Algorithm

Squarefree factorisation

Cost: $O(M(n) \log n)$ field operations for a polynomial of degree $n$ with fast gcds (a constant number of gcds of total degree $O(n)$ ), plus $p$ -th roots of at most $n$ coefficients in characteristic $p$ .

A polynomial $f$ over a field $K$ is squarefree if it is not divisible by the square of a nonconstant polynomial. Every nonzero $f$ has a unique squarefree factorisation

f = c \prod_{i \ge 1} g_i^{\,i}

with $c \in K^\times$ and monic, squarefree, pairwise coprime $g_i$ , where $g_i$ is the product of the irreducible factors of multiplicity exactly $i$ . It is the first step of every factorisation algorithm (Factoring polynomials over finite fields, Factoring polynomials over the integers) and of root finding, and it costs only gcds and exact divisions. FLINT returns the nontrivial $g_i$ together with their exponents $i$ , made monic over a field and primitive with positive leading coefficient over $\mathbb{Z}$ .

The tool is the derivative. If $f = \prod_g g^{e_g}$ over the monic irreducible $g$ , then

f' = \sum_g e_g\, g' \, \frac{f}{g}.

Lemma 1. Suppose that every irreducible $g$ is separable ( $\gcd(g, g') = 1$ ), which holds in characteristic 0 and over finite fields. Then the multiplicity of $g$ in $\gcd(f, f')$ is $e_g - 1$ if $e_g \ne 0$ in $K$ , and $e_g$ if $e_g = 0$ in $K$ (that is, if the characteristic $p$ divides $e_g$ ).

Proof. In the sum, every term except the one for $g$ is divisible by $g^{e_g}$ . The term for $g$ is divisible by exactly $g^{e_g - 1}$ if $e_g \ne 0$ in $K$ , since $g \nmid e_g g'$ , and vanishes otherwise. Hence $g^{e_g - 1} \| f'$ in the first case and $g^{e_g} \mid f'$ in the second.

Characteristic zero: Yun's algorithm#

In characteristic 0, Lemma 1 gives $\gcd(f, f') = \prod_i g_i^{\,i-1}$ , so $f / \gcd(f, f') = \prod_i g_i$ is the squarefree part of $f$ . The multiplicities can be separated by repeating this, but Yun's algorithm [Yun1976] does it with gcds whose degrees add up to $O(n)$ .

Algorithm 1: Yun's algorithm

Input: $f$ of degree $n \ge 1$ over a field of characteristic 0, monic (or primitive over a UFD).

Output: the pairs $(g_i, i)$ with $g_i \ne 1$ .

$d \gets \gcd(f, f')$ . If $d = 1$ , return $(f, 1)$ .
$v \gets f / d$ , $w \gets f' / d$ , $i \gets 1$ .
Loop:
1. $s \gets w - v'$ . If $s = 0$ , output $(v, i)$ if $v \ne 1$ and stop.
2. $u \gets \gcd(v, s)$ ; output $(u, i)$ if $u \ne 1$ .
3. $v \gets v / u$ , $w \gets s / u$ , $i \gets i + 1$ .

Theorem 2. Algorithm 1 is correct. At the start of iteration $i$ ,

v = \prod_{j \ge i} g_j, \qquad w = \sum_{j \ge i} (j - i + 1)\, g_j'\, \frac{v}{g_j},

and the gcd of iteration $i$ is $g_i$ .

Proof. For $i = 1$ , $d = \prod_j g_j^{\,j-1}$ gives $v = \prod_j g_j$ and $w = f'/d = \sum_j j g_j' v / g_j$ . Given the invariant,

s = w - v' = \sum_{j > i} (j - i)\, g_j' \frac{v}{g_j}.

Every term is divisible by $g_i$ . For $j > i$ with $g_j \ne 1$ , all terms but the $j$ -th are divisible by $g_j$ , and the $j$ -th is not, since $j - i \ne 0$ and $\gcd(g_j, g_j') = 1$ . Hence $\gcd(v, s) = g_i$ , and $v / g_i$ , $s / g_i$ satisfy the invariant for $i + 1$ . If $s = 0$ , then $g_j = 1$ for all $j > i$ , so $v = g_i$ .

The degrees of the polynomials in iteration $i$ are at most $\deg v \le \sum_{j \ge i} \deg g_j$ , and $\sum_i \sum_{j \ge i} \deg g_j = \sum_j j \deg g_j = n$ . With gcds in $O(M(m) \log m)$ for degree $m$ (Half-GCD), the whole algorithm therefore costs $O(M(n) \log n)$ , the same as one gcd of degree $n$ up to a constant factor [Yun1976] [GG2003, §14.6].

Finite fields#

In characteristic $p$ the derivative of $g^{e}$ vanishes when $p \mid e$ , and $f' = 0$ if and only if $f \in K[x^p]$ . Over a perfect field, in particular over $\mathbb{F}_q$ with $q = p^k$ , such an $f$ is a $p$ -th power: if $f = \sum_i c_i x^{ip}$ , then $f = (\sum_i c_i^{1/p} x^i)^p$ , where $c^{1/p} = c^{p^{k-1}}$ (and $c^{1/p} = c$ in a prime field). FLINT uses the following algorithm, the classical gcd iteration [Mus1971] [Knu1997, §4.6.2] combined with $p$ -th roots [GG2003, §14.6].

Algorithm 2: Squarefree factorisation over

\mathbb{F}_q

Input: $f$ of degree $n \ge 1$ over $\mathbb{F}_q$ .

Output: the pairs $(g_i, i)$ with $g_i \ne 1$ , the $g_i$ monic.

If $f' = 0$ : compute $h$ with $h^p = f$ by taking $p$ -th roots of the coefficients, factor $h$ recursively, multiply all exponents by $p$ and return.
$d \gets \gcd(f, f')$ . If $d = 1$ , return $(f, 1)$ (made monic).
$v \gets f / d$ , $i \gets 1$ . While $\deg v \ge 1$ :
1. $s \gets \gcd(v, d)$ ; $w \gets v / s$ ; output $(w, i)$ if $w \ne 1$ .
2. $d \gets d / s$ , $v \gets s$ , $i \gets i + 1$ .
If $d \ne 1$ : $d$ is a $p$ -th power; factor $d^{1/p}$ recursively and output its factors with exponents multiplied by $p$ .

Theorem 3. Algorithm 2 is correct. The exponents output in step 3 are not divisible by $p$ , those output in step 4 are, and all are distinct.

Proof. Write $f = \prod_g g^{e_g}$ . By Lemma 1, after step 2,

d = \prod_{p \nmid e_g} g^{e_g - 1} \prod_{p \mid e_g} g^{e_g}, \qquad v = \prod_{p \nmid e_g} g.

By induction on $i$ , at the start of iteration $i$ we have $v = \prod_{p \nmid e_g,\, e_g \ge i} g$ and $d = \prod_{p \nmid e_g,\, e_g \ge i} g^{e_g - i} \prod_{p \mid e_g} g^{e_g}$ . Since $v$ is squarefree and contains only factors with $p \nmid e_g$ , $s = \gcd(v, d) = \prod_{p \nmid e_g,\, e_g > i} g$ , so $w = v/s$ is the product of the factors of multiplicity exactly $i$ (which is 1 if $p \mid i$ ), and the updated $v$ and $d$ satisfy the invariant for $i + 1$ . When $v = 1$ , the remaining $d = \prod_{p \mid e_g} g^{e_g}$ is a $p$ -th power. In step 1 and step 4 the recursion is applied to a polynomial of degree at most $n/p$ , whose squarefree factorisation, with exponents multiplied by $p$ , gives the factors of $f$ with multiplicities divisible by $p$ .

As in Yun's algorithm, the degrees of the gcds in step 3 add up to at most $n$ , so the cost is $O(M(n) \log n)$ , plus at most $n/p$ $p$ -th roots of field elements per level of recursion.

The algorithm gives the factorisation with distinct exponents, not necessarily with exponents $1, 2, 3, \ldots$ in order; FLINT returns the factors in the order in which they are found.

Squarefree part and squarefreeness test#

The squarefree part $\prod_i g_i$ is $f / \gcd(f, f')$ (up to a unit) in characteristic 0, while in characteristic $p$ it is computed from the squarefree factorisation. To test whether $f$ is squarefree it suffices, over a perfect field, to test $\gcd(f, f') = 1$ : if $f' = 0$ and $\deg f \ge 1$ then $f$ is a $p$ -th power, hence not squarefree.

Use in FLINT#

gr_poly_factor_squarefree implements Algorithm 1 for unique factorisation domains of characteristic 0, after making $f$ monic (over a field) or removing its content (otherwise), and Algorithm 2 for finite fields, where the $p$ -th roots of the coefficients are computed with gr_fq_pth_root. gr_poly_squarefree_part and gr_poly_is_squarefree use the methods of the previous section. The finite field modules wrap the generic code; fmpz_poly_factor_squarefree has its own implementation of Yun's algorithm with integer gcds, and the squarefreeness tests for nmod_poly compute $\gcd(f, f')$ directly.

int gr_poly_factor_squarefree(gr_ptr c, gr_poly_vec_t fac, fmpz_vec_t exp, const gr_poly_t poly, gr_ctx_t ctx)

int gr_poly_squarefree_part(gr_poly_t res, const gr_poly_t poly, gr_ctx_t ctx)

truth_t gr_poly_is_squarefree(const gr_poly_t f, gr_ctx_t ctx)

void nmod_poly_factor_squarefree(nmod_poly_factor_t res, const nmod_poly_t f)

int nmod_poly_is_squarefree(const nmod_poly_t f)

void fmpz_mod_poly_factor_squarefree(fmpz_mod_poly_factor_t res, const fmpz_mod_poly_t f, const fmpz_mod_ctx_t ctx)

int fmpz_mod_poly_is_squarefree(const fmpz_mod_poly_t f, const fmpz_mod_ctx_t ctx)

void fmpz_poly_factor_squarefree(fmpz_poly_factor_t fac, const fmpz_poly_t F)

Referenced from#

References#

[Yun1976] D. Y. Y. Yun. On square-free decomposition algorithms. Proceedings of the 1976 ACM Symposium on Symbolic and Algebraic Computation (SYMSAC '76), 26–35. https://doi.org/10.1145/800205.806320
[GG2003] J. von zur Gathen and J. Gerhard, Modern Computer Algebra, second edition, Cambridge University Press (2003)
[Mus1971] D. R. Musser. Algorithms for polynomial factorization. PhD thesis, University of Wisconsin, 1971.
[Knu1997] Knuth, D. E. The Art of Computer Programming, volume 2: Seminumerical algorithms, 1997

Cite this page

The FLINT team, “Squarefree factorisation,” FLINT: Fast Library for Number Theory, documentation, version 3.7.0-dev, 2026. https://flintlib.org/doc/algorithms/polynomials/squarefree.html (generated 2026-11-02).

@misc{flint-doc-algorithms-polynomials-squarefree,
  author       = {{The FLINT team}},
  title        = {{Squarefree factorisation}},
  howpublished = {FLINT: Fast Library for Number Theory, documentation, version 3.7.0-dev},
  year         = {2026},
  url          = {https://flintlib.org/doc/algorithms/polynomials/squarefree.html},
  note         = {Generated 2026-11-02}
}

Edit this page · PDF version · Built with fdoc from 0a4634bf95