Formalizing Mason–Stothers Theorem and its Corollaries in Lean 4

Abstract

The ABC conjecture implies many conjectures and theorems in number theory, including the celebrated Fermat’s Last Theorem. Mason–Stothers Theorem is a function field analogue of the ABC conjecture that admits a much more elementary proof with many interesting consequences, including a polynomial version of Fermat’s Last Theorem. While years of dedicated effort are expected for a full formalization of Fermat’s Last Theorem, the simple proof of Mason–Stothers Theorem and its corollaries calls for an immediate formalization.

We formalize an elementary proof of by Snyder in Lean 4, and also formalize many consequences of Mason–Stothers, including (i) non-solvability of Fermat–Cartan equations in polynomials, (ii) non-parametrizability of a certain elliptic curve, and (iii) Davenport’s Theorem. We compare our work to existing formalizations of Mason–Stothers by Eberl in Isabelle and Wagemaker in Lean 3 respectively. Our formalization is based on the mathlib4 library of Lean 4, and is currently being ported back to mathlib4.

1 Introduction

In 1985, Oesterlé and Masser proposed the ABC conjecture [15, 16]:

Conjecture 1 (ABC conjecture)

For every positive real number $\varepsilon>0$, there exist only finitely many triples of coprime integers $(a,b,c)$ such that $a+b=c$ and

$$c>\mathrm{rad}(abc)^{1+\varepsilon}.$$

Here, $\mathrm{rad}(n)=\prod_{p|n}p$ is the product of all prime factors of $n$.

The conjecture implies many deep theorems or conjectures in number theory. For example, Fermat’s Last Theorem (FLT) for exponent $n\geq 6$ is a direct corollary of an explicit quantitative version of the ABC conjecture [9], while the currently known proof by Wiles [25] and Taylor-Wiles [22] requires heavy machinery (See [3] for detailed surveys). Also, Roth’s theorem [17] and Faltings’ theorem [8] both follow from the ABC conjecture [23]; note that the proof of each theorem earned its corresponding author a Fields medal.

In number theory, there is a strong analogy between number fields $K/\mathbb{Q}$ (e.g., rational numbers $\mathbb{Q}$) and function fields $k(X)/k$ of a smooth projective curve over $k$ (e.g., rational function field $k(t)$). Under this analogy, profound statements on integers $\mathbb{Z}$, such as the Riemann Hypothesis, the Birch and Swinnerton-Dyer conjecture, or the Langlands program, have analogous statements [5, 6, 11, 21] on the integral ring $k[t]$ of the rational function field $k(t)$. The analogs of such conjectures often turn out to be true and easier to prove in general.

In this line, Stothers proved the polynomial analog of the ABC conjecture in 1981 [20], and Mason rediscovered it in a more general form in 1983 [14], even before Osterlé and Masser proposed the ABC conjecture.

Definition 1

Let $k$ be any field. For any nonzero $f\in k[X]$, define the radical $\mathrm{rad}(f)$ of $f$ as the product of all irreducible monic factors of $f$ not counting multiplicity.

Theorem 1.1 (Mason–Stothers)

Let $k$ be any field and $a,b,c\in k[t]$ be non-zero, pairwise coprime polynomials satisfying $a+b+c=0$. Then we either have $a^{\prime}=b^{\prime}=c^{\prime}=0$ where $f^{\prime}$ denotes the (formal) derivative of $f\in k[t]$ by $t$, or

$$\max\{\deg(a),\deg(b),\deg(c)\}<\deg(\mathrm{rad}(abc)).$$

Mason and Stothers proved the theorem using algebro-geometric methods, and subsequently Snyder discovered a short and purely elementary proof [19]. Like the ABC conjecture, the Mason–Stothers theorem has a lot of interesting consequences as the following.

1. A polynomial version of Fermat’s Last Theorem. More generally, the non-solvability of the Fermat-Catalan equation $ua^{p}+vb^{q}+wc^{r}=0$ over $a,b,c\in k[t]$ with nonzero constants $u,v,w\in k$ and powers $p,q,r\in\mathbb{N}$ satisfying $1/p+1/q+1/r\leq 1$ (Theorem 2.1).

2. Non-parametrizability of the elliptic curve $y^{2}=x^{3}+1$ by rational functions $x,y\in k(t)$ (Theorem 2.2).

3. Davenport’s theorem, initially conjectured by Birch et al [2], that for any polynomials $f,g\in\mathbb{C}[t]$ we have $\deg{(f^{3}-g^{2})}\geq\frac{1}{2}\deg{f}+1$ (Theorem 2.3).

We give a fully documented Lean 4 formalization of the Mason–Stothers theorem on fields of arbitrary characteristic. Also, we formalize the aforementioned corollaries of the theorem to demonstrate the power of Mason–Stothers theorem (Theorem 2.1, 2.2, and 2.3). The code is hosted in

https://github.com/seewoo5/lean-poly-abc.

and is currently being ported to mathlib4 (see Appendix 0.A).

The Mason–Stothers theorem was already formalized by Eberl in Isabelle [7] and Wagemaker in Lean 3 [24]. We give a detailed comparison between our work and theirs in Section 7. In short, our formalization works for arbitrary characteristic (see Section 7.2), is compatible with the mathlib4 library of Lean 4, includes variants of Mason–Stothers (e.g. Theorem 6.1), and formalizes several interesting corollaries including the polynomial FLT with a slightly stronger conclusion than existing formalization (see Section 7.1).

2 Statements of the Theorem and its Corollaries

The precise statement of Mason–Stothers theorem we formalize is Theorem 1.1, which holds for arbitrary field $k$. Note that most literature either assumes that $k$ is of characteristic zero or is algebraically closed [20, 19].

If $k$ has characteristic zero, the condition $a^{\prime}=b^{\prime}=c^{\prime}=0$ in Theorem 1.1 is equivalent to $a,b,c$ being constants. If $k$ has characteristic $p>0$, then the condition $f^{\prime}=0$ for $f=a,b,c$ is equivalent to $f(t)=f_{0}(t^{p})$ being a polynomial of $t^{p}$. Indeed,

$$(a,b,c)=(-1,-x^{p},1+x^{p})=(-1,-x^{p},(1+x)^{p})$$

is an example satisfying $a+b+c=0$ and $a^{\prime}=b^{\prime}=c^{\prime}=0$, but

$$\max\{\deg(a),\deg(b),\deg(c)\}+1=p+1>2=\deg(\mathrm{rad}(abc)).$$

We now state the corollaries of the Mason–Stothers theorem mentioned in Section 1 precisely. Their proofs can be found in Section 6.

The Fermat–Catalan conjecture is a generalization of Fermat’s Last Theorem stating that the equation $a^{p}+b^{q}=c^{r}$ has only finitely many solutions $(a,b,c,p,q,r)$ in positive integers satisfying $1/p+1/q+1/r<1$ [9]. The following Theorem 2.1 is a polynomial variant which is true.

Theorem 2.1 (Fermat–Catalan Conjecture for Polynomials)

Let $k$ be any field. Let $p,q,r\geq 1$ be integers not divisible by the characteristic of $k$ such that $1/p+1/q+1/r\leq 1$. Let $u,v,w\in k$ be arbitrary nonzero constants. Then any triple $(a,b,c)$ of nonzero and pairwise coprime polynomials in $k[t]$ satisfying $ua^{p}+vb^{q}+wc^{r}=0$ should be constants $a,b,c\in k$.

Let $u=v=-w=1$ and $p=q=r=n\geq 3$ in Theorem 2.1 to recover the Fermat’s Last Theorem for polynomials.

Corollary 1 (Fermat’s Last Theorem for Polynomials)

Let $k$ be any field. Let $n\geq 3$ be any integer not divisible by the characteristic of $k$. Then any triple $(a,b,c)$ of nonzero and pairwise coprime polynomials in $k[t]$ satisfying $a^{n}+b^{n}=c^{n}$ should be constants $a,b,c\in k$.

Using Theorem 2.1, we can obtain the following corollary.

Theorem 2.2 (Non-parametrizablility of an Elliptic Curve)

Let $k$ be a field of characteristic $\neq 2,3$. If rational functions $f(t),g(t)\in k(t)$ satisfy $g(t)^{2}=f(t)^{3}+1$, then both $f(t)$ and $g(t)$ are constants in $k$.

In other words, the elliptic curve defined by the Weierstrass equation $y^{2}=x^{3}+1$ is not parametrizable by non-constant rational functions in $k(t)$.

Another interesting corollary of Mason–Stothers theorem is the following theorem by Davenport [4] initially conjectured [2] by Birch et al. This theorem motivated Stothers’ proof of the Mason–Stothers theorem [20].

Theorem 2.3 (Davenport)

Let $k$ be a field of characteristic zero. Let $f(t),g(t)\in k[t]$ be non-constant polynomials such that $f^{3}\neq g^{2}$. Then

$$\deg(f^{3}-g^{2})\geq\frac{1}{2}\deg(f)+1.$$

3 Mathematical Proof of Mason–Stothers Theorem

We summarize the proof of Mason–Stothers theorem (Theorem 1.1) in Lemmermeyer’s note [12] that we formalize.

Definition 2

The Wronskian of two polynomials $a,b\in k[t]$ is $W(a,b)=ab^{\prime}-a^{\prime}b$, where $a^{\prime}$ is the (formal) derivative of $a$ with respect to $t$.

From $a+b+c=0$ we can check that the values $W(a,b)$, $W(b,c)$, and $W(c,a)$ are all equal. Denote the common value as $W$.

Next, we observe the following property.

Lemma 1

For a nonzero polynomial $a\in k[t]$, $a/\mathrm{rad}(a)$ divides $a^{\prime}$.

For a proof, we can use the prime factorization of $a$. Let $a=up_{1}^{e_{1}}p_{2}^{e_{2}}\cdots p_{m}^{e_{m}}$ be a factorization with unit $u$ and primes $p_{i}\in k[x]$ of exponents $e_{i}>0$. Then the product rule of derivative gives $a^{\prime}=\sum_{i=1}^{m}ue_{i}p_{i}^{\prime}p_{1}^{e_{1}}p_{2}^{e_{2}}\cdots p_{i}^{e_{i}-1}\cdots p_{m}^{e_{m}}$ which is divisible by $a/\mathrm{rad}(a)=p_{1}^{e_{1}-1}p_{2}^{e_{2}-1}\cdots p_{m}^{e_{m}-1}$. An immediate corollary is that

Lemma 2

For any nonzero $a\in k[t]$, $a/\mathrm{rad}(a)$ divides $W(a,b)=ab^{\prime}-a^{\prime}b$

because $a/\mathrm{rad}(a)$ divides both $a$ and $a^{\prime}$.

The pairwise coprime polynomials $a/\mathrm{rad}(a)$, $b/\mathrm{rad}(b)$, and $c/\mathrm{rad}(c)$ all divide $W$. So their product $abc/\mathrm{rad}(abc)$ should also divide $W$. This is the key step of the proof. Divide the case into whether $W$ is zero or not. If $W=0$, then $W(a,b)=0$ implies $ab^{\prime}=a^{\prime}b$, and since $a$ and $b$ are coprime $a$ divides $a^{\prime}$ and so $a^{\prime}=0$. Likewise, from $W=0$ we also get $b^{\prime}=c^{\prime}=0$.

Now assume $W\neq 0$. Then $abc/\mathrm{rad}(abc)$ dividing $W$ implies

$$\deg(a)+\deg(b)+\deg(c)-\deg(\mathrm{rad}(abc))=\deg\left(\frac{abc}{\mathrm{rad}(abc)}\right)$$

$$\leq\deg W=\deg W(a,b)<\deg(a)+\deg(b).$$

The first inequality follows from divisibility, and the second inequality follows from the definition of Wronskian and $a\neq 0$. Hence we have $\deg(c)<\deg\mathrm{rad}(abc)$. The same argument with $W=W(b,c)$ and $W=W(c,a)$ gives

$$\max\{\deg(a),\deg(b),\deg(c)\}+1\leq\deg(\mathrm{rad}(abc)).$$

4 Basic Definitions

Now we explain our formalization of the proof of Mason–Stothers in Section 3. The frst step to develop an interface for the radical (Definition 1) and Wronskian (Definition 2) of polynomials.

4.1 Wronskian

We formalize the Wronskian $W(a,b)$ of any two polynomials $a,b\in R[X]$ with coefficients in an arbitrary commutative ring $R$.

lean variable R : Type* [CommRing R]

def wronskian (a b : R[X]) : R[X] := a * (derivative b) - (derivative a) * b

The degree of Wronskian $W(a,b)$ is strictly smaller than $\deg(a)+\deg(b)$, which was one of the last steps in our proof of Theorem 1.1.

lean theorem wronskian.natDegree_lt_add a b : R[X] (hw : wronskian a b $\neq$ 0) : natDegree (wronskian a b) ¡ natDegree a + natDegree b

For a polynomial a in Lean 4’s mathlib4, both degree a and natDegree a denote the degree of a. The difference between the two is that the natDegree has type $\mathbb{N}$ of natural numbers, while the degree has type WithBot $\mathbb{N}$ which is $\mathbb{N}$ equipped with $-\infty$. The natDegree of zero polynomial is defined as 0, while the degree of that is defined as $-\infty$. While degree is mathematically more natural, we opt to use natDegree as its type $\mathbb{N}$ is much easier to work in Lean 4 than the extended type WithBot $\mathbb{N}$.

We use that $W(a,b)=W(b,c)$ for any $a,b,c\in R[X]$ with $a+b+c=0$ in our proof. This identity actually holds for any alternating bilinear map $B:M\times M\to R$ on any $R$-module $M$. Thus we add the general theorem in the relevant place of mathlib4.

lean theorem eq_of_add_add_eq_zero [IsCancelAdd R] a b c : M (H : B.IsAlt) (hAdd : a + b + c = 0) : B a b = B b c

Note that the above theorem is stated for a slightly general class of $R$ called IsCancelAdd, where the additive structure $(R,+)$ is not necessarily a group but still satisfy the cancellation law: for any $x,y,z\in R$, $x+z=y+z\Rightarrow x=y$.

4.2 Radical

Recall that for any field $k$ and nonzero $f\in k[X]$, its radical $\mathrm{rad}(f)$ is defined as the product of all irreducible monic factors of $f$ not counting multiplicity. In fact, such a definition works over any multiplicative monoid $M$ with zero that is

1. commutative,

2. cancellative ($ab=ac\Rightarrow b=c$ for nonzero $a$),

3. a unique factorization monoid, in the sense that each nonzero element admits a unique factorization into irreducible elements, and

4. a normalization monoid, equipped with a map $u:M\setminus\{0\}\rightarrow M^{*}$ to the set of units $M^{*}$ of $M$ preserving multiplication. The map $x\mapsto u(x)^{-1}x$ is then called the normalization map.

In particular, for polynomials $M=k[X]$ the map $u:k[X]\setminus\{0\}\rightarrow k$ in (4) reads the leading coefficient of a nonzero polynomial. The corresponding normalization map $a\mapsto u(a)^{-1}a$ sends a polynomial $a$ to its scalar multiple which is monic.

In mathlib4, these assumptions can be imposed on a monoid $M$ by using the following instances on $M$.

1. CancelCommMonoidWithZero (for 1 and 2)

2. UniqueFactorizationMonoid (for 3)

3. NormalizationMonoid (for 4)

To define the radical of $a\in M$, we first extract the multiset of normalized factors of $a$ from mathlib4 using normalizedFactors a. Here, a multiset is essentially a set allowing duplicated elements. Then we convert it to a finite set (.toFinset) to get rid of duplicated elements, and multiply them all to get the radical of $a\in M$.

lean variable M : Type* [CancelCommMonoidWithZero M] [UniqueFactorizationMonoid M] [NormalizationMonoid M]

/– Prime factors of ‘a‘ are monic factors of ‘a‘ without duplication. -/ def primeFactors (a : M) : Finset M := (normalizedFactors a).toFinset

/– Radical of ‘a‘ is a product of prime factors of ‘a‘. -/ def radical (a : M) : M := (primeFactors a).prod id

In case of polynomials $M=k[X]$, any radical is a monic polynomial, and $\text{rad}(c)=1$ for any constant $c\in k$ including zero.

Radical satisfies the power law: $\text{rad}(a^{n})=\text{rad}(a)$ for $n\geq 1$. lean theorem radical_pow (a : M) n : Nat (hn : 0 ¡ n) : radical (a ^n) = radical a

Also, $\text{rad}(a)$ divides $a$. Although this is obvious from unique factorization, Lean is not aware of this intuition. A formal proof uses basic lemmas in mathlib4 to boil down the proof to that the Multiset $S$ of prime factors of $a$ contains, as a subset, the same set $S$ with duplicated elements removed. lean theorem radical_dvd_self (a : M) : radical a — a

Once we restrict our attention to a commutative domain $R$ with unique factorization, we can also prove multiplicativity of radical for coprime elements $a,b\in R$, i.e. $\text{rad}(ab)=\text{rad}(a)\text{rad}(b)$. We also have $\mathrm{rad}(-a)=\mathrm{rad}(a)$. These basic lemmas will be used frequently in the main proof.

lean variable R : Type* [CommRing R] [IsDomain R] [NormalizationMonoid R] [UniqueFactorizationMonoid R]

theorem radical_hMul a b : R (hc : IsCoprime a b) : radical (a * b) = (radical a) * (radical b)

theorem radical_neg a : R : radical (-a) = radical a

The following seems obvious to the human eye.

Lemma 3

For any field $k$ and a polynomial $a\in k[X]$ of degree $\geq 1$, the degree of its radical $\mathrm{rad}(a)$ is also at least one.

A formal proof of Lemma 3 requires more work, however. We need to explicitly take a prime factor $p$ of $a$ and show that it is also a prime factor of $\mathrm{rad}(a)$. We first show that for any element $a$ of a general monoid $M$, a prime $p$ divides $a$ if and only if it divides $\mathrm{rad}(a)$; this is done by using that the prime divisors of $a$ and $\mathrm{rad}(a)$ are the same. lean theorem prime_dvd_radical_iff a p : M (ha : a $\neq$ 0) (hp : Prime p) : p — radical a $\leftrightarrow$ p — a We then use it to show that nonzero $a\in M$ is a unit if and only if $\mathrm{rad}(a)$ is. Note that an element of $M$ is a unit if and only if it has no prime divisor. lean theorem radical_isUnit_iff a : M (h : a $\neq$ 0) : IsUnit (radical a) $\leftrightarrow$ IsUnit a Then we specialize it to $M=k[X]$ and use that nonzero $a\in k[X]$ is a unit if and only if its degree is zero, proving Lemma 3. lean lemma natDegree_radical_eq_zero_iff a : k[X] : (radical a).natDegree = 0 $\leftrightarrow$ a.natDegree = 0

The fraction $f/\mathrm{rad}(f)$ is a polynomial which will be used frequently in the proof. We define this as divRadical f in our formalization.

lean def divRadical (a : k[X]) : k[X] := a / radical a The division notation actually denotes the quotient of two polynomials as in the Euclidean division algorithm. Using that the radical divides the polynomial (radical_dvd_self), we define lemmas that introduces and eliminates divRadical f as it is multiplied by radical f. With this, we do not need to work with division explicitly and only work with multiplications, which is easier to handle with Lean 4 tactics like ring.

lean theorem eq_divRadical a x : k[X] (h : (radical a) * x = a) : x = divRadical a

theorem mul_radical_divRadical (a : k[X]) : (radical a) * (divRadical a) = a

Now we need to prove Lemma 1 that for any $a\in k[X]$, $a/\mathrm{rad}(a)$ divides $a^{\prime}$. lean theorem divRadical_dvd_derivative (a : k[X]) : (divRadical a) — (derivative a) Our formalization does not explicitly use the factorization $a=up_{1}^{e_{1}}p_{2}^{e_{2}}\cdots p_{m}^{e_{m}}$ which is somewhat cumbersome to work in Lean. Instead, we use the coprime induction in mathlib4. We first prove the result for units $a=u$ and prime powers $a=p^{e}$. Then we show that for any coprime $a,b$ satisfying the lemma, their product $ab$ also satisfies the lemma. This makes the derivative $(ab)^{\prime}=a^{\prime}b+ab^{\prime}$ much easier to manipulate than the derivative of full factorization.

5 Formalization of Mason–Stothers theorem

Finally, Mason–Stothers theorem is formalized as follows. Note that k is a field of arbitrary characteristic with only [Field k] assumed.

lean variable k : Type* [Field k]

theorem Polynomial.abc a b c : k[X] (ha : a $\neq$ 0) (hb : b $\neq$ 0) (hc : c $\neq$ 0) (hab : IsCoprime a b) (hsum : a + b + c = 0) : (derivative a = 0 $\wedge$ derivative b = 0 $\wedge$ derivative c = 0) $\vee$ Nat.max₃ a.natDegree b.natDegree c.natDegree + 1 $\leq$ (radical (a * b * c)).natDegree

We only require coprimality of $a$ and $b$, as $\gcd(b,c)=\gcd(c,a)=1$ can be deduced from $\gcd(a,b)$ and $a+b+c=0$. Because $a,b,c$ are nonzero, there is no difference in using natDegree instead of degree.

To formalize Mason–Stothers, we first formalize the proof of $abc/\mathrm{rad}(abc)|W$ mentioned as the key step of the proof in Section 3. Then we define an auxiliary lemma below that derives $\deg(c)<\deg(\mathrm{rad}(abc))$ from $abc/\mathrm{rad}(abc)|W$.

lean private theorem abc_subcall a b c w : k[X] hw : w $\neq$ 0 (wab : w = wronskian a b) (ha : a $\neq$ 0) (hb : b $\neq$ 0) (hc : c $\neq$ 0) (hab : IsCoprime a b) (hbc : IsCoprime b c) (hca : IsCoprime c a) (abc_dr_dvd_w : (a * b * c).divRadical — w) : c.natDegree + 1 $\leq$ (radical (a * b * c)).natDegree

Once the auxiliary lemma is shown, we apply this three times to the permuted triples $(a,b,c)$, $(b,c,a)$, and $(c,a,b)$ to prove the full Mason–Stothers. While it is evident that the conditions of abc_subcall are symmetric, we have to manually permute them in our formalization (e.g., change the product a * b * c to b * c * a in abc_dr_dvd_w). This, however, costs much less than repeating the whole argument for $c$ to $a$ and $b$.

6 Formalization of Corollaries

We also formalize multiple corollaries of Mason–Stothers (Theorems 2.1, 2.2 and 2.3).

6.1 Fermat–Catalan Conjecture for Polynomials (Theorem 2.1)

6.1.1 Mathematical Proof

Theorem 2.1 basically follows from Mason–Stothers applied to the triple $(ua^{p},vb^{q},wc^{r})$. Let

$$m=\max\{\deg(ua^{p}),\deg(vb^{q}),\deg(wc^{r})\}=\max\{p\deg(a),q\deg(b),r\deg(c)\}.$$

If the inequality $m<\deg(\mathrm{rad}(a^{p}b^{q}c^{r}))$ holds, then we have

$$m <\deg(\mathrm{rad}(a^{p}b^{q}c^{r}))=\deg(\mathrm{rad}(abc))\leq\deg(abc)$$

$$=\deg(a)+\deg(b)+\deg(c)=\frac{1}{p}\cdot p\deg(a)+\frac{1}{q}\cdot q\deg(b)+\frac{1}{r}\cdot r\deg(c)$$

$$\leq\left(\frac{1}{p}+\frac{1}{q}+\frac{1}{r}\right)m$$

which is a contradiction. So by Mason–Stothers it should be that $(a^{p})^{\prime}=(b^{q})^{\prime}=(c^{r})^{\prime}=0$. As none of $p,q,$ or $r$ are zero in $k$, we conclude $a^{\prime}=b^{\prime}=c^{\prime}=0$. If the characteristic of $k$ is zero, then $a^{\prime}=b^{\prime}=c^{\prime}=0$ immediately implies that $a,b,c$ are constants.

If the characteristic $\ell$ of $k$ is positive, we need an extra infinite descent argument to show that $a,b,c$ are constants. For $f=a,b,c$, that $f^{\prime}=0$ in $k[t]$ implies the existence of $f_{1}\in k[t]$ such that $f(t)=f_{1}(t^{\ell})$. Hence we have $ua_{1}(t^{\ell})^{p}+vb_{1}(t^{\ell})^{q}+wc_{1}(t^{\ell})^{r}=0$. Substitution $T=t^{\ell}$ gives $ua_{1}(T)^{p}+vb_{1}(T)^{q}+wc_{1}(T)^{r}=0$, giving rise to a new nontrivial solution $(a_{1},b_{1},c_{1})$ with strictly smaller yet nonzero degrees. Repeated application of this descent in degree leads to contradiction.

6.1.2 Formalization

The full statement of Theorem 2.1 we formalize is the following.

lean theorem Polynomial.flt_catalan p q r : $\mathbb{N}$ (hp : 0 ¡ p) (hq : 0 ¡ q) (hr : 0 ¡ r) (hineq : q * r + r * p + p * q $\leq$ p * q * r) (chp : $\neg$ringChar k — p) (chq : $\neg$ringChar k — q) (chr : $\neg$ringChar k — r) a b c : k[X] (ha : a $\neq$ 0) (hb : b $\neq$ 0) (hc : c $\neq$ 0) (hab : IsCoprime a b) u v w : k (hu : u $\neq$ 0) (hv : v $\neq$ 0) (hw : w $\neq$ 0) (heq : C u * a ^p + C v * b ^q + C w * c ^r = 0) : a.natDegree = 0 $\wedge$ b.natDegree = 0 $\wedge$ c.natDegree = 0

We state the inequality $1/p+1/q+1/r\leq 1$ as $qr+rs+sp\leq pqr$ instead, as this is expressible purely in integers which is easier to work with in Lean 4. In mathlib, for any element u : k in field $k$ the notation C u : k[X] denotes the same value in ring $k[X]$.

To formalize Theorem 2.1, we first factor the part of the proof where we show $a^{\prime}=b^{\prime}=c^{\prime}=0$. lean theorem Polynomial.flt_catalan_deriv /-…same condition as flt_catalan…-/ : derivative a = 0 $\wedge$ derivative b = 0 $\wedge$ derivative c = 0

We then formalize the infinite descent argument in Section 6.1.1 to show that the degree of $a$ is zero. If the characteristic of $k$ is nonzero, we apply a strong induction on the degree of $a$. lean theorem Polynomial.flt_catalan_aux /-…same condition as flt_catalan…-/ : a.natDegree = 0 Then we use this auxiliary step three times to formalize Theorem 2.1.

FLT for polynomial (Corollary 1) then immediately follows by considering the case when $p=q=r=n\geq 3$ and $u=v=1$, $w=-1$.

lean theorem Polynomial.flt n : $\mathbb{N}$ (hn : 3 $\leq$ n) (chn : $\neg$ringChar k — n) a b c : k[X] (ha : a $\neq$ 0) (hb : b $\neq$ 0) (hc : c $\neq$ 0) (hab : IsCoprime a b) (heq : a ^n + b ^n = c ^n) : a.natDegree = 0 $\wedge$ b.natDegree = 0 $\wedge$ c.natDegree = 0

6.2 Non-parametrizability of $y^{2}=x^{3}+1$ (Theorem 2.2)

6.2.1 Mathematical proof

As a corollary of Theorem 2.1, we can show that $y^{2}=x^{3}+1$ is not parametrizable by rational functions of $t$, similarly as in [12, Proposition 2.3.1].

Assume that a parametrization exists, so that $x=m/M$ and $y=n/N$ for some $m,n,M,N\in k[t]$ with $(m,M)=1$ and $(n,N)=1$. By clearing denominators, we obtain $n^{2}M^{3}=(m^{3}+M^{3})N^{2}$. From this one can show that $N^{2}$ and $M^{3}$ divide each other. Using the unique factorization of $N^{2}=M^{3}$, we can find $\alpha,\beta\in k^{\times}$ and $e\in k[t]$ such that $M=\alpha e^{2}$ and $N=\beta e^{3}$. Now the equation reduces to $\beta^{2}m^{3}+\alpha^{3}\beta^{2}e^{6}=\alpha^{3}n^{2}$, which is a nontrivial solution for the Fermat-Catalan equation with $(p,q,r)=(3,6,2)$. This is a contradiction as the characteristic of $k$ is not $2$ or $3$.

6.2.2 Formalization

The statement can be formalized as follows.

lean def IsConst (x : RatFunc k) := $\exists$ c : k, x = RatFunc.C c

theorem no_parametrization_y2_x3_1 (chk : $\neg$ringChar k — 6) x y : RatFunc k (eqn : y ^2 = x ^3 + 1) : IsConst x $\wedge$ IsConst y

The formalization is straightforward, but requires a large body of code for algebraic manipulation. Also, we had to formalize certain number-theoretic properties coming from that $k[t]$ is a UFD. For example, that $M^{2}$ and $N^{3}$ divides each other impling the existence of $c$ such that $M$ and $N$ are associated to $c^{3}$ and $c^{2}$ respectively, which is true for any UFD.

lean theorem associated_pow_pow_coprime_iff a b : k[X] (ha : a $\neq$ 0) (hb : b $\neq$ 0) m n : $\mathbb{N}$ (hm : m $\neq$ 0) (hn : n $\neq$ 0) (h : Associated (a ^m) (b ^n)) (hcp : m.Coprime n) : $\exists$ c : k[X], c $\neq$ 0 $\wedge$ Associated a (c ^n) $\wedge$ Associated b (c ^m)

6.3 Davenport’s Theorem (Theorem 2.3)

6.3.1 A Non-coprime Variant of Mason–Stothers Theorem

Davenport’s theorem also almost directly follows from Mason–Stothers theorem. We start with a variant of Mason–Stothers theorem by Stothers [20] that does not require coprimality of $a,b,c$.

Theorem 6.1

Let $k$ be any field of characteristic zero, and $a,b,c\in k[t]$ be non-zero polynomials satisfying $a+b+c=0$. Then we either have $a,b,c\in k$ or

$$\max\{\deg(a),\deg(b),\deg(c)\}<\deg(\mathrm{rad}(a))+\deg(\mathrm{rad}(b))+\deg(c).$$

Note that we need $k$ to have characteristic zero in Theorem 6.1. If $\textrm{char }k=p>0$, then a counterexample is $(a,b,c)=(t^{p+1},-t(1+t)^{p},t)$.

Proof

Let $d$ be the gratest common divisor of $a$ and $b$. Then $a=a_{0}d,b=b_{0}d,c=c_{0}d$ for $a_{0},b_{0},c_{0}\in k[t]$ with $a_{0}+b_{0}+c_{0}=0$. Moreover, $\gcd(a_{0},b_{0})=1$ so $a_{0},b_{0},c_{0}$ are nonzero and pairwise coprime. By applying Theorem 1.1 to $(a_{0},b_{0},c_{0})$, we either have $a_{0}^{\prime}=b_{0}^{\prime}=c_{0}^{\prime}=0$ or

$$\begin{gathered}\max\{\deg(a_{0}),\deg(b_{0}),\deg(c_{0})\}<\\ \deg(\mathrm{rad}(a_{0}))+\deg(\mathrm{rad}(b_{0}))+\deg(\mathrm{rad}(c_{0})).\end{gathered}$$ (1)

Consider the case $a_{0}^{\prime}=b_{0}^{\prime}=c_{0}^{\prime}=0$. Since $k$ have characteristic zero, $a_{0},b_{0},c_{0}\in k$. If $d\in k$ then the proof is done. Otherwise, $\deg(d)\geq 1$ by Lemma 3 so

$$\max\{\deg(a),\deg(b),\deg(c)\}=\deg(d) <\deg(\mathrm{rad}(d))+\deg(\mathrm{rad}(d))+\deg(d)$$

$$=\deg(\mathrm{rad}(a))+\deg(\mathrm{rad}(b))+\deg(c)$$

and the proof is done too.

Now consider the case where (1) is true. Then

$$\max\{\deg(a),\deg(b), \deg(c)\}=\max\{\deg(a_{0}),\deg(b_{0}),\deg(c_{0})\}+\deg(d)$$

$$<\deg(\mathrm{rad}(a_{0}))+\deg(\mathrm{rad}(b_{0}))+\deg(\mathrm{rad}(c_{0}))+\deg(d)$$

$$\leq\deg(\mathrm{rad}(a))+\deg(\mathrm{rad}(b))+\deg(c_{0})+\deg(d)$$

$$=\deg(\mathrm{rad}(a))+\deg(\mathrm{rad}(b))+\deg(c)$$

completing the proof of Theorem 6.1.

The variant Theorem 6.1 is formalized as following. lean theorem Polynomial.abc’_char0 [CharZero k] a b c : k[X] (ha : a $\neq$ 0) (hb : b $\neq$ 0) (hc : c $\neq$ 0) (hsum : a + b + c = 0) : (a.natDegree = 0 $\wedge$ b.natDegree = 0 $\wedge$ c.natDegree = 0) $\vee$ Nat.max₃ a.natDegree b.natDegree c.natDegree + 1 $\leq$ (radical a).natDegree + (radical b).natDegree + c.natDegree

6.3.2 Mathematical proof

We now prove Davenport’s theorem (Theorem 2.3), mainly following the proof in Stothers’ paper [20]. For non-constant polynomials $f,g\in k[t]$ with $f^{3}-g^{2}\neq 0$, apply Theorem 6.1 to the zero-sum triple $(-f^{3},g^{2},f^{3}-g^{2})$. The equality case

$$(f^{3})^{\prime}=(g^{2})^{\prime}=(f^{3}-g^{2})^{\prime}=0$$

cannot happen since it would imply $3f^{2}f^{\prime}=0=2gg^{\prime}$ and thus $3=0=2$.

So we get the inequality

$$\max\{3\deg(f),2\deg(g)\} \leq\max\{\deg(-f^{3}),\deg(g^{2}),\deg(f^{3}-g^{2})\}$$

$$<\deg(\mathrm{rad}(-f^{3}))+\deg(\mathrm{rad}(g^{2}))+\deg(f^{3}-g^{2})$$

$$\leq\deg(f)+\deg(g)+\deg(f^{3}-g^{2}).$$

This gives two inequalities

$$3\deg(f)+1 \leq\deg(f)+\deg(g)+\deg(f^{3}-g^{2})$$

$$2\deg(g)+1 \leq\deg(f)+\deg(g)+\deg(f^{3}-g^{2})$$

and adding these two inequalities and rearranging gives the desired inequality.

6.3.3 Formalization

The statement of Davenport’s theorem (Corollary 2.3) can be formalized as follows. lean theorem Polynomial.davenport [CharZero k] a b : k[X] (ha : a.natDegree ¿ 0) (hb : b.natDegree ¿ 0) (hnz : a ^3 - b ^2 $\neq$ 0) : a.natDegree + 2 $\leq$ 2 * (a ^3 - b ^2).natDegree

We also formalized a variant of Davenport’s theorem that allows arbitrary characteristics, with the cost of assuming coprimality of two polynomials and assuming non-vanishing derivative instead of non-constantness. Note that we cannot remove all of these assumptions; $k=\mathbb{F}_{2}$ with $(a,b)=(t^{4},t^{6}+t)$ gives a counterexample.

lean theorem Polynomial.davenport’ a b : k[X] (hab : IsCoprime a b) (haderiv : derivative a $\neq$ 0) (hbderiv : derivative b $\neq$ 0) : a.natDegree + 2 $\leq$ 2 * (a ^3 - b ^2).natDegree

7 Comparison with Previous Works

We compare our work to other formalizations of the Mason-Stothers Theorem by Eberl in Isabelle [7] and by Wagemaker in Lean 3 [10, 24]. All three formalizations, including ours, are based on the same proof by Lemmermeyer’s note [12] on the elementary proof of Snyder [19]. Unlike Snyder’s original proof [19] which assumes that $k$ is algebraically closed, all formalizations work with any field $k$ using radicals, following [12, Theorem 2.1.4, Corollary 2.1.5].

	Eberl [7]	Wagemaker [10]	Ours
Radical	radical	rad	radical
$(a,b)=1\Rightarrow\mathrm{rad}(ab)=\mathrm{rad}(a)\mathrm{rad}(b)$	radical_mult_coprime	rad_mul_eq_rad_mul_rad_of_coprime	radical_hMul
$\deg W(a,b)<\deg(a)+\deg(b)$	degree_pderiv_mult_less	degree_wron_le	natDegree_lt_add
$\frac{a}{\mathrm{rad}(a)}|a^{\prime}$ ([12, Lemma 2.1.2])	poly_div_radical_dvd_pderiv	Mason_Stothers_lemma	divRadical_dvd_derivative
Mason–Stothers	Mason_Stothers	Mason_Stothers	Polynomial.abc
	Mason_Stothers_char_0
Polynomial FLT	fermat_poly	-	Polynomial.flt
	fermat_poly_char_0

Table 1: Comparison of definitions and theorems in different formalizations of Mason–Stothers.

7.1 Eberl’s Isabelle formalization

Eberl formalized both the characteristic zero and positive case of Mason–Stothers theorem in Isabelle [7], as a part of the Archive of Formal Proofs (Isabelle-AFP) mathematics library. Consequently, their formalization is reusable with other definitions and theorems in Isabelle-AFP. We compare their formalization to ours as follows.

1. They define the radical $\mathrm{rad}(a)$ on any factoral semiring, which is a commutative ring with unique factorization. We define radical in a slightly more general setting of monoids with unique factorization.

2. They assume the coprimality condition $\mathrm{gcd}(a,b,c)=1$ in Mason–Stothers, but this is equivalent to pairwise coprimality $\mathrm{gcd}(a,b)=\mathrm{gcd}(b,c)=\mathrm{gcd}(c,a)=1$ we assume by $a+b+c=0$.

3. Their work also formalizes the polynomial version of FLT for any characteristic. They proved that, when a triple of nonzero coprime polynomials satisfy $a^{n}+b^{n}+c^{n}=0$ and at least one of $(a^{n})^{\prime}$, $(b^{n})^{\prime}$, or $(c^{n})^{\prime}$ is nonzero, then $n\leq 2$. In other words, nonzero coprime polynomials $a,b,c$ satisfying the Fermat’s equation for $n\geq 3$ should have $(a^{n})^{\prime}=(b^{n})^{\prime}=(c^{n})^{\prime}=0$. Our formalization of polynomial FLT (Corollary 1) has a strictly stronger conclusion; either the characteristic of $k$ divides $n$, or $a,b,c\in k$. This is achieved by the simple infinite descent argument in Section 6.1.1.

7.2 Wagemaker’s Lean 3 formalization

Wagemaker formalized the Mason–Stothers theorem in Lean 3, in the early days when the Lean mathlib mathematics library was taking shape [10, 24]. Consequently, Hőlzl and Wagemaker built a large body of work (“4/5 of the formalization” according to Wagemaker [24]) that formalizes many fundamental notions such as polynomials, UFDs, greatest common divisor, and coprimality [10]. Their work was then incorporated into the current mathlib/mathlib4 library of Lean 3 and 4. In particular, the design suggestions [24] in Wagemaker’s work shapes a lot of fundamental APIs in the current mathlib implementation of UFDs..

Their project was independent of Lean 3’s mathlib, however, as it was incorporated after its completion. In contrast, our work builds on the now-mature mathlib4 of Lean 4, ensuring reusability with existing definitions. In regards to the formalization of Mason–Stothers theorem, we compare their work to ours as follows.

1. They work on fields of characteristic zero only, while our formallization allows arbitrary characteristic.

2. They do not formalize further corollaries of Mason–Stothers such as polynomial FLT.

3. Their work misses the proof that a polynomial ring $R[X]$ over a unique factorization domain $R$ is also a unique factorization domain. In contrast, our formalization on Lean 4 is complete, and is based on the current mathlib4 with the proof that $R[X]$ is UFD (and many more).

4. They do not define $a/\mathrm{rad}(a)$ explicitly but instead use $\gcd(a,a^{\prime})$ to avoid polynomial division. Then they prove

   $$\deg(a)\leq\deg(\gcd(a,a^{\prime}))+\deg(\mathrm{rad}(a))$$

   as a lemma [24, Lemma 2.3.1], instead of Lemma 1 in our work.

		Eberl [7]	Wagemaker [10, 24]	Ours
Language		Isabelle	Lean 3	Lean 4
Complete		O	X	O
Mason-Stothers	$\text{char}=0$	O	O	O
	$\text{char}>0$	O	X	O
Poly-FLT	$\text{char}=0$	O	X	O
	$\text{char}>0$	O	X	O
Other corollaries		X	X	O

Table 2: Comparison of Formalizations of Mason–Stothers Theorem

8 Future Works

We suggest further directions in formalizing generalizations of Mason–Stothers theorem.

• Bayat et al. [1] extends the Mason–Stothers theorem to more than three polynomials, using the Wronskian of more than two polynomials.

• In algebraic geometry, it is known that rational maps from the projective line to a curve exist only if the curve has genus 0. This immediately proves both FLT for polynomials and non-parametrizability of any elliptic curves, as the Fermat curve $x^{n}+y^{n}=1$ ($n\geq 3$) and elliptic curve $y^{2}=x^{3}+ax+b$ have genus $>0$ [12]. However, the current mathlib4 does not have enough theorems to prove this result (e.g. Riemann–Hurwitz formula).

• Mason–Stothers theorem can be thought as the most basic case of ABC over a function field $k(C)$ of a smooth projective curve $C$, when the curve $C$ is the projective line over $k$. Mason [13] proved the following more general result:

  Theorem 8.1 (Mason)

  Let $k$ be an algebraically closed field and $C$ be a smooth projective curve over $k$. Let $a,b\in k(C)$ satisfying $a+b=1$, and $S$ be a finite subset of points in $C(k)$ containing all the zeros and poles of $a$ and $b$. Then either $a,b\in k^{\times}$ or

  $$\max\{\deg(a),\deg(b)\}\leq 2g-2+|S|.$$

  When $C=\mathbb{P}^{1}$, this reduces to the Mason–Stothers theorem: a zero-sum coprime triple $a,b,c$ of polynomials gives $(-a/c)+(-b/c)=1$, and the above inequality becomes

  $$\max\{\deg(a),\deg(b),\deg(c)\}$$

  $$=\max\left\{\deg\left(-\frac{a}{c}\right),\deg\left(-\frac{b}{c}\right)\right\}$$

  $$\leq-2+|S|=\deg(\mathrm{rad}(abc))-1,$$

  where $S=\{\text{zeros of }abc\}\cup\{\infty\}$ and $|S|=\deg(\mathrm{rad}(abc))+1$. Silverman [18] gives a short proof of the theorem using Riemann-Hurwitz formula.

Acknowledgement

We thank Kevin Buzzard for suggesting the project. Also, we thank Thomas Browning for his help in simplifying the formalization of Davenport’s theorem. We also thank the reviewers of mathlib4 who helped improving our codes and porting them into mathlib4, including Johan Commelin, Yaël Dillies, and Eric Weiser. Jineon Baek acknowledges the support from Korea Foundation for Advanced Studies during the completion of this work.

Appendix 0.A Porting to mathlib4

We are in the process of integrating our formalization of the Mason–Stothers theorem to the mathlib4 library. Table 3 lists the pull requests made so far to mathlib4 as of August 23, 2024.

Topic	PR	Descriptions
Wronskian	14281	Prove $B(a,b)+B(b,c)+B(c,a)=0$ for alternating bilinear $B$.
	14243	Define Wronskian of polynomials and prove relevant theorems.
Radical	14873	Define radical of elements.
	15531	Prove theorems on radicals of coprime elements.
Mason–Stothers	15706	Proof of Mason–Stothers theorem.
Polynomial FLT	16060	Statement of FLT for semirings, allowing nonzero unit solutions.

Table 3: List of pull requests to mathlib4.