The parallel complexity of function approximation (Q1179031)

From MaRDI portal
scientific article
Language Label Description Also known as
English
The parallel complexity of function approximation
scientific article

    Statements

    The parallel complexity of function approximation (English)
    0 references
    0 references
    26 June 1992
    0 references
    The author defines a parallel algorithm \(\phi\) with \(n\) input variables \(x_ 1,\dots,x_ n\) as a finite directed graph \((G,M)\) having the following properties (\(G\) being the set of nodes, \(M\) the set of edges, \(M\subset G\times G)\): 1. \(G\subseteq\{N\cup\{0\}\}\times N\) and \((0,1),\dots,(0,n)\in G\) (\(N\) is the set of natural numbers). 2. Every node \((j,k)\in G\) with \(j>0\) has exactly 2 incoming edges: \(((j_ l,k_ l),(j,k))\), \(((j_ r,k_ r),(j,k))\in M\), where \(jl,j_ r<j\). 3. To every node \((j,k)\), \(j>0\), belongs an operation \(op(j,k)\in\{+,-,\ast\}\). To each node \((j,k)\) of an algorithm \(\phi\) is associated recursively a function \(\psi_{(j,k)}\) of the input variables as follows: \(\psi_{(0,k)}(x_ 1,\dots,x_ n)=x_ k\) if \(1\leq k\leq n\), \(\psi_{(0,k)}(x_ 1,\dots,x_ n)=a_ k\) if \(k\geq n\), \(\psi_{(j,k)}(x_ 1,\dots,x_ n)=\psi_{(j_ l,k_ l)}(x_ 1,\dots,x_ n) op(j,k)\psi_{(j_ r,k_ r)}(x_ 1,\dots,x_ n)\) if \(j>1\). The real numbers \(a_ k\) are the input constants. The interpretation of this is that at time step \(j>0\) the processor number \(k\) performs the operation \(op(j,k)\) at the operands \(\psi_{(j_ l,k_ l)}(x_ 1,\dots,x_ n)\), \(\psi_{(j_ r,k_ r)}(x_ 1,\dots,x_ n)\). The complexity of a prallel algorithm is given by \(\hbox{comp} \phi=\max\{j\in N\mid\;\exists k\in N:\;(j,k)\in G\}\) and the parallelism of \(\phi\) is defined by \(p(\phi)=\max\{\hbox{card}\{(j,k)\in G\mid\;j=l\}\mid\;l\in N\}\). If \(f: R^ n\to R^ 1\), it is said that the parallel algorithm \(\phi\) computes the function \(f\), if there exists a \((j_ 0,k_ 0)\in G\) such that \(\psi_{(j_ \sigma,k_ \sigma)}(x)=f(x)\) for all \(x=(x_ 1,\dots,x_ n)\in R^ n\). For every number \(p\) of processors, \(1\leq p\leq\infty\), the parallel complexity of a function \(f\) is given by \(p-\hbox{comp} f=\min\{\hbox{comp} \phi\mid\;\phi\hbox{ computes }f,\;p(\phi)\leq p\}\). If \(\Lambda\) is a class of functions, then define \(p-\hbox{comp}(\Lambda)=\max\{p-\hbox{comp} q\mid\;q\in\Lambda\}\). If \(\Lambda\) is a class of functions \(f: M\to R^ 1\), \(M\subset R^ n\), then for \(f\in\Lambda\) define \(p-\hbox{comp}(f;\varepsilon)=\min\{p-\hbox{comp} \phi\mid\;\phi\hbox{ computes a } \psi\hbox{ with } |\psi(x)-f(x)|\leq\varepsilon\}\) for all \(x\in M\) and \(p-\hbox{comp}(\Lambda;\varepsilon)= \max\{p- \hbox{comp}(f;\varepsilon)\mid\;f\in\Lambda\}\). The author obtains lower and upper bounds for \(p-\hbox{comp}(P^ r)\), \(p-\hbox{comp}(P_ r)\), \(1\leq p\leq\infty\), where \(P^ r\) resp. \(P_ r\) is the set of polynomials in \(R^ n\) of the degree at most \(r\) resp. of the degree at most \(r\) in regard to separate variables. He also obtains bounds for \(p-\hbox{comp}(\Lambda:\varepsilon)\) for some classes \(\Lambda\) of continuous functions.
    0 references
    parallel complexity
    0 references
    smooth functions
    0 references
    lower and upper bounds
    0 references
    0 references

    Identifiers