1 Introduction and main results

Let G be an Abelian topological semigroup with unit, and let C(G) denote the set of complex valued continuous functions defined on G. By a polynomial we mean a function of the form \(P(\alpha _1 ,\ldots ,\alpha _k )\), where \(P\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\) and \(\alpha _1 ,\ldots ,\alpha _k\) are continuous additive functions. An exponential is a function \(m\in C(G)\) such that \(m\ne 0\) and \(m(x+y)=m(x)\cdot m(y)\) for every \(x,y\in G\). Functions of the form \(\sum _{i=1}^np_i \cdot m_i\), where \(p_i\) is a polynomial and \(m_i\) is an exponential for every \(i=1,\ldots ,n\), are called exponential polynomials. In these representations we may assume that \(p_i \ne 0\) for every i, and the exponentials \(m_1 ,\ldots ,m_n\) are distinct. One can show that such a representation is unique (see [15, Lemma 4.3, p. 41] or [6, Lemma 6]). Our starting point is the following well-known result.

Theorem A

Let G be an Abelian topological semigroup with unit. If V is a finite dimensional translation invariant linear subspace of C(G), then every element of V is an exponential polynomial. More precisely, every element of V is of the form \(\sum _{i=1}^kp_i \cdot m_i\), where \(p_1 ,\ldots ,p_k\) are polynomials, and \(m_1 ,\ldots ,m_k\) are exponentials belonging to V.

An equivalent form of the theorem above states that if a continuous function \(f\in C(G)\) satisfies the equation \(f(x+y)=\sum _{i=1}^nf_i (x)\cdot g_i (y)\) for every \(x,y\in G\) with suitable functions \(f_i , g_i :G\rightarrow {{\mathbb {C}}}\), then f is an exponential polynomial. For continuously differentiable functions on \({{\mathbb {R}}}\) this was proved by Levi-Civita in 1913 [10]. For continuous functions defined on \({{\mathbb {R}}}\) a much more general statement was proved by L. Schwartz in 1947 (see [12]). A simple proof in the case of \(G={{\mathbb {R}}}\) was given by Anselone and Korevaar in 1964; see [3] and also [9]. The generalization for Abelian (topological) semigroups was treated by Stone in 1960 and McKiernan in 1977; see [13] and [11]. The case of Abelian topological groups was also discussed by Laird and by Székelyhidi, see [8, 15, Theorem 10.1], [16]. (As for the history of Theorem A, see also [2] and the references given there.) The result was generalized for measurable (instead of continuous) functions defined on locally compact Abelian groups, see [5, 14] and [15, Theorem 10.2].

Let G be an Abelian semigroup with unit, and let \({{\mathbb {C}}}^G\) denote the set of complex valued functions defined on G. The discrete topology makes G a topological semigroup with \(C(G)={{\mathbb {C}}}^G\). A function p will be called a discrete polynomial, if it is a polynomial w.r.t. the discrete topology; that is, if \(p=P(\alpha _1 ,\ldots ,\alpha _k )\), where \(P\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\) and \(\alpha _1 ,\ldots ,\alpha _k :G\rightarrow {{\mathbb {C}}}\) are additive functions. Functions of the form \(\sum _{i=1}^kp_i \cdot m_i\), where \(p_1 ,\ldots ,p_k\) are discrete polynomials and \(m_1 ,\ldots ,m_k\) are exponentials are called discrete exponential polynomials.

From Theorem A it follows that if V is a finite dimensional translation invariant linear subspace of \({{\mathbb {C}}}^G\), then every element of V is a discrete exponential polynomial. Moreover, the exponential functions appearing in the representations of the elements of V belong to V.

Our aim is to generalize Theorem A by replacing the set of continuous (or measurable) functions by a subalgebra of \({{\mathbb {C}}}^G\). Suppose a subalgebra \({{\mathcal {A}}}\) of \({{\mathbb {C}}}^G\) is given. We say that p is an admissible polynomial with respect to \({{\mathcal {A}}}\), if \(p=P(\alpha _1 ,\ldots ,\alpha _k )\), where \(P\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\) and \(\alpha _1 ,\ldots ,\alpha _k\) are additive functions belonging to \({{\mathcal {A}}}\).

Looking for a generalization of Theorem A for function classes, the following question arises. Suppose V is a finite dimensional translation invariant linear subspace of the given algebra \({{\mathcal {A}}}\). We know that every element of V is of the form \(\sum _{i=1}^kp_i \cdot m_i\), where \(p_1 ,\ldots ,p_k\) are discrete polynomials, and \(m_1 ,\ldots ,m_k \in V\) are exponentials. When can we expect that in these representations \(p_1 ,\ldots ,p_n\) are admissible w.r.t. \({{\mathcal {A}}}\)?

If the algebra \({{\mathcal {A}}}\) contains the constant functions, then every polynomial admissible w.r.t. \({{\mathcal {A}}}\) is an element of \({{\mathcal {A}}}\). Therefore, a necessary condition for such a generalization to hold is that if p is a discrete polynomial and m is an exponential, and the translates of \(p\cdot m\) belong to \({{\mathcal {A}}}\), then \(p \in {{\mathcal {A}}}\). This is not true for every algebra. For example, let G be the additive group of \({{\mathbb {R}}}\), and let

$$\begin{aligned} {{\mathcal {A}}}=\left\{ c+\sum _{i=1}^np_i \cdot e^{ix} :n\ge 0, \ c\in {{\mathbb {R}}}, \ p_1 ,\ldots ,p_n \in {{\mathbb {C}}}[x] \right\} . \end{aligned}$$

Then \({{\mathcal {A}}}\) is an algebra over \({{\mathbb {C}}}\) containing the constant functions, the translates of \(x\cdot e^x\) belong to \({{\mathcal {A}}}\), but \(x\notin {{\mathcal {A}}}\).

We show that the following simple condition is sufficient:

$$\begin{aligned} \text {if} \ m\in {{\mathcal {A}}}\ \text {is an exponential, then} \ m^{-1}\in {{\mathcal {A}}}. \end{aligned}$$
(1)

Clearly, (1) is satisfied if \({{\mathcal {A}}}\) contains the constant functions and \({{\mathcal {A}}}\) is closed under division; that is, if \(f,g\in {{\mathcal {A}}}\) and \(g\ne 0\) implies \(f/g\in {{\mathcal {A}}}\).

In particular, if \({{\mathcal {T}}}\) is a topology on G (not necessarily compatible with the group operations), then the set of continuous functions with respect to \({{\mathcal {T}}}\) is an algebra satisfying (1), and so is the set of Borel measurable functions. Or, if \(\Sigma \) is an arbitrary \(\sigma \)-algebra on G, then the set of functions measurable with respect to \(\Sigma \) is also an algebra satisfying (1).

If G is an Abelian group and m is an exponential on G, then we have \(m^{-1}(x)=m(-x)\) for every \(x\in G\). Therefore, condition (1) is also satisfied if G is an Abelian group and the algebra \({{\mathcal {A}}}\) is symmetric; that is, if \(f \in {{\mathcal {A}}}\) implies \({\hat{f}}\in {{\mathcal {A}}}\), where \({\hat{f}} (x)=f(-x)\) for every \(x\in G\).

Our main result is the following.

Theorem 1

Let G be an Abelian semigroup with unit, and let \(\AA \subset {{\mathbb {C}}}^G\) be an algebra over \({{\mathbb {C}}}\) satisfying condition (1). If V is a finite dimensional translation invariant linear subspace of \({{\mathcal {A}}}\), then every element of V is of the form \(\sum _{i=1}^kp_i \cdot m_i\), where \(p_1 ,\ldots ,p_k\) are admissible polynomials w.r.t. \({{\mathcal {A}}}\), and \(m_1 ,\ldots ,m_k\) are exponentials belonging to V.

Theorem 1 is an easy consequence of Theorem A and the following result.

Theorem 2

Let \({{\mathcal {A}}}\) be a subalgebra of \({{\mathbb {C}}}^G\). A function f is an admissible polynomial w.r.t. \({{\mathcal {A}}}\) if and only if f is a discrete polynomial, and all translates of f belong to \({{\mathcal {A}}}\).

Note the following special case. IfGis a topological Abelian semigroup with unit, then\(f:G\rightarrow {{\mathbb {C}}}\)is a polynomial if and only iffis a continuous discrete polynomial.

Taking Theorem 2 for granted, we can prove Theorem 1 as follows. Let V be a finite dimensional translation invariant linear subspace of an algebra \({{\mathcal {A}}}\) satisfying condition (1). By Theorem A, every element \(f\in V\) is of the form \(f=\sum _{i=1}^np_i \cdot m_i\), where \(p_1 ,\ldots ,p_n\) are nonzero discrete polynomials, and \(m_1 ,\ldots ,m_n\) are distinct exponentials belonging to V. Then we have \(p_i \cdot m_i \in V\) and \(m_i \in V\) for every \(i=1,\ldots ,n\) (see [6, Lemma 6]). Fix an index i. Since \(m_i \in V\subset {{\mathcal {A}}}\), it follows from condition (1) that \(p_i \in {{\mathcal {A}}}\). By the translation invariance of V we have

$$\begin{aligned} T_h f =\sum _{i=1}^nm_i (h)\cdot T_h p_i \cdot m_i \in V \end{aligned}$$

for every \(h\in G\). This implies, by the argument above, that \(T_h p_i \in {{\mathcal {A}}}\) for every \(h\in G\). Thus the translates of \(p_i\) belong to \({{\mathcal {A}}}\). Therefore, by Theorem 2, we obtain that \(p_i\) is admissible w.r.t. \({{\mathcal {A}}}\). \(\square \)

In the argument above, we only used the ‘if’ part of Theorem 2, which is an immediate consequence of the following result.

Theorem 3

Let G be an Abelian semigroup, and let \({{\mathcal {P}}}_G\) denote the algebra of discrete polynomials on G. A subalgebra \(\AA \subset {{\mathcal {P}}}_G\) is translation invariant if and only if \(\AA =\{ 0\}\) or \({{\mathcal {A}}}\) is generated by a set of additive functions and the constant functions.

Now suppose that the translates of a discrete polynomial f belong to an algebra \({{\mathcal {A}}}\). We may assume \(f\ne 0\). Let \(\AA _f\) denote the algebra generated by the translates of f. Then \(\AA _f \subset {{\mathcal {A}}}\) and \(\AA _f\) is translation invariant. By Theorem 3, \(\AA _f\) is generated by a set of additive functions and the constant functions. Since \(f\in \AA _f\), this implies that \(f=P(b_1 ,\ldots ,b_k )\), where \(P\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\), and \(b_1 ,\ldots ,b_k\) are additive functions belonging to \(\AA _f\), hence also to \({{\mathcal {A}}}\). This means that f is admissible w.r.t. \({{\mathcal {A}}}\), proving the ‘if’ part of Theorem 2.

The proof of Theorem 1 will be completed in the next section, where we prove Theorems 2 and 3.

In Sect. 3 we also give a direct proof of Theorem 1 using none of Theorem A or Theorems 2 and 3. The reason why we present such an independent proof is that it seems to be more elementary than the existing proofs of Theorem A, which is a special case of Theorem 1. We also show that Theorems 23, and some other results we present in the next section are easy consequences of Theorem 1.

2 Proof of Theorems 2 and 3

If G is an Abelian semigroup, \(h\in G\) and \(f:G\rightarrow {{\mathbb {C}}}\), then we denote by \(T_h f\) the function \(x\mapsto f(x+h)\)\((x\in G)\). We put \(\Delta _h f=T_h f -f\) for every \(h\in G\). If \(f:G\rightarrow {{\mathbb {C}}}\), then we denote by \(V_f\) and \(\AA _f\), respectively, the linear space and the algebra generated by the set \(\{ T_h f :h\in G \}\).

The set of complex valued linear functions defined on \({{\mathbb {C}}}^n\) will be denoted by \(\Lambda _n\).

Theorem 4

  1. (i)

    Every homogeneous polynomial \(f\in {{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\) can be represented in the form \(p(\ell _1 ,\ldots ,\ell _k )\), where \(k\le n\), \(p\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\) and \(\ell _1 ,\ldots ,\ell _k \in V_f \cap \Lambda _n\).

  2. (ii)

    Every polynomial \(f\in {{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\) can be represented in the form \(p(\ell _1 ,\ldots ,\ell _k )\), where \(k\le n\), \(p\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\) and \(\ell _1 ,\ldots ,\ell _k \in \AA _f \cap \Lambda _n\).

Proof

We prove (ii) by induction on \(\mathrm{deg}\, f\) and, along the proof, we also prove (i). Let \(f\in {{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\) be arbitrary. Statement (ii) is true if f is constant. Let \(\mathrm{deg}\, f=d>0\), and suppose that (ii) is true for all polynomials of degree \(<d\). It is easy to see that the constant functions belong to \(V_f\). Indeed, \(\Delta _{h_1} \cdots \Delta _{h_d} f\) is a nonzero constant function for suitable \(h_1 ,\ldots ,h_d \in {{\mathbb {C}}}^n\), and belongs to \(V_f\). Since \(V_f\) is a linear space, it follows that \(V_f\) contains every constant function.

Obviously, \(\Lambda _n\) is an n dimensional linear space over the complex field. The set \(V_f \cap \Lambda _n\) is a linear subspace of \(\Lambda _n\), and thus its dimension, k, is at most n. Let \(b_i =\beta _{i,1} x_1 +\cdots +\beta _{i,n}x_n\)\((i=1,\ldots ,k)\) be a basis of \(V_f \cap \Lambda _n\). Since \(b_1 ,\ldots ,b_k\) are linearly independent, the rank of the matrix \(A=(\beta _{i,j})\)\((i=1,\ldots ,k, \ j=1,\ldots ,n)\) equals k, and thus A contains a \(k\times k\) submatrix with nonzero determinant. We may assume that the determinant \(|\beta _{i,j}|_{i,j=1,\ldots ,k}\) is nonzero, since otherwise we may apply a suitable permutation of the variables \(x_1 ,\ldots ,x_n\). Then there is an invertible system of linear substitutions

$$\begin{aligned} x_i =\gamma _{i,1} y_1 +\cdots +\gamma _{i,n}y_n \qquad (i=1,\ldots ,n) \end{aligned}$$
(2)

that transforms \(b_i\) to \(y_i\) for every \(i=1,\ldots ,k\). (Put \(x_i =y_i\) for every \(k<i\le n\), and solve the system of equations \(b_i =y_i\)\((i=1,\ldots ,k)\) for \(x_1 ,\ldots ,x_k\).)

Let \(p\in {{\mathbb {C}}}[y_1 ,\ldots ,y_n ]\) be the polynomial obtained from f by substitution (2). It is easy to check that substitution (2) induces a bijection from \(V_f\) onto \(V_p\). Since \(b_i \in V_f\) for every \(i=1,\ldots ,k\), it follows that \(y_1 ,\ldots ,y_k \in V_p\). Considering that \(b_1 ,\ldots ,b_k\) is a basis of \(V_f \cap \Lambda _n\), we find that a linear function belongs to \(V_p\) if and only if it involves the variables \(y_1 ,\ldots ,y_k\) only.

Let \(f_d\) and \(p_d\) denote, respectively, the sum of all terms of f and p with degree d. It is clear that \(p_d\) is obtained from \(f_d\) by substitution (2).

Let \(p=\sum _i c_i \cdot y_1^{i_1}\cdots y_n^{i_n}\), where \(i=(i_1 ,\ldots ,i_n )\) runs through the n-tuples \((i_1 ,\ldots ,i_n )\) with \(i_1 + \cdots + i_n \le d\). We claim that \(p_d \in {{\mathbb {C}}}[y_1 ,\ldots ,y_k ]\); that is, none of the variables \(y_{k+1} ,\ldots ,y_n\) occur in a term of \(p_d\) with nonzero coefficient. Indeed, suppose that \(c_i \cdot y_1^{i_1}\cdots y_n^{i_n}\) is a term in \(p_d\) such that \(c_i \ne 0\), \(i_1 +\cdots +i_n =d\), and there is an index \(j>k\) with \(i_j >0\). We may assume that \(j=n\). Let \(e_1 =(1,0 ,\ldots ,0) ,\ldots ,e_n =(0 ,\ldots ,0, 1)\) be the standard basis of \({{\mathbb {C}}}^n\), and let D denote the operator \(\Delta _{e_1}^{i_1} \cdots \Delta _{e_{n-1}}^{i_{n-1}} \Delta _{e_n}^{i_n -1}\). Clearly, \(Dp\in V_p\). Let \(c_j \cdot y_1^{j_1}\cdots y_n^{j_n}\) be an arbitrary term in p. Then its image under D is constant unless \(j_1 \ge i_1 ,\ldots ,j_{n-1}\ge i_{n-1}\), and \(j_n \ge i_n -1\), and at least one of the inequalities is strict. The images of these terms are distinct, and are of the form \(\gamma _i y_i + \delta _i\), where \(\gamma _i ,\delta _i \in {{\mathbb {C}}}\). Now one of these terms equals \(\gamma _n y_n + \delta _n\) with a nonzero \(\gamma _n\), and thus \(D p=\gamma _1 y_1 + \cdots + \gamma _n y_n +\) constant, where \(\gamma _n \ne 0\). Since the constant functions belong to \(V_p\), it follows that \(\gamma _1 y_1 + \cdots + \gamma _n y_n \in V_p \cap \Lambda _n\). This, however, contradicts the fact that all linear functions in \(V_p\) involve the variables \(y_1 ,\ldots ,y_k\) only. Therefore, we have \(p_d \in {{\mathbb {C}}}[y_1 ,\ldots ,y_k ]\).

Since \(p_d\) is obtained from \(f_d\) by substitution (2) and the substitution is invertible and transforms \(b_i\) to \(y_i\)\((1\le i\le k)\), it follows that

$$\begin{aligned} f_d =p_d (b_1 ,\ldots ,b_k ). \end{aligned}$$
(3)

Now \(b_1 ,\ldots ,b_k\) are linear functions belonging to \(V_f\). Therefore, if f is homogeneous, then \(f=f_p\), and (i) holds true.

In the general case let \({{\mathcal {L}}}_f\) denote the algebra generated by \(\AA _f \cap \Lambda _n\) and the constant functions. We have to prove \(f\in {{\mathcal {L}}}_f\).

By (3), we have \(f_d \in {{\mathcal {L}}}_f \subset \AA _f\). Putting \(g=f-f_p\) we find \(g\in \AA _f\), and thus we have \(\AA _g \subset {{\mathcal {A}}}_f\). Since \(\mathrm{deg}\, g <d\), it follows from the induction hypothesis that g belongs to the algebra generated by \(\AA _g \cap \Lambda _n\) and the constant functions. Since \(\AA _g \cap \Lambda _n \subset \AA _f \cap \Lambda _n\), we have \(g\in {{\mathcal {L}}}_f\). Therefore, by \(f_p \in {{\mathcal {L}}}_f\) we obtain \(f=f_d +g \in {{\mathcal {L}}}_f\). \(\square \)

Remark 5

In statement (ii) of Theorem 4 we cannot replace \(\AA _f \cap \Lambda _n\) by \(V_f \cap \Lambda _n\), as the following example shows. Let \(n=2\) and \(f(x,y)=x^2 +y\in {{\mathbb {C}}}[x,y]\). It is easy to check that

$$\begin{aligned} V_f =\{ a\cdot x^2 +a\cdot y +bx +c:a,b,c\in {{\mathbb {C}}}\} . \end{aligned}$$

Thus the linear functions belonging to \(V_f\) are the functions bx\((b\in {{\mathbb {C}}})\). Clearly, f is not in the algebra generated by these functions and the constants.

Lemma 6

Suppose \(c_1 ,\ldots ,c_n \in {{\mathbb {C}}}^n\) are linearly independent over \({{\mathbb {C}}}\). If a linear subspace V of \({{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\) is invariant under the translations \(T_{c_i}\)\((i=1,\ldots ,n)\), then V is invariant under all translations.

Proof

Let H denote the additive group generated by \(c_1 ,\ldots ,c_n\). Then V is invariant under translations by elements of H. Let \(e_1 ,\ldots ,e_n\) be the standard basis of \({{\mathbb {C}}}^n\), and let \(A:{{\mathbb {C}}}^n \rightarrow {{\mathbb {C}}}^n\) be the linear transformation mapping \(e_i\) to \(c_i\)\((i=1,\ldots ,n)\). Then A maps \({{\mathbb {Z}}}^n\) onto H.

The set \(V'=\{ p\circ A:p\in V\}\) is also a linear subspace of \({{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\). If \(p\in V\) and \(a\in {{\mathbb {Z}}}^n\), then \(T_a (p\circ A)(x)=p(A(x)+h)=q(A(x))\), where \(h\in H\) and \(q=T_h p\). By assumption, \(q\in V\), and thus \(T_a (p\circ A) \in V'\). Therefore, \(V'\) is invariant under translations by the elements of \({{\mathbb {Z}}}^n\). By [7, Lemma 7], this implies that \(V'\) is invariant under all translations. Thus the same is true for \(V=\{ q\circ A^{-1}:q\in V'\}\). \(\square \)

Proof of Theorem 3

Let G be an arbitrary Abelian semigroup, and suppose that the algebra \({{\mathcal {A}}}\) is generated by a set \({{\mathcal {L}}}\) of additive functions and the constant functions. If \(a\in {{\mathcal {L}}}\), then \(a(x+h)=a(x)+a(h)\) for every \(x\in G\), and thus \(T_h a =a +a(h)\in {{\mathcal {A}}}\). From this observation it is clear that \({{\mathcal {A}}}\) is translation invariant.

Now let \(\AA \subset {{\mathcal {P}}}_G\) be a nonzero translation invariant subalgebra. Then \({{\mathcal {A}}}\) contains the constant functions. Indeed, for every \(f\in {{\mathcal {A}}}\), \(f\ne 0\), already \(V_f\) (a subset of \({{\mathcal {A}}}\)) contains all constant functions.

Let \(f\in {{\mathcal {A}}}\) be arbitrary. Then \(f=p(a_1 ,\ldots ,a_n )\), where \(p\in {{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\), and \(a_1 ,\ldots ,a_n :G\rightarrow {{\mathbb {C}}}\) are additive. We may assume that \(a_1 ,\ldots ,a_n\) are linearly independent over \({{\mathbb {C}}}\). Then there are points \(h_1 ,\ldots ,h_n \in G\) such that the determinant \(|a_i (h_j )|\)\((i,j=1,\ldots ,n)\) is nonzero (see [1, Lemma 1, p. 229]). Let \(c_j =(a_1 (h_j ),\ldots ,a_k (h_j))\)\((j=1 ,\ldots ,n)\). Then the vectors \(c_1 ,\ldots ,c_n \in {{\mathbb {C}}}^n\) are linearly independent over \({{\mathbb {C}}}\).

Let V denote the set of polynomials \(q\in {{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\) such that \(q(a_1 ,\ldots ,a_n )\in {{\mathcal {A}}}\). Then V is a subalgebra of \({{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\). We prove that V is translation invariant. If \(q\in V\), then \(q(a_1 ,\ldots ,a_n )\in {{\mathcal {A}}}\) and thus, as \({{\mathcal {A}}}\) is translation invariant, we have \(T_{h_j} q(a_1 ,\ldots ,a_n )\in {{\mathcal {A}}}\) for every \(j=1,\ldots ,n\). It is clear that \(T_{h_j} q(a_1 ,\ldots ,a_n )=(T_{c_j } p)(a_1 ,\ldots ,a_n )\), and thus \(T_{c_j} q\in V\) for every \(q\in V\) and \(j=1,\ldots ,n\). By Lemma 6, this implies that V is translation invariant.

Since \(p(a_1 ,\ldots ,a_n ) =f\in {{\mathcal {A}}}\), we have \(p\in V\). Then, by the translation invariance of V, we have \(\AA _p \subset V\). By (ii) of Theorem 4, \(p=P(\ell _1 ,\ldots ,\ell _k )\), where \(k\le n\), \(P\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\), and

$$\begin{aligned} \ell _1 ,\ldots ,\ell _k \in {{\mathcal {A}}}_p \cap \Lambda _n \subset V\cap \Lambda _n . \end{aligned}$$

Now \(\ell _i \in V\) implies that \(b_i =\ell _i (a_1 ,\ldots ,a_n )\in {{\mathcal {A}}}\) for every \(i=1,\ldots ,k\). Thus we have

$$\begin{aligned} f=p(a_1 ,\ldots ,a_n )=P(\ell _1 (a_1 ,\ldots ,a_n ) ,\ldots ,\ell _k (a_1 ,\ldots ,a_n ))= P(b_1 ,\ldots ,b_k ), \end{aligned}$$

where \(b_1 ,\ldots ,b_k\) are additive functions belonging to \({{\mathcal {A}}}\). Since \(f\in {{\mathcal {A}}}\) was arbitrary, this proves that \({{\mathcal {A}}}\) equals the algebra generated by the set of additive functions belonging to \({{\mathcal {A}}}\) and the constant functions. \(\square \)

Proof of Theorem 2

The ‘if’ part was proved in Sect. 1. To prove the ‘only if’ part, let f be an admissible polynomial w.r.t. \({{\mathcal {A}}}\). Let \(f=P(\alpha _1 ,\ldots ,\alpha _k )\), where \(P\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\) and \(\alpha _1 ,\ldots ,\alpha _k\) are additive functions belonging to \({{\mathcal {A}}}\). Then f belongs to the algebra \({{\mathcal {B}}}\) generated by \(\alpha _1 ,\ldots ,\alpha _k\) and the constant functions. By Theorem 3, \({{\mathcal {B}}}\) is translation invariant. Therefore, every translate of f belongs to \({{\mathcal {B}}}\subset {{\mathcal {A}}}\). \(\square \)

3 A direct proof of Theorem 1

Let G be an Abelian semigroup with unit. A function \(f:G\rightarrow {{\mathbb {C}}}\) is a generalized polynomial, if there is an \(n\ge 0\) such that \(\Delta _{h_1} \cdots \Delta _{h_{n+1}} f=0\) for every \(h_1 ,\ldots ,h_{n+1} \in G\). The smallest n with this property is the degree of f denoted by \(\mathrm{deg}\, f\). The degree of the identically zero function is \(-1\) by definition. One can easily prove that every discrete polynomial is a generalized polynomial.

A function \(A:G^i \rightarrow {{\mathbb {C}}}\) is called i-additive, if it is additive in each of its variables, the other variables being fixed. A function \(g:G\rightarrow {{\mathbb {C}}}\) is a monomial of degree i, if there is an i-additive function A such that \(g(x)=A(x,\ldots ,x)\) for every \(x\in G\). It is well-known that every generalized polynomial can be represented in the form \(\sum _{i=0}^d f_i\), where \(f_i\) is a monomial of degree i for every \(i=1,\ldots ,d\), and \(f_0\) is constant. (See [4, Theorem 3].)

We fix an algebra \(\AA \subset {{\mathbb {C}}}^G\) satisfying condition (1). By an admissible polynomial we mean an admissible polynomial w.r.t. \({{\mathcal {A}}}\).

Lemma 7

Let f be an admissible polynomial, and let \(f=\sum _{i=0}^n f_i\), where \(f_i\) is a monomial of degree i for \(1\le i\le n\), and \(f_0\) is a constant. Then \(f_i\) is an admissible polynomial for every \(i=0,\ldots ,n\).

Proof

It is easy to see that \(f(kx)=\sum _{i=0}^n k^i \cdot f_i (x)\) for every \(x\in G\) and for every positive integer k. These equations for \(k=1,\ldots ,n+1\) constitute a linear system of equations with unknowns \(f_i (x)\)\((i=0,\ldots ,n)\). Since the determinant of this system is nonzero (being a Vandermonde determinant), it follows that each \(f_i (x)\) is a linear combination with rational coefficients of \(f(x),\ldots ,f((n+1)x)\). It is clear that each of the functions \(f(x),\ldots ,f((n+1)x)\) is an admissible polynomial, and thus the same is true of \(f_i (x)\)\((i=0,\ldots ,n)\).

\(\square \)

Lemma 8

Suppose that \({{\mathcal {A}}}\) contains the constant functions. If f is a generalized polynomial and \(x\mapsto f(2x)-2f(x)\) is an admissible polynomial, then f is the sum of an admissible polynomial and an additive function. If, in addition, \(f\in {{\mathcal {A}}}\), then f is an admissible polynomial.

Proof

Let \(f=\sum _{i=0}^n f_i\), where \(f_i\) is a monomial of degree i for every \(1\le i\le n\), and \(f_0\) is a constant. Then

$$\begin{aligned} f(2x)-2f(x)= -a_0 +\sum _{i=2}^n (2^i -2) f_i (x) \end{aligned}$$

for every \(x\in G\). Since \(f(2x)-2f(x)\) is an admissible polynomial by assumption, it follows from Lemma 7 that \(f_i\) is an admissible polynomial for every \(i\ne 1\). As \(f_1\) is additive, we obtain the first statement of the lemma. If \(f\in {{\mathcal {A}}}\), then \(f_1 =f-\sum _{i\ne 1} f_i \in {{\mathcal {A}}}\). Since \(f_1\) is additive, it follows that \(f_1\) is an admissible polynomial, and then so is f. \(\square \)

Lemma 9

Let V be a translation invariant and finite dimensional linear subspace of C(G). If \(V\ne 0\), then V contains an exponential.

Proof

Although the statement is well-known, we provide the proof for the sake of completeness. We prove by induction on the dimension of V. Let \(\mathrm{dim}\, V=n\), and suppose that either \(n=1\), or \(n>1\) and the statement is true for smaller positive dimensions.

Let \(h\in G\) be fixed. Then the translation operator \(T_{h}\) is a linear transformation mapping V into itself. Since \(\mathrm{dim}\, V<\infty \), it follows that \(T_{h}\) has an eigenvalue \(\lambda (h)\). Then \(L_h =T_h -\lambda (h) \cdot T_0\) is not invertible and, consequently, the image space \(\mathrm{Im}\, L_h\) is a proper subspace of V. It is clear that \(\mathrm{Im}\, L_h\) is a translation invariant linear subspace of V. If \(\mathrm{Im}\, L_h \ne 0\) then, by the induction hypothesis, \(\mathrm{Im}\, L_h\) contains an exponential, and we are done. Therefore, we may assume that \(\mathrm{Im}\, L_h =0\) for every h; that is, \(T_h f=\lambda (h)f\) for every \(h\in G\) and \(f\in V\).

Let \(f\in V\), \(f\ne 0\) be fixed. Then we have \(f(x+h)=\lambda (h)f(x)\) for every \(x\in G\). Thus \(f(h)=c\cdot \lambda (h)\) for every h, where \(c=f(0)\).Footnote 1 Then \(c\ne 0\), since otherwise f would be identically zero. Therefore, we find that \(\lambda \in V\), \(\lambda \ne 0\) and \(\lambda (x+h)=\lambda (h)\lambda (x)\) for every \(x,h\in G\); that is, \(\lambda \) is an exponential. \(\square \)

Proof of Theorem 1

We prove by induction on the dimension of V. The statement is true if \(V=0\). Let \(\mathrm{dim}\, V>0\), and suppose that the statement is true for smaller dimensions.

Let W denote the set of those elements of V that can be written in the form \(\sum _{i=1}^kp_i \cdot m_i\), where \(p_1 ,\ldots ,p_k\) are admissible polynomials and \(m_1 ,\ldots ,m_k\) are distinct exponentials belonging to V. It is easy to see that W is a translation invariant linear subspace of V. We have to prove that \(W=V\). Suppose this is not true, and fix an element \(f_0 \in V {\setminus } W\).

Let m be an arbitrary exponential contained in V, and let \(h\in G\) be fixed. Then \(L_h =T_{h} -m(h)\cdot T_0\) is a linear transformation mapping V into itself. Since \(\mathrm{dim}\, V<\infty \) and the kernel of \(L_h\) is nonzero as \(L_h (m) =0\), it follows that the image space \(\mathrm{Im} \, L_h\) of \(L_h\) is a proper subspace of V. Clearly, \(\mathrm{Im} \, L_h\) is translation invariant and thus, by the induction hypothesis, every element of \(\mathrm{Im} \, L_h\) is of the form \(\sum _{i=1}^kp_i \cdot m_i\), where \(p_1 ,\ldots ,p_k\) are admissible polynomials and \(m_1 ,\ldots ,m_k\) are exponentials belonging to \(\mathrm{Im} \, L_h\). Therefore, we have \(\mathrm{Im} \, L_h \subset W\) for every h.

We show that \(\langle f_0 , W\rangle =\{ c\cdot f_0 +p:c\in {{\mathbb {C}}},\ p\in W \}\) is a translation invariant linear subspace of V. It is clear that \(\langle f_0 , W\rangle \) is a linear subspace of V. Its invariance under translations follows from the facts that W is translation invariant, \(T_h f_0 =m(h)\cdot f_0 +L_h (f_0 )\), and \(L_h (f_0 )\in \mathrm{Im} \, L_h \subset W\).

If \(\langle f_0 , W\rangle \) is a proper subspace of V then, by the induction hypothesis, \(\langle f_0 , W\rangle \subset W\), and thus \(f_0 \in W\), a contradiction.

Therefore, we have \(\langle f_0 , W\rangle =V\); that is, every element of V is of the form \(c\cdot f_0 +p\)\((c\in {{\mathbb {C}}},\ p\in W)\). Let \(b_1 ,\ldots ,b_k\) be a basis of W. Since \(L_y f_0 \in W\), we have

$$\begin{aligned} f_0 (x+y)-m(y) f_0 (x)=L_y f_0 (x)=\sum _{i=1}^kb_i (x)\cdot q_i (y) \end{aligned}$$
(4)

for every \(x,y\in G\), with suitable functions \(q_i :G\rightarrow {{\mathbb {C}}}\).

We show that \(q_1 ,\ldots ,q_k \in V\). Since \(b_1 ,\ldots ,b_k\) are linearly independent, there are points \(x_1 ,\ldots ,x_k \in G\) such that the determinant \(|b_i (x_j )|\)\((i,j=1,\ldots ,k)\) is nonzero (see [1, Lemma 1, p. 229]). Applying (4) with \(x=x_j\) for every \(j=1,\ldots ,k\) we obtain a system of linear equations with unknowns \(q_i (y)\). Since the determinant of the system is nonzero, it follows that each of \(q_1 (y) ,\ldots ,q_k (y)\) is a linear combination of \(f_0 (x_j +y)-m(y)f_0 (x_j )\)\((j=1,\ldots ,k)\). The functions \(y\mapsto f_0 (x_j +y)-m(y)f_0 (x_j )\) belong to V, and then the same is true for \(q_1 ,\ldots ,q_k\).

Let \(q_i =c_i \cdot f_0 +p_i\), where \(c_i \in {{\mathbb {C}}}\) and \(p_i \in W\) for every \(i=1,\ldots ,k\). Then, by (4), we have

$$\begin{aligned} f_0 (x+y)-m(y)f_0 (x)&=\sum _{i=1}^kb_i (x)\cdot (c_i \cdot f_0 (y)+ p_i (y)) \\&=f_0 (y)\cdot \sum _{i=1}^kc_i \cdot b_i (x) + \sum _{i=1}^kb_i (x)\cdot p_i (y) \end{aligned}$$

and

$$\begin{aligned} f_0 (x+y)-m(y)f_0 (x) -m(x)f_0 (y) = B(x)\cdot f_0 (y)+\sum _{i=1}^kb_i (x)\cdot p_i (y) \end{aligned}$$
(5)

for every \(x,y\in G\), where \(B(x)=-m(x)+\sum _{i=1}^kb_i (x)\). If x is fixed, then the left hand side of (5), as a function of y, equals \(L_x f_0 -f_0 (x)\cdot m\in W\). Thus the right hand side of (5) also belongs to W. Since \(p_i \in W\) for every i, it follows that \(B(x)\cdot f_0 \in W\). Since \(f_0 \notin W\), it follows that \(B(x)=0\) for every x, and thus

$$\begin{aligned} f_0 (x+y)-m(y)f_0 (x) -m(x)f_0 (y) = \sum _{i=1}^kb_i (x)\cdot p_i (y) \end{aligned}$$
(6)

for every \(x,y\in G\).

Suppose that V contains two different exponentials \(m'\) and m. Then the argument above, with \(m'\) in place of m, gives

$$\begin{aligned} f_0 (x+y)-m'(y)f_0 (x) -m'(x)f_0 (y) = \sum _{i=1}^kb_i (x)\cdot p'_i (y) \end{aligned}$$

for every \(x,y\in G\), where \(p'_i (y) \in W\) for every \(i=1,\ldots ,k\). Subtracting (6) we get

$$\begin{aligned} (m(y) -m'(y))\cdot f_0 (x) +f_0 (y) \cdot (m(x)-m'(x)) = \sum _{i=1}^kb_i (x)\cdot (p'_i (y)- p_i (y)). \end{aligned}$$

If y is fixed, then the right hand side, as a function of x, belongs to W, and so does \(f_0 (y) \cdot (m(x)-m'(x))\). Consequently, if we fix y such that \(m(y) \ne m'(y)\), then we find \(f_0 \in W\), which is impossible.

Therefore, V contains at most one exponential. Now V contains at least one exponential by Lemma 9, and thus there is a unique exponential m contained by V. Since \(\mathrm{dim}\, W<\mathrm{dim}\, V\), it follows from the induction hypothesis that every element of W is of the form \(p\cdot m\), where p is an admissible polynomial. Let \(r_i\) and \(s_i\) be admissible polynomials such that \(b_i =r_i \cdot m\) and \(p_i =s_i \cdot m\) for every \(i=1,\ldots ,k\). Put \(P=f_0 /m\), then \(P\in {{\mathcal {A}}}\) by (1). Dividing (6) by \(m(x+y)=m(x)m(y)\) we obtain

$$\begin{aligned} P (x+y)-P (x) -P (y) = \sum _{i=1}^kr_i (x)\cdot s_i (y) \end{aligned}$$
(7)

for every \(x,y\in G\). Thus \(\Delta _y P =P(y)+\sum _{i=1}^ks_i (y) \cdot r_i\) for every y. Since \(\sum _{i=1}^ks_i (y) \cdot r_i\) is a polynomial of degree at most \(N=\max _{1\le i\le k} \mathrm{deg}\, r_i\), we have \(\Delta _{y_1} \ldots \Delta _{y_N} \Delta _{y} P=0\) for every \(y_1 ,\ldots ,y_k ,y\in G\). We obtain that P is a generalized polynomial. Applying (7) with \(y=x\) we find that \(P(2x)-2P(x)=\sum _{i=1}^kr_i (x)\cdot s_i (x)\) is an admissible polynomial. Then, since \(P\in {{\mathcal {A}}}\), Lemma 8 gives that P is an admissible polynomial. (Note that \({{\mathcal {A}}}\) contains the constant functions as \(m\cdot m^{-1}=1 \in {{\mathcal {A}}}\).) Since \(f_0 =P\cdot m\) and \(m\in V\), it follows that \(f_0 \in W\), contrary to our hypothesis. This final contradiction gives \(W=V\), completing the proof. \(\square \)

Remark 10

We show that Theorems 23 and (ii) of Theorem 4 are easy consequences of Theorem 1.

We prove the ‘if’ part of Theorem 2, the ‘only if’ part being simple. Suppose that f is a discrete polynomial, and all translates of f belong to \({{\mathcal {A}}}\). Let \(V_f\) denote the linear span of the translates of f, and let \(\AA _f\) denote the algebra generated by \(V_f\). Then \(\AA _f \subset {{\mathcal {A}}}\), and every element of \({{\mathcal {A}}}_f\) is a discrete polynomial. Therefore, the only exponential contained in \(\AA _f\) is the identically 1 function. Consequently, condition (1) is satisfied by \(\AA _f\).

Now \(V_f\) is of finite dimension, since f is a discrete polynomial. Since \(V_f \subset \AA _f\), it follows from Theorem 1 that f is an admissible polynomial with respect to \(\AA _f\) and then also with respect to \({{\mathcal {A}}}\).

Next we consider Theorem 3. We prove that if \(\AA \in {{\mathcal {P}}}_G\), \(\AA \ne 0\) and \({{\mathcal {A}}}\) is translation invariant, then \({{\mathcal {A}}}\) is generated by a set of additive functions and the constant functions. This follows from Theorem 2. First, \({{\mathcal {A}}}\) contains the constants, since already \(V_f\) contains the constants for every \(f\in {{\mathcal {A}}}\), \(f\ne 0\). If \(f\in {{\mathcal {A}}}\), then the translates of f belong to \({{\mathcal {A}}}\), and thus f is admissible. It is clear that every admissible function is in the algebra generated by the set of additive functions belonging to \({{\mathcal {A}}}\) and the constant functions.

As for (ii) of Theorem 4, let \(G={{\mathbb {C}}}^n\), and let \(f\in {{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\). Consider f as a function from \({{\mathbb {C}}}^n\) to \({{\mathbb {C}}}\). Then f is a polynomial on G, since \(x_1 ,\ldots ,x_n\), as functions defined on G, are additive. If \(\AA _f\) denotes the algebra generated by the translates of f then, by Theorem 2, f is an admissible polynomial w.r.t. \(\AA _f\). That is, \(f=Q(\alpha _1 ,\ldots ,\alpha _k )\), where \(Q\in {{\mathbb {C}}}[x_1 ,\ldots ,x_k ]\) and \(\alpha _1 ,\ldots ,\alpha _k\) are additive functions belonging to \(\AA _f\).

Since \(\AA _f \subset {{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\), it follows that \(\alpha _1 ,\ldots ,\alpha _k \in {{\mathbb {C}}}[x_1 ,\ldots ,x_n ]\). In particular, \(\alpha _1 ,\ldots ,\alpha _s\) are continuous on \({{\mathbb {C}}}^n\). Now, if a function \({{\mathbb {C}}}^n \rightarrow {{\mathbb {C}}}\) is additive and continuous then, necessarily, it is a linear function. Therefore, \(\alpha _1 ,\ldots ,\alpha _k \in \AA _f \cap \Lambda _n\), which proves (ii) of Theorem 4.