side, so to maximise the potential towards the inner part of the array.
At this point the remaining local minima have odd frequencies. Applying the step to such a node will always leave a "remainder" (like above) with the original value. This remaining node can appear anywhere, but it only makes sense to look at solutions where this remainder is on the left side or the right side of the lift (not in the middle). So for example:
4 1 1 1 1 1 2 3 4
Can resolve to one of these:
4 2 2 1 2 3 4
Or:
4 1 2 2 2 3 4
The 1 in either second or fourth position, is the above mentioned "remainder". Obviously, the second way of resolving is more promising in this example. In general, the choice is obvious when on one side there is a value that is too high to merge with, like the left-most 4 is too high for five 1 values to get to. The 4 is like a wall.
When the frequency of the local minimum is one, there is nothing we can do with it. It actually separates the array in a left and right side that do not influence each other. The same is true for the remainder element discussed above: it separates the array into two parts that do not influence each other.
So the next step in the algorithm is to find such minima (where the choice is obvious), apply that kind of step and separate the problem into two distinct problems which should be solved recursively (from the top). So in the last example, the following two problems would be solved separately:
4
2 2 3 4
Then the best of both solutions will count as the overall solution. In this case that is 5.
The most challenging part of the algorithm is to deal with those local minima for which the choice of where to put the remainder is not obvious. For instance;
3 3 1 1 1 1 1 2 3
This can go to either:
3 3 2 2 1 2 3
3 3 1 2 2 2 3
In this example the end result is the same for both options, but in bigger arrays it would be less and less obvious. So here both options have to be investigated. In general you can have many of them, like 2 in this example:
3 1 1 1 2 3 1 1 1 1 1 3
Each of these two minima has two options. This seems like to explode into too many possibilities for larger arrays. But it is not that bad. The algorithm can take opposite choices in neighbouring minima, and go alternating like this through the whole array. This way alternating sections are favoured, and get the most possible value drawn into them, while the other sections are deprived of value. Now the algorithm turns the tables, and toggles all choices so that the sections that were previously favoured are now deprived, and vice versa. The solution of both these alternatives is derived by resolving each section recursively, and then comparing the two "grand" solutions to pick the best one.
Snippet
Here is a live JavaScript implementation of the above algorithm.
Comments are provided which hopefully should make it readable.
"use strict";
function Node(val, freq) {
// Immutable plain object
return Object.freeze({
val: val,
freq: freq || 1, // Default frequency is 1.
// Max attainable value when merged:
reduced: val + (freq || 1).toString(2).length - 1
});
}
function compress(a) {
// Put repeated elements in a single node
var result = [], i, j;
for (i = 0; i < a.length; i = j) {
for (j = i + 1; j < a.length && a[j] == a[i]; j++);
result.push(Node(a[i], j - i));
}
return result;
}
function decompress(a) {
// Expand nodes into separate, repeated elements
var result = [], i, j;
for (i = 0; i < a.length; i++) {
for (j = 0; j < a[i].freq; j++) {
result.push(a[i].val);
}
}
return result;
}
function str(a) {
return decompress(a).join(' ');
}
function unstr(s) {
s = s.replace(/\D+/g, ' ').trim();
return s.length ? compress(s.split(/\s+/).map(Number)) : [];
}
/*
The function merge modifies an array in-place, performing a "step" on
the indicated element.
The array will get an element with an incremented value
and decreased frequency, unless a join occurs with neighboring
elements with the same value: then the frequencies are accumulated
into one element. When the original frequency was odd there will
be a "remainder" element in the modified array as well.
*/
function merge(a, i, leftWards, stats) {
var val = a[i].val+1,
odd = a[i].freq % 2,
newFreq = a[i].freq >> 1,
last = i;
// Merge with neighbouring nodes of same value:
if ((!odd || !leftWards) && a[i+1] && a[i+1].val === val) {
newFreq += a[++last].freq;
}
if ((!odd || leftWards) && i && a[i-1].val === val) {
newFreq += a[--i].freq;
}
// Replace nodes
a.splice(i, last-i+1, Node(val, newFreq));
if (odd) a.splice(i+leftWards, 0, Node(val-1));
// Update statistics and trace: this is not essential to the algorithm
if (stats) {
stats.total_applied_merges++;
if (stats.trace) stats.trace.push(str(a));
}
return i;
}
/* Function Solve
Parameters:
a: The compressed array to be reduced via merges. It is changed in-place
and should not be relied on after the call.
stats: Optional plain object that will be populated with execution statistics.
Return value:
The array after the best merges were applied to achieve the highest
value, which is stored in the maxValue custom property of the array.
*/
function solve(a, stats) {
var maxValue, i, j, traceOrig, skipLeft, skipRight, sections, goLeft,
b, choice, alternate;
if (!a.length) return a;
if (stats && stats.trace) {
traceOrig = stats.trace;
traceOrig.push(stats.trace = [str(a)]);
}
// Look for valleys of even size, and "lift" them
for (i = 1; i < a.length - 1; i++) {
if (a[i-1].val > a[i].val && a[i].val < a[i+1].val && (a[i].freq % 2) < 1) {
// Found an even valley
i = merge(a, i, false, stats);
if (i) i--;
}
}
// Check left-side elements with always increasing values
for (i = 0; i < a.length-1 && a[i].val < a[i+1].val; i++) {
if (a[i].freq > 1) i = merge(a, i, false, stats) - 1;
};
// Check right-side elements with always increasing values, right-to-left
for (j = a.length-1; j > 0 && a[j-1].val > a[j].val; j--) {
if (a[j].freq > 1) j = merge(a, j, true, stats) + 1;
};
// All resolved?
if (i == j) {
while (a[i].freq > 1) merge(a, i, true, stats);
a.maxValue = a[i].val;
} else {
skipLeft = i;
skipRight = a.length - 1 - j;
// Look for other valleys (odd sized): they will lead to a split into sections
sections = [];
for (i = a.length - 2 - skipRight; i > skipLeft; i--) {
if (a[i-1].val > a[i].val && a[i].val < a[i+1].val) {
// Odd number of elements: if more than one, there
// are two ways to merge them, but maybe
// one of both possibilities can be excluded.
goLeft = a[i+1].val > a[i].reduced;
if (a[i-1].val > a[i].reduced || goLeft) {
if (a[i].freq > 1) i = merge(a, i, goLeft, stats) + goLeft;
// i is the index of the element which has become a 1-sized valley
// Split off the right part of the array, and store the solution
sections.push(solve(a.splice(i--), stats));
}
}
}
if (sections.length) {
// Solve last remaining section
sections.push(solve(a, stats));
sections.reverse();
// Combine the solutions of all sections into one
maxValue = sections[0].maxValue;
for (i = sections.length - 1; i >= 0; i--) {
maxValue = Math.max(sections[i].maxValue, maxValue);
}
} else {
// There is no more valley that can be resolved without branching into two
// directions. Look for the remaining valleys.
sections = [];
b = a.slice(0); // take copy
for (choice = 0; choice < 2; choice++) {
if (choice) a = b; // restore from copy on second iteration
alternate = choice;
for (i = a.length - 2 - skipRight; i > skipLeft; i--) {
if (a[i-1].val > a[i].val && a[i].val < a[i+1].val) {
// Odd number of elements
alternate = !alternate
i = merge(a, i, alternate, stats) + alternate;
sections.push(solve(a.splice(i--), stats));
}
}
// Solve last remaining section
sections.push(solve(a, stats));
}
sections.reverse(); // put in logical order
// Find best section:
maxValue = sections[0].maxValue;
for (i = sections.length - 1; i >= 0; i--) {
maxValue = Math.max(sections[i].maxValue, maxValue);
}
for (i = sections.length - 1; i >= 0 && sections[i].maxValue < maxValue; i--);
// Which choice led to the highest value (choice = 0 or 1)?
choice = (i >= sections.length / 2)
// Discard the not-chosen version
sections = sections.slice(choice * sections.length/2);
}
// Reconstruct the solution from the sections.
a = [].concat.apply([], sections);
a.maxValue = maxValue;
}
if (traceOrig) stats.trace = traceOrig;
return a;
}
function randomValues(len) {
var a = [];
for (var i = 0; i < len; i++) {
// 50% chance for a 1, 25% for a 2, ... etc.
a.push(Math.min(/\.1*/.exec(Math.random().toString(2))[0].length,5));
}
return a;
}
// I/O
var inputEl = document.querySelector('#inp');
var randEl = document.querySelector('#rand');
var lenEl = document.querySelector('#len');
var goEl = document.querySelector('#go');
var outEl = document.querySelector('#out');
goEl.onclick = function() {
// Get the input and structure it
var a = unstr(inputEl.value),
stats = {
total_applied_merges: 0,
trace: a.length < 100 ? [] : undefined
};
// Apply algorithm
a = solve(a, stats);
// Output results
var output = {
value: a.maxValue,
compact: str(a),
total_applied_merges: stats.total_applied_merges,
trace: stats.trace || 'no trace produced (input too large)'
};
outEl.textContent = JSON.stringify(output, null, 4);
}
randEl.onclick = function() {
// Get input (count of numbers to generate):
len = lenEl.value;
// Generate
var a = randomValues(len);
// Output
inputEl.value = a.join(' ');
// Simulate click to find the solution immediately.
goEl.click();
}
// Tests
var tests = [
' ', '',
'1', '1',
'1 1', '2',
'2 2 1 2 2', '3 1 3',
'3 2 1 1 2 2 3', '5',
'3 2 1 1 2 2 3 1 1 1 1 3 2 2 1 1 2', '6',
'3 1 1 1 3', '3 2 1 3',
'2 1 1 1 2 1 1 1 2 1 1 1 1 1 2', '3 1 2 1 4 1 2',
'3 1 1 2 1 1 1 2 3', '4 2 1 2 3',
'1 4 2 1 1 1 1 1 1 1', '1 5 1',
];
var res;
for (var i = 0; i < tests.length; i+=2) {
var res = str(solve(unstr(tests[i])));
if (res !== tests[i+1]) throw 'Test failed: ' + tests[i] + ' returned ' + res + ' instead of ' + tests[i+1];
}
Enter series (space separated):<br>
<input id="inp" size="60" value="2 3 1 1 2 2"><button id="go">Solve</button>
<br>
<input id="len" size="4" value="30"><button id="rand">Produce random series of this size and solve</button>
<pre id="out"></pre>
As you can see the program produces a reduced array with the maximum value included. In general there can be many derived arrays that have this maximum; only one is given.
An O(n*m) time and space algorithm is possible, where, according to your stated limits, n <= 500 and m <= 58 (consider that even for a billion elements, m need only be about 60, representing the largest element ± log2(n)). m is representing the possible numbers 50 + floor(log2(500)):
Consider the condensed sequence, s = {[x, number of x's]}.
If M[i][j] = [num_j,start_idx] where num_j represents the maximum number of contiguous js ending at index i of the condensed sequence; start_idx, the index where the sequence starts or -1 if it cannot join earlier sequences; then we have the following relationship:
M[i][j] = [s[i][1] + M[i-1][j][0], M[i-1][j][1]]
when j equals s[i][0]
j's greater than s[i][0] but smaller than or equal to s[i][0] + floor(log2(s[i][1])), represent converting pairs and merging with an earlier sequence if applicable, with a special case after the new count is odd:
When M[i][j][0] is odd, we do two things: first calculate the best so far by looking back in the matrix to a sequence that could merge with M[i][j] or its paired descendants, and then set a lower bound in the next applicable cells in the row (meaning a merge with an earlier sequence cannot happen via this cell). The reason this works is that:
if s[i + 1][0] > s[i][0], then s[i + 1] could only possibly pair with the new split section of s[i]; and
if s[i + 1][0] < s[i][0], then s[i + 1] might generate a lower j that would combine with the odd j from M[i], potentially making a longer sequence.
At the end, return the largest entry in the matrix, max(j + floor(log2(num_j))), for all j.
JavaScript code (counterexamples would be welcome; the limit on the answer is set at 7 for convenient visualization of the matrix):
function f(str){
var arr = str.split(/\s+/).map(Number);
var s = [,[arr[0],0]];
for (var i=0; i<arr.length; i++){
if (s[s.length - 1][0] == arr[i]){
s[s.length - 1][1]++;
} else {
s.push([arr[i],1]);
}
}
var M = [new Array(8).fill([0,0])],
best = 0;
for (var i=1; i<s.length; i++){
M[i] = new Array(8).fill([0,i]);
var temp = s[i][1],
temp_odd,
temp_start,
odd = false;
for (var j=s[i][0]; temp>0; j++){
var start_idx = odd ? temp_start : M[i][j-1][1];
if (start_idx != -1 && M[start_idx - 1][j][0]){
temp += M[start_idx - 1][j][0];
start_idx = M[start_idx - 1][j][1];
}
if (!odd){
M[i][j] = [temp,start_idx];
temp_odd = temp;
} else {
M[i][j] = [temp_odd,-1];
temp_start = start_idx;
}
if (!odd && temp & 1 && temp > 1){
odd = true;
temp_start = start_idx;
}
best = Math.max(best,j + Math.floor(Math.log2(temp)));
temp >>= 1;
temp_odd >>= 1;
}
}
return [arr, s, best, M];
}
// I/O
var button = document.querySelector('button');
var input = document.querySelector('input');
var pre = document.querySelector('pre');
button.onclick = function() {
var val = input.value;
var result = f(val);
var text = '';
for (var i=0; i<3; i++){
text += JSON.stringify(result[i]) + '\n\n';
}
for (var i in result[3]){
text += JSON.stringify(result[3][i]) + '\n';
}
pre.textContent = text;
}
<input value ="2 2 3 3 2 2 3 3 5">
<button>Solve</button>
<pre></pre>
Here's a brute force solution:
function findMax(array A, int currentMax)
for each pair (i, i+1) of indices for which A[i]==A[i+1] do
currentMax = max(A[i]+1, currentMax)
replace A[i],A[i+1] by a single number A[i]+1
currentMax = max(currentMax, findMax(A, currentMax))
end for
return currentMax
Given the array A, let currentMax=max(A[0], ..., A[n])
print findMax(A, currentMax)
The algorithm terminates because in each recursive call the array shrinks by 1.
It's also clear that it is correct: we try out all possible replacement sequences.
The code is extremely slow when the array is large and there's lots of options regarding replacements, but actually works reasonbly fast on arrays with small number of replaceable pairs. (I'll try to quantify the running time in terms of the number of replaceable pairs.)
A naive working code in Python:
def findMax(L, currMax):
for i in range(len(L)-1):
if L[i] == L[i+1]:
L[i] += 1
del L[i+1]
currMax = max(currMax, L[i])
currMax = max(currMax, findMax(L, currMax))
L[i] -= 1
L.insert(i+1, L[i])
return currMax
# entry point
if __name__ == '__main__':
L1 = [2, 3, 1, 1, 2, 2]
L2 = [2, 3, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2]
print findMax(L1, max(L1))
print findMax(L2, max(L2))
The result of the first call is 4, as expected.
The result of the second call is 5 as expected; the sequence that gives the result: 2,3,1,1,2,2,2,2,2,2,2,2, -> 2,3,1,1,3,2,2,2,2,2,2 -> 2,3,1,1,3,3,2,2,2,2, -> 2,3,1,1,3,3,3,2,2 -> 2,3,1,1,3,3,3,3 -> 2,3,1,1,4,3, -> 2,3,1,1,4,4 -> 2,3,1,1,5

How to go about a d-smooth sequence algorithm

I'm really struggling to design an algorithm to find d, which is the lowest value that can be added or subtracted (at most) to make a given sequence strictly increasing.
For example.. say seq[] = [2,4,8,3,1,12]
given that sequence, the algorithm should return "5" as d because you can add or subtract at most 5 to each element such that the function is strictly increasing.
I've tried several approaches and can't seem to get a solid technique down.
I've tried looping through the seq. and checking if seq[i] < seq[i+1]. If not, it checks if d>0.. if it is, try to add/subtract it from seq[i+1]. Otherwise it calculates d by taking the difference of seq[i-1] - seq[i].
I can't get it to be stable though and Its like I keep adding if statements that are more "special cases" for unique input sequences. People have suggested using a binary search approach, but I can't make sense of applying it to this problem.
Any tips and suggestions are greatly appreciated. Thanks!
Here's my code in progress - using Python - v4
def ComputeMaxDelta3(seq):
# Create a copy to speed up comparison on modified values
aItems = seq[1:] #copies sequence elements from 1 (ignores seq[0])
# Will store the fix values for every item
# this should allocate 'length' times the 0 value
fixes = [0] * len(aItems)
print("fixes>>",fixes)
# Loop until no more fixes get applied
bNeedFix = True
while(bNeedFix):
# Hope will have no fix this turn
bNeedFix = False
# loop all subsequent item pairs (i should run from 0 to length - 2)
for i in range(0,len(aItems)-1):
# Left item
item1 = aItems[i]
# right item
item2 = aItems[i+1]
# Compute delta between left and right item
# We remember that (right >= left + 1
nDelta = item2 - (item1 + 1)
if(nDelta < 0):
# Fix the right item
fixes[i+1] -= nDelta
aItems[i+1] -= nDelta
# Need another loop
bNeedFix = True
# Compute the fix size (rounded up)
# max(s) should be int and the division should produce an int
nFix = int((max(fixes)+1)/2)
print("current nFix:",nFix)
# Balance all fixes
for i in range(len(aItems)):
fixes[i] -= nFix
print("final Fixes:",fixes)
print("d:",nFix)
print("original sequence:",seq[1:])
print("result sequence:",aItems)
return
Here's whats displayed:
Working with: [6, 2, 4, 8, 3, 1, 12]
[0]= 6 So the following numbers are the sequence:
aItems = [2, 4, 8, 3, 1, 12]
fixes>> [0, 0, 0, 0, 0, 0]
current nFix: 6
final Fixes: [-6, -6, -6, 0, 3, -6]
d: 1
original sequence: [2, 4, 8, 3, 1, 12]
result sequence: [2, 4, 8, 9, 10, 12]
d SHOULD be: 5
done!
~Note~
I start at 1 rather than 0 due to the first element being a key
As anticipated, here is (or should be) the Python version of my initial solution:
def ComputeMaxDelta(aItems):
# Create a copy to speed up comparison on modified values
aItems = aItems[:]
# Will store the fix values for every item
# this should allocate 'length' times the 0 value
fixes = [0] * len(aItems)
# Loop until no more fixes get applied
bNeedFix = True
while(bNeedFix):
# Hope will have no fix this turn
bNeedFix = False
# loop all subsequent item pairs (i should run from 0 to length - 2)
for i in range(0,len(aItems)-1):
# Left item
item1 = aItems[i]
# right item
item2 = aItems[i+1]
# Compute delta between left and right item
# We remember that (right >= left + 1
nDelta = item2 - (item1 + 1)
if(nDelta < 0):
# Fix the right item
fixes[i+1] -= nDelta
aItems[i+1] -= nDelta
# Need another loop
bNeedFix = True
# Compute the fix size (rounded up)
# max(s) should be int and the division should produce an int
nFix = (max(fixes)+1)/2 # corrected from **(max(s)+1)/2**
# Balance all fixes
for i in range(len(s)):
fixes[i] -= nFix
print("d:",nFix) # corrected from **print("d:",nDelta)**
print("s:",fixes)
return
I took your Python and fixed in order to operate exactly as my C# solution.
I don't know Python, but looking for some reference on the web, I should have found the points where your porting was failing.
If you compare your python version with mine you should find the following differences:
You saved a reference aItems into s and used it as my fixes, but fixes was meant to start as all 0.
You didn't cloned aItems over itself, then every alteration to its items was reflected outside of the method.
Your for loop was starting at index 1, whereas mine started at 0 (the very first element).
After the check for nDelta you subtracted nDelta from both s and aItems, but as I stated at points 1 and 2 they were pointing to the same items.
The ceil instruction was unnedeed because the division between two integers produces an integer, as with C#.
Please remember that I fixed the Python code basing my knowledge only on online documentation, because I don't code in that language, so I'm not 100% sure about some syntax (my main doubt is about the fixes declaration).
Regards,
Daniele.
Here is my solution:
public static int ComputeMaxDelta(int[] aItems, out int[] fixes)
{
// Create a copy to speed up comparison on modified values
aItems = (int[])aItems.Clone();
// Will store the fix values for every item
fixes = new int[aItems.Length];
// Loop until no more fixes get applied
var bNeedFix = true;
while (bNeedFix)
{
// Hope will have no fix this turn
bNeedFix = false;
// loop all subsequent item pairs
for (int ixItem = 0; ixItem < aItems.Length - 1; ixItem++)
{
// Left item
var item1 = aItems[ixItem];
// right item
var item2 = aItems[ixItem + 1];
// Compute delta between left and right item
// We remember that (right >= left + 1)
var nDelta = item2 - (item1 + 1);
if (nDelta < 0)
{
// Fix the right item
fixes[ixItem + 1] -= nDelta;
aItems[ixItem + 1] -= nDelta;
//Need another loop
bNeedFix = true;
}
}
}
// Compute the fix size (rounded up)
var nFix = (fixes.Max() + 1) / 2;
// Balance all fixes
for (int ixItem = 0; ixItem < aItems.Length; ixItem++)
fixes[ixItem] -= nFix;
return nFix;
}
The function returns the maximum computed fix gap.
As a bounus, the parameter fixes will receive the fixes for every item. These are the delta to apply to each source value in order to be sure that they will be in ascending order: some fix can be reduced but some analysis loop is required to achieve that optimization.
The following is a code to test the algorithm. If you set a breakpoint at the end of the loop, you'll be able to check the result for sequence you provided in your example.
var random = new Random((int)Stopwatch.GetTimestamp());
for (int ixLoop = -1; ixLoop < 100; ixLoop++)
{
int nCount;
int[] aItems;
// special case as the provided sample sequence
if (ixLoop == -1)
{
aItems = new[] { 2, 4, 8, 3, 1, 12 };
nCount = aItems.Length;
}
else
{
// Generates a random amount of items based on my screen's width
nCount = 4 + random.Next(21);
aItems = new int[nCount];
for (int ixItem = 0; ixItem < nCount; ixItem++)
{
// Keep the generated numbers below 30 for easier human analysis
aItems[ixItem] = random.Next(30);
}
}
Console.WriteLine("***");
Console.WriteLine(" # " + GetText(Enumerable.Range(0, nCount).ToArray()));
Console.WriteLine(" " + GetText(aItems));
int[] aFixes;
var nFix = ComputeMaxDelta(aItems, out aFixes);
// Computes the new values, that will be always in ascending order
var aNew = new int[aItems.Length];
for (int ixItem = 0; ixItem < aItems.Length; ixItem++)
{
aNew[ixItem] = aItems[ixItem] + aFixes[ixItem];
}
Console.WriteLine(" = " + nFix.ToString());
Console.WriteLine(" ! " + GetText(aFixes));
Console.WriteLine(" > " + GetText(aNew));
}
Regards,
Daniele.

Maximum Devastation to be caused if a building with height h causes all h-1 buildings to its right to collapse

In a recent interview question I got the following problem.
In a particular city we have a row of buildings with varying heights.
The collapse of a building with height h causes the next h-1 buildings on its right to collapse too.
The height of the buildings can be between 1 and 5000. Given the heights of all the buildings (arranged from left to right ie; for leftmost building index=1 and for rightmost building index=N) we needed to find out the index of the building which would cause the maximum devastation.
For example:
Input:
Number of buildings : 6
Height of Buildings:
2 1 3 3 1 6
Answer should be building at the index 3
The solution I tried was using the brute force technique with a complexity of O(N^2).
What I did was for each building in the list I found out the number of buildings that it would destroy.
Could a better solution for this question be constructed?
Simply go from the left, collapse the first building, and calculate how much total(*) damage it did.
Do this again and again from the very next building (which hasn't collapsed).
From these, pick the maximum.
Complexity: O(n).
This greedy algorithm works because the whole process is a chain reaction, if building A forces the collapse of B, then you cannot achieve better score starting from B.
(*) You can do this by maintaining one counter which stores how many buildings to the right should be collapsed. counter = max(counter - 1, height of next building).
some areas of the city function as "firewalls" - collapse stops at that point. a little thought shows that these are sections to the left of a value of 1 where height increases (to the left) no more than once per step (if you can have 0 heights that complicates things very slightly).
and the highest scoring region must start just after a firewall (since if it didn't there would be a higher region just to the left).
so scan from the right, finding these firewalls, and then find which section to the right of a firewall has the largest damage. this is O(n) because it's just linear scans (once from right to left and then once for each section, with no overlap).
actually, Karoly's answer is equivalent and simpler to implement.
Start with rightmost index.
The last building shall cause a devastation value of 1.
Iterate leftwards.
Something like (devastation from building i)
D[i] = 1 + min( N-i, max( index[i]-1, 0+D[i+1],1+D[i+2],... to index[i]-1 terms ) )
Same approach as #Karoly's answer. In ruby:
def find_max_damage_index(buildings)
max_damage = 0
max_start_index = nil
current_start_index = nil
current_h = 0
current_damage = 0
buildings.each_with_index{|h,i|
if current_h == 0 #end of batch
if current_damage > max_damage
max_damage = current_damage
max_start_index = current_start_index
end
#start new batch
current_h = h
current_damage = 1
current_start_index = i
else
current_h = h if h > current_h
current_damage += 1
end
current_h -= 1
}
#last batch
if current_damage > max_damage
max_damage = current_damage
max_start_index = current_start_index
end
return max_start_index
end
In Java, without considering subsequent collapses:
public static int collapse(int[] buildings) {
int i, maxDamage, index, currentDamage;
// do not consider the last building, it will cause only its own fall
maxDamage = 1;
index = buildings.length - 1;
for(i = buildings.length - 1; i >= 0; i--) {
// update maximum damage as the mimimum value between the building[i] (the height of the building at index i) and the remaining number of elements from i to the end of the array
currentDamage = Math.min(buildings[i], buildings.length - i);
System.out.println(currentDamage);
if(currentDamage > maxDamage) {
maxDamage = currentDamage;
index = i;
}
}
return index;
}
My final solution is different from the accepted one, which by the way I didn't fully understand.
The idea is to count starting from the rightmost position the number of buildings that the current index will collapse.
index: 7 6 5 4 3 2 1 0
height: 1 3 1 2 4 1 3 2
damage: 1 2 1 2 4 1 3 2
Then I just make a cumulative sum, starting from the rightmost position again. I add to the number of buildings the current position collapses the number of buildings that were collapsed staring from the next building to the right that didn't collapse until the end.
index: 7 6 5 4 3 2 1 0
height: 1 3 1 2 4 1 3 2
damage: 1 2 1 2 5 1 7 8
In the end, I just return the index with the maximum damage.
This solution runs in O(n) but uses an extra O(n) space.
The next code is the complete version (also works for subsequent collapses):
public static int collapse(int[] buildings) {
int i, maxIndex, max;
int damage[] = new int[buildings.length];
for(i = buildings.length - 1; i >= 0; i--) {
// compute damage for each position
damage[i] = Math.min(buildings[i], buildings.length - i);
}
for(i = buildings.length - 1; i >= 0; i--) {
// update total accumulated damage for each position
if(damage[i] > 1) {
if(damage[i] + i - 1 < buildings.length && i != (i + damage[i] - 1) ) {
damage[i] += damage[i + damage[i] - 1] - 1;
}
}
}
max = damage[0];
maxIndex = 0;
for(i = 1; i < buildings.length; i++) {
// find the maximum damage
if(damage[i] > max) {
max = damage[i];
maxIndex = i;
}
}
return maxIndex;
}

Understanding merge sort optimization: avoiding copies

I have below merge sort program in algorithms book, it is mentioned that The main problem is that merging two sorted lists requires linear extra memory, and the additional work spent copying to the temporary array and back, throughout the algorithm, has the effect of slowing down the sort considerably. This copying can be avoided by judiciously switching the roles of "a" and "tmp_array" at alternate levels of the recursion.
My question is what does author mean "copying can be avoided by judiciously switching the roles of a and tmp_array at alternate levels of the recursion" and how it is possible in following code? Request to show an example how we can achieve this?
void mergesort( input_type a[], unsigned int n ) {
input_type *tmp_array;
tmp_array = (input_type *) malloc( (n+1) * sizeof (input_type) );
m_sort( a, tmp_array, 1, n );
free( tmp_array );
}
void m_sort( input_type a[], input_type tmp_array[ ], int left, int right ) {
int center;
if( left < right ) {
center = (left + right) / 2;
m_sort( a, tmp_array, left, center );
m_sort( a, tmp_array, center+1, right );
merge( a, tmp_array, left, center+1, right );
}
}
void merge( input_type a[ ], input_type tmp_array[ ], int l_pos, int r_pos, int right_end ) {
int i, left_end, num_elements, tmp_pos;
left_end = r_pos - 1;
tmp_pos = l_pos;
num_elements = right_end - l_pos + 1;
/* main loop */
while( ( 1_pos <= left_end ) && ( r_pos <= right_end ) )
if( a[1_pos] <= a[r_pos] )
tmp_array[tmp_pos++] = a[l_pos++];
else
tmp_array[tmp_pos++] = a[r_pos++];
while( l_pos <= left_end ) /* copy rest of first half */
tmp_array[tmp_pos++] = a[l_pos++];
while( r_pos <= right_end ) /* copy rest of second half */
tmp_array[tmp_pos++] = a[r_pos++];
/* copy tmp_array back */
for(i=1; i <= num_elements; i++, right_end-- )
a[right_end] = tmp_array[right_end];
}
I'm going to assume that, without looking at this code, it is performing merge sort by declaring a contiguous block of memory the same size as the original.
So normally merge sort is like this:
split array in half
sort half-arrays by recursively invoking MergeSort on them
merge half-arrays back
I'm assuming it's recursive, so no copies will be done before we're sorting sub-arrays of size 2. Now what happens?
_ means it is memory we have available, but we don't care about the data in it
original:
8 5 2 3 1 7 4 6
_ _ _ _ _ _ _ _
Begin recursive calls:
recursive call 1:
(8 5 2 3) (1 7 4 6)
_ _ _ _ _ _ _ _
recursive call 2:
((8 5) (2 3)) ((1 7) (4 6))
_ _ _ _ _ _ _ _
recursive call 3:
(((8) (5))((2) (3)))(((1) (7))((4) (6)))
_ _ _ _ _ _ _ _
Recursive calls resolving with merging, PLUS COPYING (uses more memory, or alternatively is 'slower'):
merge for call 3, using temp space:
(((8) (5))((2) (3)))(((1) (7))((4) (6))) --\ perform merge
(( 5 8 )( 2 3 ))(( 1 7 )( 4 6 )) <--/ operation
UNNECESSARY: copy back:
(( 5 8 )( 2 3 ))(( 1 7 )( 4 6 )) <--\ copy and
_ _ _ _ _ _ _ _ --/ ignore old
merge for call 2, using temp space:
(( 5 8 )( 2 3 ))(( 1 7 )( 4 6 )) --\ perform merge
( 2 3 5 8 )( 1 4 6 7 ) <--/ operation
UNNECESSARY: copy back:
( 2 3 5 8 )( 1 4 6 7 ) <--\ copy and
_ _ _ _ _ _ _ _ --/ ignore old
merge for call 1, using temp space:
( 2 3 5 8 )( 1 4 6 7 ) --\ perform merge
1 2 3 4 5 6 7 8 <--/ operation
UNNECESSARY: copy back:
1 2 3 4 5 6 7 8 <--\ copy and
_ _ _ _ _ _ _ _ --/ ignore old
What the author is suggesting
Recursive calls resolving with merging, WITHOUT COPYING (uses less memory):
merge for call 3, using temp space:
(((8) (5))((2) (3)))(((1) (7))((4) (6))) --\ perform merge
(( 5 8 )( 2 3 ))(( 1 7 )( 4 6 )) <--/ operation
merge for call 2, using old array as temp space:
( 2 3 5 8 )( 1 4 6 7 ) <--\ perform merge
(( 5 8 )( 2 3 ))(( 1 7 )( 4 6 )) --/ operation (backwards)
merge for call 1, using temp space:
( 2 3 5 8 )( 1 4 6 7 ) --\ perform merge
1 2 3 4 5 6 7 8 <--/ operation
There you go: you don't need to do copies as long as you perform each "level" of the merge-sort tree in lock-step, as shown above.
You may have a minor issue of parity, also as demonstrated above. That is, your result may be in your temp_array. You either have three options for dealing with this:
returning the temp_array as the answer, and release the old memory (if your application is fine with that)
perform a single array copy operation, to copy temp_array back into your original array
allow yourself to consume a mere twice-as-much memory, and perform a single cycle of merges from temp_array1 to temp_array2 then back to original_array, then release temp_array2. The parity issue should be resolved.
This is not necessarily "faster":
additional work spent copying to the temporary array and back
This is actually not the core reason why it's 'faster' per se. It is obviously not asymptotically faster, nor necessarily even faster. There is a notion of latency vs. throughput. Generally running time is measured in latency, because extra garbage work (like releasing memory) may be done asynchronously. You don't necessarily need to copy "back" to the original array depending on your language. However, if you are repeating something many times on memory-bound hardware in a garbage-collected language, the garbage collection can occasionally be forced to spike if the GC algorithm is a poor choice for what you are doing (or if this is C, maybe you are waiting to allocate). Thus if you were to create extra memory in a GC language, it should not really count against you. Granted, this may cause you not to take advantage of cache properly if you use too much memory. You'd have to benchmark it yourself, very carefully for your use case.
I do not recommend creating random temporary arrays for each step though, as that would make memory O(N log(N)) and this is a trivial optimization.
Minor notes on in-placeness:
Also, the reason you can't naively do it in-place is because while you are merging two sorted sub-arrays, the new result sorted sub-array may take arbitrarily many from one input array before spontaneously swap to the other array. For example, as you can see we need a buffer because our input arrays might get split into fragments:
( 4 6 7 8 10)(1 2 3 5 9 11)(... other sub-arrays)
( 1)(6 7 8 10)(4)(2 3 5 9 11)(...
( 1 2)(7 8 10)(4 6)(3 5 9 11) ...
( 1 2 3)(8 10)(4 6 7)(5 9 11)
( 1 2 3 4(10)(8)(6 7)(5 9 11) ooph :-(
( 1 2 3 4 5)(8)(6 7)(10)(9 11) ooph
You might be able to so cleverly in-place if you do some weird variant of the kth-statistic median-of-medians algorithm, performing your merge into the middle of the two arrays rather than the start (merging from a specifically chosen element outwards left/decreasing and right/increasing simultaneously). I'm not sure how one would implement that though, or if the hunch is true.
(very minor note: Perhaps those who are familiar with sorting algorithms should be careful of comparing a traditional swap traditional swap operation involving a tmp variable in a register, which is two reads-from-cache and two writes-to-cache, to not-in-place copying to other bits of memory, without a per-operation counting argument.)
Certainly, OP's method is extremely simple to code for only twice as much memory.
Start by thinking of merge sort in this way.
0: Consider the input array A0 as a collection of ordered sequences of
length 1.
1: Merge each consecutive pair of sequences from A0, constructing a
new temporary array A1.
2: Merge each consecutive pair of sequences from A1, constructing a
new temporary array A2.
...
Finish when the last iteration results in a single sequence.
Now, you can obviously get away with just a single temporary array by doing this:
0: Consider the input array A0 as a collection of ordered sequences of
length 1.
1: Merge each consecutive pair of sequences from A0, constructing a
new temporary array A1.
2: Merge each consecutive pair of sequences from A1, overwriting A0
with the result.
3: Merge each consecutive pair of sequences from A0, overwriting A1
with the result.
...
Finish when the last iteration results in a single sequence.
Of course, you can be even smarter than this. If you want to be nicer to the cache, you might decide to sort top-down, rather than bottom-up. In this case, it hopefully becomes obvious what your textbook means when it refers to tracking the role of the arrays at different levels of recursion.
Hope this helps.
Here is my implementation without extra copies.
public static void sort(ArrayList<Integer> input) {
mergeSort(input, 0, input.size() - 1);
}
/**
* Sorts input and returns inversions number
* using classical divide and conquer approach
*
* #param input Input array
* #param start Start index
* #param end End index
* #return int
*/
private static long mergeSort(ArrayList<Integer> input, int start, int end) {
if (end - start < 1) {
return 0;
}
long inversionsNumber = 0;
// 1. divide input into subtasks
int pivot = start + (int) Math.ceil((end - start) / 2);
if (end - start > 1) {
inversionsNumber += mergeSort(input, start, pivot);
inversionsNumber += mergeSort(input, pivot + 1, end);
}
// 2. Merge the results
int offset = 0;
int leftIndex = start;
int rightIndex = pivot + 1;
while (leftIndex <= pivot && rightIndex <= end) {
if (input.get(leftIndex + offset) <= input.get(rightIndex)) {
if (leftIndex < pivot) {
leftIndex++;
} else {
rightIndex++;
}
continue;
}
moveElement(input, rightIndex, leftIndex + offset);
inversionsNumber += rightIndex - leftIndex - offset;
rightIndex++;
offset++;
}
return inversionsNumber;
}
private static void moveElement(ArrayList<Integer> input, int from, int to) {
assert 0 <= to;
assert to < from;
assert from < input.size();
int temp = input.get(from);
for (int i = from; i > to; i--) {
input.set(i, input.get(i - 1));
}
input.set(to, temp);
}
Look at the very last part of the merge function. What if, instead of copying that data, you just used the knowledge that the sorted part is now in tmp_array instead of a when the function returns, and a is available for use as a temp.
Details are left as an exercise for the reader.

Resources