Polynomial Spaces

Abhibhav Garg

April 16, 2019

Contents

1 Introduction 1

2 Polynomial Spaces 2
2.1 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.3 Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.4 Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Johnson Schemes 6

4 More Schemes 7
4.1 Hamming Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.2 Unit Sphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

5 Bibliography 8

A Existence of designs 9

1 Introduction
Over the course, we saw a number of applications of linear algebra to combinatorics. The most
notable application was the Frankl-Wilson theorem, which informally states the following: If we
have a set family such that the intersections modulo p are constrained to take atmost s values,
then the family can have atmost ns elements. This was a natural extension of the odd-town

problem. In both cases, the argument proceeded by constructing a linearly independent set of
vectors, in a space of low dimension, which naturally bounds the size of the family.
In this report, we sketch the ideas of Godsil (1988), who generalised these ideas greatly, by
introducing the framework of Polynomial Spaces. This allows one to derive many known results,
such as the Frankl Wilson theorem, in a uniform way. While the original paper first introduces
the framework completely before instantiating it, here we try and do the two simultaneously, in
the hope that this provides more motivation for some of the definitions. The instantiation we
pick as the motivating one will be that which allows us to recover the Frankl Wilson theorem.

1
Notation
Unless stated otherwise, expectations will always be taken with respect to the uniform distribu-
tion.

2 Polynomial Spaces
2.1 Basic Definitions
A polynomial space is a set Ω and a real valued function ρ : Ω×Ω → R that satisfies the following
three conditions:
• ρ(x, y) = ρ(y, x) for all x, y ∈ Ω.

• ρ(x, x) = ρ(y, y) > 0 for all x, y ∈ Ω.

• ρ(x, x) > ρ(x, y) for all x, y ∈ Ω.

We require (Ω, ρ) to satisfy some additional axioms, which will be stated later. These axioms
will always hold when Ω is finite, which is the focus of this report. The function ρ intuitively
captures the similarity between two elements of Ω. In many of the example spaces, the space
will naturally admit a metric, and ρ will just be the negative of this metric, appropriately
translated to satisfy the positivity condition. For the rest of this report, (Ω, ρ) will always
denote a polynomial space.
Our main example is the Johnson Scheme, J(n, k): Here, Ω is the set of all k sized subsets
of a set with n elements, and ρ(x, y) = |x ∩ y|.
Another example will be the Hamming Scheme: Here, Ω is the set of all n length words of
a finite alphabet Σ, and ρ(x, y) = n − |x∆y|, where |x∆y| is the usual Hamming distance.
A third example is the unit sphere. Here, Ω is the set of all unit vectors in Rn , and ρ is the
usual inner product function. This space is different from the previous two in that it is infinite,
but it will satisfy some finiteness conditions that we will see later.

2.2 Codes
We now define the notion of a code. Let Φ be a subset of Ω. Define the distance set, D(Φ) to
as D(Φ) = {ρ(x, y) | x, y ∈ Φ, x 6= y}. Let d(Φ) = |D(φ)|. This cardinality will also be called
the degree of Φ. Finally, if D(Φ) ⊆ A for some A ⊆ R, then Φ will be called an A-code. Many
problems will be reduced to finding large A-codes. In particular, consider the Johnson scheme.
Any Φ defines a set family, and the distance set is exactly the set of all possible intersections
of members of the family. If we fix an A, we can then bound the size of A-codes, which are
equivalent to bounds on set families with restricted intersections. For example, setting A to be
all positive natural numbers will recover the hypothesis of EKR, and setting A to be all natural
numbers equivalent to some λi modulo p will recover the hypothesis of Frankl Wilson.

2.3 Polynomials
We now introduce a class of functions on Ω. Let g be any univariate polynomial from the ring
of polynomials over R. For any a ∈ Ω, define the zonal polynomial ζa (g)(x) := g(ρ(a, x)). The

2
set of functions span ({ζa (g) | a ∈ Ω, deg(g) ≤ r}) is a real vector space, which we denote Z(r).
We can further also consider the ring generated by the zonal polynomials, namely all functions
generated by sums and products of zonal polynomials. We will call this the set of Polynomials,
and denote it by P . This is also a vector space. Note that Z and P will denote the vector spaces
corresponding to Ω. Vector spaces of functions over subsets Φ will be denoted ZΦ , PΦ .
We can decompose P into vector spaces P (r) for non-negative integers r, defined recursively
as

• P (0) = Z(0).

• P (1) = Z(1).

• P (k + 1) = span (gh | g ∈ P (1), h ∈ P (k)).

For a function f , the smallest k such that f ∈ P ol(k) will be called the degree of f . We also
assume the following finiteness condition:
• dim P (1) < ∞.
We now prove a couple of simple finiteness lemmas.
Lemma 2.1. If Ω is finite, then there is an integer d ≤ |Ω| such that every function on Ω lies
in Z(d).
Proof of lemma 2.1. Pick an element a ∈ Ω, and let g be the unique monic polynomial whose
roots are D(Ω \ {a}). The zonal polynomial ζa (g) then vanishes on all elements of Ω except a.
We can do this for all points, and get a set of |Ω| zonal polynomials that span the set of all
functions on Ω, proving the lemma.

Lemma 2.2. If dim(P (1)) = k + 1 for finite k, then dim(P (r)) ≤ r+k


r . Also, dim(Z(r)) ≤
r+k


r .

Proof of lemma 2.2. Pick a basis of size k + 1 for P (1). By definition, any element in P (r)
is spanned by products of r elements in P (1). Any product of r elements from P
(1) can be
written as the sum of products of r elements from the basis, of which there are r+k
r . The final
statement follows from the fact that Z(r) ⊂ P (r) by definition.

Going back to the Johnson Scheme, we have the following lemma:

n


Lemma 2.3. If (Ω, ρ) is the Johnson Scheme J(n, k), then dim(P (r)) = r for r = 0, . . . , k.
We defer the full proof of this lemma to a later section. The proof will proceed by showing
that the space P (r) is spanned by some indicator vectors (note that since Ω is finite, each vector
defines a function, and vice-versa). These vectors will be the rows of the containment matrix1 .
We now state the first application of this setting:
Theorem 2.4. Let (Ω, ρ) be a polynomial space, and let Φ be a subset of Ω with degree d. Then
|Φ| ≤ dim(Z(d)).
1
Recall that for i ≤ j, the containment matrix N (i, j) has rows spanned by all subsets of size i and columns
spanned by all subsets of size j, and N (i, j)x,y = 1 if and only if x ⊂ y.

3
 Before we prove this, let us see what this implies in the case of Johnson Schemes. As stated
before, for a set family Φ, the degree is exactly equal to all possible sizes of intersections of
its members. Therefore, the thorem says that if we have a set family whose intersections take
values from a set of size
s, then the size of the family is atmost dim(Z(s)), which, by 2.3 and 2.2
is upper bounded by ns . In other words, we recover the following theorem by Ray-Chaudhuri
and Wilson, which itself generalises EKR:

Theorem 2.5 (Theorem 3 in Ray-Chaudhuri and Wilson (1975)). Let A be a family of subsets

for all x, y ∈ A, we have |x ∩ y| ∈ {µ1 , . . . , µs }, where the µi are
of size k of [n] such that
distinct. Then |A| ≤ ns .

We now complete the proof.

Proof of theorem 2.4. Let g be the unique monic polynomial of degree d that has the elements
of D(Φ) as its roots. For all a ∈ Ω, let ha := ζa (g). By definition we have that ha ∈ Z(d)
for every a. Further, each ha vanishes on all elements of Φ except a itself. Thus the ha are
linearly independent in Z(d). Since we have |Φ| many linearly independent elements, we must
have |Φ| ≤ dim(Z(d)), completing the proof.

Next, we prove an extension of this theorem.

Theorem 2.6. Let (Ω, ρ) be a polynomial space such that ρ only takes integer valued. Let p be
a prime. Suppose Φ is a subset of Ω such that D(Φ) has exactly d0 distinct elements modulo p,
and also these elements are distinct from ρ(a, a) modulo p. Then |Φ| ≤ dim(Z(d0 )).

Again, before we prove the theorem, we see what it implies in the case of Johnson Schemes.
Using the exact same argument as before, we can see that this exactly recovers the Frank Wilson
theorem.

Proof of theorem 2.6. The proof will be very similar to that of theorem 2.4. The set D(Φ) splits
into d0 residue classes modulo p. Pick an element from each residue class, say the representative
in [0, p − 1], and let g be the unique monic degree d0 polynomial (with integer coefficients) that
vanishes on these points. We can construct the functions ha := ζa (g) as before. Each ha ∈ Z(d0 ).
We will argue that these elements are linearly independent over the reals, and this will complete
the proof, since we have |D(Φ)| linearly independent elements from Z(d0 ).
Each ha is 0 modulo p on all points in Φ except at a, since ha (x) = g(ρ(a, x)), and going
modulo p we have g(ρ(a, x)) (mod p) = g(ρ(a, x) (mod p)) = 0. Thus the ha are independent
modulo p. Since they are independent modulo p, they are also independent over the integers:
If they satisfied a non-trivial linear relationship over the integers, we could first make sure that
the coefficients are relatively prime, and then go modulo p, which will give us a non-trival
relationship (modulo p), a contradiction. This further implies that they are also independent
over the rationals, since if we had a rational linear combination that vanished, we could have
cleared out the denominators. Finally, since the coefficients are integers, this will also imply
that the functions are independent over the reals. This completes the proof.

We will now define the notion of designs. For the Johnson scheme, this notion of designs
will coincide with the combinatorial notion, which will allow us to recover another important
theorem from Ray-Chaudhuri and Wilson (1975).

4
2.4 Designs
We first introduce the additional axiom that we require on polynomial spaces: We require that
P admits a real valued inner product h·, ·i that satisfies the following:

• For all g, h ∈ P and for all a ∈ Ω we have hζa (x)g, hi = hg, ζa (x)hi.

• If f ∈ P is non-negative, then hf, 1i ≥ 0, where 1 is the constant 1 function.

If Ω is finite, the inner product will always be given by

1 X
hf, gi = E [f g] = f (x)g(x).
x∈Ω |Ω|
x∈Ω

A t-design in a polynomial space (Ω, ρ) is a finite subset Φ such that whenevr f, g are functions
with f g ∈ P (t), then we have
1 X
hf, gi = E [f g] = f (x)g(x).
x∈Φ |Φ|
x∈Φ

We have the following, when Φ is a t-design, for f, g such that f g ∈ P (t):

hf, gi = E [f g]
x∈Φ
= E [1 × (f g)] (Since the sum is finite)
x∈Φ
= h1, f gi

If we assume that hf, 1i is the average value of f over Ω, as is the case for the finite inner product,
then a t-design allows us to compute this average, which could be over a potentially infinite set,
by just averaging over a finite set. Of course there is no guarantee that t-designs will always exist,
when Ω is infinite. However, if the polynomial space Ω satisfies a slightly stronger condition,
then we can guarantee the existance of weighted t-designs (which are basically t-designs, where
distribution over elements is not uniform), even when Ω is infinite. The proof of this is pretty
neat, and involves a cone duality argument. However, since these types of polynomial spaces
are not the focus of this presentation, we move the formal statement and proof of this result to
Appendix A.
The first theorem on t-designs is the following bound, which says that t-designs cannot be
too small.

Theorem 2.7. Let Φ be a t-design in Ω. Then |Φ| ≥ dim(P ol(Ω, dt/2e).

As before, we first look at the case of Johnson Schemes. For this, we first need a definition.
A t − {n, k, λ} design in the combinatorial sense (abbreviated t-design) is a subset of k-sized
subsets of [n] such that every t sized subset of [n] is contained in exactly λ blocks.
We have the following lemma about Johnson schemes, whose proof we defer.

Lemma 2.8. If (Ω, ρ) is the Johnson Scheme J(n, k), then t-designs in this polynomial space
are equivalent to t-designs in the combinatorial sense.

5
 The above two allows us to recover the following theorem by Ray-Chaudhuri and Wilson

Theorem 2.9 (Theorem 1 in Ray-Chaudhuri
and Wilson (1975)). Let A be a t-design with
t = 2s, and such that n ≥ k + s. Then |A| ≥ ns .

We now prove the theorem.

Proof of theorem 2.7. Pick an orthogonal basis for P (dt/2e), say h1 , . . . , hn . By the definition
of t-designs, we have that

hhi , hj i = E [hi hj ] = δij .

x∈Φ

The restrictions of ha to the set Φ thus form a pairwise orthogonal set in PΦ 2 . Since Φ is finite,
we have that |Φ| = dim(P (Φ)) ≥ n, completing the proof.

In the next section we complete the proofs of the two lemma regarding Johnson Schemes,
and in the section after that we will see more instantiations of this framework.

3 Johnson Schemes
We need to prove lemma 2.3 and lemma 2.8. First we introduce a class of indicator functions.
For any set s ⊂ [n], let fs (x) = 1 if and only if s ⊂ x. This is the characteristic function for the
k subsets containing s. Further define the space I(r) = span ({fs | |s| = r}). The next lemma
captures some simple properties of I(r).

Lemma 3.1. The spaces I(r) satisfy I(r + 1) ⊆ span (I(r) · I(1)). Further, for all r ≤ s they
also satisfy I(r) ⊆ I(s).

Proof of lemma 3.1. The first lemma follows since fs ft = fs∪t when s, t are subsets. The second
follows from the identity
1 X
ft = k−|t|

fu .
u−|t| t⊆u,|s|=u

We also have the following lemma that related I(r) and P (r).

Lemma 3.2. For J(n, k) we have I(r) = P (r).

Proof of 3.2. Since I(r) and P (r) have the same recursive definition, we only need to prove the
result for r = 1. Let H be the incidence matrix of points and k subsets: the rows and indexed by
points and the columns by subsets. The matrix H T H then just has entries ρ(x, y), and thus the
rows of this matrix give elements of P (1). Infact the columns of H T H have constant sums, and
this value is non-zero. Thus every function in P (1) is in the row span. This gives P (1) ⊆ I(1).
Now using the fact that rank(H) = rank(H T H), we get the result P (1) = I(1), which completes
the proof.
2
Recall that PΦ is the ring of functions restricted to Φ.

6
 With these definitions, we are ready to prove the lemma. We will first prove lemma 2.8.

Proof of lemma 2.8. By lemma 3.2 and the argument about averages in subsection 2.4 (note
that Ω is finite here), we only need to show that

h1, f i = E [f ]
x∈Φ

for all functions f ∈ I(t), if and only if Φ is a t-design in the combinatorial sense.
The term on the right counts the number of elements of Φ that contain s, and normalises
n−t


this number by |Φ|. The term on the left is just k−t , normalised by |Ω|. These two are equal
for all t sized subsets of [n] if and only if all of them lie in the same number of elements of Φ.
This completes the proof.

We will finally prove lemma 2.3, which will conclude our discussion on Johnson schemes.

Proof of lemma 2.3. By lemma 3.2, we just have to prove that the indicator functions fs are
linearly independent. This is equivalent to proving that the incidence matrix N (r, k) is full rank.
A simple proof of this can be found in Foody and Hedayat (1977) (see lemma 5.1 within).

We conclude by very briefly discussing the other two mentioned schemes.

4 More Schemes
4.1 Hamming Schemes
A t − {q, k, λ} orthogonal array over a q-sized alphabet Σ is an N × k matrix M with entries
from Σ such that the following holds: for each s ≤ t, each s length tuple in Σs occurs λs times
as a row of the N × s matrix formed by choosing s columns, and also λ = λt . It turns out that
a t-design is the same thing as a t − {q, k, λ} orthogonal array. It also turns out that a linear
code C is a t-design if and only if its dual code has minimum distance greater than t. Thus we
get an upper bound on the size of linear codes, given a lower bound on the distance.

4.2 Unit Sphere

While the underlying space in the
case of the
unit sphere is infinite, the dimension of P (r) is
still finite, and is equal to n+r−1
r−1 + n+r−2
r−2 . This space is closed and bounded in the sense
described in Appendix A, and thus the results of that section apply. The codes and designs in
this setting are well studied, see for example Delsarte et al. (1977). In particular, (weighted)
t-designs here correspond exactly to numerical integration schemes with precision t.

7
5 Bibliography
P. Delsarte, J. M. Goethals, and J. J. Seidel. Spherical codes and designs. Geometriae Dedicata,
6(3):363–388, Sep 1977. ISSN 1572-9168. doi: 10.1007/BF03187604. URL https://doi.org/
10.1007/BF03187604.

W Foody and A Hedayat. On theory and applications of bib designs with repeated blocks. Ann
Statist, 5, 09 1977. doi: 10.1214/aos/1176343949.

C.D. Godsil. Polynomial spaces. Discrete Mathematics, 73(1):71 – 88, 1988. ISSN 0012-365X.
doi: https://doi.org/10.1016/0012-365X(88)90134-3. URL http://www.sciencedirect.
com/science/article/pii/0012365X88901343.

Dijen K. Ray-Chaudhuri and Richard M. Wilson. On t-designs. Osaka J. Math., 12(3):737–744,

1975. URL https://projecteuclid.org:443/euclid.ojm/1200758175.

8
A Existence of designs
Here we will sketch a proof of existence of weighted t-designs. First we define a weighted t-
design. A non-negative and finitely supported measure ω on Ω is called a weighted t-design if
the following holds for all functions f of degree atmost t:

h1, f i = E [f ] .
x∼ω

A null design of strength atmost t is a function ν with finite support such that hν, f i = 0 for all
functions f of degree atmost t. The difference of two weighted t-designs is always a null design.
We now define some preliminaries. A convex cone in Rn is a subset that is convex, and closed
under scaling by positive constants. Our cones will always be convex, and in the argument that
follows, we drop the adjective convex for brevity. An example of a cone is the set of all points
with all coordinates non-negative. Given a cone C, its dual cone C ∗ consists of all vectors x
such that hx, yi ≥ 0 for all y ∈ C. For cones that are closed, we have C = C ∗∗ .
We will now define a notion of closedness and boundedness for a polynomial space. This
will require a bunch of auxilary definitions. The cone we are interested in is the following. Let
U = P (t), and let U 0 denote the dual space. For every a ∈ Ω, let ω(a) : U → R be defined as
ω(a)(p) := p(a). Let C be the cone in U 0 generated by W := {ω(a) | a ∈ Ω}. The dual cone
then consists of all non-negative elements of U . Let C1 ⊂ C be the subset formed by those
elements α such that α(1) = 1. Then ω(a) ∈ C1 for all a ∈ Ω.
It is easy to check that C1 is a convex set. The claim is that the points ω(a) are extreme
points of C1 , that is points that are non themselves linear combinations of other points. To see
this, consider the function ρ(a, a) − ρ(a, x). The functional ω(b) is positive on this, unless b = a,
and thus the claim follows (Remember that the ω(a) generate C). We also have that all extreme
points of C1 must lie in W , since C and hence C1 is generated by W . Thus the set W is exactly
the set of all all extreme points of C1 .
If the set C1 is compact, then every one of its interior points will be linear combinations of
extreme points. The cone C will be closed if the set W is closed, and if this happens, the set C1
will also be closed (since it is an intersection of a closed set and a hyperplane). A polynomial
space will be called closed if the set W is closed. We can define a norm on U 0 . Pick a basis for
P (t), and for any u0 ∈ U 0 , define ku0 k to be maximum value of u0 on a basis element. The set U 0
will be bounded according to this norm if and only if the range of ρ is bounded. A polynomial
space is called bounded if this occurs.
Having defined closedness and boundedness, we can now state and prove the existence the-
orem.

Theorem A.1. Let (Ω, ρ) be a closed bounded polynomial space. If dim(P (t)) = d, then there
are weighted t-designs supported by atmost d points.

Proof of theorem A.1. Let the map λ : U → R be defined as λ(p) = h1, pi. Clearly λ ∈ C ∗∗ .
The cone C is closed, and hence C ∗∗ = C, and thus λ ∈ C, and also λ ∈ C1 . The set C1 is
a compact set in a d − 1 dimensional space, and thus we can write λ as a convex combination
of d − 1 + 1 = d extreme points. Thus λ is a convex combination of atmost d of the ω(a),
completing the proof.