Directional Stability: Definition and Applications

• Directional Stability is a concept defining robustness and convergence measured relative to a specific direction rather than an undirected neighborhood.
• It employs methodologies like directional metric regularity, slope-based and coderivative tests, and orientation-dependent energy assessments.
• Applications span perturbation analysis in control systems, numerical PDE schemes, deep learning convergence, and crystallographic interface evaluations.

Directional stability denotes a class of stability concepts in which robustness, convergence, or sensitivity is evaluated relative to a prescribed direction, orientation, or marching axis rather than only in an undirected neighborhood. The term has no single field-independent definition. In variational analysis it refers to direction-specific metric regularity and coderivative criteria for multifunctions (Huynh et al., 2015, Benko et al., 2017). In numerical PDEs it denotes stability inherited from representative ordinary differential equations posed along time or space (Jafarimoghaddam et al., 6 Jun 2025). In control and mechanics it describes stability of yaw, bearing formations, or basin escape with respect to particular physical directions (Lu et al., 2015, Zhao et al., 2015, Klinshov et al., 2015). In machine learning it refers to convergence of normalized parameter and gradient directions under gradient flow (Ji et al., 2020). In directional statistics it concerns the robustness of density ridges on the sphere under perturbations of the underlying density and under iterative ridge-finding algorithms (Zhang et al., 2021). In interface physics and heterostructure energetics it is tied to crystallographic direction, planar-front stability, and orientation-dependent interface enthalpy and band lineup [(Korzhenevskii et al., 2011); (Foster et al., 2014)].

1. Direction-dependent regularity in variational analysis

A central mathematical formulation of directional stability is directional Hölder or Lipschitz metric regularity for a multifunction $F:X \Rightarrow Y$ with closed graph, reference point $(\bar x,\bar y)$, direction $(u,v)$, and exponent $\alpha \in (0,1]$. The property requires constants $\delta,\varepsilon,\eta>0$ such that

$$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$

for points $(x,y)$ lying simultaneously in a small neighborhood of $(\bar x,\bar y)$, in a conic directional neighborhood $$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$, and under the gauge condition

$$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$

When this is imposed only for $(\bar x,\bar y)$0, the resulting property is directional Hölder metric subregularity (Huynh et al., 2015).

The directional formulation is explicitly weaker than classical, undirected metric regularity when $(\bar x,\bar y)$1, because the gauge condition is essential in the directional setting. It is also strictly stronger than directional Hölder or Lipschitz metric subregularity. In the Lipschitz case $(\bar x,\bar y)$2, the classical relation to the Aubin property of $(\bar x,\bar y)$3 and to Lyusternik–Graves-type error bounds remains the background model, while the directional theory replaces global criteria by direction-restricted slope and coderivative tests (Huynh et al., 2015).

The slope-based characterization uses the lower semicontinuous envelope

$(\bar x,\bar y)$4

together with local and nonlocal slopes. For all $(\bar x,\bar y)$5, directional Hölder metric regularity is equivalent to positivity of the directional nonlocal slope limit inferior

$(\bar x,\bar y)$6

For $(\bar x,\bar y)$7, positivity of the corresponding local slope limit inferior is also equivalent; for $(\bar x,\bar y)$8, the local slope condition is sufficient but not necessary, as illustrated by $(\bar x,\bar y)$9 (Huynh et al., 2015).

A complementary finite-dimensional framework is provided by directional limiting normal cones, subdifferentials, and coderivatives. For a closed set $(u,v)$0, the directional limiting normal cone

$(u,v)$1

refines the nondirectional Mordukhovich normal cone by retaining the approach direction. Analogously, the directional coderivative of $(u,v)$2 at $(u,v)$3 in direction $(u,v)$4 is

$(u,v)$5

This directional calculus supports direction-specific tests for metric subregularity, metric regularity, subtransversality, calmness, and the Aubin property under very weak directional qualification conditions (Benko et al., 2017).

2. Perturbation stability and sensitivity analysis

Directional stability is especially prominent in perturbation theory. In the variational-analytic setting, exclusion of directional criticality yields persistence of regularity under smooth perturbations. In Asplund spaces, the directional critical limit set $(u,v)$6 encodes scaled coderivative behavior along $(u,v)$7. If $(u,v)$8, then $(u,v)$9 is directionally Hölder metrically regular of order $\alpha \in (0,1]$0; moreover, if $\alpha \in (0,1]$1 is differentiable near $\alpha \in (0,1]$2 and satisfies

$\alpha \in (0,1]$3

on a directional neighborhood, then $\alpha \in (0,1]$4 is directionally Hölder metrically regular at $\alpha \in (0,1]$5 in direction $\alpha \in (0,1]$6 (Huynh et al., 2015).

For Lipschitz perturbations in the case $\alpha \in (0,1]$7, the modulus transforms explicitly. If $\alpha \in (0,1]$8 is directionally metrically regular at $\alpha \in (0,1]$9 in direction $\delta,\varepsilon,\eta>0$0 with modulus $\delta,\varepsilon,\eta>0$1, and if $\delta,\varepsilon,\eta>0$2 is locally Lipschitz near $\delta,\varepsilon,\eta>0$3 with Lipschitz constant $\delta,\varepsilon,\eta>0$4 and Hadamard differentiable at $\delta,\varepsilon,\eta>0$5 in direction $\delta,\varepsilon,\eta>0$6, then $\delta,\varepsilon,\eta>0$7 remains directionally metrically regular whenever $\delta,\varepsilon,\eta>0$8, with new modulus

$\delta,\varepsilon,\eta>0$9

In both the Lipschitz and Hölder perturbation results, the output direction is transformed by the first-order rule $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$0 (Huynh et al., 2015).

The same directional viewpoint enters parameterized optimization. For

$$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$1

with feasible map $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$2, solution map $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$3, and value function $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$4, directional metric regularity of $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$5 at $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$6 in direction $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$7 yields constructive feasible-direction results and directional sensitivity bounds. If $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$8 and $$\operatorname{dist}(x,F^{-1}(y)) \le \kappa [\operatorname{dist}(y,F(x))]^\alpha$$9 is directionally metrically regular at $(x,y)$0 in direction $(x,y)$1, then $(x,y)$2 is a feasible direction. In finite-dimensional $(x,y)$3, the same assumptions imply the directional Robinson condition

$(x,y)$4

Under additional assumptions, the directional Hadamard derivatives of the value function satisfy

$(x,y)$5

and if each multiplier set $(x,y)$6 is a singleton, then

$(x,y)$7

Directional regularity also yields feasibility estimates of the form

$(x,y)$8

which in turn imply directional Hölder or Lipschitz sensitivity of $(x,y)$9 (Huynh et al., 2015).

A related but distinct inverse-problem notion appears for the dissipative wave equation with boundary damping coefficient $(\bar x,\bar y)$0. For the measurement map

$(\bar x,\bar y)$1

the paper proves a directional Lipschitz stability estimate at the origin. For a fixed nontrivial $(\bar x,\bar y)$2, if the undamped solution satisfies $(\bar x,\bar y)$3 in $(\bar x,\bar y)$4, then there exist $(\bar x,\bar y)$5 and $(\bar x,\bar y)$6 such that

$(\bar x,\bar y)$7

for $(\bar x,\bar y)$8. The linearization is

$(\bar x,\bar y)$9

in $$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$0. This is directional because stability is proved only along the ray $$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$1, not uniformly over all nearby coefficients (Choulli et al., 2015).

Obstacle-type quasi-variational inequalities provide another perturbation-based setting. For the parameterized fixed-point problem

$$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$2

under positive superhomogeneity and monotonicity, the minimal and maximal solutions $$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$3 and $$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$4 are locally Lipschitz on the strictly positive cone. If $$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$5 and $$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$6, then

$$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$7

$$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$8

Under concavity hypotheses, the maximal solution map is Hadamard directionally differentiable; in the uniquely solvable case the derivative is the smallest solution of the linearized fixed-point equation

$$(\bar x,\bar y)+\operatorname{cone} B((u,v),\varepsilon)$$9

The results are stated without sign restrictions on perturbation directions and apply simultaneously to elliptic and parabolic QVIs (Christof et al., 2021).

3. Numerical and PDE notions of directional stability

In discretization of advection–diffusion equations, directional stability is introduced through a directional-ODE framework that treats discrete PDE information as a representative ODE posed along a chosen marching direction, either time or space. For the general nonlinear ADE

$$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$0

the temporal-ODE scheme freezes space and marches in time, while the spatial-ODE scheme freezes time and marches in space (Jafarimoghaddam et al., 6 Jun 2025).

For the 1D diffusion equation, the zeroth-order temporal representative ODE at node $$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$1 is

$$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$2

with exact update

$$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$3

Higher-order temporal schemes approximate the neighbor sum by a polynomial and preserve the exact exponential damping $$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$4. The paper defines directional stability as stability inherited from the exact solution of the representative ODE along the chosen direction and assessed on the discrete interval. For diffusion-dominated ODEs, the temporal kernel $$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$5 and the spatial kernels $$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$6 enforce monotone damping or regularization in the marching direction (Jafarimoghaddam et al., 6 Jun 2025).

The nonlinear diffusion representative ODE

$$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$7

admits a stationary-point analysis. Ignoring sources, the steady state satisfies $$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$8, and perturbation by $$\operatorname{dist}(y,F(x)) \le \eta \|(x,y)-(\bar x,\bar y)\|^{1/\alpha}.$$9 gives $(\bar x,\bar y)$00 with $(\bar x,\bar y)$01. Since $(\bar x,\bar y)$02 and $(\bar x,\bar y)$03, the stationary point is asymptotically stable. For the multi-stage temporal-ODE diffusion scheme, the paper proves unconditional stability in the asymptotic discrete sense $(\bar x,\bar y)$04: the limit becomes the Explicit Gauss–Seidel update for $(\bar x,\bar y)$05 and the Fully-Implicit update for $(\bar x,\bar y)$06, both known to be unconditionally convergent for diffusion-type problems. This is contrasted directly with the CFL restriction of explicit FTCS diffusion schemes (Jafarimoghaddam et al., 6 Jun 2025).

Directional solidification supplies a different PDE meaning. In a capillary-wave description of binary-alloy directional solidification, planar-front stability is analyzed relative to the pulling direction and temperature gradient. The effective Hamiltonian is written in terms of the interface position $(\bar x,\bar y)$07 and excess solute concentration $(\bar x,\bar y)$08, and the externally imposed temperature gradient $(\bar x,\bar y)$09 and pulling velocity $(\bar x,\bar y)$10 enter through the driving force. Linearization about a steadily moving planar interface leads to a universal dispersion relation for monochromatic perturbations. Instability corresponds to $(\bar x,\bar y)$11 (Korzhenevskii et al., 2011).

Within that dispersion relation, the gradient parameter $(\bar x,\bar y)$12 acts as a long-wavelength cutoff. The low-$(\bar x,\bar y)$13 branch recovers the Mullins–Sekerka instability, but for $(\bar x,\bar y)$14 the threshold shifts to a finite $(\bar x,\bar y)$15 determined by $(\bar x,\bar y)$16. A second branch has a finite gap at $(\bar x,\bar y)$17, and the neutral line $(\bar x,\bar y)$18 marks the onset of a Cahn-type oscillatory instability. Beyond threshold, the planar-front displacement $(\bar x,\bar y)$19 satisfies the nonlinear oscillator

$(\bar x,\bar y)$20

The model therefore connects directional stability of the planar front to both the Mullins–Sekerka branch and the oscillatory branch, and it links periodic interface motion to solute banding and, when low-velocity segments enter the Mullins–Sekerka unstable regime, to banded structures (Korzhenevskii et al., 2011).

4. Control, mechanics, and robustness to finite perturbations

In aircraft dynamics, directional stability is the classical yaw or weathercock stability of a vehicle with respect to sideslip. It is quantified by the yawing moment derivative with respect to sideslip, $(\bar x,\bar y)$21, with static directional stability corresponding to $(\bar x,\bar y)$22. The complete loss of a vertical stabilizer drives $(\bar x,\bar y)$23 and $(\bar x,\bar y)$24, destabilizing the Dutch-roll mode and rendering the spiral mode near-neutral or unstable (Lu et al., 2015).

One control response is differential thrust. In a damaged Boeing 747-100 model, the lateral/directional state is $(\bar x,\bar y)$25, and the damaged aircraft uses inputs $(\bar x,\bar y)$26. Differential thrust is mapped from a rudder-equivalent yawing moment by

$(\bar x,\bar y)$27

which yields

$(\bar x,\bar y)$28

for the specified flight condition. A Lyapunov-based model reference adaptive controller restores a stable flight envelope despite engine time constant and time delay, and the reported simulations show tracking errors converging to zero in approximately $(\bar x,\bar y)$29 s under the stated scenario and under $(\bar x,\bar y)$30 additive uncertainty (Lu et al., 2015).

A related robust-control design uses $(\bar x,\bar y)$31 loop-shaping. The damaged aircraft plant is again expressed in the state $(\bar x,\bar y)$32 with inputs $(\bar x,\bar y)$33, and the loop-shaping synthesis employs pre- and post-compensators $(\bar x,\bar y)$34 and $(\bar x,\bar y)$35. The reported maximum stability margin is

$(\bar x,\bar y)$36

and the closed-loop responses settle in approximately $(\bar x,\bar y)$37 s with feasible actuator effort. In that usage, directional stability means practical recovery of weathercock stability and yaw damping by a “virtual rudder” implemented through differential thrust, despite the loss of the vertical tail (Lu et al., 2015).

In multi-agent systems, directional stability appears in bearing-based formation stabilization with directed interaction topologies. Each agent obeys single-integrator dynamics $(\bar x,\bar y)$38, uses relative position measurements to its out-neighbors, and applies the distributed control law

$(\bar x,\bar y)$39

The closed-loop stacked dynamics are

$(\bar x,\bar y)$40

where $(\bar x,\bar y)$41 is the bearing Laplacian. For directed formations, stability and convergence depend on the equality

$(\bar x,\bar y)$42

which defines bearing persistence. If bearing persistence fails, undesired equilibria appear and global formation stability cannot be guaranteed. If the formation is both infinitesimally bearing rigid and bearing persistent, then

$(\bar x,\bar y)$43

so the limit formation has the same shape as the target up to translation and scaling (Zhao et al., 2015).

A dynamical-systems notion of directional stability is given by the stability threshold $(\bar x,\bar y)$44, the smallest perturbation magnitude capable of pushing a system out of an attractor’s basin. For an attractor $(\bar x,\bar y)$45,

$(\bar x,\bar y)$46

For a unit direction $(\bar x,\bar y)$47, the directional threshold is

$(\bar x,\bar y)$48

The equality

$(\bar x,\bar y)$49

identifies the most dangerous direction of perturbation. The paper contrasts this quantity with Basin Stability and emphasizes that random sampling is inefficient near threshold, deriving the asymptotic law

$(\bar x,\bar y)$50

for fixed perturbation amplitude $(\bar x,\bar y)$51 approaching the stability threshold $(\bar x,\bar y)$52 from above (Klinshov et al., 2015).

5. Directional convergence in learning and statistics

In deep learning, directional stability refers to stabilization of geometric directions during optimization even when parameter norms diverge. For deep homogeneous networks trained by gradient flow on classification losses whose minimizers are “at infinity,” the parameter trajectory satisfies

$(\bar x,\bar y)$53

and the paper proves that the normalized weights converge: $(\bar x,\bar y)$54 The normalized trajectory has finite length, so weights diverge only in norm while their direction stabilizes (Ji et al., 2020).

Homogeneity then transfers directional convergence to predictions and margins. With $(\bar x,\bar y)$55-positive homogeneity $(\bar x,\bar y)$56, the normalized margins

$(\bar x,\bar y)$57

converge for all training examples, and the margin distribution converges. Under locally Lipschitz gradients, the gradient direction also stabilizes and asymptotically aligns with the flow: $(\bar x,\bar y)$58 The analysis is based on unbounded nonsmooth Kurdyka–Łojasiewicz inequalities for definable functions, and it yields consequences for margin maximization and convergence of saliency maps (Ji et al., 2020).

For directional data on the sphere, stability concerns geometric features of an estimated density rather than optimization trajectories in parameter space. On the unit hypersphere $(\bar x,\bar y)$59, the Riemannian gradient is

$(\bar x,\bar y)$60

and the paper defines a $(\bar x,\bar y)$61-dimensional directional density ridge by

$(\bar x,\bar y)$62

Here $(\bar x,\bar y)$63 contains the eigenvectors associated with the smallest $(\bar x,\bar y)$64 eigenvalues of the spherical Hessian (Zhang et al., 2021).

The directional stability theorem states that if the regularity conditions $(\bar x,\bar y)$65 hold for $(\bar x,\bar y)$66, and if $(\bar x,\bar y)$67 is sufficiently small, then the perturbed density $(\bar x,\bar y)$68 also satisfies the ridge regularity conditions and

$(\bar x,\bar y)$69

For the plug-in directional kernel density estimator $(\bar x,\bar y)$70, this becomes

$(\bar x,\bar y)$71

The associated directional SCMS algorithm operates intrinsically on the sphere and converges linearly in geodesic distance under the stated assumptions, with contraction factor

$(\bar x,\bar y)$72

Thus, in this setting, directional stability combines perturbation robustness of ridge geometry with convergence stability of a ridge-finding iteration (Zhang et al., 2021).

6. Orientation, interfaces, and crystallographic direction

In materials theory, directional stability is tied to crystallographic orientation and interface type. For Si/ZnS abrupt and polar-compensated interfaces, the two-interface supercell energetics are written as

$(\bar x,\bar y)$73

and interface energies are compared across the $(\bar x,\bar y)$74, $(\bar x,\bar y)$75, and $(\bar x,\bar y)$76 directions (Foster et al., 2014).

The work emphasizes that directions lacking generalized mirror symmetry generate artificial bulk fields in two-slab supercells and complicate both band-offset determination and orientation-dependent stability. The adopted strategy is to place each distinct interface in a variety of supercell environments, determine valence-band offsets by linear extrapolation of the local potential from slab centers to nominal interface planes, and extract interface energies from overdetermined linear systems across several supercells. For sufficiently large supercells, the spread of offsets for a given microscopic interface is typically $(\bar x,\bar y)$77–$(\bar x,\bar y)$78 eV, much smaller than the total interfacial distribution (Foster et al., 2014).

The orientation dependence is strong. The nonpolar abrupt $(\bar x,\bar y)$79 interface has absolute interface energy $(\bar x,\bar y)$80 eV/nm$(\bar x,\bar y)$81. For $(\bar x,\bar y)$82, polar-compensated interfaces are generally more stable than abrupt ones across Zn-rich and Zn-poor conditions. For $(\bar x,\bar y)$83, reconstructions of the single-bond interfaces are stable over nearly the entire chemical-potential window $(\bar x,\bar y)$84 eV $(\bar x,\bar y)$85 eV, while abrupt triple-bond interfaces are energetically disfavored. The paper states the inequalities

$(\bar x,\bar y)$86

and

$(\bar x,\bar y)$87

for both Zn-rich and Zn-poor limits, while noting that absolute $(\bar x,\bar y)$88 interface energies cannot be extracted individually from quasi-1D supercells (Foster et al., 2014).

Band alignment is equally direction-sensitive. The valence-band offsets span roughly from $(\bar x,\bar y)$89 eV to $(\bar x,\bar y)$90 eV. For both $(\bar x,\bar y)$91 and $(\bar x,\bar y)$92 interfaces, abrupt and single-substitution compensated cases form a strongly bimodal distribution with total width greater than $(\bar x,\bar y)$93 eV: anion-terminated abrupt interfaces cluster near $(\bar x,\bar y)$94 eV, cation-terminated abrupt interfaces cluster near $(\bar x,\bar y)$95 eV, and the $(\bar x,\bar y)$96 offset $(\bar x,\bar y)$97 eV lies near the midpoint. The authors conclude that the macroscopic $(\bar x,\bar y)$98 experimental offsets imply significant selectivity among the possible microscopic interfaces (Foster et al., 2014).

A plausible implication is that, in this literature, directional stability does not merely mean dynamic return after perturbation. It can also denote orientation-dependent energetic preference, lineup robustness, and the selective stabilization of particular microscopic interface configurations.