Efficient string truncation algorithm, sequentially removing equal prefixes and suffixes

Question

Time limit per test: 5 seconds
Memory limit per test: 512 megabytes

You are given a string s of length n (n ≤ 5000). You can select any proper prefix of this string that is also its suffix and remove either selected prefix or corresponding suffix. Then you can apply an analogous operation to a resulting string and so on. What is the minimum length of the final string, that can be achieved after applying the optimal sequence of such operations?

Input
The first line of each test contains a string s that consists of small English letters.

Output
Output a single integer — the minimum length of the final string, that can be achieved after applying the optimal sequence of such operations.

Examples +-------+--------+----------------------------------+ | Input | Output | Explanation | +-------+--------+----------------------------------+ | caaca | 2 | caaca → ca|aca → aca → ac|a → ac | +-------+--------+----------------------------------+ | aabaa | 2 | aaba|a → a|aba → ab|a → ab | +-------+--------+----------------------------------+ | abc | 3 | No operations are possible | +-------+--------+----------------------------------+

Here is what I've managed to do so far:

1. Calculate the prefix function for all substrings of a given string in O(n^2)

2. Check the result of performing all the possible combinations of operations in O(n^3)

My solution passes all the tests at n ≤ 2000 but exceeds the time limit when 2000 < n ≤ 5000. Here is its code:

#include <iostream>
#include <string>

using namespace std;

const int MAX_N = 5000;

int result; // 1 less than actual

// [x][y] corresponds to substring that starts at position `x` and ends at position `x + y` =>
// => corresponding substring length is `y + 1`
int lps[MAX_N][MAX_N]; // prefix function for the substring s[x..x+y]
bool checked[MAX_N][MAX_N]; // whether substring s[x..x+y] is processed by check function

// length is 1 less than actual
void check(int start, int length) {
    checked[start][length] = true;
    if (length < result) {
        if (length == 0) {
            cout << 1; // actual length = length + 1 = 0 + 1 = 1
            exit(0); // 1 is the minimum possible result
        }
        result = length;
    }
    // iteration over all proper prefixes that are also suffixes
    // i - current prefix length
    for (int i = lps[start][length]; i != 0; i = lps[start][i - 1]) {
        int newLength = length - i;
        int newStart = start + i;
        if (!checked[start][newLength])
            check(start, newLength);
        if (!checked[newStart][newLength])
            check(newStart, newLength);
    }
}

int main()
{
    string str;
    cin >> str;
    int n = str.length();
    // lps calculation runs in O(n^2)
    for (int l = 0; l < n; l++) {
        int subLength = n - l;
        lps[l][0] = 0;
        checked[l][0] = false;
        for (int i = 1; i < subLength; ++i) {
            int j = lps[l][i - 1];
            while (j > 0 && str[i + l] != str[j + l])
                j = lps[l][j - 1];
            if (str[i + l] == str[j + l])  j++;
            lps[l][i] = j;
            checked[l][i] = false;
        }
    }
    result = n - 1;
    // checking all possible operations combinations in O(n^3)
    check(0, n - 1);
    cout << result + 1;
}

Q: Is there any more efficient solution?

Answer 1

Here's one way to get the log factor. Let dp[i][j] be true if we can reach the substring s[i..j]. Then:

dp[0][length(s)-1] ->
  true

dp[0][j] ->
  if s[0] != s[j+1]:
    false
  else:
    true if any dp[0][k]
      for j < k ≤ (j + longestMatchRight[0][j+1])

  (The longest match we can use is
   also bound by the current range.)

(Initialise left side similarly.)

Now iterate from the outside in:

for i = 1 to length(s)-2:
  for j = length(s)-2 to i:
    dp[i][j] ->
      // We removed on the right
      if s[i] != s[j+1]:
        false
      else:
        true if any dp[i][k]
          for j < k ≤ (j + longestMatchRight[i][j+1])

      // We removed on the left
      if s[i-1] != s[j]:
        true if dp[i][j]
      else:
        true if any dp[k][j]
          for (i - longestMatchLeft[i-1][j]) ≤ k < i

We can precompute the longest match for each starting pair (i, j) in O(n^2) with the recurrence,

longest(i, j) -> 
  if s[i] == s[j]:
    return 1 + longest(i + 1, j + 1)
  else:
    return 0

This would allow us to check for a substring match that starts at indexes i and j in O(1). (We need both right and left directions.)

How to get the log factor

We can think of a way to come up with a data structure that would allow us to determine if

any dp[i][k]
  for j < k ≤ (j + longestMatchRight[i][j+1])

(And similarly for the left side.)

in O(log n), considering we have already seen those values.

Here's C++ code with segment trees (for right and left queries, so O(n^2 * log n)) that includes Bananon's test generator. For 5000 "a" characters, it ran in 3.54s, 420 MB (https://ideone.com/EIrhnR). To reduce the memory, one of the segment trees is implemented on a single row (I still need to investigate doing the same with the left side queries to reduce memory even further.)

#include <iostream>
#include <string>
#include <ctime>
#include <random>
#include <algorithm>    // std::min

using namespace std;

const int MAX_N = 5000;

int seg[2 * MAX_N];
int segsL[MAX_N][2 * MAX_N];
int m[MAX_N][MAX_N][2];
int dp[MAX_N][MAX_N];
int best;

// Adapted from https://codeforces.com/blog/entry/18051
void update(int n, int p, int value) { // set value at position p
  for (seg[p += n] = value; p > 1; p >>= 1)
    seg[p >> 1] = seg[p] + seg[p ^ 1];
}
// Adapted from https://codeforces.com/blog/entry/18051
int query(int n, int l, int r) { // sum on interval [l, r)
  int res = 0;
  for (l += n, r += n; l < r; l >>= 1, r >>= 1) {
    if (l & 1) res += seg[l++];
    if (r & 1) res += seg[--r];
  }
  return res;
}
// Adapted from https://codeforces.com/blog/entry/18051
void updateL(int n, int i, int p, int value) { // set value at position p
  for (segsL[i][p += n] = value; p > 1; p >>= 1)
    segsL[i][p >> 1] = segsL[i][p] + segsL[i][p ^ 1];
}
// Adapted from https://codeforces.com/blog/entry/18051
int queryL(int n, int i, int l, int r) { // sum on interval [l, r)
  int res = 0;
  for (l += n, r += n; l < r; l >>= 1, r >>= 1) {
    if (l & 1) res += segsL[i][l++];
    if (r & 1) res += segsL[i][--r];
  }
  return res;
}

// Code by גלעד ברקן
void precalc(int n, string & s) {
  int i, j;
  for (i = 0; i < n; i++) {
    for (j = 0; j < n; j++) {
      // [longest match left, longest match right]
      m[i][j][0] = (s[i] == s[j]) & 1;
      m[i][j][1] = (s[i] == s[j]) & 1;
    }
  }

  for (i = n - 2; i >= 0; i--)
    for (j = n - 2; j >= 0; j--)
      m[i][j][1] = s[i] == s[j] ? 1 + m[i + 1][j + 1][1] : 0;

  for (i = 1; i < n; i++)
    for (j = 1; j < n; j++)
      m[i][j][0] = s[i] == s[j] ? 1 + m[i - 1][j - 1][0] : 0;
}

// Code by גלעד ברקן
void f(int n, string & s) {
  int i, j, k, longest;

  dp[0][n - 1] = 1;
  update(n, n - 1, 1);
  updateL(n, n - 1, 0, 1);

  // Right side initialisation
  for (j = n - 2; j >= 0; j--) {
    if (s[0] == s[j + 1]) {
      longest = std::min(j + 1, m[0][j + 1][1]);
      for (k = j + 1; k <= j + longest; k++)
        dp[0][j] |= dp[0][k];
      if (dp[0][j]) {
        update(n, j, 1);
        updateL(n, j, 0, 1);
        best = std::min(best, j + 1);
      }
    }
  }

  // Left side initialisation
  for (i = 1; i < n; i++) {
    if (s[i - 1] == s[n - 1]) {
      // We are bound by the current range
      longest = std::min(n - i, m[i - 1][n - 1][0]);
      for (k = i - 1; k >= i - longest; k--)
        dp[i][n - 1] |= dp[k][n - 1];
      if (dp[i][n - 1]) {
        updateL(n, n - 1, i, 1);
        best = std::min(best, n - i);
      }
    }
  }

  for (i = 1; i <= n - 2; i++) {
    for (int ii = 0; ii < MAX_N; ii++) {
      seg[ii * 2] = 0;
      seg[ii * 2 + 1] = 0;
    }
    update(n, n - 1, dp[i][n - 1]);
    for (j = n - 2; j >= i; j--) {
      // We removed on the right
      if (s[i] == s[j + 1]) {
        // We are bound by half the current range
        longest = std::min(j - i + 1, m[i][j + 1][1]);
        //for (k=j+1; k<=j+longest; k++)
        //dp[i][j] |= dp[i][k];
        if (query(n, j + 1, j + longest + 1)) {
          dp[i][j] = 1;
          update(n, j, 1);
          updateL(n, j, i, 1);
        }
      }
      // We removed on the left
      if (s[i - 1] == s[j]) {
        // We are bound by half the current range
        longest = std::min(j - i + 1, m[i - 1][j][0]);
        //for (k=i-1; k>=i-longest; k--)
        //dp[i][j] |= dp[k][j];
        if (queryL(n, j, i - longest, i)) {
          dp[i][j] = 1;
          updateL(n, j, i, 1);
          update(n, j, 1);
        }
      }
      if (dp[i][j])
        best = std::min(best, j - i + 1);
    }
  }
}

int so(string s) {
  for (int i = 0; i < MAX_N; i++) {
    seg[i * 2] = 0;
    seg[i * 2 + 1] = 0;
    for (int j = 0; j < MAX_N; j++) {
      segsL[i][j * 2] = 0;
      segsL[i][j * 2 + 1] = 0;
      m[i][j][0] = 0;
      m[i][j][1] = 0;
      dp[i][j] = 0;
    }
  }
  int n = s.length();
  best = n;
  precalc(n, s);
  f(n, s);
  return best;
}
// End code by גלעד ברקן

// Code by Bananon  =======================================================================

int result;

int lps[MAX_N][MAX_N];
bool checked[MAX_N][MAX_N];

void check(int start, int length) {
  checked[start][length] = true;
  if (length < result) {
    result = length;
  }
  for (int i = lps[start][length]; i != 0; i = lps[start][i - 1]) {
    int newLength = length - i;
    if (!checked[start][newLength])
      check(start, newLength);
    int newStart = start + i;
    if (!checked[newStart][newLength])
      check(newStart, newLength);
  }
}

int my(string str) {
  int n = str.length();
  for (int l = 0; l < n; l++) {
    int subLength = n - l;
    lps[l][0] = 0;
    checked[l][0] = false;
    for (int i = 1; i < subLength; ++i) {
      int j = lps[l][i - 1];
      while (j > 0 && str[i + l] != str[j + l])
        j = lps[l][j - 1];
      if (str[i + l] == str[j + l]) j++;
      lps[l][i] = j;
      checked[l][i] = false;
    }
  }
  result = n - 1;
  check(0, n - 1);
  return result + 1;
}

// generate =================================================================

bool rndBool() {
  return rand() % 2 == 0;
}

int rnd(int bound) {
  return rand() % bound;
}

void untrim(string & str) {
  int length = rnd(str.length());
  int prefixLength = rnd(str.length()) + 1;
  if (rndBool())
    str.append(str.substr(0, prefixLength));
  else {
    string newStr = str.substr(str.length() - prefixLength, prefixLength);
    newStr.append(str);
    str = newStr;
  }
}

void rndTest(int minTestLength, string s) {
  while (s.length() < minTestLength)
    untrim(s);
  int myAns = my(s);
  int soAns = so(s);
  cout << myAns << " " << soAns << '\n';
  if (soAns != myAns) {
    cout << s;
    exit(0);
  }
}

int main() {
  int minTestLength;
  cin >> minTestLength;
  string seed;
  cin >> seed;
  while (true)
    rndTest(minTestLength, seed);
}

And here's JavaScript code (without the log factor improvement) to show that the recurrence works. (To get the log factor, we replace the inner k loops with a single range query.)

debug = 1

function precalc(s){
  let m = new Array(s.length)
  for (let i=0; i<s.length; i++){
    m[i] = new Array(s.length)
    for (let j=0; j<s.length; j++){
      // [longest match left, longest match right]
      m[i][j] = [(s[i] == s[j]) & 1, (s[i] == s[j]) & 1]
    }
  }
  
  for (let i=s.length-2; i>=0; i--)
    for (let j=s.length-2; j>=0; j--)
      m[i][j][1] = s[i] == s[j] ? 1 + m[i+1][j+1][1] : 0

  for (let i=1; i<s.length; i++)
    for (let j=1; j<s.length; j++)
      m[i][j][0] = s[i] == s[j] ? 1 + m[i-1][j-1][0] : 0
  
  return m
}

function f(s){
  m = precalc(s)
  let n = s.length
  let min = s.length
  let dp = new Array(s.length)

  for (let i=0; i<s.length; i++)
    dp[i] = new Array(s.length).fill(0)

  dp[0][s.length-1] = 1
      
  // Right side initialisation
  for (let j=s.length-2; j>=0; j--){
    if (s[0] == s[j+1]){
      let longest = Math.min(j + 1, m[0][j+1][1])
      for (let k=j+1; k<=j+longest; k++)
        dp[0][j] |= dp[0][k]
      if (dp[0][j])
        min = Math.min(min, j + 1)
    }
  }

  // Left side initialisation
  for (let i=1; i<s.length; i++){
    if (s[i-1] == s[s.length-1]){
      let longest = Math.min(s.length - i, m[i-1][s.length-1][0])
      for (let k=i-1; k>=i-longest; k--)
        dp[i][s.length-1] |= dp[k][s.length-1]
      if (dp[i][s.length-1])
        min = Math.min(min, s.length - i)
    }
  }

  for (let i=1; i<=s.length-2; i++){
    for (let j=s.length-2; j>=i; j--){
      // We removed on the right
      if (s[i] == s[j+1]){
        // We are bound by half the current range
        let longest = Math.min(j - i + 1, m[i][j+1][1])
        for (let k=j+1; k<=j+longest; k++)
          dp[i][j] |= dp[i][k]
      }
      // We removed on the left
      if (s[i-1] == s[j]){
        // We are bound by half the current range
        let longest = Math.min(j - i + 1, m[i-1][j][0])
        for (let k=i-1; k>=i-longest; k--)
          dp[i][j] |= dp[k][j]
      }
      if (dp[i][j])
        min = Math.min(min, j - i + 1)
    }
  }

  if (debug){
    let str = ""
    for (let row of dp)
      str += row + "\n"
    console.log(str)
  }

  return min
}

function main(s){
  var strs = [
    "caaca",
    "bbabbbba",
    "baabbabaa",
    "bbabbba",
    "bbbabbbbba",
    "abbabaabbab",
    "abbabaabbaba",
    "aabaabaaabaab",
    "bbabbabbb"
  ]

  for (let s of strs){
    let t = new Date
    console.log(s)
    console.log(f(s))
    //console.log((new Date - t)/1000)
    console.log("")
  }
}

main()