Order of a Differential Operator and the Order Filtration

Q: Is the order of an operator the same as its degree?

No. In A_n the Bernstein degree of x^α∂^β is |α| + |β| , while its order is |β| . The operator x^10∂ has degree 11 and order 1 . Two different filtrations, two different numbers.

Q: Why must commutators drop the order?

Because in [PQ,a] = P[Q,a] + [P,a]Q each summand loses one order, and the Jacobi identity propagates this to [[P,Q],a] . The formal proof is a double induction on ord P + ord Q , given on the definition page.

Q: Can two operators have the same symbol?

Yes, and this is the normal situation: σ_n(P) = σ_n(Q) exactly when P - Q has order less than n . Symbols see the top order only, which is what makes them computable and also what makes them lossy.

Q: Does the order filtration have finite-dimensional pieces?

No. Already D^0(K[x]) = K[x] is infinite dimensional over K . Each piece is finitely generated as an R -module, which is the right finiteness statement here, and Hilbert functions must be taken relative to R .

Q: How does the order filtration give the characteristic variety?

Take a good filtration of a module M compatible with D^m , form gr,M over gr,D = K[x,ξ] , and take the radical of its annihilator. That ideal cuts out a subvariety of the cotangent space K^2n ; see the characteristic ideal and variety .

Q: Why is ord(PQ) = ord(P) + ord(Q) for the Weyl algebra but not in general?

Because gr,A_n is a polynomial ring and hence a domain, so the product of two non-zero symbols is non-zero. If gr,D(R) has zero divisors, principal symbols can multiply to zero and the order of the product drops.

Q: Does the definition of order depend on the ground field?

The definition does not, but the answer can. In characteristic p the operator x^m ↦ mp x^m-p on K[x] has order p and is not a polynomial in ∂ , so the graded ring is larger than K[x,ξ] .

Q: Is there an analogue of the order filtration for modules?

Yes: a filtration Γ_m of a module with D^nΓ_m ⊂ eq Γ_n+m . It is good when gr,M is finitely generated over gr,D , and good filtrations always exist for finitely generated modules.

Overview

A differential operator on $K[x]$ has an obvious order: the highest power of $∂$ appearing in it. That description depends on writing the operator in a normal form, and on there being a $∂$ to count. The inductive definition of P3 supplies an intrinsic replacement: the order of $P$ is the number of times one has to take commutators with elements of $R$ before arriving at zero.

Collecting the operators of order at most $n$ gives a chain of $R$ -modules $R = D^0(R) ⊂ eq D^1(R) ⊂ eq ⋅ s$ whose union is $D(R)$ . Two rules govern it. Composition adds order, so the chain is a filtration of the ring, not merely an increasing family. Commutators lose one order, so the successive quotients form a commutative graded ring.

That second rule is the technical fulcrum of the whole subject. It lets one attach to each operator a symbol, a genuinely commutative object, in such a way that products of operators correspond to products of symbols. Questions about a non-commutative ring become questions about a polynomial ring, where dimension theory is available; the answers are then pulled back. Characteristic varieties, dimension and multiplicity of D-modules are all built on this exchange.

For the Weyl algebra the picture is completely explicit: the order- $m$ piece is the set of operators $∑ _|β| ≤ m f_β ∂^β$ , a free $K[x_1, … ,x_n]$ -module of rank $m+nn$ , and the associated graded ring is $K[x_1, … ,x_n,ξ_1, … ,ξ_n]$ , the coordinate ring of the cotangent space. The worked example computes symbols and a bracket in $A_2$ and checks the order drop by hand.

Definition

Let $R$ be a commutative $K$ -algebra, $K$ of characteristic zero, with $R$ identified with the multiplication operators inside $End_K(R)$ , and $[P,a] = Pa - aP$ .

Order and the order filtrationCoutinho, Ch. 3 §1

With $D^-1(R) = 0$ , set $D^n(R) = P : [P,a] ∈ D^n-1(R) for all a ∈ R$ . The order of a non-zero $P ∈ D(R)$ is

ord(P) = , n ≥ 0 ;:; P ∈ D^n(R) ,.

(3.1)

The family $D^n(R)_n ≥ -1$ is the order filtration. By convention $ord(0) = - ∈ fty$ .

Symbol and associated graded ring

For $P ∈ D^n(R)$ the symbol of order $n$ is the class

σ_n(P) = P + D^n-1(R) ∈ gr^nD(R) := D^n(R)/D^n-1(R),

(3.2)

and $gr,D(R) = ⨁ _n ≥ 0gr^nD(R)$ is the associated graded ring, with multiplication induced by composition. If $ord(P) = n$ exactly, $σ_n(P) ≠ 0$ and is called the principal symbol; if $ord(P) < n$ then $σ_n(P) = 0$ , so the index must always be carried.

Note

The symbol is not a single object attached to $P$ but a pair: an integer $n$ and a class in degree $n$ . Writing $σ(P)$ without saying which $n$ is meant is a common source of error, particularly when adding operators of different orders.

Core Concepts

Order counts derivatives, not coefficients

In $A_n$ the operator $x_1^100∂_1$ has order $1$ , while $∂_1^2$ has order $2$ . The order filtration weighs $∂_i$ at $1$ and $x_i$ at $0$ . That is why its pieces are infinite dimensional over $K$ but finitely generated over $R$ : each $D^m$ is an $R$ -module of finite rank, and the coefficients are unbounded.

Why the graded ring is commutative

If $P$ has order $n$ and $Q$ order $m$ , then $P Q$ and $Q P$ both have order at most $n+m$ , and their difference $[P,Q]$ has order at most $n+m-1$ . So in $gr^n+m$ their symbols agree: $σ_n(P)σ_m(Q) = σ_m(Q)σ_n(P)$ . Non-commutativity is invisible at the top order; it is stored one level down, and that is exactly what the Poisson bracket records.

Symbols and the cotangent space

For $R = K[x_1, … ,x_n]$ the class of $∂_i$ in degree $1$ is a new variable $ξ_i$ , and $gr,A_n = K[x_1, … ,x_n,ξ_1, … ,ξ_n]$ . This is the coordinate ring of $K^2n = T^*K^n$ , the cotangent space, with $x$ the position coordinates and $ξ$ the momentum coordinates. The vanishing locus of the symbols of a system of operators is where the system degenerates; that locus is the characteristic variety, and it lives in the cotangent space precisely because the order filtration puts $ξ$ in degree $1$ and $x$ in degree $0$ .

Two filtrations, two purposes

The order filtration is intrinsic, functorial and geometric, but its pieces are infinite dimensional over $K$ . The Bernstein filtration weighs $x_i$ and $∂_i$ equally, has finite-dimensional pieces, and supports counting arguments, but exists only for the Weyl algebra and depends on coordinates. The dimension of a finitely generated module comes out the same either way, which is a theorem rather than a definition.

Construction and Proof

The two rules of (3.4) are proved on the definition page by induction on $n+m$ , using $[PQ,a] = P[Q,a] + [P,a]Q$ and the Jacobi identity. Here is what they give.

The graded ring is commutative

$gr,D(R)$ is a commutative graded $R$ -algebra, and (3.5) holds: the symbol map is multiplicative in the sense that $σ_n+m(PQ) = σ_n(P)σ_m(Q)$ for $P ∈ D^n$ , $Q ∈ D^m$ .

Proof

Multiplicativity is the definition of the induced product: the class of $P Q$ in $D^n+m/D^n+m-1$ depends only on the classes of $P$ and $Q$ , because changing $P$ by an element of $D^n-1$ changes $P Q$ by an element of $D^n+m-1$ . Commutativity is the second rule of (3.4): $PQ - QP ∈ D^n+m-1$ , so the two products have the same class in degree $n+m$ .

The induced bracket

Setting $σ_n(P), σ_m(Q) = σ_n+m-1([P,Q])$ gives a well-defined $K$ -bilinear map $gr^n × gr^m → gr^n+m-1$ which is antisymmetric, satisfies the Jacobi identity, and is a derivation in each argument. It makes $gr,D(R)$ a Poisson algebra.

Proof

Well-definedness: if $P' - P ∈ D^n-1$ then $[P',Q] - [P,Q] ∈ D^n+m-2$ by (3.4), so the class in degree $n+m-1$ is unchanged; the same on the other side. Antisymmetry and Jacobi are inherited from the commutator. The Leibniz property in each argument follows from $[P, QQ'] = [P,Q]Q' + Q[P,Q']$ on passing to symbols.

The order filtration on the Weyl algebraCoutinho (3.2.3)

For $R = K[x_1, … ,x_n]$ with $K$ of characteristic zero, $D^m(R)$ is exactly the set of operators $∑ _|β| ≤ m f_β ∂^β$ , it is free over $K[X]$ on the $m+nn$ monomials $∂^β$ with $|β| ≤ m$ , and $gr,A_n ≅ K[x_1, … ,x_n,ξ_1, … ,ξ_n]$ .

Outline

That every such sum lies in $D^m$ follows from (3.4) and $ord(∂_i) = 1$ . The reverse inclusion is the substantial half and is proved on the next page using two lemmas: an operator commuting with every $x_i$ lies in $K[X]$ , and a system of operators $P_i$ satisfying the compatibility $[P_i, x_j] = [P_j, x_i]$ admits a potential $Q$ with $P_i = [Q, x_i]$ .

Freeness is the statement that the $∂^β$ are linearly independent over $K[X]$ , which is the canonical basis theorem for $A_n$ . Given that, $gr^m$ is free over $K[X]$ on the classes of the $∂^β$ with $|β| = m$ , and the multiplication rule (3.5) makes these classes commute, identifying the graded ring with a polynomial ring in $x$ and $ξ$ .

Order of a product in a domain

Since $gr,A_n$ is a polynomial ring, it is a domain, so symbols multiply without cancellation and $ord(PQ) = ord(P) + ord(Q)$ exactly, for all non-zero $P, Q ∈ A_n$ . This is the cleanest proof that P3 is a domain. For general $R$ the graded ring need not be a domain, and then only the inequality $ord(PQ) ≤ ord(P) + ord(Q)$ survives.

Key Equations

The filtration is exhaustive and starts at $R$ :

D^-1(R) = 0, D^0(R) = R, D^1(R) = R ⊕ Der_K(R), ⋃ _n D^n(R) = D(R).

(3.3)

The two rules that make it a ring filtration with commutative graded ring:

D^n,D^m ⊂ eq D^n+m, [D^n, D^m] ⊂ eq D^n+m-1.

(3.4)

Consequently the symbol is multiplicative and the commutator descends to a bracket of degree $-1$ :

σ_n+m(PQ) = σ_n(P),σ_m(Q), σ_n+m-1([P,Q]) = σ_n(P), σ_m(Q).

(3.5)

For the Weyl algebra the pieces and the graded ring are explicit. Writing $X = (x_1, … ,x_n)$ and $ξ_i = σ_1(∂_i)$ ,

D^m(K[X]) = ∑ _|β| ≤ m f_β, ∂^β ;:; f_β ∈ K[X] , gr,A_n ≅ K[x_1, … ,x_n,ξ_1, … ,ξ_n].

(3.6)

Each piece is a free $K[X]$ -module, of rank equal to the number of multi-indices of weight at most $m$ :

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

(3.7)

On $gr,A_n$ the bracket in (3.5) is the standard Poisson bracket of the cotangent space:

f, g = ∑ _i=1^n( (∂ f)/(∂ ξ_i)(∂ g)/(∂ x_i) - (∂ f)/(∂ x_i)(∂ g)/(∂ ξ_i) ).

(3.8)

Variable Definitions

$R$: a commutative $K$ -algebra, $K$ of characteristic zero
$D^n(R)$: operators of order at most $n$ ; an $R$ -bimodule
$ord(P)$: the order of $P$ , the least $n$ with $P ∈ D^n(R)$
$σ_n(P)$: the symbol of $P$ in degree $n$ , its class modulo $D^n-1(R)$
$gr,D(R)$: the associated graded ring $⨁ _n D^n/D^n-1$
$ξ_i$: the symbol $σ_1(∂_i)$ , a coordinate on the cotangent fibre
$β$: a multi-index in $N^n$ , with $|β| = β_1 + … + β_n$
$∂^β$: the composite $∂_1^β_1 ⋅ s∂_n^β_n$
${f, g}$: the Poisson bracket on $gr,A_n$

Properties and Behaviour

Basic properties of order

$ord(P + Q) ≤ (ord P, ord Q)$ , with equality unless the orders are equal and the principal symbols cancel.
$ord(aP) ≤ ord(P)$ for $a ∈ R$ , with equality when $R$ is a domain and $a ≠ 0$ .
$ord(PQ) ≤ ord(P) + ord(Q)$ , with equality whenever $gr,D(R)$ is a domain, in particular for $D(R) = A_n$ .
$ord([P,Q]) ≤ ord(P) + ord(Q) - 1$ ; this can be strict, for instance $[∂_1,∂_2] = 0$ .

Each piece is a finitely generated $R$ -module in good cases

If $R$ is a finitely generated regular $K$ -algebra in characteristic zero, then $D^m(R)$ is a finitely generated projective $R$ -module for every $m$ , and $gr,D(R) ≅ Sym_R(Der_K(R))$ , the symmetric algebra on the module of derivations. For $R = K[X]$ this is (3.6). On singular $R$ both statements can fail.

Order and the module structure

$D^m(R)$ is closed under multiplication by $R$ on both sides, and $D^m(R) ⋅ R ⊂ eq D^m(R)$ , so the filtration is by $R$ -bimodules. This is what makes $gr,D(R)$ an $R$ -algebra rather than merely a $K$ -algebra, and it is the reason symbols can be regarded as functions on a space fibred over $Spec R$ .

Good filtrations of modules

A filtration $Γ_m$ of a $D(R)$ -module $M$ is compatible with the order filtration if $D^n Γ_m ⊂ eq Γ_n+m$ , and good if $gr M$ is finitely generated over $gr,D(R)$ . Good filtrations relative to the order filtration are what define the characteristic variety; good filtrations relative to the Bernstein filtration are what define dimension by counting.

Examples and Special Cases

$A_1$ and the Euler operator

$θ = x∂$ has order $1$ with $σ_1(θ) = xξ$ . Then $θ^2 = x∂ x ∂ = x(x∂ + 1)∂ = x^2∂^2 + x∂$ has order $2$ and $σ_2(θ^2) = x^2ξ^2 = σ_1(θ)^2$ , as multiplicativity requires. The lower-order term $x∂$ is invisible to the symbol.

Cancellation of principal symbols

$P = ∂^2 + x$ and $Q = -∂^2 + ∂$ both have order $2$ , yet $P + Q = x + ∂$ has order $1$ : the degree- $2$ symbols $ξ^2$ and $-ξ^2$ cancel. This is why $ord$ is only subadditive under addition.

The Laplacian and its characteristic variety

In $A_n$ the operator $Δ = ∂_1^2 + … + ∂_n^2$ has order $2$ and symbol $ξ_1^2 + … + ξ_n^2$ . Over $ℂ$ the zero set of that symbol is the complex quadric cone, the characteristic variety of the module $A_n/A_nΔ$ intersected with each fibre. The classical distinction between elliptic, hyperbolic and parabolic operators is a statement about this symbol, not about the operator.

An operator whose order is not visible from a presentation

Written as $∂ x - x∂$ , an expression that looks like order $1$ minus order $1$ , the operator is in fact $1$ , of order $0$ . Orders can only be read off reliably from the canonical form with all $∂$ 's to the right; any other arrangement can hide cancellation.

Worked Example

Symbols, orders and a bracket in $A_2$

Step 1 - two operators and their orders
Work in $A_2 = A_2(K)$ with generators $x_1, x_2, ∂_1, ∂_2$ . Put

$P = x_1∂_1^2 + x_2∂_2, Q = ∂_1∂_2.$

Both are already written in canonical form with all $∂$ 's on the right, so orders can be read off: $ord(P) = 2$ from the term $x_1∂_1^2$ , and $ord(Q) = 2$ . Their symbols in degree $2$ are $σ_2(P) = x_1ξ_1^2$ (the term $x_2∂_2$ has order $1$ and contributes nothing in degree $2$ ) and $σ_2(Q) = ξ_1ξ_2$ .
Step 2 - the product and its symbol
By (3.4), $PQ ∈ D^4$ , and by (3.5) its symbol in degree $4$ is the product of the symbols:

$σ_4(PQ) = (x_1ξ_1^2)(ξ_1ξ_2) = x_1ξ_1^3ξ_2 ≠ 0,$

so $ord(PQ) = 4$ exactly. Directly: $PQ = x_1∂_1^3∂_2 + x_2∂_1∂_2^2$ , whose degree- $4$ part is $x_1∂_1^3∂_2$ , confirming the symbol.
Step 3 - the commutator, term by term
Use $∂_i x_i = x_i∂_i + 1$ and that all other pairs of generators commute. First,

$∂_1∂_2,(x_1∂_1^2) = ∂_2(∂_1 x_1)∂_1^2 = ∂_2(x_1∂_1 + 1)∂_1^2 = x_1∂_1^3∂_2 + ∂_1^2∂_2,$

so $[x_1∂_1^2, ∂_1∂_2] = x_1∂_1^3∂_2 - (x_1∂_1^3∂_2 + ∂_1^2∂_2) = -∂_1^2∂_2$ . Second,

$∂_1∂_2,(x_2∂_2) = ∂_1(∂_2 x_2)∂_2 = ∂_1(x_2∂_2 + 1)∂_2 = x_2∂_1∂_2^2 + ∂_1∂_2,$

so $[x_2∂_2, ∂_1∂_2] = -∂_1∂_2$ . Adding, $[P,Q] = -∂_1^2∂_2 - ∂_1∂_2$ .
Step 4 - check the order drop
$[P,Q]$ has order $3$ , not $4$ : the rule $[D^2,D^2] ⊂ eq D^3$ is attained here, so the bound in (3.4) is sharp. The symbol in degree $3$ is $σ_3([P,Q]) = -ξ_1^2ξ_2$ .
Step 5 - compare with the Poisson bracket
Compute $σ_2(P), σ_2(Q)$ from (3.8) with $f = x_1ξ_1^2$ and $g = ξ_1ξ_2$ . The partials are $∂ f/∂ ξ_1 = 2x_1ξ_1$ , $∂ f/∂ x_1 = ξ_1^2$ , $∂ g/∂ x_1 = ∂ g/∂ x_2 = 0$ , $∂ g/∂ ξ_1 = ξ_2$ , and $f$ does not involve $x_2$ or $ξ_2$ . Hence

$f, g = (2x_1ξ_1)(0) - (ξ_1^2)(ξ_2) + 0 - 0 = -ξ_1^2ξ_2,$

which agrees with $σ_3([P,Q])$ computed in Step 4, exactly as (3.5) requires.
Step 6 - ranks of the filtration pieces
For $n = 2$ , formula (3.7) gives $rank_K[x_1,x_2]D^m = m+22$ , that is $1, 3, 6, 10$ for $m = 0,1,2,3$ . Count directly: $m = 1$ needs $1, ∂_1, ∂_2$ ; $m = 2$ adds $∂_1^2, ∂_1∂_2, ∂_2^2$ , giving $6$ ; $m = 3$ adds the four cubics $∂_1^3, ∂_1^2∂_2, ∂_1∂_2^2, ∂_2^3$ , giving $10$ . The counts match.

Result

$ord(P) = ord(Q) = 2$ , $ord(PQ) = 4$ with symbol $x_1ξ_1^3ξ_2$ , and $[P,Q] = -∂_1^2∂_2 - ∂_1∂_2$ has order $3$ with symbol $-ξ_1^2ξ_2$ , which equals the Poisson bracket of the two principal symbols. The order drop predicted by $[D^n,D^m] ⊂ eq D^n+m-1$ occurs and is exactly one.

Applications and Industry Use

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

Characteristic varieties. The symbols of the operators annihilating a module cut out the characteristic variety in the cotangent space; its dimension is the dimension of the module, and its involutivity is Gabber's theorem.
Classification of PDE. Ellipticity, hyperbolicity and the propagation of singularities are all conditions on the principal symbol, and hence on the order filtration.
Microlocal analysis. Pseudodifferential and microdifferential operators are built by allowing symbols more general than polynomials while keeping the same filtration formalism; the Sato-Kashiwara-Kawai theory begins here.
Deformation quantisation. $D(R)$ with its order filtration is a quantisation of the Poisson algebra $gr,D(R) = O(T^*X)$ ; the bracket (3.8) is the classical limit of the commutator. See the Poisson bracket on the symbol algebra.
Algorithms. Groebner bases in the Weyl algebra are computed with respect to term orders refining the order filtration, so that leading terms are symbols and the commutative theory applies.

Design Considerations

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

Order filtration or Bernstein filtration

Choose the order filtration when the question is geometric: characteristic varieties, singular supports, restriction to subvarieties, anything that should transform correctly under change of coordinates. Choose the Bernstein filtration when the question is a counting question: Hilbert polynomials over $K$ , dimension bounds, multiplicity. The invariants $d(M)$ and $e(M)$ agree, but only after a theorem, and no intermediate step of a proof may silently swap one for the other.

Where to put the coefficients

Because $D^m$ is an $R$ -module and not a $K$ -space of finite dimension, Hilbert functions relative to the order filtration must be taken over $R$ , or over $K$ after a further grading of $R$ . Implementations that assume finite-dimensional filtration pieces will not work with the order filtration; this is a frequent source of bugs when porting Bernstein-filtration code.

Left or right normal form

Order is independent of how an operator is written, but reading it off requires a normal form. The convention throughout this collection is coefficients on the left, derivatives on the right: $∑ f_β ∂^β$ . The opposite convention gives the same order but different coefficients, related by the commutation relations, and mixing the two silently is a reliable way to produce wrong symbols.

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 upright roman for operator names such as $dim$ , $deg$ and $ord$ .
ISO/IEC 40314 (MathML 3.0) encodes the formulas on this page.
There is no standard for the symbol variable. $ξ_i$ is the analyst's choice and is used here; algebraists often write $∂_i$ , $σ(∂_i)$ or $y_i$ . Papers on Groebner bases in the Weyl algebra often write $D_i$ for $∂_i$ and reserve $ξ$ for the weight vector.
Software conventions for weight vectors differ in sign and in ordering of the variables; Macaulay2, Singular and SageMath each document their own, and they are not interchangeable without translation.

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.

Ground field. Characteristic zero is needed for $D(K[X]) = A_n$ and hence for (3.6). The definition of order itself is characteristic-free, but in characteristic $p$ the operator $∂^[p]$ has order $p$ while $∂^p = 0$ , so the graded ring is not $K[x,ξ]$ .
Coefficient ring. Working over $ℚ$ keeps symbols exact and is standard for computation; extending scalars to $ℂ$ changes neither orders nor symbols, since the filtration is defined over the prime field.
Representation. An element of $D^m(K[X])$ is stored as a sparse list of pairs $(β, f_β)$ ; the order is the maximum $|β|$ present. Symbols are stored as commutative polynomials in $x$ and $ξ$ , which is why the same data structure serves both the Weyl algebra and its graded ring.
Weight vectors. Software encodes a filtration as a weight vector on $(x_1, … ,x_n,∂_1, … ,∂_n)$ . Recording which weight vector was used alongside any computed dimension is essential, since the two standard choices give different graded objects even though they agree on $d(M)$ .

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 compute the order of an element of $A_n$ given as a word in the generators, rewrite it in canonical form by repeatedly applying $∂_i x_i = x_i ∂_i + 1$ ; the order is then the largest $|β|$ with $f_β ≠ 0$ .
To compute symbols of a family of operators, take the canonical forms and keep only the top-order terms, replacing $∂_i$ by the commuting variable $ξ_i$ .
To compute the associated graded module of a submodule $N ⊂ eq A_n^r$ , take a Groebner basis of $N$ with respect to a term order refining the order filtration; the leading terms then generate $gr N$ . This step is what makes the whole computation valid, and it is why an arbitrary generating set is not enough.
Read off the characteristic ideal as the ideal of leading symbols, and compute its dimension by a commutative Hilbert-series computation.

The rewriting in step 1 is cheap; the Groebner computation in step 3 dominates, with worst-case doubly exponential behaviour in $2n$ variables. Macaulay2's Dmodules, Singular's dmod.lib and dmodapp.lib, and SageMath's ore_algebra all implement Weyl-algebra Groebner bases with a choice of weight vector, where weight $(0, … ,0,1, … ,1)$ on $(x,∂)$ gives the order filtration and $(1, … ,1,1, … ,1)$ gives the Bernstein filtration.

A cheap correctness check: for any two operators, the computed order of $[P,Q]$ must be at most $ord P + ord Q - 1$ . A commutator that comes back at the full order $ord P + ord Q$ indicates a normal-form bug, almost always a missing application of $∂_i x_i = x_i∂_i + 1$ .

Limits of Validity

Filtration pieces are not finite dimensional. $D^0(K[X]) = K[X]$ already fails. Every counting argument that needs $_K$ must use a different filtration.
$gr,D(R)$ need not be a polynomial ring, or Noetherian, or a domain. These all hold for regular $R$ in characteristic zero. On singular $R$ the graded ring can be badly behaved, and then order is not additive on products.
The order filtration does not see the size of coefficients. $x^1000∂$ and $∂$ have the same order. Any argument that needs to control coefficient growth must add a second grading, which is what the Bernstein filtration supplies.
Symbols lose information. Two operators with the same principal symbol can generate very different modules; $∂$ and $∂ + 1$ have the same symbol but $A_1/A_1∂ ≅ K[x]$ while $A_1/A_1(∂+1) ≅ K[x]e^-x$ as $A_1$ -modules, which are non-isomorphic.

Failure Modes and Common Mistakes

Reading the order off a non-canonical expression

$∂ x$ looks like order $1$ and is: it equals $x∂ + 1$ . But $∂ x - x∂$ looks like order $1$ and is $1$ , of order $0$ . Always reduce to canonical form before reading orders or symbols, and be especially careful with differences of operators of equal order.

Writing $σ(P)$ without the degree

The symbol lives in a specific graded piece. If $P$ has order $1$ then $σ_2(P) = 0$ while $σ_1(P) ≠ 0$ . When adding operators of different orders, one must fix the degree first and accept that the lower-order summand contributes zero. Formulas such as $σ(P+Q) = σ(P) + σ(Q)$ are false without that convention.

Assuming the graded ring is a polynomial ring

$gr,A_n ≅ K[x,ξ]$ is a theorem about the polynomial ring in characteristic zero, resting on the canonical basis of $A_n$ . For a singular $R$ , $gr,D(R)$ can fail to be Noetherian, and $Sym_R(Der_K(R)) → gr,D(R)$ need be neither injective nor surjective.

Substituting the Bernstein filtration into an order-filtration argument

The two filtrations have the same associated graded ring up to isomorphism of abstract rings but not as graded rings: under the Bernstein filtration $x_i$ has degree $1$ , under the order filtration degree $0$ . Statements about the characteristic variety are false for the Bernstein-graded object, and Hilbert-function counts over $K$ are meaningless for the order-graded one.

Expecting the commutator to drop more than one order

The bound $ord([P,Q]) ≤ ord P + ord Q - 1$ is sharp, as the worked example shows. A further drop happens exactly when the Poisson bracket of the principal symbols vanishes, which is a strong condition and not the generic situation.

Historical Notes

Order and symbol are classical notions from the theory of partial differential equations, where the principal symbol of a linear operator is obtained by replacing $∂/∂ x_j$ with $iξ_j$ and keeping the top-degree terms. Ellipticity, the Cauchy-Kovalevskaya theorem and the classification of second-order equations are all statements about that symbol, and they predate the algebraic theory by a century.

The algebraic recasting - order as a filtration on a ring, symbol as a class in the associated graded ring - belongs to the 1960s, with Grothendieck's inductive definition on the algebraic side and the Sato school's microlocal analysis on the analytic side. The observation that $gr$ is commutative because commutators drop order is the bridge between the two, and it is what allows a theory of dimension for modules over a non-commutative ring.

The Poisson structure induced on $gr$ was the starting point for Gabber's 1981 theorem that characteristic varieties are involutive, and, in a different direction, for deformation quantisation. In both cases the content is that the first-order departure from commutativity, which the order filtration isolates, is itself a rich structure.

Comparison

The order filtration against the Bernstein filtration on $A_n$ .
	Order filtration $D^m$	Bernstein filtration $B_m$
Weight of $x_i$	$0$	$1$
Weight of $∂_i$	$1$	$1$
$m$ -th piece	free $K[X]$ -module of rank $m+nn$	$K$ -space of dimension $m+2n2n$
Finite dimensional over $K$	no	yes
Associated graded ring	$K[x_1, … ,x_n,ξ_1, … ,ξ_n]$	$K[x_1, … ,x_n,ξ_1, … ,ξ_n]$ with a different grading
Defined for general $R$	yes	no
Used for	characteristic variety, symbols, geometry	Hilbert polynomial, dimension, multiplicity

Key Takeaways

Key points

The order of $P$ is the least $n$ with $P ∈ D^n(R)$ , equivalently the number of commutators with elements of $R$ needed to reach zero.
$D^0 = R$ , $D^1 = R ⊕ Der_K(R)$ , and the pieces are $R$ -bimodules whose union is $D(R)$ .
Composition adds orders and commutators drop the total by one; both bounds are sharp.
Therefore $gr,D(R)$ is commutative and carries a Poisson bracket $σ(P),σ(Q) = σ([P,Q])$ of degree $-1$ .
For $K[x_1, … ,x_n]$ in characteristic zero, $D^m$ is free of rank $m+nn$ over $K[X]$ and $gr,A_n = K[x,ξ]$ , the functions on the cotangent space.
Order is read off only from a canonical form; symbols must always carry the degree they live in.

FAQs

Is the order of an operator the same as its degree?

No. In $A_n$ the Bernstein degree of $x^α∂^β$ is $|α| + |β|$ , while its order is $|β|$ . The operator $x^10∂$ has degree $11$ and order $1$ . Two different filtrations, two different numbers.

Why must commutators drop the order?

Because in $[PQ,a] = P[Q,a] + [P,a]Q$ each summand loses one order, and the Jacobi identity propagates this to $[[P,Q],a]$ . The formal proof is a double induction on $ord P + ord Q$ , given on the definition page.

Can two operators have the same symbol?

Yes, and this is the normal situation: $σ_n(P) = σ_n(Q)$ exactly when $P - Q$ has order less than $n$ . Symbols see the top order only, which is what makes them computable and also what makes them lossy.

Does the order filtration have finite-dimensional pieces?

No. Already $D^0(K[x]) = K[x]$ is infinite dimensional over $K$ . Each piece is finitely generated as an $R$ -module, which is the right finiteness statement here, and Hilbert functions must be taken relative to $R$ .

How does the order filtration give the characteristic variety?

Take a good filtration of a module $M$ compatible with $D^m$ , form $gr M$ over $gr,D = K[x,ξ]$ , and take the radical of its annihilator. That ideal cuts out a subvariety of the cotangent space $K^2n$ ; see the characteristic ideal and variety.

Why is

ord(PQ) = ord(P) + ord(Q)

for the Weyl algebra but not in general?

Because $gr,A_n$ is a polynomial ring and hence a domain, so the product of two non-zero symbols is non-zero. If $gr,D(R)$ has zero divisors, principal symbols can multiply to zero and the order of the product drops.

Does the definition of order depend on the ground field?

The definition does not, but the answer can. In characteristic $p$ the operator $x^m ↦ mp x^m-p$ on $K[x]$ has order $p$ and is not a polynomial in $∂$ , so the graded ring is larger than $K[x,ξ]$ .

Is there an analogue of the order filtration for modules?

Yes: a filtration $Γ_m$ of a module with $D^nΓ_m ⊂ eq Γ_n+m$ . It is good when $gr M$ is finitely generated over $gr 𝒟$ , and good filtrations always exist for finitely generated modules.

References

S. C. Coutinho, A Primer of Algebraic D-modules, London Mathematical Society Student Texts 33, Cambridge University Press, 1995 - Ch. 3 §1 for order, Ch. 3 §2 for the identification of the filtration pieces of $A_n$ , Ch. 7 for filtered rings.
J.-E. Björk, Rings of Differential Operators, North-Holland Mathematical Library 21, North-Holland, 1979 - Ch. 1 and Ch. 2, for filtrations, symbols and graded rings.
A. Grothendieck, Éléments de géométrie algébrique IV, Publications Mathématiques de l'IHÉS 32 (1967), §16.
O. Gabber, The integrability of the characteristic variety, American Journal of Mathematics 103 (1981), 445-468 - involutivity, a statement about the Poisson bracket on the graded ring.
R. Hotta, K. Takeuchi and T. Tanisaki, D-modules, Perverse Sheaves, and Representation Theory, Progress in Mathematics 236, Birkhäuser, 2008 - Ch. 2, for characteristic varieties.
M. Saito, B. Sturmfels and N. Takayama, Gröbner Deformations of Hypergeometric Differential Equations, Algorithms and Computation in Mathematics 6, Springer, 2000 - Ch. 1, for weight vectors and filtrations in computation.
ISO 80000-2:2019, Quantities and units - Part 2: Mathematics, International Organization for Standardization.
D. R. Grayson and M. E. Stillman, Macaulay2 - documentation for the Dmodules package and its weight-vector conventions.

AI Suggested Questions

Compute the order and principal symbol of $(x_1∂_1 + x_2∂_2)^3$ in $A_2$ after reducing to canonical form.
Verify $[D^1, D^1] ⊂ eq D^1$ directly and identify it as the statement that derivations form a Lie algebra.
Find two operators in $A_1$ of order $3$ whose commutator has order $3$ , and two whose commutator has order less than $5$ by more than one.
Show that $gr,A_n$ is a polynomial ring by using the canonical basis theorem, and identify where linear independence is needed.
Give the rank of $D^m(K[x_1,x_2,x_3])$ over $K[x_1,x_2,x_3]$ for $m = 0, … ,4$ and check the values by counting monomials.
Compute the Poisson bracket of $ξ_1^2 + ξ_2^2$ with $x_1ξ_2 - x_2ξ_1$ and interpret the answer as a rotational symmetry of the Laplacian.
Explain why $A_1/A_1∂$ and $A_1/A_1(∂ + 1)$ have the same characteristic variety but are non-isomorphic modules.