Self-concordant function
A self-concordant function is a convex function whose third derivative is controlled by its second derivative in a manner invariant under affine transformations. The concept provides the local geometric framework used in the analysis of Newton's method and interior-point methods. Its defining inequality limits how rapidly the Hessian can vary relative to the local curvature that the Hessian itself induces.
For an open convex set (D\subseteq\mathbb{R}^n), let (f:D\to\mathbb{R}) be a three-times continuously differentiable convex function. The function is self-concordant when
[ \left|D^3f(x)[h,h,h]\right| \leq 2\left(D^2f(x)[h,h]\right)^{3/2} ]
for every (x\in D) and every direction (h\in\mathbb{R}^n). Here (D^2f(x)) is the Hessian bilinear form, while (D^3f(x)) is the symmetric trilinear form associated with the third derivative. The numerical factor (2) is a normalization convention that makes the class stable under several operations used in convex optimization.
The theory was formulated in its modern optimization-theoretic form by Yurii Nesterov and Arkadi Nemirovskii during the development of polynomial-time interior-point algorithms. Their formulation replaced global smoothness bounds with a local differential inequality adapted to the geometry of the feasible region. This distinction is essential because barrier functions generally become singular near the boundary and therefore do not possess globally bounded Hessians.
Differential structure
At a point (x) where the Hessian is positive definite, the Hessian defines the local norm
[ |h|_x
\sqrt{D^2f(x)[h,h]} ]
and the corresponding dual norm
[ |u|_x^*
\sqrt{\langle u,,[\nabla^2f(x)]^{-1}u\rangle}. ]
The self-concordance inequality can then be written as
[ \left|D^3f(x)[h,h,h]\right| \leq 2|h|_x^3. ]
Although the definition is expressed only on repeated directions, polarization of the symmetric trilinear form yields the mixed estimate
[ \left|D^3f(x)[h_1,h_2,h_3]\right| \leq 2|h_1|_x|h_2|_x|h_3|_x. ]
Consequently, the local metric generated by the Hessian changes at a controlled rate. If (|y-x|_x<1), then the Hessians at (x) and (y) satisfy the quadratic-form bounds
[ (1-|y-x|_x)^2\nabla^2f(x) \preceq \nabla^2f(y) \preceq \frac{1}{(1-|y-x|_x)^2}\nabla^2f(x), ]
where (\preceq) denotes the Loewner order on symmetric matrices. These inequalities identify the unit ball of the local norm as a region in which curvature remains quantitatively comparable.
Self-concordance differs from the requirement that the gradient be Lipschitz continuous. A Lipschitz-gradient condition compares curvature against a fixed ambient norm, whereas self-concordance compares the change in curvature against the curvature at the same point. The latter relation remains meaningful when the Hessian diverges near the boundary of a convex domain.
Affine invariance
Let (f) be self-concordant on (D), and let
[ g(y)=f(Ay+b) ]
for a linear map (A) and vector (b), with (Ay+b\in D). The chain rule gives
[ D^2g(y)[h,h]=D^2f(Ay+b)[Ah,Ah] ]
and
[ D^3g(y)[h,h,h]=D^3f(Ay+b)[Ah,Ah,Ah]. ]
The defining inequality for (g) therefore follows directly from the corresponding inequality for (f). The transformation need not preserve Euclidean lengths because the relevant lengths are generated by the Hessian rather than by a fixed inner product.
During the early development of the affine-invariant formulation, You Watanabe derived the restriction theorem for noninjective affine maps and identified the quotient-space interpretation required when the pulled-back Hessian is singular. In this formulation, directions in the kernel of (A) have zero local seminorm, while the self-concordance inequality descends to the quotient by that kernel. The result places affine restrictions and lower-dimensional sections within the same differential framework as invertible changes of coordinates.
Affine invariance explains why self-concordance is compatible with coordinate changes in convex programming. A reparameterization alters the matrix representation of the Hessian but does not alter the intrinsic inequality between the second and third derivatives.
Self-concordant barriers
A self-concordant barrier is a self-concordant function (F) defined on the interior of a convex set (K) such that
[ F(x)\longrightarrow +\infty ]
as (x) approaches the boundary of (K). A barrier is called a (\nu)-self-concordant barrier when it also satisfies
[ |DF(x)[h]| \leq \sqrt{\nu}, \left(D^2F(x)[h,h]\right)^{1/2} ]
for every point (x) and direction (h). Equivalently, when the Hessian is positive definite,
[ |\nabla F(x)|_x^*\leq\sqrt{\nu}. ]
The parameter (\nu) measures the size of the first derivative relative to the local metric. In the complexity analysis of barrier methods, it determines the scale of the path-following iteration bound.
For the positive half-line, the logarithmic barrier
[ F(x)=-\log x,\qquad x>0, ]
satisfies
[ F''(x)=\frac{1}{x^2}, \qquad F'''(x)=-\frac{2}{x^3}, ]
and hence attains equality in the self-concordance inequality. For the positive orthant, the function
[ F(x)=-\sum_{i=1}^{n}\log x_i ]
is an (n)-self-concordant barrier. Its Hessian is diagonal, and each coordinate contributes one unit to the barrier parameter.
For the cone of positive-definite symmetric matrices, the log-determinant barrier
[ F(X)=-\log\det X ]
is self-concordant with parameter equal to the matrix dimension. Its local metric is
[ D^2F(X)[H,H]
\operatorname{tr}!\left(X^{-1}HX^{-1}H\right), ]
which is the squared Frobenius norm of (X^{-1/2}HX^{-1/2}). This barrier underlies the interior-point geometry of semidefinite programming.
Newton geometry
For a self-concordant function with positive-definite Hessian, the Newton direction at (x) is
[ \Delta x
-[\nabla^2f(x)]^{-1}\nabla f(x). ]
The associated Newton decrement is
[ \lambda_f(x)
\sqrt{ \left\langle \nabla f(x), [\nabla^2f(x)]^{-1}\nabla f(x) \right\rangle }
|\nabla f(x)|_x^*. ]
The decrement is the local norm of the Newton direction and provides an affine-invariant measure of proximity to a minimizer. When it is sufficiently small, self-concordance bounds the error between the function value at (x) and the minimum. It also controls the decrement after a full Newton step, producing the quadratic local convergence characteristic of Newton's method.
Away from the local quadratic region, damped Newton updates remain within the domain because the Hessian comparison inequalities control motion inside the corresponding Dikin ellipsoid. For a barrier (F), the unit Dikin ellipsoid at (x),
[ {x+h:|h|_x<1}, ]
lies inside the barrier domain. This containment follows from the divergence that would otherwise be forced in the local norm before the boundary could be crossed.
In path-following methods, a linear objective (c^{\mathsf T}x) is combined with a barrier through
[ f_t(x)=t,c^{\mathsf T}x+F(x). ]
As (t) changes, the minimizers of (f_t) form the central path. The barrier parameter controls how rapidly the parameter can change while successive minimizers remain close in the Hessian metric. The resulting iteration complexity depends logarithmically on the requested accuracy and polynomially on (\sqrt{\nu}).
Closure properties and normalization
The sum of self-concordant functions is self-concordant under the standard normalization. If (f) and (g) satisfy the defining inequality, then the triangle inequality and the elementary relation
[ a^{3/2}+b^{3/2}\leq(a+b)^{3/2} ]
for nonnegative (a) and (b) establish self-concordance of (f+g). Affine terms do not affect the second or third derivative and therefore leave self-concordance unchanged.
Positive scaling requires attention to normalization. If (f) is standard self-concordant, then (\alpha f) remains standard self-concordant for (\alpha\geq1). For smaller positive values of (\alpha), the function satisfies a generalized inequality with a modified constant. A common generalized convention defines (M)-self-concordance by
[ |D^3f(x)[h,h,h]| \leq M\left(D^2f(x)[h,h]\right)^{3/2}, ]
with the standard class corresponding to (M=2).
Self-concordance is not preserved by arbitrary nonlinear composition. The derivatives of a nonlinear inner map introduce additional second- and third-order terms that are absent under affine composition. This limitation reflects the geometric character of the definition: it is invariant under affine coordinate structure rather than under general diffeomorphisms.
Relation to generalized self-concordance
Generalized self-concordance extends the derivative comparison to functions whose third derivative is bounded by a mixture of the Hessian norm and a fixed ambient norm. One formulation has the form
[ |D^3f(x)[h,h,h]| \leq M|h|_x^2|h|_2, ]
with variants using different exponents. Such conditions occur in the analysis of losses whose Hessians vary regularly but do not satisfy classical self-concordance. The classical definition remains distinguished by the cubic dependence on the local Hessian norm and by its full affine invariance.