Introduction
Causal inference [33 , 42 ] is about modeling cause and effect relations in the presence of randomness. The fundamental concept in causal inference is that of a causal model. A causal model has an associated directed graph whose vertices represent the entities between which causal relations are to be modeled, and whose directed links specify the immediate cause and effect relations. In addition to the graph, a causal model contains a set of functions compatible with the directed links, and a random noise object for each vertex. If the model is well defined, it gives rise to the observational distribution, which is a probability distribution over a collection of random variables indexed by the vertices of the graph. The question of what information about the causal model can be recovered from the observational distribution drives a major stream of research in causal inference, including the present paper.
Time is intrinsic to causal relations in that an effect cannot precede its cause, i.e., cause and effect are ordered by time. When defining a causal model which obeys the ordering of time, a discrete set of time points is often assumed. Structural equation processes [35 , 36 ] provide a broad class of linear causal models for discrete-time processes. The qualitative causal structure of an SEP is encoded by a finite directed graph. Every vertex on that graph has an independent stationary noise process, and every directed link has an associated infinite sequence, which quantifies how the past of the causing process linearly influences the present state of the caused process. The structural equation process defines a linear equation system, which is uniquely solved by the (stationary) observational process of the SEP.
In an SEP, the influence of one process on another is time invariant, i.e., the sequence describing how the latter process at a given time point is linearly influenced by the past of the former does not depend on the time point. In particular, stationary SEPs are not capable of modeling time-varying interactions. In the current paper, we extend the class of structural equation processes to a subclass of non-stationary discrete-time processes, which we call P P P -stationary, and which allow to model periodically varying interactions between processes. For the purpose of defining P P P -stationarity, we declare a subsampling to be a subset S ⊂ Z S \subset\mathbb{Z} S ⊂ Z , and we say that it has periodicity P P P if t ∈ S t \in S t ∈ S implies that also t ± P ∈ S t \pm P \in S t ± P ∈ S . Note that there are exactly P P P -many subsamplings with periodicity P P P . In this work, a process X X X is called P P P -stationary if the subprocesses X S = ( X ( t ) : t ∈ S ) X_S = (X(t) : t \in S) X S = ( X ( t ) : t ∈ S ) , where S S S indexes the P P P -periodic subsamplings, are jointly stationary. To give an example, imagine we measure the temperature at some fixed location twice a year, once during summer and once during winter. The resulting temperature time series will turn out non-stationary if at that location winter and summer are significantly different from each other. Nonetheless, the time series consisting only of the winter and summer measurements, respectively, might still be stationary, as we would expect summer and winter, respectively, to be similar from year to year. If, in addition, the dependence between winter and summer does not change over time, the entire time series, which includes both winter and summer values, will be 2-stationary.
In this paper, we define a class of linear causal models for P P P -stationary processes whose components are indexed by some finite index set V V V . To stay in line with previous terminology [35 , 34 , 36 ] , we term them P P P -stationary SEPs. A P P P -stationary SEP consists of a graph G = ( V × { 0 , … , P − 1 } , D ) G = (V \times\{0, \ldots, P - 1\}, D) G = ( V × { 0 , … , P − 1 } , D ) , a link filter for every link on G G G , and a noise process for every vertex on G G G . A vertex on G G G represents a P P P -periodic subsampling of some component process v v v . Consequently, a link between two vertices expresses a linear influence of one P P P -periodic subprocess onto another P P P -periodic subprocess.
Our main motivation to introduce P P P -stationary SEPs is to provide a graphical model for handling temporally incomplete observations. For the biyearly temperature measurements, the subprocess consisting only of the summer or winter values is a simple example of temporally incomplete observation. However, our notion of temporal incompleteness allows for much more general observational scenarios.
The underlying observation to our approach is that any temporally incomplete observation of an P P P -stationary SEP can be represented, for an appropriate integer R R R , as a finite collection of R R R -periodic subsamplings. Then we will see that every P P P -stationary SEP has an equivalent R P RP R P -stationary SEP, which is equipped with a latent structure rendering every missing subsample as latent. In this setting, we analyze the problem of identifying causal effects, i.e., link filters, in P P P -stationary SEPs from temporally incomplete observations. Following previous work [34 , 36 ] , we will study the problem in the frequency domain, where the algebraic relations are more tractable. The identification of a given effect in the P P P -stationary SEP using a set of instrumental processes requires certain entries in the spectral density. If the observations are temporally incomplete, one faces the problem that the spectral density can generally not be computed, so identification via instrumental processes seems infeasible. However, by providing a linear recovery formula (Proposition 12), we can characterize (Definition 13) the temporally incomplete observational settings for which it is possible to recover a given entry in the spectral density (using that linear formula). Accordingly, in a setting of temporally incomplete observations, we detect a given effect to be identifiable if there is a suitable set of instrumental processes while the temporal incompleteness of the observations is such that all the entries in the spectral density that are required for identification via that set of instrumental processes can be recovered (Theorem 18 ). For technical reasons, we restrict our analysis to P P P -stationary structural vector autoregressive processes, i.e., SVAR processes whose coefficients are allowed to vary periodically over time with periodicity P P P .
To exemplify the problem and our way of approaching it, consider the simple setup of three stationary SVAR processes, whose causal relations are given by u → v → w u \to v \to w u → v → w . Suppose we are interested in the effect associated with the link v → w v \to w v → w . If we had temporally complete observations for v v v and w w w , we could identify this effect from the spectral density via linear regression in the frequency domain. However, this approach cannot be pursued if the observations for v v v and w w w are temporally incomplete, since in this case the spectral density cannot be computed. Assuming temporally complete observations for the instrumental process u u u , it will follow from the theory we develop in this paper that the spectral dependence between u u u and v v v , and between u u u and w w w can be recovered, so that the effect v → w v \to w v → w can be identified by the usual instrumental process formula for the frequency domain. A more interesting case, discussed in Example 9 and pictured in Figure 7 , demonstrates that our findings allow to identify effects even if none of the involved processes is fully observable.
Figure 7. a.) The summary graph for a 2-stationary SVAR process. A vertex a k a_k a k represents the process defl 2 ∘ sh k ( X a ) \operatorname{defl}_2 \circ\operatorname{sh}^k(X_a) defl 2 ∘ sh k ( X a ) , and a bidirectional link represents two directed links, i.e., one in each direction. b.) The observational structure, where a vertex of the form a k , j a_{k,j} a k , j represents the subprocess defl 4 ∘ sh 2 k + j ( X a ) \operatorname{defl}_4 \circ\operatorname{sh}^{2k+j}(X_a) defl 4 ∘ sh 2 k + j ( X a ) . The fully and palely colored nodes represent the observed and unobserved subprocesses, respectively. The process l l l is completely unobserved.
The rest of the paper is organized as follows. In § 2 , we recall some basic terminology for linear causal models of discrete-time stationary processes. In § 3 , we define P P P -stationary SEPs and SVAR processes and show how an P P P -stationary SVAR process can equivalently by described as an P P P -stationary SEP. Furthermore, we define the embedding operator that identifies for every P P P -stationary SEP an equivalent R P RP R P -stationary SEP. In Section 4, we present the algebraic structure of P P P -stationary SEPs in the frequency domain. The main result of this section will be used to determine if a given temporally incomplete observational pattern allows to linearly recover a specified entry in the spectral density. In § 4 , we formalize the problem of identifying causal effects in P P P -stationary SEPs from temporally incomplete data in the frequency domain, i.e., we will make precise what we mean by a temporally incomplete observations and we will provide a notion of rational identifiability for P P P -stationary SVAR processes (under a given temporally incomplete observational pattern). Using those terms, we will formulate the identifiability result.
Related work
We approach causal inference in the framework of Structural Causal Models presented and developed in, e.g., [33 , 28 , 6 , 42 ] . Since we focus on linear models, much of the work and theory developed here is inspired and influenced by the theory and terminology used in the literature on linear Structural Equation Modeling , see, e.g., [4 , 41 , 46 , 47 ] , and especially those contributions that take an algebraic perspective on the matter, see, e.g., [12 , 13 , 19 , 11 , 44 ] .
Our paper is particularly related to those contributions which incorporate time as a discrete index into the modeling of causal relations. Early contributions include those related to Granger Causality , see [22 , 23 , 25 , 26 , 27 ] , which underwent a reformulation in terms of concepts and terminology from graphical modeling, see [8 , 9 , 14 , 18 , 15 , 17 , 16 ] . This line of work has then been further developed, for instance, in [40 , 39 , 38 , 21 , 30 ] .
The theory presented in the present paper builds on [35 , 37 , 34 ] , which develop the notion of (stationary) structural equation processes to quantify process-level causal relations (in the frequency domain). An algebraic approach to structure and effect identification from observational data, in the spirit of the contributions mentioned above, was developed in [36 ] .
A specific from of temporally incomplete observations are undersampled time series, i.e., when observations are gathered less frequently than causal effects happen. An introduction to the problems related to inferring causal relations from undersampled time series can be found in [10 ] . Specific approaches to inferring the causal structure from an undersampled time series model have been developed in, e.g., [31 , 29 ] . The special case where the underlying time series model is linear is investigated, for example, in [24 , 45 ] . Specifically, they consider SVAR processes of order one (causal effects happen either instantaneously or with a time lag of one time unit) generated by a non-Gaussian white noise process. In [45 ] , the observations are allowed to be gathered at two different sampling rates, which is also known as the mixed-frequency case. Another approach to the identification of VAR models (without instantaneous effects) is presented in [2 , 1 ] , where the authors consider stable VAR processes of some arbitrary finite order p p p , and which are assumed to be sampled at two different frequencies. The fast frequency is assumed to be the actual time scale of the process, i.e., the frequency with which causal effects happen. Under the assumption that at least one process is observed at the actual time scale, they show that for almost every possible VAR parameter configuration, that parameter can be uniquely recovered from the second moments of the observed process (which is generally temporally incomplete).
In this work, we consider linear time series models whose coefficients are allowed to vary periodically over time. Such models generally yield non-stationary processes. What we call an P P P -stationary SVAR process in this work is a special instance of a time varying VAR process, see, e.g., [43 , 7 ] . A recent contribution to non-stationary undirected graphical modeling can be found in [3 ] . Our approach to weaken the stationarity assumption in a discrete-time setting is similar to the one taken in [20 ] , which develops a causal discovery algorithm for so-called semi-stationary Markov chains with an underlying periodically varying causal structure.
Basic notation
Let V V V be some finite set and K \mathbb{K} K a field, then K V \mathbb{K}^{V} K V denotes the K \mathbb{K} K -vector space of dimension ∣ V ∣ |V| ∣ V ∣ and whose coordinates are indexed by V V V . Let U U U be another finite set, then K V × U \mathbb{K}^{V \times U} K V × U denotes the space of matrices whose rows are indexed by V V V and whose columns are indexed by U U U . In case V = U V = U V = U , we write I V I_V I V for the identity matrix. For a matrix A = [ A v , u ] v ∈ V , u ∈ U ∈ K V × U A = [A_{v,u}]_{v\in V,u\in U} \in\mathbb{K}^{V\times U} A = [ A v , u ] v ∈ V , u ∈ U ∈ K V × U , and subsets V ′ ⊂ V V' \subset V V ′ ⊂ V and U ′ ⊂ U U' \subset U U ′ ⊂ U , we define the submatrix of A A A corresponding to the entries V ′ × U ′ V' \times U' V ′ × U ′ as [ A ] V ′ , U ′ : = [ A v , u ] v ∈ V ′ , u ∈ U ′ ∈ K V ′ × U ′ [A]_{V',U'} := [A_{v,u}]_{v\in V',u\in U'} \in\mathbb{K}^{V'\times U'} [ A ] V ′ , U ′ := [ A v , u ] v ∈ V ′ , u ∈ U ′ ∈ K V ′ × U ′ . The transpose of A A A is denoted A ⊤ : = [ A v , u ] u ∈ U , v ∈ V ∈ K U × V A^\top:= [A_{v,u}]_{u\in U,v\in V} \in\mathbb{K}^{U\times V} A ⊤ := [ A v , u ] u ∈ U , v ∈ V ∈ K U × V . Since we will often encounter block-matrices, it will be useful to have a dedicated notation for them. Let U , V , I , J U, V, I, J U , V , I , J be finite sets, then define BMat ( K , ( U , V ) , ( I , J ) ) \operatorname{BMat}(\mathbb{K},(U,V),(I,J)) BMat ( K , ( U , V ) , ( I , J )) to be the set of matrices A A A which have the following two level block structure:
A = [ A u , v ] u ∈ U , v ∈ V , A u , v = [ A u , v ( i , j ) ] i ∈ I , j ∈ J ∈ K I × J . A = [A_{u,v}]_{u\in U,v\in V}, \qquad A_{u,v} = [A_{u,v}^{(i,j)}]_{i\in I,j\in J} \in\mathbb{K}^{I\times J}. A = [ A u , v ] u ∈ U , v ∈ V , A u , v = [ A u , v ( i , j ) ] i ∈ I , j ∈ J ∈ K I × J .
If U = V U = V U = V or I = J I = J I = J , we will abbreviate the corresponding set of block-matrices by BMat ( K , V , ( I , J ) ) \operatorname{BMat}(\mathbb{K},V,(I,J)) BMat ( K , V , ( I , J )) and BMat ( K , ( U , V ) , I ) \operatorname{BMat}(\mathbb{K},(U,V),I) BMat ( K , ( U , V ) , I ) respectively, and if both U = V U = V U = V and I = J I = J I = J , then we abbreviate the corresponding set of block-matrices by BMat ( K , U , I ) \operatorname{BMat}(\mathbb{K},U,I) BMat ( K , U , I ) .
For two integers p , q ∈ Z p,q \in\mathbb{Z} p , q ∈ Z , define [ p , q ] : = { i ∣ p ≤ i ≤ q } [p,q] := \{i \mid p \le i \le q\} [ p , q ] := { i ∣ p ≤ i ≤ q } , and for n ≥ 1 n \ge1 n ≥ 1 , we write [ n ] [n] [ n ] to abbreviate [ 1 , n ] [1,n] [ 1 , n ] . For a positive integer P ≥ 1 P \ge1 P ≥ 1 , we define the set of P P P -multiples to be P Z : = { k P ∣ k ∈ Z } P\mathbb{Z} := \{kP \mid k \in\mathbb{Z}\} P Z := { k P ∣ k ∈ Z } , and for m ∈ [ 0 , P − 1 ] m \in[0,P-1] m ∈ [ 0 , P − 1 ] , we define P Z − m : = { k P − m ∣ k ∈ Z } P\mathbb{Z}-m := \{kP-m \mid k \in\mathbb{Z}\} P Z − m := { k P − m ∣ k ∈ Z } .
For K \mathbb{K} K either R \mathbb{R} R or C \mathbb{C} C , the space ℓ p ( Z , K ) \ell^p(\mathbb{Z},\mathbb{K}) ℓ p ( Z , K ) consists of all sequences λ : Z → K \lambda:\mathbb{Z}\to\mathbb{K} λ : Z → K such that ∥ λ ∥ p p : = ∑ k ∈ Z ∣ λ ( k ) ∣ p < ∞ \|\lambda\|_p^p := \sum_{k\in\mathbb{Z}}|\lambda(k)|^p < \infty ∥ λ ∥ p p := ∑ k ∈ Z ∣ λ ( k ) ∣ p < ∞ . We use ℓ p ( Z , K V × U ) \ell^p(\mathbb{Z},\mathbb{K}^{V\times U}) ℓ p ( Z , K V × U ) to denote the space of sequences Λ : Z → K V × U \Lambda:\mathbb{Z}\to\mathbb{K}^{V\times U} Λ : Z → K V × U such that [ Λ ] v , u : = ( k ∈ Z ↦ [ Λ ( k ) ] v , u ) ∈ ℓ p ( Z , K ) [\Lambda]_{v,u}:=(k\in\mathbb{Z}\mapsto[\Lambda(k)]_{v,u})\in\ell^p(\mathbb{Z},\mathbb{K}) [ Λ ] v , u := ( k ∈ Z ↦ [ Λ ( k ) ] v , u ) ∈ ℓ p ( Z , K ) for all ( v , u ) ∈ V × U (v,u)\in V\times U ( v , u ) ∈ V × U . In this work, we will call any element in Λ ∈ ℓ 1 ( Z , K V × U ) \Lambda\in\ell^1(\mathbb{Z},\mathbb{K}^{V\times U}) Λ ∈ ℓ 1 ( Z , K V × U ) such that Λ ( k ) = 0 \Lambda(k)=0 Λ ( k ) = 0 if k < 0 k<0 k < 0 , a filter , a term often used in the time series literature. For two sequences λ , μ ∈ ℓ p ( Z , K ) \lambda,\mu\in\ell^p(\mathbb{Z},\mathbb{K}) λ , μ ∈ ℓ p ( Z , K ) , the convolution resp. tilted convolution are defined as follows:
λ ∗ μ : = ( k ↦ ∑ l ∈ Z λ ( l ) μ ( k − l ) ) , λ ∗ ′ μ : = ( k ↦ ∑ l ∈ Z λ ( k + l ) μ ( l ) ) . \lambda* \mu:= \left(k\mapsto\sum_{l\in\mathbb{Z}}\lambda(l)\mu(k-l)\right), \qquad\lambda*' \mu:= \left(k\mapsto\sum_{l\in\mathbb{Z}}\lambda(k+l)\mu(l)\right). λ ∗ μ := ( k ↦ l ∈ Z ∑ λ ( l ) μ ( k − l ) ) , λ ∗ ′ μ := ( k ↦ l ∈ Z ∑ λ ( k + l ) μ ( l ) ) .
The transpose of a sequence Λ \Lambda Λ is the element-wise transpose, Λ ⊤ : = ( k ↦ Λ ( k ) ⊤ ) \Lambda^\top:= (k\mapsto\Lambda(k)^\top) Λ ⊤ := ( k ↦ Λ ( k ) ⊤ ) . If V = U V=U V = U , then I V I_V I V denotes the identity filter , i.e., I V ( 0 ) = I V I_V(0)=I_V I V ( 0 ) = I V , and I V ( k ) = 0 I_V(k)=0 I V ( k ) = 0 for all k ≠ 0 k\ne0 k = 0 . Furthermore, for a fixed set S S S and a subset S ′ ⊂ S S'\subset S S ′ ⊂ S , we write 1 S ′ : S → { 0 , 1 } \mathbf{1}_{S'}:S\to\{0,1\} 1 S ′ : S → { 0 , 1 } for the indicator function , i.e., 1 S ′ ( x ) = 1 \mathbf{1}_{S'}(x)=1 1 S ′ ( x ) = 1 if and only if x ∈ S ′ x\in S' x ∈ S ′ .
Preliminaries
In this section, we review the basic terminology for directed graphs, discrete-time stationary processes, and linear causal models of processes, i.e., structural equation processes [35 , 34 , 36 ] .
Directed graphs
A directed graph G = ( V , D ) G=(V,D) G = ( V , D ) consists of a countable set (finite or infinite) of vertices V V V and a set of directed links D ⊂ V × V ∖ { ( v , v ) ∣ v ∈ V } D\subset V\times V\setminus\{(v,v)\mid v\in V\} D ⊂ V × V ∖ {( v , v ) ∣ v ∈ V } . To express that ( v , w ) ∈ D (v,w)\in D ( v , w ) ∈ D is a directed link, we write v → w v\to w v → w for short, and v ↛ w v\not\to w v → w otherwise. Note that we exclude self-links, i.e., links of the form v → v v\to v v → v . The parent set of v ∈ V v\in V v ∈ V is defined to be pa ( v ) : = { u ∈ V ∣ u → v } \operatorname{pa}(v):=\{u\in V\mid u\to v\} pa ( v ) := { u ∈ V ∣ u → v } . A path on G G G is a finite sequence π = ( v i → w i ) i ∈ [ n ] \pi=(v_i\to w_i)_{i\in[n]} π = ( v i → w i ) i ∈ [ n ] of links such that either w i = v i + 1 w_i=v_{i+1} w i = v i + 1 or w i = w i + 1 w_i=w_{i+1} w i = w i + 1 or v i = v i + 1 v_i=v_{i+1} v i = v i + 1 for all $1\le i<n$. A path π \pi π is called directed if w i = v i + 1 w_i=v_{i+1} w i = v i + 1 for all 1 ≤ i < n 1\le i<n 1 ≤ i < n . By convention, we also consider empty paths, i.e., for every vertex v ∈ V v\in V v ∈ V , we denote by ϵ v \epsilon_v ϵ v the path from v v v to itself that does not contain any link. For two vertices v , w ∈ V v,w\in V v , w ∈ V , we will denote the set of all directed paths from v v v to w w w by P G ( v , w ) P_G(v,w) P G ( v , w ) . A graph G G G is called acyclic if P G ( v , v ) = { ϵ v } P_G(v,v)=\{\epsilon_v\} P G ( v , v ) = { ϵ v } for every $v\in V. The set of ancestors and descendants , respectively, of a vertex v ∈ V v\in V v ∈ V are defined as follows:
anc G ( v ) : = { u ∈ V ∣ P G ( u , v ) ≠ ∅ } des G ( v ) : = { w ∈ V ∣ P G ( v , w ) ≠ ∅ } . \operatorname{anc}_G(v):=\{u\in V\mid P_G(u,v)\ne\varnothing\}
\qquad
\operatorname{des}_G(v):=\{w\in V\mid P_G(v,w)\ne\varnothing\}. anc G ( v ) := { u ∈ V ∣ P G ( u , v ) = ∅ } des G ( v ) := { w ∈ V ∣ P G ( v , w ) = ∅ } .
The strongly connected component of a vertex v ∈ V v\in V v ∈ V is the intersection anc G ( v ) ∩ des G ( v ) \operatorname{anc}_G(v)\cap\operatorname{des}_G(v) anc G ( v ) ∩ des G ( v ) . Note that the vertex set of any directed graph can be partitioned into its strongly connected components.
Given a subset U ⊂ V U\subset V U ⊂ V , we write pa G ( U ) : = ⋃ u ∈ U pa G ( u ) \operatorname{pa}_G(U):=\bigcup_{u\in U}\operatorname{pa}_G(u) pa G ( U ) := ⋃ u ∈ U pa G ( u ) to refer to the parents of the vertices in U U U , and analogously for the ancestors anc G ( U ) \operatorname{anc}_G(U) anc G ( U ) and the descendants des G ( U ) \operatorname{des}_G(U) des G ( U ) .
A trek between two vertices v , w ∈ V v,w\in V v , w ∈ V on G G G is a tuple τ = ( τ L , top ( τ ) , τ R ) \tau=(\tau_L,\operatorname{top}(\tau),\tau_R) τ = ( τ L , top ( τ ) , τ R ) , where top ( τ ) ∈ V \operatorname{top}(\tau)\in V top ( τ ) ∈ V and τ L ∈ P G ( top ( τ ) , v ) \tau_L\in P_G(\operatorname{top}(\tau),v) τ L ∈ P G ( top ( τ ) , v ) and τ R ∈ P G ( top ( τ ) , w ) \tau_R\in P_G(\operatorname{top}(\tau),w) τ R ∈ P G ( top ( τ ) , w ) . The set of treks between two vertices v , w ∈ V v,w\in V v , w ∈ V on G G G is denoted by T G ( v , w ) T_G(v,w) T G ( v , w ) .
For a subset U ⊂ V U\subset V U ⊂ V , the restriction of G G G to U U U is the directed graph G U = ( U , D U ) G_U=(U,D_U) G U = ( U , D U ) , where D U = D ∩ U × U D_U=D\cap U\times U D U = D ∩ U × U . Furthermore, suppose ( G i = ( V i , D i ) ) i ∈ I (G_i=(V_i,D_i))_{i\in I} ( G i = ( V i , D i ) ) i ∈ I is a discrete collection of directed graphs, then ⋃ i ∈ I G i = ( ⋃ i ∈ I V i , ⋃ i ∈ I D i ) \bigcup_{i\in I}G_i=\left(\bigcup_{i\in I}V_i,\bigcup_{i\in I}D_i\right) ⋃ i ∈ I G i = ( ⋃ i ∈ I V i , ⋃ i ∈ I D i ) is the union graph and, similarly, ⋂ i ∈ I G i = ( ⋂ i ∈ I V i , ⋂ i ∈ I D i ) \bigcap_{i\in I}G_i=\left(\bigcap_{i\in I}V_i,\bigcap_{i\in I}D_i\right) ⋂ i ∈ I G i = ( ⋂ i ∈ I V i , ⋂ i ∈ I D i ) is the intersection graph.
Given a finite set V V V , a directed graph G = ( V × Z , D ) G=(V\times\mathbb{Z},D) G = ( V × Z , D ) is called a (directed) time series graph if for any link v ( s ) → w ( t ) v(s)\to w(t) v ( s ) → w ( t ) , it holds that s ≤ t s\leq t s ≤ t , i.e., links do not point backwards in time. A time series graph G G G is said to be stationary if v ( t − k ) → w ( t ) v(t-k)\to w(t) v ( t − k ) → w ( t ) , for some t ∈ Z t\in\mathbb{Z} t ∈ Z , implies v ( s − k ) → w ( s ) v(s-k)\to w(s) v ( s − k ) → w ( s ) for all s ∈ Z s\in\mathbb{Z} s ∈ Z . Given a stationary time series graph G G G , we write v → k w v\to_k w v → k w to abbreviate that v ( t − k ) → w ( t ) v(t-k)\to w(t) v ( t − k ) → w ( t ) for every t ∈ Z t\in\mathbb{Z} t ∈ Z . A stationary time series graph G = ( V × Z , D ) G=(V\times\mathbb{Z},D) G = ( V × Z , D ) can be reduced to a finite graph G = ( V , D ) G=(V,D) G = ( V , D ) that contains the directed link v → w v\to w v → w , where v ≠ w v\neq w v = w , if and only if and there is some k ≥ 0 k\geq0 k ≥ 0 such that v → k w v\to_k w v → k w holds on G G G . Following [35 , 34 , 36 ] , we call this finite reduction the process graph of G G G .
Stationary discrete-time processes
Let V V V be a finite set and let X = ( X v ( t ) ∣ v ∈ V , t ∈ Z ) X=(X_v(t)\mid v\in V,t\in\mathbb{Z}) X = ( X v ( t ) ∣ v ∈ V , t ∈ Z ) be a V V V -variate discrete-time stochastic process. In this work, we use the word process to refer to a discrete-time stochastic process. For subsets U ⊆ V U\subseteq V U ⊆ V and T ⊆ Z T\subseteq\mathbb{Z} T ⊆ Z , we define the associated subprocess X U ( T ) : = ( X u ( t ) ∣ u ∈ U , t ∈ T ) X_U(T):=(X_u(t)\mid u\in U,t\in T) X U ( T ) := ( X u ( t ) ∣ u ∈ U , t ∈ T ) . For a subset t ⊂ Z t\subset\mathbb{Z} t ⊂ Z and some s ∈ Z s\in\mathbb{Z} s ∈ Z , we define t − s : = { t − s ∣ t ∈ t } t-s:=\{t-s\mid t\in t\} t − s := { t − s ∣ t ∈ t } . A process X X X is called stationary if for every finite t ⊂ Z t\subset\mathbb{Z} t ⊂ Z and every s ∈ Z s\in\mathbb{Z} s ∈ Z , the random vectors X ( t ) X(t) X ( t ) and X ( t − s ) X(t-s) X ( t − s ) follow the same distribution. Otherwise, the process X X X is called non-stationary . For pairwise disjoint subsets U 1 , … , U n ⊂ V U_1,\ldots,U_n\subset V U 1 , … , U n ⊂ V , the associated subprocesses X U 1 , … , X U n X_{U_1},\ldots,X_{U_n} X U 1 , … , X U n are called mutually independent if for any finite t ⊂ Z t\subset\mathbb{Z} t ⊂ Z the random vectors X U 1 ( t ) , … , X U n ( t ) X_{U_1}(t),\ldots,X_{U_n}(t) X U 1 ( t ) , … , X U n ( t ) are mutually independent. The process X X X is called a white noise process if for any two finite subsets t , s ⊂ Z t,s\subset\mathbb{Z} t , s ⊂ Z such that t ∩ s = ∅ t\cap s=\varnothing t ∩ s = ∅ , the random vectors X ( t ) X(t) X ( t ) and X ( s ) X(s) X ( s ) are independent. The mean of a stationary process X X X is μ : = E [ X ( t ) ] \mu:=\mathbb{E}[X(t)] μ := E [ X ( t )] and the autocovariance sequence (ACS) is given by ACS ( X ) : = ( k ↦ E [ X ( t ) X ( t − k ) ⊤ ] ) \operatorname{ACS}(X):=(k\mapsto\mathbb{E}[X(t)X(t-k)^\top]) ACS ( X ) := ( k ↦ E [ X ( t ) X ( t − k ) ⊤ ]) .
Structural Equation Processes
A structural equation process (SEP) S = ( G , Λ , X I ) S=(G,\Lambda,X^{\mathrm I}) S = ( G , Λ , X I ) consists of: a finite directed graph G = ( V , D ) G=(V,D) G = ( V , D ) ; a matrix-valued filter Λ ∈ ℓ 1 ( Z , R V × V ) \Lambda\in\ell^1(\mathbb{Z},\mathbb{R}^{V\times V}) Λ ∈ ℓ 1 ( Z , R V × V ) such that v ↛ w v\not\to w v → w implies that Λ v , w \Lambda_{v,w} Λ v , w is the zero-filter; and a V V V -variate mutually independent process X I X^{\mathrm I} X I . A solution to S S S is a process X X X such that it satisfies X = Λ ⊤ ∗ X + X I X=\Lambda^\top*X+X^{\mathrm I} X = Λ ⊤ ∗ X + X I . An SEP S S S has a (unique) solution if the filter Λ \Lambda Λ is such that S ( Λ ) : = ∑ k ≥ 0 Λ k S(\Lambda):=\sum_{k\geq0}\Lambda^k S ( Λ ) := ∑ k ≥ 0 Λ k exists and is absolutely summable. In this case, the unique solution is given by the process S ( Λ ⊤ ) ∗ X I S(\Lambda^\top)*X^{\mathrm I} S ( Λ ⊤ ) ∗ X I .
One way to construct an SEP is to start with an underlying structural vector autoregressive (SVAR) process, which consists of a stationary time series graph G = ( V × Z , D ) G=(V\times\mathbb{Z},D) G = ( V × Z , D ) ; a filter Φ : Z → R V × V \Phi:\mathbb{Z}\to\mathbb{R}^{V\times V} Φ : Z → R V × V such that v ̸ → k w v\not\to_k w v → k w implies Φ v , w ( k ) = 0 \Phi_{v,w}(k)=0 Φ v , w ( k ) = 0 ; and a V V V -variate mutually independent white noise process η \eta η . A stationary process X X X is said to be an SVAR process compatible with ( G , Φ , η ) (G,\Phi,\eta) ( G , Φ , η ) if it satisfies X = Φ ∗ X + η X=\Phi*X+\eta X = Φ ∗ X + η . Such a process X X X exists uniquely (up to a zero-set) if the coefficients Φ \Phi Φ are stable , see [32 , 36 ] . Additionally, every stable (in the sense of [36 ] ) SVAR process can be represented as an SEP S = ( G , Λ , X I ) S=(G,\Lambda,X^{\mathrm I}) S = ( G , Λ , X I ) , where G G G is the process graph of G G G , see [35 , 34 , 36 ] , meaning that the SVAR- and SEP solution coincide.
A structural equation model for P P P -stationary processes
We begin this section by introducing an embedding operator for sequences and processes. This operator will be central to the development of the main results of this paper, since it will allow us to handle temporally incomplete observations. Using the embedding operator, we introduce P P P -stationary processes and filters to define P P P -stationary structural equation processes as a general class of linear causal models for P P P -stationary processes. In the final part of this section, we will see that this class encompasses structural vector autoregressive processes whose coefficients can vary P P P -periodically over time. This observation will be important, since in developing the main results (stated for such P P P -stationary SVAR processes) we will use the structure of an P P P -stationary SEP as well as the additional structure of an P P P -stationary SVAR processes.
An embedding operator
To prepare the construction of the embedding operator, we introduce the following three operators on the space of filters. Let λ : Z → K V × U \lambda:\mathbb{Z}\to\mathbb{K}^{V\times U} λ : Z → K V × U be a sequence. For j ∈ Z j\in\mathbb{Z} j ∈ Z , the j j j -shifted sequence is defined as sh j ( λ ) : = ( k ↦ λ ( k − j ) ) \operatorname{sh}^{j}(\lambda):=(k\mapsto\lambda(k-j)) sh j ( λ ) := ( k ↦ λ ( k − j )) . For a positive integer N N N , we have the N N N -deflated sequence defl N ( λ ) : = ( k ↦ λ ( k N ) ) \operatorname{defl}_{N}(\lambda) := (k \mapsto\lambda(kN)) defl N ( λ ) := ( k ↦ λ ( k N )) and the N N N -inflated sequence infl N ( λ ) : = ( k ↦ 1 Z ( k ) λ ( 1 Z ( k ) k N − 1 ) ) \operatorname{infl}_{N}(\lambda) := (k \mapsto1_{\mathbb{Z}}(k)\lambda(1_{\mathbb{Z}}(k)kN^{-1})) infl N ( λ ) := ( k ↦ 1 Z ( k ) λ ( 1 Z ( k ) k N − 1 )) . Based on these operators, we define for a positive integer N N N two linear embedding operators
e N , e N ′ : ℓ 2 ( C U × V ) → ℓ 2 ( BMat ( C , [ 0 , N − 1 ] , ( U , V ) ) ) . e_{N},e'_{N}:\ell^{2}(\mathbb{C}^{U\times V})\to\ell^{2}(\operatorname{BMat}(\mathbb{C},[0,N-1],(U,V))). e N , e N ′ : ℓ 2 ( C U × V ) → ℓ 2 ( BMat ( C , [ 0 , N − 1 ] , ( U , V ))) .
Those operators map a sequence λ ∈ ℓ 2 ( C U × V ) \lambda\in\ell^{2}(\mathbb{C}^{U\times V}) λ ∈ ℓ 2 ( C U × V ) to the following sequences:
e N ( λ ) : = [ defl N ∘ sh i − j ( λ ) ] i , j ∈ [ 0 , N − 1 ] , e N ′ ( λ ) : = [ defl N ∘ sh j − i ( λ ) ] i , j ∈ [ 0 , N − 1 ] . e_{N}(\lambda):=\left[\operatorname{defl}_{N}\circ\operatorname{sh}^{i-j}(\lambda)\right]_{i,j\in[0,N-1]},\qquad e'_{N}(\lambda):=\left[\operatorname{defl}_{N}\circ\operatorname{sh}^{j-i}(\lambda)\right]_{i,j\in[0,N-1]}. e N ( λ ) := [ defl N ∘ sh i − j ( λ ) ] i , j ∈ [ 0 , N − 1 ] , e N ′ ( λ ) := [ defl N ∘ sh j − i ( λ ) ] i , j ∈ [ 0 , N − 1 ] .
Analogously, we define for a V V V -variate process X \mathsf{X} X , the embedded process e N ( X ) : = [ defl N ∘ sh i ( X ) ] i ∈ [ 0 , N − 1 ] e_{N}(\mathsf{X}):=\left[\operatorname{defl}_{N}\circ\operatorname{sh}^{i}(\mathsf{X})\right]_{i\in[0,N-1]} e N ( X ) := [ defl N ∘ sh i ( X ) ] i ∈ [ 0 , N − 1 ] , which is a process that consists of N ∣ V ∣ N|V| N ∣ V ∣ component processes. In view of the theory we develop in this work, the following two lemmas will be important.
Lemma 1. Let λ ∈ ℓ 2 ( C U × V ) \lambda\in\ell^{2}(\mathbb{C}^{U\times V}) λ ∈ ℓ 2 ( C U × V ) , then for every i ∈ [ 0 , N − 1 ] i\in[0,N-1] i ∈ [ 0 , N − 1 ] it holds that
sh i ( λ ) = ∑ j = 0 N − 1 sh j ∘ infl N ( defl N ∘ sh i − j ( λ ) ) . (1) \operatorname{sh}^{i}(\lambda)=\sum_{j=0}^{N-1}\operatorname{sh}^{j}\circ\operatorname{infl}_{N}\left(\operatorname{defl}_{N}\circ\operatorname{sh}^{i-j}(\lambda)\right). \tag*{(1)} sh i ( λ ) = j = 0 ∑ N − 1 sh j ∘ infl N ( defl N ∘ sh i − j ( λ ) ) . (1)
Lemma 2. Let λ ∈ ℓ 2 ( C W × V ) \lambda\in\ell^{2}(\mathbb{C}^{W\times V}) λ ∈ ℓ 2 ( C W × V ) and μ ∈ ℓ 2 ( C V × U ) \mu\in\ell^{2}(\mathbb{C}^{V\times U}) μ ∈ ℓ 2 ( C V × U ) , then it holds that
e N ( λ ∗ μ ) = e N ( λ ) ∗ e N ( μ ) , e N ( λ ∗ ′ μ ) = e N ( λ ) ∗ ′ e N ′ ( μ ) . e_{N}(\lambda*\mu)=e_{N}(\lambda)*e_{N}(\mu),\qquad e_{N}(\lambda*'\mu)=e_{N}(\lambda)*'e'_{N}(\mu). e N ( λ ∗ μ ) = e N ( λ ) ∗ e N ( μ ) , e N ( λ ∗ ′ μ ) = e N ( λ ) ∗ ′ e N ′ ( μ ) . Furthermore, for a V V V -variate process X \mathsf{X} X , it holds that e N ( λ ∗ X ) = e N ( λ ) ∗ e N ( X ) e_{N}(\lambda*\mathsf{X})=e_{N}(\lambda)*e_{N}(\mathsf{X}) e N ( λ ∗ X ) = e N ( λ ) ∗ e N ( X ) .
P-stationary processes
Definition 1 (P-stationary process). Let P P P be a positive integer. A process X = ( X v ∣ v ∈ V ) \mathsf{X}=(\mathsf{X}_{v}\mid v\in V) X = ( X v ∣ v ∈ V ) is called P P P -stationary if the stacked process ( e P ( X v ) ∣ v ∈ V ) (e_{P}(\mathsf{X}_{v})\mid v\in V) ( e P ( X v ) ∣ v ∈ V ) consisting of ∣ V ∣ P |V|P ∣ V ∣ P many components, is stationary. For two distinct components u , v ∈ V u,v\in V u , v ∈ V , the associated component processes are said to independent if the stationary processes e P ( X v ) e_{P}(\mathsf{X}_{v}) e P ( X v ) and e P ( X u ) e_{P}(\mathsf{X}_{u}) e P ( X u ) are independent.
Convention 1. In this work, we will implicitly consider any V V V -variate P P P -stationary process X \mathsf{X} X in its stationary from. That means, when introducing X \mathsf{X} X as P P P -stationary, we will formally refer by X \mathsf{X} X to the stationary process ( e P ( X v ) ∣ v ∈ V ) (e_{P}(\mathsf{X}_{v})\mid v\in V) ( e P ( X v ) ∣ v ∈ V ) .
Definition 2 (P-stationary ACS). Let X \mathsf{X} X be a V V V -variate P P P -stationary process. The P P P -stationary autocovariance sequence (ACS) of X \mathsf{X} X is the ACS of the stationary process ( e P ( X v ) ∣ v ∈ V ) (e_{P}(\mathsf{X}_{v})\mid v\in V) ( e P ( X v ) ∣ v ∈ V ) . In particular, the P P P -stationary ACS of X \mathsf{X} X is block-structured such that ACS ( X ) : Z → BMat ( R , V , [ 0 , P − 1 ] ) \operatorname{ACS}(\mathsf{X}):\mathbb{Z}\to\operatorname{BMat}(\mathbb{R},V,[0,P-1]) ACS ( X ) : Z → BMat ( R , V , [ 0 , P − 1 ]) .
Definition 3 (Filtering P-stationary processes). Let V V V and W W W be finite sets, a block-structured filter of the form Λ : Z → BMat ( R , ( V , W ) , [ 0 , P − 1 ] ) \Lambda:\mathbb{Z}\to\operatorname{BMat}(\mathbb{R},(V,W),[0,P-1]) Λ : Z → BMat ( R , ( V , W ) , [ 0 , P − 1 ]) is said to be a P P P -stationary filter. Given a V V V -variate P P P -stationary process X \mathsf{X} X , the process Λ T ∗ X \Lambda^{\mathsf{T}}*\mathsf{X} Λ T ∗ X is said to be the Λ \Lambda Λ -filtering of X \mathsf{X} X .
Lemma 3. Let X \mathsf{X} X be a V V V -variate P P P -stationary process, and Λ : Z → BMat ( R , ( V , W ) , [ 0 , P − 1 ] ) \Lambda:\mathbb{Z}\to\operatorname{BMat}(\mathbb{R},(V,W),[0,P-1]) Λ : Z → BMat ( R , ( V , W ) , [ 0 , P − 1 ]) a P P P -stationary filter, then it holds that ACS ( Λ T ∗ X ) = Λ T ∗ ACS ( X ) ∗ ′ Λ \operatorname{ACS}(\Lambda^{\mathsf{T}}*\mathsf{X})=\Lambda^{\mathsf{T}}*\operatorname{ACS}(\mathsf{X})*'\Lambda ACS ( Λ T ∗ X ) = Λ T ∗ ACS ( X ) ∗ ′ Λ .
Question 1. As of Remark 1, for a P P P -stationary process X \mathsf{X} X and P P P -stationary filter Λ \Lambda Λ , the processes X \mathsf{X} X and Y : = Λ ∗ X \mathsf{Y}:=\Lambda*\mathsf{X} Y := Λ ∗ X are also R P RP R P -stationary. How to describe the linear relation between X \mathsf{X} X and Y \mathsf{Y} Y in terms of an R P RP R P -stationary filter?
To answer this question, we make use of the embedding operator constructed in the previous subsection. Let Λ \Lambda Λ be a P P P -stationary filter, then we define the block-wise embeddings as follows:
E R ( Λ ) : = [ e R ( Λ v , w ) ] v , w ∈ V ′ , E R ′ ( Λ ) : = [ e R ′ ( Λ v , w ) ] v , w ∈ V ′ . E_{R}(\Lambda):=\left[e_{R}(\Lambda_{v,w})\right]_{v,w\in V'},\qquad E'_{R}(\Lambda):=\left[e'_{R}(\Lambda_{v,w})\right]_{v,w\in V'}. E R ( Λ ) := [ e R ( Λ v , w ) ] v , w ∈ V ′ , E R ′ ( Λ ) := [ e R ′ ( Λ v , w ) ] v , w ∈ V ′ .
Note that E R ( Λ ) E_{R}(\Lambda) E R ( Λ ) has a three-level block structure, where the first level is indexed by V V V , the second by [ 0 , R − 1 ] [0,R-1] [ 0 , R − 1 ] and the third by [ 0 , P − 1 ] [0,P-1] [ 0 , P − 1 ] . Given a V V V -variate P P P -stationary process X \mathsf{X} X in its stationary form (see Convention 1), the embedding applied to X \mathsf{X} X takes the following stacked form:
E R ( X ) = [ e R ∘ e P ( X v ) ] v ∈ V . E_{R}(\mathsf{X})=\left[e_{R}\circ e_{P}(\mathsf{X}_{v})\right]_{v\in V}. E R ( X ) = [ e R ∘ e P ( X v ) ] v ∈ V .
In particular, E R ( X ) E_{R}(\mathsf{X}) E R ( X ) is the stationary form, see Convention 1, of X \mathsf{X} X read as an R P RP R P -stationary process. The following states that the block-wise embedding is compatible with convolution, filtering of processes and the computation of the ACS.
Proposition 4. Let Λ : Z → BMat ( R , ( V , W ) ) , [ 0 , P − 1 ] ) \Lambda: \mathbb{Z} \to\operatorname{BMat}(\mathbb{R},(V,W)),[0,P-1]) Λ : Z → BMat ( R , ( V , W )) , [ 0 , P − 1 ]) and Γ : Z → BMat ( R , ( U , V ) , [ 0 , P − 1 ] ) \Gamma: \mathbb{Z} \to\operatorname{BMat}(\mathbb{R},(U,V),[0,P-1]) Γ : Z → BMat ( R , ( U , V ) , [ 0 , P − 1 ]) be two P P P -stationary filters, then it holds that
E R ( Γ ∗ Λ ) = E R ( Γ ) ∗ E R ( Λ ) , E R ( Γ ∗ ′ Λ ) = E R ( Γ ) ∗ ′ E R ′ ( Λ ) . E_{\mathbb{R}}(\Gamma* \Lambda) = E_{\mathbb{R}}(\Gamma) * E_{\mathbb{R}}(\Lambda), \qquad E_{\mathbb{R}}(\Gamma*' \Lambda) = E_{\mathbb{R}}(\Gamma) *' E'_{\mathbb{R}}(\Lambda). E R ( Γ ∗ Λ ) = E R ( Γ ) ∗ E R ( Λ ) , E R ( Γ ∗ ′ Λ ) = E R ( Γ ) ∗ ′ E R ′ ( Λ ) .
Furthermore, for a V V V -variate P P P -stationary process X \mathbf{X} X , it holds that E R ( Λ T ∗ X ) = E R ( Λ T ) ∗ E R ( X ) E_{\mathbb{R}}(\Lambda^{\mathsf{T}} * \mathbf{X}) = E_{\mathbb{R}}(\Lambda^{\mathsf{T}}) * E_{\mathbb{R}}(\mathbf{X}) E R ( Λ T ∗ X ) = E R ( Λ T ) ∗ E R ( X ) , and ACS ( E R ( X ) ) = E R ( ACS ( X ) ) \operatorname{ACS}(E_{\mathbb{R}}(\mathbf{X})) = E_{\mathbb{R}}(\operatorname{ACS}(\mathbf{X})) ACS ( E R ( X )) = E R ( ACS ( X )) .
P-stationary structural equation processes
For a finite set X X X , we call a filter Λ : Z → R X × X \Lambda: \mathbb{Z} \to\mathbb{R}^{X \times X} Λ : Z → R X × X compatible with a finite directed graph G = ( X , D ) G = (X,D) G = ( X , D ) if v ↛ w v \not\to w v → w implies that Λ v , w \Lambda_{v,w} Λ v , w is the zero filter.
Definition 4 (P P P -stationary SEP). Let P P P be a positive integer. A P P P -stationary structural equation process (SEP) over some finite index set V V V is given by a triple S = ( G , Λ , X I ) \mathcal{S} = (G,\Lambda,\mathbf{X}^{\mathrm{I}}) S = ( G , Λ , X I ) consisting of: a finite directed graph G = ( V × [ 0 , P − 1 ] , D ) G = (V \times[0,P-1],D) G = ( V × [ 0 , P − 1 ] , D ) ; a P P P -stationary filter Λ : Z → BMat ( R , V , [ 0 , P − 1 ] ) \Lambda: \mathbb{Z} \to\operatorname{BMat}(\mathbb{R},V,[0,P-1]) Λ : Z → BMat ( R , V , [ 0 , P − 1 ]) compatible with G G G ; and a zero-mean V V V -variate P P P -stationary process X I \mathbf{X}^{\mathrm{I}} X I . A P P P -stationary process X \mathbf{X} X is said to be a solution for the SEP S \mathcal{S} S if it satisfies the structural equation X = Λ T ∗ X + X I \mathbf{X} = \Lambda^{\mathsf{T}} * \mathbf{X} + \mathbf{X}^{\mathrm{I}} X = Λ T ∗ X + X I . The P P P -stationary SEP S \mathcal{S} S is called stable if the filter S ( Λ ) = ∑ k ≥ 0 Λ k S(\Lambda) = \sum_{k \ge0} \Lambda^k S ( Λ ) = ∑ k ≥ 0 Λ k is absolutely summable.
For a stable P P P -stationary SEP S = ( G , Λ , X I ) \mathcal{S} = (G,\Lambda,\mathbf{X}^{\mathrm{I}}) S = ( G , Λ , X I ) , the process X = S ( Λ T ) ∗ X I \mathbf{X} = S(\Lambda^{\mathsf{T}}) * \mathbf{X}^{\mathrm{I}} X = S ( Λ T ) ∗ X I is the solution to the SEP S \mathcal{S} S . In this case, Lemma 3 implies that the P P P -stationary ACS of the solution X \mathbf{X} X is given by
ACS ( X ) = S ( Λ ) T ∗ ACS ( X I ) ∗ S ( Λ ) . (2) \operatorname{ACS}(\mathbf{X}) = S(\Lambda)^{\mathsf{T}} * \operatorname{ACS}(\mathbf{X}^{\mathrm{I}}) * S(\Lambda). \tag*{(2)} ACS ( X ) = S ( Λ ) T ∗ ACS ( X I ) ∗ S ( Λ ) . (2)
So far, the internal dynamics have been taken to be some general P P P -stationary process. The simplest instance of an P P P -stationary processes is an P P P -stationary white noise process.
Definition 5 (P P P -stationary white noise). A P P P -stationary univariate process η \eta η is called a white noise process if the stationary process e P ( η ) e_P(\eta) e P ( η ) is a white noise process such that all P P P components are mutually independent.
In this work, the internal dynamics will always arise as filtered white noise processes. In the following, we define their general structure. For that purpose, we will use V I = { v I ∣ v ∈ V } V^{\mathrm{I}} = \{v^{\mathrm{I}} \mid v \in V\} V I = { v I ∣ v ∈ V } to denote a (disjoint) a copy of V V V , i.e., V I ∩ V = ∅ V^{\mathrm{I}} \cap V = \varnothing V I ∩ V = ∅ .
Definition 6 (IDFS). A V V V -variate and P P P -stationary internal dynamic filtering structure (IDFS) is given by a P P P -stationary SEP S I = ( G I , ι , η ) \mathcal{S}^{\mathrm{I}} = (G^{\mathrm{I}},\iota,\eta) S I = ( G I , ι , η ) , where G I = ( V I ∪ V × [ 0 , P − 1 ] , D I ) G^{\mathrm{I}} = (V^{\mathrm{I}} \cup V \times[0,P-1],D^{\mathrm{I}}) G I = ( V I ∪ V × [ 0 , P − 1 ] , D I ) is such that every link is of the form ( v I , m ) → ( v , n ) (v^{\mathrm{I}},m) \to(v,n) ( v I , m ) → ( v , n ) for some v ∈ V v \in V v ∈ V and some m , n ∈ [ 0 , P − 1 ] m,n \in[0,P-1] m , n ∈ [ 0 , P − 1 ] ; a P P P -stationary filter ι : Z → BMat ( R , V I , V ) , [ 0 , P − 1 ] ) \iota: \mathbb{Z} \to\operatorname{BMat}(\mathbb{R},V^{\mathrm{I}},V),[0,P-1]) ι : Z → BMat ( R , V I , V ) , [ 0 , P − 1 ]) compatible with G I G^{\mathrm{I}} G I ; and a V V V -variate P P P -stationary mutually independent zero-mean white noise process η \eta η . The solution process, X I = ι T ∗ η \mathbf{X}^{\mathrm{I}} = \iota^{\mathsf{T}} * \eta X I = ι T ∗ η , will be referred to as the internal dynamics .
Note that the assumption on G I G^{\mathrm{I}} G I implies that the components of X I \mathbf{X}^{\mathrm{I}} X I are mutually independent. In Figure 1 , we display the graphical data of an 2-stationary SEP in conjunction with that of an IDFS.
Figure 1. For V = { u , v } V = \{u,v\} V = { u , v } , the graph specifies a 2-stationary SEP together with an IDFS. A vertex of the form a k a_k a k and a k I a_k^{\mathrm{I}} a k I , respectively, represents the subprocess defl 2 ∘ sh k ( X a ) \operatorname{defl}_2 \circ\operatorname{sh}_k(\mathbf{X}_a) defl 2 ∘ sh k ( X a ) and defl 2 ∘ sh k ( η a ) \operatorname{defl}_2 \circ\operatorname{sh}_k(\eta_a) defl 2 ∘ sh k ( η a ) , respectively.
P P P -stationary SVAR processes
We will now introduce P P P -stationary time series graphs and P P P -stationary SVAR processes. Then we will construct for a given P P P -stationary SVAR process its equivalent P P P -stationary SEP representation.
Definition 7 (P P P -stationary time series graph). A time series graph G = ( V × Z , D ) \mathcal{G} = (V \times\mathbb{Z}, \mathcal{D}) G = ( V × Z , D ) is called P P P -stationary if for any directed link v ( t − k ) → w ( t ) v(t-k) \to w(t) v ( t − k ) → w ( t ) on G \mathcal{G} G , it holds that for every l ∈ Z l \in\mathbb{Z} l ∈ Z also v ( t − k + l P ) → w ( t + l P ) v(t-k+lP) \to w(t+lP) v ( t − k + l P ) → w ( t + l P ) is a directed link on G \mathcal{G} G .
Given a P P P -stationary time series graph G = ( V × Z , D ) \mathcal{G} = (V \times\mathbb{Z}, \mathcal{D}) G = ( V × Z , D ) , we define the following sets of links:
D ( i , j ) : = { v ( t − k ) → w ( t ) ∣ t ∈ P Z − j , t − k ∈ P Z − i } , i , j ∈ [ 0 , P − 1 ] . \mathcal{D}^{(i,j)} := \{v(t-k) \to w(t) \mid t \in P\mathbb{Z}-j,\ t-k \in P\mathbb{Z}-i\}, \qquad i,j \in[0,P-1]. D ( i , j ) := { v ( t − k ) → w ( t ) ∣ t ∈ P Z − j , t − k ∈ P Z − i } , i , j ∈ [ 0 , P − 1 ] .
It follows from G \mathcal{G} G being P P P -stationary that these subsets are disjoint and cover the entire set D \mathcal{D} D . For a link v ( t − k ) → w ( t ) ∈ D ( i , j ) v(t-k) \to w(t) \in\mathcal{D}^{(i,j)} v ( t − k ) → w ( t ) ∈ D ( i , j ) , we write v → ( i , j ) l w v \xrightarrow{(i,j)}_{l} w v ( i , j ) l w , where l l l is such that k = i − j + l P k=i-j+lP k = i − j + l P . The order of a P P P -stationary time series graph is defined to be
max u , v ∈ V , i , j ∈ [ 0 , P − 1 ] { l ∣ u → ( i , j ) l v } . \max_{u,v\in V,\ i,j\in[0,P-1]} \{l \mid u \xrightarrow{(i,j)}_{l} v\}. u , v ∈ V , i , j ∈ [ 0 , P − 1 ] max { l ∣ u ( i , j ) l v } .
Throughout this work, we assume the order of any time series graph to be finite. We call a P P P -stationary filter Φ : Z → BMat ( R , V , [ 0 , P − 1 ] ) \Phi:\mathbb{Z} \to\operatorname{BMat}(\mathbb{R},V,[0,P-1]) Φ : Z → BMat ( R , V , [ 0 , P − 1 ]) compatible with a P P P -stationary time series graph G \mathcal{G} G if u ̸ → ( i , j ) v u \not\xrightarrow{(i,j)} v u ( i , j ) v implies that Φ u , v ( i , j ) = 0 \Phi^{(i,j)}_{u,v}=0 Φ u , v ( i , j ) = 0 .
Definition 8 (P P P -stationary SVAR process). A P P P -stationary structural vector autoregressive (SVAR) model consists of: a P P P -stationary time series graph G = ( V × Z , D ) \mathcal{G}=(V \times\mathbb{Z},\mathcal{D}) G = ( V × Z , D ) ; a P P P -stationary filter Φ : Z → BMat ( R , V , [ 0 , P − 1 ] ) \Phi:\mathbb{Z} \to\operatorname{BMat}(\mathbb{R},V,[0,P-1]) Φ : Z → BMat ( R , V , [ 0 , P − 1 ]) compatible with G \mathcal{G} G ; and a zero-mean P P P -stationary mutually independent white noise process η \eta η . A P P P -stationary process X X X solves the SVAR model if it satisfies X = Φ T ∗ X + η X=\Phi^{\mathsf{T}}*X+\eta X = Φ T ∗ X + η .
In the following, R G \mathcal{R}^{\mathcal{G}} R G will be the vector space of filters Φ : Z → BMat ( R , V , [ 0 , P − 1 ] ) \Phi:\mathbb{Z} \to\operatorname{BMat}(\mathbb{R},V,[0,P-1]) Φ : Z → BMat ( R , V , [ 0 , P − 1 ]) compatible with the time series graph G \mathcal{G} G . Note that R G \mathcal{R}^{\mathcal{G}} R G is finite dimensional, since we assume G \mathcal{G} G to have finite order.
Example 1. Consider the finite excerpt of the graph G = ( { u , v } × Z , D ) \mathcal{G}= (\{u,v\} \times\mathbb{Z},\mathcal{D}) G = ({ u , v } × Z , D ) depicted in Figure 2 together with a filter Φ \Phi Φ compatible with G \mathcal{G} G . Following the notation introduced above, the 2-stationary graph G \mathcal{G} G contains the following relations, each with an associated entry in Φ \Phi Φ . The auto-dependencies of the process u u u are given as follows:
Figure 2. For t ∈ 2 Z t \in2\mathbb{Z} t ∈ 2 Z , the graph shows an excerpt of a 2-stationary time series graph.
u → ( 0 , 1 ) 1 u : ϕ u , u ( 0 , 1 ) ( 1 ) , u → ( 1 , 0 ) 0 u : ϕ u , u ( 1 , 0 ) ( 0 ) . u \xrightarrow{(0,1)}_{1} u:\ \phi_{u,u}^{(0,1)}(1), \qquad u \xrightarrow{(1,0)}_{0} u:\ \phi_{u,u}^{(1,0)}(0). u ( 0 , 1 ) 1 u : ϕ u , u ( 0 , 1 ) ( 1 ) , u ( 1 , 0 ) 0 u : ϕ u , u ( 1 , 0 ) ( 0 ) .
Similarly, one can read off the relations associated with the auto-causal links of v v v and the links arising from the lagged 2-stationary causal links from u u u on v v v .
In the remainder of this subsection, we will reformulate a given P P P -stationary SVAR process as an equivalent P P P -stationary SEP, meaning that the solution processes of the SVAR model and the associated SEP will coincide. For the remainder of this section, we fix an P P P -stationary SVAR model ( G , Φ , η ) (\mathcal{G}, \Phi, \eta) ( G , Φ , η ) .
Figure 3. a.) The summary graph derived from the graph in Figure 2 . b.) The graph G ∪ G I \mathcal{G} \cup\mathcal{G}^{\mathrm{I}} G ∪ G I derived from a.). In both graphs, a vertex of the form a k a_k a k represents the subprocess defl 2 ∘ sh k ( X a ) \operatorname{defl}_2 \circ\operatorname{sh}^{k}(X_a) defl 2 ∘ sh k ( X a ) , and a k I a_k^{\mathrm{I}} a k I represents defl 2 ∘ sh k ( η a ) \operatorname{defl}_2 \circ\operatorname{sh}^{k}(\eta_a) defl 2 ∘ sh k ( η a ) .
We need to define two graphs. Namely, the graph G \mathcal{G} G defining the P P P -stationary SEP and the graph G I \mathcal{G}^{\mathrm{I}} G I defining the internal dynamics. We will derive both of them from the following.
Definition 9 (P -stationary summary graph). Let G \mathcal{G} G be a P P P -stationary time series graph. Its summary graph, denoted smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) , is a directed graph over the vertex set V × [ 0 , P − 1 ] V \times[0, P-1] V × [ 0 , P − 1 ] that connects two distinct vertices by ( v , i ) → ( w , j ) (v,i) \to(w,j) ( v , i ) → ( w , j ) if and only if there is a k ≥ 0 k \ge0 k ≥ 0 such that v → k ( i , j ) w v \xrightarrow{k}^{(i,j)} w v k ( i , j ) w holds on G \mathcal{G} G .
To define the structure for the internal dynamics of v ∈ V v \in V v ∈ V , let G v \mathcal{G}_v G v be the restriction of G \mathcal{G} G to { v } × Z \{v\} \times\mathbb{Z} { v } × Z . The summary graphs ( smry ( G v ) ∣ v ∈ V ) (\operatorname{smry}(\mathcal{G}_v) \mid v \in V) ( smry ( G v ) ∣ v ∈ V ) determine G I \mathcal{G}^{\mathrm{I}} G I in the following way:
P smry ( G v ) ( ( v , i ) , ( v , j ) ) ≠ ∅ ⟺ ( v I , i ) → ( v , j ) . P_{\operatorname{smry}(\mathcal{G}_v)}((v,i),(v,j)) \ne\varnothing\Longleftrightarrow(v^{\mathrm{I}},i) \to(v,j). P smry ( G v ) (( v , i ) , ( v , j )) = ∅ ⟺ ( v I , i ) → ( v , j ) .
Finally, the process graph G \mathcal{G} G of the time series graph G \mathcal{G} G is defined as follows: for v ≠ w v \ne w v = w , there is a directed link ( v , i ) → ( w , j ) (v,i) \to(w,j) ( v , i ) → ( w , j ) on G \mathcal{G} G if and only if there is some l ∈ [ 0 , P − 1 ] l \in[0,P-1] l ∈ [ 0 , P − 1 ] such that: ( v , i ) → ( w , l ) (v,i) \to(w,l) ( v , i ) → ( w , l ) is a directed link on smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) , and ( w I , l ) → ( w , j ) (w^{\mathrm{I}},l) \to(w,j) ( w I , l ) → ( w , j ) is a link on G I \mathcal{G}^{\mathrm{I}} G I .
We define the internal dynamics of v ∈ V v \in V v ∈ V to be the P P P -stationary process X v I X_v^{\mathrm{I}} X v I that satisfies
X v I = Φ v , v T ∗ X v I + η v . (3) X_v^{\mathrm{I}} = \Phi_{v,v}^{\mathrm{T}} * X_v^{\mathrm{I}} + \eta_v. \tag*{(3)} X v I = Φ v , v T ∗ X v I + η v . (3)
The filter ι v \iota_v ι v generating the internal dynamics (meaning that X v I = ι v T ∗ η v X_v^{\mathrm{I}} = \iota_v^{\mathrm{T}} * \eta_v X v I = ι v T ∗ η v ) is implicitly defined by the following linear relation:
ι v = I P + ι v ∗ Φ v , v ⟺ ι v ∗ ( I P − Φ v , v ) = I P . (4) \iota_v = I_P + \iota_v * \Phi_{v,v} \Longleftrightarrow\iota_v * (I_P - \Phi_{v,v}) = I_P. \tag*{(4)} ι v = I P + ι v ∗ Φ v , v ⟺ ι v ∗ ( I P − Φ v , v ) = I P . (4)
In fact, the solution to the above equation is given by ι v = S ( Φ v , v ) \iota_v = S(\Phi_{v,v}) ι v = S ( Φ v , v ) . In order to make sure that ι v \iota_v ι v is a filter, i.e., absolutely summable, we need to impose a condition on Φ v , v \Phi_{v,v} Φ v , v . Its explicit formulation generalizes the stability notion used in [34 , 36 ] and will be given at the end of this subsection.
Example 2. We revisit the example associated with the time series graph of order one shown in Figure 2 . Equation (4) gives the following recursive relations for the internal dynamics of any order one 2-stationary SVAR process:
ι ( 0 ) = ( I 2 − Φ ( 0 ) ) − 1 , ι ( k ) = ι ( 0 ) ( Φ ( 1 ) ι ( 0 ) ) k . \iota(0) = (I_2 - \Phi(0))^{-1}, \qquad\iota(k) = \iota(0)(\Phi(1)\iota(0))^k. ι ( 0 ) = ( I 2 − Φ ( 0 ) ) − 1 , ι ( k ) = ι ( 0 ) ( Φ ( 1 ) ι ( 0 ) ) k .
For the process u u u , with α 0 : = ϕ u , u ( 1 , 0 ) ( 0 ) \alpha_0 := \phi_{u,u}^{(1,0)}(0) α 0 := ϕ u , u ( 1 , 0 ) ( 0 ) and α 1 : = ϕ u , u ( 0 , 1 ) ( 1 ) \alpha_1 := \phi_{u,u}^{(0,1)}(1) α 1 := ϕ u , u ( 0 , 1 ) ( 1 ) , we get
ι u ( 0 ) = [ 1 0 α 0 1 ] , ι u ( k ) = [ α 0 k α 1 k α 0 k − 1 α 1 k α 0 k + 1 α 1 k α 0 k α 1 k ] . \iota_u(0) = \begin{bmatrix} 1 & 0 \\ \alpha_0 & 1 \end{bmatrix}, \qquad
\iota_u(k) = \begin{bmatrix} \alpha_0^k\alpha_1^k & \alpha_0^{k-1}\alpha_1^k \\ \alpha_0^{k+1}\alpha_1^k & \alpha_0^k\alpha_1^k \end{bmatrix}. ι u ( 0 ) = [ 1 α 0 0 1 ] , ι u ( k ) = [ α 0 k α 1 k α 0 k + 1 α 1 k α 0 k − 1 α 1 k α 0 k α 1 k ] .
For the process v v v , with γ 0 , 1 = ϕ v , v ( 0 , 1 ) ( 1 ) \gamma_{0,1} = \phi_{v,v}^{(0,1)}(1) γ 0 , 1 = ϕ v , v ( 0 , 1 ) ( 1 ) and γ 0 , 0 = ϕ v , v ( 0 , 0 ) ( 1 ) \gamma_{0,0} = \phi_{v,v}^{(0,0)}(1) γ 0 , 0 = ϕ v , v ( 0 , 0 ) ( 1 ) and γ 1 , 1 = ϕ v , v ( 1 , 1 ) ( 1 ) \gamma_{1,1} = \phi_{v,v}^{(1,1)}(1) γ 1 , 1 = ϕ v , v ( 1 , 1 ) ( 1 ) , we get
ι v ( 0 ) = [ 1 0 0 1 ] , ι v ( k ) = [ γ 0 , 0 k γ 0 , 1 ∑ j = 1 k − 1 γ 0 , 0 j γ 1 , 1 k − 1 − j 0 γ 1 , 1 k ] . \iota_v(0) = \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix}, \qquad
\iota_v(k) = \begin{bmatrix} \gamma_{0,0}^k & \gamma_{0,1}\sum_{j=1}^{k-1}\gamma_{0,0}^j\gamma_{1,1}^{k-1-j} \\ 0 & \gamma_{1,1}^k \end{bmatrix}. ι v ( 0 ) = [ 1 0 0 1 ] , ι v ( k ) = [ γ 0 , 0 k 0 γ 0 , 1 ∑ j = 1 k − 1 γ 0 , 0 j γ 1 , 1 k − 1 − j γ 1 , 1 k ] .
The following states that the filters ι v \iota_v ι v , defined in Equation (3) are compatible with the graph G \mathcal{G} G .
Proposition 5. Let v ∈ V v \in V v ∈ V and i , j ∈ [ 0 , P − 1 ] i,j \in[0,P-1] i , j ∈ [ 0 , P − 1 ] be such that ( v i , i ) ↛ ( v , j ) (v^i,i) \not\to(v,j) ( v i , i ) → ( v , j ) , then ι v ( i , j ) = 0 \iota_v^{(i,j)} = 0 ι v ( i , j ) = 0 .
It remains to construct the link filters compatible with the process graph G = ( V × [ 0 , P − 1 ] , D ) G = (V \times[0,P-1],D) G = ( V × [ 0 , P − 1 ] , D ) . For two components u , v ∈ V u,v \in V u , v ∈ V we define the P P P -stationary filter
λ u , v : = { Φ u , v ∗ ι v u ≠ v , 0 u = v . (5) \lambda_{u,v} := \begin{cases} \Phi_{u,v} * \iota_v & u \ne v, \\ 0 & u = v. \end{cases} \tag*{(5)} λ u , v := { Φ u , v ∗ ι v 0 u = v , u = v . (5)
Compatibility of the link filters with the process graph is a direct consequence of Proposition 5 and the definition of the process graph.
Corollary 6. The P P P -stationary filter Λ = [ λ u , v ] u , v ∈ V \Lambda= [\lambda_{u,v}]_{u,v\in V} Λ = [ λ u , v ] u , v ∈ V is compatible with the process graph G G G .
Example 3. For the 2-stationary SVAR model with underlying time series graph shown in Figure 2 , we set β 0 : = ϕ u , v ( 0 , 0 ) ( 0 ) \beta_0 := \phi_{u,v}^{(0,0)}(0) β 0 := ϕ u , v ( 0 , 0 ) ( 0 ) and β 1 : = ϕ u , v ( 0 , 1 ) ( 1 ) \beta_1 := \phi_{u,v}^{(0,1)}(1) β 1 := ϕ u , v ( 0 , 1 ) ( 1 ) . Then the 2-stationary link filter for u → v u \to v u → v is
λ u , v ( k ) = Φ u , v ( 0 ) ι v ( k ) + Φ u , v ( 1 ) ι v ( k − 1 ) = [ β 0 γ 0 , 0 k β 0 γ 0 , 1 ∑ j = 0 k − 1 γ 0 , 0 j γ 1 , 1 k − 1 − j + β 1 γ 1 , 1 k − 1 0 0 ] . \begin{aligned}
\lambda_{u,v}(k) = \Phi_{u,v}(0)\iota_v(k) + \Phi_{u,v}(1)\iota_v(k-1) \\
&= \begin{bmatrix} \beta_0\gamma_{0,0}^k & \beta_0\gamma_{0,1}\sum_{j=0}^{k-1}\gamma_{0,0}^j\gamma_{1,1}^{k-1-j} + \beta_1\gamma_{1,1}^{k-1} \\ 0 & 0 \end{bmatrix}.
\end{aligned} λ u , v ( k ) = Φ u , v ( 0 ) ι v ( k ) + Φ u , v ( 1 ) ι v ( k − 1 ) = [ β 0 γ 0 , 0 k 0 β 0 γ 0 , 1 ∑ j = 0 k − 1 γ 0 , 0 j γ 1 , 1 k − 1 − j + β 1 γ 1 , 1 k − 1 0 ] .
Note that if G G G is acyclic, and Φ \Phi Φ is such that every ι v \iota_v ι v , as in Equation (4) , is absolutely summable, then S ( Λ ) S(\Lambda) S ( Λ ) is absolutely summable. However, if G G G is cyclic, then each ι v \iota_v ι v being absolutely summable does not suffice to ensure absolute summability of S ( Λ ) S(\Lambda) S ( Λ ) . In the following, we provide a set of filters Φ \Phi Φ that yield a stable SEP representation, even if G G G is cyclic. Let V ~ 1 , … , V ~ K \widetilde{V}_1,\ldots,\widetilde{V}_K V 1 , … , V K be the strongly connected components of the summary graph smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) , then we define the following subset of G \mathcal{G} G -compatible filters:
R s t b G : = { Φ ∈ R G | ∥ [ Φ ⊤ ] V ~ i , V ~ i ∥ 1 < 1 , for all i ∈ [ K ] } . (6) \mathbb{R}^{\mathcal{G}}_{\mathrm{stb}} := \left\{ \Phi\in\mathbb{R}^{\mathcal{G}} \,\middle|\, \left\|[\Phi^\top]_{\widetilde{V}_i,\widetilde{V}_i}\right\|_1 < 1,\ \text{for all } i \in[K] \right\}. \tag*{(6)} R stb G := { Φ ∈ R G [ Φ ⊤ ] V i , V i 1 < 1 , for all i ∈ [ K ] } . (6)
with ∥ Φ ∥ 1 : = ∥ ∑ k ∈ Z ∣ Φ ( k ) ∣ ∥ 1 \|\Phi\|_1 := \left\|\sum_{k\in\mathbb{Z}}|\Phi(k)|\right\|_1 ∥Φ ∥ 1 := ∑ k ∈ Z ∣Φ ( k ) ∣ 1 , where on the right, absolute values are taken entry-wise and the norm is the operator norm w.r.t. ∥ ∙ ∥ 1 \|\bullet\|_1 ∥ ∙ ∥ 1 . The following states that the solution of the SVAR model coincides with the solution of the associated SEP. Its proof is deferred to the supplementary material.
Proposition 7. Let ( G , Φ , η ) (\mathcal{G},\Phi,\eta) ( G , Φ , η ) be a P P P -stationary SVAR model and Φ ∈ R s t b G \Phi\in\mathbb{R}^{\mathcal{G}}_{\mathrm{stb}} Φ ∈ R stb G , then the associated P P P -stationary SEP ( G , Λ , ι ⊤ ∗ η ) (G,\Lambda,\iota^\top* \eta) ( G , Λ , ι ⊤ ∗ η ) is stable, and furthermore, a P P P -stationary process X X X is a solution to the SVAR model if and only if it is a solution to the P P P -stationary SEP model.
The RP-stationary SEP representation of a P-stationary SVAR model
Given a stable P P P -stationary SEP S = ( G , Λ , X I ) S = (G,\Lambda,X^{\mathsf I}) S = ( G , Λ , X I ) , the properties of the embedding E R E_R E R allow one to identify an R P RP R P -stationary SEP S R = ( G R , E R ( Λ ) , E R ( X I ) ) S_R = (G_R,E_R(\Lambda),E_R(X^{\mathsf I})) S R = ( G R , E R ( Λ ) , E R ( X I )) that is equivalent to S S S , in the sense that their solutions coincide. If the SEP S S S has an underlying P P P -stationary SVAR model V = ( G , Φ , η ) \mathcal{V} = (\mathcal{G},\Phi,\eta) V = ( G , Φ , η ) , then it is possible to construct an R P RP R P -stationary time series graph G R \mathcal{G}_R G R such that the embedded SVAR model V R = ( G R , E R ( Φ ) , E R ( η ) ) \mathcal{V}_R = (\mathcal{G}_R,E_R(\Phi),E_R(\eta)) V R = ( G R , E R ( Φ ) , E R ( η )) and V \mathcal{V} V are equivalent to each other. However, we will not need the embedded time series graph G R \mathcal{G}_R G R in the present work. Instead, we will make use of the following: Since the embedding preserves the identity filter and is compatible with convolution, see Proposition 4, it holds that
E R ( ι v ) = I R + E R ( Φ v ) ∗ E R ( ι v ) , E R ( λ u , v ) = E R ( Φ u , v ) ∗ E R ( ι v ) . (7) E_R(\iota_v) = I_R + E_R(\Phi_v) * E_R(\iota_v), \qquad E_R(\lambda_{u,v}) = E_R(\Phi_{u,v}) * E_R(\iota_v). \tag*{(7)} E R ( ι v ) = I R + E R ( Φ v ) ∗ E R ( ι v ) , E R ( λ u , v ) = E R ( Φ u , v ) ∗ E R ( ι v ) . (7)
To construct the frequency domain representation of a P P P -stationary SEP, we briefly recall the Fourier transformation, see, e.g., [5 ] . Then we will see that the frequency domain representation of an P P P -stationary SVAR process is parameterized over the field of rational functions with real coefficients. In the final part of this section, we will derive a linear relationship between the P P P -stationary spectrum, i.e., the Fourier transformed ACS, and its R P RP R P -stationary embedding. In the next section, we will use this relationship to decide whether a given entry in the P P P -stationary spectrum can be linearly recovered from a temporally incomplete observational pattern.
Throughout this section, ( G , Φ , η ) (G,\Phi,\eta) ( G , Φ , η ) specifies a P P P -stationary SVAR model over a finite set of vertices V V V . Furthermore, we will assume throughout this section that Φ ∈ R s t b G \Phi\in\mathbb{R}_{\mathrm{stb}}^{\mathcal{G}} Φ ∈ R stb G . If not stated otherwise, we will refer by S = ( G , Λ , ι T ∗ η ) S=(G,\Lambda,\iota^{\mathsf{T}}*\eta) S = ( G , Λ , ι T ∗ η ) to the associated P P P -stationary SEP representation.
Fourier transformation
Let S 1 S^1 S 1 be the set of all z ∈ C z \in\mathbb{C} z ∈ C such that ∣ z ∣ = 1 \lvert z\rvert=1 ∣ z ∣ = 1 . The linear and continuous Fourier transform is a Hilbert isomorphism between the space of complex-valued square summable sequences, denoted by ℓ 2 ( Z ; C ) \ell^2(\mathbb{Z};\mathbb{C}) ℓ 2 ( Z ; C ) , and the space L 2 ( S 1 ; C ) L^2(S^1;\mathbb{C}) L 2 ( S 1 ; C ) consisting of the (equivalence classes of) square integrable functions f : S 1 → C f:S^1 \to\mathbb{C} f : S 1 → C . The Fourier transform takes a sequence λ \lambda λ to the function F ( λ ) = ( z ↦ ∑ k ∈ Z λ ( k ) z k ) \mathcal{F}(\lambda)=\left(z\mapsto\sum_{k\in\mathbb{Z}}\lambda(k)z^k\right) F ( λ ) = ( z ↦ ∑ k ∈ Z λ ( k ) z k ) .
It will be useful to recall that the Fourier transform is compatible with the (tilted) convolution of sequences in the following sense: Let λ , μ ∈ ℓ 2 ( Z ; C ) \lambda,\mu\in\ell^2(\mathbb{Z};\mathbb{C}) λ , μ ∈ ℓ 2 ( Z ; C ) be sequences, then it holds that F ( λ ∗ μ ) = F ( λ ) F ( μ ) \mathcal{F}(\lambda*\mu)=\mathcal{F}(\lambda)\mathcal{F}(\mu) F ( λ ∗ μ ) = F ( λ ) F ( μ ) and F ( λ ∗ ′ μ ) = F ( λ ) F ( μ ) ∗ \mathcal{F}(\lambda*'\mu)=\mathcal{F}(\lambda)\mathcal{F}(\mu)^* F ( λ ∗ ′ μ ) = F ( λ ) F ( μ ) ∗ , where right hand sides are point-wise products of functions, and ∗ * ∗ in superscript means taking complex conjugate point-wise, see [34 ] .
For V V V a finite set, we denote by I V I_V I V the identity mapping in L 2 ( S 1 ; C V × V ) L^2(S^1;\mathbb{C}^{V\times V}) L 2 ( S 1 ; C V × V ) , i.e., I V ( z ) = I V I_V(z)=I_V I V ( z ) = I V is the identity matrix for every z ∈ S 1 z\in S^1 z ∈ S 1 . We omit the subscript if there is no ambiguity about the index set V V V of the underlying vector space.
The frequency domain analogues of the j j j -shift and P P P -inflation operators (for sequences) map a function f ∈ L 1 ( S 1 ; C ) f\in L^1(S^1;\mathbb{C}) f ∈ L 1 ( S 1 ; C ) to sh j ( f ) : = ( z ↦ z j f ( z ) ) \operatorname{sh}^j(f):=\left(z\mapsto z^j f(z)\right) sh j ( f ) := ( z ↦ z j f ( z ) ) and infl P ( f ) : = ( z ↦ f ( z P ) ) \operatorname{infl}_P(f):=\left(z\mapsto f\left(z^P\right)\right) infl P ( f ) := ( z ↦ f ( z P ) ) , respectively. These mappings are linear and related to their analogues in the space of sequences as follows:
F ( sh j ( λ ) ) = sh j ( F ( λ ) ) , F ( infl P ( λ ) ) = infl P ( F ( λ ) ) . (8) \mathcal{F}(\operatorname{sh}^j(\lambda))=\operatorname{sh}^j(\mathcal{F}(\lambda)),\qquad\mathcal{F}(\operatorname{infl}_P(\lambda))=\operatorname{infl}_P(\mathcal{F}(\lambda)). \tag*{(8)} F ( sh j ( λ )) = sh j ( F ( λ )) , F ( infl P ( λ )) = infl P ( F ( λ )) . (8)
The embedding operators e N , e N ′ e_N,e'_N e N , e N ′ for sequences, defined in § 3 , can be translated to the frequency domain. Specifically, since F \mathcal{F} F is an isomorphism, we can define for f ∈ L 2 ( S 1 ; C V × V ) f\in L^2(S^1;\mathbb{C}^{V\times V}) f ∈ L 2 ( S 1 ; C V × V ) and some positive integer N N N the following:
e N ( f ) : = F ∘ e N ∘ F − 1 ( f ) , e N ′ ( f ) : = F ∘ e N ′ ∘ F − 1 ( f ) . (9) e_N(f):=\mathcal{F}\circ e_N\circ\mathcal{F}^{-1}(f),\qquad e'_N(f):=\mathcal{F}\circ e'_N\circ\mathcal{F}^{-1}(f). \tag*{(9)} e N ( f ) := F ∘ e N ∘ F − 1 ( f ) , e N ′ ( f ) := F ∘ e N ′ ∘ F − 1 ( f ) . (9)
Note that e N ( f ) , e N ′ ( f ) ∈ L 2 ( S 1 ; BMat ( C , [ 0 , P − 1 ] ( V , U ) ) ) e_N(f),e'_N(f)\in L^2(S^1;\operatorname{BMat}(\mathbb{C},[0,P-1](V,U))) e N ( f ) , e N ′ ( f ) ∈ L 2 ( S 1 ; BMat ( C , [ 0 , P − 1 ] ( V , U ))) , i.e., they are block-structured. Furthermore, for a block-structured function F : L 2 ( S 1 ; BMat ( C , V , [ 0 , P − 1 ] ) ) F:L^2(S^1;\operatorname{BMat}(\mathbb{C},V,[0,P-1])) F : L 2 ( S 1 ; BMat ( C , V , [ 0 , P − 1 ])) , we define the block-wise operators E N ( F ) = [ e N ( F v , w ) ] v , w ∈ V E_N(F)=[e_N(F_{v,w})]_{v,w\in V} E N ( F ) = [ e N ( F v , w ) ] v , w ∈ V and E N ′ ( F ) = [ e N ′ ( F v , w ) ] v , w ∈ V E'_N(F)=[e'_N(F_{v,w})]_{v,w\in V} E N ′ ( F ) = [ e N ′ ( F v , w ) ] v , w ∈ V .
To derive the main result of this section, we will use that the embedding has the following properties.
Lemma 8. Let f , g ∈ L 2 ( S 1 ; C V × V ) f,g\in L^2(S^1;\mathbb{C}^{V\times V}) f , g ∈ L 2 ( S 1 ; C V × V ) , then it holds that e N ( f g ) = e N ( f ) e N ( g ) e_N(fg)=e_N(f)e_N(g) e N ( f g ) = e N ( f ) e N ( g ) . Furthermore, e N ( I ) = e N ′ ( I ) = I e_N(I)=e'_N(I)=I e N ( I ) = e N ′ ( I ) = I , i.e., the operators preserve the identity function. In particular, if f f f is point-wise invertible, then it holds that e N ( f − 1 ) = e N ( f ) − 1 e_N(f^{-1})=e_N(f)^{-1} e N ( f − 1 ) = e N ( f ) − 1 . Setting f ~ : = e N ( f ) \tilde{f}:=e_N(f) f ~ := e N ( f ) , it holds that f ~ i , j = f ~ i + 1 , j + 1 \tilde{f}_{i,j}=\tilde{f}_{i+1,j+1} f ~ i , j = f ~ i + 1 , j + 1 for all i , j ∈ [ 0 , N − 2 ] i,j\in[0,N-2] i , j ∈ [ 0 , N − 2 ] , and f ~ N − 1 , k − 1 = sh 1 ( f ~ 0 , k ) \tilde{f}_{N-1,k-1}=\operatorname{sh}^{1}(\tilde{f}_{0,k}) f ~ N − 1 , k − 1 = sh 1 ( f ~ 0 , k ) for every k ∈ [ 0 , N − 1 ] k\in[0,N-1] k ∈ [ 0 , N − 1 ] . Finally, e N ′ ( f ) = [ f ~ j , i ] i , j ∈ [ 0 , N − 1 ] e'_N(f)=[\tilde{f}_{j,i}]_{i,j\in[0,N-1]} e N ′ ( f ) = [ f ~ j , i ] i , j ∈ [ 0 , N − 1 ] , where the right hand side is the block-transpose of f ~ \tilde{f} f ~ .
The statements in Lemma 8 follow directly from the definition of the embedding in the sequence space and Lemma 2 and the basic properties of the Fourier transform.
Suppose X X X is a P P P -stationary process and let Σ \Sigma Σ be its P P P -stationary ACS. If Σ \Sigma Σ is absolutely summable, then we define the spectral density of X X X to be S : = F ( Σ ) \mathcal{S}:=\mathcal{F}(\Sigma) S := F ( Σ ) .
P-stationary SVAR frequency domain representation
Let S = ( G , Λ , ι ⊤ ∗ η ) \mathcal{S}=(\mathcal{G},\Lambda,\iota^\top*\eta) S = ( G , Λ , ι ⊤ ∗ η ) be a stable P P P -stationary SEP, then we define H : = F ( Λ ) H:=\mathcal{F}(\Lambda) H := F ( Λ ) to be the block-matrix of link functions, and f = F ( ι ) f=\mathcal{F}(\iota) f = F ( ι ) to be the block-matrix of internal functions. Equation (2) in combination with the basic properties of the Fourier transform allow to factorize the spectrum of the solution process X X X and the internal dynamics X I = ι ⊤ ∗ η X^{I}=\iota^\top*\eta X I = ι ⊤ ∗ η , respectively, as follows:
S = ( I − H ) − ⊤ S I ( I − H ) − ∗ , S I = S I = f ⊤ Ω f ∗ . \mathcal{S}=(I-H)^{-\top}\mathcal{S}^{I}(I-H)^{-*},\qquad\mathcal{S}^{I}=S^{I}=f^\top\Omega f^*. S = ( I − H ) − ⊤ S I ( I − H ) −∗ , S I = S I = f ⊤ Ω f ∗ .
where Ω \Omega Ω denotes the constant function that maps every z ∈ S 1 z\in S^{1} z ∈ S 1 to the diagonal matrix ACS ( η ) ( 0 ) \operatorname{ACS}(\eta)(0) ACS ( η ) ( 0 ) . In the following, ( G , H , S I ) (\mathcal{G},H,\mathcal{S}^{I}) ( G , H , S I ) will be referred to as the spectral representation of S \mathcal{S} S .
We will now see that the spectral representation of the P P P -stationary SEP associated with the SVAR model ( G , Φ , η ) (\mathcal{G},\Phi,\eta) ( G , Φ , η ) can be formulated over the rational functions. First, recall that the space of polynomials with real coefficients in one indeterminate z z z is
R [ z ] : = { ∑ k = 0 n α k z k ∣ α k ∈ R , α n ≠ 0 } . \mathbb{R}[z]:=\left\{\sum_{k=0}^{n}\alpha_k z^k\mid\alpha_k\in\mathbb{R},\alpha_n\ne0\right\}. R [ z ] := { k = 0 ∑ n α k z k ∣ α k ∈ R , α n = 0 } .
The field of fractions of R [ z ] \mathbb{R}[z] R [ z ] is the field of rational functions and will be denoted R ( z ) \mathbb{R}(z) R ( z ) , on which there is an involution ( ∙ ) ∗ : R ( z ) → R ( z ) (\bullet)^*:\mathbb{R}(z)\to\mathbb{R}(z) ( ∙ ) ∗ : R ( z ) → R ( z ) , i.e., a self-inverse isomorphism that respects addition and multiplication of rational functions, with the additional property that r ∗ ( z ) = r ( z ∗ ) r^*(z)=r(z^*) r ∗ ( z ) = r ( z ∗ ) , for every z ∈ S 1 z\in S^{1} z ∈ S 1 , see [36 ] . For v ∈ V v\in V v ∈ V , it follows from the definition of the internal filter (Equation (3) ) and the compatibility of the Fourier transform with convolution that
f v ( Φ ) = ( I P − F ( Φ v , v ) ) − 1 ∈ R ( z ) P × P . (10) f_v(\Phi)=\left(I_P-\mathcal{F}(\Phi_{v,v})\right)^{-1}\in\mathbb{R}(z)^{P\times P}. \tag*{(10)} f v ( Φ ) = ( I P − F ( Φ v , v ) ) − 1 ∈ R ( z ) P × P . (10)
This matrix has entries in R ( z ) \mathbb{R}(z) R ( z ) since F ( Φ v , v ) \mathcal{F}(\Phi_{v,v}) F ( Φ v , v ) is a matrix of polynomials. Combining Equation (10) with the definition of the link filter, i.e., Equation (5) , yields that the matrix-valued link functions associated with two distinct processes u ≠ v u\ne v u = v is given by the matrix of rational functions
h u , v ( Φ ) = F ( Φ u , v ) f v = F ( Φ u , v ) ( I P − F ( Φ v , v ) ) − 1 ∈ R ( z ) P × P . (11) h_{u,v}(\Phi)=\mathcal{F}(\Phi_{u,v})f_v=\mathcal{F}(\Phi_{u,v})\left(I_P-\mathcal{F}(\Phi_{v,v})\right)^{-1}\in\mathbb{R}(z)^{P\times P}. \tag*{(11)} h u , v ( Φ ) = F ( Φ u , v ) f v = F ( Φ u , v ) ( I P − F ( Φ v , v ) ) − 1 ∈ R ( z ) P × P . (11)
Example 4. For the time series graph from Figure 2 , and notation as in Example 2 and Example 3 the internal function of the component v v v takes the following form
f v ( z ) = 1 ( 1 − γ 0 , 0 z ) ( 1 − γ 1 , 1 z ) [ 1 − γ 1 , 1 z γ 0 , 1 0 1 − γ 0 , 0 z ] , f_v(z)=\frac{1}{(1-\gamma_{0,0}z)(1-\gamma_{1,1}z)}
\begin{bmatrix}
1-\gamma_{1,1}z & \gamma_{0,1}\\
0 & 1-\gamma_{0,0}z
\end{bmatrix}, f v ( z ) = ( 1 − γ 0 , 0 z ) ( 1 − γ 1 , 1 z ) 1 [ 1 − γ 1 , 1 z 0 γ 0 , 1 1 − γ 0 , 0 z ] ,
and the 2-stationary link function of u → v u\to v u → v is
h u , v ( z ) = [ β 0 β 1 z 0 0 ] f v ( z ) = 1 ( 1 − γ 0 , 0 z ) ( 1 − γ 1 , 1 z ) [ β 0 ( 1 − γ 1 , 1 z ) β 0 γ 0 , 1 + β 1 ( z − γ 0 , 0 z 2 ) 0 0 ] . h_{u,v}(z)=
\begin{bmatrix}
\beta_0 & \beta_1z\\
0 & 0
\end{bmatrix}
f_v(z)
=
\frac{1}{(1-\gamma_{0,0}z)(1-\gamma_{1,1}z)}
\begin{bmatrix}
\beta_0(1-\gamma_{1,1}z) & \beta_0\gamma_{0,1}+\beta_1(z-\gamma_{0,0}z^2)\\
0 & 0
\end{bmatrix}. h u , v ( z ) = [ β 0 0 β 1 z 0 ] f v ( z ) = ( 1 − γ 0 , 0 z ) ( 1 − γ 1 , 1 z ) 1 [ β 0 ( 1 − γ 1 , 1 z ) 0 β 0 γ 0 , 1 + β 1 ( z − γ 0 , 0 z 2 ) 0 ] .
For the following, we define K G : = R ( Φ , Ω , z ) \mathbb{K}_{\mathcal{G}}:=\mathbb{R}(\Phi,\Omega,z) K G := R ( Φ , Ω , z ) to be the field of rational functions with indeterminates Φ , Ω \Phi,\Omega Φ , Ω and z z z , where Φ \Phi Φ is the vector of indeterminates where each entry corresponds to a unique link on G \mathcal{G} G . Similarly, we treat Ω \Omega Ω as a vector of indeterminates.
Remark 3. In the next section, it will be useful to consider the frequency domain objects associated with a time series graph G \mathcal{G} G as parameterized rational functions, i.e.,
H , f , S I , S ∈ BMat ( K G , V , [ 0 , P − 1 ] ) . H,f,\mathcal{S}^{I},\mathcal{S}\in\operatorname{BMat}(\mathbb{K}_{\mathcal{G}},V,[0,P-1]). H , f , S I , S ∈ BMat ( K G , V , [ 0 , P − 1 ]) .
RP-stationary frequency representation of a P-stationary SVAR process
Let ( G , Φ , η ) (\mathcal{G},\Phi,\eta) ( G , Φ , η ) be an P P P -stationary SVAR model. The goal of this subsection is to show that the associated P P P -stationary spectrum S \mathcal{S} S can be linearly recovered (over R ( z ) \mathbb{R}(z) R ( z ) ) from its R P RP R P -stationary embedding E R ( S ) \operatorname{E}_{R}(\mathcal{S}) E R ( S ) . This linear relationship will be central in the derivation of the identifiability criterion for link functions in P P P -stationary SVAR processes under temporally incomplete observations.
To capture the general structure of the embedded spectrum E R ( S ) E_R(S) E R ( S ) , we define the following matrices over R ( z ) \mathbb{R}(z) R ( z ) . For a non-zero r ∈ R ( z ) r \in\mathbb{R}(z) r ∈ R ( z ) and P × P P \times P P × P -dimensional matrices A 0 , … , A R − 1 A_0,\ldots,A_{R-1} A 0 , … , A R − 1 with entries in R ( z ) \mathbb{R}(z) R ( z ) , we define the block-matrix
A ( A 0 , … , A R − 1 ) : = [ A i , j ] i , j ∈ [ 0 , R − 1 ] , A i , j : = { A j − i j ≥ i , r A P − i + j j < i . (12) \mathcal{A}(A_0,\ldots,A_{R-1}) := [A_{i,j}]_{i,j\in[0,R-1]}, \qquad A_{i,j} := \begin{cases} A_{j-i} & j \ge i, \\ rA_{P-i+j} & j < i. \end{cases} \tag*{(12)} A ( A 0 , … , A R − 1 ) := [ A i , j ] i , j ∈ [ 0 , R − 1 ] , A i , j := { A j − i r A P − i + j j ≥ i , j < i . (12)
We will denote the set of all block-matrices that possess the structure (12) by T r ( R , P ) ⊂ BMat ( R ( z ) , [ 0 , R − 1 ] , [ 0 , P − 1 ] ) \mathcal{T}_r(R,P) \subset\operatorname{BMat}(\mathbb{R}(z),[0,R-1],[0,P-1]) T r ( R , P ) ⊂ BMat ( R ( z ) , [ 0 , R − 1 ] , [ 0 , P − 1 ]) . The following proposition states some useful properties of T r ( R , P ) \mathcal{T}_r(R,P) T r ( R , P ) . We provide the proof in the supplementary material.
Proposition 9. The set T r ( R , P ) \mathcal{T}_r(R,P) T r ( R , P ) is multiplicative and additive, i.e., for B , C ∈ T r ( R , P ) \mathbf{B},\mathbf{C} \in\mathcal{T}_r(R,P) B , C ∈ T r ( R , P ) , it holds that B + C , B C ∈ T r ( R , P ) \mathbf{B}+\mathbf{C},\mathbf{B}\mathbf{C} \in\mathcal{T}_r(R,P) B + C , BC ∈ T r ( R , P ) . Furthermore, if B \mathbf{B} B is invertible, then B − 1 ∈ T r ( R , P ) \mathbf{B}^{-1} \in\mathcal{T}_r(R,P) B − 1 ∈ T r ( R , P ) .
For A = [ A i , j ] i , j ∈ [ 0 , R − 1 ] ∈ T ( R , P ) \mathbf{A}=[A_{i,j}]_{i,j\in[0,R-1]}\in\mathcal{T}(R,P) A = [ A i , j ] i , j ∈ [ 0 , R − 1 ] ∈ T ( R , P ) , we define A ⊤ R = [ A j , i ] i , j ∈ [ 0 , R − 1 ] \mathbf{A}^{\top_R}=[A_{j,i}]_{i,j\in[0,R-1]} A ⊤ R = [ A j , i ] i , j ∈ [ 0 , R − 1 ] to be the block-transpose, i.e., the arrangement of the blocks is transposed while the arrangement of the entries within the blocks is left unchanged.
Lemma 10. For a positive integer P P P , and a non-zero r ∈ R ( z ) r\in\mathbb{R}(z) r ∈ R ( z ) , there are the following bijections
( ∙ ) ⊤ R : T r ( R , P ) ⟶ T r − 1 ( R , P ) , A ↦ A ⊤ R , ( ∙ ) ∗ : T r ( R , P ) ⟶ T r ∗ ( R , P ) , A ↦ A ∗ , (\bullet)^{\top_R}:\mathcal{T}_r(R,P)\longrightarrow\mathcal{T}_{r^{-1}}(R,P),\quad A\mapsto A^{\top_R},\qquad(\bullet)^*:\mathcal{T}_r(R,P)\longrightarrow\mathcal{T}_{r^*}(R,P),\quad A\mapsto A^*, ( ∙ ) ⊤ R : T r ( R , P ) ⟶ T r − 1 ( R , P ) , A ↦ A ⊤ R , ( ∙ ) ∗ : T r ( R , P ) ⟶ T r ∗ ( R , P ) , A ↦ A ∗ , where transposition induces an anti-isomorphism and conjugation induces an isomorphism (w.r.t. to matrix multiplication).
For k ∈ Z k\in\mathbb{Z} k ∈ Z , we will abbreviate T ( z ↦ z k ) ( R , P ) \mathcal{T}_{(z\mapsto z^k)}(R,P) T ( z ↦ z k ) ( R , P ) by T k ( R , P ) \mathcal{T}_k(R,P) T k ( R , P ) . In the following, we will call a matrix F ∈ BMat ( R ( z ) , V , [ 0 , R P − 1 ] ) F\in\operatorname{BMat}(\mathbb{R}(z),V,[0,RP-1]) F ∈ BMat ( R ( z ) , V , [ 0 , R P − 1 ]) block-wise T r ( R , P ) \mathcal{T}_r(R,P) T r ( R , P ) if for every u , v ∈ V u,v\in V u , v ∈ V , it holds that F u , v ∈ T r ( R , P ) F_{u,v}\in\mathcal{T}_r(R,P) F u , v ∈ T r ( R , P ) . For the following proposition, note that the identities in (7) imply that the embedded link functions E R ( H ) E_R(H) E R ( H ) and internal functions E R ( f ) E_R(f) E R ( f ) are matrices over R ( z ) \mathbb{R}(z) R ( z ) .
Proposition 11. The embedded matrix of link function E R ( H ) E_R(H) E R ( H ) , the embedded internal spectrum E R ( S I ) E_R(S^I) E R ( S I ) , and the embedded spectrum E R ( S ) E_R(S) E R ( S ) are block-wise T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) .
Proof. That E R ( H ) E_R(H) E R ( H ) is block-wise T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) is a direct consequence of Lemma 8. That the embedded internal spectrum is block diagonal follows from the structure of the filter ι \iota ι and the fact that the components of the white noise process η \eta η are mutually independent. To see that each block on the diagonal is in T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) , fix some v ∈ V v\in V v ∈ V . First, we apply Proposition 4 to Equation (2) and then apply the Fourier transform and use its basic properties mentioned at the beginning of this section to get the following factorization:
e R ( S v I ) = e R ( f v ⊤ ) e R ( Ω ) e R ( f v ) ∗ . e_R(S^I_v)=e_R(f_v^\top)e_R(\Omega)e_R(f_v)^*. e R ( S v I ) = e R ( f v ⊤ ) e R ( Ω ) e R ( f v ) ∗ . It follows from Lemma 8 and Lemma 10 that every factor in this factorization is an element in T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) . Hence, it follows from Proposition 9 that E R ( S I ) E_R(S^I) E R ( S I ) is block-wise T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) .
It remains to show that E R ( S ) E_R(S) E R ( S ) is block-wise T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) . For the following, we set H ~ : = ( I − H ) − 1 \widetilde{H}:=(I-H)^{-1} H := ( I − H ) − 1 . Since the embedding E R ( ∙ ) E_R(\bullet) E R ( ∙ ) commutes with taking inverse, it follows from what we have just shown, and the Cayley-Hamilton theorem combined with Proposition 9 that E R ( H ~ ⊤ ) E_R(\widetilde{H}^{\top}) E R ( H ⊤ ) and E R ( H ~ ∗ ) E_R(\widetilde{H}^*) E R ( H ∗ ) are block-wise T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) . To see that the embedded spectrum E R ( S ) E_R(S) E R ( S ) is block-wise T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) , we factorize
E R ( S ) = E R ( H ~ ⊤ ) E R ( S I ) E R ( H ~ ∗ ) . (13) E_R(S)=E_R(\widetilde{H}^{\top})E_R(S^I)E_R(\widetilde{H}^*). \tag*{(13)} E R ( S ) = E R ( H ⊤ ) E R ( S I ) E R ( H ∗ ) . (13) Since each factor is block-wise T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) , it follows from Proposition 9 that the product, i.e., E R ( S ) E_R(S) E R ( S ) is block-wise T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) .
Example 5. Considering the same setting and notation as in Example 2, the embedded internal functions for the process u u u and v v v , respectively, are
e 2 ( f u ) ( z ) = [ 1 − α 0 0 0 0 1 − α 1 0 0 0 1 − α 0 − α 1 z 0 0 1 ] − 1 , e 2 ( f v ) ( z ) = [ 1 0 − γ 0 , 0 − γ 0 , 1 0 1 0 − γ 1 , 1 − γ 0 , 0 z − γ 0 , 1 z 1 0 0 − γ 1 , 1 z 0 1 ] − 1 . e_2(f_u)(z)=
\begin{bmatrix}
1 & -\alpha_0 & 0 & 0\\
0 & 1 & -\alpha_1 & 0\\
0 & 0 & 1 & -\alpha_0\\
-\alpha_1 z & 0 & 0 & 1
\end{bmatrix}^{-1},
\qquad
e_2(f_v)(z)=
\begin{bmatrix}
1 & 0 & -\gamma_{0,0} & -\gamma_{0,1}\\
0 & 1 & 0 & -\gamma_{1,1}\\
-\gamma_{0,0}z & -\gamma_{0,1}z & 1 & 0\\
0 & -\gamma_{1,1}z & 0 & 1
\end{bmatrix}^{-1}. e 2 ( f u ) ( z ) = 1 0 0 − α 1 z − α 0 1 0 0 0 − α 1 1 0 0 0 − α 0 1 − 1 , e 2 ( f v ) ( z ) = 1 0 − γ 0 , 0 z 0 0 1 − γ 0 , 1 z − γ 1 , 1 z − γ 0 , 0 0 1 0 − γ 0 , 1 − γ 1 , 1 0 1 − 1 .
We will now establish a linear relationship between the P P P -stationary spectrum and its R P RP R P -stationary embedding. First, recall that the shift and inflation operators sh j \operatorname{sh}^{j} sh j and infl R \operatorname{infl}_{R} infl R map rational functions to rational functions. Hence, entry-wise application yield the following homomorphisms of rings, i.e., they are compatible with matrix addition and multiplication:
sh j : T k ( R , P ) → T k ( R , P ) , infl R : T k ( R , P ) → T k R ( R , P ) . \operatorname{sh}^{j}: \mathcal{T}_{k}(R,P) \to\mathcal{T}_{k}(R,P), \qquad\operatorname{infl}_{R}: \mathcal{T}_{k}(R,P) \to\mathcal{T}_{kR}(R,P). sh j : T k ( R , P ) → T k ( R , P ) , infl R : T k ( R , P ) → T k R ( R , P ) .
To state the linear relationship in a way that will be useful in the next section, we define for a given set I ⊂ [ 0 , R − 1 ] 2 I \subset[0,R-1]^{2} I ⊂ [ 0 , R − 1 ] 2 the associated mapping
φ I : I → [ 0 , R − 1 ] , ( i , j ) ↦ ( j − i ) 1 i ≤ j ( i , j ) + ( R − i + j ) 1 i > j ( i , j ) . (14) \varphi_{I}: I \to[0,R-1], \qquad(i,j) \mapsto(j-i)\mathbf{1}_{i\leq j}(i,j) + (R-i+j)\mathbf{1}_{i>j}(i,j). \tag*{(14)} φ I : I → [ 0 , R − 1 ] , ( i , j ) ↦ ( j − i ) 1 i ≤ j ( i , j ) + ( R − i + j ) 1 i > j ( i , j ) . (14)
Proposition 12. Let u , v ∈ V u,v \in V u , v ∈ V and I ⊂ [ 0 , R − 1 ] 2 I \subset[0,R-1]^{2} I ⊂ [ 0 , R − 1 ] 2 be such that the mapping φ I \varphi_{I} φ I is a bijection, then for S ~ = e R ( S u , v ) \widetilde{S} = e_{R}(S_{u,v}) S = e R ( S u , v ) , it holds that
S u , v = ∑ ( i , j ) ∈ I sh j − i ∘ infl R ( S ~ i , j ) . S_{u,v} = \sum_{(i,j)\in I} \operatorname{sh}^{j-i} \circ\operatorname{infl}_{R}(\widetilde{S}_{i,j}). S u , v = ( i , j ) ∈ I ∑ sh j − i ∘ infl R ( S i , j ) .
Proof. First, it follows from Lemma 1 and the two identities in (8) that
S u , v = ∑ j = 0 R − 1 sh j ∘ infl R ( S ~ 0 , j ) . (15) S_{u,v} = \sum_{j=0}^{R-1} \operatorname{sh}^{j} \circ\operatorname{infl}_{R}(\widetilde{S}_{0,j}). \tag*{(15)} S u , v = j = 0 ∑ R − 1 sh j ∘ infl R ( S 0 , j ) . (15)
Secondly, let us define I ≤ = { ( i , j ) ∈ I ∣ i ≤ j } I_{\leq} = \{(i,j) \in I \mid i \leq j\} I ≤ = {( i , j ) ∈ I ∣ i ≤ j } and I > = { ( i , j ) ∈ I ∣ i > j } I_{>} = \{(i,j) \in I \mid i > j\} I > = {( i , j ) ∈ I ∣ i > j } . Note that it follows from Lemma 8 that for ( i , j ) ∈ I ≤ (i,j) \in I_{\leq} ( i , j ) ∈ I ≤ , it holds that S ~ i , j = S ~ 0 , j − i \widetilde{S}_{i,j} = \widetilde{S}_{0,j-i} S i , j = S 0 , j − i and for ( i , j ) ∈ I > (i,j) \in I_{>} ( i , j ) ∈ I > , it holds that S ~ i , j = sh ( S ~ 0 , R − i + j ) \widetilde{S}_{i,j} = \operatorname{sh}(\widetilde{S}_{0,R-i+j}) S i , j = sh ( S 0 , R − i + j ) . Using the assumption on I I I , the identity infl R ∘ sh j = sh R j ∘ infl R \operatorname{infl}_{R} \circ\operatorname{sh}^{j} = \operatorname{sh}^{Rj} \circ\operatorname{infl}_{R} infl R ∘ sh j = sh R j ∘ infl R and the decomposition (15) , we compute
S u , v = ∑ ( i , j ) ∈ I ≤ sh j − i ∘ infl R ( S ~ 0 , j − i ) + ∑ ( i , j ) ∈ I > sh R − i + j ∘ infl R ( S ~ 0 , R − i + j ) = ∑ ( i , j ) ∈ I ≤ sh j − i ∘ infl R ( S ~ i , j ) + ∑ ( i , j ) ∈ I > sh R − i + j ∘ infl R ( sh − 1 ( S ~ i , j ) ) = ∑ ( i , j ) ∈ I sh j − i ( S ~ i , j ) . \begin{aligned}
S_{u,v}
&= \sum_{(i,j)\in I_{\leq}} \operatorname{sh}^{j-i}\circ\operatorname{infl}_{R}(\widetilde{S}_{0,j-i}) + \sum_{(i,j)\in I_{>}} \operatorname{sh}^{R-i+j}\circ\operatorname{infl}_{R}(\widetilde{S}_{0,R-i+j}) \\
&= \sum_{(i,j)\in I_{\leq}} \operatorname{sh}^{j-i}\circ\operatorname{infl}_{R}(\widetilde{S}_{i,j}) + \sum_{(i,j)\in I_{>}} \operatorname{sh}^{R-i+j}\circ\operatorname{infl}_{R}(\operatorname{sh}^{-1}(\widetilde{S}_{i,j})) \\
&= \sum_{(i,j)\in I} \operatorname{sh}^{j-i}(\widetilde{S}_{i,j}).
\end{aligned} S u , v = ( i , j ) ∈ I ≤ ∑ sh j − i ∘ infl R ( S 0 , j − i ) + ( i , j ) ∈ I > ∑ sh R − i + j ∘ infl R ( S 0 , R − i + j ) = ( i , j ) ∈ I ≤ ∑ sh j − i ∘ infl R ( S i , j ) + ( i , j ) ∈ I > ∑ sh R − i + j ∘ infl R ( sh − 1 ( S i , j )) = ( i , j ) ∈ I ∑ sh j − i ( S i , j ) .
Identification under temporally incomplete observations
Throughout this section, ( G , Φ , η ) (\mathcal{G},\Phi,\eta) ( G , Φ , η ) denotes a P P P -stationary SVAR process whose components are indexed by a finite set V V V , and Φ ∈ R s t b G \Phi\in\mathbb{R}^{G}_{\mathrm{stb}} Φ ∈ R stb G . First, we make precise what we mean by temporally incomplete. Then we will derive a simple criterion to check whether a given entry in the P P P -stationary spectrum can be recovered from a given temporally incomplete observation. Furthermore, we provide a graphical criterion to determine the invertibility of a given subspectrum, which is in fashion of [44 , 36 ] , and which we will use to derive a instrumental process based graphical criterion to decide the identifiability of a set of link functions under temporally incomplete observations.
Semi-periodic subsets of Z \mathbb{Z} Z
Definition 10 (Semi-periodic set). Let P P P be a non-negative integer. A subset O ⊂ Z O \subset\mathbb{Z} O ⊂ Z is said to be P P P -periodic if there is some m ∈ [ 0 , P − 1 ] m \in[0,P-1] m ∈ [ 0 , P − 1 ] such that O = P Z − m O = P\mathbb{Z} - m O = P Z − m . A set of integers O O O is called periodic if there exists an integer P P P such that O O O is P P P -periodic. A semi-periodic set is a finite union of periodic sets. A semi-periodic set O O O is said to have common periodicity P P P if it can be expressed as a union of P P P -periodic sets.
Given positive integers P 1 , … , P d P_{1},\ldots,P_{d} P 1 , … , P d , recall that their least common multiple is given by the smallest positive integer k k k such that k ∈ P i Z k \in P_{i}\mathbb{Z} k ∈ P i Z for every i ∈ [ d ] i \in[d] i ∈ [ d ] .
Lemma 13. Let O = ⋃ i = 1 d P i Z − m i \mathcal{O} = \bigcup_{i=1}^{d} P_i\mathbb{Z} - m_i O = ⋃ i = 1 d P i Z − m i be a semi-periodic set, where P i > 0 P_i > 0 P i > 0 and m i ∈ [ 0 , P i − 1 ] m_i \in[0, P_i - 1] m i ∈ [ 0 , P i − 1 ] for all i ∈ [ d ] i \in[d] i ∈ [ d ] . Then the least common multiple of the P i P_i P i ’s is the smallest common periodicity of O \mathcal{O} O .
Proof. The claim follows from the following simple observation. Suppose O = P Z − m \mathcal{O} = P\mathbb{Z} - m O = P Z − m is a P P P -periodic set, then O \mathcal{O} O can be written as semi-periodic set with common periodicity R P RP R P , where R R R is a positive integer. Specifically, it holds that O = ⋃ k = 0 R − 1 R P Z − m − k P \mathcal{O} = \bigcup_{k=0}^{R-1} RP\mathbb{Z} - m - kP O = ⋃ k = 0 R − 1 R P Z − m − k P .
Lemma 14. Let O = ( O v ∣ v ∈ V ) \mathcal{O} = (O_v \mid v \in V) O = ( O v ∣ v ∈ V ) be a finite collection of semi-periodic sets and let P v P_v P v be the smallest common periodicity of O v O_v O v . For P P P the least common multiple of the positive P v P_v P v ’s, it holds that P P P is the smallest number such that each O v O_v O v has common periodicity P P P .
Definition 11. Let O = ( O v ∣ v ∈ V ) \mathcal{O} = (O_v \mid v \in V) O = ( O v ∣ v ∈ V ) be a finite collection semi periodic sets. A positive integer P P P is said to be a common periodicity for O \mathcal{O} O if P P P is a common periodicity for every O v O_v O v . The minimal periodicity of O \mathcal{O} O is the smallest common periodicity.
In the following, we will represent finite collections of semi-periodic sets as follows. Let O = ( O v ∣ v ∈ V ) \mathcal{O} = (O_v \mid v \in V) O = ( O v ∣ v ∈ V ) be a collection of semi-periodic sets with common periodicity P P P , then define for every v ∈ V v \in V v ∈ V the set I v : = { i ∈ [ 0 , P − 1 ] ∣ P Z − i ⊆ O v } I_v := \{i \in[0, P - 1] \mid P\mathbb{Z} - i \subseteq O_v\} I v := { i ∈ [ 0 , P − 1 ] ∣ P Z − i ⊆ O v } . We will refer to I = ( I v ∣ v ∈ V ) \mathcal{I} = (I_v \mid v \in V) I = ( I v ∣ v ∈ V ) as the finite representation of O \mathcal{O} O with respect to the common periodicity P P P .
Example 6. Consider V = { u , v } V = \{u, v\} V = { u , v } , with semi-periodic sets
O u = 2 Z ∪ 4 Z − 1 , O v = 3 Z − 1 ∪ 2 Z − 1. O_u = 2\mathbb{Z} \cup4\mathbb{Z} - 1,\qquad O_v = 3\mathbb{Z} - 1 \cup2\mathbb{Z} - 1. O u = 2 Z ∪ 4 Z − 1 , O v = 3 Z − 1 ∪ 2 Z − 1. The minimal common periodicity of O u O_u O u is 4 and the minimal common periodicity of O v O_v O v is 6. Consequently, the minimal common periodicity for ( O u , O v ) (O_u,O_v) ( O u , O v ) is 12, giving the representation
I u = { 0 , 2 , 3 , 4 , 6 , 7 , 8 , 10 , 11 } , I v = { 1 , 2 , 3 , 5 , 7 , 8 , 9 , 11 } . I_u = \{0,2,3,4,6,7,8,10,11\},\qquad I_v = \{1,2,3,5,7,8,9,11\}. I u = { 0 , 2 , 3 , 4 , 6 , 7 , 8 , 10 , 11 } , I v = { 1 , 2 , 3 , 5 , 7 , 8 , 9 , 11 } . Spectral identifiability under temporally incomplete observations {#sec:5.2} Definition 12. Let X = ( X v ) v ∈ V \mathbf{X} = (X_v)_{v\in V} X = ( X v ) v ∈ V be a V V V -variate process. A collection of semi-periodic sets O = ( O v ∣ v ∈ V ) \mathcal{O} = (O_v \mid v \in V) O = ( O v ∣ v ∈ V ) specifies the observational structure of the process X \mathbf{X} X if X v ( t ) X_v(t) X v ( t ) is observable if and only if t ∈ O v t \in O_v t ∈ O v . The observable process is X ( O ) = ( X v ( O v ) ) v ∈ V \mathbf{X}(\mathcal{O}) = (X_v(O_v))_{v\in V} X ( O ) = ( X v ( O v ) ) v ∈ V . The observational structure O \mathcal{O} O is called temporally incomplete if there is at least one v ∈ V v \in V v ∈ V such that O v ≠ Z O_v \ne\mathbb{Z} O v = Z .
Let X \mathbf{X} X be an P P P -stationary SVAR process compatible with ( G , Φ , η ) (G,\Phi,\eta) ( G , Φ , η ) , and let O \mathcal{O} O be the observational structure for X \mathbf{X} X . Furthermore, let R P > 0 RP > 0 R P > 0 be the least common multiple of P P P and the minimal periodicity of O \mathcal{O} O . In what follows, I \mathcal{I} I will be the finite representation of O \mathcal{O} O w.r.t. the common periodicity R P RP R P .
Let u , v ∈ V u,v \in V u , v ∈ V be the indices of two component processes. To ease notation, we write S ~ : = e R ( S u , v ) \widetilde{S} := e_R(S_{u,v}) S := e R ( S u , v ) for the embedded spectrum, which is an element in T 1 ( R , P ) \mathcal{T}_1(R,P) T 1 ( R , P ) , see Proposition 11. That means, S ~ \widetilde{S} S is entirely specified by the submatrices S ~ 0 , 0 , … , S ~ R − 1 \widetilde{S}_{0,0},\ldots,\widetilde{S}_{R-1} S 0 , 0 , … , S R − 1 . If each of these blocks can be recovered from the observational subspectrum S ~ ( I ) : = [ S ~ ] I u , I v \widetilde{S}(\mathcal{I}) := [\widetilde{S}]_{I_u,I_v} S ( I ) := [ S ] I u , I v , then one can use Proposition 12 to get the P P P -stationary spectral dependence S u , v S_{u,v} S u , v .
In view of our main result, we will need to decide whether a specific entry in the P P P -stationary spectrum, i.e., S u , v ( j u , j v ) S_{u,v}^{(j_u,j_v)} S u , v ( j u , j v ) , where j u , j v ∈ [ 0 , P − 1 ] j_u,j_v \in[0,P-1] j u , j v ∈ [ 0 , P − 1 ] , can be recovered from the observational subspectrum S ~ ( I ) \widetilde{S}(\mathcal{I}) S ( I ) . For the purpose of deriving an associated criterion, we represent every i ∈ [ 0 , R P − 1 ] i \in[0,RP-1] i ∈ [ 0 , R P − 1 ] by the tuple ( ⌊ i / P ⌋ , i m o d P ) ∈ [ 0 , R − 1 ] × [ 0 , P − 1 ] (\lfloor i/P\rfloor,i \bmod P) \in[0,R-1] \times[0,P-1] (⌊ i / P ⌋ , i mod P ) ∈ [ 0 , R − 1 ] × [ 0 , P − 1 ] . Consequently, we represent the sets I u , I v I_u,I_v I u , I v as subsets of [ 0 , R − 1 ] × [ 0 , P − 1 ] [0,R-1] \times[0,P-1] [ 0 , R − 1 ] × [ 0 , P − 1 ] . Furthermore, for j ∈ [ 0 , P − 1 ] j \in[0,P-1] j ∈ [ 0 , P − 1 ] , we define the set I u ( j ) : = { i ∈ [ 0 , R − 1 ] ∣ ( i , j ) ∈ I u } I_u(j) := \{i \in[0,R-1] \mid(i,j) \in I_u\} I u ( j ) := { i ∈ [ 0 , R − 1 ] ∣ ( i , j ) ∈ I u } and, analogously, for j ∈ [ 0 , P − 1 ] j \in[0,P-1] j ∈ [ 0 , P − 1 ] , we write φ u , v ( j u , j v ) \varphi_{u,v}^{(j_u,j_v)} φ u , v ( j u , j v ) for the mapping φ I u ( j u ) × I v ( j v ) \varphi_{I_u(j_u)\times I_v(j_v)} φ I u ( j u ) × I v ( j v ) , as defined in (15) . Note that it follows from Proposition 12 that the entry S u , v ( j u , j v ) S_{u,v}^{(j_u,j_v)} S u , v ( j u , j v ) can be recovered if φ u , v ( j u , j v ) \varphi_{u,v}^{(j_u,j_v)} φ u , v ( j u , j v ) maps surjectively into [ 0 , R − 1 ] [0,R-1] [ 0 , R − 1 ] . Based on this observation, we introduce the following terminology.
Definition 13. For u , v ∈ V u,v \in V u , v ∈ V and j u , j v ∈ [ 0 , R − 1 ] j_u,j_v \in[0,R-1] j u , j v ∈ [ 0 , R − 1 ] , the entry S u , v ( j u , j v ) S_{u,v}^{(j_u,j_v)} S u , v ( j u , j v ) is linearly identifiable under the observational structure I \mathcal{I} I if the mapping φ u , v ( j u , j v ) \varphi_{u,v}^{(j_u,j_v)} φ u , v ( j u , j v ) is a surjection into [ 0 , R − 1 ] [0,R-1] [ 0 , R − 1 ] . The P P P -stationary spectrum S u , v S_{u,v} S u , v is said to be identifiable if all its entries are identifiable.
Proposition 15 (Identifiability conditions). Let j u , j v ∈ [ 0 , P − 1 ] j_u,j_v \in[0,P-1] j u , j v ∈ [ 0 , P − 1 ] , then ∣ I u ( j u ) ∩ I v ( j v ) ∣ ≥ R |I_u(j_u) \cap I_v(j_v)| \ge R ∣ I u ( j u ) ∩ I v ( j v ) ∣ ≥ R is necessary and ∣ I u ( j u ) ∣ + ∣ I v ( j v ) ∣ > R |I_u(j_u)| + |I_v(j_v)| > R ∣ I u ( j u ) ∣ + ∣ I v ( j v ) ∣ > R is sufficient for identifiability of the entry S u , v ( j u , j v ) S_{u,v}^{(j_u,j_v)} S u , v ( j u , j v ) .
Invertible subspectra
Proof. We begin with the necessary condition. Denoting the image of φ u , v ( j u , j v ) \varphi_{u,v}^{(j_u,j_v)} φ u , v ( j u , j v ) by im u , v ( j u , j v ) \operatorname{im}_{u,v}^{(j_u,j_v)} im u , v ( j u , j v ) , the proposition follows from
∣ im u , v ( j u , j v ) ∣ ≤ ∣ I u ( j u ) ∣ ∣ I v ( j v ) ∣ . \left|\operatorname{im}_{u,v}^{(j_u,j_v)}\right| \leq\left|I_u(j_u)\right|\left|I_v(j_v)\right|. im u , v ( j u , j v ) ≤ ∣ I u ( j u ) ∣ ∣ I v ( j v ) ∣ .
We continue with the sufficient condition. It suffices to consider the case P = 1 P = 1 P = 1 , where I u , I v ⊂ [ 0 , R − 1 ] I_u, I_v \subset[0, R - 1] I u , I v ⊂ [ 0 , R − 1 ] , and we set φ : = φ u , v \varphi:= \varphi_{u,v} φ := φ u , v . We show the claim by contradiction. Suppose there is some l ∈ [ 0 , R − 1 ] l \in[0, R - 1] l ∈ [ 0 , R − 1 ] such that l ∉ im ( φ ) = φ ( I u , I v ) l \notin\operatorname{im}(\varphi) = \varphi(I_u, I_v) l ∈ / im ( φ ) = φ ( I u , I v ) . This means that for every ( i , j ) ∈ I u × I v (i,j) \in I_u \times I_v ( i , j ) ∈ I u × I v , we have
l ≠ φ ( i , j ) = { j − i i ≤ j , R − i + j i > j . l \ne\varphi(i,j) =
\begin{cases}
j - i & i \leq j, \\
R - i + j & i > j.
\end{cases} l = φ ( i , j ) = { j − i R − i + j i ≤ j , i > j .
To arrive at a contradiction, we define the set A = { i + l ∣ i ∈ I u ∩ [ 0 , R − l ] } A = \{i + l \mid i \in I_u \cap[0, R - l]\} A = { i + l ∣ i ∈ I u ∩ [ 0 , R − l ]} and the set B = { i + l − R ∣ i ∈ [ R − l + 1 , R − 1 ] ∩ I u } B = \{i + l - R \mid i \in[R - l + 1, R - 1] \cap I_u\} B = { i + l − R ∣ i ∈ [ R − l + 1 , R − 1 ] ∩ I u } . First, we observe that A ∩ B = ∅ A \cap B = \varnothing A ∩ B = ∅ . If there was an element in the intersection, then there would be i ∈ [ 0 , R − l ] i \in[0, R - l] i ∈ [ 0 , R − l ] and i ′ ∈ [ R − l + 1 , R − 1 ] i' \in[R - l + 1, R - 1] i ′ ∈ [ R − l + 1 , R − 1 ] such that
i + l = i ′ + l − R ⟺ R = i ′ − i . i + l = i' + l - R \Longleftrightarrow R = i' - i. i + l = i ′ + l − R ⟺ R = i ′ − i .
Since the equation on the right hand cannot hold, it follows that A ∩ B = ∅ A \cap B = \varnothing A ∩ B = ∅ . Furthermore, note that ∣ A ∪ B ∣ = ∣ I u ∣ |A \cup B| = |I_u| ∣ A ∪ B ∣ = ∣ I u ∣ and l ∉ im ( φ ) l \notin\operatorname{im}(\varphi) l ∈ / im ( φ ) means that A ∪ B ∩ I v = ∅ A \cup B \cap I_v = \varnothing A ∪ B ∩ I v = ∅ , which contradicts the assumption ∣ I v ∣ + ∣ I u ∣ = ∣ I v ∣ + ∣ A ∪ B ∣ > R |I_v| + |I_u| = |I_v| + |A \cup B| > R ∣ I v ∣ + ∣ I u ∣ = ∣ I v ∣ + ∣ A ∪ B ∣ > R . Hence, the mapping φ \varphi φ must be surjective. □
Example 7. In this example, we classify for R = 2 R = 2 R = 2 all the partially observed scenarios where the P P P -stationary S u , v S_{u,v} S u , v is identifiable. The embedded spectrum S ~ = e 2 ( S u , v ) \widetilde{S} = e_2(S_{u,v}) S = e 2 ( S u , v ) is block-structured as follows:
S ~ = [ S ~ 0 , 0 S ~ 0 , 1 S ~ 1 , 0 S ~ 1 , 1 ] , S ~ 1 , 0 = sh 1 ( S ~ 0 , 1 ) . \widetilde{S} =
\begin{bmatrix}
\widetilde{S}_{0,0} & \widetilde{S}_{0,1} \\
\widetilde{S}_{1,0} & \widetilde{S}_{1,1}
\end{bmatrix},
\qquad
\widetilde{S}_{1,0} = \operatorname{sh}^{1}(\widetilde{S}_{0,1}). S = [ S 0 , 0 S 1 , 0 S 0 , 1 S 1 , 1 ] , S 1 , 0 = sh 1 ( S 0 , 1 ) .
Using Proposition 12, we get the following expressions for the P P P -stationary spectrum:
S u , v ( z ) = S ~ 0 , 0 ( z 2 ) + z S ~ 0 , 1 ( z 2 ) = S ~ 0 , 0 ( z 2 ) + z − 1 S ~ 1 , 0 ( z 2 ) . S_{u,v}(z) = \widetilde{S}_{0,0}(z^{2}) + z\widetilde{S}_{0,1}(z^{2}) = \widetilde{S}_{0,0}(z^{2}) + z^{-1}\widetilde{S}_{1,0}(z^{2}). S u , v ( z ) = S 0 , 0 ( z 2 ) + z S 0 , 1 ( z 2 ) = S 0 , 0 ( z 2 ) + z − 1 S 1 , 0 ( z 2 ) .
Using Proposition 15 , we see that identifiability of S u , v S_{u,v} S u , v requires, for all j u , j v ∈ [ 0 , P − 1 ] j_u, j_v \in[0, P - 1] j u , j v ∈ [ 0 , P − 1 ] that ∣ I u ( j u ) ∣ ∣ I v ( j v ) ∣ ≥ 2 |I_u(j_u)||I_v(j_v)| \geq2 ∣ I u ( j u ) ∣∣ I v ( j v ) ∣ ≥ 2 . In particular, for every j u , j v ∈ [ 0 , 1 ] j_u, j_v \in[0, 1] j u , j v ∈ [ 0 , 1 ] , both I u ( j u ) I_u(j_u) I u ( j u ) and I v ( j v ) I_v(j_v) I v ( j v ) must contain at least one element while at least one of them must contain two elements. In other words, for R = 2 R = 2 R = 2 , the necessary condition and the sufficient condition in Proposition 15 are equivalent. Examples for observational structures, with R = 2 = P R = 2 = P R = 2 = P , can be found in Figure 4 , where S u , v S_{u,v} S u , v is identifiable under the observational structure depicted in a.) and b.), and not identifiable under the observational structures shown in c.) and d.).
Figure 4. For R = 2 = P R = 2 = P R = 2 = P , each figure shows an observational structure of a 2-variate process, where u i , j u_{i,j} u i , j represents the subprocess defl 4 ∘ sh 2 i + j ( X u ) \operatorname{defl}_{4} \circ\operatorname{sh}^{2i+j}(X_u) defl 4 ∘ sh 2 i + j ( X u ) , and analogously for v v v . The fully and palely colored nodes correspond to the observed and unobserved subprocesses, respectively.
Invertible subspectra
In this subsection, we will read all frequency domain objects associated with an P P P -stationary SVAR process as matrices over K G \mathbb{K}_{\mathcal{G}} K G , i.e., H , f , S ′ , S ∈ BMat ( K G , V , [ 0 , P − 1 ] ) H,f,S^{\prime},S \in\operatorname{BMat}(\mathbb{K}_{\mathcal{G}},V,[0,P-1]) H , f , S ′ , S ∈ BMat ( K G , V , [ 0 , P − 1 ]) , see Remark 3. The goal of this subsection is to derive a graphical criterion for the summary graph smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) , see Definition 9, to decide whether a given square shaped submatrix of S S S is invertible, as a matrix over K G \mathbb{K}_{\mathcal{G}} K G . In deriving the criterion, we will exploit that the process graph imposes algebraic relations in the spectrum. To capture this relationship, we establish a trek-rule similar to [44 , 36 ] for P P P -stationary processes.
For a directed path π = ( ( v i , m i ) → ( v i + 1 , m i + 1 ) ) ∈ P G ( ( v , m ) , ( w , n ) ) \pi=((v_i,m_i)\to(v_{i+1},m_{i+1}))\in\mathcal{P}_{\mathcal{G}}((v,m),(w,n)) π = (( v i , m i ) → ( v i + 1 , m i + 1 )) ∈ P G (( v , m ) , ( w , n )) , its corresponding path-function is defined as follows:
h ( π ) : = ∏ i = 1 n h v i , v i + 1 ( m i , m i + 1 ) ∈ K G . h^{(\pi)}:=\prod_{i=1}^{n}h_{v_i,v_{i+1}}^{(m_i,m_{i+1})}\in\mathbb{K}_{\mathcal{G}}. h ( π ) := i = 1 ∏ n h v i , v i + 1 ( m i , m i + 1 ) ∈ K G .
Furthermore, for a trek τ ∈ T G ∪ G I ( ( v , m ) , ( w , n ) ) \tau\in\mathcal{T}_{\mathcal{G}\cup\mathcal{G}^{\mathrm{I}}}((v,m),(w,n)) τ ∈ T G ∪ G I (( v , m ) , ( w , n )) with top ( τ ) ∈ V I × [ 0 , P − 1 ] \operatorname{top}(\tau)\in V^{\mathrm{I}}\times[0,P-1] top ( τ ) ∈ V I × [ 0 , P − 1 ] , it follows from the structure of G I \mathcal{G}^{\mathrm{I}} G I that τ \tau τ must have the following form
( v , m ) ← ⋯ ← ( u , i ) ← ( u I , k ) → ( u , j ) → ⋯ → ( w , n ) . (v,m)\leftarrow\cdots\leftarrow(u,i)\leftarrow(u^{\mathrm{I}},k)\to(u,j)\to\cdots\to(w,n). ( v , m ) ← ⋯ ← ( u , i ) ← ( u I , k ) → ( u , j ) → ⋯ → ( w , n ) .
So for any trek τ ∈ T G ∪ G I ( ( v , m ) , ( ( w , n ) ) \tau\in\mathcal{T}_{\mathcal{G}\cup\mathcal{G}^{\mathrm{I}}}((v,m),((w,n)) τ ∈ T G ∪ G I (( v , m ) , (( w , n )) , we define the associated trek-function as follows:
S ( τ ) : = { h ( π L ) f u ( k , i ) ω u ( k ) ( f u ( k , j ) h ( π R ) ) ∗ top ( τ ) = ( u I , k ) ∈ V I × [ 0 , P − 1 ] , 0 top ( τ ) ∈ V × [ 0 , P − 1 ] , (16) S^{(\tau)}:=\begin{cases}
h^{(\pi_{\mathrm{L}})}f_{u}^{(k,i)}\omega_{u}^{(k)}\left(f_{u}^{(k,j)}h^{(\pi_{\mathrm{R}})}\right)^{*} & \operatorname{top}(\tau)=(u^{\mathrm{I}},k)\in V^{\mathrm{I}}\times[0,P-1],\\
0 & \operatorname{top}(\tau)\in V\times[0,P-1],
\end{cases} \tag*{(16)} S ( τ ) := { h ( π L ) f u ( k , i ) ω u ( k ) ( f u ( k , j ) h ( π R ) ) ∗ 0 top ( τ ) = ( u I , k ) ∈ V I × [ 0 , P − 1 ] , top ( τ ) ∈ V × [ 0 , P − 1 ] , (16)
where ω u ( k ) ∈ K G \omega_{u}^{(k)}\in\mathbb{K}_{\mathcal{G}} ω u ( k ) ∈ K G is the monomial of degree one associated with the variance of η u ( k ) \eta_{u}^{(k)} η u ( k ) . The following describes the spectrum of a P P P -stationary SVAR process in terms of trek functions.
Proposition 16 (P P P -stationary trek-rule). Let G \mathcal{G} G be a P P P -stationary time series graph, and let G G G be the associated P P P -stationary process graph, and let G I G^{\mathrm{I}} G I the associated graph for the P P P -stationary internal dynamics. For every ( v , m ) , ( w , n ) ∈ V × [ 0 , P − 1 ] (v,m),(w,n)\in V\times[0,P-1] ( v , m ) , ( w , n ) ∈ V × [ 0 , P − 1 ] , it holds that
S v , w ( m , n ) = ∑ τ ∈ T G ∪ G I ( ( v , m ) , ( w , n ) ) S ( τ ) . S_{v,w}^{(m,n)}=\sum_{\tau\in\mathcal{T}_{\mathcal{G}\cup\mathcal{G}^{\mathrm{I}}}((v,m),(w,n))}S^{(\tau)}. S v , w ( m , n ) = τ ∈ T G ∪ G I (( v , m ) , ( w , n )) ∑ S ( τ ) .
Let X , Y ⊂ V × [ 0 , P − 1 ] X,Y\subset V\times[0,P-1] X , Y ⊂ V × [ 0 , P − 1 ] be (possibly intersecting) subsets with ∣ X ∣ = ∣ Y ∣ = N |X|=|Y|=N ∣ X ∣ = ∣ Y ∣ = N . A trek system [44 ] on the summary graph smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) , between X X X and Y Y Y , is a tuple of treks T = ( τ i ∣ i ∈ [ N ] ) T=(\tau_i\mid i\in[N]) T = ( τ i ∣ i ∈ [ N ]) , such that there is a bijection α : [ N ] → [ N ] \alpha:[N]\to[N] α : [ N ] → [ N ] and τ i ∈ T smry ( G ) ( x i , y α ( i ) ) \tau_i\in\mathcal{T}_{\operatorname{smry}(\mathcal{G})}(x_i,y_{\alpha(i)}) τ i ∈ T smry ( G ) ( x i , y α ( i ) ) . This system T T T is said to be without intersection if none of the τ i \tau_i τ i intersect at any vertex. In the following, T smry ( G ) ( X , Y ) \mathcal{T}_{\operatorname{smry}(\mathcal{G})}(X,Y) T smry ( G ) ( X , Y ) will denote the set of trek systems without intersection.
Proposition 17. Let G \mathcal{G} G be a P P P -stationary time series graph and smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) its summary graph. Let X , Y ⊂ V × [ 0 , P − 1 ] X,Y\subset V\times[0,P-1] X , Y ⊂ V × [ 0 , P − 1 ] be two subsets with equal cardinality, then
T smry ( G ) ( X , Y ) ≠ ∅ ⟹ det ( [ S ] X , Y ) ≠ 0. \mathcal{T}_{\operatorname{smry}(\mathcal{G})}(X,Y)\ne\varnothing\Longrightarrow\det([S]_{X,Y})\ne0. T smry ( G ) ( X , Y ) = ∅ ⟹ det ([ S ] X , Y ) = 0.
Remark 4. Whether [ S G ] X , Y [S_{\mathcal{G}}]_{X,Y} [ S G ] X , Y is invertible cannot be decided by identifying non-intersecting trek systems on the process graph G G G . We demonstrate this using { u , v } \{u,v\} { u , v } -variate 2-stationary SVAR process with the following non-zero coefficients:
ϕ u , v ( 0 , 0 ) ( 0 ) = β 0 , ϕ u , v ( 1 , 0 ) ( 1 ) = β 1 , ϕ v , v ( 0 , 1 ) ( 0 ) = γ . \phi_{u,v}^{(0,0)}(0)=\beta_0,\qquad\phi_{u,v}^{(1,0)}(1)=\beta_1,\qquad\phi_{v,v}^{(0,1)}(0)=\gamma. ϕ u , v ( 0 , 0 ) ( 0 ) = β 0 , ϕ u , v ( 1 , 0 ) ( 1 ) = β 1 , ϕ v , v ( 0 , 1 ) ( 0 ) = γ .
Consequently, the associated summary graph is specified by the set links { u 0 → v 0 , u 1 → v 0 , v 0 → v 1 } \{u_0\to v_0,u_1\to v_0,v_0\to v_1\} { u 0 → v 0 , u 1 → v 0 , v 0 → v 1 } , which means that T smry ( G ) ( { u 0 , u 1 } , { v 0 , v 1 } ) = ∅ \mathcal{T}_{\operatorname{smry}(\mathcal{G})}(\{u_0,u_1\},\{v_0,v_1\})=\varnothing T smry ( G ) ({ u 0 , u 1 } , { v 0 , v 1 }) = ∅ . On the other hand, the process graph G G G contains the links u 0 → v 0 u_0\to v_0 u 0 → v 0 and u 1 → v 1 u_1\to v_1 u 1 → v 1 , which means that T G ( { u 0 , u 1 } , { v 0 , v 1 } ) ≠ ∅ \mathcal{T}_{G}(\{u_0,u_1\},\{v_0,v_1\})\ne\varnothing T G ({ u 0 , u 1 } , { v 0 , v 1 }) = ∅ . But the link function associated with u → v u\to v u → v has rank one, as can be seen from
h u , v ( z ) = [ β 0 0 β 1 z 0 ] [ 1 γ 0 1 ] . h_{u,v}(z)=
\begin{bmatrix}
\beta_0 & 0\\
\beta_1z & 0
\end{bmatrix}
\begin{bmatrix}
1 & \gamma\\
0 & 1
\end{bmatrix}. h u , v ( z ) = [ β 0 β 1 z 0 0 ] [ 1 0 γ 1 ] .
Consequently, for X = { u } × { 0 , 1 } X=\{u\}\times\{0,1\} X = { u } × { 0 , 1 } and Y = { v } × { 0 , 1 } Y=\{v\}\times\{0,1\} Y = { v } × { 0 , 1 } , the associated subspectrum [ S ] X , Y = h u , v ⊤ Ω u h u , v ∗ [S]_{X,Y}=h_{u,v}^{\top}\Omega_u h_{u,v}^{*} [ S ] X , Y = h u , v ⊤ Ω u h u , v ∗ has rank one, i.e., it is not invertible. This shows why the criterion in Proposition 17 needs to be stated for systems of treks on smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) rather than on G G G .
Identification with generalized instrumental processes
Let O = ( O v ∣ v ∈ V ) ⊂ V × Z \mathbf{O} = (O_v \mid v \in V) \subset V \times\mathbb{Z} O = ( O v ∣ v ∈ V ) ⊂ V × Z be an observational structure, and R P RP R P the least common sampling rate that is a multiple of P P P , and I \mathbf{I} I the finite representation of O \mathbf{O} O with respect to R P RP R P .
Question 2. For a process y ∈ V y \in V y ∈ V , does the observable spectrum S ~ ( I ) \widetilde{S}(\mathbf{I}) S ( I ) provide enough information to identify the link-functions h v , y ( m , n ) h^{(m,n)}_{v,y} h v , y ( m , n ) , for all ( v , m ) ∈ pa G ( ( y , n ) ) (v,m) \in\operatorname{pa}_{G}((y,n)) ( v , m ) ∈ pa G (( y , n )) ?
To address Question 2 algebraically, we use the following notion of identifiability, which is similar to the identifiability terms used in, e.g., [19 , 44 , 36 ] .
Definition 14 (Rational identifiability). Let I \mathbf{I} I be as above, and let v , w ∈ V v,w \in V v , w ∈ V be two process indices such that ( v , i ) → ( w , j ) (v,i) \to(w,j) ( v , i ) → ( w , j ) for some i , j ∈ [ 0 , P − 1 ] i,j \in[0,P-1] i , j ∈ [ 0 , P − 1 ] . The link ( v , i ) → ( w , j ) (v,i) \to(w,j) ( v , i ) → ( w , j ) is said to be (rationally) identifiable under I \mathbf{I} I if there is a rational identification function
ξ v , w : K G I × I → K G , \xi_{v,w}: \mathbb{K}_{G}^{\mathbf{I}\times\mathbf{I}} \to\mathbb{K}_{G}, ξ v , w : K G I × I → K G , such that ξ v , w ( infl R ( S ~ ( I ) ) ) = h v , w ( i , j ) ∈ K G \xi_{v,w}(\operatorname{infl}_{R}(\widetilde{S}(\mathbf{I}))) = h^{(i,j)}_{v,w} \in\mathbb{K}_{G} ξ v , w ( infl R ( S ( I ))) = h v , w ( i , j ) ∈ K G , where S ~ ( I ) = [ E R ( S ) ] I , I \widetilde{S}(\mathbf{I}) = [E_{R}(S)]_{\mathbf{I},\mathbf{I}} S ( I ) = [ E R ( S ) ] I , I is the observational spectrum. The P P P -stationary relation v → w v \to w v → w is said to be identifiable if every link ( v , i ) → ( w , j ) (v,i) \to(w,j) ( v , i ) → ( w , j ) is identifiable.
Theorem 18. Let n ∈ [ 0 , P − 1 ] n \in[0,P-1] n ∈ [ 0 , P − 1 ] and y ∈ V y \in V y ∈ V be a process index. Suppose there is a subset X ⊂ des G ( ( y , n ) ) c X \subset\operatorname{des}_{G}((y,n))^{c} X ⊂ des G (( y , n ) ) c , i.e., non-descendants of ( y , n ) (y,n) ( y , n ) w.r.t. G G G , such that T smry ( G ) ( X , pa G ( ( y , n ) ) ) ≠ ∅ T_{\operatorname{smry}(G)}(X,\operatorname{pa}_{G}((y,n))) \ne\varnothing T smry ( G ) ( X , pa G (( y , n ))) = ∅ , and such that S x , w ( i , j ) S^{(i,j)}_{x,w} S x , w ( i , j ) is identifiable under the observational structure I \mathbf{I} I , for every ( x , i ) ∈ X (x,i) \in X ( x , i ) ∈ X and ( w , j ) ∈ pa G ( ( y , n ) ) ∪ { ( y , n ) } (w,j) \in\operatorname{pa}_{G}((y,n)) \cup\{(y,n)\} ( w , j ) ∈ pa G (( y , n )) ∪ {( y , n )} . Then for every ( w , m ) ∈ pa G ( ( y , n ) ) (w,m) \in\operatorname{pa}_{G}((y,n)) ( w , m ) ∈ pa G (( y , n )) , the link ( w , m ) → ( y , n ) (w,m) \to(y,n) ( w , m ) → ( y , n ) is rationally identifiable under the observational structure I \mathbf{I} I .
Before proving the theorem, we briefly comment on its interpretation. Theorem 18 states that for any given time series graph G G G , with summary graph smry ( G ) \operatorname{smry}(G) smry ( G ) , a link on G G G (recall that G G G is determined by smry ( G ) \operatorname{smry}(G) smry ( G ) ) is identifiable, as of Definition 14, if there is a set of instrumental (sub-)processes satisfying the conditions of Theorem 18 . Due to our notion of identifiability, it is important to note that R s t b G \mathbb{R}^{G}_{\mathrm{stb}} R stb G is an open and non-empty semi-algebraic subset of R G \mathbb{R}^{G} R G , meaning that it has non-zero Lebesgue measure. On the other hand, every SVAR processes (seen as a point in R s t b G \mathbb{R}^{G}_{\mathrm{stb}} R stb G ), for which identifiability fails even though the criterion of Theorem 18 is satisfied, belongs to the set of singular points of the identification function from Definition 14, and this set of singular points has Lebesgue measure zero. To summarize, if for a given set of links, the conditions of Theorem 18 are satisfied, then for almost every stable and compatible SVAR process, the associated identification formula, applied to the observational spectrum, yields the respective link functions.
Proof of Theorem 18 . Let us write X = { ( u i , m i ) ∣ i ∈ [ N ] } X = \{(u_i,m_i) \mid i \in[N]\} X = {( u i , m i ) ∣ i ∈ [ N ]} and W : = pa G ( ( y , n ) ) = { ( w i , q i ) ∣ i ∈ [ N ] } W := \operatorname{pa}_{G}((y,n)) = \{(w_i,q_i) \mid i \in[N]\} W := pa G (( y , n )) = {( w i , q i ) ∣ i ∈ [ N ]} . Now note that every trek τ ∈ T GUG ( ( u i , m i ) , ( y , n ) ) \tau\in\mathcal{T}_{\operatorname{GUG}}((u_i,m_i),(y,n)) τ ∈ T GUG (( u i , m i ) , ( y , n )) with non-zero trek function is of the form
( u i , m i ) ← ⋯ ← ( z , l ′ ) ← ( z ′ , l ) → ( z , l ′ ′ ) → ⋯ → ( w j , q j ) → ( y , n ) . (17) (u_i,m_i) \leftarrow\cdots\leftarrow(z,l') \leftarrow(z',l) \to(z,l'') \to\cdots\to(w_j,q_j) \to(y,n). \tag*{(17)} ( u i , m i ) ← ⋯ ← ( z , l ′ ) ← ( z ′ , l ) → ( z , l ′′ ) → ⋯ → ( w j , q j ) → ( y , n ) . (17)
Hence, the trek rule stated in Proposition 16 yields the following linear relationship
[ S ] { ( y , n ) } , X = [ H ] W , { ( y , k ) } T [ S ] W , X . [S]_{\{(y,n)\},X}=[H]^{\mathsf{T}}_{W,\{(y,k)\}}[S]_{W,X}. [ S ] {( y , n )} , X = [ H ] W , {( y , k )} T [ S ] W , X .
Using Proposition 17 with the assumption T s m r y ( G ) ( X , W ) ≠ ∅ \mathrm{T}_{\mathrm{smry}(G)}(X,W)\ne\varnothing T smry ( G ) ( X , W ) = ∅ , we conclude that
[ H ] W , ( y , k ) T = [ S ] { ( y , n ) } , X [ S ] W , X − 1 , [H]^{\mathsf{T}}_{W,(y,k)}=[S]_{\{(y,n)\},X}[S]^{-1}_{W,X}, [ H ] W , ( y , k ) T = [ S ] {( y , n )} , X [ S ] W , X − 1 ,
where by assumption all the terms on the right hand side are identifiable under I. □
Example 8. We consider the scenario of a 2-stationary process consisting of three observed processes u , v , w u,v,w u , v , w , whose causal structure is encoded by the chain u → v → w u\to v\to w u → v → w . In Figure 5 , we show three different summary graphs encoding the chain u → v → w u\to v\to w u → v → w . In Figure 6 , we show four different observational structures for the processes u , v , w u,v,w u , v , w . We will now go through these cases and check whether Theorem 18 can be applied to identify the 2-stationary effect v → w v\to w v → w .
Figure 5. Three different summary graphs for the 2-stationary causal chain u → v → w u \to v \to w u → v → w . A vertex of the form a k a_k a k represents the subprocess defl 2 ∘ sh k ( X a ) \operatorname{defl}_2 \circ\operatorname{sh}^{k}(X_a) defl 2 ∘ sh k ( X a ) .
Figure 6. Four observational structures for three 2-stationary processes u , v , w u,v,w u , v , w . A vertex of the form a k , j a_{k,j} a k , j represents the subprocess defl 2 ∘ sh 2 k + j ( X a ) \operatorname{defl}_{2}\circ\operatorname{sh}_{2k+j}(X_a) defl 2 ∘ sh 2 k + j ( X a ) . The fully and palely colored nodes represent the observed and unobserved subprocesses, respectively.
a.) At the level of the process graph G G G , we have pa G ( w i ) = { v 1 } \operatorname{pa}_{G}(w_i)=\{v_1\} pa G ( w i ) = { v 1 } , for i = 0 , 1 i=0,1 i = 0 , 1 . Under none of the observational structures, the process v v v is completely observable, so we need to use the process u u u as an instrumental process. Since P s m r y ( G ) ( u 0 , v 1 ) ≠ ∅ \mathrm{P}_{\mathrm{smry}(G)}(u_0,v_1)\ne\varnothing P smry ( G ) ( u 0 , v 1 ) = ∅ , and since by Proposition 15 , the entries S u , v ( 0 , 1 ) , S u , w ( 0 , 0 ) , S u , w ( 0 , 1 ) S^{(0,1)}_{u,v},S^{(0,0)}_{u,w},S^{(0,1)}_{u,w} S u , v ( 0 , 1 ) , S u , w ( 0 , 0 ) , S u , w ( 0 , 1 ) are identifiable under all of the observational scenarios shown in Figure 6 , we can apply Theorem 18 to infer that v → w v\to w v → w is identifiable. Specifically, the identification formulas are as follows:
h v , w ( 1 , 0 ) = S w , u ( 0 , 0 ) S v , u ( 1 , 0 ) , h^{(1,0)}_{v,w}=\frac{S^{(0,0)}_{w,u}}{S^{(1,0)}_{v,u}}, h v , w ( 1 , 0 ) = S v , u ( 1 , 0 ) S w , u ( 0 , 0 ) ,
h v , w ( 1 , 1 ) = S w , u ( 1 , 0 ) S v , u ( 1 , 0 ) . h^{(1,1)}_{v,w}=\frac{S^{(1,0)}_{w,u}}{S^{(1,0)}_{v,u}}. h v , w ( 1 , 1 ) = S v , u ( 1 , 0 ) S w , u ( 1 , 0 ) .
b.) In this case, we have that pa G ( w i ) = { v 0 , v 1 } \operatorname{pa}_{G}(w_i)=\{v_0,v_1\} pa G ( w i ) = { v 0 , v 1 } , for i = 0 , 1 i=0,1 i = 0 , 1 . Hence, we need both v 0 v_0 v 0 and v 1 v_1 v 1 as instruments for the identification of v → w v\to w v → w . However, T s m r y ( G ) ( { u 0 , u 1 } , { v 0 , v 1 } ) = ∅ \mathrm{T}_{\mathrm{smry}(G)}(\{u_0,u_1\},\{v_0,v_1\})=\varnothing T smry ( G ) ({ u 0 , u 1 } , { v 0 , v 1 }) = ∅ . Hence, Theorem 18 cannot be used to infer the identifiability of v → w v\to w v → w .
c.) In this case, it also holds that pa G ( w i ) = { v 0 , v 1 } \operatorname{pa}_{G}(w_i)=\{v_0,v_1\} pa G ( w i ) = { v 0 , v 1 } , for i = 0 , 1 i=0,1 i = 0 , 1 . But now we have that T s m r y ( G ) ( { u 0 , u 1 } , { v 0 , v 1 } ) ≠ ∅ \mathrm{T}_{\mathrm{smry}(G)}(\{u_0,u_1\},\{v_0,v_1\})\ne\varnothing T smry ( G ) ({ u 0 , u 1 } , { v 0 , v 1 }) = ∅ , as u 0 → v 1 u_0\to v_1 u 0 → v 1 and u 1 → v 0 u_1\to v_0 u 1 → v 0 yield a non-intersecting trek system. In order to apply Theorem 18 , we need an observational structure under which S v , u S_{v,u} S v , u and S w , u S_{w,u} S w , u are identifiable. This is the case only for the observational structures shown in a.) and b.) of Figure 6 . For the reasoning behind the identifiability of the spectral dependencies, the reader may review Example 7.
Example 9. In Figure 7 , there are four partially observed processes u u u , u ′ u' u ′ , v v v , w w w and one latent process l l l . The 2-stationary summary graph for those processes is depicted in Figure 7 a.). The observational structure for is shown in Figure 7 b.). Note that all of the processes are only partially observable. For the identification of v → w v \to w v → w , the process u u u , i.e., u 0 u_0 u 0 and u 1 u_1 u 1 , could serve as an instrumental process, since u 0 ← l 0 → v 0 u_0 \leftarrow l_0 \to v_0 u 0 ← l 0 → v 0 and u 1 ← l 1 ← v 1 u_1 \leftarrow l_1 \leftarrow v_1 u 1 ← l 1 ← v 1 is a non-intersecting trek system, i.e., T s m r y ( G ) ( { u 0 , u 1 } , { v 0 , v 1 } ) ≠ ∅ T_{\mathrm{smry}}(G)(\{u_0,u_1\},\{v_0,v_1\}) \ne\varnothing T smry ( G ) ({ u 0 , u 1 } , { v 0 , v 1 }) = ∅ . However, u 1 u_1 u 1 is not observable, so S u , v ( 1 , i ) S_{u,v}^{(1,i)} S u , v ( 1 , i ) is not identifiable, as of Definition 13, for i = 0 , 1 i=0,1 i = 0 , 1 . Considering the summary graph, we notice that also u 0 u_0 u 0 and u 1 ′ u'_1 u 1 ′ can be used as an instrumental processes for the identification of v → w v \to w v → w , since u 0 ← l 0 → v 0 → w 0 u_0 \leftarrow l_0 \to v_0 \to w_0 u 0 ← l 0 → v 0 → w 0 and u 1 ′ ← u 1 ← l 1 → v 1 u'_1 \leftarrow u_1 \leftarrow l_1 \to v_1 u 1 ′ ← u 1 ← l 1 → v 1 constitute a non-intersecting trek system. The terms S u , v ( 0 , i ) S_{u,v}^{(0,i)} S u , v ( 0 , i ) and S u , w ( 0 , i ) S_{u,w}^{(0,i)} S u , w ( 0 , i ) and S u ′ , v ( 1 , i ) S_{u',v}^{(1,i)} S u ′ , v ( 1 , i ) and S u ′ , w ( 1 , i ) S_{u',w}^{(1,i)} S u ′ , w ( 1 , i ) are identifiable for i = 0 , 1 i=0,1 i = 0 , 1 , see Proposition 15 . Hence, Theorem 18 implies that the 2-stationary effect v → w v \to w v → w is identifiable.
Remark 5. In Example 9, we used subprocesses of two different component processes to identify the 2-stationary effect v → w v \to w v → w , which was possible because we assumed a 2-stationary process. If the process was stationary instead of 2-stationary, and if we had the same observational structure, then we could not use Theorem 18 for the identification of the then stationary effect v → w v \to w v → w , since neither S u , v S_{u,v} S u , v nor S u ′ , v S_{u',v} S u ′ , v are identifiable, as of Definition 13. In principle, we could make use of the embedding operation, i.e., considering the equivalent 2-stationary embedding and try to use a similar algebraic structure as above (in this case, the embedded time series graph G 2 G_2 G 2 , mentioned earlier, will matter). However, this requires an adaptation of Proposition 17 (since this proposition is not stated for embedded processes), which will require additional considerations, since the embedding introduces additional relations in the link functions. This problem seems to be an interesting subject for future work.
Discussion
Summary
In this work, we introduced a linear causal model for P P P -stationary processes, i.e., non-stationary discrete-time processes having all their P P P -periodic subsamplings jointly stationary. The associated graph contains a unique vertex for every P P P -periodic subsampling of each component process. A directed link between two subsamplings of two different component processes encodes an immediate influence, and this influence is quantified by the associated link filter, which describes how the past of the subsampled driving process linearly influences the subsampled target process. In addition, every process index has an associated P P P -stationary noise process, which we called the internal dynamics. We assumed the internal dynamics of two distinct components to be independent of each other. If the link filters are sufficiently small, the data of an P P P -stationary SEP give rise to a unique P P P -stationary observational process. Then we showed that structural vector autoregressive processes, whose coefficients are allowed to vary periodically over time with periodicity P P P , belong to the class of P P P -stationary structural equation processes by constructing an P P P -stationary SEP that is equivalent to the P P P -stationary SVAR process.
Our main interest was to infer the link filters in an P P P -stationary SEP associated with an P P P -stationary SVAR process when the observations are temporally incomplete, assuming that the summary graph of the SVAR process is known. The approach we took was based on three observations: First, every observational structure can be seen as a union of R P RP R P -periodic subsamplings, for appropriately selected R R R . Second, every P P P -stationary SEP has an equivalent representation as an R P RP R P -stationary SEP. Third, the problem can equivalently be solved in the frequency domain, where the algebraic relations are simpler. The first two observations allowed us to model an P P P -stationary SVAR processes with temporally incomplete observations as an R P RP R P -stationary SEP, where every missing subsample is treated as a latent process. However, we did not infer the link filters on the embedded R P RP R P -stationary SEP. Instead, we used the third observation to derive a linear relation between between the P P P -stationary spectrum and the R P RP R P -stationary spectrum, i.e., the embedded spectrum. Based on this relationship, we could determine which entries in the P P P -stationary spectrum can be recovered under the given observational structure. Hence, we arrived at the following reformulation of the initial problem: Is a link function identifiable from those entries in the P P P -stationary spectrum that are identifiable under the given observational structure? Based on this reformulation, we could derive a simple identification criterion for an instrumental process setting.
Future work
There are several lines in which the work presented in this paper can be extended. Instrumental processes provide a simple way to identify causal effects in latent process settings. However, in the literature on structural equation modeling, i.e., the case of i.i.d. observations, several other techniques have been explored to identify effects when the latent variables confound the observed ones in more interesting ways. In future work, one could try to identify link functions when the missing observations introduce more challenges. In this analysis it might be worthwhile to take into account that temporally incomplete observations hide entries in the spectrum in a slightly different way than when entire processes are unobserved. For example, to identify the spectral dependence between two processes, it suffices that only one of them is fully observable while observations of the other can be arbitrarily rare. However, when a process is not fully observed, the corresponding entry on the diagonal of the spectrum is not identifiable, no matter how rarely observations are missing.
In the current paper, we only used the spectrum of the process, i.e., the cumulants of order two for the identification of causal effects. In ongoing work, we consider the case where the internal dynamics are filtered non-Gaussian white noise processes. Inspired by recent results in causal inference for linear non-Gaussian systems, we imagine that these ideas could also be applied to frequency domain versions of the higher order cumulants for the identification of the causal structure, i.e., the underlying directed graph, as well as the link functions, under more challenging observational structures.
Finally, in this work we assumed to know the periodicity P P P with which the coefficients in the underlying SVAR process are allowed to vary. In practice, this number P P P would have to be estimated. A possibility for future work could be to derive an estimator and an associated test for P P P based on the linear relationship between the embedded R P RP R P -stationary spectrum and the P P P -stationary spectrum.
Funding This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 883818).
Statement on AI use The author consulted AI (Claude) when looking for related research, but for nothing else.
Regarding the embedding operator
We will use the following two Lemmas to prove Lemma 1 and Lemma 2 and Proposition 4.
Lemma 19. Let λ , μ ∈ ℓ 1 ( Z ; C ) \lambda,\mu\in\ell^{1}(\mathbb{Z};\mathbb{C}) λ , μ ∈ ℓ 1 ( Z ; C ) and i ∈ Z i \in\mathbb{Z} i ∈ Z , then for convolution, it holds that
sh i ( λ ∗ μ ) = sh i ( λ ) ∗ μ = λ ∗ sh i ( μ ) , \operatorname{sh}^{i}(\lambda* \mu) = \operatorname{sh}^{i}(\lambda) * \mu= \lambda* \operatorname{sh}^{i}(\mu), sh i ( λ ∗ μ ) = sh i ( λ ) ∗ μ = λ ∗ sh i ( μ ) , and for tilted convolution, it holds that
sh i ( λ ∗ ′ μ ) = sh i ( λ ) ∗ ′ μ = λ ∗ ′ sh − i ( μ ) . \operatorname{sh}^{i}(\lambda*' \mu) = \operatorname{sh}^{i}(\lambda) *' \mu= \lambda*' \operatorname{sh}^{-i}(\mu). sh i ( λ ∗ ′ μ ) = sh i ( λ ) ∗ ′ μ = λ ∗ ′ sh − i ( μ ) . Proof. We fix an integer t ∈ Z t \in\mathbb{Z} t ∈ Z , and note that sh i ( λ ∗ μ ) ( t ) = λ ∗ μ ( t − i ) \operatorname{sh}^{i}(\lambda* \mu)(t) = \lambda* \mu(t-i) sh i ( λ ∗ μ ) ( t ) = λ ∗ μ ( t − i ) . We now show the identities by direct computation:
λ ∗ μ ( t − i ) = ∑ s ∈ Z λ ( s ) μ ( t − s − i ) = ∑ s ∈ Z λ ( s ) sh i ( μ ) ( t − s ) = λ ∗ sh i ( μ ) ( t ) . \lambda* \mu(t-i) = \sum_{s\in\mathbb{Z}} \lambda(s)\mu(t-s-i) = \sum_{s\in\mathbb{Z}} \lambda(s)\operatorname{sh}^{i}(\mu)(t-s) = \lambda* \operatorname{sh}^{i}(\mu)(t). λ ∗ μ ( t − i ) = s ∈ Z ∑ λ ( s ) μ ( t − s − i ) = s ∈ Z ∑ λ ( s ) sh i ( μ ) ( t − s ) = λ ∗ sh i ( μ ) ( t ) . Furthermore, using the index transformation s = s ′ − i s=s'-i s = s ′ − i on the right hand side of the first equation, we get that
λ ∗ μ ( t − i ) = ∑ s ′ ∈ Z λ ( s ′ − i ) μ ( t − s ′ ) = sh i ( λ ) ∗ μ ( t ) . \lambda* \mu(t-i) = \sum_{s'\in\mathbb{Z}} \lambda(s'-i)\mu(t-s') = \operatorname{sh}^{i}(\lambda) * \mu(t). λ ∗ μ ( t − i ) = s ′ ∈ Z ∑ λ ( s ′ − i ) μ ( t − s ′ ) = sh i ( λ ) ∗ μ ( t ) . Regarding tilted convolution, we compute
λ ∗ ′ μ ( t − i ) = ∑ s ∈ Z λ ( s + t − i ) μ ( s ) = sh i ( λ ) ∗ ′ μ ( t ) , \lambda*' \mu(t-i) = \sum_{s\in\mathbb{Z}} \lambda(s+t-i)\mu(s) = \operatorname{sh}^{i}(\lambda) *' \mu(t), λ ∗ ′ μ ( t − i ) = s ∈ Z ∑ λ ( s + t − i ) μ ( s ) = sh i ( λ ) ∗ ′ μ ( t ) , and using the index shift s = s ′ + i s=s'+i s = s ′ + i on the right hand side of the second equation, we get that
λ ∗ ′ μ ( t − i ) = ∑ s ′ ∈ Z λ ( s ′ + t ) μ ( s ′ + i ) = λ ∗ ′ sh − i ( μ ) ( t ) . \lambda*' \mu(t-i) = \sum_{s'\in\mathbb{Z}} \lambda(s'+t)\mu(s'+i) = \lambda*' \operatorname{sh}^{-i}(\mu)(t). λ ∗ ′ μ ( t − i ) = s ′ ∈ Z ∑ λ ( s ′ + t ) μ ( s ′ + i ) = λ ∗ ′ sh − i ( μ ) ( t ) .
Lemma 20. Let λ , μ ∈ ℓ 1 ( Z ; C ) \lambda,\mu\in\ell^{1}(\mathbb{Z};\mathbb{C}) λ , μ ∈ ℓ 1 ( Z ; C ) and let R > 0 R>0 R > 0 be a positive integer, then it holds that
defl R ( λ ∗ μ ) = ∑ m = 0 R − 1 defl R ∘ sh − m ( λ ) ∗ defl R ∘ sh m ( μ ) , \operatorname{defl}_{R}(\lambda* \mu) = \sum_{m=0}^{R-1} \operatorname{defl}_{R} \circ\operatorname{sh}^{-m}(\lambda) * \operatorname{defl}_{R} \circ\operatorname{sh}^{m}(\mu), defl R ( λ ∗ μ ) = m = 0 ∑ R − 1 defl R ∘ sh − m ( λ ) ∗ defl R ∘ sh m ( μ ) , and it also holds that
defl R ( λ ∗ ′ μ ) = ∑ m = 0 R − 1 defl R ∘ sh − m ( λ ) ∗ ′ defl R ∘ sh − m ( μ ) . \operatorname{defl}_{R}(\lambda*' \mu) = \sum_{m=0}^{R-1} \operatorname{defl}_{R} \circ\operatorname{sh}^{-m}(\lambda) *' \operatorname{defl}_{R} \circ\operatorname{sh}^{-m}(\mu). defl R ( λ ∗ ′ μ ) = m = 0 ∑ R − 1 defl R ∘ sh − m ( λ ) ∗ ′ defl R ∘ sh − m ( μ ) . Proof. For t ∈ Z t \in\mathbb{Z} t ∈ Z , we show the claim by direct computation:
λ ∗ μ ( t R ) = ∑ s ∈ Z λ ( s ) μ ( t R − s ) = ∑ m = 0 R − 1 ∑ s ′ ∈ Z λ ( s ′ R + m ) μ ( ( t − s ′ ) R − m ) = ∑ m = 0 R − 1 ∑ s ′ ∈ Z defl R ∘ sh − m ( λ ) ( s ′ ) defl R ∘ sh m ( μ ) ( t − s ′ ) = ∑ m = 0 R − 1 defl R ∘ sh − m ( λ ) ∗ defl R ∘ sh m ( μ ) ( t ) . \begin{aligned}
\lambda* \mu(tR) &= \sum_{s\in\mathbb{Z}} \lambda(s)\mu(tR-s) \\
&= \sum_{m=0}^{R-1}\sum_{s'\in\mathbb{Z}} \lambda(s'R+m)\mu((t-s')R-m) \\
&= \sum_{m=0}^{R-1}\sum_{s'\in\mathbb{Z}} \operatorname{defl}_{R}\circ\operatorname{sh}^{-m}(\lambda)(s')\operatorname{defl}_{R}\circ\operatorname{sh}^{m}(\mu)(t-s') \\
&= \sum_{m=0}^{R-1}\operatorname{defl}_{R}\circ\operatorname{sh}^{-m}(\lambda)*\operatorname{defl}_{R}\circ\operatorname{sh}^{m}(\mu)(t).
\end{aligned} λ ∗ μ ( tR ) = s ∈ Z ∑ λ ( s ) μ ( tR − s ) = m = 0 ∑ R − 1 s ′ ∈ Z ∑ λ ( s ′ R + m ) μ (( t − s ′ ) R − m ) = m = 0 ∑ R − 1 s ′ ∈ Z ∑ defl R ∘ sh − m ( λ ) ( s ′ ) defl R ∘ sh m ( μ ) ( t − s ′ ) = m = 0 ∑ R − 1 defl R ∘ sh − m ( λ ) ∗ defl R ∘ sh m ( μ ) ( t ) .
Now consider the tilted convolution
λ ∗ ′ μ ( t R ) = ∑ s ∈ Z λ ( s + t R ) μ ( s ) = ∑ m = 0 R − 1 ∑ s ′ ∈ Z λ ( m + ( s ′ + t ) R ) μ ( s ′ R + m ) = ∑ m = 0 R − 1 defl R ∘ sh − m ( λ ) ∗ ′ defl R ∘ sh − m ( μ ) ( t ) . \begin{aligned}
\lambda*' \mu(tR) = \sum_{s\in\mathbb{Z}} \lambda(s+tR)\mu(s) \\
&= \sum_{m=0}^{R-1}\sum_{s'\in\mathbb{Z}} \lambda(m+(s'+t)R)\mu(s'R+m) \\
&= \sum_{m=0}^{R-1} \operatorname{defl}_{R}\circ\operatorname{sh}^{-m}(\lambda) *' \operatorname{defl}_{R}\circ\operatorname{sh}^{-m}(\mu)(t).
\end{aligned} λ ∗ ′ μ ( tR ) = s ∈ Z ∑ λ ( s + tR ) μ ( s ) = m = 0 ∑ R − 1 s ′ ∈ Z ∑ λ ( m + ( s ′ + t ) R ) μ ( s ′ R + m ) = m = 0 ∑ R − 1 defl R ∘ sh − m ( λ ) ∗ ′ defl R ∘ sh − m ( μ ) ( t ) .
Proof of Lemma 1. We need to show Equation (1) . Let us define λ ~ : = e R ( λ ) \tilde{\lambda}:=e_R(\lambda) λ ~ := e R ( λ ) , and let λ ~ i , j : = defl R ∘ sh i − j ( λ ) \tilde{\lambda}_{i,j}:=\operatorname{defl}_{R}\circ\operatorname{sh}^{i-j}(\lambda) λ ~ i , j := defl R ∘ sh i − j ( λ ) . First, note that for i , j ∈ [ 0 , R − 2 ] i,j\in[0,R-2] i , j ∈ [ 0 , R − 2 ] , it holds that
λ ~ i , j = λ ~ i + 1 , j + 1 , λ ~ R − 1 , j − 1 = sh 1 ( λ ~ 0 , j ) , \tilde{\lambda}_{i,j}=\tilde{\lambda}_{i+1,j+1},
\qquad
\tilde{\lambda}_{R-1,j-1}=\operatorname{sh}^{1}(\tilde{\lambda}_{0,j}), λ ~ i , j = λ ~ i + 1 , j + 1 , λ ~ R − 1 , j − 1 = sh 1 ( λ ~ 0 , j ) ,
and note that for any sequence λ \lambda λ it holds that
( sh j ∘ str R ( λ ) ) ( k ) = str R ( λ ) ( k − j ) = 1 R Z ( k − j ) λ ( 1 R Z ( k − j ) k − j R ) . \left(\operatorname{sh}^{j}\circ\operatorname{str}_{R}(\lambda)\right)(k)
=\operatorname{str}_{R}(\lambda)(k-j)
=\mathbf{1}_{R\mathbb{Z}}(k-j)\lambda\left(\mathbf{1}_{R\mathbb{Z}}(k-j)\frac{k-j}{R}\right). ( sh j ∘ str R ( λ ) ) ( k ) = str R ( λ ) ( k − j ) = 1 R Z ( k − j ) λ ( 1 R Z ( k − j ) R k − j ) .
In particular,
( sh j ∘ str R ( λ ~ 0 , j ) ) ( k ) = 1 R Z ( k − j ) λ ~ 0 , j ( k − j R ) = 1 R Z ( k − j ) ( defl R ∘ sh − j ( λ ) ) ( 1 R Z ( k − j ) k − j R ) = 1 R Z ( k − j ) λ ( k ) . \begin{aligned}
\left(\operatorname{sh}^{j}\circ\operatorname{str}_{R}(\tilde{\lambda}_{0,j})\right)(k)
=\mathbf{1}_{R\mathbb{Z}}(k-j)\tilde{\lambda}_{0,j}\left(\frac{k-j}{R}\right) \\
&=\mathbf{1}_{R\mathbb{Z}}(k-j)(\operatorname{defl}_{R}\circ\operatorname{sh}^{-j}(\lambda))\left(\mathbf{1}_{R\mathbb{Z}}(k-j)\frac{k-j}{R}\right) \\
&=\mathbf{1}_{R\mathbb{Z}}(k-j)\lambda(k).
\end{aligned} ( sh j ∘ str R ( λ ~ 0 , j ) ) ( k ) = 1 R Z ( k − j ) λ ~ 0 , j ( R k − j ) = 1 R Z ( k − j ) ( defl R ∘ sh − j ( λ )) ( 1 R Z ( k − j ) R k − j ) = 1 R Z ( k − j ) λ ( k ) .
This implies the claimed identity for i = 0 i=0 i = 0 , i.e.,
μ = ∑ j = 0 R − 1 sh j ∘ str R ( λ ~ 0 , j ) . (18) \mu=\sum_{j=0}^{R-1}\operatorname{sh}^{j}\circ\operatorname{str}_{R}(\tilde{\lambda}_{0,j}). \tag*{(18)} μ = j = 0 ∑ R − 1 sh j ∘ str R ( λ ~ 0 , j ) . (18)
Before we proceed with i > 0 i>0 i > 0 , we note that for any i , j , R ∈ Z i,j,R\in\mathbb{Z} i , j , R ∈ Z with R > 0 R>0 R > 0 , the following relations hold:
sh i ∘ sh j = sh i + j , str R ∘ sh j = sh j R ∘ str R . (19) \operatorname{sh}^{i}\circ\operatorname{sh}^{j}=\operatorname{sh}^{i+j},
\qquad
\operatorname{str}_{R}\circ\operatorname{sh}^{j}=\operatorname{sh}^{jR}\circ\operatorname{str}_{R}. \tag*{(19)} sh i ∘ sh j = sh i + j , str R ∘ sh j = sh j R ∘ str R . (19)
Furthermore, note that sh j \operatorname{sh}^{j} sh j and str R \operatorname{str}_{R} str R are linear operators on the space of filters. Using the two identities in (18), we proceed with i > 0 i>0 i > 0 by direct computation
∑ j = 0 R − 1 sh j ∘ str R ( λ ~ i , j ) = ∑ j = 0 i − 1 sh j ∘ str R ( sh ( λ ~ 0 , R − i + j ) ) + ∑ j = i R − 1 sh j ∘ str R ( λ ~ 0 , j − i ) = ∑ j = 0 i − 1 sh j + R ∘ str R ( λ ~ 0 , R − i + j ) + ∑ j = i R − 1 sh j + ( j − i ) ∘ str R ( λ ~ 0 , j − i ) = sh i ( ∑ j = 0 i − 1 sh R − i + j ∘ str R ( λ ~ 0 , R − i + j ) + ∑ j = i R − 1 sh j − i ∘ str R ( λ ~ 0 , j − i ) ) = sh i ( λ ~ ) . \begin{aligned}
\sum_{j=0}^{R-1}\operatorname{sh}^{j}\circ\operatorname{str}_{R}(\tilde{\lambda}_{i,j})
&=\sum_{j=0}^{i-1}\operatorname{sh}^{j}\circ\operatorname{str}_{R}\left(\operatorname{sh}(\tilde{\lambda}_{0,R-i+j})\right)
+\sum_{j=i}^{R-1}\operatorname{sh}^{j}\circ\operatorname{str}_{R}(\tilde{\lambda}_{0,j-i}) \\
&=\sum_{j=0}^{i-1}\operatorname{sh}^{j+R}\circ\operatorname{str}_{R}(\tilde{\lambda}_{0,R-i+j})
+\sum_{j=i}^{R-1}\operatorname{sh}^{j+(j-i)}\circ\operatorname{str}_{R}(\tilde{\lambda}_{0,j-i}) \\
&=\operatorname{sh}^{i}\left(\sum_{j=0}^{i-1}\operatorname{sh}^{R-i+j}\circ\operatorname{str}_{R}(\tilde{\lambda}_{0,R-i+j})
+\sum_{j=i}^{R-1}\operatorname{sh}^{j-i}\circ\operatorname{str}_{R}(\tilde{\lambda}_{0,j-i})\right) \\
&=\operatorname{sh}^{i}(\tilde{\lambda}).
\end{aligned} j = 0 ∑ R − 1 sh j ∘ str R ( λ ~ i , j ) = j = 0 ∑ i − 1 sh j ∘ str R ( sh ( λ ~ 0 , R − i + j ) ) + j = i ∑ R − 1 sh j ∘ str R ( λ ~ 0 , j − i ) = j = 0 ∑ i − 1 sh j + R ∘ str R ( λ ~ 0 , R − i + j ) + j = i ∑ R − 1 sh j + ( j − i ) ∘ str R ( λ ~ 0 , j − i ) = sh i ( j = 0 ∑ i − 1 sh R − i + j ∘ str R ( λ ~ 0 , R − i + j ) + j = i ∑ R − 1 sh j − i ∘ str R ( λ ~ 0 , j − i ) ) = sh i ( λ ~ ) .
For the last equality, we used (18) and the index transformation j ′ = j + R − i j'=j+R-i j ′ = j + R − i for the first sum and j ′ = j + i j'=j+i j ′ = j + i for the second sum.
Proof of Lemma 2. We begin with the identity for convolution. Let us define λ ~ : = e R ( λ ) \tilde{\lambda}:=e_R(\lambda) λ ~ := e R ( λ ) , and μ ~ = e R ( μ ) \tilde{\mu}=e_R(\mu) μ ~ = e R ( μ ) and γ ~ = e R ( γ : = λ ∗ μ ) \tilde{\gamma}=e_R(\gamma:=\lambda*\mu) γ ~ = e R ( γ := λ ∗ μ ) . Furthermore, we set λ ~ i , j = defl R ∘ sh i − j ( λ ) \tilde{\lambda}_{i,j}=\operatorname{defl}_{R}\circ\operatorname{sh}^{i-j}(\lambda) λ ~ i , j = defl R ∘ sh i − j ( λ ) and $\tilde{\mu}_{i,j}=\operatorname{defl}_{R}\circ sh i − j ( μ ) \operatorname{sh}^{i-j}(\mu) sh i − j ( μ ) and γ ~ i , j = defl R ∘ sh i − j ( γ ) \widetilde{\gamma}_{i,j} = \operatorname{defl}_{R} \circ\operatorname{sh}^{i-j}(\gamma) γ i , j = defl R ∘ sh i − j ( γ ) as the i , j i,j i , j -th block entry in λ ~ \widetilde{\lambda} λ resp. μ ~ \widetilde{\mu} μ and γ ~ \widetilde{\gamma} γ . We show the claimed identity by direct computation:
γ ~ i , j = defl R ∘ sh i − j ( λ ∗ μ ) = defl R ( sh i ( λ ) ∗ sh − j ( μ ) ) = ∑ m = 0 R − 1 defl R ∘ sh − m ( sh i ( λ ) ) ∗ defl R ∘ sh m ( sh − j ( μ ) ) = ∑ m = 0 R − 1 λ ~ i , m ∗ μ ~ m , j . \begin{aligned}
\widetilde{\gamma}_{i,j} = \operatorname{defl}_{R} \circ\operatorname{sh}^{i-j}(\lambda* \mu) \\
&= \operatorname{defl}_{R}(\operatorname{sh}^{i}(\lambda) * \operatorname{sh}^{-j}(\mu)) \\
&= \sum_{m=0}^{R-1} \operatorname{defl}_{R} \circ\operatorname{sh}^{-m}(\operatorname{sh}^{i}(\lambda)) * \operatorname{defl}_{R} \circ\operatorname{sh}^{m}(\operatorname{sh}^{-j}(\mu)) \\
&= \sum_{m=0}^{R-1} \widetilde{\lambda}_{i,m} * \widetilde{\mu}_{m,j}.
\end{aligned} γ i , j = defl R ∘ sh i − j ( λ ∗ μ ) = defl R ( sh i ( λ ) ∗ sh − j ( μ )) = m = 0 ∑ R − 1 defl R ∘ sh − m ( sh i ( λ )) ∗ defl R ∘ sh m ( sh − j ( μ )) = m = 0 ∑ R − 1 λ i , m ∗ μ m , j .
which shows that γ ~ = λ ~ ∗ μ ~ \widetilde{\gamma} = \widetilde{\lambda} * \widetilde{\mu} γ = λ ∗ μ .
Let us now proceed with the case of tilted convolution. For that purpose, we set μ ~ ′ = e R ′ ( μ ) \widetilde{\mu}' = e'_{R}(\mu) μ ′ = e R ′ ( μ ) with μ ~ i , j ′ = defl R ∘ sh j − i ( μ ) \widetilde{\mu}'_{i,j} = \operatorname{defl}_{R} \circ\operatorname{sh}^{j-i}(\mu) μ i , j ′ = defl R ∘ sh j − i ( μ ) and γ ~ ′ = e R ( γ ′ : = λ ∗ ′ μ ) \widetilde{\gamma}' = e_{R}(\gamma' := \lambda*' \mu) γ ′ = e R ( γ ′ := λ ∗ ′ μ ) with γ ~ i , j ′ = defl R ∘ sh i − j ( γ ′ ) \widetilde{\gamma}'_{i,j} = \operatorname{defl}_{R} \circ\operatorname{sh}^{i-j}(\gamma') γ i , j ′ = defl R ∘ sh i − j ( γ ′ ) . We show the claimed identity by direct computation:
γ ~ i , j ′ = defl R ∘ sh i − j ( λ ∗ ′ μ ) = defl R ( sh i ( λ ) ∗ ′ sh j ( μ ) ) = ∑ m = 0 R − 1 defl R ∘ sh − m ( sh i ( λ ) ) ∗ ′ defl R ∘ sh − m ( sh j ( μ ) ) = ∑ m = 0 R − 1 λ ~ i , m ∗ ′ μ ~ m , j ′ . \begin{aligned}
\widetilde{\gamma}'_{i,j} = \operatorname{defl}_{R} \circ\operatorname{sh}^{i-j}(\lambda*' \mu) \\
&= \operatorname{defl}_{R}(\operatorname{sh}^{i}(\lambda) *' \operatorname{sh}^{j}(\mu)) \\
&= \sum_{m=0}^{R-1} \operatorname{defl}_{R} \circ\operatorname{sh}^{-m}(\operatorname{sh}^{i}(\lambda)) *' \operatorname{defl}_{R} \circ\operatorname{sh}^{-m}(\operatorname{sh}^{j}(\mu)) \\
&= \sum_{m=0}^{R-1} \widetilde{\lambda}_{i,m} *' \widetilde{\mu}'_{m,j}.
\end{aligned} γ i , j ′ = defl R ∘ sh i − j ( λ ∗ ′ μ ) = defl R ( sh i ( λ ) ∗ ′ sh j ( μ )) = m = 0 ∑ R − 1 defl R ∘ sh − m ( sh i ( λ )) ∗ ′ defl R ∘ sh − m ( sh j ( μ )) = m = 0 ∑ R − 1 λ i , m ∗ ′ μ m , j ′ .
The identity for embedding the filtered process follows by the same arguments as the identity for embedding the convolution of two sequences. To avoid repetition, we will not detail the verification.
Proof of Proposition 4. The first three identities follow directly from Lemma 2. So we only need to show that the computation of the ACS is compatible with the embedding. We denote the P P P -stationary ACS of X X X by Σ \Sigma Σ . Let v ′ , v ∈ V v', v \in V v ′ , v ∈ V be two indices, then we compute for t , k ∈ Z t,k \in\mathbb{Z} t , k ∈ Z as follows:
[ ACS ( e R ( X ) ) ] v ′ , v ( k ) = E [ e R ( X v ′ ( t ) ) e R ( X v ( t − k ) ) ⊤ ] = [ E [ ( X v ′ ( − i + t R ) ) ( X v ( − j + ( t − k ) R ) ) ⊤ ] ] i , j ∈ [ 0 , R − 1 ] = [ Σ v ′ , v ( j − i + k R ) ] i , j ∈ [ 0 , R − 1 ] = e R ( Σ v ′ , v ) ( k ) . \begin{aligned}
[\operatorname{ACS}(e_{R}(X))]_{v',v}(k) = \mathbb{E}\left[e_{R}(X_{v'}(t))e_{R}(X_{v}(t-k))^{\top}\right] \\
&= \left[\mathbb{E}\left[(X_{v'}(-i+tR))(X_{v}(-j+(t-k)R))^{\top}\right]\right]_{i,j\in[0,R-1]} \\
&= \left[\Sigma_{v',v}(j-i+kR)\right]_{i,j\in[0,R-1]} \\
&= e_{R}(\Sigma_{v',v})(k).
\end{aligned} [ ACS ( e R ( X )) ] v ′ , v ( k ) = E [ e R ( X v ′ ( t )) e R ( X v ( t − k ) ) ⊤ ] = [ E [ ( X v ′ ( − i + tR )) ( X v ( − j + ( t − k ) R ) ) ⊤ ] ] i , j ∈ [ 0 , R − 1 ] = [ Σ v ′ , v ( j − i + k R ) ] i , j ∈ [ 0 , R − 1 ] = e R ( Σ v ′ , v ) ( k ) .
Representing a P P P -stationary SVAR representation as P P P -stationary SEP
Notation for time series graphs
Let G = ( V × Z , D ) \mathcal{G} = (V \times\mathbb{Z}, \mathcal{D}) G = ( V × Z , D ) be a P P P -stationary time series graph and let Φ ∈ R G \Phi\in\mathbb{R}^{\mathcal{G}} Φ ∈ R G be a compatible filter. For v , w ∈ V v,w \in V v , w ∈ V and m , n ∈ [ 0 , P − 1 ] m,n \in[0,P-1] m , n ∈ [ 0 , P − 1 ] , we write P v , w ( m , n ) ( k ) \mathcal{P}^{(m,n)}_{v,w}(k) P v , w ( m , n ) ( k ) to denote the set of directed paths on G \mathcal{G} G starting at v ( ( t − k ) P − m ) v((t-k)P-m) v (( t − k ) P − m ) and ending at w ( t P − n ) w(tP-n) w ( tP − n ) . For a directed path
π = ( v i ( s i P − l i ) → v i + 1 ( s i + 1 P − l i + 1 ) ∣ i ∈ [ N − 1 ] ) , \pi= (v_i(s_iP-l_i) \to v_{i+1}(s_{i+1}P-l_{i+1}) \mid i \in[N-1]), π = ( v i ( s i P − l i ) → v i + 1 ( s i + 1 P − l i + 1 ) ∣ i ∈ [ N − 1 ]) ,
define the associated path filter as follows:
ϕ ( π ) : = ( ∏ i = 1 N − 1 ϕ v i , v i + 1 ( l i , l i + 1 ) ( s i + 1 − s i ) ) 1 { k } ( j ) . \phi^{(\pi)} := \left(\prod_{i=1}^{N-1}\phi^{(l_i,l_{i+1})}_{v_i,v_{i+1}}(s_{i+1}-s_i)\right)\mathbf{1}_{\{k\}}(j). ϕ ( π ) := ( i = 1 ∏ N − 1 ϕ v i , v i + 1 ( l i , l i + 1 ) ( s i + 1 − s i ) ) 1 { k } ( j ) .
Given two directed paths, π ∈ P u , v ( m , n ) ( k ) \pi\in\mathcal{P}_{u,v}^{(m,n)}(k) π ∈ P u , v ( m , n ) ( k ) and π ′ ∈ P v , w ( n , q ) ( k ′ ) \pi' \in\mathcal{P}_{v,w}^{(n,q)}(k') π ′ ∈ P v , w ( n , q ) ( k ′ ) , it holds that
ϕ ( π ∗ π ′ ) = ϕ ( π ) ∗ ϕ ( π ′ ) , (20) \phi^{(\pi* \pi')} = \phi^{(\pi)} * \phi^{(\pi')}, \tag*{(20)} ϕ ( π ∗ π ′ ) = ϕ ( π ) ∗ ϕ ( π ′ ) , (20)
where π ∗ π ′ ∈ P u , w ( m , q ) ( k + k ′ ) \pi* \pi' \in\mathcal{P}_{u,w}^{(m,q)}(k+k') π ∗ π ′ ∈ P u , w ( m , q ) ( k + k ′ ) is the concatenation of π \pi π and π ′ \pi' π ′ . Finally, we define the union
P v , w ( m , n ) : = ⋃ k ∈ Z P v , w ( m , n ) ( k ) , \mathcal{P}_{v,w}^{(m,n)} := \bigcup_{k\in\mathbb{Z}} \mathcal{P}_{v,w}^{(m,n)}(k), P v , w ( m , n ) := k ∈ Z ⋃ P v , w ( m , n ) ( k ) ,
and associated with this set, we define
T v , w ( m , n ) ( Φ ) : = ∑ π ∈ P v , w ( m , n ) ϕ ( π ) = ∑ k ∈ Z ∑ π ∈ P v , w ( m , n ) ( k ) ϕ ( π ) . (21) T_{v,w}^{(m,n)}(\Phi) := \sum_{\pi\in\mathcal{P}_{v,w}^{(m,n)}} \phi^{(\pi)} = \sum_{k\in\mathbb{Z}} \sum_{\pi\in\mathcal{P}_{v,w}^{(m,n)}(k)} \phi^{(\pi)}. \tag*{(21)} T v , w ( m , n ) ( Φ ) := π ∈ P v , w ( m , n ) ∑ ϕ ( π ) = k ∈ Z ∑ π ∈ P v , w ( m , n ) ( k ) ∑ ϕ ( π ) . (21)
which we arrange into the P P P -stationary sequence T ( Φ ) : Z → BMat ( R , V , [ 0 , P − 1 ] ) T(\Phi): \mathbb{Z} \to\operatorname{BMat}(\mathbb{R},V,[0,P-1]) T ( Φ ) : Z → BMat ( R , V , [ 0 , P − 1 ]) . Note that T ( Φ ) T(\Phi) T ( Φ ) is well-defined if the spectral radius of the matrix Φ ( 0 ) \Phi(0) Φ ( 0 ) is less than one.
Lemma 21. It holds that
I + Φ ∗ T ( Φ ) = T ( Φ ) . \mathbb{I} + \Phi* T(\Phi) = T(\Phi). I + Φ ∗ T ( Φ ) = T ( Φ ) .
Before proving the lemma, we recall a basic property about directed paths. In order to state it, we define for ( v , m ) ∈ V × [ 0 , P − 1 ] (v,m) \in V \times[0,P-1] ( v , m ) ∈ V × [ 0 , P − 1 ] the set
C ( v , m ) : = { ( u , q ) ∈ V × [ 0 , P − 1 ] ∣ ∃ k : v → ( m , q ) k u } . C(v,m) := \left\{(u,q) \in V \times[0,P-1] \mid\exists k : v \xrightarrow{(m,q)}_{k} u \right\}. C ( v , m ) := { ( u , q ) ∈ V × [ 0 , P − 1 ] ∣ ∃ k : v ( m , q ) k u } .
Lemma 22. For m , n ∈ [ 0 , P − 1 ] m,n \in[0,P-1] m , n ∈ [ 0 , P − 1 ] and v , w ∈ V v,w \in V v , w ∈ V , there is the following bijection of sets
P v , w ( m , n ) ≅ { ⋃ ( u , q ) ∈ C ( v , m ) P u , w ( q , n ) ( v , m ) ≠ ( w , n ) , ⋃ ( u , q ) ∈ C ( v , m ) P u , w ( q , n ) ∪ { ϵ ( v , m ) } ( v , m ) = ( w , n ) . \mathcal{P}_{v,w}^{(m,n)} \cong
\begin{cases}
\bigcup_{(u,q)\in C(v,m)} \mathcal{P}_{u,w}^{(q,n)} & (v,m) \ne(w,n),\\
\bigcup_{(u,q)\in C(v,m)} \mathcal{P}_{u,w}^{(q,n)} \cup\{\epsilon_{(v,m)}\} & (v,m) = (w,n).
\end{cases} P v , w ( m , n ) ≅ { ⋃ ( u , q ) ∈ C ( v , m ) P u , w ( q , n ) ⋃ ( u , q ) ∈ C ( v , m ) P u , w ( q , n ) ∪ { ϵ ( v , m ) } ( v , m ) = ( w , n ) , ( v , m ) = ( w , n ) .
Proof. Let
π = ( v i ( s i P − l i ) → v i + 1 ( s i + 1 P − l i + 1 ) ∣ i ∈ [ N ] ) ∈ P v , w ( m , n ) , \pi= \left(v_i(s_iP-l_i) \to v_{i+1}(s_{i+1}P-l_{i+1}) \mid i \in[N]\right) \in\mathcal{P}_{v,w}^{(m,n)}, π = ( v i ( s i P − l i ) → v i + 1 ( s i + 1 P − l i + 1 ) ∣ i ∈ [ N ] ) ∈ P v , w ( m , n ) ,
then the claimed bijection maps π \pi π to the path one obtains by removing the first link. Note that this map takes the empty path to the empty path. It is clear that this gives a bijection.
Proof of Lemma 21. We prove the claimed identity element-wise and we abbreviate T ( Φ ) T(\Phi) T ( Φ ) by T T T . Let therefore v , w ∈ V v,w \in V v , w ∈ V and m , n ∈ [ 0 , P − 1 ] m,n \in[0,P-1] m , n ∈ [ 0 , P − 1 ] , then
1 { ( v , m ) } ( w , n ) + [ Φ ∗ T ] v , w ( m , n ) = 1 { ( v , m ) } ( w , n ) + ∑ u ∈ V ∑ q ∈ [ 0 , P − 1 ] Φ v , u ( m , q ) ∗ T u , w ( q , n ) = 1 { ( v , m ) } ( w , n ) + ∑ ( u , q ) ∈ C ( v , m ) Φ v , u ( m , q ) ∗ T u , w ( q , n ) = T v , w ( m , n ) , \begin{aligned}
\mathbb{1}_{\{(v,m)\}}(w,n) + [\Phi* T]_{v,w}^{(m,n)}
&= \mathbb{1}_{\{(v,m)\}}(w,n) + \sum_{u\in V}\sum_{q\in[0,P-1]} \Phi_{v,u}^{(m,q)} * T_{u,w}^{(q,n)} \\
&= \mathbb{1}_{\{(v,m)\}}(w,n) + \sum_{(u,q)\in C(v,m)} \Phi_{v,u}^{(m,q)} * T_{u,w}^{(q,n)} \\
&= T_{v,w}^{(m,n)},
\end{aligned} 1 {( v , m )} ( w , n ) + [ Φ ∗ T ] v , w ( m , n ) = 1 {( v , m )} ( w , n ) + u ∈ V ∑ q ∈ [ 0 , P − 1 ] ∑ Φ v , u ( m , q ) ∗ T u , w ( q , n ) = 1 {( v , m )} ( w , n ) + ( u , q ) ∈ C ( v , m ) ∑ Φ v , u ( m , q ) ∗ T u , w ( q , n ) = T v , w ( m , n ) ,
where the second equality follows from compatibility of Φ \Phi Φ with G G G and the third equality follows from Lemma 22 and (20) .
Proposition 23. For Φ ∈ R G \Phi\in\mathbb{R}^{\mathcal{G}} Φ ∈ R G such that the spectral radius of Φ ( 0 ) \Phi(0) Φ ( 0 ) is less than one, it holds that
S ( Φ ) = T ( Φ ) . (22) S(\Phi) = T(\Phi). \tag*{(22)} S ( Φ ) = T ( Φ ) . (22)
Lemma 24. Let Λ : Z → R N × N \Lambda: \mathbb{Z} \to\mathbb{R}^{N\times N} Λ : Z → R N × N be filter such that I − Λ ( 0 ) I-\Lambda(0) I − Λ ( 0 ) is invertible and Λ ( k ) = 0 \Lambda(k)=0 Λ ( k ) = 0 for k < 0 k<0 k < 0 , and A : Z → R N × N A:\mathbb{Z}\to\mathbb{R}^{N\times N} A : Z → R N × N . Then the following implication holds:
A = Λ ∗ A ⟹ A = 0. A = \Lambda* A \Longrightarrow A=0. A = Λ ∗ A ⟹ A = 0.
Proof. We show the claim by induction. First, consider k = 0 k = 0 k = 0 , then
( Λ ∗ A ) ( 0 ) = A ( 0 ) ⟺ ( I − Λ ( 0 ) ) A ( 0 ) = 0 ⟺ A ( 0 ) = 0 , (\Lambda* A)(0) = A(0) \Longleftrightarrow(I - \Lambda(0))A(0) = 0 \Longleftrightarrow A(0) = 0, ( Λ ∗ A ) ( 0 ) = A ( 0 ) ⟺ ( I − Λ ( 0 )) A ( 0 ) = 0 ⟺ A ( 0 ) = 0 ,
where for the first equality we used that Λ ( k ) = 0 \Lambda(k) = 0 Λ ( k ) = 0 for k < 0 k < 0 k < 0 , and for the last equivalence we used that I − Λ ( 0 ) I - \Lambda(0) I − Λ ( 0 ) is invertible.
Assuming A ( 0 ) = A ( ± 1 ) ⋯ = A ( ± k ) = 0 A(0) = A(\pm1) \cdots= A(\pm k) = 0 A ( 0 ) = A ( ± 1 ) ⋯ = A ( ± k ) = 0 , we compute
A ( k + 1 ) = ∑ l = 0 k + 1 Λ ( l ) A ( k + 1 − l ) = Λ ( 0 ) A ( k + 1 ) , A(k + 1) = \sum_{l=0}^{k+1} \Lambda(l)A(k + 1 - l) = \Lambda(0)A(k + 1), A ( k + 1 ) = l = 0 ∑ k + 1 Λ ( l ) A ( k + 1 − l ) = Λ ( 0 ) A ( k + 1 ) ,
and therefore
0 = ( I − Λ ( 0 ) ) A ( k + 1 ) ⟺ A ( k + 1 ) = 0. 0 = (I - \Lambda(0))A(k + 1) \Longleftrightarrow A(k + 1) = 0. 0 = ( I − Λ ( 0 )) A ( k + 1 ) ⟺ A ( k + 1 ) = 0.
The case for A ( − k − 1 ) A(-k - 1) A ( − k − 1 ) works analogously.
Proof of Proposition 23. The claim follows from the following identities:
S ( Φ ) = I + Φ ∗ S ( Φ ) , T ( Φ ) = I + Φ ∗ T ( Φ ) , S(\Phi) = I + \Phi* S(\Phi), \qquad T(\Phi) = I + \Phi* T(\Phi), S ( Φ ) = I + Φ ∗ S ( Φ ) , T ( Φ ) = I + Φ ∗ T ( Φ ) ,
which yield the equality
S ( Φ ) − T ( Φ ) = Φ ∗ ( S ( Φ ) − T ( Φ ) ) , S(\Phi) - T(\Phi) = \Phi* (S(\Phi) - T(\Phi)), S ( Φ ) − T ( Φ ) = Φ ∗ ( S ( Φ ) − T ( Φ )) ,
which implies, by Lemma 24, that S ( Φ ) = T ( Φ ) S(\Phi) = T(\Phi) S ( Φ ) = T ( Φ ) .
The internal and link filters as sums of path filters
Let v , w ∈ V v, w \in V v , w ∈ V and m , n ∈ [ 0 , P − 1 ] m, n \in[0, P - 1] m , n ∈ [ 0 , P − 1 ] be such that ( v , m ) → ( w , n ) (v, m) \to(w, n) ( v , m ) → ( w , n ) is a link on the process graph, then write P v → w ( m , n ) ( k ) ⊂ P v , w ( m , n ) ( k ) \mathcal{P}^{(m,n)}_{v\to w}(k) \subset\mathcal{P}^{(m,n)}_{v,w}(k) P v → w ( m , n ) ( k ) ⊂ P v , w ( m , n ) ( k ) for the set of directed paths on G \mathcal{G} G of the form
π = ( v i ( s i P − l i ) → v i + 1 ( s i + 1 P − l i + 1 ) ∣ i ∈ [ N − 1 ] ) , \pi= (v_i(s_iP - l_i) \to v_{i+1}(s_{i+1}P - l_{i+1}) \mid i \in[N - 1]), π = ( v i ( s i P − l i ) → v i + 1 ( s i + 1 P − l i + 1 ) ∣ i ∈ [ N − 1 ]) ,
where v i = w v_i = w v i = w for i > 1 i > 1 i > 1 and l 1 = m l_1 = m l 1 = m and l N = n l_N = n l N = n and s N − s 1 = k s_N - s_1 = k s N − s 1 = k , and define
P v → w ( m , n ) : = ⋃ k ∈ Z P v → w ( m , n ) ( k ) . \mathcal{P}^{(m,n)}_{v\to w} := \bigcup_{k \in\mathbb{Z}} \mathcal{P}^{(m,n)}_{v\to w}(k). P v → w ( m , n ) := k ∈ Z ⋃ P v → w ( m , n ) ( k ) .
The following is a special instance of Proposition 23.
Proposition 25. For v ≠ w ∈ V v \ne w \in V v = w ∈ V and m , n ∈ [ 0 , P − 1 ] m, n \in[0, P - 1] m , n ∈ [ 0 , P − 1 ] , the associated internal resp. link filters can be described as sums of TSG path filters in the following way:
ι v ( m , n ) = ∑ π ∈ P v → v ( m , n ) ϕ ( π ) , λ v , w ( m , n ) = ∑ π ∈ P v → w ( m , n ) ϕ ( π ) . \iota^{(m,n)}_v = \sum_{\pi\in\mathcal{P}^{(m,n)}_{v\to v}} \phi^{(\pi)}, \qquad\lambda^{(m,n)}_{v,w} = \sum_{\pi\in\mathcal{P}^{(m,n)}_{v\to w}} \phi^{(\pi)}. ι v ( m , n ) = π ∈ P v → v ( m , n ) ∑ ϕ ( π ) , λ v , w ( m , n ) = π ∈ P v → w ( m , n ) ∑ ϕ ( π ) .
In particular, v m I ↛ v n ⟹ ι v ( m , n ) = 0 v^I_m \not\to v_n \Longrightarrow\iota^{(m,n)}_v = 0 v m I → v n ⟹ ι v ( m , n ) = 0 , and v m ↛ w n ⟹ λ v , w ( m , n ) = 0 v_m \not\to w_n \Longrightarrow\lambda^{(m,n)}_{v,w} = 0 v m → w n ⟹ λ v , w ( m , n ) = 0 .
Proof. Let G v \mathcal{G}_v G v be the restriction of G \mathcal{G} G to the vertex set { v } × Z \{v\} \times\mathbb{Z} { v } × Z , and let Φ v ∈ R G v \Phi_v \in\mathbb{R}^{\mathcal{G}_v} Φ v ∈ R G v be the restriction of Φ \Phi Φ to G v \mathcal{G}_v G v , then identity regarding the internal filter follows by applying Proposition 23 to G v \mathcal{G}_v G v and Φ v \Phi_v Φ v . The identity regarding the link filters follows then directly from the identity for the internal filters and the definition of the link filters, i.e., λ v , w = Φ v , w ∗ ι w \lambda_{v,w} = \Phi_{v,w} * \iota_w λ v , w = Φ v , w ∗ ι w .
Path-wise decomposition of the power sum of the link filter
Let G = ( V × [ 0 , P − 1 ] , D ) \mathcal{G} = (V \times[0, P - 1], D) G = ( V × [ 0 , P − 1 ] , D ) be the process graph associated with the time series graph G \mathcal{G} G . We denote by P ( ( v , m ) , ( w , n ) ) \mathcal{P}((v, m), (w, n)) P (( v , m ) , ( w , n )) the set of directed paths on G \mathcal{G} G starting at ( v , m ) (v, m) ( v , m ) and ending at ( w , n ) (w, n) ( w , n ) . For a directed path π = ( ( v i , l i ) → ( v i + 1 , l i + 1 ) ∣ i ∈ [ N − 1 ] ) \pi= ((v_i, l_i) \to(v_{i+1}, l_{i+1}) \mid i \in[N - 1]) π = (( v i , l i ) → ( v i + 1 , l i + 1 ) ∣ i ∈ [ N − 1 ]) on G \mathcal{G} G , we define its associated path filter
λ ( π ) : = λ v 1 , v 2 ( l 1 , l 2 ) ∗ ⋯ ∗ λ v N − 1 , v N ( l N − 1 , l N ) . \lambda^{(\pi)} := \lambda^{(l_1,l_2)}_{v_1,v_2} * \cdots* \lambda^{(l_{N-1},l_N)}_{v_{N-1},v_N}. λ ( π ) := λ v 1 , v 2 ( l 1 , l 2 ) ∗ ⋯ ∗ λ v N − 1 , v N ( l N − 1 , l N ) .
For two directed paths, π ∈ P ( ( u , m ) , ( v , n ) ) \pi\in\mathcal{P}((u, m), (v, n)) π ∈ P (( u , m ) , ( v , n )) and π ′ ∈ P ( ( v , n ) , ( w , q ) ) \pi' \in\mathcal{P}((v, n), (w, q)) π ′ ∈ P (( v , n ) , ( w , q )) , it holds that
λ ( π ∗ π ′ ) = λ ( π ) ∗ λ ( π ′ ) , \lambda^{(\pi* \pi')} = \lambda^{(\pi)} * \lambda^{(\pi')}, λ ( π ∗ π ′ ) = λ ( π ) ∗ λ ( π ′ ) ,
where π ∗ π ′ ∈ P ( ( u , m ) , ( w , q ) ) \pi* \pi' \in\mathcal{P}((u, m), (w, q)) π ∗ π ′ ∈ P (( u , m ) , ( w , q )) is the concatenation of π \pi π and π ′ \pi' π ′ .
Stability of P-stationary SVAR processes
Lemma 26. It holds that
S ( Λ ) v , w ( m , n ) = ∑ π ∈ P ( ( v , m ) , ( w , n ) ) λ ( π ) . S(\Lambda)^{(m,n)}_{v,w}=\sum_{\pi\in\mathrm{P}((v,m),(w,n))}\lambda^{(\pi)}. S ( Λ ) v , w ( m , n ) = π ∈ P (( v , m ) , ( w , n )) ∑ λ ( π ) .
Proof. For k ≥ 0 k\ge0 k ≥ 0 , define P k ⊂ P ( ( v , m ) , ( w , n ) ) \mathrm{P}^{k}\subset\mathrm{P}((v,m),(w,n)) P k ⊂ P (( v , m ) , ( w , n )) to be the subset of directed paths π \pi π consisting of exactly k k k links. Using the fact that Λ \Lambda Λ is compatible with G \mathcal{G} G , one can show by induction that
[ Λ k ] v , w ( m , n ) = ∑ π ∈ P k λ ( π ) . [\Lambda^{k}]^{(m,n)}_{v,w}=\sum_{\pi\in\mathrm{P}^{k}}\lambda^{(\pi)}. [ Λ k ] v , w ( m , n ) = π ∈ P k ∑ λ ( π ) .
Then the claim follows as
P ( ( v , m ) , ( w , n ) ) = ⋃ k ≥ 0 P k . \mathrm{P}((v,m),(w,n))=\bigcup_{k\ge0}\mathrm{P}^{k}. P (( v , m ) , ( w , n )) = k ≥ 0 ⋃ P k .
:::
To describe S ( Λ ) S(\Lambda) S ( Λ ) in terms of path filters on G \mathcal{G} G , we introduce the following sets of directed paths. For v , w ∈ V v,w\in V v , w ∈ V and m , n ∈ [ 0 , P − 1 ] m,n\in[0,P-1] m , n ∈ [ 0 , P − 1 ] , define P ^ v , w ( m , n ) ⊂ P v , w ( m , n ) \widehat{\mathrm{P}}^{(m,n)}_{v,w}\subset\mathrm{P}^{(m,n)}_{v,w} P v , w ( m , n ) ⊂ P v , w ( m , n ) to be the set of directed paths π \pi π of the form π = π ′ ∗ π ′ ′ \pi=\pi'\ast\pi'' π = π ′ ∗ π ′′ , where π ′ ∈ P v , v ′ ( m , n ′ ) \pi'\in\mathrm{P}^{(m,n')}_{v,v'} π ′ ∈ P v , v ′ ( m , n ′ ) and π ′ ′ ∈ P v ′ , w ( n ′ , n ) \pi''\in\mathrm{P}^{(n',n)}_{v',w} π ′′ ∈ P v ′ , w ( n ′ , n ) , for some v ′ ∈ V ∖ { v } v'\in V\setminus\{v\} v ′ ∈ V ∖ { v } and n ′ ∈ [ 0 , P − 1 ] n'\in[0,P-1] n ′ ∈ [ 0 , P − 1 ] .
Lemma 27. It holds that
S ( Λ ) v , w ( m , n ) = ∑ π ∈ P ^ v , w ( m , n ) ϕ ( π ) . (23) S(\Lambda)^{(m,n)}_{v,w}=\sum_{\pi\in\widehat{\mathrm{P}}^{(m,n)}_{v,w}}\phi^{(\pi)}. \tag*{(23)} S ( Λ ) v , w ( m , n ) = π ∈ P v , w ( m , n ) ∑ ϕ ( π ) . (23)
Proof. For a path π = ( ( v i , l i ) → ( v i + 1 , l i + 1 ) ∣ i ∈ [ N − 1 ] ) ∈ P ( ( v , m ) , ( w , n ) ) \pi=((v_i,l_i)\to(v_{i+1},l_{i+1})\mid i\in[N-1])\in\mathrm{P}((v,m),(w,n)) π = (( v i , l i ) → ( v i + 1 , l i + 1 ) ∣ i ∈ [ N − 1 ]) ∈ P (( v , m ) , ( w , n )) , we define
P π : = ∏ i ∈ [ N − 1 ] P v i → v i + 1 ( l i , l i + 1 ) . \mathbf{P}^{\pi}:=\prod_{i\in[N-1]}\mathbf{P}^{(l_i,l_{i+1})}_{v_i\to v_{i+1}}. P π := i ∈ [ N − 1 ] ∏ P v i → v i + 1 ( l i , l i + 1 ) .
Note that concatenation of paths on G \mathcal{G} G yields the injective map
∗ : P π ⟶ P ^ v , w ( m , n ) , ( π i ∣ i ∈ [ N − 1 ] ) ↦ π 1 ∗ ⋯ ∗ π N − 1 . \ast:\mathbf{P}^{\pi}\longrightarrow\widehat{\mathrm{P}}^{(m,n)}_{v,w},\qquad(\pi_i\mid i\in[N-1])\mapsto\pi_1\ast\cdots\ast\pi_{N-1}. ∗ : P π ⟶ P v , w ( m , n ) , ( π i ∣ i ∈ [ N − 1 ]) ↦ π 1 ∗ ⋯ ∗ π N − 1 .
Furthermore, Proposition 25 in combination with the definition of λ ( π ) \lambda^{(\pi)} λ ( π ) and P π \mathbf{P}^{\pi} P π gives that
λ ( π ) = ∑ ( π i ∣ i ∈ [ N − 1 ] ) ∈ P π ϕ ( π 1 ) ∗ ⋯ ∗ ϕ ( π N − 1 ) ⏟ ϕ ( π 1 ∗ ⋯ ∗ π N − 1 ) . (24) \lambda^{(\pi)}=
\sum_{(\pi_i\mid i\in[N-1])\in\mathbf{P}^{\pi}}
\underbrace{\phi^{(\pi_1)}\ast\cdots\ast\phi^{(\pi_{N-1})}}_{\phi^{(\pi_1\ast\cdots\ast\pi_{N-1})}}. \tag*{(24)} λ ( π ) = ( π i ∣ i ∈ [ N − 1 ]) ∈ P π ∑ ϕ ( π 1 ∗ ⋯ ∗ π N − 1 ) ϕ ( π 1 ) ∗ ⋯ ∗ ϕ ( π N − 1 ) . (24)
On the other hand, for every path π ^ ∈ P ^ v , w ( m , n ) \widehat{\pi}\in\widehat{\mathrm{P}}^{(m,n)}_{v,w} π ∈ P v , w ( m , n ) , there is a unique π ∈ P ( ( v , m ) , ( w , n ) ) \pi\in\mathrm{P}((v,m),(w,n)) π ∈ P (( v , m ) , ( w , n )) such that π ^ ∈ P π \widehat{\pi}\in\mathbf{P}^{\pi} π ∈ P π . In particular, concatenation on G \mathcal{G} G yields a bijection
⋃ π ∈ P ( ( v , m ) , ( w , n ) ) P π → ≅ P ^ v , w ( m , n ) . (25) \bigcup_{\pi\in\mathrm{P}((v,m),(w,n))}\mathbf{P}^{\pi}\xrightarrow{\ \cong\ }\widehat{\mathrm{P}}^{(m,n)}_{v,w}. \tag*{(25)} π ∈ P (( v , m ) , ( w , n )) ⋃ P π ≅ P v , w ( m , n ) . (25)
Combining the identity (23) with the bijection (24) and Lemma 26 proves the identity (22).
Proposition 28. It holds that
S ( Φ ) = ι ∗ S ( Λ ) . S(\Phi)=\iota\ast S(\Lambda). S ( Φ ) = ι ∗ S ( Λ ) .
Proof. The proposition follows immediately by combining (20) with the observation that concatenation on G \mathcal{G} G yields a bijection
P v , w ( m , n ) ≅ ⋃ m ′ ∈ [ 0 , P − 1 ] P v , v ( m , m ′ ) × P ^ v , w ( m ′ , n ) . \mathbf{P}^{(m,n)}_{v,w}\cong\bigcup_{m'\in[0,P-1]}\mathbf{P}^{(m,m')}_{v,v}\times\widehat{\mathbf{P}}^{(m',n)}_{v,w}. P v , w ( m , n ) ≅ m ′ ∈ [ 0 , P − 1 ] ⋃ P v , v ( m , m ′ ) × P v , w ( m ′ , n ) .
Proof of Proposition 7. The solution of the SVAR process is S ( Φ ) ⊤ ∗ η S(\Phi)^{\top}\ast\eta S ( Φ ) ⊤ ∗ η , and the solution of the associated SEP is S ( Λ ) ⊤ ∗ ι ⊤ ∗ η S(\Lambda)^{\top}\ast\iota^{\top}\ast\eta S ( Λ ) ⊤ ∗ ι ⊤ ∗ η . Hence, the solutions coincide by Proposition 28.
Stability of P P P -stationary SVAR processes
Let ( G , Φ , η ) (\mathcal{G}, \Phi, \eta) ( G , Φ , η ) be a P P P -stationary SVAR model and ( G , Λ , Ξ 1 ) (\mathcal{G}, \Lambda, \Xi^{1}) ( G , Λ , Ξ 1 ) the associated SEP. We wish to develop a criterion for the parameter Φ \Phi Φ that, if satisfied, implies stability of Λ \Lambda Λ . For that purpose, we will develop an alternative description of S ( Φ ) S(\Phi) S ( Φ ) based on which we develop a criterion for Φ \Phi Φ to check if S ( Φ ) S(\Phi) S ( Φ ) is absolutely summable. Then we will use Proposition Proposition 28 to conclude that Λ \Lambda Λ is stable.
For the alternative formulation of S ( Φ ) S(\Phi) S ( Φ ) , we construct a P P P -stationary Ψ : Z → BMat ( R , V , [ 0 , P − 1 ] ) \Psi:\mathbb{Z}\to\operatorname{BMat}(\mathbb{R},V,[0,P-1]) Ψ : Z → BMat ( R , V , [ 0 , P − 1 ]) , which is compatible with the summary graph smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) . The entry of Ψ \Psi Ψ corresponding to v , w ∈ V , W v,w\in V,W v , w ∈ V , W and m , n ∈ [ 0 , P − 1 ] m,n\in[0,P-1] m , n ∈ [ 0 , P − 1 ] is defined as follows:
Ψ v , w ( m , n ) : = { S ( Φ v , v ( m , m ) ) m = n and v = w , Φ v , w ( m , n ) ∗ Ψ w , w ( n , n ) m ≠ n or v ≠ w . \Psi_{v,w}^{(m,n)} :=
\begin{cases}
S\left(\Phi_{v,v}^{(m,m)}\right) & m=n\text{ and }v=w,\\
\Phi_{v,w}^{(m,n)} * \Psi_{w,w}^{(n,n)} & m\ne n\text{ or }v\ne w.
\end{cases} Ψ v , w ( m , n ) := { S ( Φ v , v ( m , m ) ) Φ v , w ( m , n ) ∗ Ψ w , w ( n , n ) m = n and v = w , m = n or v = w .
Note that Ψ v , w ( m , n ) \Psi_{v,w}^{(m,n)} Ψ v , w ( m , n ) is non-zero only if ( v , m ) → ( w , n ) (v,m)\to(w,n) ( v , m ) → ( w , n ) on the auxiliary graph smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) . In order to describe the entries of Ψ \Psi Ψ as sums of path filters on the time series graph G \mathcal{G} G , we define P ~ v → w ( m , n ) ( k ) ⊂ P v → w ( m , n ) ( k ) \widetilde{\mathcal{P}}_{v\to w}^{(m,n)}(k)\subset\mathcal{P}_{v\to w}^{(m,n)}(k) P v → w ( m , n ) ( k ) ⊂ P v → w ( m , n ) ( k ) to be the directed paths on G \mathcal{G} G of the following form:
π = ( v i ( s i P − l i ) → v i + 1 ( s i + 1 P − l i + 1 ) ∣ i ∈ [ N − 1 ] ) , \pi=\left(v_i(s_iP-l_i)\to v_{i+1}(s_{i+1}P-l_{i+1})\mid i\in[N-1]\right), π = ( v i ( s i P − l i ) → v i + 1 ( s i + 1 P − l i + 1 ) ∣ i ∈ [ N − 1 ] ) ,
where v 1 = v v_1=v v 1 = v and l 1 = m l_1=m l 1 = m , and v i = w v_i=w v i = w and l i = n l_i=n l i = n for every i > 1 i>1 i > 1 . By the same principle as in the previous section, we define
P ~ v → w ( m , n ) : = ⋃ k ∈ Z P ~ v → w ( m , n ) ( k ) . \widetilde{\mathcal{P}}_{v\to w}^{(m,n)}:=\bigcup_{k\in\mathbb{Z}}\widetilde{\mathcal{P}}_{v\to w}^{(m,n)}(k). P v → w ( m , n ) := k ∈ Z ⋃ P v → w ( m , n ) ( k ) .
Lemma 29. For v , w ∈ V v,w\in V v , w ∈ V and m , n ∈ [ 0 , P − 1 ] m,n\in[0,P-1] m , n ∈ [ 0 , P − 1 ] , it holds that
Ψ v , w ( m , n ) = ∑ π ∈ P ~ v → w ( m , n ) ϕ ( π ) . \Psi_{v,w}^{(m,n)}=\sum_{\pi\in\widetilde{\mathcal{P}}_{v\to w}^{(m,n)}}\phi^{(\pi)}. Ψ v , w ( m , n ) = π ∈ P v → w ( m , n ) ∑ ϕ ( π ) .
Proof. To prove the claim, we inspect the filter S = S ( Φ v , v ( m , m ) ) S=S\left(\Phi_{v,v}^{(m,m)}\right) S = S ( Φ v , v ( m , m ) ) . In fact, it follows from Proposition Proposition 23 that
S = ∑ π ∈ P ~ v → v ( m , m ) ϕ ( π ) . S=\sum_{\pi\in\widetilde{\mathcal{P}}_{v\to v}^{(m,m)}}\phi^{(\pi)}. S = π ∈ P v → v ( m , m ) ∑ ϕ ( π ) .
Specifically, this follows by applying Proposition Proposition 23 to the P P P -stationary filter
Φ ′ ( p , q ) = { Φ v , v ( m , m ) ( p , q ) = ( m , m ) , 0 otherwise , \Phi'^{(p,q)}=
\begin{cases}
\Phi_{v,v}^{(m,m)} & (p,q)=(m,m),\\
0 & \text{otherwise},
\end{cases} Φ ′ ( p , q ) = { Φ v , v ( m , m ) 0 ( p , q ) = ( m , m ) , otherwise ,
which is compatible with the time series graph G v ′ = ( { v } × Z , D ) \mathcal{G}'_v=(\{v\}\times\mathbb{Z},D) G v ′ = ({ v } × Z , D ) , where there is v ⟶ ( p , q ) v v\overset{(p,q)}{\longrightarrow}v v ⟶ ( p , q ) v if and only if p = q = m p=q=m p = q = m and v ⟶ ( m , m ) v v\overset{(m,m)}{\longrightarrow}v v ⟶ ( m , m ) v is a lagged relation on G \mathcal{G} G , and by the very definition of Ψ \Psi Ψ , this observation suffices to show the statement of the lemma. □
Let π = ( ( v i , l i ) → ( v i + 1 , l i + 1 ) ∣ i ∈ [ N − 1 ] ) ∈ P smry ( G ) ( ( v , m ) , ( w , n ) ) \pi=((v_i,l_i)\to(v_{i+1},l_{i+1})\mid i\in[N-1])\in\mathcal{P}_{\operatorname{smry}(\mathcal{G})}((v,m),(w,n)) π = (( v i , l i ) → ( v i + 1 , l i + 1 ) ∣ i ∈ [ N − 1 ]) ∈ P smry ( G ) (( v , m ) , ( w , n )) , then define the associated path filter
Ψ ( π ) = Ψ v 1 , v 1 ( l 1 , l 1 ) ∗ Ψ v 1 , v 2 ( l 1 , l 2 ) ∗ ⋯ ∗ Ψ v N − 1 , v N ( l N − 1 , l N ) = ∑ π ~ ∈ P ~ π ϕ ( π ~ ) . (26) \Psi^{(\pi)}=\Psi_{v_1,v_1}^{(l_1,l_1)}*\Psi_{v_1,v_2}^{(l_1,l_2)}*\cdots*\Psi_{v_{N-1},v_N}^{(l_{N-1},l_N)}
=\sum_{\widetilde{\pi}\in\widetilde{\mathcal{P}}^\pi}\phi^{(\widetilde{\pi})}. \tag*{(26)} Ψ ( π ) = Ψ v 1 , v 1 ( l 1 , l 1 ) ∗ Ψ v 1 , v 2 ( l 1 , l 2 ) ∗ ⋯ ∗ Ψ v N − 1 , v N ( l N − 1 , l N ) = π ∈ P π ∑ ϕ ( π ) . (26)
where P ~ π = P v 1 , v 1 ( l 1 , l 1 ) × ∏ i ∈ [ N − 1 ] P ~ v i → v i + 1 ( l i , l i + 1 ) \widetilde{\mathcal{P}}^\pi=\mathcal{P}_{v_1,v_1}^{(l_1,l_1)}\times\prod_{i\in[N-1]}\widetilde{\mathcal{P}}_{v_i\to v_{i+1}}^{(l_i,l_{i+1})} P π = P v 1 , v 1 ( l 1 , l 1 ) × ∏ i ∈ [ N − 1 ] P v i → v i + 1 ( l i , l i + 1 ) and
ϕ ( π ~ ) = ϕ ( π ~ 0 ) ∗ ⋯ ∗ ϕ ( π ~ N − 1 ) . \phi^{(\widetilde{\pi})}=\phi^{(\widetilde{\pi}_0)}*\cdots*\phi^{(\widetilde{\pi}_{N-1})}. ϕ ( π ) = ϕ ( π 0 ) ∗ ⋯ ∗ ϕ ( π N − 1 ) .
Proposition 30. Given the notation as above, it holds that
S ( Ψ ) = S ( Φ ) . S(\Psi)=S(\Phi). S ( Ψ ) = S ( Φ ) .
Proof. Combining the fact that
S ( Ψ ) v , w ( m , n ) = ∑ π ∈ P s m r y ( G ) ( ( v , m ) , ( w , n ) ) Ψ ( π ) , S(\Psi)_{v,w}^{(m,n)}=\sum_{\pi\in\mathrm{P}_{\mathrm{smry}(\mathcal{G})}((v,m),(w,n))}\Psi^{(\pi)}, S ( Ψ ) v , w ( m , n ) = π ∈ P smry ( G ) (( v , m ) , ( w , n )) ∑ Ψ ( π ) ,
with the observation that concatenation of paths on G \mathcal{G} G induces a bijection of sets
⋃ π ∈ P ( ( v , m ) , ( w , n ) ) Π ~ π ≅ P v , w ( m , n ) , \bigcup_{\pi\in\mathrm{P}((v,m),(w,n))}\widetilde{\Pi}^{\pi}\cong\mathrm{P}_{v,w}^{(m,n)}, π ∈ P (( v , m ) , ( w , n )) ⋃ Π π ≅ P v , w ( m , n ) ,
and the Equation (26) , we get the statement of the proposition
S ( Ψ ) v , w ( m , n ) = ∑ π ∈ P s m r y ( G ) ( ( v , m ) , ( w , n ) ) ψ ( π ) = ∑ π ∈ P v , w ( m , n ) ϕ ( π ) = S ( Φ ) v , w ( m , n ) . S(\Psi)_{v,w}^{(m,n)}
=\sum_{\pi\in\mathrm{P}_{\mathrm{smry}(\mathcal{G})}((v,m),(w,n))}\psi^{(\pi)}
=\sum_{\pi\in\mathrm{P}_{v,w}^{(m,n)}}\phi^{(\pi)}
=S(\Phi)_{v,w}^{(m,n)}. S ( Ψ ) v , w ( m , n ) = π ∈ P smry ( G ) (( v , m ) , ( w , n )) ∑ ψ ( π ) = π ∈ P v , w ( m , n ) ∑ ϕ ( π ) = S ( Φ ) v , w ( m , n ) .
In the following, V ~ 1 , … , V ~ N ⊂ V × [ 0 , P − 1 ] \widetilde{V}_1,\ldots,\widetilde{V}_N\subset V\times[0,P-1] V 1 , … , V N ⊂ V × [ 0 , P − 1 ] denote the strongly connected components of the summary graph s m r y ( G ) \mathrm{smry}(\mathcal{G}) smry ( G ) , i.e., ( v , m ) , ( w , n ) ∈ V ~ i (v,m),(w,n)\in\widetilde{V}_i ( v , m ) , ( w , n ) ∈ V i for some i ∈ [ N ] i\in[N] i ∈ [ N ] if and only if ( v , m ) ∈ anc s m r y ( G ) ( w , n ) ∩ des s m r y ( G ) ( w , n ) (v,m)\in\operatorname{anc}_{\mathrm{smry}(\mathcal{G})}(w,n)\cap\operatorname{des}_{\mathrm{smry}(\mathcal{G})}(w,n) ( v , m ) ∈ anc smry ( G ) ( w , n ) ∩ des smry ( G ) ( w , n ) .
Proposition 31. Let Φ ∈ R G \Phi\in\mathbb{R}^{\mathcal{G}} Φ ∈ R G be such that ∥ [ Φ ⊤ ] V ~ i , V ~ i ∥ 1 < 1 \lVert[\Phi^{\top}]_{\widetilde{V}_i,\widetilde{V}_i}\rVert_1<1 ∥[ Φ ⊤ ] V i , V i ∥ 1 < 1 for all i ∈ [ N ] i\in[N] i ∈ [ N ] , then S ( Φ ) S(\Phi) S ( Φ ) as well as S ( Λ ) S(\Lambda) S ( Λ ) are absolutely summable. Furthermore, the Fourier transformed filter F ( S ( Φ ) ) \mathcal{F}(S(\Phi)) F ( S ( Φ )) takes the form of a rational function that is well-defined for every z ∈ C z\in\mathbb{C} z ∈ C such that ∣ z ∣ ≤ 1 \lvert z\rvert\leq1 ∣ z ∣ ≤ 1 .
To alleviate the proof of Proposition 31, we first prove a weaker version of it, which we state in the following lemma. For the proof of that lemma, it will be useful to recall the following fact about the operator norm with respect to the norm ∥ ⋅ ∥ 1 \lVert\cdot\rVert_1 ∥ ⋅ ∥ 1 . Let A ∈ C V × V A\in\mathbb{C}^{V\times V} A ∈ C V × V be a matrix, then it holds that
∥ A ∥ 1 = max v ∈ R V ∖ { 0 } ∥ A ∥ 1 ∥ v ∥ 1 = max i ∈ V ∑ j ∈ V ∣ A i , j ∣ . (27) \lVert A\rVert_1=\max_{v\in\mathbb{R}^{V}\setminus\{0\}}\frac{\lVert A\rVert_1}{\lVert v\rVert_1}=\max_{i\in V}\sum_{j\in V}\lvert A_{i,j}\rvert. \tag*{(27)} ∥ A ∥ 1 = v ∈ R V ∖ { 0 } max ∥ v ∥ 1 ∥ A ∥ 1 = i ∈ V max j ∈ V ∑ ∣ A i , j ∣ . (27)
Note that ∥ A ⊤ ∥ 1 < 1 \lVert A^{\top}\rVert_1<1 ∥ A ⊤ ∥ 1 < 1 implies that the spectral radius of A A A is less than one. Suppose A v = λ v Av=\lambda v A v = λ v , then ∥ A v ∥ 1 = ∣ λ ∣ ∥ v ∥ 1 \lVert Av\rVert_1=\lvert\lambda\rvert\lVert v\rVert_1 ∥ A v ∥ 1 = ∣ λ ∣ ∥ v ∥ 1 . Hence, ∣ λ ∣ < 1 \lvert\lambda\rvert<1 ∣ λ ∣ < 1 follows from
∥ A v ∥ 1 = ∑ i ∈ V ∣ ∑ j ∈ V A i , j v j ∣ ≤ ∑ j ∈ V ∣ v j ∣ ∑ i ∈ V ∣ A i , j ∣ ≤ ∥ v ∥ 1 ∥ A ⊤ ∥ 1 < ∥ v ∥ 1 . \lVert Av\rVert_1=\sum_{i\in V}\left|\sum_{j\in V}A_{i,j}v_j\right|\leq\sum_{j\in V}\lvert v_j\rvert\sum_{i\in V}\lvert A_{i,j}\rvert\leq\lVert v\rVert_1\lVert A^{\top}\rVert_1<\lVert v\rVert_1. ∥ A v ∥ 1 = i ∈ V ∑ j ∈ V ∑ A i , j v j ≤ j ∈ V ∑ ∣ v j ∣ i ∈ V ∑ ∣ A i , j ∣ ≤ ∥ v ∥ 1 ∥ A ⊤ ∥ 1 < ∥ v ∥ 1 .
Recall from the main part that for a matrix-valued filter Φ \Phi Φ , we defined
∥ Φ ∥ 1 : = ∥ ∑ k ∈ Z ∣ Φ ( k ) ∣ ∥ 1 . \lVert\Phi\rVert_1:=\left\lVert\sum_{k\in\mathbb{Z}}\lvert\Phi(k)\rvert\right\rVert_1. ∥ Φ ∥ 1 := k ∈ Z ∑ ∣ Φ ( k )∣ 1 .
Furthermore, note that ∑ k ∈ Z ∣ Φ ( k ) ∣ = F ( ∣ Φ ∣ ) ( 1 ) \sum_{k\in\mathbb{Z}}\lvert\Phi(k)\rvert=\mathcal{F}(\lvert\Phi\rvert)(1) ∑ k ∈ Z ∣ Φ ( k )∣ = F (∣ Φ ∣) ( 1 ) , where ∣ Φ ∣ ( k ) = ∣ Φ ( k ) ∣ \lvert\Phi\rvert(k)=\lvert\Phi(k)\rvert ∣ Φ ∣ ( k ) = ∣ Φ ( k )∣ , for every k ∈ Z k\in\mathbb{Z} k ∈ Z .
Lemma 32. If ∥ Φ ⊤ ∥ 1 < 1 \lVert\Phi^{\top}\rVert_1<1 ∥ Φ ⊤ ∥ 1 < 1 , then the filter S ( Φ ) S(\Phi) S ( Φ ) is absolutely summable, and S ( F ( Φ ) ) ( z ) S(\mathcal{F}(\Phi))(z) S ( F ( Φ )) ( z ) takes the form of a rational function in z z z that is well-defined for every z ∈ C z\in\mathbb{C} z ∈ C with ∣ z ∣ ≤ 1 \lvert z\rvert\leq1 ∣ z ∣ ≤ 1 .
Proof. The preceding observation in combination with the assumption of the lemma give, for ∣ z ∣ ≤ 1 \lvert z\rvert\leq1 ∣ z ∣ ≤ 1 , the chain of inequalities
∥ F ( Φ ⊤ ) ( z ) ∥ 1 ≤ ∥ F ( ∣ Φ ∣ ⊤ ) ( 1 ) ∥ 1 ≤ ∥ Φ ⊤ ∥ 1 < 1. \lVert\mathcal{F}(\Phi^{\top})(z)\rVert_1\leq\lVert\mathcal{F}(\lvert\Phi\rvert^{\top})(1)\rVert_1\leq\lVert\Phi^{\top}\rVert_1<1. ∥ F ( Φ ⊤ ) ( z ) ∥ 1 ≤ ∥ F (∣ Φ ∣ ⊤ ) ( 1 ) ∥ 1 ≤ ∥ Φ ⊤ ∥ 1 < 1.
In particular, the spectral radius of F ( Φ ) ( z ) \mathcal{F}(\Phi)(z) F ( Φ ) ( z ) is less than one, for every ∣ z ∣ ≤ 1 \lvert z\rvert\leq1 ∣ z ∣ ≤ 1 , meaning that
S ( F ( Φ ) ( z ) ) = S ( F ( Φ ) ( z ) ) = ( I − F ( Φ ) ( z ) ) − 1 S(\mathcal{F}(\Phi)(z))=S(\mathcal{F}(\Phi)(z))=(I-\mathcal{F}(\Phi)(z))^{-1} S ( F ( Φ ) ( z )) = S ( F ( Φ ) ( z )) = ( I − F ( Φ ) ( z ) ) − 1
exists and takes the form of a rational function in z z z because Φ \Phi Φ being compatible with the finite order time series graph G \mathcal{G} G implies that F ( Φ ) \mathcal{F}(\Phi) F ( Φ ) is a polynomial in z z z . Finally, noting that S ( ∣ Φ ∣ ) = ∣ S ( ∣ Φ ∣ ) ∣ S(\lvert\Phi\rvert)=\lvert S(\lvert\Phi\rvert)\rvert S (∣ Φ ∣) = ∣ S (∣ Φ ∣)∣ , we use the observations preceding the lemma to conclude with the absolute summability of S ( Φ ) S(\Phi) S ( Φ ) , which is implied by
∥ S ( Φ ) ∥ 1 ≤ ∥ S ( ∣ Φ ∣ ) ∥ 1 = ∥ F ( S ( ∣ Φ ∣ ) ) ( 1 ) ∥ 1 = ∥ ∑ k ≥ 0 F ( ∣ Φ ∣ ) k ( 1 ) ∥ 1 < ∞ . \lVert S(\Phi)\rVert_1\leq\lVert S(\lvert\Phi\rvert)\rVert_1=\lVert\mathcal{F}(S(\lvert\Phi\rvert))(1)\rVert_1=\left\lVert\sum_{k\geq0}\mathcal{F}(\lvert\Phi\rvert)^k(1)\right\rVert_1<\infty. ∥ S ( Φ ) ∥ 1 ≤ ∥ S (∣ Φ ∣) ∥ 1 = ∥ F ( S (∣ Φ ∣)) ( 1 ) ∥ 1 = k ≥ 0 ∑ F (∣ Φ ∣ ) k ( 1 ) 1 < ∞.
Proof of Proposition 31. Since the assumption on Φ \Phi Φ is stated with respect to the operator norm, which satisfies Equation (27) , we may assume without loss of generality that Φ \Phi Φ is entry-wise non-negative, i.e., for all k ∈ Z k \in\mathbb{Z} k ∈ Z , every entry of the matrix Φ ( k ) \Phi(k) Φ ( k ) is non-negative.
Let V ~ 1 , … , V ~ N \widetilde{V}_1,\ldots,\widetilde{V}_N V 1 , … , V N be the strongly connected components of the summary graph smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) , and assume that they are ordered such that for every ( v , m ) ∈ V ~ i (v,m) \in\widetilde{V}_i ( v , m ) ∈ V i and ( w , n ) ∈ V ~ j (w,n) \in\widetilde{V}_j ( w , n ) ∈ V j with i < j i<j i < j , it holds that P smry ( G ) ( ( v , m ) , ( w , n ) ) = ∅ \mathcal{P}_{\operatorname{smry}(\mathcal{G})}((v,m),(w,n))=\varnothing P smry ( G ) (( v , m ) , ( w , n )) = ∅ .
Now let i ≤ j i \leq j i ≤ j , and ( v , m ) ∈ V ~ i (v,m) \in\widetilde{V}_i ( v , m ) ∈ V i and ( w , n ) ∈ V ~ j (w,n) \in\widetilde{V}_j ( w , n ) ∈ V j . Furthermore, let us denote by I ( i , j ) = { x 1 < ⋯ < x k ∣ x l ∈ [ i , j ] , x 1 = i , x k = j } I(i,j)=\{x_1<\cdots<x_k\mid x_l\in[i,j],x_1=i,x_k=j\} I ( i , j ) = { x 1 < ⋯ < x k ∣ x l ∈ [ i , j ] , x 1 = i , x k = j } . Then note that concatenation of paths on smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) yields an isomorphism of sets:
P smry ( G ) ( V ~ i , V ~ j ) ≅ ⋃ x ∈ I ( i , j ) P smry ( G ) ( V ~ x 1 , V ~ x 1 ) × ∏ l = 2 k D V ~ x l − 1 , V ~ x l ′ × P smry ( G ) ( V ~ x l , V ~ x l ) , (28) \mathcal{P}_{\operatorname{smry}(\mathcal{G})}(\widetilde{V}_i,\widetilde{V}_j)\cong\bigcup_{\mathbf{x}\in I(i,j)}\mathcal{P}_{\operatorname{smry}(\mathcal{G})}(\widetilde{V}_{x_1},\widetilde{V}_{x_1})\times\prod_{l=2}^{k}D'_{\widetilde{V}_{x_{l-1}},\widetilde{V}_{x_l}}\times\mathcal{P}_{\operatorname{smry}(\mathcal{G})}(\widetilde{V}_{x_l},\widetilde{V}_{x_l}), \tag*{(28)} P smry ( G ) ( V i , V j ) ≅ x ∈ I ( i , j ) ⋃ P smry ( G ) ( V x 1 , V x 1 ) × l = 2 ∏ k D V x l − 1 , V x l ′ × P smry ( G ) ( V x l , V x l ) , (28)
where D V ~ x , V ~ y ′ = D ∩ ( V ~ x × V ~ y ) D'_{\widetilde{V}_x,\widetilde{V}_y}=D\cap(\widetilde{V}_x\times\widetilde{V}_y) D V x , V y ′ = D ∩ ( V x × V y ) , and P smry ( G ) ( V ~ x , V ~ y ) = ⋃ ( v , w ) ∈ V ~ x × V ~ y P smry ( G ) ( v , w ) \mathcal{P}_{\operatorname{smry}(\mathcal{G})}(\widetilde{V}_x,\widetilde{V}_y)=\bigcup_{(v,w)\in\widetilde{V}_x\times\widetilde{V}_y}\mathcal{P}_{\operatorname{smry}(\mathcal{G})}(v,w) P smry ( G ) ( V x , V y ) = ⋃ ( v , w ) ∈ V x × V y P smry ( G ) ( v , w ) .
By combining Lemma 29 with the bijection (28) , we get that
[ S ( Ψ ) ] V ~ i , V ~ j = ∑ i = x 1 < ⋯ < x k = j [ S ( Ψ ) ] V ~ x 1 , V ~ x 1 ∗ ∏ l = 2 k [ Ψ ] V ~ x l − 1 , V ~ x l ∗ [ S ( Ψ ) ] V ~ x l , V ~ x l . (29) [S(\Psi)]_{\widetilde{V}_i,\widetilde{V}_j}=\sum_{i=x_1<\cdots<x_k=j}[S(\Psi)]_{\widetilde{V}_{x_1},\widetilde{V}_{x_1}}*\prod_{l=2}^{k}[\Psi]_{\widetilde{V}_{x_{l-1}},\widetilde{V}_{x_l}}*[S(\Psi)]_{\widetilde{V}_{x_l},\widetilde{V}_{x_l}}. \tag*{(29)} [ S ( Ψ ) ] V i , V j = i = x 1 < ⋯ < x k = j ∑ [ S ( Ψ ) ] V x 1 , V x 1 ∗ l = 2 ∏ k [ Ψ ] V x l − 1 , V x l ∗ [ S ( Ψ ) ] V x l , V x l . (29)
The right hand side is a finite sum in which every summand is convolution of finitely many factors. In particular, to prove the proposition, it suffices to show that each of those factors is absolutely summable. In order to do that, note that the definition of strictly connected components and Proposition 30 imply the chain of identities
[ S ( Ψ ) ] V ~ x , V ~ x = S ( [ Ψ ] V ~ x , V ~ x ) = S ( [ Φ ] V ~ x , V ~ x ) , [S(\Psi)]_{\widetilde{V}_x,\widetilde{V}_x}=S([\Psi]_{\widetilde{V}_x,\widetilde{V}_x})=S([\Phi]_{\widetilde{V}_x,\widetilde{V}_x}), [ S ( Ψ ) ] V x , V x = S ([ Ψ ] V x , V x ) = S ([ Φ ] V x , V x ) ,
where the assumption on Φ \Phi Φ together with Lemma 32 yields that the filter on the right hand side is absolutely summable. To see that Ψ \Psi Ψ is absolutely summable, note that the assumption on Φ \Phi Φ implies that ∥ Φ v , v ⊤ ∥ 1 < 1 \|\Phi_{v,v}^{\top}\|_1<1 ∥ Φ v , v ⊤ ∥ 1 < 1 . Hence, the definition of Ψ \Psi Ψ in combination with Lemma 32 (applied to an appropriate subgraph of G \mathcal{G} G ) implies that Ψ \Psi Ψ is absolutely summable. To summarize, the filter S ( Ψ ) S(\Psi) S ( Ψ ) is absolutely summable. Finally, combining Proposition 30 with Proposition 28 gives that ι ∗ S ( Λ ) = S ( Φ ) = S ( Ψ ) \iota* S(\Lambda)=S(\Phi)=S(\Psi) ι ∗ S ( Λ ) = S ( Φ ) = S ( Ψ ) . Since we assumed Φ \Phi Φ to be entry-wise non-negative, it follows that also Λ \Lambda Λ and ι \iota ι are entry-wise non-negative. Furthermore, since ι ( 0 ) \iota(0) ι ( 0 ) is the identity matrix, it holds entry-wise that 0 ≤ S ( Λ ) ( k ) ≤ S ( Φ ) ( k ) 0\leq S(\Lambda)(k)\leq S(\Phi)(k) 0 ≤ S ( Λ ) ( k ) ≤ S ( Φ ) ( k ) , for every k ∈ Z k\in\mathbb{Z} k ∈ Z . In particular, the absolute summability of S ( Φ ) S(\Phi) S ( Φ ) implies that S ( Λ ) S(\Lambda) S ( Λ ) is absolutely summable as well.
Finally, that F ( S ( Φ ) ) \mathcal{F}(S(\Phi)) F ( S ( Φ )) takes the form of a rational function that is well defined for every ∣ z ∣ ≤ 1 |z|\leq1 ∣ z ∣ ≤ 1 follows by applying the Fourier transform to Equality (29) , as it results in a sum of products of rational functions, and all of those are well-defined for ∣ z ∣ ≤ 1 |z|\leq1 ∣ z ∣ ≤ 1 by Lemma 32. ∎
We conclude this section with the observation that S ( H ) S(H) S ( H ) admits an explicit parameterization in terms of the SVAR filter Φ \Phi Φ . Associated with Φ \Phi Φ , we define Φ I : Z → Bmat ( R , V , [ 0 , P − 1 ] ) \Phi^{\mathrm{I}}: \mathbb{Z}\to\operatorname{Bmat}(\mathbb{R},V,[0,P-1]) Φ I : Z → Bmat ( R , V , [ 0 , P − 1 ]) to be the P P P -stationary filter whose block entry Φ u , v I \Phi^{\mathrm{I}}_{u,v} Φ u , v I is equal to Φ v , v \Phi_{v,v} Φ v , v if u = v u=v u = v , and equal to the zero-filter otherwise.
Corollary 33. Suppose ( G , Φ , η ) (\mathcal{G},\Phi,\eta) ( G , Φ , η ) is a P P P -stationary SVAR model, where Φ \Phi Φ is as in Proposition 31, then
S ( H ) = ( I − F ( Φ I ) ) ( I − F ( Φ ) ) − 1 . S(H)=(I-\mathcal{F}(\Phi^{\mathrm{I}}))(I-\mathcal{F}(\Phi))^{-1}. S ( H ) = ( I − F ( Φ I )) ( I − F ( Φ ) ) − 1 .
The P P P -stationary SEP representation of a stationary SVAR process
Proof of Proposition 9. First, it follows from the definition that T r ( R , P ) \mathcal{T}_r(\mathbb{R},P) T r ( R , P ) is closed under addition and scalar multiplication. Let A , B ∈ T r ( R , P ) A,B\in\mathcal{T}_r(\mathbb{R},P) A , B ∈ T r ( R , P ) .
MULTIPLICATIVE: We now show that A B ∈ T r ( R , P ) AB\in\mathcal{T}_r(\mathbb{R},P) A B ∈ T r ( R , P ) . Let A 0 , … , A R − 1 ∈ R ( z ) P × P A_0,\ldots,A_{R-1}\in\mathbb{R}(z)^{P\times P} A 0 , … , A R − 1 ∈ R ( z ) P × P and B 0 , … , B R − 1 ∈ R ( z ) P × P B_0,\ldots,B_{R-1}\in\mathbb{R}(z)^{P\times P} B 0 , … , B R − 1 ∈ R ( z ) P × P be the matrices specifying A A A resp. B B B . Let us compute, for i , j ∈ [ 0 , P − 1 ] i,j\in[0,P-1] i , j ∈ [ 0 , P − 1 ] :
C j : = [ A B ] 0 , j = ∑ l = 0 R − 1 A 0 , l B l , j = ∑ l = 0 j A l B j − l + ∑ l = j + 1 R − 1 r A l B R − l + j (30) C_j:=[AB]_{0,j}=\sum_{l=0}^{R-1}A_{0,l}B_{l,j}=\sum_{l=0}^{j}A_lB_{j-l}+\sum_{l=j+1}^{R-1}rA_lB_{R-l+j} \tag*{(30)} C j := [ A B ] 0 , j = l = 0 ∑ R − 1 A 0 , l B l , j = l = 0 ∑ j A l B j − l + l = j + 1 ∑ R − 1 r A l B R − l + j (30)
Let i , j ∈ [ 0 , R − 1 ] i,j \in[0,R-1] i , j ∈ [ 0 , R − 1 ] and assume that i ≤ j i \leq j i ≤ j , we wish to show that [ A B ] i , j = γ j − i [AB]_{i,j}=\gamma_{j-i} [ A B ] i , j = γ j − i .
[ A B ] i , j = ∑ l = 0 i − 1 A i , l B l , j + ∑ l = i j A i , l B l , j + ∑ l = j + 1 R − 1 A i , l B l , j = ∑ l = 0 i − 1 r A R − i + l B j − l + ∑ l = i j A l − i B j − l + ∑ l = j + 1 R − 1 A l − i r B R − l + j \begin{aligned}
[AB]_{i,j}=\sum_{l=0}^{i-1}A_{i,l}B_{l,j}+\sum_{l=i}^{j}A_{i,l}B_{l,j}+\sum_{l=j+1}^{R-1}A_{i,l}B_{l,j} \\
&=\sum_{l=0}^{i-1}rA_{R-i+l}B_{j-l}+\sum_{l=i}^{j}A_{l-i}B_{j-l}+\sum_{l=j+1}^{R-1}A_{l-i}rB_{R-l+j}
\end{aligned} [ A B ] i , j = l = 0 ∑ i − 1 A i , l B l , j + l = i ∑ j A i , l B l , j + l = j + 1 ∑ R − 1 A i , l B l , j = l = 0 ∑ i − 1 r A R − i + l B j − l + l = i ∑ j A l − i B j − l + l = j + 1 ∑ R − 1 A l − i r B R − l + j
Now make the following substitutions: l = − R + i + l ′ l=-R+i+l' l = − R + i + l ′ for the first summand, l = l ′ + i l=l'+i l = l ′ + i for the other two summands, which yields:
[ A B ] i , j = ∑ l ′ = R − i R − 1 r A l ′ B R − l ′ + ( j − i ) + ∑ l ′ = 0 j − i A l ′ B ( j − i ) − l ′ + ∑ l ′ = j − i + 1 R − i − 1 A l ′ r B R − l ′ + ( j − i ) = ∑ l ′ = 0 j − i A l ′ B ( j − i ) − l ′ + ∑ l ′ = j − i + 1 R − 1 r A l ′ B R − l ′ + ( j − i ) = C j − i . \begin{aligned}
[AB]_{i,j}=\sum_{l'=R-i}^{R-1}rA_{l'}B_{R-l'+(j-i)}+\sum_{l'=0}^{j-i}A_{l'}B_{(j-i)-l'}+\sum_{l'=j-i+1}^{R-i-1}A_{l'}rB_{R-l'+(j-i)} \\
&=\sum_{l'=0}^{j-i}A_{l'}B_{(j-i)-l'}+\sum_{l'=j-i+1}^{R-1}rA_{l'}B_{R-l'+(j-i)} \\
&=C_{j-i}.
\end{aligned} [ A B ] i , j = l ′ = R − i ∑ R − 1 r A l ′ B R − l ′ + ( j − i ) + l ′ = 0 ∑ j − i A l ′ B ( j − i ) − l ′ + l ′ = j − i + 1 ∑ R − i − 1 A l ′ r B R − l ′ + ( j − i ) = l ′ = 0 ∑ j − i A l ′ B ( j − i ) − l ′ + l ′ = j − i + 1 ∑ R − 1 r A l ′ B R − l ′ + ( j − i ) = C j − i .
Now assume i > j i>j i > j and compute
[ A B ] i , j = ∑ l = 0 j A i , l B l , j + ∑ l = j + 1 i − 1 A i , l B l , j + ∑ l = i P − 1 A i , l B l , j = ∑ l = 0 j r A R − i + l B j − l + ∑ l = j + 1 i − 1 r A R − i + l r B R − l + j + ∑ l = i R − 1 A l − i r B R − l + j = ∑ l ′ = R − i R − i + j r A l ′ B R − i + j − l ′ + ∑ l ′ = R − i + j + 1 R − 1 r A l ′ r B 2 R − i + j − l ′ + ∑ l ′ = 0 R − i − 1 A l ′ r B R − i + j − l ′ = r ( ∑ l ′ = 0 R − i + j A l ′ B R − i + j − l ′ + ∑ l ′ = R − i + j + 1 R − 1 r A l ′ B 2 R − i + j − l ′ ) = r C R − i + j . \begin{aligned}
[AB]_{i,j}=\sum_{l=0}^{j}A_{i,l}B_{l,j}+\sum_{l=j+1}^{i-1}A_{i,l}B_{l,j}+\sum_{l=i}^{P-1}A_{i,l}B_{l,j} \\
&=\sum_{l=0}^{j}rA_{R-i+l}B_{j-l}+\sum_{l=j+1}^{i-1}rA_{R-i+l}rB_{R-l+j}+\sum_{l=i}^{R-1}A_{l-i}rB_{R-l+j} \\
&=\sum_{l'=R-i}^{R-i+j}rA_{l'}B_{R-i+j-l'}+\sum_{l'=R-i+j+1}^{R-1}rA_{l'}rB_{2R-i+j-l'}+\sum_{l'=0}^{R-i-1}A_{l'}rB_{R-i+j-l'} \\
&=r\left(\sum_{l'=0}^{R-i+j}A_{l'}B_{R-i+j-l'}+\sum_{l'=R-i+j+1}^{R-1}rA_{l'}B_{2R-i+j-l'}\right) \\
&=rC_{R-i+j}.
\end{aligned} [ A B ] i , j = l = 0 ∑ j A i , l B l , j + l = j + 1 ∑ i − 1 A i , l B l , j + l = i ∑ P − 1 A i , l B l , j = l = 0 ∑ j r A R − i + l B j − l + l = j + 1 ∑ i − 1 r A R − i + l r B R − l + j + l = i ∑ R − 1 A l − i r B R − l + j = l ′ = R − i ∑ R − i + j r A l ′ B R − i + j − l ′ + l ′ = R − i + j + 1 ∑ R − 1 r A l ′ r B 2 R − i + j − l ′ + l ′ = 0 ∑ R − i − 1 A l ′ r B R − i + j − l ′ = r l ′ = 0 ∑ R − i + j A l ′ B R − i + j − l ′ + l ′ = R − i + j + 1 ∑ R − 1 r A l ′ B 2 R − i + j − l ′ = r C R − i + j .
This shows that the set T P \mathcal{T}^{P} T P is multiplicative.
CLOSED UNDER INVERSION: Let A ∈ T r ( R , P ) A\in T_r(R,P) A ∈ T r ( R , P ) be invertible over R ( z ) \mathbb{R}(z) R ( z ) , i.e., det ( A ) ≠ 0 \det(A)\ne0 det ( A ) = 0 as a rational function. Let p A ( λ ) = det ( λ I − A ) p_A(\lambda)=\det(\lambda I-A) p A ( λ ) = det ( λ I − A ) be the characteristic polynomial over R ( z ) \mathbb{R}(z) R ( z ) taking the form
p A ( λ ) = λ N + f N − 1 λ N − 1 + ⋯ + f 1 λ + f 0 , p_A(\lambda)=\lambda^N+f_{N-1}\lambda^{N-1}+\cdots+f_1\lambda+f_0, p A ( λ ) = λ N + f N − 1 λ N − 1 + ⋯ + f 1 λ + f 0 ,
where N = R P N=RP N = R P and f i ∈ R ( z ) f_i\in\mathbb{R}(z) f i ∈ R ( z ) . Note that f 0 = ( − 1 ) N det ( A ) f_0=(-1)^N\det(A) f 0 = ( − 1 ) N det ( A ) . The Cayley Hamilton Theorem now implies that p A ( A ) = 0 p_A(A)=0 p A ( A ) = 0 , which gives the following expression for the inverse
A − 1 = ( − 1 ) N + 1 det ( A ) − 1 ( A N − 1 + f N − 1 A N − 2 + ⋯ + f 2 A + f 1 I ) . A^{-1}=(-1)^{N+1}\det(A)^{-1}\left(A^{N-1}+f_{N-1}A^{N-2}+\cdots+f_2A+f_1I\right). A − 1 = ( − 1 ) N + 1 det ( A ) − 1 ( A N − 1 + f N − 1 A N − 2 + ⋯ + f 2 A + f 1 I ) .
Since T r ( R , P ) T_r(R,P) T r ( R , P ) is closed under multiplication, addition and scalar multiplication, it follows that A − 1 A^{-1} A − 1 is an element in T r ( R , P ) T_r(R,P) T r ( R , P ) .
Graphical criterion for non-vanishing minors
Proof of Proposition 17. Let T = ( τ 1 , … , τ N ) ∈ T s m r y ( G ) ( X , Y ) T=(\tau_1,\ldots,\tau_N)\in T_{\mathrm{smry}(\mathcal{G})}(X,Y) T = ( τ 1 , … , τ N ) ∈ T smry ( G ) ( X , Y ) be a non-intersecting trek system, where X = ( ( v i , m i ) ∣ i ∈ [ N ] ) X=((v_i,m_i)\mid i\in[N]) X = (( v i , m i ) ∣ i ∈ [ N ]) and Y = ( ( y i , n i ) ∣ i ∈ [ N ] ) . Y=((y_i,n_i)\mid i\in[N]). Y = (( y i , n i ) ∣ i ∈ [ N ]) . Now define a time series subgraph G ~ ≤ G \widetilde{\mathcal{G}}\leq\mathcal{G} G ≤ G . First, we will show that T T T can always be adapted such that every vertex visited by some unique τ i \tau_i τ i is visited by τ i \tau_i τ i exactly once. Note that T T T being non-intersecting implies that every vertex is visited by at most one τ i \tau_i τ i . Let v v v be a vertex appearing on trek τ i = ( π i R , top i , π i L ) \tau_i=(\pi_i^R,\operatorname{top}_i,\pi_i^L) τ i = ( π i R , top i , π i L ) , where
π i R = top i → v 1 R → ⋯ → v M i R \pi_i^R=\operatorname{top}_i\to v_1^R\to\cdots\to v_{M_i}^R π i R = top i → v 1 R → ⋯ → v M i R
π i L = top i → v 1 L → ⋯ → v Q i L . \pi_i^L=\operatorname{top}_i\to v_1^L\to\cdots\to v_{Q_i}^L. π i L = top i → v 1 L → ⋯ → v Q i L .
Then define a R = min { i ∣ v i R = v } a_{R}=\min\{i\mid v_{i}^{R}=v\} a R = min { i ∣ v i R = v } and a L a_{L} a L analogously. Furthermore, define b R = max { i ∣ v i R = v } b_{R}=\max\{i\mid v_{i}^{R}=v\} b R = max { i ∣ v i R = v } and b L b_{L} b L analogously. If v v v appears on only one side, say the left hand side, then replace τ i \tau_{i} τ i by τ i ′ = ( π i L ′ , top i , π i R ) \tau_{i}^{\prime}=(\pi_{i}^{L^{\prime}},\operatorname{top}_{i},\pi_{i}^{R}) τ i ′ = ( π i L ′ , top i , π i R ) , where
π i L ′ = { top i → v 1 L → ⋯ → v a L L → v b L + 1 L → ⋯ → v Q i L b L < Q i , top i → v 1 L → ⋯ → v a L L b L = Q i . \pi_{i}^{L^{\prime}}=
\begin{cases}
\operatorname{top}_{i}\to v_{1}^{L}\to\cdots\to v_{a_{L}}^{L}\to v_{b_{L}+1}^{L}\to\cdots\to v_{Q_{i}}^{L} & b_{L}<Q_{i},\\
\operatorname{top}_{i}\to v_{1}^{L}\to\cdots\to v_{a_{L}}^{L} & b_{L}=Q_{i}.
\end{cases} π i L ′ = { top i → v 1 L → ⋯ → v a L L → v b L + 1 L → ⋯ → v Q i L top i → v 1 L → ⋯ → v a L L b L < Q i , b L = Q i .
If v v v appears on both sides, then we replace τ i \tau_{i} τ i by τ i ′ = ( π i L ′ , v , π i R ′ ) \tau_{i}^{\prime}=(\pi_{i}^{L^{\prime}},v,\pi_{i}^{R^{\prime}}) τ i ′ = ( π i L ′ , v , π i R ′ ) , where
π i L ′ = v → v b L + 1 L → ⋯ → v Q i L , \pi_{i}^{L^{\prime}}=v\to v_{b_{L}+1}^{L}\to\cdots\to v_{Q_{i}}^{L}, π i L ′ = v → v b L + 1 L → ⋯ → v Q i L ,
π i R ′ = v → v b R + 1 R → ⋯ → v M i R . \pi_{i}^{R^{\prime}}=v\to v_{b_{R}+1}^{R}\to\cdots\to v_{M_{i}}^{R}. π i R ′ = v → v b R + 1 R → ⋯ → v M i R .
Note that the trek τ i ′ \tau_{i}^{\prime} τ i ′ visits the vertex v v v exactly once. Proceeding this way, any non-intersecting trek system can be reduced to a trek system which visits every vertex at most once. So from now on we assume T T T to be such trek system.
For every link ( u , l ) → ( u ′ , l ′ ) (u,l)\to(u^{\prime},l^{\prime}) ( u , l ) → ( u ′ , l ′ ) appearing on some trek τ i \tau_{i} τ i , we add the relation u → ( l , l ′ ) k u ′ u\xrightarrow{(l,l^{\prime})}_{k}u^{\prime} u ( l , l ′ ) k u ′ to smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) , where
k = min { j ∣ u → ( l , l ′ ) j u ′ on G } . k=\min\{j\mid u\xrightarrow{(l,l^{\prime})}_{j}u^{\prime}\text{ on }\mathcal{G}\}. k = min { j ∣ u ( l , l ′ ) j u ′ on G } .
Furthermore, let smry ( G ~ ) \operatorname{smry}(\widetilde{\mathcal{G}}) smry ( G ) be the summary graph of G ~ \widetilde{\mathcal{G}} G . By construction and the assumption that T T T is non-intersecting, it follows that smry ( G ~ ) \operatorname{smry}(\widetilde{\mathcal{G}}) smry ( G ) is an acyclic subgraph of smry ( G ) \operatorname{smry}(\mathcal{G}) smry ( G ) . Secondly, we have that
T smry ( G ) ( ( v , m ) , ( w , n ) ) = { { τ i } ( v , m ) = ( v i , m i ) and ( w , n ) = ( w α ( i ) , n α ( i ) ) , { ϵ ( v , m ) } ( v , m ) = ( w , n ) , ∅ otherwise . T_{\operatorname{smry}(\mathcal{G})}((v,m),(w,n))=
\begin{cases}
\{\tau_{i}\} & (v,m)=(v_{i},m_{i})\text{ and }(w,n)=(w_{\alpha(i)},n_{\alpha(i)}),\\
\{\epsilon_{(v,m)}\} & (v,m)=(w,n),\\
\varnothing& \text{otherwise}.
\end{cases} T smry ( G ) (( v , m ) , ( w , n )) = ⎩ ⎨ ⎧ { τ i } { ϵ ( v , m ) } ∅ ( v , m ) = ( v i , m i ) and ( w , n ) = ( w α ( i ) , n α ( i ) ) , ( v , m ) = ( w , n ) , otherwise .
In particular, any two vertices are connected by at most one trek. Applying Proposition 16 to the subgraph G ~ \widetilde{\mathcal{G}} G , we get the following expression for the spectrum
[ S G ~ ] X , Y = ( 1 { α ( i ) } ( j ) S G ~ ( τ i ) ∣ i , j ∈ [ N ] ) ∈ K G ~ X × Y . [S_{\widetilde{\mathcal{G}}}]_{X,Y}=\left(\mathbf{1}_{\{\alpha(i)\}}(j)S_{\widetilde{\mathcal{G}}}^{(\tau_{i})}\mid i,j\in[N]\right)\in K_{\widetilde{\mathcal{G}}}^{X\times Y}. [ S G ] X , Y = ( 1 { α ( i )} ( j ) S G ( τ i ) ∣ i , j ∈ [ N ] ) ∈ K G X × Y .
By construction, we have that
S G ~ ( τ i ) = ω z i ( k i ) F ( ψ π i L ′ ) ( F ( ψ π i R ′ ) ) ∗ , S_{\widetilde{\mathcal{G}}}^{(\tau_{i})}=\omega_{z_{i}}^{(k_{i})}\mathcal{F}(\psi^{\pi_{i}^{L^{\prime}}})(\mathcal{F}(\psi^{\pi_{i}^{R^{\prime}}}))^{*}, S G ( τ i ) = ω z i ( k i ) F ( ψ π i L ′ ) ( F ( ψ π i R ′ ) ) ∗ ,
where top i = ( z i , k i ) \operatorname{top}_{i}=(z_{i},k_{i}) top i = ( z i , k i ) and F ( ψ π i L ′ ) = ∏ j = 1 K i − 1 F ( Φ v j L , v j + 1 L ( l j , l j + 1 ) ) \mathcal{F}(\psi^{\pi_{i}^{L^{\prime}}})=\prod_{j=1}^{K_{i-1}}\mathcal{F}(\Phi_{v_{j}^{L},v_{j+1}^{L}}^{(l_{j},l_{j+1})}) F ( ψ π i L ′ ) = ∏ j = 1 K i − 1 F ( Φ v j L , v j + 1 L ( l j , l j + 1 ) ) and analogously for F ( ψ π i R ) \mathcal{F}(\psi^{\pi_{i}^{R}}) F ( ψ π i R ) . Note that all of those are non-zero polynomials in the indeterminates of the SVAR coefficients, so that
det ( [ S G ~ ] X , Y ) = ( − 1 ) Π ∏ i = 1 N S G ~ ( τ i ) \det([S_{\widetilde{\mathcal{G}}}]_{X,Y})=(-1)^{\Pi}\prod_{i=1}^{N}S_{\widetilde{\mathcal{G}}}^{(\tau_{i})} det ([ S G ] X , Y ) = ( − 1 ) Π i = 1 ∏ N S G ( τ i )
is a product of non-zero rational functions, and therefore non-zero overall, and since this implies that also det ( [ S G ] X , Y ) ∈ K G \det([S_{\mathcal{G}}]_{X,Y})\in K_{\mathcal{G}} det ([ S G ] X , Y ) ∈ K G is non-zero, we conclude that [ S G ] X , Y ∈ K G X , Y [S_{\mathcal{G}}]_{X,Y}\in K_{\mathcal{G}}^{X,Y} [ S G ] X , Y ∈ K G X , Y is invertible.