---
title: 'Tangential Cone Condition: Theory & Applications'
url: https://www.emergentmind.com/topics/tangential-cone-condition
type: topic
---

# Tangential Cone Condition: Theory & Applications

to=arxiv_search.search  福利彩票天天json
{"query":"all:\"tangential cone condition\" OR ti:\"tangential cone condition\"","max_results":10,"sort_by":"submittedDate","sort_order":"descending"}
to=arxiv_search.search  天天中彩票无法්ඩict  大发快三有json
{"query":"ti:\"tangential cone condition\" OR abs:\"tangential cone condition\"","max_results":10,"sort_by":"submittedDate","sort_order":"descending"}
to=arxiv_search.search  ปมถวายสัตย์  天天中彩票微信
{"query":"\"tangential cone condition\"","max_results":10,"sort_by":"relevance"}
The tangential cone condition denotes a family of first-order regularity conditions that compare a nonlinear object with its tangent or linearized approximation. In nonlinear inverse problems, it is imposed on a forward operator \(F\) and bounds the Taylor remainder by the size of the nonlinear increment. In variational analysis and constrained optimization, closely related conditions identify the tangent cone with a linearized cone or require compatibility of tangent cones with facial structure. In geometric topology and differential geometry, the same phrase refers to coincidence or control of tangent and paratangent cones. Across these settings, the common theme is that first-order infinitesimal data must capture the relevant local behavior with sufficient accuracy [1707.07589].

## 1. Operator-theoretic formulation in inverse problems

For a nonlinear operator equation
\[
F(x)=y,\qquad F:D(F)\subseteq X\to Y,
\]
between Banach spaces, the paper on IRGNM in Banach space imposes the tangential cone condition in the form
\[
\|F(\tilde{x}) - F(x) - F'(x)(\tilde{x}-x)\| \;\leq\; c_{tc} \,\|F(\tilde{x}) - F(x)\| \quad \text{for all } x,\tilde{x}\in\mathcal{B}_R,
\]
where \(\mathcal{B}_R=\{x\in \mathcal{D}(F):R(x)\le r\}\), \(F'(x)\) is the Gâteaux derivative, and \(c_{tc}<\frac13\) [1707.07589]. The same paper identifies this as the “Scherzer condition” and uses it as the central nonlinearity assumption in place of stronger restrictions on \(F\).

In the EIT setting, the strong tangential cone condition is written as
\[
\|F(\tilde{x}) - F(x) - F'(x)(\tilde{x} - x)\|_Y \;\le\; \eta\,\|F(\tilde{x}) - F(x)\|_Y,
\]
with \(\eta\in(0,1)\), while the weak tangential cone condition is
\[
\big(F(\tilde{x}) - F(x) - F'(x)(\tilde{x}-x),\, F(\tilde{x}) - F(x)\big)_Y \;\le\; \eta\,\|F(\tilde{x}) - F(x)\|_Y^2.
\]
The same source also records the relaxed condition
\[
\big( F'(x)(\tilde{x}-x),\, F(\tilde{x}) - F(x) \big)_Y \ge 0,
\]
called a weaker “quasi-cone” condition [2105.02635].

These formulations all quantify the same idea: the linearization \(F'(x)(\tilde{x}-x)\) must approximate the nonlinear increment \(F(\tilde{x})-F(x)\) with an error that is controlled by the increment itself. The precise constants and norms vary with the analytic framework.

## 2. Variants, strength, and convergence consequences

The tangential condition was introduced in Hanke, Neubauer, and Scherzer as a sufficient condition for convergence of the Landweber iteration for solving ill-posed problems, and later work extends its use to IRGNM, Tikhonov, and Ivanov schemes [1908.01239]. In the Banach-space IRGNM analysis, the assumptions are that \(X,Y\) are Banach spaces, the sublevel sets \(\mathcal{B}_r\) are compact with respect to some topology \(\mathcal{T}\), \(F\) is Gâteaux differentiable on \(\mathcal{B}_R\), and both \(x\mapsto F(x)\) and \(x\mapsto F'(x)\) are \(\mathcal{T}\)-to-norm continuous on \(\mathcal{B}_R\). Under these hypotheses and the tangential cone condition with \(c_{tc}<1/3\), the iterates of both IRGNM Tikhonov and IRGNM Ivanov are well-defined, the discrepancy-principle stopping index is finite, and as \(\delta\to0\) the stopped iterates converge subsequentially in \(\mathcal{T}\) to solutions of \(F(x)=y\); if the solution is unique in \(\mathcal{B}_R\), then convergence is to that solution, with \(k_*=\mathcal{O}(\log(1/\delta))\) [1707.07589].

For Landweber-type methods, the EIT paper summarizes the standard implications: strong TCC with \(\eta\le 1/2\) implies strong convergence of nonlinear Landweber under additional standard conditions, weak TCC implies nonexpansivity and weak subsequential convergence, and the quasi-cone condition ensures that iterates stay in a neighborhood of the solution and also gives weak subsequential convergence [2105.02635]. The parabolic PDE paper frames the same condition as a local convexity condition for the residual \(\theta\mapsto\|F(\theta)-y\|^2\), with
\[
\|F(\theta)-F(\tilde\theta)-F'(\theta)(\theta-\tilde\theta)\|_{\mathcal{Y}}
\leq
c_{tc}\,\|F(\theta)-F(\tilde\theta)\|_{\mathcal{Y}}
\]
for \(\theta,\tilde\theta\) in a local ball, together with a uniform bound
\[
\|F'(\theta)\|_{\mathcal{L}(\mathcal{X},\mathcal{Y})}\leq C_F
\]
on that ball [1908.01239].

The relationship between TCC and other nonlinearity assumptions is also explicit. The IRGNM paper states that local invariance of the range of \(F'(x)^*\) is sufficient for the tangential cone condition, and therefore is slightly stronger [1707.07589]. This suggests that TCC occupies an intermediate position: weaker than adjoint-range invariance, but still strong enough to drive convergence proofs without source conditions.

## 3. Verification in PDE coefficient identification and EIT

For time-dependent benchmark inverse problems, the parabolic PDE paper verifies the tangential cone condition for several model classes: identification of a potential \(c(x)\), identification of a diffusion coefficient \(a(x)\), inverse source with quadratic gradient nonlinearity, and inverse source with cubic reaction nonlinearity [1908.01239]. The analysis is carried out in both an all-at-once formulation and a reduced formulation.

In the all-at-once setting, one solves for both parameter \(\theta\) and state \(u\), with
\[
F(\theta,u)=
\begin{pmatrix}
\dot{u}-f(\theta,u)\\
u(0)-u_0\\
\mathcal{C}u
\end{pmatrix},
\]
and the general all-at-once tangential cone condition is
\[
\begin{aligned}
&\|f(\tilde\theta,\tilde u)-f(\theta,u)-f_\theta'(\theta,u)(\tilde\theta-\theta)-f_u'(\theta,u)(\tilde u-u)\|_{\mathcal{W}^*}\\
&\le c_{tcc}^{AAO}\Bigl( \|\dot{\tilde u}-\dot u-f(\tilde\theta,\tilde u)+f(\theta,u)\|_{\mathcal{W}^*}^2 +\|\tilde u(0)-u(0)\|_H^2 +\|\mathcal{C}(\tilde u-u)\|_{\mathcal{Y}}^2 \Bigr)^{1/2}.
\end{aligned}
\]
For practical verification, the paper uses the simplified form
\[
\|f(\tilde\theta,\tilde u)-f(\theta,u)-f_\theta'(\theta,u)(\tilde\theta-\theta)-f_u'(\theta,u)(\tilde u-u)\|_{\mathcal{W}^*}
\le
c_{tcc}^{AAO}\|\mathcal{C}(\tilde u-u)\|_{\mathcal{Y}}.
\]
Under stability of the linearized state equation, this implies the reduced tangential cone condition for \(F=\mathcal{C}\circ S\), with
\[
c_{tcc}^{Re}=C_{lin}\,c_{tcc}^{AAO}
\]
[1908.01239].

For electrical impedance tomography, the forward operator is
\[
F(\gamma)=\Lambda_\gamma-\Lambda_1,
\]
where \(\Lambda_\gamma\) is the Dirichlet-to-Neumann map for
\[
(\gamma\nabla u)=0 \quad \text{in }\Omega,\qquad u|_{\partial\Omega}=f.
\]
The derivative is
\[
\langle\Lambda'_\gamma(w)f,g\rangle = \int_{\Omega} w\,\nabla u_{\gamma,f}\cdot\nabla u_{\gamma,g}\,dx.
\]
A key estimate in that paper is the Löwner-order bound
\[
0 \;\preceq\; \Lambda_\gamma - \Lambda_{\gamma^\dagger} - \Lambda'_\gamma(\gamma-\gamma^\dagger) \;\preceq\; \Lambda'_\gamma\!\left(\frac{|\gamma-\gamma^\dagger|^2}{\gamma^\dagger}\right),
\]
which yields the norm estimate
\[
\|F(\gamma) - F(\gamma^\dagger) - F'[\gamma](\gamma-\gamma^\dagger)\|_Y^2
\le
\big\|F'[\gamma]\Big(\frac{|\gamma-\gamma^\dagger|^2}{\gamma^\dagger}\Big)\big\|_Y^2.
\]
From this, the paper derives sufficient criteria for weak and strong TCC. In particular, if
\[
\|F'[\gamma](|\gamma-\gamma^\dagger|^2/\gamma^\dagger)\|_Y
\le \zeta\,\|F'[\gamma](\gamma-\gamma^\dagger)\|_Y
\]
with \(\zeta<1\), then the strong TCC holds with \(\eta=\frac{\zeta}{1-\zeta}\); and for conductivities with \(\gamma\ge\gamma^\dagger\) or \(\gamma\le\gamma^\dagger\) almost everywhere and sufficiently small \(\|\gamma-\gamma^\dagger\|_\infty\), the strong TCC holds locally [2105.02635].

## 4. Tangent cone versus linearized cone in optimization and variational analysis

In Banach-space constrained optimization, the tangential cone condition appears as equality between the tangent cone and the linearized cone. For a feasible set
\[
\mathcal{F} := \{ x\in E \mid h_i(x)=0,\ i\in I_0;\ \ h_i(x)\le 0,\ i\in I\},
\]
with \(E\) a real Banach space and \(h_i\) continuously Fréchet differentiable, the tangent cone is
\[
\begin{aligned}
T_C(x_0) := \Big\{d\in E \ \Big|\ &\exists \varepsilon>0,\ \exists r:(0,\varepsilon]\to E \text{ with } \|r(t)\|/t \to 0 \ (t\downarrow 0),\\
&\text{such that } x_0 + t d + r(t) \in C \ \text{for all } t\in(0,\varepsilon]\Big\},
\end{aligned}
\]
and the linearized cone is
\[
\Gamma_{\mathcal{F}}(x_0) := \big\{d\in E \mid \langle D h_i(x_0), d\rangle = 0 \ \forall i\in I_0,\  \langle D h_i(x_0), d\rangle \le 0 \ \forall i\in I(x_0)\big\}.
\]
Under the relaxed constant rank constraint qualification, the paper proves the Abadie condition
\[
\Gamma_{\mathcal{F}}(x_0) = T_{\mathcal{F}}(x_0)
\]
[1905.05581].

The relaxed constant rank condition requires that for every index set
\[
I_0\subset J\subset I_0\cup I(\bar x),
\]
the rank of \(\{Dh_i(x)\}_{i\in J}\) is constant on a neighborhood of \(\bar x\). Under this hypothesis, the paper proves that every \(d\in\Gamma_{\mathcal{F}}(x_0)\) admits a correction \(r(t)\) with \(\|r(t)\|/t\to0\) such that the trajectory \(x_0+td+r(t)\) remains feasible for small \(t\). It also derives existence of Lagrange multipliers for local minima, using the explicit polar of the linearized cone
\[
\Gamma_{\mathcal{F}}(x_0)^\circ =
\Big\{ d^* \in E^* \,\Big|\, d^* = \sum_{i\in I_0\cup I(x_0)} \lambda_i Dh_i(x_0),\ \lambda_i\ge 0\ (i\in I(x_0)),\ \lambda_i\in\mathbb{R}\ (i\in I_0)\Big\}.
\]
Here the tangential cone condition is not an inequality on \(F\), but an exact linearization property for the feasible set itself. This suggests a direct analogy with inverse-problem TCC: in both settings, first-order information is required to reproduce the true infinitesimal geometry without loss.

## 5. Geometric cone conditions: manifolds, convex sets, and determinantal varieties

A distinct geometric usage occurs in the characterization of \(C^1\) manifolds by tangent and paratangent cones. For \(F\subset\mathbb{R}^n\) and \(x\in F\), the paper defines
\[
\mathrm{Tan}^-(F,x) :=\liminf_{\lambda\to0^+} \frac{F-x}{\lambda}, \qquad
\mathrm{Tan}^+(F,x) :=\limsup_{\lambda\to0^+} \frac{F-x}{\lambda},
\]
and
\[
\mathrm{pTan}^-(F,x) :=\liminf_{\lambda\to0^+,\,F\ni y\to x}\frac{F-y}{\lambda}, \qquad
\mathrm{pTan}^+(F,x) :=\limsup_{\lambda\to0^+,\,F\ni y\to x}\frac{F-y}{\lambda}.
\]
These satisfy
\[
\mathrm{pTan}^-(F,x)\subset \mathrm{Tan}^-(F,x)\subset \mathrm{Tan}^+(F,x)\subset \mathrm{pTan}^+(F,x).
\]
The main characterization is that a non-empty set \(F\subset\mathbb{R}^n\) is a \(C^1\) submanifold of \(\mathbb{R}^n\) iff \(F\) is locally compact and
\[
\mathrm{pTan}^-(F,x)=\mathrm{pTan}^+(F,x)\quad\text{for every }x\in F,
\]
which forces all four cones to coincide and to be vector subspaces [1202.2760].

In convex analysis, the paper on facially dual complete cones introduces tangential exposure:
\[
\mT(x;C)\cap \lspan(F-x) = \mT(x;F)\quad \forall F\lhd C,\ \forall x\in F.
\]
For cones, this becomes
\[
\mT(x;C)\cap \lspan F = \mT(x;F).
\]
The paper proves that if a closed convex cone \(K\) is facially dual complete, then for every face \(F\lhd K\) and every \(x\in F\),
\[
\mT(x;K)\cap \lspan F = \mT(x;F),
\]
so every facially dual complete cone is tangentially exposed; it then defines lexicographic tangent cones by recursive iteration of the tangent-cone operation and proves that strong tangential exposure is sufficient for facial dual completeness [1704.06368].

For low-rank matrix geometry, the determinantal-variety paper studies the set
\[
R_{\le r}^{m\times n}:=\{X\in\mathbb{R}^{m\times n}\mid \operatorname{rk}X\le r\}.
\]
At a rank-\(r\) point \(X\), the Bouligand tangent cone to \(R_{\le r_1}^{m\times n}\) is
\[
T_{R_{\le r_1}^{m\times n}}(X) = T_{R_r^{m\times n}}(X) \;\oplus\; N_{R_r^{m\times n}}(X)\cap R_{\le r_1-r}^{m\times n},
\]
equivalently,
\[
\eta\in T_{R_{\le r_1}^{m\times n}}(X)
\quad\text{iff}\quad
\eta = [U\;U_\perp]
\begin{bmatrix}
A & B\\
C & D
\end{bmatrix}
[V\;V_\perp]^\top,\qquad \operatorname{rk}D\le r_1-r.
\]
The paper proves that the tangent cone correspondence \(X\mapsto T_{R_{\le r}^{m\times n}}(X)\) is continuous relative to \(R_{\le r}^{m\times n}\) at every point of rank exactly \(r\), while relative to the smooth stratum \(R_r^{m\times n}\) it is neither inner semicontinuous nor outer semicontinuous at lower-rank points [2201.03979]. Here again, the operative issue is whether tangent geometry varies in a controlled manner.

## 6. Broader mathematical uses and interpretive cautions

The phrase “tangent cone” also appears in the Tangent Cone theorem for cohomology jump loci. For a finite-type CW-complex \(X\), if \(X\) is \(q\)-formal, then for all \(i\le q\),
\[
T_1(\mathcal{V}^i(X)) = TC_1(\mathcal{V}^i(X)) = \mathcal{R}^i(X),
\]
where \(\mathcal{V}^i(X)\) is the characteristic variety and \(\mathcal{R}^i(X)\) is the resonance variety [1502.02279]. In that setting, “satisfying the tangent cone condition” means that the local analytic geometry of \(\mathcal{V}^i(X)\) at the identity is completely determined by resonance; failure of equality is used as a non-formality obstruction. Although this is not the same condition as the operator inequality of inverse problems, it is another first-order exactness principle.

A common misconception is that “tangential cone condition” has a single universal definition. The literature represented here uses the phrase in several technically distinct senses: an operator inequality in inverse problems, an equality \(T_C(x_0)=\Gamma_C(x_0)\) in constrained optimization, coincidence of paratangent cones in manifold theory, compatibility of tangent cones with facial structure in convex geometry, and equality of analytic and exponential tangent cones in cohomology jump loci. A plausible implication is that the phrase should always be read relative to the ambient framework.

Another misconception is that first-order tangential regularity automatically propagates to higher-order or stratified settings. The convex-geometric examples show that tangential exposure need not imply strong tangential exposure, and that strong tangential exposure is sufficient but not necessary for facial dual completeness [1704.06368]. The determinantal-variety analysis likewise shows continuity of tangent cones at maximal allowed rank and loss of semicontinuity at lower-rank points [2201.03979]. These examples indicate that tangential control can be sharply local, stratified, and hierarchy-dependent.

Taken together, these works support a unifying interpretation: a tangential cone condition is a mechanism for certifying that infinitesimal data—derivatives, tangent cones, paratangent cones, or tangent spaces—are faithful enough to govern local behavior. The exact form of that fidelity depends on whether the objective is algorithmic convergence, geometric regularity, dual stability, or topological rigidity.

Source: https://www.emergentmind.com/topics/tangential-cone-condition