Ridge regression

Ridge regression is a method of estimating the coefficients of multiple-regression models in scenarios where the independent variables are highly correlated. It has been used in many fields including econometrics, chemistry, and engineering. Also known as Tikhonov regularization, named for Andrey Tikhonov, it is a method of regularization of ill-posed problems. It is particularly useful to mitigate the problem of multicollinearity in linear regression, which commonly occurs in models with large numbers of parameters. In general, the method provides improved efficiency in parameter estimation problems in exchange for a tolerable amount of bias (see bias–variance tradeoff).

The theory was first introduced by Hoerl and Kennard in 1970 in their Technometrics papers "Ridge regressions: biased estimation of nonorthogonal problems" and "Ridge regressions: applications in nonorthogonal problems". This was the result of ten years of research into the field of ridge analysis.

Ridge regression was developed as a possible solution to the imprecision of least square estimators when linear regression models have some multicollinear (highly correlated) independent variables—by creating a ridge regression estimator (RR). This provides a more precise ridge parameters estimate, as its variance and mean square estimator are often smaller than the least square estimators previously derived.

Overview
In the simplest case, the problem of a near-singular moment matrix $$\mathbf{X}^\mathsf{T}\mathbf{X}$$ is alleviated by adding positive elements to the diagonals, thereby decreasing its condition number. Analogous to the ordinary least squares estimator, the simple ridge estimator is then given by $$\hat{\beta}_{R} = \left(\mathbf{X}^{\mathsf{T}} \mathbf{X} + \lambda \mathbf{I}\right)^{-1} \mathbf{X}^{\mathsf{T}} \mathbf{y}$$ where $$\mathbf{y}$$ is the regressand, $$\mathbf{X}$$ is the design matrix, $$\mathbf{I}$$ is the identity matrix, and the ridge parameter $$\lambda \geq 0$$ serves as the constant shifting the diagonals of the moment matrix. It can be shown that this estimator is the solution to the least squares problem subject to the constraint $$\beta^\mathsf{T}\beta = c$$, which can be expressed as a Lagrangian: $$\min_{\beta} \, \left(\mathbf{y} - \mathbf{X} \beta\right)^\mathsf{T} \left(\mathbf{y} - \mathbf{X} \beta\right) + \lambda \left(\beta^\mathsf{T}\beta - c\right)$$ which shows that $$\lambda$$ is nothing but the Lagrange multiplier of the constraint. Typically, $$\lambda$$ is chosen according to a heuristic criterion, so that the constraint will not be satisfied exactly. Specifically in the case of $$\lambda = 0$$, in which the constraint is non-binding, the ridge estimator reduces to ordinary least squares. A more general approach to Tikhonov regularization is discussed below.

History
Tikhonov regularization was invented independently in many different contexts. It became widely known through its application to integral equations in the works of Andrey Tikhonov    and David L. Phillips. Some authors use the term Tikhonov–Phillips regularization. The finite-dimensional case was expounded by Arthur E. Hoerl, who took a statistical approach, and by Manus Foster, who interpreted this method as a Wiener–Kolmogorov (Kriging) filter. Following Hoerl, it is known in the statistical literature as ridge regression, named after ridge analysis ("ridge" refers to the path from the constrained maximum).

Tikhonov regularization
Suppose that for a known matrix $$A$$ and vector $$\mathbf{b}$$, we wish to find a vector $$\mathbf{x}$$ such that $$A\mathbf{x} = \mathbf{b},$$ where $$\mathbf{x}$$ and $$\mathbf{b}$$ may be of different sizes and $$A$$ may be non-square.

The standard approach is ordinary least squares linear regression. However, if no $$\mathbf{x}$$ satisfies the equation or more than one $$\mathbf{x}$$ does—that is, the solution is not unique—the problem is said to be ill posed. In such cases, ordinary least squares estimation leads to an overdetermined, or more often an underdetermined system of equations. Most real-world phenomena have the effect of low-pass filters in the forward direction where $$A$$ maps $$\mathbf{x}$$ to $$\mathbf{b}$$. Therefore, in solving the inverse-problem, the inverse mapping operates as a high-pass filter that has the undesirable tendency of amplifying noise (eigenvalues / singular values are largest in the reverse mapping where they were smallest in the forward mapping). In addition, ordinary least squares implicitly nullifies every element of the reconstructed version of $$\mathbf{x}$$ that is in the null-space of $$A$$, rather than allowing for a model to be used as a prior for $$\mathbf{x}$$. Ordinary least squares seeks to minimize the sum of squared residuals, which can be compactly written as $$\left\|A\mathbf{x} - \mathbf{b}\right\|_2^2,$$ where $$\|\cdot\|_2$$ is the Euclidean norm.

In order to give preference to a particular solution with desirable properties, a regularization term can be included in this minimization: $$\left\|A\mathbf{x} - \mathbf{b}\right\|_2^2 + \left\|\Gamma \mathbf{x}\right\|_2^2$$ for some suitably chosen Tikhonov matrix $$\Gamma $$. In many cases, this matrix is chosen as a scalar multiple of the identity matrix ($$\Gamma = \alpha I$$), giving preference to solutions with smaller norms; this is known as $L_{2}$ regularization. In other cases, high-pass operators (e.g., a difference operator or a weighted Fourier operator) may be used to enforce smoothness if the underlying vector is believed to be mostly continuous. This regularization improves the conditioning of the problem, thus enabling a direct numerical solution. An explicit solution, denoted by $$\hat{x}$$, is given by $$\hat{x} = \left(A^\mathsf{T} A + \Gamma^\mathsf{T} \Gamma\right)^{-1} A^\mathsf{T} \mathbf{b}.$$ The effect of regularization may be varied by the scale of matrix $$\Gamma$$. For $$\Gamma = 0$$ this reduces to the unregularized least-squares solution, provided that (ATA)−1 exists.

$L_{2}$ regularization is used in many contexts aside from linear regression, such as classification with logistic regression or support vector machines, and matrix factorization.

Application to existing fit results
Since Tikhonov Regularization simply adds a quadratic term to the objective function in optimization problems, it is possible to do so after the unregularised optimisation has taken place. E.g., if the above problem with $$\Gamma = 0$$ yields the solution $$\hat{x}_0$$, the solution in the presence of $$\Gamma \ne 0$$ can be expressed as: $$\hat{x} = B \hat{x}_0,$$ with the "regularisation matrix" $$B = \left(A^\mathsf{T} A + \Gamma^\mathsf{T} \Gamma\right)^{-1} A^\mathsf{T} A$$.

If the parameter fit comes with a covariance matrix of the estimated parameter uncertainties $$V_0$$, then the regularisation matrix will be $$B = (V_0^{-1} + \Gamma^\mathsf{T}\Gamma)^{-1} V_0^{-1},$$ and the regularised result will have a new covariance $$V = B V_0 B^\mathsf{T}.$$

In the context of arbitrary likelihood fits, this is valid, as long as the quadratic approximation of the likelihood function is valid. This means that, as long as the perturbation from the unregularised result is small, one can regularise any result that is presented as a best fit point with a covariance matrix. No detailed knowledge of the underlying likelihood function is needed.

Generalized Tikhonov regularization
For general multivariate normal distributions for $$\mathbf x$$ and the data error, one can apply a transformation of the variables to reduce to the case above. Equivalently, one can seek an $$\mathbf x$$ to minimize $$\left\|A \mathbf x - \mathbf b\right\|_P^2 + \left\|\mathbf x - \mathbf x_0\right\|_Q^2,$$ where we have used $$\left\|\mathbf{x}\right\|_Q^2$$ to stand for the weighted norm squared $$\mathbf{x}^\mathsf{T} Q \mathbf{x}$$ (compare with the Mahalanobis distance). In the Bayesian interpretation $$P$$ is the inverse covariance matrix of $$\mathbf b$$, $$\mathbf x_0$$ is the expected value of $$\mathbf x$$, and $$Q$$ is the inverse covariance matrix of $$\mathbf x$$. The Tikhonov matrix is then given as a factorization of the matrix $$Q = \Gamma^\mathsf{T} \Gamma$$ (e.g. the Cholesky factorization) and is considered a whitening filter.

This generalized problem has an optimal solution $$\mathbf x^*$$ which can be written explicitly using the formula $$\mathbf x^* = \left(A^\mathsf{T} PA + Q\right)^{-1} \left(A^\mathsf{T} P \mathbf{b} + Q \mathbf{x}_0\right),$$ or equivalently, when Q is not a null matrix: $$\mathbf x^* = \mathbf x_0 + \left(A^\mathsf{T} P A + Q \right)^{-1} \left(A^\mathsf{T} P \left(\mathbf b - A \mathbf x_0\right)\right).$$

Lavrentyev regularization
In some situations, one can avoid using the transpose $$A^\mathsf{T}$$, as proposed by Mikhail Lavrentyev. For example, if $$A$$ is symmetric positive definite, i.e. $$A = A^\mathsf{T} > 0$$, so is its inverse $$A^{-1}$$, which can thus be used to set up the weighted norm squared $$\left\|\mathbf x\right\|_P^2 = \mathbf x^\mathsf{T} A^{-1} \mathbf x$$ in the generalized Tikhonov regularization, leading to minimizing $$\left\|A \mathbf x - \mathbf b\right\|_{A^{-1}}^2 + \left\|\mathbf x - \mathbf x_0 \right\|_Q^2$$ or, equivalently up to a constant term, $$\mathbf x^\mathsf{T} \left(A+Q\right) \mathbf x - 2 \mathbf x^\mathsf{T} \left(\mathbf b + Q \mathbf x_0\right).$$

This minimization problem has an optimal solution $$\mathbf x^*$$ which can be written explicitly using the formula $$\mathbf x^* = \left(A + Q\right)^{-1} \left(\mathbf b + Q \mathbf x_0\right),$$ which is nothing but the solution of the generalized Tikhonov problem where $$A = A^\mathsf{T} = P^{-1}.$$

The Lavrentyev regularization, if applicable, is advantageous to the original Tikhonov regularization, since the Lavrentyev matrix $$A + Q$$ can be better conditioned, i.e., have a smaller condition number, compared to the Tikhonov matrix $$A^\mathsf{T} A + \Gamma^\mathsf{T} \Gamma.$$

Regularization in Hilbert space
Typically discrete linear ill-conditioned problems result from discretization of integral equations, and one can formulate a Tikhonov regularization in the original infinite-dimensional context. In the above we can interpret $$A$$ as a compact operator on Hilbert spaces, and $$x$$ and $$b$$ as elements in the domain and range of $$A$$. The operator $$A^* A + \Gamma^\mathsf{T} \Gamma $$ is then a self-adjoint bounded invertible operator.

Relation to singular-value decomposition and Wiener filter
With $$\Gamma = \alpha I$$, this least-squares solution can be analyzed in a special way using the singular-value decomposition. Given the singular value decomposition $$A = U \Sigma V^\mathsf{T}$$ with singular values $$\sigma _i$$, the Tikhonov regularized solution can be expressed as $$\hat{x} = V D U^\mathsf{T} b,$$ where $$D$$ has diagonal values $$D_{ii} = \frac{\sigma_i}{\sigma_i^2 + \alpha^2}$$ and is zero elsewhere. This demonstrates the effect of the Tikhonov parameter on the condition number of the regularized problem. For the generalized case, a similar representation can be derived using a generalized singular-value decomposition.

Finally, it is related to the Wiener filter: $$\hat{x} = \sum _{i=1}^q f_i \frac{u_i^\mathsf{T} b}{\sigma_i} v_i,$$ where the Wiener weights are $$f_i = \frac{\sigma _i^2}{\sigma_i^2 + \alpha^2}$$ and $$q$$ is the rank of $$A$$.

Determination of the Tikhonov factor
The optimal regularization parameter $$\alpha$$ is usually unknown and often in practical problems is determined by an ad hoc method. A possible approach relies on the Bayesian interpretation described below. Other approaches include the discrepancy principle, cross-validation, L-curve method, restricted maximum likelihood and unbiased predictive risk estimator. Grace Wahba proved that the optimal parameter, in the sense of leave-one-out cross-validation minimizes $$G = \frac{\operatorname{RSS}}{\tau^2} = \frac{\left\|X \hat{\beta} - y\right\|^2}{ \left[\operatorname{Tr}\left(I - X\left(X^\mathsf{T} X + \alpha^2 I\right)^{-1} X^\mathsf{T}\right)\right]^2},$$ where $$\operatorname{RSS}$$ is the residual sum of squares, and $$\tau$$ is the effective number of degrees of freedom.

Using the previous SVD decomposition, we can simplify the above expression: $$\operatorname{RSS} = \left\| y - \sum_{i=1}^q (u_i' b) u_i \right\|^2 + \left\| \sum _{i=1}^q \frac{\alpha^2}{\sigma_i^2 + \alpha^2} (u_i' b) u_i \right\|^2,$$ $$\operatorname{RSS} = \operatorname{RSS}_0 + \left\| \sum_{i=1}^q \frac{\alpha^2}{\sigma_i^2 + \alpha^2} (u_i' b) u_i \right\|^2,$$ and $$\tau = m - \sum_{i=1}^q \frac{\sigma_i^2}{\sigma_i^2 + \alpha^2} = m - q + \sum_{i=1}^q \frac{\alpha^2}{\sigma _i^2 + \alpha^2}.$$

Relation to probabilistic formulation
The probabilistic formulation of an inverse problem introduces (when all uncertainties are Gaussian) a covariance matrix $$ C_M$$ representing the a priori uncertainties on the model parameters, and a covariance matrix $$ C_D$$ representing the uncertainties on the observed parameters. In the special case when these two matrices are diagonal and isotropic, $$ C_M = \sigma_M^2 I $$ and $$ C_D = \sigma_D^2 I $$, and, in this case, the equations of inverse theory reduce to the equations above, with $$ \alpha = {\sigma_D}/{\sigma_M} $$.

Bayesian interpretation
Although at first the choice of the solution to this regularized problem may look artificial, and indeed the matrix $$\Gamma$$ seems rather arbitrary, the process can be justified from a Bayesian point of view. Note that for an ill-posed problem one must necessarily introduce some additional assumptions in order to get a unique solution. Statistically, the prior probability distribution of $$x$$ is sometimes taken to be a multivariate normal distribution. For simplicity here, the following assumptions are made: the means are zero; their components are independent; the components have the same standard deviation $$\sigma _x$$. The data are also subject to errors, and the errors in $$b$$ are also assumed to be independent with zero mean and standard deviation $$\sigma _b$$. Under these assumptions the Tikhonov-regularized solution is the most probable solution given the data and the a priori distribution of $$x$$, according to Bayes' theorem.

If the assumption of normality is replaced by assumptions of homoscedasticity and uncorrelatedness of errors, and if one still assumes zero mean, then the Gauss–Markov theorem entails that the solution is the minimal unbiased linear estimator.