A symmetry may change coordinates while leaving some polynomial quantities unchanged. Invariant theory asks which quantities survive, how to generate all of them, and which relations prevent those generators from varying independently.
The subject joins representation theory with polynomial algebra. Reynolds averaging extracts invariant polynomials from ordinary ones. Molien’s formula counts independent homogeneous invariants degree by degree. A presentation by generators and relations then describes the entire invariant ring. These are connected tasks, but they are not interchangeable: counting three invariants does not identify them, and listing three generators does not prove they have no relations. [1–3]
This guide works with a finite group G acting linearly on a finite-dimensional complex vector space V. The field, finiteness and linearity assumptions remain in force unless explicitly changed. The examples are complete small calculations, not claims that every invariant ring has the same simple form. For the representation operations behind polynomial spaces, use the tensor products, duality and symmetric powers guide.
The polynomial action · Reynolds averaging · A complete swap example · Molien’s formula · A cyclic example with a relation · Practice and solutions · BTT Mathematics Hub
The question is what remains unchanged
Swap the coordinates x and y. The individual coordinate x changes to y, but x+y and xy remain the same. Negate x while keeping y unchanged. Now x+y is generally changed, while x² and y remain fixed. Negate both x and y. Then xy becomes invariant again.
The answer therefore depends on the actual action, not merely the name or order of the group. Three actions involving a group of order two can produce different sets of invariant polynomials.
We also need to distinguish an invariant polynomial from a fixed vector in V. The action (x,y)↦(−x,−y) has no nonzero fixed vector, but it has many nonconstant invariant polynomials, including x², xy and y². The polynomials live in a different representation space.
The action on polynomials is a dual action
Write ρ(g) for the matrix acting on V. A polynomial f is a function on V. The induced left action is
(g·f)(v)=f(ρ(g−1)v).
The inverse ensures that g·(h·f)=(gh)·f. Degree-one coordinate functions form V*, the dual space, and the full polynomial ring is C[V]=Sym(V*). The action preserves polynomial degree.
A polynomial is invariant when g·f=f for every g. Equivalently, f(ρ(g)v)=f(v) for every g and v, because inversion permutes the group. The invariant polynomials form the subring C[V]G.
The ring property follows directly: sums and products of unchanged functions remain unchanged. Constants are always invariant. The homogeneous degree-d invariants form a finite-dimensional vector space, even though the full invariant ring usually has infinitely many linearly independent polynomials.
Generators are different from a vector-space basis
For the sign action x↦−x on one coordinate, every invariant polynomial contains only even powers. The invariant ring is C[x²]. One algebra generator, x², produces all its elements by addition and multiplication.
As a vector space, however, the ring has the infinite basis 1,x²,x⁴,x⁶,… . Saying that the ring has one generator does not mean that it is one-dimensional.
A homogeneous generating set also does not automatically provide independent coordinates. Distinct polynomial expressions in the generators may represent the same invariant. Such identities are the relations that a complete presentation must retain.
Reynolds averaging projects onto the invariant part
For a finite group over C, define the Reynolds operator by
R(f)=(1/|G|)Σg∈Gg·f.
Applying h to this sum replaces the summation index g by hg, so h·R(f)=R(f). Thus the output is invariant. If f was invariant already, all |G| summands equal f, so R(f)=f. Consequently R²=R: it is a projection. Reynolds operators and their role in constructing invariants are standard tools of finite-group invariant theory. [1–3]
The operator is linear and degree-preserving. Therefore averaging all degree-d monomials spans the space of homogeneous degree-d invariants. Some averages vanish, and different monomials can have the same average, so this spanning collection need not be a basis.
A zero average does not mean the action has no invariants. It means that this particular input has no component in the invariant subspace of that degree.
Averaging is not multiplication
For x↦−x, R(x)=0 but R(x²)=x². Therefore R(x·x) differs from R(x)R(x). Reynolds averaging is not generally a ring homomorphism.
It does satisfy a more limited multiplication rule: if a is invariant, then R(af)=aR(f). Pulling a out of each group-transformed summand is legitimate because g·a=a.
This distinction is essential in computations. Averaging the factors of a product separately can destroy an invariant that appears only when they are multiplied. The projection acts on the full polynomial representation, not independently on each occurrence of a variable.
A complete example: exchanging two coordinates
Let G={e,s}, where s(x,y)=(y,x). The invariants are the symmetric polynomials in two variables. Set e1=x+y and e2=xy.
Every symmetric polynomial is a linear combination of diagonal monomials xaya and paired monomials xayb+xbya with a>b. The first type equals e2a. The second equals e2b(xa−b+ya−b).
Define pk=xk+yk. We have p0=2, p1=e1, and
pk=e1pk−1−e2pk−2.
Expanding the right side cancels the two mixed terms and leaves xk+yk. Induction therefore expresses every pk, and hence every symmetric polynomial, in e1 and e2.
Compute with the generators
The recurrence gives p2=e1²−2e2 and p3=e1³−3e1e2. One more step gives
x⁴+y⁴=e1⁴−4e1²e2+2e2².
Similarly x³y+xy³=xy(x²+y²)=e2(e1²−2e2). These are exact identities, so substituting any pair of complex numbers provides a local arithmetic check.
Reynolds averaging illustrates the same structure from another direction. R(x²)=(x²+y²)/2=(e1²−2e2)/2, while R(xy)=xy=e2. Two different degree-two invariant directions have appeared.
Prove that there are no hidden relations in the swap example
Consider the products e1ae2b. Under lexicographic order with x before y, their leading monomials are xa+byb. Different pairs (a,b) have different leading monomials.
No nonzero finite linear combination of these products can therefore vanish. This proves that e1 and e2 are algebraically independent. Combined with the generation argument, we obtain C[x,y]S2=C[e1,e2], a polynomial ring with generator degrees 1 and 2.
Both halves of the conclusion matter. Generation says that nothing is missing. Algebraic independence says that no relation has been concealed. Checking only one of them would not establish the complete description.
Count by degree with a Hilbert series
Let hd be the dimension of the homogeneous degree-d invariant space. Its Hilbert series is H(t)=Σd≥0hdtd. Here t is a formal bookkeeping variable: the coefficient of td records a dimension.
For the swap example, a basis in degree d consists of e1ae2b with a+2b=d. There are floor(d/2)+1 choices. Summing over a and b gives
H(t)=1/[(1−t)(1−t²)].
In degree 4, the three basis elements are e1⁴, e1²e2 and e2². They explain h4=3. The coefficient does not mean that the full ring has only three invariants.
Molien’s formula counts invariants without first finding generators
Let σ(g) be the action on V*, the degree-one polynomial space. For a finite group over C, Molien’s formula states
H(t)=(1/|G|)Σg∈G1/det(I−tσ(g)).
Since σ(g) is the inverse transpose of ρ(g), and inversion permutes the finite group, the averaged formula can equivalently be written with ρ(g) in the determinant. The dual-action convention must still be remembered when interpreting individual summands. Molien’s theorem and its representation-theoretic derivation are presented in the references. [1–3]
The formula gives an exact generating series. It does not by itself produce an invariant polynomial, prove that a proposed list generates the ring, or identify every relation among that list.
Why determinants appear in a polynomial counting problem
Diagonalise the finite-order action on V*, with eigenvalues λ1,…,λn. A monomial with exponents a1,…,an is an eigenvector in the polynomial representation with eigenvalue λ1a1···λnan.
Summing these eigenvalues by total degree gives the product of geometric series ∏j(1−λjt)−1. That is exactly 1/det(I−tσ(g)). Its coefficient of td is the trace of g on degree-d polynomials.
Average these traces over the group. The averaged operator is the Reynolds projection on that homogeneous space. In characteristic zero an idempotent projection has trace equal to the dimension of its image. The coefficient therefore counts the degree-d invariants. This combines symmetric-power eigenvalues with the projection argument, explaining the architecture of Molien’s formula.
Check Molien against the swap calculation
The identity on C² has two eigenvalues 1. The coordinate swap has eigenvalues 1 and −1. Therefore
H(t)=(1/2)[1/(1−t)²+1/(1−t²)]=1/[(1−t)(1−t²)].
The rational expression agrees with the series obtained from the two explicit algebraically independent generators. We now have two independently organised derivations: a polynomial-generation proof and a representation-theoretic count.
A single reflection (x,y)↦(−x,y) produces the same Hilbert series. In that coordinate system the invariant ring is C[x²,y]. The equality of series is unsurprising because the reflection and swap actions are equivalent after a linear change of coordinates. The generating expressions differ because the coordinates differ.
A group of the same order with a different invariant ring
Now let the nonidentity element act by (x,y)↦(−x,−y). A monomial xayb is invariant exactly when a+b is even. Set A=x², B=xy and C=y².
If a and b are both even, the monomial is a product of A and C. If both are odd, it is B times such a product. These are exactly the possibilities when a+b is even. Thus A,B,C generate the invariant ring.
They are not algebraically independent: AC=B². This is an explicit relation of degree 4 in the original variables, with each generator assigned degree 2.
Prove that this one relation is enough
Using B²=AC, reduce any expression in A,B,C to F(A,C)+B·K(A,C). After substitution, the first part consists of monomials with both x and y exponents even. The second consists of monomials with both exponents odd.
Those two sets cannot cancel one another. Within each set, distinct powers of A and C also give distinct monomials. The reduced expression is therefore unique.
We have proved the presentation C[x,y]{±I}≅C[A,B,C]/(AC−B²). The relation is not merely one identity we happened to notice; it generates all the algebraic relations in this example.
Its Hilbert series records the relation
The unique normal form gives H(t)=(1+t²)/(1−t²)². Equivalently, H(t)=(1−t⁴)/(1−t²)³. The latter expression resembles three freely generated degree-two variables with one degree-four relation removed, while the former comes directly from the two parity classes in the normal form.
Molien confirms it: H(t)=(1/2)[1/(1−t)²+1/(1+t)²]. In degree 2k, all 2k+1 monomials xay2k−a are invariant; in odd degree none are. Hence the series begins 1+3t²+5t⁴+7t⁶+···.
There are three displayed generators but only two independent parameters in the resulting algebraic description. The equation AC=B² ties the coordinates together. Generator count, vector-space dimension by degree and algebraic dimension answer different questions.
A cyclic action of order three
Let ω be a primitive cube root of unity and let g act on points by g(x,y)=(ωx,ω−1y). Then g³ is the identity. A monomial xayb is invariant exactly when a−b is divisible by 3.
The dual action changes the exponent sign in its eigenvalue, but the condition for eigenvalue 1 is the same. Define u=x³, v=y³ and w=xy. These are invariant, with degrees 3,3 and 2.
To prove generation, remove min(a,b) copies of xy from an invariant monomial. The remaining power is entirely in x or entirely in y, and its exponent is divisible by 3. It is therefore a power of u or v. Every invariant monomial, and hence every invariant polynomial, is generated by u,v,w.
One relation and three normal-form layers
The generators satisfy uv=w³. Reduce every power w³ using uv. Every expression becomes F0(u,v)+wF1(u,v)+w²F2(u,v).
After substituting x and y, the three layers have both exponents congruent to 0,1 or 2 modulo 3 respectively. They cannot cancel across layers. Within one layer, distinct monomials uavb remain distinct. This proves uniqueness.
C[x,y]C3≅C[u,v,w]/(uv−w³).
The relation has degree 6 under the weights deg u=deg v=3 and deg w=2. Treating u,v,w as if they all had degree 1 would give the wrong Hilbert series for the original polynomial grading.
Compute the cyclic Molien series two ways
The normal form gives H(t)=(1+t²+t⁴)/(1−t³)². Equivalently, it is (1−t⁶)/[(1−t³)²(1−t²)].
For either nonidentity group element, the eigenvalues are ω and ω². Their denominator is (1−ωt)(1−ω²t)=1+t+t². Therefore Molien gives
H(t)=(1/3)[1/(1−t)²+2/(1+t+t²)].
The two rational functions simplify to the same expression. The coefficients from degree 0 through degree 8 are 1,0,1,2,1,2,3,2,3. Each can also be counted directly by listing a+b=d with a−b divisible by 3.
At degree 6, the invariant monomials are x⁶,x³y³,y⁶. Their generator expressions include u²,uv,v², while w³ is the same invariant as uv, not a fourth independent one. This is exactly where the relation first affects the degree count.
A rational denominator does not automatically identify generators
The same Hilbert series can be written in several rational forms. For example, multiplying numerator and denominator by 1+t changes the displayed factors without changing the series. A denominator containing three factors does not, by itself, prove that three algebra generators of the corresponding degrees have been constructed.
Likewise, subtracting one relation from a polynomial ring has a simple Hilbert-series effect in our one-relation examples because we have proved exact normal forms. Arbitrary lists of relations may themselves satisfy dependencies. One cannot always subtract their degrees independently and obtain the correct count.
A complete presentation needs a surjective map from a polynomial algebra in proposed generators to the invariant ring, together with a proof of its kernel. A Hilbert-series comparison can help certify that proof, but it must be applied to the actual graded algebra being presented.
Finite generation gives a stopping principle
For a finite linear group over C, the invariant ring is finitely generated. Noether’s degree bound says that homogeneous invariants of degree at most |G| suffice to generate it. The bound and its role in algorithmic construction are described in Sturmfels’s invariant-theory lectures. [3]
This does not say that every minimal generating set contains |G| elements, or that all generators have degree exactly |G|. It supplies an upper bound on degrees, not a count of generators and not a claim of computational cheapness.
Reynolds averages of all monomials up to that bound span the required invariant pieces. In principle they therefore generate the ring. In practice this can be a very redundant collection, so linear dependence and algebraic dependence tests remain important.
What a completeness check must establish
Suppose candidate invariants f1,…,fr generate a subalgebra B of C[V]G. Compute the homogeneous dimensions of B and compare them with Molien’s dimensions. A mismatch at any degree proves that the proposed description is incomplete or wrongly calculated.
Agreement through a few small degrees is not a proof of equality in all degrees. A generator might first appear later. A proof requires an applicable degree bound, an exact Hilbert-series identity for the candidate algebra, a normal-form argument or another valid global criterion.
Our swap, inversion and cyclic examples used explicit generation and normal forms. Molien then provided an independent count. That sequence is stronger than presenting a plausible rational function and reading imagined generators from its denominator.
Invariants describe orbits, but one invariant is rarely enough
An orbit consists of all points ρ(g)v obtained from v. Every invariant polynomial is constant on each orbit. For a finite group over C, the full collection of invariant polynomials separates different orbits.
A direct argument explains the last statement. Two different finite orbits are disjoint finite sets. Polynomial interpolation on their union produces a polynomial equal to zero on the first and one on the second. Averaging it leaves those values unchanged and makes it invariant. The resulting invariant distinguishes the orbits.
For simultaneous inversion, (1,0) and (0,1) have the same value of x²+y², but they are not in the same orbit. The tuple (x²,xy,y²) distinguishes them: its values are (1,0,0) and (0,0,1). Preserving one useful quantity is weaker than completely identifying an orbit.
Quotient coordinates must retain their relations
For the inversion action, map (x,y) to (A,B,C)=(x²,xy,y²). The image satisfies AC=B². Recording three output coordinates without their relation introduces combinations that did not arise independently.
The quotient description intentionally forgets which point in an orbit was used. It does not forget every relation among the surviving coordinates. That is the distinction between controlled identification and uncontrolled information loss.
This connects invariant theory to the rest of the series. Fourier analysis reorganises an entire function into symmetry components. Invariant theory retains the trivial component in each polynomial degree, then studies how those components multiply.
Invariant and equivariant are different requirements
An invariant scalar function satisfies f(gv)=f(v). An equivariant map between two representation spaces satisfies F(gv)=gF(v), using the appropriate action on its output. The first remains unchanged; the second changes in a prescribed compatible way.
The identity map V→V is equivariant, but its coordinate functions are not generally invariant. A scalar polynomial such as x² under x↦−x is invariant, but it does not by itself give an equivariant vector output with the same sign action.
Choosing the correct requirement depends on the task. A label intended to identify an orbit should be invariant. A transformed vector intended to preserve directional information may need equivariance instead. These design choices should be made before discarding representation components.
The characteristic-zero boundary is essential
If the field characteristic divides |G|, the scalar 1/|G| does not exist. The Reynolds formula above is unavailable in that form. The ordinary trace argument and the displayed complex Molien calculation cannot simply be transferred unchanged.
This does not mean that invariant polynomials disappear. It means that the averaging mechanism and the associated proof have lost a required hypothesis. The modular representation route explains the analogous failure of averaging in complete reducibility.
The finite-group hypothesis also matters. Infinite groups may require different averaging, algebraic assumptions or analytic machinery. This article does not use a finite sum as a substitute for those additional theories.
A calculation workflow that preserves the mathematical question
First write the action matrices and verify their group relations. Specify whether coordinates transform as vectors or polynomial functions, and record the polynomial grading. Then identify a few invariant candidates by direct substitution or Reynolds averaging.
Compute Molien’s series to know the expected dimensions. In each degree, compare averaged polynomials by their coefficient vectors rather than counting how many expressions were printed. Build candidate algebra generators and test whether new invariant directions are products of earlier ones.
Finally prove generation and identify relations. Use exact arithmetic for roots of unity and rational coefficients when available. A decimal residual can help detect a numerical issue, but it is not a proof that a symbolic relation holds identically.
Practice with worked explanations
1. Average a monomial under a swap
Find R(x³y) when x and y are exchanged. Solution: R(x³y)=(x³y+xy³)/2. In elementary generators it is e2(e1²−2e2)/2. The output is homogeneous of degree 4, just like the input.
2. A failed multiplicativity claim
For x↦−x, compare R(x²) and R(x)². Solution: R(x²)=x² while R(x)²=0. Reynolds averaging is linear and idempotent, not a general ring homomorphism.
3. Count swap invariants in degree seven
Find the dimension in degree 7 for C[x,y]S2. Solution: solve a+2b=7 in nonnegative integers. The possibilities have b=0,1,2,3, giving dimension 4. A basis is e1⁷,e1⁵e2,e1³e2²,e1e2³.
4. Reflection parity
Under (x,y)↦(−x,y), is x³y² invariant? Solution: no. It changes sign because the exponent of x is odd. Its Reynolds average is zero. The polynomial x⁴y² is invariant.
5. Simultaneous inversion parity
Under (x,y)↦(−x,−y), is x³y invariant? Solution: yes, because the total degree is 4. In the generators A=x²,B=xy,C=y² it equals AB. The condition differs from a reflection that negates only x.
6. Reduce a relation
Reduce B⁵ using AC=B². Solution: B⁵=B(AC)²=A²BC². This has at most one factor B and is in the normal form F(A,C)+BK(A,C).
7. Cyclic invariant test
For g(x,y)=(ωx,ω−1y), decide whether x⁵y² is invariant. Solution: 5−2=3 is divisible by 3, so it is invariant. It equals uw², where u=x³ and w=xy. Its degree is 3+2·2=7.
8. Count before and after a relation
Why are u²,uv,v²,w³ not four independent degree-six invariants in the C3 example? Solution: uv=w³. After identifying that equality, the three independent monomials are x⁶,x³y³,y⁶. The Hilbert coefficient in degree 6 is therefore 3.
9. Check the constant coefficient
What must the constant term of the Molien series be? Solution: 1. At t=0 every determinant denominator is 1, and averaging |G| copies of 1 gives 1. This agrees with the one-dimensional space of constant polynomials.
10. A fixed-vector misconception
If V has no nonzero G-fixed vector, must C[V]G contain only constants? Solution: no. Simultaneous inversion on C² has fixed-vector space zero but invariant polynomials x²,xy,y². Polynomial invariants live in higher symmetric powers of the dual space.
11. An incomplete orbit label
Does x²+y² distinguish every orbit of the simultaneous-inversion action? Solution: no. Points (1,0) and (0,1) have equal value 1 but are not related by multiplication by −1. The full generator tuple separates these two orbits.
12. What does a degree bound actually bound?
For a finite group of order 8 over C, does Noether’s bound require exactly eight generators? Solution: no. It says that invariants of degree at most 8 can generate the ring. The number of necessary generators and their relations depend on the action.
A learning sequence from unchanged quantities to complete algebra
Start with the three order-two actions and require the learner to test the same monomial under each one. This makes dependence on the actual representation visible. Next ask for a Reynolds average and an example showing why averaging cannot be distributed over an arbitrary product.
Develop the swap ring through the power-sum recurrence, then compare it with simultaneous inversion. Both have small generating sets, but only one of the displayed presentations is freely generated. Ask the learner to explain that difference using an explicit relation and a unique normal form.
Introduce Molien only after the degree-counting question is clear. Then use the C3 example to compare three routes: congruences on monomials, normal forms in generators, and a determinant average. Agreement among these routes tests whether the learner is preserving the action, the grading and the relations together.
Mathematical references
[1] Richard P. Stanley, Invariants of Finite Groups and Their Applications to Combinatorics, for the invariant-ring viewpoint, Hilbert series and Molien’s theorem. [2] Dmitri Panyushev, Lectures on Representations of Finite Groups and Invariant Theory, for the representation-theoretic derivation of invariant dimensions and Molien’s formula. [3] Bernd Sturmfels, Algorithmic Invariant Theory, for Reynolds construction, the characteristic-zero degree bound and computational completeness questions. The examples above include explicit generation and relation proofs rather than relying only on series matching.
Continue Representation Mathematics — Batch 03
For complete symmetry-adapted function decompositions, read Finite-Group Fourier Analysis. For irreducible permutation actions and polynomial tensor connections, use Symmetric Group Representations and Young Tableaux. For vector spaces connected by maps and the distinction between simple and indecomposable objects, continue to Quiver Representations and Gabriel’s Theorem. Return to the BTT Mathematics Learning Hub.