---
title: Dynamic Semi-Convexity Condition
url: https://www.emergentmind.com/topics/dynamic-semi-convexity-condition
type: topic
---

# Dynamic Semi-Convexity Condition

Dynamic semi-convexity is a parameter-dependent lower-curvature condition in which the admissible amount of semi-convexity is not fixed a priori, but varies with the governing parameters of a problem. In the special Lagrangian equation, this notion is realized by a phase- and dimension-dependent threshold
\[
u+\frac{1}{2}\tan(\theta)\,|x|^2\;\text{convex},\qquad \theta=\frac{\tfrac{\pi}{2}-\Theta}{n-1},
\]
for viscosity solutions of
\[
F(D^2u):=\sum_{i=1}^n \tan^{-1}(\lambda_i)=\Theta,
\]
where $\lambda_i$ are the Hessian eigenvalues and $\Theta$ is the phase. In that setting, the condition is “dynamic” because the threshold $\kappa=\tan\theta$ tightens or relaxes as $\Theta$ and $n$ vary, and because it is exactly the threshold needed for phase-lowering rotations to preserve viscosity subsolutions and supersolutions. This mechanism yields analyticity, interior derivative estimates, sharp counterexamples, and a Liouville theorem in the subcritical almost-negative regime [2510.17202].

## 1. General meaning of dynamic semi-convexity

Semi-convexity, in its standard form, requires that
\[
x\mapsto f(x)+\frac{\mu}{2}\|x\|^2
\]
be convex on a convex domain $D$. When $f$ is twice differentiable, this is equivalent to the Hessian lower bound
\[
\nabla^2 f(x)\succeq -\mu I \quad \text{for all }x\in D.
\]
A dynamic or parametric variant replaces the constant $\mu$ by a parameter-dependent function, for example requiring that for each $t$,
\[
x\mapsto f(x,t)+\frac{\mu(t)}{2}\|x\|^2
\]
be convex in $x$. This yields a time-dependent DC decomposition
\[
f(x,t)=g(x,t)-h(x,t),\qquad g(x,t)=f(x,t)+\frac{\bar\mu}{2}\|x\|^2,\quad h(x,t)=\frac{\bar\mu}{2}\|x\|^2,
\]
whenever $\mu(t)\le \bar\mu$ uniformly [1708.06999].

In the special Lagrangian setting, the same structural idea becomes quantitatively sharper. The lower-curvature allowance is not arbitrary: it is locked to the phase $\Theta$ and the dimension $n$ through
\[
\theta=\frac{\tfrac{\pi}{2}-\Theta}{n-1},\qquad \kappa=\tan\theta.
\]
The term “dynamic” therefore refers to a threshold that must be adjusted as the ambient PDE parameters change, rather than to time evolution [2510.17202].

A related but distinct usage appears in online optimization. There, a practical dynamic semi-convexity condition is formulated through per-round smoothness and an error-bound condition, namely
\[
f_t(x)-\min f_t \ge \frac{\beta_t}{2}\,\mathrm{dist}(x,\mathcal{X}_t^*)^2,\qquad \beta_t\ge \beta_{\min}>0,
\]
together with slowly drifting solution sets. This condition supports contraction toward $\mathcal{X}_t^*$ and dynamic regret bounds depending on path-length or squared path-length [1608.03933].

| Context | Dynamic semi-convexity form | Main role |
|---|---|---|
| Special Lagrangian PDE | $u+\frac12\tan(\theta)|x|^2$ convex | Preserves viscosity structure under rotation |
| DC representation | $f(x,t)+\frac{\mu(t)}{2}\|x\|^2$ convex in $x$ | Produces dynamic DC decompositions |
| Dynamic regret | Error-bound/semi-strong convexity with $\beta_t$ | Gives contraction toward time-varying minimizers |

## 2. Special Lagrangian formulation and the phase-dependent threshold

For a potential $u$ on $\Omega\subset \mathbb{R}^n$, the special Lagrangian equation is
\[
F(D^2u):=\sum_{i=1}^n \tan^{-1}(\lambda_i)=\Theta,
\]
with phase $\Theta\in(-n\pi/2,n\pi/2)$. The paper isolates the subcritical phase range
\[
\Theta\in\bigl(-(n-2)\tfrac{\pi}{2},\;\tfrac{\pi}{2}\bigr),
\]
and defines
\[
\theta:=\frac{\tfrac{\pi}{2}-\Theta}{n-1}\in(0,\tfrac{\pi}{2}).
\]
This parameter distributes the gap from $\pi/2$ uniformly across the $n-1$ non-maximal eigen-directions [2510.17202].

The associated dynamic semi-convexity condition is
\[
u+\frac{1}{2}\tan(\theta)\,|x|^2\;\text{ is convex on }B_1,
\]
equivalently
\[
D^2u\ge -\tan(\theta)\,I
\quad\text{in the sense of distributions}.
\]
Its dependence on $\Theta$ is monotone. As $\Theta\uparrow \pi/2$, one has $\theta\downarrow 0$, so the admissible semi-convexity becomes increasingly restrictive. As $\Theta\downarrow -(n-2)\pi/2$, one has $\theta\uparrow \pi/2$, and the allowable semi-convexity becomes much larger. The threshold is therefore neither a purely local regularity assumption nor a generic lower Hessian bound; it is tied to the phase geometry of the equation itself [2510.17202].

A central algebraic identity is
\[
\frac{\pi}{2}-\tan^{-1}(\lambda_1)
=
\sum_{i=2}^n\bigl(\tan^{-1}(\lambda_i)+\theta\bigr).
\tag{1}
\]
This quantitatively links the maximal Hessian eigenvalue to the remaining eigenvalues shifted by $\theta$. In the later analysis, that identity provides the eigenvalue rigidity needed to rule out blow-up of the largest rotated eigenvalue [2510.17202].

The viscosity framework is standard: if a quadratic polynomial $P$ touches $u$ from above near $x_0$, then $F(D^2P)\ge \Theta$; if it touches from below, then $F(D^2P)\le \Theta$. Ellipticity is expressed by
\[
F(M+N)\ge F(M)\qquad\text{whenever }N\ge 0.
\]
At smooth points, the linearization agrees with the Laplace–Beltrami operator on the gradient graph $(x,Du(x))$ with metric $g=I+(D^2u)^2$, so directional second derivatives $u_{ee}$ are subsolutions of the linearized operator [2510.17202].

## 3. Rotation, touching preservation, and phase descent

The decisive mechanism is a Lewy–Yuan-type rotation of the gradient graph. For $0<\phi<\pi/2-\theta$,
\[
\bar x+i\bar y=e^{-i\phi}(x+i\,Du(x)),\qquad
\bar x=cx+s\,Du(x),\;\;\bar y=-sx+c\,Du(x),
\]
with $s=\sin\phi$ and $c=\cos\phi$. The rotated potential is defined by the Legendre-transform formula
\[
\bar u(\bar x)=\frac{c}{2s}|\bar x|^2-\frac{1}{s}\Big(su(x)+\frac{c}{2}|x|^2\Big)^*(\bar x).
\]
The rotation preserves touching, and the eigen-angles satisfy
\[
\tan^{-1}(\bar\lambda_i)=\tan^{-1}(\lambda_i)-\phi.
\tag{2}
\]
This allows the phase to be lowered while retaining viscosity control [2510.17202].

Under the phase range and dynamic semi-convexity hypothesis, both subsolutions and supersolutions are preserved:
\[
u\text{ subsolution of }F(D^2u)=\Theta
\Longrightarrow
\bar u\text{ subsolution of }F(D^2\bar u)=\Theta-n\phi,
\]
\[
u\text{ supersolution of }F(D^2u)=\Theta
\Longrightarrow
\bar u\text{ supersolution of }F(D^2\bar u)=\Theta-n\phi.
\]
The paper emphasizes that this preservation is sharp in the phase/semi-convexity range considered; outside that regime, rotations can fail to preserve subsolutions [2510.17202].

A specific choice of angle drives the problem into a negative supercritical regime:
\[
\phi=\frac{\pi}{2}-\theta-\delta,\qquad
\delta:=\frac{(\frac{\pi}{2}-\theta)}{2n}
=
\frac{(n-2)\frac{\pi}{2}+\Theta}{2n(n-1)}\in(0,\tfrac{\pi}{4n}),
\]
so that
\[
\Theta-n\phi=-(n-2)\frac{\pi}{2}-n\delta.
\tag{3}
\]
For this rotated phase, $-\bar u$ solves a concave, uniformly elliptic PDE with convex superlevel set. Evans–Krylov then yields $C^{2,\alpha}$ regularity, and Morrey’s theorem upgrades the rotated solution to analyticity. Because the rotation map is distance-increasing and locally bi-Lipschitz, the regularity can be transferred back to the original function $u$ [2510.17202].

## 4. Eigenvalue rigidity, analyticity, and quantitative interior estimates

The dynamic semi-convexity hypothesis does more than provide a lower Hessian bound. Combined with the phase splitting in equation (1), it enforces a rigidity phenomenon at any putative blow-up point of the maximal rotated eigenvalue. If
\[
\tan^{-1}(\bar\lambda_1)=\frac{\pi}{2}-\phi
\]
at some point, then the remaining rotated eigenvalues must saturate the lower bound:
\[
\tan^{-1}(\bar\lambda_i)\equiv-(\theta+\phi)\qquad\text{for }i=2,\dots,n.
\]
By convexity of level sets for the rotated equation of $-\bar u$, sums of the largest $k$ eigenvalues are viscosity subsolutions of the linearized operator; the strong maximum principle then forces constancy of the smaller eigenvalues, and the equation forces constancy of $\bar\lambda_1$. This contradicts Alexandrov’s theorem that $D^2u$ is finite almost everywhere. Consequently,
\[
D^2\bar u<\tan(\pi/2-\phi)\,I
\]
everywhere [2510.17202].

The inverse Hessian transform is
\[
D^2u(x)=
(sI+c\,D^2\bar u(\bar x))\,(cI-s\,D^2\bar u(\bar x))^{-1}.
\tag{4}
\]
Since the right-hand side is locally bounded once $D^2\bar u$ is controlled, the original potential acquires a locally bounded Hessian and hence analyticity. This is the content of the regularity theorem: if
\[
\Theta\in\bigl(-(n-2)\pi/2,\pi/2\bigr),
\qquad
\theta=\frac{\pi/2-\Theta}{n-1},
\]
and $u$ is a viscosity solution in $B_1$ with
\[
u+\frac12\tan\theta\,|x|^2 \text{ convex},
\]
then $u$ is analytic. Moreover, for $k\ge 2$,
\[
|D^ku(0)|
\le
\exp\Big(C(n,k,\Theta)\big(1+\|Du\|_{L^\infty(B_1)}\big)\Big).
\tag{5}
\]
The dependence is exponential in the Lipschitz norm and depends only on $n,k,\Theta$ [2510.17202].

The analysis also yields pointwise control of the non-maximal eigenvalues:
\[
-\tan(\theta)\le \lambda_i<1\qquad\text{for all }i\ge 2.
\tag{6}
\]
This estimate feeds into a volume bound for the rotated image:
\[
|\bar x(B_1)|\le C(1+L),\qquad L=\|Du\|_{L^\infty(B_1)}.
\tag{7}
\]
A chain-of-balls argument in the rotated domain, combined with the weak Harnack inequality, gives for nonnegative supersolutions $w$ of the linearized equation at $\bar u$,
\[
\int_{\bar x(B_{1/2})} w^p \le e^{C(1+L)}\,w^p\big(\bar x(0)\big),
\tag{8}
\]
for universal $p>0,C$. Applying this to
\[
w:=\sum_{i=2}^n\big(\bar\lambda_i+\tan(\theta+\phi)\big)\ge 0
\]
produces the exponential Hessian control [2510.17202].

## 5. Sharpness, counterexamples, and Liouville rigidity

The phase range and the dynamic threshold are both sharp. First, if
\[
\Theta\in[\pi/2,(n-2)\pi/2)
\]
and $\varepsilon>0$, there exist singular viscosity solutions of $F(D^2u)=\Theta$ such that $u+\varepsilon|x|^2$ is convex. Thus, once the phase crosses $\pi/2$, no negative uniform lower bound on $D^2u$ is sufficient to force regularity. Second, if
\[
\Theta\in\bigl(-(n-2)\pi/2,\pi/2\bigr)
\]
and $\varepsilon>0$, there exist singular viscosity solutions such that
\[
u+\frac12\bigl(\tan\theta+\varepsilon\bigr)|x|^2\;\text{ is convex},
\]
that is,
\[
D^2u\ge -(\tan\theta+\varepsilon)I.
\]
Hence the threshold $\tan\theta$ cannot be relaxed. In both regimes, the examples are Lipschitz but not $C^1$ and have non-minimal gradient graphs [2510.17202].

The construction begins from a rank-$(n-1)$ model
\[
\Phi(x)=\frac{\lambda x_1^2}{2(1+x_3)}+\frac{\lambda x_2^2}{2(1-x_3)}+\sum_{i\ge4}\Big(\frac{a_i}{2}x_i^2+\frac{x_i^4}{12}\Big),
\]
for which $F(D^2\Phi)$ has a nondegenerate local minimum at the origin. One then solves the constant-phase equation near a small convex set $K_\varepsilon$, glues analytically to a function $w$ with one-sided sign on $\det D^2w$, and applies the Legendre transform to define $u=-w^*$. Away from the image hypersurface, $u$ is analytic and solves
\[
F(D^2u)=c^*_\epsilon+\frac{\pi}{2}\Big(2-n+2\#\{i\ge4:a_i<0\}\Big),
\]
with one Hessian eigenvalue tending to $+\infty$ and the others controlled [2510.17202].

The exponential dependence in equation (5) is also optimal. In dimension $n=2$ and phase $\Theta=\pi/2$, the explicit solution
\[
g(s)=s\sinh^{-1}(s)-\sqrt{1+s^2},\qquad
u(x,y)=e^{-M}\cos(y)\,g\!\Big(\frac{e^M x}{\cos y}\Big)
\]
satisfies
\[
\det D^2u=1,\qquad \|\nabla u\|_{L^\infty}\lesssim M,\qquad u_{xx}(0,0)=e^M.
\]
Therefore any interior Hessian estimate must be at least exponential in $\|Du\|_{L^\infty}$. For $\Theta\neq 0$ in $n=2$, a partial Legendre–Lewy–Wang–Yuan transform yields a rotated potential solving
\[
F(D^2\bar u)=\frac{\pi}{2}-\theta,
\]
and the exponential behavior persists. Adding quadratic directions with coefficient $-\tan\theta$ or $+A$ embeds this optimality into higher dimensions [2510.17202].

The same dynamic mechanism yields a Liouville theorem. If
\[
\Theta\in\bigl(-(n-2)\pi/2,\pi/2\bigr),\qquad
u+\frac12\tan\theta\,|x|^2\text{ convex},
\]
and $u$ is an entire viscosity solution of the special Lagrangian equation, then $u$ is a quadratic polynomial. After rotation to a negative supercritical phase, Evans–Krylov and Schauder imply decay of the Hölder seminorm of $D^2\bar u$ on large balls, forcing $\bar u$ to be quadratic; the inverse rotation then gives the same conclusion for $u$ [2510.17202].

## 6. Broader formulations beyond the special Lagrangian equation

Outside special Lagrangian geometry, dynamic semi-convexity appears as a reusable structural pattern. In the DC setting, one formulation requires that for each parameter value $t$,
\[
x\mapsto f(x,t)+\frac{\mu(t)}{2}\|x\|^2
\]
be convex. Uniform control $\mu(t)\le \bar\mu$ produces a time-uniform decomposition into convex components. In the smooth case, if
\[
\sup_{x\in D}\|\nabla^2 f(x)\|=L<\infty,
\]
then
\[
f_1(x):=L\|x\|^2-f(x)
\]
is convex, so
\[
f(x)=L\|x\|^2-f_1(x)
\]
is a DC decomposition. The same paper characterizes DC functions in terms of bounded variation of directional derivatives along circle or plane arcs, which serves as a nonsmooth analogue of curvature control [1708.06999].

In online learning, the analogous role is played by semi-strong convexity, defined by
\[
f(x)-\min_{z\in\mathcal{X}}f(z)\ge \frac{\beta}{2}\,\big\|x-\Pi_{\mathcal{X}^*}(x)\big\|^2.
\]
This is weaker than strong convexity because the minimizer set may be non-singleton and the curvature is only enforced relative to distance from the solution set. With $L$-smooth losses and one gradient query per round, dynamic regret scales with the path-length of the comparator sequence. With multiple gradient queries per round,
\[
z_t^{j+1}=\Pi_{\mathcal{X}}\bigl(z_t^j-\eta\,\nabla f_t(z_t^j)\bigr),\qquad \eta\le 1/L,
\]
the paper derives bounds of order
\[
O\bigl(\min\{P_T,S_T\}\bigr)
\]
under strong convexity, and under semi-strong convexity when the aggregate gradients at minimizers are suitably small. In the self-concordant case, a damped Newton scheme achieves an analogous dependence in local Hessian norms [1608.03933].

These broader formulations indicate that dynamic semi-convexity is not a single invariant definition across all fields. Rather, it denotes a family of parameter-dependent lower-curvature conditions whose exact form is dictated by the surrounding analytic mechanism. In special Lagrangian theory, the threshold is phase-driven and rotation-compatible; in DC analysis, it is tied to quadratic regularization and variation bounds; in dynamic regret, it appears as an error-bound condition that contracts iterates toward moving solution sets. A plausible implication is that the common mathematical content is not merely “semi-convexity with varying constants,” but semi-convexity calibrated to the transformation or stability principle that a problem requires [2510.17202] [1708.06999] [1608.03933].

Source: https://www.emergentmind.com/topics/dynamic-semi-convexity-condition