This program solves a version of the subset-sum problem: given a set of non-negative integers \$S=\{a_1,a_2,\cdots,a_n\}\$ (they may not be distinct) and a non-negative integer \$s\$, decide whether there exists a subset of \$S\$ (let's call it \$T\$) such that \$\sum\limits_{b\in T}b=s\$.

Let the sum of all elements in \$S\$ be \$t\$ (\$t=\sum\limits_{j=1}^n a_j\$). Consider the polynomial $$f(x)=\prod\limits_{j=1}^{n}\left(1+x^{a_j}\right)$$ which can be expanded to the form $$f(x)=c_0+c_1x^1+c_2x^2+\cdots+c_tx^t\quad\left(t=\sum\limits_{i=1}^n a_i\right)$$ It is easy to see that the problem is equivalent to deciding whether \$c_s=0\$. Let \$m=\max\{s,t-s\}+1\$ and \$\zeta=e^{2\pi i/m}\$, then we have $$\sum\limits_{k=0}^{m-1}{\left(\zeta^p\right)}^k=\begin{cases} 0,&\text{if }m\text{ divides }k\\ m,&\text{if }m\text{ does not divide }k \end{cases}$$ where \$p\in\mathbb{Z}\$. Therefore we can isolate the coefficient \$c_s\$ in \$f(x)\$. Consider $$g(x)=x^{-s}f(x)=c_0x^{-s}+c_1x^{1-s}+\cdots+c_s+\cdots+c_tx^{t-s}$$We have $$\sum\limits_{k=0}^{m-1}g\left(\zeta^k\right)=\sum\limits_{j=0}^{t}c_j\sum\limits_{k=0}^{m-1}{\left(\zeta^k\right)}^{j-s}=mc_s$$ Therefore $$c_s=\sum\limits_{k=0}^{m-1}\zeta^{-ks}f\left(\zeta^k\right)$$ So we only need to compute \$c_s\$ and see if \$|c_s|>\frac{1}{2}\$ (we do not directly compare \$c_s\$ to zero since there are rounding errors during floating-point computation). When we compute \$f\left(\zeta^k\right)\$, we do not directly pass \$\zeta^k\$ as the argument; instead, we pass its argument \$2\pi ik/m\$, and then we can compute \${\left(\zeta^k\right)}^{a_j}\$ with the Euler formula $$e^{2\pi ia_j k/m}=\cos(2\pi a_j k/m)+i\sin(2\pi a_j k/m)$$ without directly computing the exponent of a complex number.

Here is the Python code:

import math
from typing import *

class SubsetSumSolver:
    def __init__(self, data: List[int], desired_sum: int):
        self.data = data # a
        self.desired_sum = desired_sum # s
        self.len_data = len(data) # n
        self.sum_data = sum(data) # t
    def evaluate_polynomial(self, theta: float) -> complex:
        result = 1.
        for element in self.data: # element: a_j
            arg = element * theta
            result *= 1 + math.cos(arg) + 1j * math.sin(arg)
        return result
    def solve(self) -> bool:
        if self.desired_sum > self.sum_data:
            return False
        if self.desired_sum == self.sum_data:
            return True
        if self.desired_sum == 0:
            return True
        order = max(self.desired_sum,
            self.sum_data - self.desired_sum) + 1 # m
        result = 0.
        double_pi = 2 * math.pi
        for k in range(order):
            theta = double_pi * k / order
            f_result = self.evaluate_polynomial(theta)
            theta_s = -double_pi * k * self.desired_sum / order
            term_s = math.cos(theta_s) + 1j * math.sin(theta_s)
            result += term_s * f_result
        result /= order
        return abs(result) > 0.5

a = [2, 3, 5, 6, 8, 13, 27, 36]
s = 67 # test data
solver = SubsetSumSolver(a, s)
print(solver.solve()) # True

I have a question here: If \$n\$ is very large, will it be possible that this code doesn't work because of huge rounding errors?


1 Answer 1


Don't use wildcard imports

... unless you actually use all imported names. You only use List from typing, so just import that. Or, even better, don't and use list for type hinting directly.

Stop writing classes

SubsetSumSolver consists of three methods: __init__, solve and a helper function evaluate_polynomial. You can implement the solver as a free function called subset_sum taking data and desired_sum as arguments. Your class does not store any state relevant to the solved problem. Once the calculation is done and the method solve has returned, the state of the class is meaningless. This should be a hint, that a class was unnecessary in the first place. You can also see that by your invocation of the class: Instatiate, method call, destroy (or, in Python, garbage-collect) is an anti-pattern in my book.


... is hard. data is one of the worst names for a variable, since it does not convey at all what it represents. All variables in any program in any language contain "data". So calling a variable like that is meaningless. In your case data is the set of integers that we want to check whether they fulfill the subset sum problem. So let's call it integers.

Use fitting types.

Subset-sum, as the name suggests, operates on sets. Why do you then pass a list of ints? Is it meaningful to pass a list like [1, 1, 1, 1, 1] to a problem operating on sets? If your answer is no, then adjust your program to use the fitting datatype, namely set.

Float precision

As always with floating point numbers, precision will eventually become an issue, yes.

  • \$\begingroup\$ Thanks for your answer. Actually, the subset-sum problem does not necessarily require the input integers to be distinct, so the "set" here is not the mathematical concept "set" (maybe the wording of my question was a little ambiguous). Therefore, I think it's appropriate to pass a list of ints instead of a set of ints, and yes, a list like [1, 1, 1, 1, 1] is possible. \$\endgroup\$
    – Soha
    Commented Jan 4, 2023 at 9:24
  • \$\begingroup\$ @Soha yes, the generic term from computational complexity theory can also be applied to other data containers. But in your question you explicitly used the word "set" and the mathematical definition of a set. So, at any rate, you should reconsider what kind of problem you actually want to solve here. If you use the word "set" in combination with "{...}", I automatically am in ZF(C) mode. \$\endgroup\$ Commented Jan 4, 2023 at 9:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.