The D-module Attached to a Differential Equation

Q: What does it mean when the module is zero?

That 1 is a consequence of the system, so the only solution in any module is 0 . This is the algebraic form of an inconsistent system, and it can happen without any equation looking contradictory, since commutators of the given operators are also consequences.

Q: Can inhomogeneous equations be handled the same way?

Not by the same quotient. The solution set of Pf = g is a coset of the solution set of Pf = 0 , so it is not a module. One standard device replaces the pair by a short exact sequence and records the inhomogeneity as an extension class; another enlarges the system so that the inhomogeneous term becomes one of the unknowns.

Q: How does the order of the equation show up in the module?

Not directly. The Bernstein degree of P appears as the multiplicity of A_n/A_nP , and the order of P appears as the holonomic rank after the coefficients are localised to rational functions. The order itself is not an invariant of the module, since a change of generator can change it.

Q: Why is one partial differential equation so much worse behaved than one ordinary one?

Because A_n/A_nP has dimension 2n-1 . For n = 1 this is 1 , the holonomic value, and the solution space is finite-dimensional. For n ≥ 2 it exceeds n , the module is not holonomic, and there is no finiteness theorem to appeal to - as the rotation operator x_1∂_2 - x_2∂_1 , with its infinite-dimensional polynomial solution space, shows.

Q: Where do distributions and hyperfunctions enter?

Only as choices of the target module S . Nothing about M changes when the target is enlarged, but the solution space can jump from zero to non-zero, which is what happens for xf = 0 when S is enlarged to contain the Dirac delta.

Overview

A linear differential equation with polynomial coefficients, $∑ _α g_α ∂^α f = 0$ with $g_α ∈ K[x_1, … ,x_n]$ , contains two quite different kinds of information. There is the operator $P = ∑ _α g_α ∂^α$ , a finite piece of algebraic data living in the Weyl algebra $A_n$ ; and there are the functions that satisfy the equation, which may be polynomials, smooth functions, holomorphic functions, distributions, or objects with no classical meaning at all.

The algebraic theory of differential equations begins by separating the two. This page carries out the first half of that separation: to a system $P_1 f = … = P_m f = 0$ it attaches a single left $A_n$ -module $M = A_n / ∑ _i A_n P_i$ , built entirely out of the operators and mentioning no function whatever. The second half - recovering the solutions in a chosen space by taking homomorphisms out of $M$ - is on the solutions page.

The construction is short, but it is not arbitrary. The left ideal $J = ∑ _i A_n P_i$ is precisely the set of differential consequences of the system: the operators that can be deduced from the given ones by composing on the left and adding. Two systems with the same consequences give literally the same module, which is exactly the identification a solution theory ought to make. The module is what survives when a system is stripped of everything that depends on how it was written down.

What the module gains in return is that it is an object of a good category. It is finitely generated, indeed cyclic; it has a dimension and a multiplicity; it can be pushed and pulled along maps of affine spaces. Numerical facts about the original equation, such as how many independent solutions to expect, become invariants of $M$ .

Definition

Throughout, $K$ is a field of characteristic zero, $K[X] = K[x_1, … ,x_n]$ , and $A_n = A_n(K)$ is the $n$ -th Weyl algebra, generated by $x_1, … ,x_n$ and $∂_1, … ,∂_n$ subject to $[∂_i,x_j] = δ_ij$ and the commuting relations among the $x$ 's and among the $∂$ 's.

The module of a systemCoutinho, Ch. 6 §1

Let $P_1, … ,P_m ∈ A_n$ and consider the homogeneous linear system

P_1(f) = P_2(f) = … = P_m(f) = 0.

(6.1)

Let $J = A_n P_1 + … + A_n P_m$ be the left ideal of $A_n$ generated by the $P_i$ . The $D$ -module of the system (6.1) is the left $A_n$ -module

M = A_n / J,

(6.2)

with its canonical generator $u = 1 + J$ . By construction $P_i u = 0$ for every $i$ , and $M = A_n u$ is cyclic.

Degenerate case

$M = 0$ if and only if $J = A_n$ , that is, if and only if $1$ is a differential consequence of the system. Such a system is inconsistent: since $1 ⋅ f = f$ , its only solution in any $A_n$ -module is $0$ . For instance $P_1 = x∂$ and $P_2 = ∂ x$ generate the whole of $A_1$ , because $∂ x - x ∂ = 1$ .

Annihilator form of the same idea

If $S$ is any left $A_n$ -module and $s ∈ S$ , the annihilator $Ann(s) = Q ∈ A_n : Qs = 0$ is a left ideal, and $A_n s ≅ A_n/Ann(s)$ . So every cyclic module is the module of some system, and the system is recovered as any finite generating set of $Ann(s)$ - which exists because $A_n$ is Noetherian.

Core Concepts

Why a left ideal and not a right or a two-sided one

If $f$ satisfies $Pf = 0$ then $f$ satisfies $(QP)f = 0$ for every operator $Q$ , because $(QP)f = Q(Pf) = Q(0) = 0$ . Nothing of the kind holds on the other side: from $Pf = 0$ one cannot conclude $(PQ)f = 0$ , since $Q f$ need not be a solution. The set of operators that every solution of (6.1) must satisfy is therefore closed under left multiplication and addition, and that is exactly the definition of a left ideal.

Two-sided ideals are useless here for a sharper reason: $A_n$ is a simple ring, so its only two-sided ideals are $0$ and $A_n$ . Quotienting by the two-sided ideal generated by a non-zero $P$ would always give the zero module, destroying all information. The whole theory depends on working with one-sided ideals in a noncommutative ring.

What the module remembers, and what it forgets

The passage from the list $(P_1, … ,P_m)$ to $M$ forgets the list and remembers the ideal. Adding a consequence to the system changes nothing: the system ${P, Q P}$ has the same ideal as ${P}$ . Multiplying an equation by a unit of $A_n$ , or replacing the list by any other generating set of $J$ , changes nothing either. This is the correct notion of equivalence for systems, and it is imposed automatically rather than by hand.

What is forgotten is any individual solution. $M$ is built from operators, and the same $M$ serves for polynomial solutions, smooth solutions and distributional solutions alike. To get solutions back one must name a target module $S$ and compute $Hom_A_n(M,S)$ ; different choices of $S$ give genuinely different answers for the same $M$ .

The generator is the unknown function

It is useful to read the canonical generator $u$ as the unknown $f$ itself, subjected to the relations of the system and to nothing else. Formally, $u$ is the universal solution: it satisfies the system in $M$ , and every solution in every module is obtained from $u$ by a unique homomorphism. In this reading $M$ is the free object on one unknown modulo the equations, exactly as a group given by generators and relations is a free group modulo relations.

Construction and Proof

Two things need to be checked: that the construction has a universal property, and that nothing is lost when a module is turned back into a system.

Universal property of the canonical generator

Let $M = A_n/J$ with $J = ∑ _i A_n P_i$ and let $u$ be the canonical generator. For every left $A_n$ -module $S$ and every $s ∈ S$ with $P_i s = 0$ for $i = 1, … ,m$ , there is a unique homomorphism of left $A_n$ -modules $φ : M \to S$ with $φ(u) = s$ .

Proof

Uniqueness is immediate: $M = A_n u$ , so $φ(Qu) = Qφ(u) = Qs$ is forced for every $Q ∈ A_n$ . For existence, define $φ : A_n → S$ by $φ(Q) = Qs$ ; this is a homomorphism of left modules because $φ(RQ) = RQs = R φ(Q)$ . Its kernel contains each $P_i$ , hence contains the left ideal they generate, hence contains $J$ . So $\tilde{φ}$ factors through $A_n/J = M$ , and the induced map sends $u = 1 + J$ to $s$ .

Taking $S$ and $s$ arbitrary, this proposition is already the solution theorem of the next page in disguise: homomorphisms out of $M$ and solutions in $S$ are the same data. It also shows that $M$ does not depend on the choice of generators of $J$ , since the universal property mentions only $J$ .

From a module back to a system

Suppose $M$ is a cyclic left $A_n$ -module with generator $u$ . Then $Ann(u)$ is a left ideal, and $A_n$ is left Noetherian, so $Ann(u) = ∑ _i=1^m A_n P_i$ for finitely many operators $P_i$ . The system $P_1 f = … = P_m f = 0$ has $M$ as its module. The dictionary is therefore a genuine two-way translation, at the cost of one choice: the generator.

A different generator gives a different ideal

Take $M = K[x]$ as an $A_1$ -module. With the generator $u = 1$ we get $Ann(1) = A_1∂$ , so $K[x] ≅ A_1/A_1∂$ : the equation is $f' = 0$ .

But $x$ also generates $K[x]$ , because $∂ ⋅ x = 1$ . Its annihilator contains $∂^2$ and $x∂ - 1$ , and in fact equals $A_1∂^2 + A_1(x∂ - 1)$ ; the system is now $f'' = 0$ , $xf' = f$ . A dimension count confirms there is nothing more in the annihilator: the symbols $ξ^2$ and $x ξ$ cut the associated graded ring $K[x,ξ]$ down to the span of the monomials $x^a$ and $ξ$ , whose count in degree at most $m$ is $m+2$ , and $m+2$ is also $_K B_m! ⋅ ! x$ , the dimension of the corresponding filtration piece of $K[x]$ .

So one module carries two visibly different systems. The module is the invariant; the equations are a presentation of it.

Key Equations

An operator of $A_n$ acts on a function by the rule that gives the Weyl algebra its name:

P = ∑ _α ∈ N^n g_α(x), ∂^α, P(f) = ∑ _α g_α(x), ∂^|α| f∂ x^α,

(6.3)

the sum being finite; here $α$ is a multi-index and $g_α ∈ K[X]$ .

The ideal of consequences and the module are

J = ∑ _i=1^m A_n P_i, M = A_n/J = A_n u, P_i u = 0.

(6.4)

Cyclicity is the statement that $M$ has a presentation with a single generator, that is an exact sequence

A_n^m ⟶ ⋅ ,P A_n ⟶ M ⟶ 0, (Q_1, … ,Q_m) ↦ ∑ _i Q_i P_i.

(6.5)

The universal property, which is proved above and exploited on the next page, reads

Hom_A_n(M, S) ; ≅ ; , s ∈ S : P_1 s = … = P_m s = 0 ,

(6.6)

as $K$ -vector spaces, for every left $A_n$ -module $S$ .

For a single non-zero operator in one variable the invariants of $M$ can be written down. If $P ∈ A_n$ is non-zero of Bernstein degree $d$ , so that its symbol $σ(P) ∈ K[x_1, … ,x_n,ξ_1, … ,ξ_n]$ is homogeneous of degree $d$ , then

gr(A_n/A_nP) ; ≅ ; K[x_1, … ,x_n,ξ_1, … ,ξ_n]/(σ(P)), d(M) = 2n-1, e(M) = d.

(6.7)

Variable Definitions

$K$: the ground field, of characteristic zero
$K[X]$: the polynomial ring $K[x_1, … ,x_n]$ , the default space of solutions
$A_n$: the $n$ -th Weyl algebra over $K$
$P_i$: the operators of the system, elements of $A_n$
$J$: the left ideal $∑ _i A_n P_i$ of differential consequences
$M$: the module $A_n/J$ of the system
$u$: the canonical generator $1 + J$ of $M$ , the universal unknown
$S$: a chosen left $A_n$ -module in which solutions are sought
$σ(P)$: the symbol of $P$ : its top-degree part with $∂_i$ replaced by the commuting variable $ξ_i$
$d(M),, e(M)$: the dimension and the multiplicity of the module, computed from a good filtration

Properties and Behaviour

Finiteness

$M = A_n/J$ is cyclic, hence finitely generated, and since $A_n$ is left Noetherian $J$ is finitely generated, so $M$ is finitely presented. Every quotient and every submodule of $M$ is again finitely generated.

Dimension and multiplicity of a single equation

Let $0 ≠ P ∈ A_n$ have Bernstein degree $d$ and let $M = A_n/A_nP$ . Then $d(M) = 2n - 1$ and $e(M) = d$ .

Sketch: the associated graded ring of $A_n$ for the Bernstein filtration is the polynomial ring in $2n$ variables, which is a domain, so degrees add: $(QP) = Q + P$ . Hence $A_nP ∩ B_m = B_m-dP$ , the associated graded of $A_nP$ is the principal ideal $(σ(P))$ , and $gr M$ is the coordinate ring of a hypersurface of degree $d$ in $2n$ variables. Its Hilbert polynomial has degree $2n-1$ and leading coefficient $d/(2n-1)!$ .

One variable is special

For $n = 1$ and $0 ≠ P ∈ A_1$ , the module $A_1/A_1P$ has $d(M) = 1$ , so it is holonomic, with multiplicity equal to the Bernstein degree of $P$ . For $n ≥ 2$ a single non-trivial equation gives $d(M) = 2n-1 > n$ , so a single equation in several variables is never holonomic. This is the algebraic reason why one ordinary differential equation has a finite-dimensional solution space while one partial differential equation generally does not.

Adding equations can only shrink the module

If $J ⊂ eq J'$ then there is a canonical surjection $A_n/J ↠ A_n/J'$ , and by (6.6) the solution space of the larger system embeds in that of the smaller one. Imposing more equations makes the module smaller and the solution space smaller, in the same direction - the correspondence between systems and modules is order reversing on ideals and order preserving on modules.

Examples and Special Cases

The polynomial ring

$n = 1$ , $P = ∂$ . Then $M = A_1/A_1∂ ≅ K[x]$ , with $u$ corresponding to the constant $1$ . Indeed every operator reduces modulo $A_1∂$ to a polynomial in $x$ , and the map $Q ↦ Q(1)$ is an isomorphism. See the polynomial module. Polynomial solutions of $f' = 0$ : the constants, a one-dimensional space.

An equation with no classical solution

$n = 1$ , $P = x$ . The equation $xf = 0$ has no non-zero classical solution at all: not a polynomial, not a continuous function, not a smooth one, because a function vanishing away from the origin vanishes identically. Yet $M = A_1/A_1x$ is a perfectly good non-zero module, with $K$ -basis the classes of $∂^b$ , $b ≥ 0$ . Its solution appears once the target module is enlarged: the Dirac delta satisfies $xδ = 0$ , and $Ann(δ) = A_1x$ , so $A_1δ ≅ M$ . This is worked out in the delta page and the microfunctions page.

The exponential

$n = 1$ , $P = ∂ - 1$ . Again no non-zero polynomial solves $f' = f$ , but $(x)$ does in the module of holomorphic functions. The module $A_1/A_1(∂-1)$ is a twist of $K[x]$ by an automorphism of $A_1$ , and is simple.

Two commuting equations in two variables

$n = 2$ , $P_1 = ∂_1$ , $P_2 = ∂_2$ . Then $J = A_2∂_1 + A_2∂_2$ and $M = A_2/J ≅ K[x_1,x_2]$ . Polynomial solutions: the constants. Here two equations in two variables produce a holonomic module of dimension $2$ , whereas either equation alone would give dimension $3$ .

An equation with infinitely many independent solutions

$n = 2$ , $P = x_1∂_2 - x_2∂_1$ , the generator of infinitesimal rotations. A polynomial $f$ satisfies $Pf = 0$ precisely when it is invariant under rotation, and over a field where this can be said cleanly the solutions are the polynomials in $x_1^2 + x_2^2$ . The solution space is infinite-dimensional, in agreement with the corollary above: $d(M) = 3$ , so $M$ is not holonomic.

Worked Example

The Euler operator: $M = A_1/A_1(x∂ - λ)$

Step 1 - the module and a basis
Fix $λ ∈ K$ and let $P = x∂ - λ$ , so the equation is $xf' = λ f$ . Put $M = A_1/A_1P$ and $u = 1 + A_1P$ , so that $x∂, u = λ u$ .

The Bernstein symbol of $P$ is $σ(P) = xξ$ , of degree $2$ . By (6.7), $gr M ≅ K[x,ξ]/(xξ)$ , whose monomial basis consists of the $x^a$ with $a ≥ 0$ and the $ξ^b$ with $b ≥ 1$ . Lifting, a $K$ -basis of $M$ is

$, x^a u : a ≥ 0 , ; ∪ ; , ∂^b u : b ≥ 1 ,.$

Every mixed monomial reduces: $x∂ u = λ u$ , and $∂ x u = (x∂ + 1)u = (λ+1)u$ .
Step 2 - the Hilbert function, checked by hand
Filter $M$ by $Γ_m = B_m u$ . Counting the basis elements of degree at most $m$ gives $(m+1) + m = 2m+1$ . Check the first three values directly.

$m = 0$ : $Γ_0 = Ku$ , dimension $1$ . $m = 1$ : $B_1 = ⟨ 1, x, ∂ ⟩$ and $B_1 ∩ A_1P = 0$ because every non-zero element of $A_1P$ has degree at least $2$ ; so $Γ_1 = 3$ . $m = 2$ : $B_2$ has the six monomials $1, x, ∂, x^2, x∂, ∂^2$ , and $B_2 ∩ A_1P = KP$ ; so $Γ_2 = 5$ . The values $1, 3, 5$ match $2m+1$ .
Step 3 - dimension and multiplicity
The Hilbert polynomial is $χ(m) = 2m+1$ , of degree $1$ with leading coefficient $2$ . Hence $d(M) = 1$ and $e(M) = 1! ⋅ 2 = 2$ , confirming (6.7) with $n = 1$ and $d = 2$ . Since $d(M) = n = 1$ , the module is holonomic.
Step 4 - polynomial solutions
By (6.6), $Hom_A_1(M, K[x])$ is the space of $f ∈ K[x]$ with $xf' = λ f$ . Write $f = ∑ _a c_a x^a$ ; then $xf' = ∑ _a a, c_a x^a$ , so the equation says $a,c_a = λ c_a$ for every $a$ . Thus $c_a = 0$ unless $a = λ$ .

So the polynomial solution space is $0$ unless $λ$ is a non-negative integer, in which case it is the line spanned by $x^λ$ . A single one-dimensional module can therefore have solution space of dimension $1$ or $0$ depending on a parameter, while the module itself is non-zero for every $λ$ .
Step 5 - solutions elsewhere
Over $K = C$ , take instead $S = H(U)$ with $U = C ∖ (- ∈ fty,0]$ , where a branch of $z^λ$ is defined. Then $z^λ$ solves $zf' = λ f$ for every $λ$ , and $Hom_A_1(M,H(U))$ is one-dimensional for every $λ$ . Enlarging the target module has changed the answer without changing $M$ .

Result

$M = A_1/A_1(x∂-λ)$ is holonomic with $d(M) = 1$ and $e(M) = 2$ , independently of $λ$ . Its polynomial solution space is one-dimensional when $λ ∈ N$ and zero otherwise; its solution space in holomorphic functions on a slit plane is one-dimensional for every $λ$ . The module encodes the equation; the target module decides which solutions exist.

Applications and Industry Use

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

Turning a system into a module is the step that makes symbolic computation with differential equations possible, because ideal membership in $A_n$ is decidable while asking whether a function is a solution is not.

Deciding consequences. Whether an operator $Q$ is a consequence of a system is the question $Q ∈ J$ , answered by a Gröbner basis computation in the Weyl algebra. Whether a system is contradictory is the question $1 ∈ J$ .
Counting solutions in advance. The invariants $d(M)$ and $e(M)$ , and the holonomic rank after localising the coefficients, bound the number of independent solutions of the system near a generic point, without solving anything.
Automatic proof of identities. Zeilberger's method works by attaching a module to a summand or integrand and manipulating the annihilating ideal; see Zeilberger's method and holonomic functions.
Special functions. A special function is stored in computer algebra as a generator together with its annihilating ideal - that is, as a cyclic $D$ -module with a chosen generator - which is a finite and exact description of an object with no closed form.
Linear systems theory. In the behavioural approach to multidimensional linear systems, a system of linear equations with operator coefficients is identified with the finitely presented module it presents, and system-theoretic properties such as controllability and autonomy become module-theoretic ones. The same translation, one level of generality up, is the one made here.

Design Considerations

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

Three modelling choices are made before the module is written down, and each is reversible only at a cost.

Which ring of operators

Working in $A_n$ keeps the coefficients polynomial and the ring Noetherian and simple, which is what makes the dimension theory work. Working instead in $B_n = K(x) ⟨ ∂_1, … ,∂_n ⟩$ , with rational function coefficients, makes the ring a left and right Euclidean domain in the case $n = 1$ and turns a holonomic module into a finite-dimensional vector space over $K(x)$ whose dimension is the holonomic rank. The price is that all information carried by the singular locus is discarded. For analytic questions one moves further, to rings of operators with convergent or formal power series coefficients.

Which side

The convention here is left modules acting on functions, which forces left ideals. Right modules appear naturally for densities and for integration, and the transposition anti-automorphism converts one to the other; see side-changing functors. Mixing the conventions in one computation is the single most common source of sign errors.

One unknown or several

A system in a single unknown function gives a cyclic module. A system in several unknowns gives a quotient of a free module of higher rank, described by a presentation matrix; that is the subject of the presentation matrix page. Since every holonomic module is in fact cyclic, the distinction is one of convenience rather than of substance in the holonomic case.

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.

All the basic questions about $M = A_n/J$ reduce to Gröbner basis computations in the Weyl algebra, which is one of the few noncommutative rings where the theory works essentially as in the commutative case.

Fix a monomial order on the $x^α∂^β$ compatible with the chosen filtration, and compute a Gröbner basis $G$ of $J$ from the given $P_i$ .
Consistency: $M = 0$ if and only if $G$ contains a unit, that is if and only if $1$ reduces to $0$ .
Membership: $Q$ is a consequence of the system if and only if $Q$ reduces to $0$ modulo $G$ .
Invariants: the leading terms of $G$ generate the associated graded ideal, so the Hilbert polynomial of the commutative quotient gives $d(M)$ and $e(M)$ directly.
Solutions: polynomial and rational solutions of the system can be enumerated by further algorithms once the module is known to be holonomic.

Costs are those of Gröbner bases in $2n$ variables, doubly exponential in the worst case. Implementations include the Dmodules package of Macaulay2, dmod.lib in Singular, the ore_algebra package in SageMath and HolonomicFunctions.m in Mathematica. A useful sanity check on any implementation: a reported dimension below $n$ for a non-zero module contradicts Bernstein's inequality and indicates a bug.

Limits of Validity

The construction is narrower than the phrase differential equation suggests, and the hypotheses matter.

Linear only. Nonlinear equations have no left ideal of consequences: the set of operators annihilating a solution is not stable under the operations one would need. Nonlinear problems belong to differential algebra and to the theory of $D$ -schemes, not here.
Homogeneous only. An inhomogeneous equation $Pf = g$ has a solution set that is a coset, not a subspace. Module theory handles it through the module $A_n/A_nP$ together with a class in an $Ext$ group, or by enlarging the system; the naive quotient encodes only the homogeneous part.
Polynomial coefficients. Equations with analytic or smooth coefficients are not operators in $A_n$ . The formalism transfers to the corresponding operator rings, but the finiteness statements that make the theory usable have to be re-proved there.
Solutions require a target. $M ≠ 0$ does not mean the system has a non-zero solution. The equation $xf = 0$ gives a non-zero module and no classical solution at all. The module is a question, not an answer.

Failure Modes and Common Mistakes

Using the right ideal, or the two-sided ideal

Only left multiplication preserves solutions: $Pf = 0$ implies $QPf = 0$ , never $PQf = 0$ . Forming $PA_n$ or $A_nPA_n$ gives an object with no relation to the equation, and in the two-sided case the simplicity of $A_n$ makes the quotient $0$ for every non-zero $P$ . In an implementation, the check is whether the reduction routine multiplies candidate cofactors on the correct side.

Believing the module determines the equations

It determines the left ideal, given the generator. Change the generator and the ideal changes, as the example of $K[x]$ generated by $1$ or by $x$ shows: the systems $f' = 0$ and $f'' = 0, xf' = f$ present isomorphic modules. Any invariant of the equation that is to be read off the module must be checked to be independent of the presentation.

Reading the order of the operator as the number of solutions

For an ordinary equation of order $k$ with non-vanishing leading coefficient, the classical solution space near a non-singular point does have dimension $k$ . But the polynomial solution space can be smaller (it is $0$ for $f' = f$ ), and in several variables the count fails completely: one equation in $n ≥ 2$ variables typically has an infinite-dimensional solution space. The invariant that generalises the order is the holonomic rank, not the order.

Forgetting that a system of one equation may already be inconsistent

Inconsistency is not a property of visibly contradictory equations. The system $x∂, ∂ x$ looks harmless but generates all of $A_1$ , since the commutator of the two generators is $1$ . Always test $1 ∈ J$ before interpreting a computed invariant of $M$ ; the zero module has every dimension convention attached to it and no content.

Historical Notes

The idea of studying a system of linear partial differential equations through the module it presents grew out of two traditions. On the analytic side, Ehrenpreis and Palamodov in the late 1950s and 1960s treated constant-coefficient systems by identifying them with modules over a polynomial ring and applying commutative algebra, culminating in the fundamental principle. On the formal side, Riquier, Janet and later Spencer had developed the involutive theory of over-determined systems, in which passing to the consequences of a system - completion to involution - is precisely the step that produces the ideal used here.

The variable-coefficient version required a noncommutative ring, and it appeared in Kashiwara's 1970 master's thesis, where the module attached to a system was made the primary object and the solution space became a $Hom$ group, with the derived functors carrying the rest of the information. Sato's algebraic analysis and the Sato-Kashiwara-Kawai theory of microfunctions gave the analytic counterpart. Independently, Bernstein's work on analytic continuation of $f^s$ made the Weyl algebra itself a subject of study.

The algebraic presentation used here, in which everything is done over $A_n$ with no sheaves and no analysis, is the one adopted in Coutinho's primer; it is the shortest honest route to the main theorems and the one that computer algebra systems implement.

Comparison

The dictionary between the classical and the module-theoretic descriptions of a linear system.
Classical object	Module-theoretic object	Comment
Equation $Pf = 0$	Left ideal $A_nP$	the equation is the operator, up to left multiples
System $P_1f = … = P_mf = 0$	Left ideal $J = ∑ _i A_nP_i$	generators of the ideal are the equations
Differential consequence	Element of $J$	left multiply and add
The system itself	Module $M = A_n/J$	independent of how the system was written
The unknown function	Canonical generator $u = 1 + J$	the universal solution
A solution in a space $S$	Element of $Hom_A_n(M,S)$	one homomorphism per solution
Change of unknown	Change of generator of $M$	changes the ideal, not the module
Order of the equation	Multiplicity, or holonomic rank	the order is not an invariant of $M$
Inconsistent system	$M = 0$	equivalently $1 ∈ J$

Key Takeaways

Key points

A homogeneous linear system $P_1f = … = P_mf = 0$ with polynomial coefficients determines the left ideal $J = ∑ _i A_nP_i$ and the cyclic module $M = A_n/J$ .
The left ideal is exactly the set of differential consequences of the system, because $Pf = 0$ implies $QPf = 0$ but not $PQf = 0$ .
The canonical generator $u = 1+J$ is a universal solution: homomorphisms $M \to S$ correspond bijectively to solutions of the system in $S$ .
Every cyclic module arises this way, and the system is recovered from a finite generating set of the annihilator of a generator; changing the generator changes the system but not the module.
$M = 0$ exactly when $1 ∈ J$ , the algebraic form of inconsistency; a non-zero module still need not have solutions in a given target.
For a single non-zero $P ∈ A_n$ of Bernstein degree $d$ , $A_n/A_nP$ has dimension $2n-1$ and multiplicity $d$ : holonomic if and only if $n = 1$ .

FAQs

Why quotient the whole Weyl algebra rather than take the set of solutions directly?

Because the set of solutions depends on where you look for them, and is often infinite-dimensional or empty, while the quotient is a finitely presented algebraic object that exists before any function space is chosen. All the classical solution spaces are then recovered from it as $Hom$ groups.

Is the module determined by the system, or the system by the module?

The system determines the module completely. The module determines the system only after a generator is chosen, and different generators give different, equally valid, systems presenting the same module.

What does it mean when the module is zero?

That $1$ is a consequence of the system, so the only solution in any module is $0$ . This is the algebraic form of an inconsistent system, and it can happen without any equation looking contradictory, since commutators of the given operators are also consequences.

Can inhomogeneous equations be handled the same way?

Not by the same quotient. The solution set of $Pf = g$ is a coset of the solution set of $Pf = 0$ , so it is not a module. One standard device replaces the pair by a short exact sequence and records the inhomogeneity as an extension class; another enlarges the system so that the inhomogeneous term becomes one of the unknowns.

Does the construction need the coefficients to be polynomials?

The construction itself needs only a ring of operators. The theory needs more: over $A_n$ one has the Noetherian property, simplicity, and a dimension theory, and these are what turn the definition into theorems. Over rings of operators with smooth coefficients some of this fails.

How does the order of the equation show up in the module?

Not directly. The Bernstein degree of $P$ appears as the multiplicity of $A_n/A_nP$ , and the order of $P$ appears as the holonomic rank after the coefficients are localised to rational functions. The order itself is not an invariant of the module, since a change of generator can change it.

Why is one partial differential equation so much worse behaved than one ordinary one?

Because $A_n/A_nP$ has dimension $2n-1$ . For $n = 1$ this is $1$ , the holonomic value, and the solution space is finite-dimensional. For $n ≥ 2$ it exceeds $n$ , the module is not holonomic, and there is no finiteness theorem to appeal to - as the rotation operator $x_1∂_2 - x_2∂_1$ , with its infinite-dimensional polynomial solution space, shows.

Where do distributions and hyperfunctions enter?

Only as choices of the target module $S$ . Nothing about $M$ changes when the target is enlarged, but the solution space can jump from zero to non-zero, which is what happens for $xf = 0$ when $S$ is enlarged to contain the Dirac delta.

References

S. C. Coutinho, A Primer of Algebraic D-modules, London Mathematical Society Student Texts 33, Cambridge University Press, 1995 - Ch. 6 §1, with Theorem (6.1.2) and the exercises of Ch. 6 §4.
M. Kashiwara, Algebraic Study of Systems of Partial Differential Equations, master's thesis, University of Tokyo 1970; English translation, Mémoires de la Société Mathématique de France 63, 1995.
J.-E. Björk, Rings of Differential Operators, North-Holland Mathematical Library 21, North-Holland, 1979 - Ch. 1 and Ch. 5.
R. Hotta, K. Takeuchi and T. Tanisaki, D-modules, Perverse Sheaves, and Representation Theory, Progress in Mathematics 236, Birkhäuser, 2008 - Ch. 1, for systems and their modules.
M. Saito, B. Sturmfels and N. Takayama, Gröbner Deformations of Hypergeometric Differential Equations, Algorithms and Computation in Mathematics 6, Springer, 2000 - Ch. 1, for the algorithmic side.
L. Ehrenpreis, Fourier Analysis in Several Complex Variables, Wiley-Interscience, 1970, and V. P. Palamodov, Linear Differential Operators with Constant Coefficients, Springer, 1970 - the constant-coefficient prehistory.
U. Oberst, Multidimensional constant linear systems, Acta Applicandae Mathematicae 20 (1990), 1-175 - systems as finitely presented modules in control theory.
Macaulay2 Dmodules package documentation, and Singular dmod.lib, for computing with $A_n/J$ .
ISO 80000-2:2019, Quantities and units - Part 2: Mathematics, International Organization for Standardization.

AI Suggested Questions

Compute the annihilator in $A_1$ of the polynomial $x^2$ inside $K[x]$ and compare the resulting system with $f' = 0$ .
Show that $A_1/A_1(x∂ - λ)$ is simple exactly when $λ$ is not an integer, and describe the submodules otherwise.
Verify that $A_2/A_2(x_1∂_2 - x_2∂_1)$ has dimension $3$ by computing its Hilbert polynomial.
Give an example of two operators in $A_2$ that generate the unit left ideal without either being a unit.
Work out the module attached to the Cauchy-Riemann system and identify the operator that annihilates both unknowns.
Explain why the order of an operator is not an invariant of the module it presents, with an explicit change of generator.
Compute $d(M)$ and $e(M)$ for $M = A_2/A_2(∂_1^2 + ∂_2^2)$ and check the general formula.