measuring Statistical Dependence with Hilbert-Schmidt Norm

From statwiki
Jump to navigation Jump to search

"Hilbert-Schmist Norm of the Cross-Covariance operator" is proposed as an independence criterion in reproducing kernel Hilbert spaces (RKHSs). The measure is refereed to as Hilbert-Schmidt Independence Criterion, or HSIC. An empirical estimate of this measure is introduced which may be well used in many practical application such as independent Component Analysis (ICA), Maximum Variance Unfolding (MVU), feature extraction, feature selection, ... .

Introduction

In this article, a criterion is introduced which is used to measure the dependence between two multivariate random variables. More specifically, we are looking for an appropriate function of two random variables whose output is zero if the involving random variables are independent and achieves a high value if they are statistically dependent. If instead of "independence" we were looking for "uncorrelation" the situation would be much easier to handle. Because "correlation" of two random variables involves only expected values, as opposed to concept of independence between two random variables. However, fortunately, based on the following theorem we still can effectively tackle the problem of independence between two random variables:

Theorem: Two random variables [math]\displaystyle{ \,\!x }[/math] and [math]\displaystyle{ \,\!y }[/math] are independent if and only if any bounded continuous function of the two random variables are uncorrelated. (have zero covariance).

For example, for random variables [math]\displaystyle{ \,\!x }[/math] and [math]\displaystyle{ \,\!y }[/math] to be independent, all pairs [math]\displaystyle{ (x,y),\, (x^2,y),\,(x,e^y), \,(\sin(x),\log(y)),\, ... }[/math] have to be uncorrelated. That is why we need to deal with functions of random variables when examining the independence property. and That is why Hilbert spaces and concepts from "Functional Analysis" are involved in this article (Of course, no one likes to torture him/herself by using complicated mathematical expressions ;)).

How the rest of this note is organized

In the following, we first give an introductory note on the concepts from functional analysis and Hilbert spaces required in this paper. (No prior knowledge of these concepts is assumed here). Then we introduce the independence measure, namely, Hilbert-Schmidt Independence Criterion and explain why we expect it to work properly. Thirdly, we introduce an statistics which may be used to estimate the HSIC and discuss its convergence behavior. Finally, through some examples, we demonstrate how this criterion may be effectively employed to solve many well-known problems.

Simple introduction to Reproducing Kernel Hilbert Spaces

Basic concepts

Let [math]\displaystyle{ \mathcal{F} }[/math] denote the set containing all real-valued functions defined on domain [math]\displaystyle{ \,\!\mathcal{X} }[/math]. By real-valued, we mean the output of these functions is always a real number. (Note that this definition makes good sense considering the methodology mentioned in the introduction section.) We may note that this space is linear, as for any [math]\displaystyle{ f,g\in\mathcal{F} }[/math] and [math]\displaystyle{ a,b\in\mathbb{R} }[/math], [math]\displaystyle{ \,\!af+bg }[/math] will be a member of [math]\displaystyle{ \mathcal{F} }[/math], as well. On the other hand, we may define the concept of real-valued inner product between the elements of [math]\displaystyle{ \mathcal{F} }[/math]. The simplest way to do that probably is the familiar:
[math]\displaystyle{ \langle a,b\rangle=\int_{}^{}{f(t)g(t)dt} }[/math]
However, other definitions for the inner product is also possible. The only required conditions are that:
1. [math]\displaystyle{ \langle f,f\rangle\geq0,\,\forall\, f\in\mathcal{F} }[/math] with [math]\displaystyle{ \langle f,f\rangle=0 }[/math] iff [math]\displaystyle{ \,\!f=0 }[/math].
2. [math]\displaystyle{ \langle f,g\rangle=\langle g,f\rangle,\, \forall \,f,g\in \mathcal{F} }[/math].
3. [math]\displaystyle{ \langle af+bg,h\rangle=a\langle f,h\rangle+b\langle g,h\rangle,\,\forall\,f,g,h\in\mathcal{F} }[/math] and all real [math]\displaystyle{ \,\!a }[/math] and [math]\displaystyle{ \,\!b }[/math].

Endowed with the concept of inner product, the set [math]\displaystyle{ \mathcal{F} }[/math] here is, in fact, a Hilbert space. (In general, a Hilbert space is a (complete) linear space endowed with an inner product defined on the elements with the above properties. However, in this article we only deal with those Hilbert spaces whose elements are real-valued functions on [math]\displaystyle{ \mathcal{X} }[/math].)
A mapping from the Hilbert space to real numbers [math]\displaystyle{ L:\mathcal{F}\rightarrow \mathbb{R} }[/math] is called a functional. A functional is called linear if the underlying mapping is linear, i.e.:
[math]\displaystyle{ \,\!L(af+bg)=aL(f)+bL(g) }[/math]
for all [math]\displaystyle{ \,\!f }[/math] and [math]\displaystyle{ \,\!g }[/math] in [math]\displaystyle{ \mathcal{F} }[/math] and all real [math]\displaystyle{ \,\!a }[/math] and [math]\displaystyle{ \,\!b }[/math]. A simple (and yet very important) example of a functional is a function which gets an element of [math]\displaystyle{ \mathcal{F} }[/math] (which is a real-valued function on [math]\displaystyle{ \mathcal{X} }[/math]) as input and returns its value at a specific fixed point [math]\displaystyle{ \,\!x }[/math]. This functional is called Dirac evaluation functional and is denoted by [math]\displaystyle{ \delta_x:\mathcal{F} \rightarrow \mathbb{R} }[/math]. Thus we have [math]\displaystyle{ \,\!\delta_x(f)=f(x) }[/math]. It can be easily seen that [math]\displaystyle{ \,\!\delta_x }[/math] is a linear functional.
There is theorem on linear functionals named "Riesz representation Theorem" which plays a key role in functional analysis:
Theorem (Riesz representation): If [math]\displaystyle{ \,\!L }[/math] is a bounded linear functional on a Hilbert space [math]\displaystyle{ \mathcal{F} }[/math], then there exists a unique element [math]\displaystyle{ \,\!h_f }[/math] in [math]\displaystyle{ \mathcal{F} }[/math] such that [math]\displaystyle{ L(f)=\langle f,h_f\rangle }[/math] for all [math]\displaystyle{ \,\!f }[/math] in [math]\displaystyle{ \mathcal{F} }[/math].
As mentioned above, [math]\displaystyle{ \,\!\delta_x }[/math] is linear functional and thus according to Riesz theorem there is an element of [math]\displaystyle{ \mathcal{F} }[/math], denoted by [math]\displaystyle{ \,\!k_x }[/math] such that:
[math]\displaystyle{ \delta_x(f)=f(x)=\langle f,k_x\rangle }[/math] for all [math]\displaystyle{ f\in\mathcal{F} }[/math].

Reproducing Kernel Hilbert Spaces (RKHS)

Reproducing Kernel Hilbert Spaces are a subtype of Hilbert spaces introduced above. A Hilbert space is a RKHS if for all [math]\displaystyle{ x\in \mathcal{X} }[/math], the corresponding Dirac evaluation operator [math]\displaystyle{ \delta_x:\mathcal{F} \rightarrow \mathbb{R} }[/math] is a continuous linear functional. Based on this, Reproducing Kernel Hilbert Spaces, in fact, represent a family of Hilbert spaces with smooth real-valued functions as their elements.
To every RKHS we may assign a symmetric positive definite function [math]\displaystyle{ k:\,\mathcal{X}\times\mathcal{X}\rightarrow\mathbb{R} }[/math], named kernel. Using the Dirac evaluation functional, the kernel function is defined as:
[math]\displaystyle{ \,\!k(x,y):=k_x(y) }[/math]
It is not hard to see that:
[math]\displaystyle{ k(x,y)=k_x(y)=\langle k_x,k_y\rangle }[/math]
It may be shown that there is a unique correspondence between a RKHS and its kernel. In other words, given a Hilbert space, its kernel function would be identified uniquely; and inversely, if [math]\displaystyle{ \,\!k }[/math] is symmetric positive definite function of the form [math]\displaystyle{ k:\mathcal{X}\times \mathcal{X}\rightarrow \mathbb{R} }[/math], then there is a unique Hilbert space of functions on [math]\displaystyle{ \mathcal{X} }[/math] for which [math]\displaystyle{ \,\!k }[/math] is a reproducing kernel.
The elements of the RKHSs we consider here are always the same, i.e., the smooth real-valued functions on a set [math]\displaystyle{ \mathcal{X} }[/math]. However, they may have different inner products defined and thus different types of kernels.

The independence measure

Now we turn our attention back to the the problem of measuring independence between two (generally multivariate) random variables [math]\displaystyle{ \,\!x }[/math] and [math]\displaystyle{ \,\!y }[/math]. Induced by the idea described in the introduction section, we define RKHS [math]\displaystyle{ \mathcal{F} }[/math] from [math]\displaystyle{ \mathcal{X} }[/math] to [math]\displaystyle{ \mathbb{R} }[/math] containing all continuous bounded real-valued functions of [math]\displaystyle{ \,\!x }[/math], and RKHS [math]\displaystyle{ \mathcal{G} }[/math] from [math]\displaystyle{ \mathcal{Y} }[/math] to [math]\displaystyle{ \mathbb{R} }[/math] containing all continuous bounded real-valued functions of [math]\displaystyle{ \,\!y }[/math]. Here, [math]\displaystyle{ \mathcal{X} }[/math] and [math]\displaystyle{ \mathcal{Y} }[/math] denotes the support (set of possible values) of random variables [math]\displaystyle{ \,\!x }[/math] and [math]\displaystyle{ \,\!y }[/math] respectively. As mentioned before, we are particularly interested in the cross-covariance between elements of [math]\displaystyle{ \mathcal{F} }[/math] and [math]\displaystyle{ \mathcal{G} }[/math]:
[math]\displaystyle{ \text{cov}(f(x),g(y))=\mathbf{E}_{x,y}[f(x)g(y)]-\mathbf{E}_{x}[f(x)]\mathbf{E}_{y}[g(y)] }[/math]
In functional Analysis, an operator is a name given to a mapping which maps elements from one Hilbert space to elements of the other one. It may be shown that there exist a unique operator [math]\displaystyle{ \,\!C_{x,y}:\mathcal{F}\rightarrow \mathcal{G} }[/math] mapping elements of [math]\displaystyle{ \mathcal{F} }[/math] to elements of [math]\displaystyle{ \mathcal{G} }[/math] such that:
[math]\displaystyle{ \langle f,C_{x,y}(g)\rangle=\text{cov}(f,g) }[/math]
for all [math]\displaystyle{ f\in \mathcal{F} }[/math] and [math]\displaystyle{ g\in \mathcal{G} }[/math]. This operator is called Cross-covariance operator.
We may define the concept of norm for an operator. For example, consider an operator in the form of a matrix [math]\displaystyle{ C_{m\times n} }[/math] mapping vectors of size [math]\displaystyle{ n\times 1 }[/math] in [math]\displaystyle{ \mathbb{R}^n }[/math] to vectors of size [math]\displaystyle{ m\times 1 }[/math] in [math]\displaystyle{ \mathbb{R}^m }[/math]. Then the Frobenius norm of this matrix may be defined as the norm of the corresponding operator.
There are different norms defined for operators. One of them is called Hilbert-Schmidt (HS) norm and is defined as follows:
[math]\displaystyle{ \|C\|^2_{HS}:=\sum_{i,j}\langle Cv_i,u_j \rangle^2_{\mathcal{F}} }[/math]
where [math]\displaystyle{ \,\!u_j }[/math] and [math]\displaystyle{ \,\!v_i }[/math] are orthogonal bases of [math]\displaystyle{ \mathcal{F} }[/math] and [math]\displaystyle{ \mathcal{G} }[/math], respectively. It is easy to see that the Frobenius norm on matrices may be considered a spacial case of this norm.

Now, we are in a position to define our measure for dependence of two random variables. The measure is the "Hilbert-Schmidt norm of the cross-covariance operator":
[math]\displaystyle{ \text{HSIC}(p_{xy},\mathcal{F},\mathcal{G}):=\|C_{xy}\|^2_{HS} }[/math]
We may note that if [math]\displaystyle{ \|C_{xy}\|^2_{HS} }[/math] is zero, then the value of [math]\displaystyle{ \langle f,C_{x,y}(g)\rangle }[/math], i.e., [math]\displaystyle{ \,\!\text{cov}(f,g) }[/math], will be always zero for any [math]\displaystyle{ f\in\ \mathcal{F} }[/math] and [math]\displaystyle{ g\in \mathcal{G} }[/math] and thus based on the theorem given in the introduction section, the random variables [math]\displaystyle{ \,\!x }[/math] and [math]\displaystyle{ \,\!y }[/math] will be independent. We will return to this important fact again in the section "Independence test using HSIC" below.

HSIC advantages

HSIC has several advantages: First, Compared to the other measures, it is more straightforward and extra regularization terms are not needed. Second, as the sample size increases, any existing dependence with high probability will be detected. This is not true in the case of any other kernel dependence test and happens just for the covariance constraints. Third, the finite sample bias of the estimate is negligible and has the order of [math]\displaystyle{ m^{-1} }[/math]. This is not always true for the other kernel dependence tests. Last, It has been shown in an ICA experiment that this new independence test is better than the others competing with the best of the ICA methods.

HSIC in terms of kernels

To compute HSIC we need to express it in terms of kernel functions. It can be shown that this can be achieved via the following identity:
[math]\displaystyle{ \text{HSIC}(p_{xy},\mathcal{F},\mathcal{G})=\mathbf{E}_{x,x',y,y'}[k(x,x')l(y,y')]+\mathbf{E}_{x,x'}[k(x,x')]\mathbf{E}_{y,y'}[l(y,y')]-2\mathbf{E}_{x,y}[\mathbf{E}_{x'}[k(x,x')]\mathbf{E}_{y'}[l(y,y')]] }[/math]

Because the data is random, the kernels are random so we use expected values.
Following is the proof of this theorem. However, the reader may escape this proof and keep track of the rest of the article without losing the main idea of HSIC.

Proof

First we need to introduce some concepts:

Mean

Mean elements of [math]\displaystyle{ \mathcal{F} }[/math] and [math]\displaystyle{ \mathcal{G} }[/math] are defined as those elements of these spaces for which

[math]\displaystyle{ \langle\mu_x,f \rangle_{\mathcal{F}}=\mathbf{E}_x[\langle\phi(x),f \rangle_{\mathcal{F}}]=\mathbf{E}_x[f(x)] }[/math]
[math]\displaystyle{ \langle\mu_y,g \rangle_{\mathcal{G}}=\mathbf{E}_y[\langle\psi(y),g \rangle_{\mathcal{G}}]=\mathbf{E}_y[g(x)] }[/math]
Note that based on Reisz representation theorem this mean always uniquely exists. One can simply show that [math]\displaystyle{ \|\mu_x\|^2_{\mathcal{F}} }[/math] may be calculated by applying expectation twice as follows:
[math]\displaystyle{ \|\mu_x\|^2_{\mathcal{F}}=\mathbf{E}_{x,x'}[\langle \phi(x),\phi(x')\rangle_{\mathcal{F}}]=\mathbf{E}_{x,x'}[k(x,x')] }[/math]

Tensor Product Operator

We may employ any [math]\displaystyle{ f\in \mathcal{F} }[/math] and [math]\displaystyle{ g\in \mathcal{G} }[/math] to define a tensor product operator [math]\displaystyle{ f\otimes g:\mathcal{G}\rightarrow\mathcal{F} }[/math] as follows:
[math]\displaystyle{ (f\otimes g)h:=f\langle g,h\rangle_{\mathcal{G}} \quad }[/math] for all [math]\displaystyle{ h\in\mathcal{G} }[/math]

Hilbert-Schmidt operators

A Hilbert-Schmidt (HS) operator is a linear operator [math]\displaystyle{ C:\mathcal{G}\rightarrow\mathcal{F} }[/math] if its HS norm exists. We may define an inner product between two HS operator as follows:
[math]\displaystyle{ \langle C,D\rangle_{HS}:=\sum_{i,j}{\langle Cv_i,u_i\rangle_{\mathcal{F}}\langle Dv_i,u_i\rangle_{\mathcal{F}}} }[/math]
We may simply proof the following lemma for inner product of two tensor product operators:
[math]\displaystyle{ \langle f_1\otimes g_1,f_2\otimes g_2\rangle_HS=\langle f_1,f_2\rangle_F\langle g_1,g_2\rangle_G }[/math]
Using this lemma, we can simply show the norm of [math]\displaystyle{ f\otimes g }[/math] equals
[math]\displaystyle{ \|f\otimes g\|^2_{HS}=\|f\|^2_{\mathcal{F}}\; \|g\|^2_{\mathcal{G}} }[/math]

Cross-covariance Operator

The cross-covariance operator, introduced above, may be equivalently expressed as follows:
[math]\displaystyle{ C_{xy}:=\underbrace{\mathbf{E}_{x,y}[\phi(x)\otimes\psi(y)]}_{:=\tilde{C}_{xy}}-\underbrace{\mu_x\otimes\mu_y}_{:=M_{xy}} }[/math]
This expression comes from the linearity of expectation. Auxiliary variables [math]\displaystyle{ \tilde{C}_{xy} }[/math] and [math]\displaystyle{ \,\!M_{xy} }[/math] will be used as the basis of our measure of dependence.

Body of the proof

Expanding [math]\displaystyle{ \,\!C_{xy} }[/math] using the above formula, we have:
[math]\displaystyle{ \|C_{xy}\|^2_{HS}=\langle \bar{C}_{xy}-M_{xy},\bar{C}_{xy}-M_{xy}\rangle_{HS}=\mathbf{E}_{x,y,x',y'}[\langle \phi(x)\otimes\psi(y),\phi(x')\otimes\psi(y')\rangle_{HS}] }[/math]
[math]\displaystyle{ -2\mathbf{E}_{x,y}[\langle \mu_x\otimes\mu_y,\phi(x')\otimes\psi(y')\rangle_{HS}]+\langle \mu_x\otimes\mu_y,\mu_x\otimes\mu_y\rangle_{HS} }[/math]
applying the above lemma and using definitions of [math]\displaystyle{ \,\!\mu_x }[/math] and [math]\displaystyle{ \,\!\mu_y }[/math] and the fact that [math]\displaystyle{ \langle \phi(x)\phi(x')\rangle_F=k(x,x') }[/math], we arrive at the kernel representation of HSIC.

Empirical Criterion

Definition (Empirical HSIC)

Let [math]\displaystyle{ Z:=\{(x_1,y_1),\cdots,(x_m,y_m)\}\subseteq \mathcal{X}\times\mathcal{Y} }[/math] be a series of [math]\displaystyle{ \,\!m }[/math] independent observations drawn from [math]\displaystyle{ p_{x\,\!y} }[/math]. An estimator of HSIC, is given by
[math]\displaystyle{ \text{HSIC}(Z,\mathcal{F},\mathcal{G}):=(m-1)^{-2}\textbf{tr}(KHLH) }[/math]
where [math]\displaystyle{ H, K, L\in\mathbb{R}^{m\times m}, K_{ij}:=k(x_i,x_j), L_{ij}:=l(y_i,y_j) \,\, \text{and} \, \, H_{ij}:=\delta_{ij}-m^{-1} }[/math].

Bias of Estimator

It may bee shown that the bias of the above empirical estimation is of the order [math]\displaystyle{ \,\!O(m^{-1}) }[/math]:
Theorem: Let [math]\displaystyle{ \mathbf{E}_Z }[/math] denote the expectation taken over [math]\displaystyle{ \,\!m }[/math] independent copies [math]\displaystyle{ \,\!(x_i,y_i) }[/math] drawn from [math]\displaystyle{ p_{\,\!xy} }[/math]. Then
[math]\displaystyle{ \text{HSIC}(p_{xy},\mathcal{F},\mathcal{G})=\mathbf{E}_Z[\text{HSIC}(Z,\mathcal{F},\mathcal{G})]+O(m^{-1}) }[/math].

Bound on Empirical HSIC

Theorem: Assume that [math]\displaystyle{ \,\!k }[/math] and [math]\displaystyle{ \,\!l }[/math] are bounded almost everywhere by 1, and are non-negative. Then for [math]\displaystyle{ \,\!m\gt 1 }[/math] and all [math]\displaystyle{ \,\!\delta\gt 0 }[/math], with probablity at least [math]\displaystyle{ \,\!1-\delta }[/math], for all [math]\displaystyle{ \,\!p_{xy} }[/math] :
[math]\displaystyle{ |\text{HSIC}(p_{xy},\mathcal{F},\mathcal{G})-\text{HSIC}(Z,\mathcal{F},\mathcal{G})|\leq\sqrt{\frac{log(6/\delta)}{\alpha^2m}}+\frac{C}{m} }[/math]
where [math]\displaystyle{ \,\!\alpha\gt 0.24 }[/math] and [math]\displaystyle{ \,\!C }[/math] are constants.

Independence Test using HSIC

Theorem: Denote by [math]\displaystyle{ \mathcal{F} }[/math], [math]\displaystyle{ \mathcal{G} }[/math] RKHSs with universal kernels [math]\displaystyle{ \,\!k }[/math], [math]\displaystyle{ \,\!l }[/math] on the compact domains [math]\displaystyle{ \mathcal{X} }[/math] and [math]\displaystyle{ \mathcal{Y} }[/math] respectively. We assume without loss of generality that [math]\displaystyle{ \|f\|_{\infty}\leq 1 }[/math] and [math]\displaystyle{ \|g\|_{\infty}\leq 1 }[/math] for all [math]\displaystyle{ f \in \mathcal{F} }[/math] and [math]\displaystyle{ g \in \mathcal{G} }[/math]. Then [math]\displaystyle{ \|C_{xy}\|_{HS} =0 }[/math] if and only if [math]\displaystyle{ \,\!x }[/math] and [math]\displaystyle{ \,\!y }[/math] are independent.
Based on this result, to maximize the dependence between two kernels we need to increase the value of the empirical estimate, i.e., [math]\displaystyle{ \,\!\textbf{tr}(KHLH) }[/math].
It can be shown that if even one of the kernels [math]\displaystyle{ \,\!K }[/math] or [math]\displaystyle{ \,\!L }[/math] is already centered, we may drop the centering matrices [math]\displaystyle{ \,\!H }[/math] and simply use the objective function [math]\displaystyle{ \,\!\textbf{tr}(KL) }[/math].

Efficient computation in HSIC

Computational cost is a factor in choosing HSIC. Using a low rank decomposition of the Gram matrices via an incomplete Choleskey decomposition results a cost saving according to the following lemma:
Lemma 2 (Efficinet approximation to HSIC): Let [math]\displaystyle{ K\approx AA^T \, and \, L \approx BB^T, \, where \, A\in R^{m*d_f} \, and \, B\in R^{m*d_g} }[/math]. then we may approximate [math]\displaystyle{ tr(HKHL) \, in \, o(m({d_f}^2+{d_g}^2)) }[/math] time.

Related work

One of the methods to find dependence between functions in RKHS, as proposed by Gretton et al ,the constarined covariance (COCO) was to use the largest singular value of the cross-covariance operator, which behaves identically to the correlation operator at independence and no regularization is required.
In this paper, the concept of COCO is extended by using the entire spectrum of the cross-covariance operator to determine when all its singular values are zero, rather than looking only at the largest singular value.

More detiled explanation

The above theorem shows that correlation can be used to study independence. In other words, for two random variables [math]\displaystyle{ X \, }[/math] and [math]\displaystyle{ Y \, }[/math], the functional correlation operator (the word functional here refers to the fact that the correlation is computed between a function of X and a function of Y, instead of between X and Y) sheds light on whether [math]\displaystyle{ X \, }[/math] and [math]\displaystyle{ Y \, }[/math] are independent (or how independent they are). Similarly, the functional cross-covariance operator also gives us information about the degree of dependence between [math]\displaystyle{ X \, }[/math] and [math]\displaystyle{ Y \, }[/math].

Below are some examples of how the functional correlation operator and the functional cross-covariance operator used as a statistics to test independence. For easier reading we will omit the word functional in the following examples; readers should keep in mind that the operators indeed act on functions of the random variables.

Kernel Canonical Correlation(KCC)

In this approach, a regularized correlation operator is derived from the covariance operator of [math]\displaystyle{ X \, }[/math], the covariance operator of [math]\displaystyle{ Y \, }[/math] and the cross-covariance operator between [math]\displaystyle{ X \, }[/math] and [math]\displaystyle{ Y \, }[/math].

Constrained Covariance (COCO)

In this approach, the largest singular value of the cross-covariance operator is used to test independence. Compared to KCC, this method has the advantage that no regularization is needed.

This paper (HSIC)

The paper is indeed an extension of COCO, in the sense that the paper looks at the entire spectrum of the cross-covariance operator to determine whether all its singular values are zero, instead of looking at only the largest singular value, which is what COCO does. It is thus intuitively obvious that a more robust indication of independence than COCO can be obtained. More specifically, the measure of independence is the sum of the squared singular values of the cross-covariance operator(its squared Hilbert-Schmidt norm), which the authors termed HSIC(Hilbert Schmidt Independence Criterion).

Kernel Canonical Correlation Analysis (CCA)

In Classical CCA linear mapping [math]\displaystyle{ a^{T}X }[/math] and [math]\displaystyle{ b^{T}Y }[/math] are obtained, respectively for random variables [math]\displaystyle{ X }[/math] and [math]\displaystyle{ Y }[/math], such that the linear projections have maximum correlation.
Kernel CCA extends this method to estimate functions [math]\displaystyle{ f\in \mathcal{F} }[/math] and [math]\displaystyle{ g\in \mathcal{G} }[/math] such that the correlation of [math]\displaystyle{ f(X) }[/math] and [math]\displaystyle{ g(Y) }[/math] is maximized.
to find [math]\displaystyle{ f }[/math] and [math]\displaystyle{ g }[/math] the following optimization problem must be solved:

[math]\displaystyle{ max_{(f \in \mathcal{F} , g \in \mathcal{G})(f \neq 0 , g \neq 0)} \frac{Cov(f(x),g(Y))}{(Var(f(X))_{1/2} (Var(g(y))_{1/2}} }[/math]

For finite independent sample [math]\displaystyle{ {(X_{1},Y_{1}),(X_{2},Y_{2}),...,(X_{n},Y_{n})} }[/math] estimators for functions [math]\displaystyle{ f }[/math] and [math]\displaystyle{ g }[/math] can be obtained by optimizing the empirical version of the optimization problem:

[math]\displaystyle{ max_{(f \in \mathcal{F} , g \in \mathcal{G})(f \neq 0 , g \neq 0)} \frac{\widehat{Cov}(f(x),g(Y))}{(\widehat{Var}(f(X)+\epsilon_{n}||f||_{\mathcal{F}}^{2})_{1/2} (\widehat{Var}(g(y)+\epsilon_{n}||g||_{\mathcal{G}}^{2})_{1/2}} }[/math]

for which:

[math]\displaystyle{ \widehat{Cov}(f(x),g(Y))= \frac {1}{n} \sum_{i=1}^{n} \left( f(X_{i})- \sum_{j=1}^{n}f(X_{j}) \right) \left( g(Y_{i})- \sum_{j=1}^{n}g(Y_{j}) \right) }[/math]
[math]\displaystyle{ \widehat{Var}(f(X))= \frac {1}{n} \sum_{i=1}^{n} \left( f(X_{i})- \sum_{j=1}^{n}f(X_{j}) \right)^{2} }[/math] [math]\displaystyle{ \widehat{Var}(g(Y))= \frac {1}{n} \sum_{i=1}^{n} \left( g(Y_{i})- \sum_{j=1}^{n}g(Y_{j}) \right)^{2} }[/math]

Representation of Kernel CCA by Cross-Covariance Operator

Optimization problem for kernel CCA can be formulated using Cross-Covariance Operator for [math]\displaystyle{ (X,Y) }[/math]

[math]\displaystyle{ Sup_{(f \in \mathcal{F} , g \in \mathcal{G})} \langle g,\Sigma_{XY} f \rangle_{\mathcal{G}} }[/math] subject to [math]\displaystyle{ \langle f,\Sigma_{XX} f \rangle_{\mathcal{F}}=1 }[/math] and [math]\displaystyle{ \langle g,\Sigma_{YY} g \rangle_{\mathcal{G}}=1 }[/math]

Solution of this problem is the eigenfunction of the largest eigenvalue of the following eigenproblem

[math]\displaystyle{ \Sigma_{YX} f = \rho_{1} \Sigma_{YY} g }[/math]

[math]\displaystyle{ \Sigma_{YX} g = \rho_{1} \Sigma_{XX} f }[/math]