Counting singlets.
How a column of numbers from group theory led me to a Laguerre polynomial
In this exploration
A counting question
How many independent ways are there to combine quarks and antiquarks into something color-neutral? With the three colors of SU(3), three quarks combine in exactly one way, and four quarks with one antiquark in three ways. This exploration follows how a pattern in a table of numbers led me to a formula that answers the question for any number of colors, a formula I have not found in the literature.
Let me state the question in general. SU() is the symmetry group of colors.1 A quark here is something that can carry any one of the colors, and an antiquark is an object carrying an anticolor. Combine quarks with antiquarks. How many linearly independent color-neutral states, or singlets, can they form? I label the answer as . These words may not make sense to you now; they will be defined in the next section.
I needed these numbers for my research. In a qubit-regularized gauge theory, each link of the lattice holds a small, fixed number of states, and the gauge-invariant states at a site are built from singlets of the representations on the links that meet there. In the previous exploration, these were singlets of a few spin-½ flux ends in SU(2), and I could count them by hand. For SU(), a computer package such as GroupMath returns the numbers one at a time.2 For practical purposes, that is all I need. However, my curiosity demanded a neater formula for in terms of and .
The figure below lets you choose and and read off , together with a closed formula that I will arrive at by the end.
Singlets of n + m quarks and m antiquarks
k = n + m = 5 quarks and m = 2 antiquarks: the singlets of 3⊗5 ⊗ 3⊗2
Closed form: m! Lm(n)(1) = 2! L2(3)(1) = 11
Inside the theorem’s range (m ≤ n + 1): the count equals the closed form
Selection rule: with 2 antiquarks, every other k with m < k < n + m (here k = 3 and 4) gives 0 singlets.
SU(3), 5 quarks and 2 antiquarks: 11 independent singlets. The closed form gives 11 too; m ≤ n + 1 is inside the theorem’s range.
Each count is the number of independent singlets in the product, computed by singlets.py; the closed form is the alternating sum of the theorem. The page reads both from that script’s output and computes neither. The dots stand for the particles only; they do not show how the singlets are built.
What is being counted
Here is the one paragraph of linear algebra I need. The color state of one quark is a vector in , and each element of SU() acts on it as the matrix . A rule that turns group elements into matrices in this way is a representation; this one is the fundamental representation, written . An antiquark carries the dual representation , in which acts by its complex conjugate .3 The states of quarks and antiquarks form the tensor product , a space of dimension , on which acts as on every quark and as on every antiquark. A singlet is a vector that every in SU() leaves fixed. The singlets form a subspace, and its dimension is the number of independent color-neutral combinations.4
Not every number of quarks allows a singlet. The matrix , with , belongs to SU(), and it multiplies every vector of by the same number, .5 A singlet must be left fixed, so singlets exist only when is a multiple of . When quarks outnumber antiquarks by at least one and at most , that is, for , only survives. That is the case I count:
For SU(3), counts the singlets of three quarks, and those of four quarks and an antiquark.
The data
I first computed these counts with GroupMath. The table shows for small and .6 I have left out the cells with , where the pattern below breaks down.
| n \ m | 0 | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|---|
| 1 | 1 | 1 | 1 | |||
| 2 | 1 | 2 | 5 | 14 | ||
| 3 | 1 | 3 | 11 | 47 | 225 | |
| 4 | 1 | 4 | 19 | 104 | 641 | 4364 |
| 5 | 1 | 5 | 29 | 191 | 1405 | 11389 |
| 6 | 1 | 6 | 41 | 314 | 2661 | 24694 |
| 7 | 1 | 7 | 55 | 479 | 4577 | 47579 |
The row begins with three numbers I had already worked out for SU(3): one singlet for three quarks, three for four quarks and an antiquark, and eleven for five quarks and two antiquarks. Read along a row, and the numbers grow quickly with no rule in sight. Read down a column instead. Fix , let grow from , and the entries follow a polynomial in . The first six are
I fitted these polynomials from the counts up to . What single formula produces all of them?
Following the numbers
Since I am not Ramanujan and I haven’t memorized enough special functions, the pattern in those polynomials is not immediately obvious to me. What follows more or less follows my original steps of discovery that led me to a conjecture: take differences, notice a pattern, build a generating function.
Differences
Differences are the discrete version of derivatives. A polynomial of degree becomes a constant after differences, so repeated differences reveal the degree of a column without any guess at its formula. Take and . Subtract each entry from the next, do the same to the new row, and repeat. In the table, each entry is the number above and to its right minus the number directly above it; for instance, .
| Row | n = 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|
| 14 | 47 | 104 | 191 | 314 | 479 | 692 | |
| 33 | 57 | 87 | 123 | 165 | 213 | ||
| 24 | 30 | 36 | 42 | 48 | |||
| 6 | 6 | 6 | 6 |
The third differences are constant, and the constant is . So the column comes from a cubic, and the constant means its leading coefficient is 1.7 Newton’s forward-difference formula then reads off the first entry of each row,
The same happened in every column up to : the -th differences are the constant , so has degree and leading coefficient 1.
A pattern in the first differences
The first differences hold more than the degree. Look at them again: 33, 57, 87, 123, 165, 213. Every one is a multiple of 3, and dividing by 3 gives 11, 19, 29, 41, 55, 71. That is the column, starting at . In other words, . The same holds in every column, with the factor in place of 3. For instance, . As an identity between polynomials, for every ,
I will call this pattern D, for differences. Who would have thought one column of the table knows so much about its neighbor!
The figure below lets you repeat both steps. Choose and use the Take differences button until the rows collapse to the constant . Then turn on the Compare with m × previous column toggle to see pattern D.
Differences down a column
| n | 2 | 3 | 4 | 5 | 6 | 7 | 8 | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Δ⁰N(n, 3) | 14 | 47 | 104 | 191 | 314 | 479 | 692 | |||||||
| Δ¹ | 33 | 57 | 87 | 123 | 165 | 213 | ||||||||
| 3 × N(n + 1, 2)m × previous column | 333 × 11 | 573 × 19 | 873 × 29 | 1233 × 41 | 1653 × 55 | 2133 × 71 | ||||||||
| Δ² | 24 | 30 | 36 | 42 | 48 | |||||||||
| Δ³m! = 6 | 6 | 6 | 6 | 6 | ||||||||||
Δ³ subtracts neighboring entries of Δ²: 4 entries. Δ³ is constant: every entry is 6 = 3!. So on this window N(n, 3) is a polynomial in n of degree 3 with leading coefficient 1.
Beneath Δ¹: 3 × N(n + 1, 2), the m = 2 column taken one step along in n. All 6 entries agree with Δ¹: they match, as pattern D predicts.
Each row lists the differences of neighboring entries in the row above, so its entries sit between the two they come from. The first row is read from the output of singlets.py; the page computes only these subtractions and the products m × N in the comparison row. The window starts at n = m − 1 (n = 1 for m = 1), where the column follows its polynomial.
A generating function
A factor of that appears each time steps down by one is the signature of an exponential generating function, a power series whose -th term is divided by . So for each I packed the whole sequence into one function,
Multiply pattern D by and sum over . The factor cancels against the factorial, and the whole pattern becomes one line, . Hence
But what is ? Here comes the exciting part. If pattern D holds, every polynomial in the family is fixed by a single sequence of integers, the constant terms:8
A quick search on the On-Line Encyclopedia of Integer Sequences (OEIS) identified this sequence as A009940: , where is the -th Laguerre polynomial. Its exponential generating function is , and so
Laguerre
I recognized the shape of this function (or, rather, now it is easier to look through the list of well-known generating functions). The generalized Laguerre polynomials , a classical family of orthogonal polynomials, have the generating function9
At and this matches term by term, so . That was my conjecture:
That was the arduous but exciting short hike I took to reach the conjecture. Not to rob others of the same fun, but I thought I could contribute to the mathematical community by saving people some time: just as I first searched the coefficients on OEIS, I wished I didn’t have to go through A009940 to get to my generating function.
Thus, in March 2024, I submitted the coefficients of in my polynomials, , to the OEIS. The sequence is now A371114, times the coefficient of in . The formula listed there, in terms of A009940, is exactly what the generating function predicts.10
Still, the data stop at and small . Neither pattern D nor the conjecture was proven, and I could not see why the polynomials stop matching the counts once exceeds .
And then AI enters the picture.
From conjecture to a proposed theorem
I did not find the proof myself: it was written by Claude, an AI model, in October 2026. To be honest, I have not checked it step by step. In my foreseeable future, my interest in physics will occupy so much of my attention that I won’t have time to check it. Nevertheless, I publish it as a tentative proof because I believe putting the idea to the public is better than letting it stay on my computer. If the proof holds, it turns the conjecture into the following theorem.
Theorem (proof by Claude, not yet independently checked). For every and ,
and the other values of in the window give no singlets at all.
The script at the end of this article confirms the formula numerically for every and , independently of the proof; the proof, if it holds, extends it to all .
The alternating sum shows directly that is a polynomial in of degree , with integer coefficients and leading coefficient 1, as the differences suggested. The generating function follows from the theorem for every , and pattern D follows from . So if the proof holds, the pattern and the generating function are theorems too.
The range cannot be extended while fixing the rest of the theorem. At , the count exceeds the formula by exactly the Catalan number . For SU(2), for instance, . This is why the table stops at . The formula does not fail at random: just outside its range it misses by a well-known sequence, for to , and the proof below traces the miss to a single Young diagram. The first figure shows this excess when you set .
The proposed proof, in outline
Write and . A Young diagram is a partition of , drawn as left-justified rows of boxes, and is the number of standard Young tableaux of shape : fillings of the boxes with that increase along each row and down each column. The proof turns the count into a sum over such tableaux and then collapses the sum.
(i) Schur–Weyl duality. Singlets in are the SU()-maps from to . Since , the fully antisymmetric part of , is trivial for SU(), one may replace by , and the maps are then the same as U(n)-maps. Schur–Weyl duality splits into irreducible representations of U(n), each paired with an irreducible representation of the symmetric group and labeled by a Young diagram with at most rows. Schur’s lemma then reduces to a question about the symmetric group.
(ii) Kostka numbers. After transposing the diagrams, that question has the answer
where the Kostka number is the number of semistandard tableaux of shape and content . Two standard facts do the rest. For content , the ones must fill the start of row 1, so , the number of standard fillings of the skew shape left after removing the first boxes of row 1. And , the dimension of a permutation module of .
(iii) The sieve lemma. For with ,
where is 1 if holds and 0 otherwise. If , only the term survives. If with , then forces . This is the only place where is used. The skew shape then splits into a single row of boxes and the rest of the diagram, and the left side becomes a multiple of .
(iv) The alternating sum. Substituting the lemma into the Kostka sum, swapping the two sums and using both Kostka facts gives
(v) Laguerre. The explicit form of turns this sum into . Equivalently, the alternating sum has exactly the generating function found from the data.
At , the condition fails for exactly one diagram, , two rows of boxes. For it the left side of the lemma is , while the right side is 0, so the count picks up an extra .
There is also a heuristic reading of the theorem. The term, , counts an obvious spanning set of singlets: put one -tensor, the fully antisymmetric tensor with indices, on of the quarks, and pair each of the remaining quarks with an antiquark by a Kronecker delta. These singlets are linearly dependent, and the alternating terms remove the overcounting.
New, as far as I can tell
The ingredients are all on record: the Laguerre polynomials and their generating function are classical, and the constant terms were already in the OEIS. What I have not found in the literature is the statement that connects them to singlet counts, for . The formula is new to me, and I would be glad to learn of an earlier source.
The same singlet counts can also be written as integrals over the group,11
where is the Haar measure, the uniform measure on the group, normalized so that the group has total volume 1. The Weyl integration formula turns this into an integral over the eigenvalues of . The same counts give the number of independent tensor structures in the group integrals of the kind Zuber revisited for SU(). Such integrals appear in strong-coupling expansions of baryonic Wilson loops.
Now I will say something stupid that may frustrate a mathematician: the proof indirectly proves/derives pattern D, but does not explain why. Is there a bijection, a matching of the singlets or of the tableaux that count them, that shows directly why equals times ? Such a bijection would explain the pattern that started this search in terms of the singlets themselves. How can I “see” this fact? Perhaps the tableau sum is where I would start.12
Reproduce it
The script I used is singlets.py. It needs only Python 3, with no extra packages, and python3 singlets.py finishes in under a second. It checks eight things:
- a fast formula for as a sum over Young tableaux, against a direct evaluation of the group integral, together with the selection rule;
- the data table;
- the difference table for ;
- the polynomials to , fitted from the counts and checked against their expected coefficients;
- pattern D;
- the constant terms against A009940, and the generating function;
- the theorem for every and ;
- the Catalan excess at .
Further reading
- J.-B. Zuber, “Revisiting SU(N) integrals,” Journal of Physics A: Mathematical and Theoretical 50, 015203 (2017), arXiv:1611.00236. It studies SU() group integrals whose independent tensor structures these counts enumerate.
- R. M. Fonseca, “GroupMath: A Mathematica package for group theory calculations”, Computer Physics Communications, 108085 (2021), arXiv:2011.01764. The package I used for the original counts.
- W. Fulton and J. Harris, Representation Theory: A First Course (Springer, 1991), Lectures 4 and 6. A standard reference for Schur–Weyl duality, Young tableaux and Kostka numbers.
- D. Bump, Lie Groups, 2nd ed., Graduate Texts in Mathematics 225 (Springer, 2013), Chapters 2 and 17. A standard reference for Schur orthogonality of characters, which writes the counts as group integrals, and for the Weyl integration formula.
- DLMF, Eq. 18.12.13. The generating function of the Laguerre polynomials.
- OEIS, A009940: , the constant terms of the polynomials.
- R. X. Siew, OEIS, A371114: times the coefficient of in , the linear coefficients of the polynomials.
Footnotes
-
SU() is the group of unitary matrices with determinant 1. The quarks and antiquarks are names for its two simplest representations, defined in the next section. Only for are they the quarks of particle physics. ↩
-
GroupMath, by R. M. Fonseca, is a Mathematica package for group theory calculations; see Further reading. ↩
-
For a unitary matrix, , which is how acts on linear functions of a vector. That is why the antiquark representation is called the dual. Writing , the antiquarks live in the dual space . ↩
-
Equivalently, the singlets in are the linear maps from to that commute with every . In index language, counts the linearly independent invariant tensors with fundamental and antifundamental indices. ↩
-
These matrices form the center of SU(), the elements that commute with every other element. Each quark contributes a factor and each antiquark a factor . ↩
-
The values shown are computed by the script linked at the end of this article. It does not use GroupMath, and its SU(3) counts agree with the original GroupMath values. ↩
-
For a polynomial of degree with leading coefficient , the -th difference is the constant . Newton’s formula rebuilds the polynomial from the first entry of each row. ↩
-
Here does not correspond to a special unitary group; is simply the polynomial evaluated at zero. ↩
-
Explicitly, . The generating function is given in the NIST Digital Library of Mathematical Functions (DLMF), Eq. 18.12.13, and is valid for . The ordinary Laguerre polynomials are the case . ↩
-
The formula is , the coefficient form of . ↩
-
The integrand is the trace of the matrix by which acts on . Averaging that matrix over the group projects every vector onto its singlet part, and the trace of this projection is the number of singlets. This is Schur orthogonality for the characters of a compact group; see Bump, Lie Groups, Chapter 2. The eigenvalue form is the Weyl integration formula, Chapter 17 of the same book. ↩
-
A natural place to look is the tableau sum , valid for every , where is the number of rows of and adds a full column of height to . The question is then whether some bijection between pairs of standard Young tableaux accounts for pattern D. ↩