Brstjjy

Part of a series on Statistics
Regression analysis
Models
Linear regression Simple regression Polynomial regression General linear model
Generalized linear model Discrete choice Logistic regression Multinomial logit Mixed logit Probit Multinomial probit Ordered logit Ordered probit Poisson
Multilevel model Fixed effects Random effects Mixed model
Nonlinear regression Nonparametric Semiparametric Robust Quantile Isotonic Principal components Least angle Local Segmented
Errors-in-variables
Estimation
Least squares Linear Non-linear
Ordinary Weighted Generalized
Partial Total Non-negative Ridge regression Regularized
Least absolute deviations Iteratively reweighted Bayesian Bayesian multivariate
Background
Regression model validation Mean and predicted response Errors and residuals Goodness of fit Studentized residual Gauss–Markov theorem
Statistics portal
v t e

Tikhonov regularization, named for Andrey Tikhonov, is the most commonly used method of regularization of ill-posed problems. In statistics, the method is known as ridge regression, in machine learning it is known as weight decay, and with multiple independent discoveries, it is also variously known as the Tikhonov–Miller method, the Phillips–Twomey method, the constrained linear inversion method, and the method of linear regularization. It is related to the Levenberg–Marquardt algorithm for non-linear least-squares problems.

Suppose that for a known matrix A{displaystyle A} and vector b{displaystyle mathbf {b} }, we wish to find a vector x{displaystyle mathbf {x} } such that

Ax=b.{displaystyle Amathbf {x} =mathbf {b} .}

The standard approach is ordinary least squares linear regression. However, if no x{displaystyle mathbf {x} } satisfies the equation or more than one x{displaystyle 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 (over-fitted), or more often an underdetermined (under-fitted) system of equations. Most real-world phenomena have the effect of low-pass filters in the forward direction where A{displaystyle A} maps x{displaystyle mathbf {x} } to b{displaystyle 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 x{displaystyle mathbf {x} } that is in the null-space of A{displaystyle A}, rather than allowing for a model to be used as a prior for x{displaystyle mathbf {x} }.
Ordinary least squares seeks to minimize the sum of squared residuals, which can be compactly written as

‖Ax−b‖22,{displaystyle |Amathbf {x} -mathbf {b} |_{2}^{2},}

where ‖⋅‖2{displaystyle |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:

‖Ax−b‖22+‖Γx‖22{displaystyle |Amathbf {x} -mathbf {b} |_{2}^{2}+|Gamma mathbf {x} |_{2}^{2}}

for some suitably chosen Tikhonov matrix Γ{displaystyle Gamma }. In many cases, this matrix is chosen as a multiple of the identity matrix (Γ=αI{displaystyle Gamma =alpha I}), giving preference to solutions with smaller norms; this is known as L₂ regularization.^[1] 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 x^{displaystyle {hat {x}}}, is given by

x^=(A⊤A+Γ⊤Γ)−1A⊤b.{displaystyle {hat {x}}=(A^{top }A+Gamma ^{top }Gamma )^{-1}A^{top }mathbf {b} .}

The effect of regularization may be varied by the scale of matrix Γ{displaystyle Gamma }. For Γ=0{displaystyle Gamma =0} this reduces to the unregularized least-squares solution, provided that (A^TA)⁻¹ exists.

L₂ regularization is used in many contexts aside from linear regression, such as classification with logistic regression or support vector machines,^[2] and matrix factorization.^[3]

History

Tikhonov regularization has been invented independently in many different contexts.
It became widely known from its application to integral equations from the work of
Andrey Tikhonov^{[4][5][6][7][8]} and David L. Phillips.^[9] Some authors use the term Tikhonov–Phillips regularization.
The finite-dimensional case was expounded by Arthur E. Hoerl, who took a statistical approach,^[10] and by Manus Foster, who interpreted this method as a Wiener–Kolmogorov (Kriging) filter.^[11] Following Hoerl, it is known in the statistical literature as ridge regression.^[12]

Generalized Tikhonov regularization

For general multivariate normal distributions for x{displaystyle x} and the data error, one can apply a transformation of the variables to reduce to the case above. Equivalently, one can seek an x{displaystyle x} to minimize

‖Ax−b‖P2+‖x−x0‖Q2,{displaystyle |Ax-b|_{P}^{2}+|x-x_{0}|_{Q}^{2},}

where we have used ‖x‖Q2{displaystyle |x|_{Q}^{2}} to stand for the weighted norm x⊤Qx{displaystyle x^{top }Qx} (compare with the Mahalanobis distance). In the Bayesian interpretation P{displaystyle P} is the inverse covariance matrix of b{displaystyle b}, x0{displaystyle x_{0}} is the expected value of x{displaystyle x}, and Q{displaystyle Q} is the inverse covariance matrix of x{displaystyle x}. The Tikhonov matrix is then given as a factorization of the matrix Q=Γ⊤Γ{displaystyle Q=Gamma ^{top }Gamma } (e.g. the Cholesky factorization) and is considered a whitening filter.

This generalized problem has an optimal solution x∗{displaystyle x^{*}} which can be solved explicitly using the formula

x∗=(A⊤PA+Q)−1(A⊤Pb+Qx0),{displaystyle x^{*}=(A^{top }PA+Q)^{-1}(A^{top }Pb+Qx_{0}),}

or equivalently

x∗=x0+(A⊤PA+Q)−1(A⊤P(b−Ax0)).{displaystyle x^{*}=x_{0}+(A^{top }PA+Q)^{-1}(A^{top }P(b-Ax_{0})).}

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{displaystyle A} as a compact operator on Hilbert spaces, and x{displaystyle x} and b{displaystyle b} as elements in the domain and range of A{displaystyle A}. The operator A∗A+Γ⊤Γ{displaystyle A^{*}A+Gamma ^{top }Gamma } is then a self-adjoint bounded invertible operator.

Relation to singular-value decomposition and Wiener filter

With Γ=αI{displaystyle 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=UΣV⊤{displaystyle A=USigma V^{top }}

with singular values σi{displaystyle sigma _{i}}, the Tikhonov regularized solution can be expressed as

x^=VDU⊤b,{displaystyle {hat {x}}=VDU^{top }b,}

where D{displaystyle D} has diagonal values

Dii=σiσi2+α2{displaystyle 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. ^[13]

Finally, it is related to the Wiener filter:

x^=∑i=1qfiui⊤bσivi,{displaystyle {hat {x}}=sum _{i=1}^{q}f_{i}{frac {u_{i}^{top }b}{sigma _{i}}}v_{i},}

where the Wiener weights are fi=σi2σi2+α2{displaystyle f_{i}={frac {sigma _{i}^{2}}{sigma _{i}^{2}+alpha ^{2}}}} and q{displaystyle q} is the rank of A{displaystyle A}.

Determination of the Tikhonov factor

The optimal regularization parameter α{displaystyle 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^[14][15]

G=RSSτ2=‖Xβ^−y‖2[Tr⁡(I−X(XTX+α2I)−1XT)]2,{displaystyle G={frac {operatorname {RSS} }{tau ^{2}}}={frac {|X{hat {beta }}-y|^{2}}{[operatorname {Tr} (I-X(X^{T}X+alpha ^{2}I)^{-1}X^{T})]^{2}}},}

where RSS{displaystyle operatorname {RSS} } is the residual sum of squares, and τ{displaystyle tau } is the effective number of degrees of freedom.

Using the previous SVD decomposition, we can simplify the above expression:

RSS=‖y−∑i=1q(ui′b)ui‖2+‖∑i=1qα2σi2+α2(ui′b)ui‖2,{displaystyle 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},}

RSS=RSS0+‖∑i=1qα2σi2+α2(ui′b)ui‖2,{displaystyle 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

τ=m−∑i=1qσi2σi2+α2=m−q+∑i=1qα2σi2+α2.{displaystyle 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 CM{displaystyle C_{M}} representing the a priori uncertainties on the model parameters, and a covariance matrix CD{displaystyle C_{D}} representing the uncertainties on the observed parameters.^[16] In the special case when these two matrices are diagonal and isotropic, CM=σM2I{displaystyle C_{M}=sigma _{M}^{2}I} and CD=σD2I{displaystyle C_{D}=sigma _{D}^{2}I}, and, in this case, the equations of inverse theory reduce to the equations above, with α=σD/σM{displaystyle 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 Γ{displaystyle 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{displaystyle 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 σx{displaystyle sigma _{x}}. The data are also subject to errors, and the errors in b{displaystyle b} are also assumed to be independent with zero mean and standard deviation σb{displaystyle sigma _{b}}. Under these assumptions the Tikhonov-regularized solution is the most probable solution given the data and the a priori distribution of x{displaystyle x}, according to Bayes' theorem.^[17]

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 estimator. ^[18]

See also

• 
  LASSO estimator is another regularization method in statistics.


• Matrix regularization

References

• 
  Press, W. H.; Teukolsky, S. A.; Vetterling, W. T.; Flannery, B. P. (2007). "Section 19.5. Linear Regularization Methods". Numerical Recipes: The Art of Scientific Computing (3rd ed.). New York: Cambridge University Press. ISBN 978-0-521-88068-8.