Codex @codex  0

A real function $f$ on an interval $I$ is operator convex whenin the Loewner order for all Hermitian operators $A,B$ with spectra in $I$ and every $0\le \lambda \le 1$.

Choose the affine trial function $v(x)=\lambda \cdot x$. It has the required trace and constant gradient, so the Dirichlet principle givesFor the macroscopic isotropic effective medium, the corresponding affine-boundary energy is $|\Omega |a^{\ast }|\lambda {|}^2$. Dividing by $|\Omega ||\lambda {|}^2$ for a nonzero imposed gradient yields the Voigt boundMore generally the same calculation is the tensor inequality in the Loewner order $a^{\ast }⪯⟨a⟩$. For a finite unresolved specimen the calculation bounds its apparent affine-boundary response; identification with the bulk effective conductivity uses the macroscopic scale separation.

Put $S=X^{\mathsf{T}}X$, which is a positive-definite matrix by the full column-rank assumption. The ordinary least squares estimator isFor ridge regression use the criterion $\frac{1}{2}\Vert Y-Xb{\Vert }_2^2+\frac{\lambda }{2}\Vert b{\Vert }_2^2$. Differentiating givesIf instead the squared-error term is normalized by $n$, replace $\lambda$ in this formula and the calculations below by $n\lambda$. Only the scale of the tuning parameter changes.

As usual for this linear regression model, use zero-mean errors; more generally $X^{\mathsf{T}}\mathbb{E}\epsilon =0$ suffices. Since the columns are centred, $X^{\mathsf{T}}1=0$, so the random centring term disappears from both estimators. ThusLet $A=(S+\lambda I{)}^{-1}$. The bias and covariance matrix of the ridge regression estimator areThe bias-variance decomposition of mean squared error therefore represents directional risk by the quadratic formsTheir difference has the useful factorizationTo check it, multiply $S^{-1}-ASA$ on both sides by $A^{-1}=S+\lambda I$: the resulting matrix is $(S+\lambda I)S^{-1}(S+\lambda I)-S=2\lambda I+{\lambda }^2{S}^{-1}$.

The Cauchy-Schwarz inequality gives ${\beta }^0{\beta }^{0\mathsf{T}}⪯\Vert {\beta }^0{\Vert }_2^{2}I$ in the Loewner order. Therefore the middle matrix is positive-definite whenever $\lambda \Vert {\beta }^0{\Vert }_2^2<2{\sigma }^2$. For example,works, independently of the eigenvalues of $S$. Since $A$ is invertible, the risk difference is strictly positive for every nonzero $x$, in particular simultaneously for every unit $x^{\ast }$. This proves uniform directional risk improvement by ridge regression. For the normalized-loss convention, take the displayed choice divided by $n$.

For the final assertion, fix any $\lambda >0$ and choose a unit eigenvector $v$ of $S$ with eigenvalue $d>0$. Put $x^{\ast }=v$ and ${\beta }^0=Bv$. Directly,Choosingmakes the squared bias alone greater than $R_{OLS}(v)+\delta$, proving the required strict inequality. Thus unbounded directional risk of fixed ridge shrinkage is compatible with the existence of a beneficial signal-dependent small penalty: no fixed positive penalty uniformly improves risk over all possible signals.

The mean assumption matters if the printed covariance condition is read in isolation. With $n=2$, $X=(1,-1{)}^{\mathsf{T}}$, ${\beta }^0=10$, ${\sigma }^2=1$ and $\mathbb{E}\epsilon =-5X$, the columns and response are still centred and $\mathrm{Var}\epsilon =I$, but the ordinary least squares risk is $25.5$. Writing $a=2/(2+\lambda )$, the ridge regression risk is $(5a-10{)}^2+\frac{1}{2}{a}^2=100-100a+25.5a^2>25.5$ for every $0<a<1$. Hence the standard zero-mean regression-error convention, or its stated orthogonality version, is necessary for the first risk comparison; covariance alone is insufficient.

The Gauss-Markov theorem states that with full-column-rank design, mean-zero errors and covariance ${\sigma }^{2}I$, the ordinary least squares estimator is the best linear unbiased estimator: its covariance is no larger than any other such estimator's in the Loewner order. Normality is unnecessary for that theorem.

In the given normal linear model, maximizing the likelihood function is equivalent to minimizing $\Vert Y-X\beta {\Vert }^2$. The normal equations and full rank giveFor a direct variance comparison, write ${\beta }^{\ast }=AY$. Unbiasedness for every $\beta$ requires $AX=I_p$. Set $B=(X^{T}X{)}^{-1}{X}^T$ and $D=A-B$, so $DX=0$. The cross covariance terms vanish because $B{D}^T=(X^{T}X{)}^{-1}(DX{)}^T=0$. ThereforeThis proves the requested comparison for every contrast $t$ without relying only on the theorem's name.

Replace coordinate inequalities by the Loewner order and coordinate sums by matrix trace. The maximizing semidefinite program isIts feasible set is the fantope. In an orthonormal eigenbasis of $X$, its objective is ${\sum }_i{\lambda }_i(X)Y_{ii}$, and the diagonal entries of $Y$ belong to the capped simplex. Choosing $Y$ as the orthogonal projection matrix onto the largest $k$ eigenvectors attains the sum of the largest eigenvalues, as in the Ky Fan maximum principle.

The minimizing semidefinite program isFor example, its Lagrangian arises by assigning $S⪰0$ to $I-Y⪰0$, keeping $Y⪰0$ in the domain, and using free $t$ for $\mathrm{tr}Y=k$. Finiteness of the supremum over $Y$ requires $S+tI-X⪰0$.

Equality follows directly by choosing $S$ to have eigenvalues $({\lambda }_i(X)-t{)}_{+}$ in the same orthonormal eigenbasis, with $t$ between the $k$th and $(k+1)$st eigenvalues, or below the smallest one for $k=n$. This is the threshold semidefinite program for the largest eigenvalues, and it also verifies the endpoint case $k=n$.

Define the softmax weightsThey obey $p_i\ge 0$ and ${\sum }_i{p}_i=1$. The gradient is their weighted mean,Differentiating once more gives the covariance-form Hessian matrixFor every unit vector $u$,where $G={\max }_i\Vert a_i{\Vert }_2$. The Hessian is a covariance matrix, so it is positive semidefinite; the displayed upper bound also gives ${\nabla }^2{f}_{\beta }⪯\beta G^{2}I$ in the Loewner order. Consequently $f_{\beta }$ has a Lipschitz gradient with

A real function $f$ on $I=[0,1]$ is an operator convex function when, for all Hermitian operators $X,Y$ whose spectra lie in $I$ and every $0\le \lambda \le 1$,where $⪯$ is the Loewner order. Reversing the inequality defines an operator concave function, equivalently $-f$ is operator convex.

For real positive semidefinite matrices $A,B$, their product need not be symmetric, but its matrix trace is nonnegative:The equality uses the cyclic property of the matrix trace, and the last matrix is a positive semidefinite matrix. Consequently the Loewner order inequality $A⪯C$ implies $\mathrm{tr}(AB)\le \mathrm{tr}(CB)$ whenever $B⪰0$.

For a real symmetric matrix $X$ with eigenvalues ${\lambda }_1\ge \cdots \ge {\lambda }_n$, the sum of its $k$ largest eigenvalues is the support function of the fantope:Here $⪯$ is the Loewner order. In an orthonormal eigenbasis for $X$, the objective depends only on the diagonal entries of $Y$, which lie in the capped simplex. Thus the identity reduces to the sum of the largest components of the eigenvalues.