15-2 Longest palindrome subsequence

A palindrome is a nonempty string over some alphabet that reads the same forward and backward. Examples of palindromes are all strings of length $1$, $\text{civic}$, $\text{racecar}$, and $\text{aibohphobia}$ (fear of palindromes).

Give an efficient algorithm to find the longest palindrome that is a subsequence of a given input string. For example, given the input $\text{character}$, your algorithm should return $\text{carac}$. What is the running time of your algorithm?

We solve the longest palindrome subsequence ($\text{LPS}$) problem in a manner similar to how we compute the longest common subsequence in Section 15.4.

Step 1: Characterizing a longest palindrome subsequence

The $\text{LPS}$ problem has an optimal-substructure property, where the subproblems correspond to pairs of indices, starting and ending, of the input sequence.

For a sequence $X = \langle x_1, x_2, \ldots, x_n\rangle$, we denote the subsequence starting at $x_i$ and ending at $x_j$ by $X_{ij} = \langle x_i, x_{i + 1}, \ldots, x_j \rangle$.

Theorem (Optimal substructure of an LPS)

Let $X = \langle x_1, x_2, \ldots, x_n \rangle$ be the input sequence, and let $Z = \langle z_1, z_2, \ldots, z_m \rangle$ be any $\text{LPS}$ of $X$.

  1. If $n = 1$, then $m = 1$ and $z_1 = x_1$.
  2. If $n = 2$ and $x_1 = x_2$, then $m = 2$ and $z_1 = z_2 = x_1 = x_2$.
  3. If $n = 2$ and $x_1 \ne x_2$, then $m = 1$ and $z_1$ is equal to either $x_1$ or $x_n$.
  4. If $n > 2$ and $x_1 = x_n$, then $m > 2$, $z_1 = z_m = x_1 = x_n$, and $Z_{2, m - 1}$ is an $\text{LPS}$ of $X_{2, n - 1}$.
  5. If $n > 2$ and $x_1\ne x_n$, then $z_1 \ne x_1$ implies that $Z_{1, m}$ is an $\text{LPS}$ of $X_{2, n}$.
  6. If $n > 2$ and $x_1\ne x_n$, then $z_m \ne x_n$ implies that $Z_{1, m}$ is an $\text{LPS}$ of $X_{1, n - 1}$.


Properties 1., 2., and 3. follow trivially from the definition of $\text{LPS}$.

  1. If $n > 2$ and $x_1 = x_n$, then we can choose $x_1$ and $x_n$ as the ends of $Z$ and at least one more element of $X$ as part of $Z$. Thus, it follows that $m > 2$. If $z_1 \ne x_1$, then we could append $x_1 = x_n$ to the ends of $Z$ to obtain a palindrome subsequence of $X$ with length $m + 2$, contradicting the supposition that $Z$ is a longest palindrome subsequence of $X$. Thus, we must have $z_1 = x_1 (= x_n = zm)$. Now, $Z_{2, m - 1}$ is a length-$(m - 2)$ palindrome subsequence of $X_{2, n - 1}$. We wish to show that it is an $\text{LPS}$. Suppose for the purpose of contradiction that there exists a palindrome subsequence $W$ of $X_{2, n - 1}$ with length greater than $m - 2$. Then, appending $x_1 = x_n$ to the ends of $W$ produces a palindrome subsequence of $X$ whose length is greater than $m$, which is a contradiction.
  2. If $z_1 \ne x_1$, then $Z$ is a palindrome subsequence of $X_{2, n}$. If there were a palindrome subsequence $W$ of $X_{2, n}$ with length greater than $m$, then $W$ would also be a palindrome subsequence of $X$, contradicting the assumption that $Z$ is an $\text{LPS}$ of $X$.
  3. The proof is symmetric to (2).

The way that the theorem characterizes longest palindrome subsequences tells us that an $\text{LPS}$ of a sequence contains within it an $\text{LPS}$ of a subsequence of the sequence. Thus, the $\text{LPS}$ problem has an optimal-substructure property.

Step 2: A recursive solution

The theorem implies that we should examine either one or two subproblems when finding an $\text{LPS}$ of $X = \langle x_1, x_2, \ldots, x_n \rangle$, depending on whether $x_1 = x_n$.

Let us define $p[i, j]$ to be the length of an $\text{LPS}$ of the subsequence $X_{ij}$. If $i = j$, the $\text{LPS}$ has length $1$. If $j = i + 1$, then the $\text{LPS}$ has length either $1$ or $2$, depending on whether $x_i = x_j$. The optimal substructure of the $\text{LPS}$ problem gives the following recursive formula:

$$ p[i, j] = \begin{cases} 1 & \text{if $i = j$}, \\ 2 & \text{if $j = i + 1$ and $x_i = x_j$}, \\ 1 & \text{if $j = i + 1$ and $x_i \ne x_j$}, \\ p[i + 1, j - 1] + 2 & \text{if $j > i + 1$ and $x_i = x_j$}, \\ \max(p[i, j - 1], p[i + 1, j]) & \text{if $j > i + 1$ and $x_i \ne x_j$}. \end{cases} $$

Step 3: Computing the length of an LPS

Procedure $\text{LONGEST-PALINDROME}$ takes a sequence $X = \langle x_1, x_2, \ldots, x_n \rangle$ as input. The procedure fills cells $p[i, i]$, where $1 \le i \le n$, and $p[i, i + 1]$, where $1 \le i \le n - 1$, as the base cases. It then starts filling cells $p[i, j]$, where $j > i + 1$. The procedure fills the $p$ table row by row, starting with row $n - 2$ and moving toward row $1$. (Rows $n - 1$ and $n$ are already filled as part of the base cases.) Within each row, the procedure fills the entries from left to right. The procedure also maintains the table $b[1..n, 1..n]$ to help us construct an optimal solution. Intuitively, $b[i, j]$ points to the table entry corresponding to the optimal subproblem solution chosen when computing $p[i, j]$. The procedure returns the $b$ and $p$ tables, $p[1, n]$ contains the length of an $\text{LPS}$ of $X$. The running time of $\text{LONGEST-PALINDROME}$ is clearly $\Theta(n^2)$.

    n = X.length
    let b[1..n, 1..n] and p[0..n, 0..n] be new tables
    for i = 1 to n - 1
        p[i, i] = 1
        j = i + 1
        if x[i] == x[j]
            p[i, j] = 2
            b[i, j] = "↙"
        else p[i, j] = 1
            b[i, j] = "↓"
    p[n, n] = 1
    for i = n - 2 downto 1
        for j = i + 2 to n
            if x[i] == x[j]
                p[i, j] = p[i + 1, j - 1] + 2
                b[i, j] = "↙"
            else if p[i + 1, j]  p[i, j - 1]
                p[i, j] = p[i + 1, j]
                b[i, j] = "↓"
            else p[i, j] = p[i, j - 1]
                b[i, j] = "←"
    return p and b

Step 4: Constructing an LPS

The $b$ table returned by $\text{LONGEST-PALINDROME}$ enables us to quickly construct an $\text{LPS}$ of $X = \langle x_1, x_2, \ldots, x_m\rangle$. We simply begin at $b[1, n]$ and trace through the table by following the arrows. Whenever we encounter a "$\swarrow$" in entry $b[i, j]$, it implies that $x_i = y_j$ are the first and last elements of the $\text{LPS}$ that $\text{LONGEST-PALINDROME}$ found. The following recursive procedure returns a sequence $S$ that contains an $\text{LPS}$ of $X$. The initial call is $\text{GENERATE-LPS}(b, X, 1, X.length, \langle \rangle)$, where $\langle\rangle$ denotes an empty sequence. Within the procedure, the symbol $||$ denotes concatenation of a symbol and a sequence.

GENERATE-LPS(b, X, i, j, S)
    if i > j
        return S
    else if i == j
        return S || x[i]
    else if b[i, j] == "↙"
        return x[i] || GENERATE-LPS(b, X, i + 1, j - 1, S) || x[i]
    else if b[i, j] == "↓"
        return GENERATE-LPS(b, X, i + 1, j, S)
    else return GENERATE-LPS(b, X, i, j - 1, S)