The Weyl Algebra as the Ring of Differential Operators on Affine Space

Q: Why is the inclusion A_n ⊂ eq D(K[X]) easy?

Because multiplication operators have order 0 , each ∂_i has order 1 since [∂_i, f] = ∂_i(f) is again a multiplication, and composition adds orders. So every word in the generators has finite order.

Q: Where exactly does the proof use characteristic zero?

In one place: the potential Q'' = ∑ _β (β_k+1)^-1g_β ∂^β+e_k divides by positive integers. In characteristic p that inverse does not exist when p ∣ β_k + 1 , and the theorem genuinely fails.

Q: How does the theorem determine the order filtration?

The induction proves D^m = C_m level by level, so the order of an operator in canonical form is the largest |β| with f_β ≠ 0 . Coefficients are irrelevant to order: x_1^1000∂_1 has order 1 .

Q: Is the potential Q unique?

Only up to K[X] . If Q and Q' both work then [Q - Q', x_i] = 0 for all i , so Q - Q' is multiplication by a polynomial by the rigidity lemma. That is precisely the constant of integration, and it is fixed in the theorem by evaluating at 1 .

Q: What replaces the theorem in characteristic p?

D(K[X]) is generated by all divided power operators ∂_i^[k] , which act on monomials by binomial coefficients. The resulting ring is not Noetherian and not finitely generated as a K -algebra, and it is a union of an increasing chain of matrix-like subalgebras over Frobenius powers.

Overview

The Weyl algebra $A_n(K)$ is defined concretely: it is the subalgebra of $End_K(K[x_1, … ,x_n])$ generated by multiplication by the variables and by the partial derivatives. The ring P2 is defined abstractly, by an induction that mentions only commutators. This page proves that the two agree when $R$ is a polynomial ring over a field of characteristic zero.

The inclusion $A_n ⊂ eq D(K[X])$ is easy: multiplication operators have order $0$ , the $∂_i$ have order $1$ , and composition adds orders. Everything is in the reverse inclusion. An operator of order $m$ is given to us as a $K$ -linear map with a commutator property, and we have to manufacture from it an explicit expression $∑ _|β| ≤ m f_β ∂^β$ .

Two lemmas do the work. The first says that an operator commuting with multiplication by every coordinate must be multiplication by a polynomial - a rigidity statement. The second is a discrete Poincaré lemma: given operators $P_1, … ,P_n$ satisfying the symmetry $[P_i, x_j] = [P_j, x_i]$ , there is a single $Q$ with $P_i = [Q, x_i]$ for all $i$ . This is the algebraic form of the statement that a curl-free polynomial vector field on affine space has a potential, and it is exactly where characteristic zero is used: the construction of $Q$ divides by positive integers.

The theorem also pins down the order filtration: $D^m(K[X])$ is the set of operators with $|β| ≤ m$ , a free $K[X]$ -module of rank $m+nn$ . In characteristic $p$ the theorem is false, and the extra operators are the divided powers.

Definition

Let $K$ be a field of characteristic zero, $X = (x_1, … ,x_n)$ , and $K[X] = K[x_1, … ,x_n]$ . Multi-indices $α, β$ range over $N^n$ , with $|β| = β_1 + … + β_n$ and $∂^β = ∂_1^β_1 ⋅ s∂_n^β_n$ ; $e_i$ is the $i$ -th standard multi-index.

The spaces $C_m$

For $m ≥ 0$ set

C_m = ∑ _|β| ≤ m f_β, ∂^β ;:; f_β ∈ K[X] ⊂ eq A_n(K),

(3.14)

so $C_0 = K[X]$ , $C_1 = K[X] ⊕ Der_K(K[X])$ by the description of derivations of a polynomial ring, and $A_n = ⋃ _m C_m$ . By the canonical basis theorem the monomials $x^α ∂^β$ are linearly independent, so $C_m$ is a free $K[X]$ -module on the $∂^β$ with $|β| ≤ m$ .

The Weyl algebra is the ring of differential operators of $K[X]$ Coutinho (3.2.3)

Let $K$ have characteristic zero. Then

D (K[x_1, … ,x_n] ) = A_n(K), D^m (K[x_1, … ,x_n] ) = C_m for every m ≥ 0.

(3.15)

In particular the order of $∑ f_β ∂^β$ is $|β| : f_β ≠ 0$ , and $D^m$ is free over $K[X]$ of rank $m+nn$ .

Note

The identity $C_m = C_m+1 ∩ D^m(K[X])$ is what makes the induction in the proof work: it says an element of $C_m+1$ that happens to have order at most $m$ already lies in $C_m$ . Once the theorem is proved this is a triviality, but it is used along the way in the form $[C_m, x_i] ⊂ eq C_m-1$ .

Core Concepts

What has to be produced

An abstract operator of order $m$ comes with no coefficients. All we know is that commuting with any coordinate lowers the order. The proof turns that into coefficients by descending induction: form $P_i = [P, x_i]$ , which has order $m-1$ and by induction is an explicit expression; then reconstruct $P$ from the $P_i$ up to a constant of integration, which is a polynomial.

Why the reconstruction is a potential problem

The $n$ operators $P_i = [P, x_i]$ are the algebraic analogue of the partial derivatives of a function. They are not arbitrary: since $ad_x_i$ and $ad_x_j$ commute, they satisfy $[P_i, x_j] = [P_j, x_i]$ , the analogue of the symmetry of mixed partials. The reconstruction lemma is the converse - the analogue of the statement that a curl-free field on affine space is a gradient. Coutinho points out this parallel explicitly, and it is the right way to remember the lemma.

The one identity that runs the machine

Everything reduces to $[∂^β, x_i] = β_i,∂^β - e_i$ . Commuting with $x_i$ differentiates the symbol with respect to $ξ_i$ ; building a potential integrates it. The integration step divides by $β_i + 1$ , which is where characteristic zero enters and where the theorem fails in characteristic $p$ .

Why the rigidity lemma is needed at all

Reconstruction from the $P_i$ can only determine $P$ up to something whose commutator with every $x_i$ vanishes. Lemma 2.1 says that ambiguity is exactly $K[X]$ , no more. Without it, the difference $Q - P$ could in principle be some exotic operator, and the induction would not close.

Construction and Proof

The easy inclusion first: $∂_i ∈ D^1$ and $f ∈ D^0$ for $f ∈ K[X]$ , so by the composition rule $D^nD^m ⊂ eq D^n+m$ we get $C_m ⊂ eq D^m(K[X])$ for all $m$ . What follows proves the reverse.

Rigidity: operators commuting with all coordinatesCoutinho (3.2.1)

Let $P ∈ D(K[X])$ satisfy $[P, x_i] = 0$ for $i = 1, … ,n$ . Then $P ∈ K[X]$ , that is, $P$ is multiplication by a polynomial.

Proof

It is enough to show $[P, f] = 0$ for every $f ∈ K[X]$ , since then $P$ has order $0$ and $D^0(K[X]) = K[X]$ . Because $f ↦ [P,f]$ is additive, it suffices to treat monomials $f = x^α$ , and we induct on $|α|$ . For $|α| ≤ 1$ the claim is the hypothesis. For $|α| ≥ 2$ pick $i$ with $α_i ≠ 0$ and write $x^α = x_i ⋅ x^α - e_i$ ; then

[P, x^α] = [P, x_i],x^α - e_i + x_i,[P, x^α - e_i] = 0 + 0 = 0,

the first term by hypothesis and the second by the inductive hypothesis.

Existence of a potentialCoutinho (3.2.2)

Let $r ≥ 1$ and let $P_1, … ,P_n ∈ C_r-1$ satisfy $[P_i, x_j] = [P_j, x_i]$ for all $1 ≤ i,j ≤ n$ . Then there exists $Q ∈ C_r$ with $[Q, x_i] = P_i$ for $i = 1, … ,n$ .

Proof

We construct $Q$ by descending induction on $k$ , maintaining the statement: there is $Q' ∈ C_r$ with $[Q', x_i] = P_i$ for all $i > k$ . For $k = n$ the statement is vacuous, satisfied by $Q' = 0$ . Suppose it holds for $k$ ; we produce a $Q$ that works for $k - 1$ , that is, additionally for $i = k$ .

Put $G = [Q', x_k] - P_k$ . Both terms lie in $C_r-1$ , so $G ∈ C_r-1$ . For $i > k$ , using that $ad_x_i$ and $ad_x_k$ commute and then the hypothesis on the $P$ 's,

[G, x_i] = [[Q',x_k],x_i ] - [P_k, x_i] = [[Q',x_i],x_k ] - [P_i, x_k] = [P_i, x_k] - [P_i, x_k] = 0.

Now write $G = ∑ _β g_β ∂^β$ in canonical form. By (3.16), $[G, x_i] = ∑ _β β_i, g_β, ∂^β - e_i$ , and since the $∂^γ$ are linearly independent over $K[X]$ , the vanishing of $[G,x_i]$ forces $β_i = 0$ whenever $g_β ≠ 0$ . Doing this for every $i > k$ shows that all multi-indices occurring in $G$ are supported in the first $k$ coordinates.

Define $Q^{''}$ by (3.19). Every $β$ occurring has $|β| ≤ r-1$ , so $|β + e_k| ≤ r$ and $Q'' ∈ C_r$ . The division by $β_k + 1$ is legitimate because $char K = 0$ . By (3.16), $[Q'', x_k] = ∑ _β (β_k+1)/(β_k+1) g_β ∂^β = G$ , and $[Q'', x_i] = 0$ for $i > k$ because the multi-indices $β + e_k$ have $i$ -th entry zero.

Set $Q = Q' - Q''$ . Then $[Q, x_i] = [Q', x_i] = P_i$ for $i > k$ , and $[Q, x_k] = [Q', x_k] - G = P_k$ . This is the statement for $k-1$ , and the descending induction terminates at $k = 0$ with the required $Q$ .

Proof of the theorem

We show $D^m(K[X]) ⊂ eq C_m$ by induction on $m$ . For $m = 0$ , $D^0(K[X]) = K[X] = C_0$ . For $m = 1$ , $D^1(K[X]) = K[X] ⊕ Der_K(K[X])$ and every derivation of $K[X]$ is $∑ _i D(x_i)∂_i$ , so $D^1 = C_1$ .

Assume $D^k(K[X]) = C_k$ for all $k ≤ m-1$ , and take $P ∈ D^m(K[X])$ . Set $P_i = [P, x_i] ∈ D^m-1(K[X]) = C_m-1$ . These satisfy the compatibility (3.18), so the potential lemma with $r = m$ supplies $Q ∈ C_m$ with $[Q,x_i] = P_i$ for every $i$ . Then $[Q - P, x_i] = 0$ for every $i$ , and $Q - P$ lies in $D(K[X])$ , so by the rigidity lemma $Q - P ∈ K[X] = C_0$ . Hence $P = Q - (Q-P) ∈ C_m$ .

Taking the union over $m$ gives $D(K[X]) = ⋃ _m C_m = A_n(K)$ , and the filtration statement is what the induction proved.

The Poincaré lemma in disguise

Restricted to order $1$ , the potential lemma reads: if $F_1, … ,F_n ∈ K[X]$ satisfy $∂ F_i/∂ x_j = ∂ F_j/∂ x_i$ for all $i, j$ , then there is $g ∈ K[X]$ with $F_i = ∂ g/∂ x_i$ . That is precisely the statement that a closed polynomial $1$ -form on affine space is exact. The general case is the same statement one order up, for symbols, and the proof - integrate in one variable at a time, correcting as you go - is the classical proof.

Key Equations

The commutation relation of the Weyl algebra, in multi-index form:

[∂_i, x_j] = δ_ij, [∂^β, x_i] = β_i,∂^β - e_i, [x^α, ∂_i] = -α_i, x^α - e_i.

(3.16)

Commuting with a coordinate drops the order by exactly one, and the operation is the symbol derivative:

[, ⋅ ,, x_i] : C_m ⟶ C_m-1, σ_m-1 ([P, x_i] ) = (∂)/(∂ ξ_i),σ_m(P).

(3.17)

The compatibility condition satisfied by the family $P_i = [P, x_i]$ , which comes from the commutativity of the operators $ad_x_i$ :

[P_i, x_j] = [[P,x_i],x_j ] = [[P,x_j],x_i ] = [P_j, x_i] (1 ≤ i,j ≤ n).

(3.18)

The explicit potential built in Lemma 2.2 from an operator $G = ∑ _β g_β ∂^β$ whose multi-indices are supported in the first $k$ coordinates:

Q'' = ∑ _β (1)/(β_k + 1), g_β, ∂^β + e_k, [Q'', x_k] = G.

(3.19)

Finally, the rank of each filtration piece as a free $K[X]$ -module:

rank_K[X] D^m(K[X]) = #β ∈ N^n : |β| ≤ m = m+nn.

(3.20)

Variable Definitions

$K$: the ground field, of characteristic zero
$K[X]$: the polynomial ring $K[x_1, … ,x_n]$
$A_n(K)$: the $n$ -th Weyl algebra over $K$
$∂_i$: partial differentiation with respect to $x_i$ , acting on $K[X]$
$β$: a multi-index in $N^n$ ; $|β|$ is the sum of its entries
$e_i$: the multi-index with $1$ in position $i$ and $0$ elsewhere
$C_m$: the operators $∑ _|β| ≤ m f_β ∂^β$
$D^m(K[X])$: the operators of order at most $m$ , defined by iterated commutators
$P_i$: the commutator $[P, x_i]$ , playing the role of a partial derivative of $P$
$Q$: a potential: an operator with $[Q, x_i] = P_i$ for all $i$

Properties and Behaviour

The order filtration is the derivative degree

For $P = ∑ _β f_β ∂^β ≠ 0$ in $A_n$ , $ord(P) = |β| : f_β ≠ 0$ , and $D^m$ is free over $K[X]$ of rank $m+nn$ . The coefficients play no part: $x_1^1000∂_1$ has order $1$ . See the order filtration.

The graded ring

$gr,A_n ≅ K[x_1, … ,x_n,ξ_1, … ,ξ_n]$ with $ξ_i = σ_1(∂_i)$ . Since this is a domain, orders add on products and $A_n$ is a domain. Since it is Noetherian, $A_n$ is Noetherian. Both standard properties of $A_n$ descend from the theorem via the filtration.

Intrinsic characterisation of the generators

$K[X]$ and $Der_K(K[X])$ generate $D(K[X])$ . This is the smooth case of the general statement for regular rings, and it fails on singular varieties; see differential operators on an affine variety.

Consequences for modules

Since $D(K[X]) = A_n$ , the notion of a D-module on affine space is unambiguous: a module over the concrete Weyl algebra and a module over the abstractly defined operator ring are the same thing. All the constructions in this collection - the polynomial ring as a D-module, localisations, modules attached to equations - therefore have both a concrete and an intrinsic reading.

What is not claimed

The theorem says nothing about $D(R)$ for other $R$ , and its proof does not transfer: both lemmas use the free structure of $K[X]$ heavily - linear independence of the $∂^β$ over $K[X]$ in the potential lemma, and the monomial basis in the rigidity lemma. On a quotient ring neither is available.

Examples and Special Cases

$n = 1$

$D(K[x]) = A_1(K)$ , generated by $x$ and $∂$ with $[∂, x] = 1$ , and $D^m$ is free of rank $m+1$ over $K[x]$ on $1, ∂, … , ∂^m$ . The potential lemma is trivial here - with one variable there is no compatibility condition - and the whole proof reduces to integrating a single operator once.

The compatibility condition is not vacuous

Take $n = 2$ , $P_1 = ∂_2$ , $P_2 = 0$ . Then $[P_1, x_2] = 1$ but $[P_2, x_1] = 0$ , so the condition fails and no potential exists: any $Q$ with $[Q,x_2] = 0$ has all its multi-indices free of $β_2$ , and then $[Q,x_1]$ cannot involve $∂_2$ . The condition is exactly what rules this out.

Characteristic $p$ : the theorem fails

Over $K$ of characteristic $p$ , the map $∂^[p] : x^m ↦ mpx^m-p$ is a $K$ -linear endomorphism of $K[x]$ of order $p$ . Yet $∂^p = 0$ on $K[x]$ in that characteristic, so $A_1$ contains no operator of order $p$ built from $∂$ alone in the required way, and one checks $∂^[p] ∉ A_1$ . The full ring $D(K[x])$ is generated by all the divided powers $∂^[k]$ , $k ≥ 0$ , and is strictly larger than $A_1$ ; see the positive characteristic page.

Laurent polynomials

For $R = K[x^ ± 1]$ the same argument, applied after localisation, gives $D(R) = R ⟨ ∂ ⟩$ , the localisation of $A_1$ at the powers of $x$ . Smoothness, not polynomiality, is what the proof really uses once one is willing to localise.

Worked Example

Running the potential construction in $A_2$

Step 1 - the data and the compatibility check
Take $n = 2$ , $r = 2$ , and

$P_1 = 2x_1∂_1 + ∂_2, P_2 = ∂_1,$

both in $C_1$ . The lemma requires $[P_1, x_2] = [P_2, x_1]$ . Using $[∂_i, x_j] = δ_ij$ : $[P_1, x_2] = 2x_1[∂_1,x_2] + [∂_2,x_2] = 0 + 1 = 1$ , and $[P_2, x_1] = [∂_1, x_1] = 1$ . They agree, so a potential must exist.
Step 2 - the case $k = 2$
Start with $Q' = 0$ , which handles the empty condition $i > 2$ . Then $G = [Q', x_2] - P_2 = -∂_1$ . Its multi-index is $β = (1,0)$ , already supported in the first $2$ coordinates, so no reduction is needed. Formula (3.19) with $k = 2$ gives

$Q'' = (1)/(β_2 + 1)(-1),∂^β + e_2 = (1)/(0+1)(-1),∂_1∂_2 = -∂_1∂_2,$

and $Q_(1) = Q' - Q'' = ∂_1∂_2$ . Check: $[∂_1∂_2, x_2] = ∂_1[∂_2,x_2] = ∂_1 = P_2$ .
Step 3 - the case $k = 1$
Now $Q' = ∂_1∂_2$ and $G = [Q', x_1] - P_1$ . Since $[∂_1∂_2, x_1] = [∂_1,x_1]∂_2 = ∂_2$ ,

$G = ∂_2 - (2x_1∂_1 + ∂_2) = -2x_1∂_1.$

As the proof predicts, $[G, x_2] = -2x_1[∂_1, x_2] = 0$ , so $G$ is supported in the first coordinate only. Formula (3.19) with $k = 1$ and $β = (1,0)$ gives

$Q'' = (1)/(β_1 + 1)(-2x_1),∂^β + e_1 = (1)/(2)(-2x_1),∂_1^2 = -x_1∂_1^2.$
Step 4 - the potential
Set $Q = Q' - Q'' = ∂_1∂_2 + x_1∂_1^2 ∈ C_2$ . Verify both required identities directly, using $[∂_1^2, x_1] = 2∂_1$ :

$[Q, x_1] = ∂_2 + x_1,[∂_1^2, x_1] = ∂_2 + 2x_1∂_1 = P_1,$

$[Q, x_2] = [∂_1∂_2, x_2] + x_1[∂_1^2, x_2] = ∂_1 + 0 = P_2.$

Both check out, and $Q$ has order $2$ as the lemma promises.
Step 5 - recovering an abstract operator
Suppose now $P ∈ D^2(K[x_1,x_2])$ is given abstractly and happens to have $[P,x_1] = P_1$ and $[P,x_2] = P_2$ . By the rigidity lemma $Q - P$ is multiplication by a polynomial $f$ , and evaluating at $1 ∈ K[X]$ identifies it: $f = Q(1) - P(1)$ . Here $Q(1) = (∂_1∂_2 + x_1∂_1^2)(1) = 0$ , so $f = -P(1)$ and

$P = ∂_1∂_2 + x_1∂_1^2 + P(1).$

The abstract operator has been written explicitly, which is exactly what the theorem asserts is always possible.
Step 6 - where characteristic zero was used
Twice, both in Step 3: the coefficient $1β_1+1 = 12$ requires $2$ to be invertible. In characteristic $2$ the same data admits no potential of the form produced here, and indeed the theorem is false: $D(K[x])$ then contains the divided power operator $x^m ↦ m2x^m-2$ , which is not in $A_1$ .

Result

$Q = ∂_1∂_2 + x_1∂_1^2$ is the potential for the compatible pair $P_1 = 2x_1∂_1 + ∂_2$ , $P_2 = ∂_1$ , verified by $[Q,x_1] = 2x_1∂_1 + ∂_2$ and $[Q,x_2] = ∂_1$ . Any abstract $P ∈ D^2$ with those commutators equals $Q + P(1)$ , so it lies in $C_2$ - the inductive step of the theorem, carried out on a concrete case.

Applications and Industry Use

In a mathematics topic, this section covers downstream use inside mathematics, computing and engineering rather than a manufactured product.

Legitimising D-module theory on affine space. The theorem is why a module over $A_n$ may be called a D-module: the concrete algebra is the intrinsic one, so the theory does not depend on a choice of coordinates for its definition, only for its notation.
Transfer to smooth varieties. Covering a smooth variety by affine charts isomorphic to open subsets of $K^n$ , and using that $𝒟$ localises, gives the sheaf $D_X$ with the expected local description. Every local computation with $D_X$ is a computation in a localised Weyl algebra.
Symmetries and invariants. Because the operator ring is intrinsic, any automorphism of $K[X]$ induces an automorphism of $A_n$ ; this is the source of the link between automorphisms of the Weyl algebra and polynomial automorphisms, and hence of the Dixmier conjecture and its relation to the Jacobian conjecture.
Computer algebra. Implementations represent operators in the canonical form $∑ f_β ∂^β$ ; the theorem guarantees no operator escapes that representation, so the data structure is complete.
Systems of PDE with polynomial coefficients. Any linear system with polynomial coefficients is an $A_n$ -module, and the theorem says nothing is lost by that encoding: every intrinsically defined operator on polynomials is already of that shape.

Design Considerations

For a mathematical object, design considerations are the modelling choices: which ring, which filtration, which category to work in.

Where to place the coefficients

The theorem is stated with coefficients on the left, $∑ f_β ∂^β$ . Coefficients on the right, $∑ ∂^β g_β$ , give the same ring and the same orders, with coefficients related by the commutation relations. Fixing the convention once matters more than which one is chosen, because symbols and Groebner-basis leading terms are computed from the canonical form.

How much smoothness is really needed

The proof uses the polynomial ring specifically, but the conclusion holds for any regular finitely generated $K$ -algebra in characteristic zero, with a different and harder argument. When designing a statement, ask whether polynomiality or regularity is the real hypothesis; for anything that will later be localised or globalised, regularity is the right one.

Which lemma to generalise

Of the two lemmas, the rigidity lemma generalises easily - it only needs the ring to be generated by the elements one commutes with. The potential lemma is the delicate one: it is a statement about the vanishing of a cohomology group, and on a variety with non-trivial topology or singularities the analogous group need not vanish. That is the structural reason the theorem is about affine space.

Standards and Codes

For mathematics, the relevant standards are notation, numeric and markup standards together with reference implementations.

ISO 80000-2 fixes $∂$ for partial differentiation and $ℕ$ for the natural numbers including zero, which is the convention used for multi-indices here.
ISO/IEC 40314 (MathML 3.0) encodes the mathematics on this page.
Naming: the ring is called the Weyl algebra after Hermann Weyl's use of the canonical commutation relations in quantum mechanics; in the physics literature the same algebra appears as the algebra generated by position and momentum operators, with $[ p, q] = -iℏ$ . The factor of $i ℏ$ is a rescaling and changes nothing algebraically over a field of characteristic zero.
Software conventions: Macaulay2 writes dx for $∂_x$ , Singular writes Dx, SageMath writes Dx in ore_algebra. All use the left-coefficient canonical form.

Material Selection

The material of a mathematical construction is its numeric substrate: the ground field, the coefficient ring and the representation used to store it.

Characteristic zero is essential. The division by $β_k + 1$ in (3.19) fails when the characteristic divides it. In characteristic $p$ the correct statement is that $D(K[X])$ is generated by the divided powers $∂_i^[k]$ and is a strictly larger, non-Noetherian ring.
Any field of characteristic zero works. Nothing in either lemma needs $K$ algebraically closed or complete; $ℚ$ , number fields, $ℝ$ and $ℂ$ all behave identically, and the theorem commutes with extension of scalars.
Finitely many variables. The proof inducts over $n$ coordinates in the potential lemma. For infinitely many variables the notion of order still makes sense, but the descending induction does not terminate and the statement must be reformulated.
Representation. Store an operator as a sparse map from multi-indices $β$ to polynomials $f_β$ . By the theorem this is a complete and faithful encoding of $D(K[X])$ , and the order is the maximum $|β|$ present.

Computational Notes

Read this as the manufacturing section of the template: how the object is actually built by machine, at what cost, and where the computation stops being decidable.

To decide whether a given $K$ -linear map $P$ on $K[X]$ is a differential operator of order at most $m$ , compute the iterated commutators with the $x_i$ up to depth $m+1$ ; the map is in $D^m$ exactly when all of them vanish.
To produce the coefficients, run the induction of the proof: recursively express $[P,x_i]$ in canonical form, apply the potential construction (3.19), and finish with the constant $P(1)$ .
For an operator already given as a word in $x_i$ and $∂_i$ , no such work is needed: rewrite with $∂_i x_i = x_i∂_i + 1$ until all $∂$ 's stand to the right, and read the order off the result.

Step 2 is a linear-algebra recursion whose cost is dominated by the number of multi-indices, $m+nn$ , and it is rarely needed in practice: operators arriving from applications are already presented as elements of $A_n$ . The theorem's computational value is negative rather than positive - it certifies that nothing is missing from the representation, so no algorithm needs to search outside $A_n$ .

Macaulay2's Dmodules, Singular's dmod.lib, and SageMath's ore_algebra all take $A_n$ over $ℚ$ as their model of differential operators on affine space, with the canonical form as the internal representation. A useful invariant to assert while testing: the order of $[P, x_i]$ must be exactly $ord(P) - 1$ unless the symbol of $P$ is independent of $ξ_i$ , which by (3.17) is the statement that $∂ σ(P)/∂ ξ_i = 0$ .

Limits of Validity

Characteristic zero. The theorem is false in characteristic $p$ ; the failure is not a technicality but changes the ring.
Polynomial ring only. The proof does not extend to quotients. For $R = K[X]/J$ both lemmas break: the monomials are no longer independent and the potential construction has no well-defined target.
No statement about other regular rings from this proof. The result for general regular $K$ -algebras is true but is a different theorem, proved by localising to the smooth local case and using a regular system of parameters.
Finiteness of the number of variables. The descending induction in the potential lemma runs over $i = n, n-1, … , 1$ and needs $n$ finite.

Failure Modes and Common Mistakes

Assuming the easy inclusion is the whole theorem

$A_n ⊂ eq D(K[X])$ is immediate from the composition rule. The content is the reverse inclusion, and it is not formal: it needs both lemmas and it needs characteristic zero. A proof that only checks that $x_i$ and $∂_i$ have finite order has proved nothing.

Dropping the compatibility condition in the potential lemma

Without $[P_i,x_j] = [P_j,x_i]$ there is generally no potential; the pair $P_1 = ∂_2$ , $P_2 = 0$ is a counterexample. The condition is automatic for the family $P_i = [P,x_i]$ arising in the proof, which is why it is easy to forget that it is a hypothesis.

Carrying the theorem into characteristic $p$

In characteristic $p$ the operator $x^m ↦ mpx^m-p$ is a differential operator on $K[x]$ that is not in $A_1$ . Statements such as \"a D-module is a module over the Weyl algebra\" silently become false, and results depending on simplicity of $A_n$ fail as well since $x^p$ and $∂^p$ become central.

Extending the theorem to singular coordinate rings

The analogous statement for $O(X)$ with $X$ singular is false. On the cusp $K[t^2,t^3]$ there is an operator of order $2$ that is not a polynomial in the ring and its derivations. Anything proved here should be quoted with the hypothesis that the variety is smooth.

Confusing order with degree in the theorem's statement

$D^m(K[X]) = C_m$ bounds $|β|$ , not $|α| + |β|$ . The coefficients $f_β$ are unrestricted, so $D^m$ is infinite dimensional over $K$ for every $m ≥ 0$ . The Bernstein filtration is the one that bounds both.

Historical Notes

The Weyl algebra entered mathematics through quantum mechanics: the canonical commutation relation $[ p, q] = -iℏ$ of Heisenberg and Born, formalised by Hermann Weyl in the 1920s, is the relation $[∂, x] = 1$ up to a constant. The elementary observation that this relation cannot be realised by finite matrices - take traces of $[∂, x] = 1$ - is the statement that $A_1$ has no finite-dimensional representations in characteristic zero.

The intrinsic definition of a differential operator came much later, with Grothendieck's EGA IV in 1967. The theorem on this page is the compatibility statement between the two developments: the algebra that physics wrote down by hand is the algebra that algebraic geometry produces from the polynomial ring with no input beyond its ring structure. In the algebraic literature it is usually attributed to Grothendieck as the affine-space case of the description of $𝒟$ for smooth schemes.

The failure in characteristic $p$ was noticed early and turned into a theory rather than an obstacle: the divided power operators form the ring of Hasse-Schmidt differential operators, and the resulting theory of $F$ -modules and Frobenius descent, developed by Berthelot, Lyubeznik, Smith and others, is now a substantial subject in its own right. Coutinho's Chapter 3 restricts to characteristic zero and is explicit that this is a hypothesis, not a convenience.

Comparison

The two lemmas, what they do, and what they need.
	Rigidity lemma (3.2.1)	Potential lemma (3.2.2)
Statement	$[P,x_i] = 0$ for all $i$ implies $P ∈ K[X]$	compatible $P_i$ admit $Q$ with $[Q,x_i] = P_i$
Role in the proof	pins down the constant of integration	reconstructs the operator from its commutators
Classical analogue	a function with zero gradient is constant	a curl-free field has a potential
Needs char $0$	no	yes, division by $β_k + 1$
Needs the canonical basis	no	yes, to force multi-indices to drop
Generalises to regular rings	readily	only with more work

Key Takeaways

Key points

For $K$ of characteristic zero, $D(K[x_1, … ,x_n]) = A_n(K)$ and $D^m = C_m$ , the operators with $|β| ≤ m$ .
The inclusion $A_n ⊂ eq D$ is immediate; the content is the converse.
The rigidity lemma says an operator commuting with every coordinate is multiplication by a polynomial.
The potential lemma says a family $P_1, … ,P_n$ with $[P_i,x_j] = [P_j,x_i]$ is $[Q,x_i]$ for a single $Q$ of one order higher - a discrete Poincaré lemma.
Characteristic zero is used exactly once, in dividing by $β_k + 1$ when building the potential, and the theorem fails without it.
Consequences: the order of $∑ f_β∂^β$ is $|β|$ , $D^m$ is free of rank $m+nn$ over $K[X]$ , and $gr,A_n = K[x,ξ]$ .
The theorem is about affine space; it fails on singular varieties and holds for general regular rings only by a different proof.

FAQs

Why is the inclusion

A_n ⊂ eq D(K[X])

easy?

Because multiplication operators have order $0$ , each $∂_i$ has order $1$ since $[∂_i, f] = ∂_i(f)$ is again a multiplication, and composition adds orders. So every word in the generators has finite order.

Where exactly does the proof use characteristic zero?

In one place: the potential $Q'' = ∑ _β (β_k+1)^-1g_β ∂^β+e_k$ divides by positive integers. In characteristic $p$ that inverse does not exist when $p ∣ β_k + 1$ , and the theorem genuinely fails.

What is the compatibility condition really saying?

That mixed second commutators agree, which is the algebraic form of the symmetry of mixed partial derivatives. It is automatic for $P_i = [P,x_i]$ because $ad_x_i$ and $ad_x_j$ commute, and it is necessary for a potential to exist.

Does the theorem hold for

K[x]/(f)

Not in general. It holds when the quotient is regular, which for a hypersurface means smooth. For the cusp $K[t^2,t^3]$ it fails; see differential operators on an affine variety.

How does the theorem determine the order filtration?

The induction proves $D^m = C_m$ level by level, so the order of an operator in canonical form is the largest $|β|$ with $f_β ≠ 0$ . Coefficients are irrelevant to order: $x_1^1000∂_1$ has order $1$ .

Is the potential

Q

unique?

Only up to $K[X]$ . If $Q$ and $Q^{'}$ both work then $[Q - Q', x_i] = 0$ for all $i$ , so $Q - Q'$ is multiplication by a polynomial by the rigidity lemma. That is precisely the constant of integration, and it is fixed in the theorem by evaluating at $1$ .

What replaces the theorem in characteristic

p

$D(K[X])$ is generated by all divided power operators $∂_i^[k]$ , which act on monomials by binomial coefficients. The resulting ring is not Noetherian and not finitely generated as a $K$ -algebra, and it is a union of an increasing chain of matrix-like subalgebras over Frobenius powers.

Does this mean D-modules on affine space are just

A_n

-modules?

Yes, and that is the point of the theorem. The intrinsic definition and the concrete one give the same category, so results proved about $A_n$ -modules are results about D-modules on $K^n$ with no further justification needed.

References

S. C. Coutinho, A Primer of Algebraic D-modules, London Mathematical Society Student Texts 33, Cambridge University Press, 1995 - Ch. 3 §2, Lemmas (3.2.1) and (3.2.2) and Theorem (3.2.3).
A. Grothendieck, Éléments de géométrie algébrique IV, Publications Mathématiques de l'IHÉS 32 (1967), §16 - the description of $𝒟$ for smooth schemes.
J. C. McConnell and J. C. Robson, Noncommutative Noetherian Rings, Graduate Studies in Mathematics 30, revised edition, American Mathematical Society, 2001 - Ch. 15, for regular rings.
J.-E. Björk, Rings of Differential Operators, North-Holland Mathematical Library 21, North-Holland, 1979 - Ch. 1.
J. Dixmier, Sur les algèbres de Weyl, Bulletin de la Société Mathématique de France 96 (1968), 209-242.
P. Berthelot and A. Ogus, Notes on Crystalline Cohomology, Princeton University Press, 1978 - for divided power structures and differential operators in characteristic $p$ .
R. Hotta, K. Takeuchi and T. Tanisaki, D-modules, Perverse Sheaves, and Representation Theory, Progress in Mathematics 236, Birkhäuser, 2008 - Ch. 1, for $D_X$ on smooth varieties.
ISO 80000-2:2019, Quantities and units - Part 2: Mathematics, International Organization for Standardization.
ISO/IEC 40314:2016, Information technology - Mathematical Markup Language (MathML) Version 3.0, International Organization for Standardization.

AI Suggested Questions

Prove $[∂^β, x_i] = β_i ∂^β - e_i$ by induction on $|β|$ from $[∂_i, x_j] = δ_ij$ .
Run the potential construction for $n = 2$ with $P_1 = x_2∂_2$ , $P_2 = x_2∂_1$ , first checking whether the compatibility condition holds.
Show that the rigidity lemma fails if one only assumes $[P, x_1] = 0$ in two variables, by exhibiting an operator that is not a polynomial.
Write out the potential lemma for $r = 1$ and identify it as the statement that a closed polynomial $1$ -form on affine space is exact.
Verify that $∂^[2] : x^m ↦ m2x^m-2$ is a differential operator of order $2$ on $K[x]$ in characteristic $2$ , and that it is not in $A_1$ .
Compute the rank of $D^m(K[x_1,x_2,x_3])$ over $K[x_1,x_2,x_3]$ for $m ≤ 3$ and check against $m+33$ .
Explain why the potential lemma is a cohomological statement, and identify the group whose vanishing it expresses.