THE ALGEBRAIC STRUCTURE IN SIGNAL PROCESSING: TIME AND SPACE Markus Puschel¨ and Jose´ M. F. Moura Electrical and Computer Engineering Carnegie Mellon University Email: {pueschel,moura}@ece.cmu.edu ABSTRACT visualizations are undirected. We thus call them space models. Fi- nally, we briefly discuss the extension of our approach to 2-D SP The assumptions underlying linear signal processing (SP) produce for separable and nonseparable time and space models. more structure than vector spaces. We capture this structure by de- Organization. We introduce the concept of algebra and mod- scribing the space of filters as an algebra and the space of signals as ule in Section 2. In Section 3, we motivate why linear SP is al- the associated module. We formulate an algebraic approach to SP gebraic in nature and introduce the concept of a signal model that that is axiomatically based on the concept of a signal model. Sig- underlies our algebraic approach. In Sections 4 and 5, we show the nal models for time are visualized as directed graphs. We construct signal models for infinite and finite SP in time and space. Section 6 corresponding models for undirected graphs, which we hence call briefly discusses the extension to separable and nonseparable 2-D space models, and show that, in particular, the 16 DCTs and DSTs SP. Finally, section 7 concludes the paper. are Fourier transforms for these finite space models. Finally, we discuss the extension of our theory to separable and nonseparable 2-D SP. 2. BACKGROUND: ALGEBRAS AND MODULES In this section, we introduce two concepts from abstract algebra: 1. INTRODUCTION Algebra and module. These are needed to identify the algebraic structure in linear signal processing. Both are vector spaces, but Linear signal processing (henceforth simply called SP) is a well- possess additional structure. In Section 3, we will explain that in developed theory and a comprehensive set of tools for processing linear signal processing filter spaces are algebras and signal spaces 1-D time signals, continuous or discrete, of infinite or finite du- are modules. For every given pair of algebra and module, there is ration. Each of these four cases has specific notions of filtering, an associated notion of Fourier transform that we introduce. We Fourier transform, and other key concepts needed in SP. focus on regular modules (the algebra and the module are the same Many other transforms are used in SP, most prominently the set but with different properties) of polynomial algebras, which are class of trigonometric transforms that include the discrete cosine particularly relevant in signal processing, as we will explain later. and sine transforms (DCTs and DSTs). These compute a spec- Algebra (filter space). An algebra A is a vector space that trum of sorts, so conceptually they are Fourier transforms; but, is also a ring, i.e., it permits multiplication and the distributivity questions remain: for which type of signals? what are their infi- law holds. Examples include the field of complex numbers C, nite counterparts? and, just as for the DFT, what are the associ- the space of polynomials in one variable C[x], and the space of ated notions of filtering, convolution, and spectrum? The current Laurent series in x with absolute summable coefficient sequences. derivation of the DCTs based on random processes, [1], or their in- Another example, particularly important in SP, is introduced next. terpretation as Karhunen-Loeve` transforms of random fields, [2], Let p(x) ∈ C[x] be a fixed polynomial of degree deg(p) = n, do not answer these questions. then the set of polynomials In 2-D, SP is usually developed in a separable way, which as- sumes the signal resides on a rectangular lattice. For many ap- C[x]/p(x) = {q(x) ∈ C[x] | deg(q) < deg(p)} (1) plications, this is appropriate. Mersereau et al. [3] provide some of degree smaller than n with addition and multiplication modulo answers for more general lattices. We would like, for example, to p is called a polynomial algebra. In C[x]/p(x), p(x) = 0 holds. explore generic questions including what are the DCT equivalents As a vector space, C[x]/p(x) has dimension deg(p) = n. We on nonseparable lattices. write elements of A as h, the common letter used for filters in SP. In this paper, we give answers to these questions in an overview Module (signal space). Assume that an algebra A has been on recent work [4, 5]. Our approach is in two steps. First, we chosen. An A-module, M is a vector space that permits an oper- identify the algebraic structure inherent to SP that goes beyond ation of the A on M, written as “·” (multiplication), such that the vector spaces to algebras and modules. We argue that polynomial distributivity law holds (as well as some other properties that we algebras are key concepts in SP. Second, we formulate a general omit). Formally, for h ∈ A, s, s , s ∈M, framework for SP, based on the notion of a signal model that we 1 2 define as a triple of an algebra of filters, a module of signals, and h · s ∈M, h · (s1 + s2) = h · s1 + h · s2. a mapping Φ that generalizes the z-transform. Once a model is given, other concepts such as the Fourier transform are automati- For a given algebra, one choice of module is the regular mod- cally defined. We describe the signal models for infinite and finite ule M = A with the operation of A on M being the multiplica- time SP and visualize them as directed graphs. Then we identify tion in A. Almost all modules considered in this paper are regular. the models underlying the DCTs and DSTs and show that their We will focus on algebras and modules that are C-vector spaces in this paper, but the theory extends to other basefields. We denote This work was supported by NSF through award 0310941. signals in M as s. Fourier transform. Assume that an algebra A and an A- • / • / • / • / • / • module M are given. Then the associated Fourier transform (if x−2 x−1 x0 x1 x2 x3 it exists) is a linear mapping1 that decomposes M into a direct Fig. 1. Visualization of the infinite discrete time model (4). sum of smallest submodules invariant under A. We omit a formal definition, as we are only interested in one special case discussed next. as explained below. Once a signal model is chosen, filtering is C Let A = [x]/p(x) and let M = A be the regular mod- well-defined (the operation of A on M) and the main ingredients ule. Further, assume p(x) is of degree n and separable, i.e., it has for SP follow: spectrum, Fourier transform (see Section 2), fre- pairwise distinct zeros α = (α0,...,αn−1). Then, the associated quency response, as well as others [4]. Fourier transform is given by the Chinese remainder theorem as Example. Consider the space of finite energy sequences s ∈ V = ℓ2(Z). A signal model for V is given by ∆ : C[x]/p(x) → C[x]/(x − α0) ⊕ . ⊕ C[x]/(x − αn−1), s(x) 7→ (s(α0), . , s(αn−1)). ℓ 1 Z A = {Pℓ∈Z hℓx | (hℓ) ∈ ℓ ( )}, Since ∆ is a linear mapping, we can express it as a matrix w.r.t. cho- ℓ 2 Z (4) M = {Pℓ∈Z hℓx | (hℓ) ∈ ℓ ( )}, sen bases. Let b = (p0, . , pn−1) be a basis of M, and choose s ℓ 0 Φ = V →M, 7→ s = s(x) = Pℓ∈Z sℓx . (x ) as basis in each C[x]/(x − αk). Then ∆ corresponds to The mapping Φ is the ordinary z-transform3. For this reason we P = [p (α )] , (2) b,α ℓ k 0≤k,ℓ<n call this signal model the infinite discrete time model. Note that 4 which we also call a Fourier transform for the A-module M with the module is not regular, since A 6= M. 0 Shift. An important concept in our general framework is the basis b. Choosing arbitrary bases (akx ), ak ∈ C, in C[x]/(x − shift operator. We formally define the shift as the chosen generator αk) yields the most general form of Fourier transform for M (with basis b): of the algebra A. For example, the 1-D model (4) has only one shift x. With this definition, we can classify all models that lead F = diag ≤ (1/ak)Pb,α. (3) 0 k<n to shift-invariant SP as those with commutative algebras. If the algebra has only one shift, it is necessarily commutative as in the 3. THE ALGEBRAIC STRUCTURE IN SIGNAL case of (4). PROCESSING In the case of finite sequences, e.g., s ∈ Cn, M and A are finite-dimensional. Finite-dimensional commutative algebras are This section introduces the signal model that formalizes the con- precisely the polynomial algebras; in the case of one shift x they nection between algebras and modules and SP. have the form in (1). SP is algebraic. We identify the basic assumptions imposed In the remainder of the paper, we identify several signal mod- on the filter and signal space in SP. Clearly, the set of filters is els that occur in SP. Lacking space, we sketch only their derivation a vector space (scalar multiplication = amplification by α ∈ C; and focus on providing intuition about them. A detailed exposition addition = parallel connection) but is even an algebra A, since se- is found in [4, 5]. quential connection of filters, i.e., multiplication of filters, is per- missible. The set of signals is also a vector space in linear SP.