---
title: Tangential Subdifferentials
url: https://www.emergentmind.com/topics/tangential-subdifferentials
type: topic
---

# Tangential Subdifferentials

Tangential subdifferentials are generalized first-order objects that encode nonsmooth behavior through tangent geometry. In recent variational-analysis literature, the term appears in several closely related senses: an epigraph-based subdifferential built from the contingent cone to $\operatorname{epi} f$, a directional limiting subdifferential that filters generalized gradients by the actual direction of approach, a support-function-based subdifferential for tangentially convex functions, and, under partial smoothness, the projection of the limiting subdifferential onto the tangent space of an active manifold [1106.2338][1712.04704][2501.15540][2509.03205]. The common theme is a decomposition of first-order information into tangential and normal parts, with the tangential part governing motion along feasible sets or active manifolds and the normal part encoding deviation away from them.

## 1. Foundational formulations

A basic tangential construction begins with the epigraph of an extended-real-valued function. For $f:X\to \mathbb{R}\cup\{+\infty\}$, with epigraphical mapping $E(x)=[f(x),+\infty)$, the tangent cone to $\operatorname{epi} f$ induces the contingent epigraphical directional derivative
\[
f^\circ(\bar x;h):=\inf\{r\in\mathbb{R}:(h,r)\in T_{\operatorname{epi} f}(\bar x,f(\bar x))\}.
\]
The corresponding tangential subdifferential is
\[
\partial_T f(\bar x):=\{v\in X^*:\langle v,h\rangle \le f^\circ(\bar x;h)\ \text{for all }h\in X\}.
\]
Equivalently, in terms of the graphical derivative of $E$,
\[
f^\circ(\bar x;h)=\inf E'(h),\qquad E'(h):=DE(\bar x\mid f(\bar x))(h),
\]
and $\partial_T f(\bar x)$ is the set of dual elements dominated by $\inf E'(h)$ in every direction [1106.2338].

In finite dimensions, the same paper shows that tangential and normal-cone approaches are equivalent under its generalized-derivative framework: graphical derivatives of the epigraph mapping and coderivatives of the epigraphical map yield equivalent characterizations of tangential subdifferentials. When $\operatorname{epi} f$ is Clarke regular at $(\bar x,f(\bar x))$, the tangential subdifferential coincides with the Clarke subdifferential, and for convex $f$ it agrees with the usual convex subdifferential [1106.2338].

A second foundational formulation is used for tangentially convex functions. If $\mathcal J:\mathbb{R}^n\to \mathbb{R}\cup\{+\infty\}$ is directionally differentiable at $k$ and the map $d\mapsto \mathcal J'(k;d)$ is convex, then $\mathcal J$ is tangentially convex at $k$, and its tangential subdifferential is
\[
\partial^{T}\mathcal J(k):=\{\xi\in\mathbb{R}^n:\langle \xi,d\rangle \le \mathcal J'(k;d)\ \text{for all }d\in\mathbb{R}^n\}.
\]
This set is nonempty, compact, and convex, and the directional derivative is its support function:
\[
\mathcal J'(k;d)=\max_{\xi\in \partial^{T}\mathcal J(k)}\langle \xi,d\rangle \qquad \forall d\in\mathbb{R}^n.
\]
In the smooth case, $\partial^{T}\mathcal J(k)=\{\nabla \mathcal J(k)\}$; for convex $\mathcal J$ on an open domain, $\partial^{T}\mathcal J(k)=\partial_{cvx}\mathcal J(k)$; and under Clarke regularity it coincides with the Clarke subdifferential [2509.03205].

## 2. Directional and constrained tangential subdifferentials

A directional refinement replaces non-directional limiting objects by approach-direction-dependent ones. For a closed set $S\subset \mathbb{R}^n$, the directional limiting normal cone is
\[
N_S(x;d):=\{v\in\mathbb{R}^n\mid \exists t_k\downarrow 0,\ u_k\to d,\ v_k\to v,\ x+t_k u_k\in S,\ v_k\in N_S(x+t_k u_k)\}.
\]
It satisfies $N_S(x;0)=N_S(x)$, $N_S(x;d)\subset N_S(x)$, and $N_S(x;d)=\emptyset$ if $d\notin T_S(x)$ [1712.04704].

For functions, the directional limiting subdifferential is defined on epigraph directions:
\[
\partial f(x;(h,v)):=\{\xi\in \mathbb{R}^n:(\xi,-1)\in N_{\operatorname{epi} f}((x,f(x));(h,v))\}.
\]
The paper also uses an analytic directional subdifferential $da\,f(x;h)$ and proves that, under calmness of $f$ in direction $h$,
\[
da\,f(x;h)=\bigcup_{v\in Df(x)(h)} \partial f(x;(h,v)).
\]
If $f$ is directionally differentiable at $x$ along $h$, then $v=f'(x;h)$ [1712.04704].

This framework yields a constrained or tangential subdifferential relative to a set $S$. Since
\[
\partial \delta_S(x;(h,0))=N_S(x;h),
\]
one may regard
\[
\partial_{\mathrm{tan}} f(x\mid S;h,v):=\partial (f+\delta_S)(x;(h,v))
\]
as the tangential subdifferential along feasible direction $h\in T_S(x)$. The directional sum rule gives
\[
\partial(f+\delta_S)(x;(h,v))\subset \partial f(x;(h,v))+N_S(x;h),
\]
and, under calmness and weak directional qualification conditions, equality often holds [1712.04704].

The directional viewpoint sharpens classical subdifferential calculus because constraint contributions vanish in interior-pointing directions. For the half-space $S=\{x:\langle c,x\rangle \le 0\}$ with $\langle c,x\rangle=0$, one has
\[
N_S(x;h)=\{0\}\ \text{if}\ \langle c,h\rangle<0,\qquad 
N_S(x;h)=\mathbb{R}_+c\ \text{if}\ \langle c,h\rangle=0.
\]
Thus, for strictly interior feasible directions, the tangential subdifferential of $f+\delta_S$ reduces to the unconstrained directional subdifferential of $f$ [1712.04704].

## 3. Partial smoothness and manifold-projected tangential components

Within partial smoothness, tangential subdifferentials arise from an active manifold. A function $f:\mathbb{R}^n\to \overline{\mathbb{R}}$ is partly smooth at $\bar x$ relative to a set $M$ containing $\bar x$ if $M$ is a $C^p$-smooth manifold around $\bar x$ and four conditions hold: smoothness of $f|_M$, prox-regularity and regularity near $\bar x$ with nonempty subdifferentials, sharpness, and continuity of $\partial f|_M$. The sharpness condition is
\[
T_M(\bar x)=\operatorname{LinHull}(\partial f(\bar x))^\perp,
\]
equivalently
\[
N_M(\bar x)=\operatorname{Lin}(\partial f(\bar x)).
\]
For set-valued operators $A$, the operator version of partial smoothness replaces $\partial f$ by $A$ and requires
\[
N_M(\bar x)=\operatorname{Lin}(A(\bar x)),
\]
together with continuity of $\operatorname{proj}_{T_M(x)}(A(x))$ along $M$ [2501.15540].

The operator framework provides a canonical “smooth representative”
\[
\widetilde A(x)=\operatorname{proj}_{T_M(x)}(A(x)),
\]
which is single-valued and continuous along $M$ near $\bar x$. For functions, with $A=\partial f$, this motivates the tangential subdifferential
\[
\partial_T f(x):=\operatorname{proj}_{T_M(x)}(\partial f(x)).
\]
Under partial smoothness, $\partial_T f(x)$ is single-valued and equals the Riemannian gradient of the restriction:
\[
\partial_T f(x)=\nabla(f|_M)(x).
\]
The normal component is generated by the subdifferential span, yielding
\[
\partial f(x)\subset \nabla(f|_M)(x)+N_M(x),
\]
with equality at $\bar x$ and locally along $M$ under the sharpness and continuity assumptions [2501.15540].

This tangential-normal decomposition drives identifiability. The paper introduces the local union
\[
U:=\bigcup_{x\in M\cap B_\varepsilon(\bar x)}\bigl(x+\gamma A_\varepsilon(x)\bigr),
\]
where $A_\varepsilon$ is an $\varepsilon$-localization around $(\bar x,\bar u)$ and $\gamma\in (0,1/r)$. It proves that $\operatorname{Span}(U)=\mathbb{R}^n$ and
\[
J_{\gamma A_\varepsilon}(U)=M\cap B_\varepsilon(\bar x),
\]
and if $\bar z=\bar x+\gamma \bar u\in \operatorname{int}(U)$ with $d=\operatorname{dist}(\bar z,\partial U)$, then any sequence $x^k\to \bar x$ satisfying
\[
\limsup_k \operatorname{dist}(\bar u,A_\varepsilon(x^k))< d/\gamma
\]
eventually lies in $M$ [2501.15540]. This removes the need for exact dual convergence and allows identification under non-vanishing errors and degeneracy.

The examples are explicit. For $f(x)=\|x\|_1$ and support $S=\{i:\bar x_i\neq 0\}$, the active manifold is
\[
M=\{x\in \mathbb{R}^n:\operatorname{supp}(x)\subseteq S\},
\]
the tangent subspace is the coordinate subspace on $S$, and
\[
\partial_T\|x\|_1(\bar x)=\operatorname{proj}_{T_M(\bar x)}(\partial \|\bar x\|_1)
\]
equals $\operatorname{sign}(\bar x)_S$ on $S$ and $0$ on $S^c$, which is exactly $\nabla(\|\cdot\|_1|_M)(\bar x)$. For $f(x)=\|x\|_0$, the tangential component vanishes while the normal component is $\operatorname{span}\{e_i:i\in S^c\}$ [2501.15540].

## 4. Tangential transversality and Clarke subdifferential calculus

Not all uses of “tangential” introduce a new subdifferential. In the Banach-space transversality framework of Bivas, Krastanov, and Ribarska, the central objects are uniform tangent sets and strong tangential transversality, and the subdifferentials that appear are Clarke’s $\partial^C$ and the singular $\partial^0$ [1810.01814].

For a closed set $S\subset X$ and $x_0\in S$, a bounded set $D_S(x_0)\subset X$ is a uniform tangent set if, for each $\varepsilon>0$, there exists $\delta>0$ such that for every $v\in D_S(x_0)$ and every $x\in S\cap (x_0+\delta B)$ one can find $\lambda>0$ with
\[
S\cap (x+t(v+\varepsilon B))\neq \emptyset\qquad \forall t\in [0,1].
\]
Such sets are contained in the Clarke tangent cone $T^C_S(x_0)$, and under separability or hypertangent assumptions they generate that cone [1810.01814].

For two closed sets $A,B\subset X$, strong tangential transversality at $x_0\in A\cap B$ means that there exist uniform tangent sets $D_A(x_0)$ and $D_B(x_0)$ and some $p>0$ such that
\[
p\overline B\subset \operatorname{co}\bigl(D_A(x_0)-D_B(x_0)\bigr).
\]
This implies tangential transversality and transversality of the Clarke tangent cones. The payoff is the Clarke normal intersection property
\[
N^C_{A\cap B}(x_0)\subset N^C_A(x_0)+N^C_B(x_0),
\]
with the sum on the right weak* closed [1810.01814].

The same paper applies this to split epigraphs
\[
C_1=\{(x,r_1,r_2): r_1\ge f_1(x)\},\qquad
C_2=\{(x,r_1,r_2): r_2\ge f_2(x)\},
\]
and derives a Clarke sum rule:
\[
\partial^C(f_1+f_2)(x_0)\subset \partial^C f_1(x_0)+\partial^C f_2(x_0)
\]
whenever $C_1$ and $C_2$ satisfy the required normal intersection property; strong tangential transversality is a sufficient condition. The paper also shows that cone transversality alone may fail to guarantee weak* closedness of $N^C_A+N^C_B$, as illustrated by Example 3.19 in $X=c_0(\mathbb{N})$ [1810.01814].

A recurrent misconception is therefore that “tangential subdifferential” must always denote a distinct subdifferential construction. The transversality paper shows a different usage: “tangential” can instead designate primal tangent-set hypotheses that enable Clarke normal-cone and Clarke subdifferential calculus.

## 5. Tangential subdifferentials in nonsmooth mathematical programming

The most explicit optimization-oriented use of the term appears in nonsmooth mathematical programs with equilibrium constraints. In the MPEC setting
\[
\min \mathcal J(k)\quad \text{subject to}\quad
\ell(k)\le 0,\ \bar h(k)=0,\ G(k)\ge 0,\ H(k)\ge 0,\ G(k)^\top H(k)=0,
\]
the paper works with functions that are tangentially convex at the reference point and builds all constraint qualifications and stationarity concepts from tangential subdifferentials [2509.03205].

The fundamental calculus is support-function based. If $f$ and $g$ are tangentially convex at $x$, then
\[
(f+g)'(x;d)=f'(x;d)+g'(x;d)
\]
for all $d$, and consequently
\[
\partial^{T}(f+g)(x)=\partial^{T}f(x)+\partial^{T}g(x).
\]
Positive scaling and affine perturbation behave analogously. The same paper also defines $\partial^{T}$-pseudoconvexity and $\partial^{T}$-quasiconvexity through sign conditions on $\langle \xi,t-k\rangle$ for all $\xi\in \partial^{T}\mathcal J(k)$ [2509.03205].

Constraint qualifications are expressed through dual cones built from unions of tangential subdifferentials of active constraint functions. With $\Pi(k^*)$ and $\Psi(k^*)$ defined in this way, the generalized standard Abadie CQ is
\[
\Pi(k^*)\subset T(K,k^*),
\]
the MPEC Abadie CQ is
\[
\Psi(k^*)\subset T(K,k^*),
\]
and the MPEC Zangwill CQ is
\[
\Psi(k^*)\subset \operatorname{cl} Dcon(K,k^*).
\]
Because $Dcon(K,k^*)\subset Acon(K,k^*)\subset T(K,k^*)$ and $T(K,k^*)$ is closed, MPEC Zangwill CQ implies MPEC-ACQ. The paper further proves that its MPEC weak reverse convex CQ implies the MPEC Zangwill CQ [2509.03205].

Stationarity is likewise formulated entirely in terms of tangential subdifferentials. GA-stationarity requires multipliers $\lambda$ and $\mu$ satisfying an inclusion of the form
\[
0\in \partial^{T}\mathcal J(k^*)+\text{multiplier-weighted sums of } \partial^{T}\ell_i(k^*),\ \partial^{T}\bar h_j(k^*),\ \partial^{T}(\pm G_i)(k^*),\ \partial^{T}(\pm H_i)(k^*),
\]
together with nonnegativity and complementarity-index zeroing conditions. GS-stationarity strengthens this by imposing $\mu_i^G=\mu_i^H=0$ for all degenerate indices $i\in \Omega$ [2509.03205].

Necessary and sufficient optimality results follow. If $k^*$ is a local minimizer, $\mathcal J$ is tangentially convex and locally Lipschitz near $k^*$, GS-ACQ holds, and the cone
\[
\Delta:=\operatorname{cone}(\ell)+\operatorname{cone}(\bar h)+\operatorname{cone}(G_\Theta)+\operatorname{cone}(H_\Upsilon)+\operatorname{cone}((GH)_\Omega)
\]
is closed, then $k^*$ is GS-stationary. Under MPEC-ACQ and tangential convexity of the effective data, $k^*$ is GA-stationary. Conversely, if $k^*$ is GA-stationary, $\mathcal J$ is $\partial^{T}$-pseudoconvex, the active constraints satisfy the stated $\partial^{T}$-quasiconvexity assumptions, and the multiplier-index sets obey
\[
\Omega_{\mu}^{G}\cup \Omega_{\mu}^{H}\cup \Theta_{\mu}^{+}\cup \Upsilon_{\mu}^{+}=\emptyset,
\]
then $k^*$ is a global minimizer [2509.03205].

## 6. Relations, examples, and recurring distinctions

Several standard examples clarify how the different frameworks relate. For $f(x)=|x|$ at $0$, the epigraph tangent cone is
\[
T_{\operatorname{epi} f}(0,0)=\{(h,r):r\ge |h|\},
\]
hence $f^\circ(0;h)=|h|$ and
\[
\partial_T f(0)=[-1,1].
\]
For $f(x)=\|x\|_2$ at the origin, one obtains $\partial_T f(0)=\mathbb{B}_2$; for $f(x)=x^2\sin(1/x)$ with $f(0)=0$, one gets $\partial_T f(0)=\{0\}$ [1106.2338].

In the support-function framework, $\mathcal J(k_1,k_2)=|k_1|+k_2^2$ at $(0,0)$ satisfies $\mathcal J'(0;d)=|d_1|$, and the paper records
\[
\partial^{T}\mathcal J(0)=\{(1,0),(-1,0),(0,0)\}.
\]
In the MPEC example with objective $\mathcal J(k_1,k_2)=|k_1|+k_2^3$ and feasible set $\{(k_1,k_2):k_2=0,\ k_1\ge 0\}$, the tangential subdifferentials
\[
\partial^{T}\mathcal J(0)=\{(1,0),(-1,0)\},\quad
\partial^{T}\ell(0)=\{(0,1),(0,-1)\},\quad
\partial^{T}(-G)(0)=\{(-1,0)\},\quad
\partial^{T}(-H)(0)=\{(0,-1)\}
\]
support a direct verification of GS-stationarity [2509.03205].

In the partial-smoothness framework, the representative example is the $\ell_1$ norm. There the tangential subdifferential is not a support-function dual object but the tangent-space projection of the limiting subdifferential:
\[
\partial_T\|x\|_1(\bar x)=\operatorname{proj}_{T_M(\bar x)}(\partial \|\bar x\|_1),
\]
which equals the Riemannian gradient of the restriction to the active manifold. A different phenomenon occurs for $\|x\|_0$, where the tangential component vanishes and the normal component carries the structure [2501.15540].

The directional framework adds a further distinction: tangential information depends on the chosen feasible direction $h$. For indicators, 
\[
\partial \delta_S(x;(h,0))=N_S(x;h),
\]
so the tangential contribution of the constraint can be zero in interior-pointing directions and nontrivial in boundary-preserving directions. This directional selectivity is precisely what makes the directional limiting approach sharper than non-directional limiting calculus [1712.04704].

Across these formulations, several equivalences recur under regularity. In the smooth case, the tangential subdifferential collapses to the gradient; for convex functions it coincides with the convex subdifferential; and under Clarke regularity it coincides with the Clarke subdifferential [1106.2338][2509.03205]. What changes from one framework to another is not the objective of encoding first-order nonsmooth information, but the geometric mechanism used to extract the tangential part: contingent epigraph geometry, direction-filtered limiting normals, projection onto an active tangent manifold, or support-function dualization of a sublinear directional derivative.

Source: https://www.emergentmind.com/topics/tangential-subdifferentials