Symmetric boolean functions as Zhegalkin polynomials

Question

Let \$\mathbb{B} = \mathbb{Z}_2 = \{0, 1\}\$ be the set of booleans. A symmetric boolean function in \$n\$ arguments is a function \$f_S : \mathbb{B}^n \to \mathbb{B}\$ that checks if the number of its true arguments is in \$S\$, i. e. a function \$f_S\$ such that

$$f_S(\alpha_1, \alpha_2, \ldots, \alpha_n) = \left(\sum_{1 \le i \le n}\alpha_{i}\right) \in S$$

or equivalently

$$f_S(\alpha_1, \alpha_2, \ldots, \alpha_n) = |\{ i : \alpha_i \}| \in S. $$

These functions are called symmetric boolean function because they are all the boolean function that do not depend on argument order.

The set of boolean numbers forms a field, conventionally denoted \$GF(2)\$, where multiplication is logical and (\$\wedge\$) and addition is exclusive or (\$\oplus\$). This field has the properties that \$-\alpha = \alpha\$, \$\alpha / \beta = \alpha \wedge \beta\$ (for all \$\beta \neq 0\$), \$\alpha \oplus \alpha = 0\$ and \$\alpha \wedge \alpha = \alpha\$ for all \$\alpha\$ in addition to the field properties. For convenience, I am going to write \$\oplus\$ as \$+\$ and leave out \$\wedge\$ in the next paragraphs. For example, I will write \$\alpha \oplus \beta \oplus \alpha \wedge \beta\$ as \$\alpha + \beta + \alpha\beta\$.

Every boolean function can be expressed as a polynomial over the field of boolean numbers. Such a polynomial is called a Zhegalkin polynomial, the representation is canonical, i. e. there is exactly one polynomial that represents a given boolean function. Here are some examples for boolean functions and their corresponding Zhegalkin polynomials:

• \$\alpha \vee \beta = \alpha + \beta + \alpha\beta\$
• \$\alpha \wedge \neg\beta = \alpha + \alpha\beta\$
• \$(\alpha \to \beta) \oplus \gamma = 1 + \alpha + \gamma + \alpha\beta\$
• \$\neg(\alpha \wedge \beta) \vee \gamma = 1 + \alpha\beta + \alpha\beta\gamma\$
• \$f_{\{1, 4\}}(\alpha, \beta, \gamma, \delta) = \alpha + \beta + \gamma + \delta + \alpha\beta\gamma + \alpha\beta\delta + \alpha\gamma\delta + \beta\gamma\delta + \alpha\beta\gamma\delta\$
  where the triple terms \$(\alpha\beta\gamma + \alpha\beta\delta + \alpha\gamma\delta + \beta\gamma\delta)\$ are required because otherwise we would have \$f(1,1,1,0) = 1\$, and \$3 \not\in \{1, 4\}\$.

One observation we can make is that due to symmetry, the Zhegalkin polynomial for a symmetric boolean function always has the form

$$f_S(\alpha_1, \alpha_2, \ldots, \alpha_n) = \sum_{i \in T_{S,n}}t_{n,i}$$

where \$\sum\$ denotes summation with exclusive or and \$t_{n, i}\$ is the sum of all product terms that use \$i\$ out of the \$n\$ arguments to \$f_S\$ and \$T_{S,n} \subseteq \{0, 1, \ldots, n\}\$ is a set describing those \$t_{n, i}\$ we need to construct \$f_S(\alpha_1, \alpha_2, \ldots, \alpha_n)\$. Here are some examples for \$t_{n, i}\$ so you can get an idea what they look like:

• \$t_{4, 1} = \alpha + \beta + \gamma + \delta\$
• \$t_{3, 2} = \alpha\beta + \alpha\gamma + \beta\gamma\$
• \$t_{4, 4} = \alpha\beta\gamma\delta\$
• for all \$n\$: \$t_{n, 0} = 1\$

We can thus describe an \$f_S\$ in \$n\$ arguments completely by \$n\$ and \$T_{S, n}\$. Here are some test cases:

• \$T_{\{1, 4\}, 4} = \{1, 3, 4\}\$
• \$T_{\{2, 3, 4\}, 4} = \{2, 4\}\$
• \$T_{\{2\}, 4} = \{2, 3\}\$
• \$T_{\{1, 2\}, 2} = \{1, 2\}\$ (logical or)
• \$T_{\{2, 3\}, 3} = \{2\}\$ (median function)

The task in this challenge is to compute \$T_{S, n}\$ given \$S\$ and \$n\$. You can choose an appropriate input and output format as you like. This challenge is code-golf, the shortest solution in octets wins. As a special constraint, your solution must run in polynomial time \$O(n^k)\$ where \$n\$ is the \$n\$ from the input.

Answer 1

Dyalog APL, 15 bytes

{2|⍺⌹⍉∘.!⍨0,⍳⍵}

Matrix division, basically. Takes input like 0 0 1 1 1 {2|⍺⌹⍉∘.!⍨0,⍳⍵} 4 for T₄({2,3,4}).

{              }        Dyadic function:
              ⍵         Right argument          4
             ⍳          Index vector            1 2 3 4
           0,           Append 0                0 1 2 3 4
      ∘.                Outer product           ——————————
          ⍨             |With itself              1 0 0 0 
        !               |By binomial coeff.       2 1 0 0
     ⍉                  Matrix transpose          3 3 1 0
                        Result at this step: ->   4 6 4 1
   ⍺                    Left argument
    ⌹                   Matrix division (A⌹B is A^-1 B, not A B^-1)
                        (implicitly converts ⍺ to column vector)
 2|                     Modulo 2

Since the transposed matrix of binomial coefficients is lower triangular with determinant 1, its inverse will only contain integers. So to matrix-divide over F₂ we can just matrix-divide over the reals, then mod by 2.

Matrix division is an O(n^3) operation, and calculating each of the O(N^2) binomial coefficients is well below O(n^3), so this runs in O(N^5).

Try it on TryAPL.

Answer 2

CJam (25 24 bytes)

{_,{1$_,,2$~f&:!.*:^t}/}

This is an anonymous function which takes an array of 0 or 1 which serves as an indicator function for \$S\$ and returns an array of 0 or 1 which serves as an indicator function for \$T\$. E.g. to find \$T_{\{1, 4\}, 4}\$ you pass [0 1 0 0 1] and it returns [0 1 0 1 1].

Online demo

Explanation

Recall that \$t_{n,i}\$ is the sum of all product terms that use \$i\$ out of the \$n\$ arguments to \$f_S\$. Suppose that \$s\$ of the \$n\$ arguments are true. Then we have three cases for \$t_{n,i}(\alpha_0, \ldots, \alpha_{n-1})\$:

• \$i > s\$ : none of the terms is true, so \$t_{n, i}(\alpha_0, \ldots, \alpha_{n-1}) = 0\$
• \$i = s\$ : exactly one of the terms is true, so \$t_{n, i}(\alpha_0, \ldots, \alpha_{n-1}) = 1\$
• \$i < s\$ : there are \$\binom{s}{i}\$ terms which are true, so \$t_{n, i}(\alpha_0, \ldots, \alpha_{n-1}) = \binom{s}{i} \bmod 2\$

If we let \$\sigma\$ be the indicator function for \$S\$ and \$\tau\$ be the indicator function for \$T_{S, n})\$, we therefore get the matrix identity

$$\begin{pmatrix}1 & 0 & 0 & \ldots & 0 \\ \binom{1}{0} & 1 & 0 & \ldots & 0 \\ \binom{2}{0} & \binom{2}{1} & 1 & \ldots & 0 \\ \vdots & \vdots & \vdots & \ddots & 0 \\ \binom{n}{0} & \binom{n}{1} & \binom{n}{2} & \ldots & 1 \end{pmatrix} \begin{pmatrix}\tau(0) \\ \tau(1) \\ \tau(2) \\ \vdots \\ \tau(n) \end{pmatrix} = \begin{pmatrix}\sigma(0) \\ \sigma(1) \\ \sigma(2) \\ \vdots \\ \sigma(n) \end{pmatrix} $$

which is dead easy to solve because the coefficient matrix is already in row-reduced echelon form.

Note that this is quite a well-known matrix, because it's the binomial coefficients modulo 2, which famously forms a Sierpiński triangle. I calculate them using Lucas' theorem and bit-twiddling as 2$~f&:!

Answer 3

R, 115 bytes

function(S,n)which(abs(solve(sapply(1:n,function(k)(sapply(1:n,function(j)choose(j,k)%%2))),is.element(1:n,S)))==1)

The key is realizing that the sum of t's that evaluate to a T value is just the choose function. A slightly ungolfed version that doesn't calculate in polynomial time is below, but illustrates the derivation of the t's a little better (just run the middle line).

function(S,n)which(abs(solve(
sapply(1:n,function(k)(sapply(1:n,function(j)sum(apply(combn(c(rep(T,j),rep(F,max(0,k-j))),k),2,all))%%2))),
is.element(1:n,S)))==1)