The general theory of canonical correlation and its relation to functional analysis

Abstract

The classical theory of canonical correlation is concerned with a standard description of the relationship between any linear combination of ρ random variablesxs, and any linear combination ofqrandom variablesytinsofar as this relation can be described in terms of correlation. Lancaster [1] has extended this theory, forp=q= 1, to include a description of the correlation of any function of a random variablexand any function of a random variabley(both functions having finite variance) for a class of joint distributions ofxandywhich is very general. It is the purpose of this paper to derive Lancaster's results from general theorems concerning the spectral decomposition of operators on a Hilbert space. These theorems lend themselves easily to the generalisation of the theory to situations wherepandqare not finite. In the case of Gaussian, stationary, processes this generalisation is equivalent to the classical spectral theory and corresponds to a canonical reduction of a (finite) sample of data which is basic. The theory also then extends to any number of processes. In the Gaussian case, also, the present discussion-is connected with the results of Gelfand and Yaglom [2] relating to the amount of information in one random process about another.

Modeling and Forecasting Daily Electricity Load Curves: A Hybrid Approach

“…Our method is closely related to the canonical correlation analysis yet we focus on regressing Y i (·) on X i (·), and thus Y i (·) and X i (·) are not treated on an equal footing, which is different from, and much simpler than, the canonical correlation analysis. The literature on functional canonical correlation analysis includes Hannan (1961), Silverman (1996), He et al (2003), Cupidon et al (2008), Eubank and Hsing (2008) and Yang et al (2011).…”

Optical control of scattering, absorption and lineshape in nanoparticles

“…The spatial correlations that appear in this theory can be interpreted geometrically. The principal cosine between s n and i n is defined by the overlap integral, or correlation [19], as cos (ξ n ) ≡ i n · s n , with the arbitrary phases of i n , s n chosen so that the integral is either positive or null. The corresponding function sin (ξ n ) is the orthogonal distance between s n and i n .…”

The geometrical nature of optical resonances: from a sphere to fused dimer nanoparticles

“…The statistical correlation [8] and the orthogonal distance between s n and i n in H, which are calculated through the overlap integral of the modes on the surface of the particle [4], are the so-called principal cosines and sines, cos(ξ n ) and sin(ξ n ). The angles ξ are invariant under unitary transformation and characterize the geometry of the subspaces of the internal and scattered solutions in H. This geometry is induced by the particular particle through the surface integrals [4].…”