Preprint
Article

This version is not peer-reviewed.

Separation Laws for Flip Bifurcations and Endogenous Switching Boundaries in Gradient Maps

Submitted:

10 October 2026

Posted:

12 October 2026

You are already at the latest version

Abstract
We study how radial gradient clipping affects the period-two orbit created by a supercritical flip bifurcation near a nondegenerate minimum with a simple maximal Hessian eigenvalue. Clipping leaves the local flip threshold unchanged. For a small clipping threshold \(c\), the stable smooth two-cycle first reaches the clipping boundary at a parameter distance \(Kc^2 + O(c^4)\) from the flip, with an explicit coefficient determined by the potential and the clipping norm. The two cycle points have opposite gradients and therefore reach the boundary simultaneously. A clipped two-cycle cannot have one point strictly inside the smooth region and the other strictly inside the clipped region. Every two-cycle with both points strictly clipped has an exact unit Floquet multiplier and is nonhyperbolic. A solvable planar model illustrates the transition from a stable smooth two-cycle to a neutral family of clipped two-cycles.
Keywords: 
;  ;  ;  ;  ;  ;  

1. Introduction

Finite-step gradient descent, G η ( x ) = x − η ∇ V ( x ) , is a discrete dynamical system whose behavior near the boundary of linear stability can differ from its gradient-flow approximation [1,2,3,4]. At a nondegenerate local minimum with a simple maximal Hessian eigenvalue λ * , the first stability threshold is η f = 2 / λ * . The corresponding multiplier reaches − 1 , and a nondegenerate supercritical flip creates a stable period-two branch [5,6,7]. Such oscillations are relevant to the edge-of-stability phenomenon [8]. Recent analyses describe their nonlinear stability and amplitude through higher derivatives and coupling between Hessian eigendirections [9,10,11].
Gradient clipping controls large updates in optimization [12,13,14,15]. We study radial clipping in a norm ∥ · ∥ :
F η , c ( x ) = x − η C c ( ∇ V ( x ) ) , C c ( g ) = min { 1 , c / ∥ g ∥ } g ,
with C c ( 0 ) = 0 . Its switching boundary Σ c = { x : ∥ ∇ V ( x ) ∥ = c } is generated by the same potential as the smooth dynamics. For each fixed c > 0 , clipping is inactive near a critical point. It therefore preserves the local flip, but can change the newborn oscillation once its amplitude reaches Σ c . We ask when this first contact occurs and what restrictions radial clipping imposes on subsequent two-cycles.

1.1. Main Results and Their Relation to Existing Theory

Under the hypotheses in Section 2, we establish three connected results. First, the clipped and smooth maps have identical local germs at the minimum, and the stable smooth two-cycle has leading amplitude A η − η f , where A 2 = 3 λ * 2 / Γ and Γ is the explicit third- and fourth-derivative expression in (2.22). Second, its first clipping contact satisfies
η sc ( c ) = η f + K c 2 + O ( c 4 ) , K = Γ 3 λ * 4 ∥ v * ∥ 2 .
The two cycle points have exactly opposite gradients and hence reach the boundary simultaneously. Together with the smooth signed square-root parametrization, this balance removes the term of order η − η f from the gradient exposure and yields the even-power remainder in (1.1). Third, radial clipping excludes strict smooth–clipped two-cycle itineraries, and every strictly fully clipped two-cycle has an exact unit Floquet multiplier. Thus any continuation into that regime is nonhyperbolic; its existence and nonlinear stability require additional information. A solvable planar model realizes a neutral family of clipped two-cycles after contact.
Smooth flip theory supplies the local birth mechanism [5,6,16,17]; general dynamical-systems background is given in [18,19,20]. Piecewise-smooth and border-collision theory supplies the broader switching framework [21,22,23,24,25,26,27]. Related treatments of discontinuous and nonsmooth dynamics provide further context for switching systems [28,29,30,31]. In particular, quadratic tangency between period-doubling and border-collision loci is already known [32,33,34,35,36]. Our contribution is the explicit coefficient for the endogenous gradient boundary, the improved remainder, and the exact algebraic and spectral rigidity imposed by radial clipping. We retain the smooth derivation to fix conventions and identify the coefficient entering the contact law.

1.2. Organization

Section 2 gives the setting and local stability. Section 3 constructs the smooth two-cycle. Section 4 combines switching geometry with the separation law, and Section 5 proves post-contact degeneracy. Section 6 gives the solvable model, followed by the conclusions.

2. Setting, Hypotheses and Local Stability

Throughout the paper, 〈 · , · 〉 and ∥ · ∥ 2 denote the standard Euclidean inner product and norm on R d , respectively. All gradients, Hessians, and spectral quantities are understood with respect to this Euclidean structure.
Let U ⊂ R d be an open set and let
V : U → R
be a smooth potential. For a step size η > 0 , the associated gradient-descent map is
G η ( x ) = x − η ∇ V ( x ) .
Our aim is to study how a smooth flip bifurcation of (2.1) interacts with the switching boundary generated intrinsically by gradient clipping.

2.1. Clipped Gradient Maps

Let ∥ · ∥ be a norm on R d . For a clipping threshold c > 0 , define
C c ( g ) = g , ∥ g ∥ ≤ c , c ∥ g ∥ g , ∥ g ∥ > c , g ∈ R d .
Equivalently,
C c ( g ) = min 1 , c ∥ g ∥ g ,
with the convention C c ( 0 ) = 0 .
The clipped gradient map is
F η , c ( x ) = x − η C c ( ∇ V ( x ) ) .
The maps take values in R d ; U need not be forward invariant. All orbit statements concern iterates that remain in U. Under (H2) and (H5), the sufficiently small positive levels of ∥ ∇ V ∥ near x * are C 3 hypersurfaces: the Hessian is invertible there, and the norm derivative at a nonzero vector is nonzero by Euler’s identity. No global regularity of remote components of Σ c is asserted.
The standard Euclidean gradient-clipping rule corresponds to
∥ · ∥ = ∥ · ∥ 2 .
Associated with (2.3) is the endogenous switching hypersurface
Σ c = x ∈ U : ∥ ∇ V ( x ) ∥ = c .
It separates the local phase space into the smooth region
Ω c − = x ∈ U : ∥ ∇ V ( x ) ∥ < c
and the clipped region
Ω c + = x ∈ U : ∥ ∇ V ( x ) ∥ > c .
Accordingly,
F η , c ( x ) = G η ( x ) = x − η ∇ V ( x ) , x ∈ Ω c − ,
whereas in Ω c + ,
F η , c ( x ) = x − η c ∇ V ( x ) ∥ ∇ V ( x ) ∥ .
Thus F η , c is continuous and piecewise smooth, while its switching geometry is not prescribed independently of the dynamics: the boundary Σ c is generated by the same potential V that defines the gradient map.

2.2. The Distinguished Critical Point

We focus on the dynamics near an isolated nondegenerate local minimum x * of V.
Set
H : = D 2 V ( x * ) .
Since H is symmetric, choose an orthonormal eigenbasis { v i } i = 1 d satisfying
H v i = λ i v i , 0 < λ 1 ≤ λ 2 ≤ … ≤ λ d .
We denote the maximal eigenvalue and its corresponding unit eigenvector by
λ * : = λ d , v * : = v d .
The linearization of the smooth gradient map at x * is
D G η ( x * ) = I − η H ,
and hence its multipliers are
ρ i ( η ) = 1 − η λ i , i = 1 , … , d .
The distinguished step size at which the maximal-curvature multiplier reaches − 1 is
η f : = 2 λ * .
At η = η f ,
ρ d ( η f ) = − 1 .
The parameter
μ : = η − η f
will be used as the local bifurcation parameter throughout the paper.

2.3. Higher-Order Tensors and the Flip Coefficient

We write
T : = D 3 V ( x * ) , Q : = D 4 V ( x * ) ,
viewed as symmetric multilinear forms.
For u , w ∈ R d , let
T ( u , w ) ∈ R d
denote the Euclidean Riesz representative of the linear functional z ↦ T [ u , w , z ] ; namely,
T ( u , w ) , z = T [ u , w , z ] for all z ∈ R d .
Similarly, Q ( u , w , q ) denotes the vector determined by 〈 Q ( u , w , q ) , z 〉 = Q [ u , w , q , z ] for every z. Thus parentheses denote vector contractions, while brackets denote scalar multilinear evaluations.
Define
g * : = T ( v * , v * )
and
h * : = H − 1 g * .
The nonlinear quantity governing the local flip bifurcation is
Γ : = 3 T [ v * , v * , h * ] − Q [ v * , v * , v * , v * ] .
Equivalently, using the eigenbasis (2.10),
Γ = 3 ∑ i = 1 d T [ v * , v * , v i ] 2 λ i − Q [ v * , v * , v * , v * ] .
With the flip-normal-form convention used in this paper, define
ℓ : = η f 6 Γ = Γ 3 λ * .
Thus ℓ > 0 corresponds to the supercritical case in which a stable small-amplitude period-two branch is created for η > η f .
The expression (2.23) is the multivariate gradient-map form of the nonlinear flip criterion: the cubic coupling of the critical direction to all Hessian eigendirections contributes to the sign of the bifurcation coefficient, rather than only the derivatives taken purely along v * .

2.4. Standing Hypotheses

The following assumptions are imposed unless stated otherwise.
  • Regularity. The potential satisfies
    V ∈ C ∞ ( U , R ) .
  • Nondegenerate strict local minimum. The distinguished point x * ∈ U satisfies
    ∇ V ( x * ) = 0 , H = D 2 V ( x * ) ≻ 0 .
  • Simple maximal curvature. The largest Hessian eigenvalue is simple:
    λ i < λ d = λ * , 1 ≤ i < d .
  • Supercritical flip nondegeneracy. The coefficient defined in (2.22) satisfies
    Γ > 0 ,
    or equivalently
    ℓ > 0 .
  • Smooth clipping geometry. The clipping norm ∥ · ∥ is of class C 3 on R d ∖ { 0 } .

2.5. Immediate Spectral Consequences

Assumptions (H2)–(H3) imply that the spectral hypotheses required for a simple codimension-one flip bifurcation are automatic for the gradient map.
Indeed, for i = 1 , … , d − 1 ,
0 < λ i < λ * ,
and therefore, at η = η f ,
− 1 < 1 − 2 λ i λ * < 1 .
Hence − 1 is the unique multiplier of D G η f ( x * ) on the unit circle.
Moreover,
d d η ρ d ( η ) = − λ * ,
so that
d d η ρ d ( η ) η = η f = − λ * ≠ 0 .
Thus the critical multiplier crosses − 1 transversally as η passes through η f .
Thus, under the spectral assumptions (H2)–(H3), the sign of Γ provides the remaining nonlinear nondegeneracy condition that distinguishes the supercritical and subcritical flip cases.

2.6. Local Identity and Fixed Points

Proposition 
(Fixed-point invariance). For η , c > 0 , Fix ( F η , c ) = Crit ( V ) .
Proof. 
Radial clipping vanishes precisely when its argument vanishes; therefore F η , c ( x ) = x is equivalent to ∇ V ( x ) = 0 . □
Theorem 
(Local identity and stability). For each fixed c > 0 there is a radius r c > 0 , independent of η, such that
F η , c = G η on B r c ( x * ) , D F η , c ( x * ) = I − η H .
Under(H1)–(H3), x * is locally asymptotically stable for 0 < η < η f and unstable for η > η f . At η f , its only unit-circle multiplier is the simple − 1 .
Proof. 
Choose r 0 > 0 with B r 0 ( x * ) ¯ ⊂ U and put L * = sup y ∈ B r 0 ( x * ) ∥ D 2 V ( y ) ∥ op , where ∥ A ∥ op = sup ∥ z ∥ 2 = 1 ∥ A z ∥ uses the Euclidean norm in the domain and the clipping norm in the range. The integral formula for the gradient gives ∥ ∇ V ( x ) ∥ ≤ L * ∥ x − x * ∥ 2 . Hence r c = min { r 0 , c / ( 2 L * ) } ensures that clipping is inactive, proving the local identity. The multipliers are 1 − η λ i : all have modulus less than one below 2 / λ * , whereas 1 − η λ * < − 1 above it. Simplicity of λ * gives the assertion at the threshold. Linearization alone does not determine stability at η f . □
Corollary 
(Local normal-form invariance). The local jets, center-manifold equations and smooth flip coefficients of F η , c and G η coincide. In particular, ℓ is independent of c.
Proof. 
The maps agree on an open neighborhood of x * , so their derivatives and local invariance equations agree. A sufficiently small center manifold of the smooth map gives a common choice; uniqueness of center manifolds is not required. □
The identity also gives equality of local stable sets defined by orbits that remain in this neighborhood, and, when the maps are local diffeomorphisms, equality of unstable-manifold germs. The radius is O ( c ) as c → 0 : local clipping independence does not supply a uniform neighborhood in this limit. Positive definiteness of H excludes minimizer manifolds and additional center directions; multiple critical eigenvalues and degenerate flip coefficients are outside the present scope.

3. Smooth Flip Bifurcation and the Emerging Period-Two Branch

In this section we study the smooth gradient map
G η ( x ) = x − η ∇ V ( x )
near the critical step size
η f = 2 λ * .
The local mechanism considered here is the classical simple flip bifurcation of a smooth discrete map [5,6]. For gradient-descent dynamics, Mulayoff and Stich [9] recently derived a multivariate nonlinear stability criterion involving the second-, third-, and fourth-order derivatives of the potential, while Gan [10] formulated edge-of-stability dynamics in a systematic bifurcation-theoretic framework.
Accordingly, the existence of a smooth period-two branch is not the new contribution of the present paper. We reproduce the local reduction because its explicit coefficient enters the switching-separation law derived later; we therefore give a self-contained center-manifold and normal-form derivation in the notation of the present paper.
By the local identity theorem in Section 2, the resulting local bifurcation structure is inherited exactly by the clipped map F η , c as long as the bifurcating orbit remains inside the smooth region Ω c − .
Throughout this section we write
μ = η − η f .
Thus the bifurcation point corresponds to μ = 0 .
Under (H2)–(H3), the multiplier associated with v * is
ρ * ( μ ) = 1 − ( η f + μ ) λ * = − 1 − λ * μ ,
whereas all remaining multipliers lie strictly inside the unit disk at μ = 0 .
The nonlinear character of the bifurcation is governed by the coefficient
ℓ = Γ 3 λ * ,
defined in (2.24). Assumption (H4) gives
ℓ > 0 ,
which corresponds to the supercritical case.

3.1. Parameter-Dependent Center Reduction

Translate the critical point to the origin, y = x − x * , and define
Φ μ ( y ) : = G η f + μ ( x * + y ) − x * .
Then Φ μ ( 0 ) = 0 for every sufficiently small μ , and at μ = 0 ,
D Φ 0 ( 0 ) = I − η f H ,
whose unique critical multiplier is − 1 with eigenvector v * ; the complementary spectrum lies strictly inside the unit disk.
To incorporate the bifurcation parameter, consider the extended map
( y , μ ) ⟼ Φ μ ( y ) , μ .
Its two-dimensional local center manifold is associated with the critical state direction v * and the neutral parameter direction, and after restricting to a sufficiently small neighborhood can be represented as
y = z v * + h ( z , μ ) ,
with
h ( z , μ ) ∈ v * ⊥ , h ( 0 , μ ) = 0 , D z h ( 0 , 0 ) = 0 , h ( z , μ ) = O z 2 + | μ | | z | .
The dynamics on the center manifold are described by the scalar map
z n + 1 = f μ ( z n ) .

3.2. Flip Normal Form

We next record the local normal form required for the asymptotic analysis.
Proposition 
(Reduced flip normal form). Assume(H1)–(H4). There exists a local near-identity state coordinate, preserving z = 0 and leaving the physical parameter μ = η − η f unchanged, in which the reduced map (3.8) takes the form
f μ ( z ) = − 1 + λ * μ z + ℓ z 3 + R ( z , μ ) ,
where
R ( z , μ ) = O | z | 4 + | μ | | z | 2 + μ 2 | z |
as ( z , μ ) → ( 0 , 0 ) .
The cubic coefficient is
ℓ = η f 6 3 T [ v * , v * , H − 1 g * ] − Q [ v * , v * , v * , v * ] ,
or equivalently
ℓ = Γ 3 λ * .
Proof. 
The critical multiplier − 1 is simple and crosses the unit circle transversally (Section 2.5), so the standard parameter-dependent center-manifold and flip-normal-form reductions apply; see [5,37,38]. The linear parameter coefficient is ρ * ( μ ) = − 1 − λ * μ by (3.2), and the quadratic term is eliminated by the standard near-identity transformation for a simple flip; the remaining cubic coefficient is the first flip coefficient.
For the gradient map
G η ( x ) = x − η ∇ V ( x ) ,
one has
D 2 G η f ( x * ) [ u , w ] = − η f T ( u , w ) ,
D 3 G η f ( x * ) [ u , w , q ] = − η f Q ( u , w , q ) .
Writing J = I − η f H , B 0 = D 2 G η f ( x * ) , and C 0 = D 3 G η f ( x * ) , the standard flip coefficient is
ℓ = 1 6 〈 v * , C 0 [ v * , v * , v * ] 〉 − 1 2 〈 v * , B 0 [ v * , ( J − I ) − 1 B 0 [ v * , v * ] ] 〉 .
Here the left and right critical eigenvectors coincide because J is symmetric, and J − I = − η f H is invertible. Substitution gives
ℓ = η f 2 T [ v * , v * , H − 1 g * ] − η f 6 Q [ v * , v * , v * , v * ] .
Using
η f = 2 λ *
and the definition of Γ in (2.22) yields
ℓ = Γ 3 λ * .
The remainder estimate (3.10) follows from the regularity in (H1) and the parameter-dependent near-identity normal-form transformation. The Taylor remainders also satisfy ∂ z R = O ( | z | 3 + | μ | | z | + μ 2 ) and ∂ z R 2 = O ( | z | 3 + | μ | | z | + μ 2 ) for the second iterate below. □

3.3. The Second Iterate

The bifurcating period-two orbit is most conveniently obtained from the fixed points of the second iterate.
Lemma 
(Expansion of the second iterate). Under the hypotheses of Proposition 3.1,
f μ 2 ( z ) − z = 2 λ * μ z − 2 ℓ z 3 + R 2 ( z , μ ) ,
where
R 2 ( z , μ ) = O | z | 4 + | μ | | z | 2 + μ 2 | z | .
Proof. 
Write
a μ : = 1 + λ * μ .
Then
f μ ( z ) = − a μ z + ℓ z 3 + R ( z , μ ) .
Ignoring terms of order contained in (3.16), one obtains
f μ 2 ( z ) = − a μ − a μ z + ℓ z 3 + ℓ − a μ z 3 + R 2 ( z , μ ) = a μ 2 z − ℓ a μ ( 1 + a μ 2 ) z 3 + R 2 ( z , μ ) .
Since
a μ 2 = 1 + 2 λ * μ + O ( μ 2 )
and
a μ ( 1 + a μ 2 ) = 2 + O ( μ ) ,
we obtain
f μ 2 ( z ) − z = 2 λ * μ z − 2 ℓ z 3 + R 2 ( z , μ ) .
The stated remainder estimate follows by substituting (3.10) into the composition. □
Equation (3.15) exhibits explicitly the balance responsible for the square-root amplitude law:
λ * μ ∼ ℓ z 2 .

3.4. Existence and Amplitude of the Period-Two Branch

We now obtain the local period-two branch together with its leading amplitude.
Theorem 
(Supercritical period-two branch). Assume(H1)–(H4). There exists μ 0 > 0 such that for every
0 < μ < μ 0 ,
the smooth gradient map G η f + μ possesses a unique local period-two cycle
P 2 ( μ ) = x − ( μ ) , x + ( μ )
in a sufficiently small neighborhood of x * .
The two points satisfy
x ± ( μ ) = x * ± A μ 1 / 2 v * + O ( μ ) , μ → 0 + ,
where
A = λ * ℓ
and therefore
A 2 = 3 λ * 2 Γ .
The labeling may be chosen so that
G η f + μ ( x + ( μ ) ) = x − ( μ ) , G η f + μ ( x − ( μ ) ) = x + ( μ ) .
Proof. 
We first work in the reduced coordinate z.
Set
μ = ε 2 , ε > 0 ,
and seek nonzero solutions of
f ε 2 2 ( z ) − z = 0
in the form
z = ε w .
By Lemma 3.2,
f ε 2 2 ( ε w ) − ε w = 2 λ * ε 3 w − 2 ℓ ε 3 w 3 + O ( ε 4 ) = 2 ε 3 w ( λ * − ℓ w 2 ) + O ( ε ) .
Define the rescaled bifurcation equation
S ( w , ε ) : = f ε 2 2 ( ε w ) − ε w 2 ε 3 , ε > 0 .
Under (H1) the rescaled map extends as a C 3 function to ε = 0 : since V ∈ C ∞ , the gradient map G η is smooth, and the parameter-dependent flip reduction can be chosen with the finite C 6 regularity needed here. Thus the reduced map f μ and the numerator in (3.24) are therefore C 6 in ( ε , w ) and vanishes to third order in ε , so dividing by 2 ε 3 yields a C 3 function. We set
S ( w , 0 ) = w ( λ * − ℓ w 2 ) .
Since ℓ > 0 , the nonzero zeros of S ( · , 0 ) are
w ± 0 = ± λ * ℓ .
Furthermore,
∂ w S ( w ± 0 , 0 ) = λ * − 3 ℓ ( w ± 0 ) 2 = − 2 λ * ≠ 0 .
The implicit function theorem therefore yields two branches
w ± ( ε ) = w ± 0 + O ( ε ) .
Consequently,
z ± ( μ ) = ε w ± ( ε ) = ± λ * ℓ μ + O ( μ ) .
The fixed point z = 0 is locally the only fixed point of f μ near the origin. Indeed,
∂ z ( f 0 ( z ) − z ) | z = 0 = − 2 ≠ 0 ,
so the implicit function theorem excludes nearby nonzero period-one fixed points. Hence the two nonzero solutions z ± ( μ ) of
f μ 2 ( z ) = z
have prime period two.
After the normal-form state change, the physical embedding is x = x * + χ ( z , μ ) , where χ ( z , μ ) = z v * + O ( z 2 + | μ | | z | ) . Thus
x ± ( μ ) = x * + χ ( z ± ( μ ) , μ ) .
Since
χ ( z , μ ) − z v * = O ( z 2 + | μ | | z | )
and
z ± ( μ ) = O ( μ ) ,
we have
χ ( z ± ( μ ) , μ ) − z ± ( μ ) v * = O ( μ ) .
Therefore
x ± ( μ ) = x * ± λ * ℓ μ v * + O ( μ ) ,
which proves (3.19).
Finally, the two nonzero points form a single two-cycle, so after possibly interchanging the labels we obtain (3.22).
For completeness, these branches exhaust the nearby nonzero roots. Divide the second-iterate equation by z and extend smoothly at zero. Its derivative in μ at the origin is 2 λ * > 0 , so it determines μ = ϕ ( z ) , where ϕ ( z ) = ( ℓ / λ * ) z 2 + O ( z 3 ) . The function ϕ is strictly convex near zero and has its unique minimum there; hence each sufficiently small positive μ has exactly two nonzero roots. Every sufficiently small periodic orbit is a full local orbit and belongs to the chosen center manifold. Thus the two roots exhaust the local two-cycle. The direct reduction in Section 4.3 independently gives the same local uniqueness without choosing a center manifold. □

3.5. Stability of the Emerging Cycle

We next verify the local stability of the branch.
Theorem 
(Stability of the supercritical period-two cycle). Under(H1)–(H4), the period-two cycle P 2 ( μ ) given by Theorem 3.3 is locally asymptotically stable for all sufficiently small μ > 0 .
Along the critical center direction, the multiplier of the two-step map satisfies
ρ cyc c ( μ ) = 1 − 4 λ * μ + O ( μ 3 / 2 ) , μ → 0 + .
Proof. 
Differentiate the expansion (3.15):
( f μ 2 ) ′ ( z ) = 1 + 2 λ * μ − 6 ℓ z 2 + O | z | 3 + | μ | | z | + μ 2 .
At either point z = z ± ( μ ) , Theorem 3.3 gives
z ± ( μ ) 2 = λ * ℓ μ + O ( μ 3 / 2 ) .
Hence
( f μ 2 ) ′ ( z ± ( μ ) ) = 1 + 2 λ * μ − 6 λ * μ + O ( μ 3 / 2 ) = 1 − 4 λ * μ + O ( μ 3 / 2 ) .
Thus the center multiplier lies strictly inside the unit disk for all sufficiently small μ > 0 .
At μ = 0 , every transverse multiplier of the two-step map equals
ρ i ( η f ) 2 , i = 1 , … , d − 1 ,
and by (2.28),
| ρ i ( η f ) | < 1 .
Therefore
ρ i ( η f ) 2 < 1 .
By continuity, all transverse multipliers remain strictly inside the unit disk along the sufficiently small period-two branch.
The reduction principle then implies local asymptotic stability of the full period-two cycle. □

3.6. Transfer to the Clipped Map Before Switching

We finally transfer the preceding smooth bifurcation result to the clipped map.
Corollary 
(Pre-contact period-two cycle for the clipped map). Assume(H1)–(H5)and fix c > 0 . Then there exists
μ c > 0
such that for every
0 < μ < μ c ,
the clipped map
F η f + μ , c
possesses the same locally asymptotically stable period-two cycle P 2 ( μ ) as the smooth gradient map G η f + μ .
Moreover,
x ± ( μ ) ∈ Ω c − , 0 < μ < μ c .
Proof. 
For the fixed value c > 0 , let r c > 0 be given by Theorem 2.2. By Theorem 3.3,
x ± ( μ ) → x * as μ → 0 + .
Hence there exists μ c > 0 such that
x ± ( μ ) ∈ B r c ( x * ) ⊂ Ω c −
for every 0 < μ < μ c .
On this ball,
F η f + μ , c = G η f + μ
by Theorem 2.2. Therefore the two maps have exactly the same period-two cycle and the same local derivatives along that cycle. Its local asymptotic stability follows from Theorem 3.4. □
The two endpoint branches and their square-root amplitude are visualized in Figure 1, using the exact planar model of Section 6. The two branches represent one orbit whose points are exchanged, rather than two distinct period-two orbits.

4. Switching Geometry and the Separation Law

We now study how the smooth period-two branch constructed in Section 3 approaches the endogenous switching hypersurface
Σ c = x ∈ U : ∥ ∇ V ( x ) ∥ = c .
A key structural feature of gradient maps appears at this point. For an arbitrary period-two cycle of gradient descent, the gradients at its two points are exactly opposite. Consequently, their gradient norms are identical for every norm ∥ · ∥ .
This identity has two important consequences.
First, the two points of the smooth period-two orbit necessarily reach the clipping boundary simultaneously. Thus, unlike a generic piecewise-smooth map, the present gradient-generated switching problem does not exhibit a splitting of the two first-contact parameters.
Second, the common gradient magnitude along the bifurcating cycle has a particularly simple expansion in the square-root parameter ε = η − η f . This structure will yield the quadratic separation law in Section 4.

4.1. Exact Gradient Balance on a Period-Two Cycle

We begin with an exact identity that does not rely on the local bifurcation expansion. An equivalent balance relation appears in the recent characterization of gradient-descent period-two orbits by Litman [11]. Since this identity is the starting point of the switching analysis below, we include its elementary proof for completeness.
Proposition 
(Exact gradient balance). Let η > 0 , and suppose that { x − , x + } is a period-two cycle of the smooth gradient map
G η ( x ) = x − η ∇ V ( x ) ,
that is,
G η ( x + ) = x − , G η ( x − ) = x + , x + ≠ x − .
Then
∇ V ( x + ) = x + − x − η , ∇ V ( x − ) = − x + − x − η .
In particular,
∇ V ( x + ) = − ∇ V ( x − ) .
Consequently, for every norm ∥ · ∥ ,
∥ ∇ V ( x + ) ∥ = ∥ ∇ V ( x − ) ∥ .
Proof. 
The first identity in (4.1) gives
x − = x + − η ∇ V ( x + ) ,
and hence
∇ V ( x + ) = x + − x − η .
Similarly,
x + = x − − η ∇ V ( x − )
implies
∇ V ( x − ) = x − − x + η = − x + − x − η .
This proves (4.2) and (4.3).
Since every norm is even,
∥ − g ∥ = ∥ g ∥ ,
and therefore
∥ ∇ V ( x − ) ∥ = ∥ − ∇ V ( x + ) ∥ = ∥ ∇ V ( x + ) ∥ .
□

4.2. Midpoint and Half-Difference Coordinates

For the bifurcating period-two cycle
P 2 ( μ ) = { x − ( μ ) , x + ( μ ) } , μ = η − η f > 0 ,
define its midpoint and half-difference by
m ( μ ) : = x + ( μ ) + x − ( μ ) 2 − x * ,
d ( μ ) : = x + ( μ ) − x − ( μ ) 2 .
Thus
x ± ( μ ) = x * + m ( μ ) ± d ( μ ) .
Theorem 3.3 implies
m ( μ ) = O ( μ ) , d ( μ ) = A μ v * + O ( μ ) .
The period-two equations yield considerably more structure.
Lemma 
(Midpoint–difference equations). For every smooth period-two cycle of G η written in the form (4.7),
∇ V ( x * + m + d ) + ∇ V ( x * + m − d ) = 0 ,
∇ V ( x * + m + d ) − ∇ V ( x * + m − d ) = 4 η d .
Proof. 
By Proposition 4.1,
∇ V ( x * + m + d ) = 2 η d
and
∇ V ( x * + m − d ) = − 2 η d .
Adding and subtracting these identities gives (4.9) and (). □

4.3. Square-Root Parametrization and Parity

The following direct reduction justifies both the regularity and the parity; exchanging the two points alone would not establish smoothness in the signed square-root parameter.
The implicit function theorem applied to (4.9), whose midpoint derivative at the origin is 2 H , gives a unique smooth even function m = M ( d ) , with
M ( d ) = − 1 2 H − 1 T ( d , d ) + O ( ∥ d ∥ 2 4 ) .
Insert this function into the difference equation and denote its left-hand side minus 4 d / ( η f + μ ) by R ( d , μ ) . This residual is odd in d. Write d = t v * + w , w ⊥ v * . The perpendicular equation has derivative 2 ( H | v * ⊥ − λ * I ) in w at the origin, which is invertible by (H3). It therefore determines uniquely w = w ( t , μ ) ; this function is odd in t, and w t ( 0 , μ ) = 0 , so w = t 3 W ( t 2 , μ ) with smooth W. In dimension one this perpendicular equation is absent.
Dividing the remaining scalar equation by 2 t and extending smoothly to t = 0 gives, with s = t 2 ,
λ * − 2 η f + μ − Γ 6 s + O ( s 2 + | μ | s ) = 0 .
Here the cubic contribution is 1 6 Q [ v * , v * , v * , v * ] − 1 2 T [ v * , v * , H − 1 g * ] = − Γ / 6 . The derivative in μ at ( s , μ ) = ( 0 , 0 ) is λ * 2 / 2 . Hence the scalar equation determines μ = ψ ( s ) with ψ ′ ( 0 ) = Γ / ( 3 λ * 2 ) > 0 . Its inverse satisfies s ( μ ) = A 2 μ + O ( μ 2 ) . Every sufficiently small nontrivial cycle is captured by this reduction, since the perpendicular solution at t = 0 is zero. This also proves local uniqueness of the unordered cycle.
Set t ( ε ) = ε s ( ε 2 ) / ε 2 , using the positive smooth extension with value A 2 for the quotient at zero. Then t and d are smooth and odd, and M ( d ) is smooth and even.
Introduce the signed square-root parameter
ε 2 = μ , η = η f + ε 2 .
The two-cycle equations are invariant under the transformation
( d , ε ) ⟼ ( − d , − ε ) ,
which merely exchanges the two points of the cycle.
The construction above gives a parametrization with
m ( − ε ) = m ( ε ) , d ( − ε ) = − d ( ε ) .
Hence the midpoint is even and the half-difference is odd in ε .
Proposition 
(Refined two-cycle expansion). Assume(H1)–(H4). The period-two branch can be parametrized by the signed variable ε, with μ = ε 2 , such that
m ( ε ) = ε 2 m 2 + O ( ε 4 ) ,
d ( ε ) = A ε v * + O ( ε 3 ) ,
where
m 2 = − A 2 2 H − 1 g * = − A 2 2 h * .
Consequently,
x ± ( μ ) = x * − A 2 2 μ h * ± A μ v * + O ( μ 3 / 2 ) .
Proof. 
The parity relations (4.12) imply that the local expansions have the form
m ( ε ) = ε 2 m 2 + O ( ε 4 )
and
d ( ε ) = ε d 1 + O ( ε 3 ) .
By Theorem 3.3,
d 1 = A v * .
It remains to determine m 2 . Taylor expansion of the gradient about x * gives
∇ V ( x * + y ) = H y + 1 2 T ( y , y ) + 1 6 Q ( y , y , y ) + O ( ∥ y ∥ 2 4 ) .
Apply (4.17) to y = m + d and y = m − d and add the resulting expressions. Using (4.9), we obtain
2 H m + T ( d , d ) + O ( ε 4 ) = 0 .
Here we used
m = O ( ε 2 ) , d = O ( ε ) ,
and the cancellation of all terms odd in d.
Substituting
m = ε 2 m 2 + O ( ε 4 )
and
d = A ε v * + O ( ε 3 )
into (4.18) gives
2 H m 2 + A 2 T ( v * , v * ) = 0 .
Since H is invertible,
m 2 = − A 2 2 H − 1 T ( v * , v * ) = − A 2 2 h * .
This proves (4.15) and hence (4.16). □

4.4. A Consistency Derivation of the Flip Amplitude

The midpoint–difference equations also provide a direct, coordinate-free derivation of the leading oscillation amplitude.
Proposition 
(Amplitude identity). Under(H1)–(H4), the leading amplitude satisfies
A 2 Γ = 3 λ * 2 .
Equivalently,
A 2 = 3 λ * 2 Γ ,
in agreement with (3.21).
Proof. 
Expand the difference equation () to cubic order.
Using
m = ε 2 m 2 + O ( ε 4 ) , d = A ε v * + O ( ε 3 ) ,
Taylor expansion gives
∇ V ( x * + m + d ) − ∇ V ( x * + m − d ) = 2 H d + 2 T ( m , d ) + 1 3 Q ( d , d , d ) + O ( ε 4 ) .
On the other hand,
4 η = 4 η f + ε 2 = 4 η f − 4 η f 2 ε 2 + O ( ε 4 ) .
Take the Euclidean inner product of () with v * and compare the coefficients of ε 3 .
The unknown O ( ε 3 ) correction in d cancels because
4 η f = 2 λ * .
Thus,
2 A T [ v * , v * , m 2 ] + A 3 3 Q [ v * , v * , v * , v * ] = − 4 A η f 2 .
Since
m 2 = − A 2 2 h *
and
η f = 2 λ * ,
equation (4.22) becomes
− A 3 T [ v * , v * , h * ] + A 3 3 Q [ v * , v * , v * , v * ] = − A λ * 2 .
Dividing by A > 0 and rearranging gives
A 2 3 T [ v * , v * , h * ] − Q [ v * , v * , v * , v * ] = 3 λ * 2 .
By definition of Γ ,
A 2 Γ = 3 λ * 2 .
□

4.5. Gradient Exposure Along the Bifurcating Cycle

Define the common gradient exposure of the smooth period-two cycle by
E ( μ ) : = ∥ ∇ V ( x + ( μ ) ) ∥ = ∥ ∇ V ( x − ( μ ) ) ∥ .
The equality is exact by Proposition 4.1.
Using the half-difference coordinate, one obtains the exact formula
E ( μ ) = 2 η f + μ ∥ d ( μ ) ∥ .
This identity is considerably stronger than a direct Taylor expansion of ∇ V at the two cycle points.
Theorem 
(Asymptotic switching geometry of the two-cycle). Assume(H1)–(H5). As μ → 0 + ,
E ( μ ) = B μ 1 / 2 + O ( μ 3 / 2 ) ,
where
B = λ * A ∥ v * ∥ .
Equivalently,
B 2 = 3 λ * 4 ∥ v * ∥ 2 Γ .
In particular, there is no term of order μ in (4.25).
Proof. 
Set
ε = μ .
By Proposition 4.1,
d ( ε ) = A ε v * + O ( ε 3 ) .
Using positive homogeneity of the norm,
∥ d ( ε ) ∥ = ε A v * + O ( ε 2 ) = A ∥ v * ∥ ε + O ( ε 3 ) .
Furthermore,
2 η f + ε 2 = 2 η f + O ( ε 2 ) = λ * + O ( ε 2 ) .
Substituting (4.28) and (4.29) into (4.24) yields
E ( μ ) = λ * A ∥ v * ∥ ε + O ( ε 3 ) .
Since ε = μ ,
E ( μ ) = B μ 1 / 2 + O ( μ 3 / 2 ) ,
where
B = λ * A ∥ v * ∥ .
Finally, using
A 2 = 3 λ * 2 Γ
gives
B 2 = λ * 2 A 2 ∥ v * ∥ 2 = 3 λ * 4 ∥ v * ∥ 2 Γ .
□

4.6. Simultaneous First Contact

We now record the most important geometric consequence of the exact gradient balance.
Theorem 
(Simultaneous switching contact). Let P 2 ( μ ) = { x − ( μ ) , x + ( μ ) } be the smooth period-two branch from Theorem 3.3. For every sufficiently small μ > 0 ,
x + ( μ ) ∈ Σ c ⟺ x − ( μ ) ∈ Σ c .
Likewise,
x + ( μ ) ∈ Ω c − ⟺ x − ( μ ) ∈ Ω c − ,
x + ( μ ) ∈ Ω c + ⟺ x − ( μ ) ∈ Ω c + .
Hence the two points of a smooth gradient-descent period-two cycle cannot cross a radial clipping boundary sequentially.
Proof. 
By Proposition 4.1,
∥ ∇ V ( x + ( μ ) ) ∥ = ∥ ∇ V ( x − ( μ ) ) ∥ .
Each of the three statements follows immediately by comparing this common value with the clipping threshold c. □
Corollary 
(No pre-contact splitting of the two-cycle). There do not exist two distinct first-contact parameters
η sc + ( c ) ≠ η sc − ( c )
for the two points of the smooth period-two branch. Whenever the first switching contact exists, it is a simultaneous two-point contact.
Proof. 
If one point lies on Σ c , then the other lies on Σ c by Theorem 4.6. □

4.7. Exposure Factorization and Unique Contact

Lemma 
(Exposure factorization and monotonicity). For sufficiently small μ ≥ 0 ,
E ( μ ) = μ β ( μ ) , β ∈ C 1 , β ( 0 ) = B > 0 .
Consequently, E is strictly increasing for small μ > 0 .
Proof. 
The odd C 3 half-difference has the form d ( ε ) = ε q ( ε 2 ) with q ∈ C 1 and q ( 0 ) = A v * . By (4.24), take β ( μ ) = 2 ∥ q ( μ ) ∥ / ( η f + μ ) . The norm is smooth at the nonzero vector q ( 0 ) , so β is C 1 and β ( 0 ) = λ * A ∥ v * ∥ = B . For μ > 0 , E ′ ( μ ) = β ( μ ) / ( 2 μ ) + μ β ′ ( μ ) > 0 near zero, since β is bounded below and β ′ is bounded. □
Define μ sc ( c ) by E ( μ sc ( c ) ) = c and set η sc ( c ) = η f + μ sc ( c ) . These are the unique local first-contact parameters for small c > 0 .

4.8. Quadratic Separation with an Explicit Coefficient

Theorem 
(Flip-to-switching separation law). Under(H1)–(H5), the local first-contact parameter exists and is unique for sufficiently small c > 0 . As c → 0 + ,
η sc ( c ) − η f = K c 2 + O ( c 4 ) , K = 1 B 2 = ℓ λ * 3 ∥ v * ∥ 2 = Γ 3 λ * 4 ∥ v * ∥ 2 > 0 .
Proof. 
Continuity, strict monotonicity and E ( 0 ) = 0 give local existence and uniqueness. At contact the factorization gives c 2 = μ sc β ( μ sc ) 2 . Since β ( 0 ) = B > 0 , first μ sc = O ( c 2 ) and then, using β ( μ ) = B + O ( μ ) , c 2 = B 2 μ sc + O ( c 4 ) . Thus μ sc = c 2 / B 2 + O ( c 4 ) . Substitute B 2 = λ * 2 A 2 ∥ v * ∥ 2 , A 2 = λ * / ℓ and ℓ = Γ / ( 3 λ * ) . □
In an orthonormal Hessian eigenbasis the coefficient is
K = 1 3 λ * 4 ∥ v * ∥ 2 3 ∑ i = 1 d T [ v * , v * , v i ] 2 λ i − Q [ v * , v * , v * , v * ] .
For Euclidean clipping ∥ v * ∥ 2 = 1 , so K 2 = ℓ / λ * 3 = Γ / ( 3 λ * 4 ) . The coefficient combines linear curvature, nonlinear saturation, and the clipping norm in the critical direction. At fixed curvature, a larger ℓ > 0 gives a smaller oscillation and later contact. The exponent two agrees with the generic tangency mechanism [33]; the explicit coefficient, simultaneous contact, and O ( c 4 ) remainder use the gradient structure.

4.9. Dynamical Meaning of the Contact

For 0 < μ < μ sc ( c ) both points lie in Ω c − , so the stable smooth cycle is also a cycle of F η , c . At equality both lie on Σ c ; the clipping operator still equals the identity there, and the cycle equations remain valid. For μ > μ sc ( c ) within the local interval, the smooth reference branch lies in Ω c + and generally ceases to be a cycle of the clipped map. These conclusions follow by comparing the strictly increasing exposure with c. At the boundary the map is generally nondifferentiable; smooth-side stability does not assert stability for all perturbations of the boundary cycle. Figure 2 illustrates the geometry and parameter separation.
Figure 3 shows the corresponding exact boundary and two-cycle positions in the planar model of Section 6. Figure 4 compares its exact contact law with the quadratic asymptotic formula and shows how the common exposure determines the contact.

5. Post-Contact Degeneracy of Clipped Two-Cycles

Section 4 shows that the smooth period-two branch created by the supercritical flip reaches the switching hypersurface simultaneously at both of its points.
The radial clipping structure imposes a further restriction: a genuine period-two orbit cannot have one point in the smooth region and the other in the strictly clipped region. This excludes strict mixed itineraries, but does not assert existence of a continuation or exclude cycles involving boundary points. The results below concern cycles whose two points both lie in Ω c + .
The post-contact results obtained below are not consequences of the generic border-collision theory cited in Section 1. They rely specifically on the radial structure of the clipping operator.
The purpose of this section is to characterize the universal structure of such fully clipped period-two orbits. The main result is that every fully clipped period-two orbit is necessarily nonhyperbolic: its two-step monodromy matrix has an exact multiplier equal to one.
This neutral multiplier is not a consequence of the local flip normal form. It arises directly from the radial normalization inherent in gradient clipping.

5.1. Fully Clipped Period-Two Orbits

We first record the exact algebraic structure of a fully clipped period-two orbit.
Proposition 
(Exact geometry of a fully clipped two-cycle). Let
P 2 = { y − , y + }
be a genuine period-two orbit of
F η , c ( x ) = x − η C c ( ∇ V ( x ) ) ,
and suppose
y − , y + ∈ Ω c + .
Then
∥ y + − y − ∥ = η c .
Moreover, there exist
u ∈ R d , ∥ u ∥ = 1 ,
and numbers
r + , r − > c
such that
y + − y − = η c u ,
∇ V ( y + ) = r + u ,
∇ V ( y − ) = − r − u .
Proof. 
The period-two relations are
y − = y + − η C c ( ∇ V ( y + ) ) ,
y + = y − − η C c ( ∇ V ( y − ) ) .
Therefore
C c ( ∇ V ( y + ) ) = y + − y − η ,
C c ( ∇ V ( y − ) ) = − y + − y − η .
Since both points lie in the strictly clipped region,
C c ( ∇ V ( y ± ) ) = c .
Taking norms in (5.7) gives
∥ y + − y − ∥ = η c .
Define
u = y + − y − η c .
Then
∥ u ∥ = 1 .
Because radial clipping preserves direction,
∇ V ( y + ) ∥ ∇ V ( y + ) ∥ = u
and
∇ V ( y − ) ∥ ∇ V ( y − ) ∥ = − u .
Setting
r ± = ∥ ∇ V ( y ± ) ∥
gives the result. □

5.2. Exclusion of Mixed Period-Two Itineraries

We now establish the restriction imposed by radial clipping.
Proposition 
(No mixed smooth–clipped period-two orbit). A radial clipped gradient map cannot possess a genuine period-two orbit
{ y − , y + }
such that exactly one point lies in Ω c − and the other lies in Ω c + .
More generally, if one point is strictly smooth, the other is strictly smooth as well: their clipped gradient outputs are opposite and have the same norm less than c. Otherwise both points belong to Σ c ∪ Ω c + . A boundary–strictly-clipped pair is not excluded by this argument, and a classical two-step derivative need not exist for such a pair.
Proof. 
Suppose, for contradiction, that
y − ∈ Ω c − , y + ∈ Ω c + .
Then the period-two equations imply
∇ V ( y − ) + C c ( ∇ V ( y + ) ) = 0 .
Taking the clipping norm gives
∥ ∇ V ( y − ) ∥ = C c ( ∇ V ( y + ) ) = c ,
which contradicts
∥ ∇ V ( y − ) ∥ < c .
The opposite mixed itinerary is ruled out identically. For the more general assertion, the opposite clipped outputs have equal norm, and ∥ C c ( g ) ∥ = min { ∥ g ∥ , c } . If one output has norm less than c, so does the other, forcing both original gradients to have norm less than c. □

5.3. Differential Structure in the Clipped Region

We now examine the differential structure of the radial clipping operator in the strictly clipped region. This structure will be the key ingredient in the spectral analysis of fully clipped period-two orbits.
Let
ρ ( g ) : = ∥ g ∥ .
For any nonzero vector g ∈ R d , write
u : = g ρ ( g ) , ρ ( u ) = 1 ,
and define the covector
ϕ u : = D ρ ( u ) .
Since ρ is positively homogeneous of degree one, Euler’s identity gives
ϕ u [ u ] = D ρ ( u ) [ u ] = ρ ( u ) = 1 .
Moreover, the derivative of a positively homogeneous function of degree one is homogeneous of degree zero. Hence, since g = ρ ( g ) u with ρ ( g ) > 0 ,
D ρ ( g ) = D ρ ( u ) = ϕ u .
In the strictly clipped region
ρ ( g ) > c ,
the clipping operator takes the form
C c ( g ) = c g ρ ( g ) .
For any h ∈ R d , differentiation yields
D C c ( g ) [ h ] = c ρ ( g ) h − c ρ ( g ) 2 g D ρ ( g ) [ h ] .
Using
g = ρ ( g ) u
and (5.10), we obtain
D C c ( g ) [ h ] = c ρ ( g ) h − u ϕ u [ h ] .
Equivalently,
D C c ( g ) = c ρ ( g ) I − u ⊗ ϕ u .
The operator
P u : = I − u ⊗ ϕ u
has a simple geometric interpretation. By (5.9),
P u u = u − u ϕ u [ u ] = 0 ,
and
ϕ u P u = ϕ u − ϕ u [ u ] ϕ u = 0 .
Thus P u projects onto
T u { ρ = 1 } = ker ϕ u ,
the tangent space of the unit ρ -sphere at u, along the radial direction spanned by u.
In particular, multiplying (5.11) from the left by ϕ u gives
ϕ u D C c ( g ) = c ρ ( g ) ϕ u − ϕ u [ u ] ϕ u = 0 .
Equation (5.12) expresses the differential degeneracy created by radial clipping: infinitesimal variations in the gradient-magnitude direction are annihilated, whereas only variations tangent to the clipping sphere contribute to the differential of the clipped update.
For every x ∈ Ω c + , the chain rule therefore gives
D F η , c ( x ) = I − η D C c ( ∇ V ( x ) ) D 2 V ( x ) .
Proposition 
(Pointwise neutral covector). Let x ∈ Ω c + and set
u = ∇ V ( x ) ∥ ∇ V ( x ) ∥ .
Then
ϕ u D F η , c ( x ) = ϕ u .
In particular,
1 ∈ σ D F η , c ( x ) .
Proof. 
Using (5.13),
ϕ u D F η , c ( x ) = ϕ u − η ϕ u D C c ( ∇ V ( x ) ) D 2 V ( x ) .
By (5.12),
ϕ u D C c ( ∇ V ( x ) ) = 0 .
Hence the second term vanishes and
ϕ u D F η , c ( x ) = ϕ u .
Thus ϕ u is a left eigenvector of D F η , c ( x ) associated with the eigenvalue 1. Therefore
1 ∈ σ D F η , c ( x ) .
□

5.4. Universal Unit Multiplier for Fully Clipped Two-Cycles

We now obtain the principal post-contact result.
Theorem 
(Universal unit multiplier).
Let
P 2 = { y − , y + }
be any fully clipped period-two orbit of F η , c .
Let
M c : = D F η , c ( y − ) D F η , c ( y + )
be the corresponding two-step monodromy matrix.
Then
1 ∈ σ ( M c ) .
More precisely, if u is the common clipping direction defined by Proposition 5.1, then
ϕ u D F η , c ( y + ) = ϕ u
and
ϕ u D F η , c ( y − ) = ϕ u .
Hence
ϕ u M c = ϕ u .
Proof. 
By Proposition 5.1, there exist r + , r − > c and a vector u with ∥ u ∥ = 1 such that
∇ V ( y + ) = r + u , ∇ V ( y − ) = − r − u .
Since every norm is even,
ρ ( − z ) = ρ ( z ) ,
and therefore
D ρ ( − u ) = − D ρ ( u ) .
Thus
ϕ − u = − ϕ u .
At y + , the normalized gradient direction is u. Hence Proposition 5.3 gives
ϕ u D F η , c ( y + ) = ϕ u .
At y − , the normalized gradient direction is − u . Applying Proposition 5.3 with − u gives
ϕ − u D F η , c ( y − ) = ϕ − u .
Since ϕ − u = − ϕ u , this is equivalent to
ϕ u D F η , c ( y − ) = ϕ u .
Consequently,
ϕ u M c = ϕ u D F η , c ( y − ) D F η , c ( y + ) = ϕ u .
Thus ϕ u is a nonzero left eigenvector of M c associated with the eigenvalue 1. Therefore
1 ∈ σ ( M c ) .
□
Corollary 
(Fully clipped period-two cycles are nonhyperbolic). Every fully clipped period-two orbit of a radial clipped gradient map is nonhyperbolic.
Proof. 
Theorem 5.4 gives a multiplier equal to 1, which lies on the unit circle. □
Corollary 
(No linear exponential attraction). A fully clipped period-two orbit cannot be linearly exponentially asymptotically stable.
Proof. 
Linear exponential asymptotic stability requires every Floquet multiplier of the period-two orbit to lie strictly inside the unit disk. Theorem 5.4 provides an exact multiplier equal to 1. □

5.5. The Spectral Transition at First Contact

Before contact, the smooth period-two orbit is hyperbolically attracting along its critical direction.
By Theorem 3.4,
ρ sm c ( μ ) = 1 − 4 λ * μ + O ( μ 3 / 2 ) .
At the first switching contact,
μ sc ( c ) = K c 2 + O ( c 4 ) .
Hence
ρ sm c μ sc ( c ) = 1 − 4 λ * K c 2 + O ( c 3 ) .
Since
K = Γ 3 λ * 4 ∥ v * ∥ 2 ,
this becomes
ρ sm c μ sc ( c ) = 1 − 4 Γ 3 λ * 3 ∥ v * ∥ 2 c 2 + O ( c 3 ) .
Thus the smooth cycle reaches the switching boundary while still weakly attracting:
ρ sm c < 1 .
Any fully clipped period-two orbit beyond contact, however, satisfies
ρ clip c = 1
in at least one direction.
This yields the following conclusion.
Corollary 
(Loss of hyperbolicity under fully clipped continuation).
Suppose that a period-two orbit admits a continuation from the first switching contact into the fully clipped regime. Then every such fully clipped period-two continuation is necessarily nonhyperbolic.
In the smooth-side limit at contact, the critical multiplier satisfies
ρ sm c = 1 − 4 λ * K c 2 + O ( c 3 ) ,
and therefore
1 − ρ sm c = 4 λ * K c 2 + O ( c 3 ) .
Thus the smooth-side limiting cycle is still weakly attracting. This statement concerns the smooth reference derivative, not a classical derivative of the clipped map on the boundary. By Theorem 5.4, every fully clipped period-two orbit carries an exact unit multiplier.

5.6. Degeneracy of the Period-Two Continuation Equation

A period-two point of the clipped map is a solution of
P ( y , η , c ) = F η , c 2 ( y ) − y = 0 .
If the two-cycle were hyperbolic, one would normally attempt to continue it by applying the implicit function theorem to
P = 0 .
The universal unit multiplier prevents this.
Proposition 
(Singularity of the post-contact continuation problem). Let y belong to a fully clipped period-two orbit. Then
D y P ( y , η , c ) = D F η , c 2 ( y ) − I
is singular.
Consequently, assumptions(H1)–(H5)do not yield a standard implicit-function continuation of the period-two branch through the fully clipped regime.
Proof. 
By Theorem 5.4,
1 ∈ σ D F η , c 2 ( y ) .
Hence
0 ∈ σ D F η , c 2 ( y ) − I ,
so the derivative of the period-two fixed-point equation is singular. □

5.7. Euclidean Clipping

For the Euclidean norm,
∥ · ∥ = ∥ · ∥ 2 ,
the geometry becomes especially transparent.
If
u = ∇ V ( x ) ∥ ∇ V ( x ) ∥ 2 ,
then in the clipped region
D C c ( ∇ V ( x ) ) = c ∥ ∇ V ( x ) ∥ 2 I − u u ⊤ .
Hence
D F η , c ( x ) = I − η c ∥ ∇ V ( x ) ∥ 2 I − u u ⊤ D 2 V ( x ) .
The radial covector satisfies
u ⊤ D F η , c ( x ) = u ⊤ .
For a fully clipped period-two orbit with gradient directions ± u ,
u ⊤ D F η , c ( y − ) D F η , c ( y + ) = u ⊤ .
Thus the unit multiplier has a direct geometric interpretation: Euclidean clipping removes the derivative of the update in the radial gradient-magnitude direction.
An exact unit multiplier does not alone imply nonlinear instability or a family of cycles. It makes D y ( F η , c 2 ( y ) − y ) singular, so the standing hypotheses do not guarantee a standard implicit-function continuation after contact. Additional nonlinear information must determine existence and stability. The planar model below supplies one realization. Also, unit eigenvalues of individual Jacobians do not imply a unit multiplier for arbitrary longer orbits: the common covector argument uses the opposite gradient directions of a two-cycle.

6. A Completely Solvable Planar Gradient Model

We now present a two-dimensional model for which the smooth flip bifurcation, the first clipping contact, the separation coefficient, and the post-contact dynamics can all be computed explicitly.
Consider the potential
V ( x , y ) = λ 2 x 2 + ν 2 y 2 − b 4 x 4 + d 6 x 6 ,
where
λ > ν > 0 , b > 0 , d > 0 , b 2 < 4 d λ .
We use the standard Euclidean clipping norm throughout this section.
The gradient is
∇ V ( x , y ) = λ x − b x 3 + d x 5 ν y .
For convenience, define
g ( x ) : = λ x − b x 3 + d x 5 = x λ − b x 2 + d x 4 .
The condition b 2 < 4 d λ implies
λ − b s + d s 2 > 0 for every s ≥ 0 ,
and hence
x g ( x ) > 0 for every x ≠ 0 .
Thus ( 0 , 0 ) is the unique critical point of V and is a strict global minimum.

6.1. Verification of the Standing Hypotheses

At the origin,
H = D 2 V ( 0 , 0 ) = λ 0 0 ν .
Since
λ > ν > 0 ,
the maximal Hessian eigenvalue is
λ * = λ
and its normalized eigenvector is
v * = e 1 = 1 0 .
The flip threshold is therefore
η f = 2 λ .
Since the potential is even in x,
D 3 V ( 0 , 0 ) = 0 .
Moreover,
D 4 V ( 0 , 0 ) [ e 1 , e 1 , e 1 , e 1 ] = − 6 b .
Hence
Γ = − D 4 V ( 0 , 0 ) [ e 1 4 ] = 6 b > 0 .
Thus the flip is supercritical.
The first flip coefficient is
ℓ = Γ 3 λ = 2 b λ .
The general amplitude formula gives
A 2 = 3 λ 2 Γ = λ 2 2 b ,
or
A = λ 2 b .
Finally, the general Euclidean separation coefficient becomes
K 2 = Γ 3 λ 4 = 2 b λ 4 .
We shall recover all of these quantities below from the exact dynamics.

6.2. Exact Smooth Gradient Dynamics

The ordinary gradient-descent map is
G η ( x , y ) = ( 1 − η λ ) x + η b x 3 − η d x 5 ( 1 − η ν ) y .
The line
L = { ( x , 0 ) : x ∈ R }
is invariant. On L, the dynamics reduce to the odd scalar map
f η ( x ) = ( 1 − η λ ) x + η b x 3 − η d x 5 .
Because f η is odd, a symmetric period-two orbit has the form
{ − a , a } , a > 0 .
The condition
f η ( a ) = − a
is equivalent to
λ − 2 η = b a 2 − d a 4 .
Set
q : = a 2 .
Then
d q 2 − b q + λ − 2 η = 0 .
The small-amplitude root is therefore
q sm ( η ) = b − b 2 − 4 d λ − 2 η 2 d .
The positive small root is defined for 2 / λ < η < 2 / ( λ − b 2 / ( 4 d ) ) . The other quadratic root does not tend to zero at the flip; the uniqueness statements in this paper concern the local branch.
Thus the smooth period-two branch is given explicitly by
P 2 , sm ( η ) = ( − a sm ( η ) , 0 ) , ( a sm ( η ) , 0 ) ,
where
a sm ( η ) = q sm ( η ) .

6.3. Exact Verification of the Square-Root Law

Let
μ = η − 2 λ .
Then
λ − 2 η = λ 2 2 μ − λ 3 4 μ 2 + O ( μ 3 ) .
Solving (6.21) gives
a sm ( η ) 2 = λ 2 2 b μ + λ 3 ( d λ − b 2 ) 4 b 3 μ 2 + O ( μ 3 ) .
Taking the square root yields
a sm ( η ) = λ 2 b μ 1 / 2 1 + λ ( d λ − b 2 ) 4 b 2 μ + O ( μ 2 ) .
Hence
a sm ( η ) = A μ + O ( μ 3 / 2 ) ,
with
A = λ 2 b ,
exactly as predicted by (6.16).
Because the model is reflection-symmetric, the midpoint of the smooth two-cycle remains exactly at the critical point:
( a sm , 0 ) + ( − a sm , 0 ) 2 = ( 0 , 0 ) .

6.4. Stability of the Smooth Period-Two Branch

The derivative of the scalar map is
f η ′ ( x ) = 1 − η λ + 3 η b x 2 − 5 η d x 4 .
Using the cycle equation (6.21), one obtains
f η ′ ( a ) = − 1 + 2 η a 2 b − 2 d a 2 .
Since f η ′ is even, the critical multiplier of the two-step map is
ρ x sm = − 1 + 2 η a 2 b − 2 d a 2 2 .
For μ → 0 + ,
ρ x sm = 1 − 4 λ μ + O ( μ 2 ) ,
in agreement with Theorem 3.4.
The transverse multiplier is
ρ y sm = ( 1 − η ν ) 2 .
Since 0 < ν < λ , at η = η f one has | 1 − 2 ν / λ | < 1 , so the transverse multiplier stays inside the unit disk. Consequently, the smooth period-two cycle is locally asymptotically stable for all sufficiently small η − η f > 0 .

6.5. Exact Gradient Exposure

Figure 5 visualizes one-step exchange, the resulting alternation, and convergence to the smooth cycle. The clipped example in that figure illustrates the exact family constructed below.
At the positive point of the cycle,
∇ V ( a sm , 0 ) = g ( a sm ) 0 .
By the exact period-two relation,
g ( a sm ) = 2 a sm η .
Hence the common Euclidean gradient exposure is
E ( η ) = 2 a sm ( η ) η .
Using (6.30),
E ( μ ) = λ 2 2 b μ 1 / 2 1 + λ ( d λ − 3 b 2 ) 4 b 2 μ + O ( μ 2 ) .
In particular,
B = λ 2 2 b ,
and therefore
1 B 2 = 2 b λ 4 ,
again agreeing with the general separation coefficient (6.17).

6.6. Exact Switching-Contact Equation

The first contact occurs when
2 a sm ( η ) η = c .
Hence at contact
a sm = η c 2 .
Substituting this into (6.21) gives the exact scalar equation
Ψ ( η , c ) : = λ − 2 η − b 4 η 2 c 2 + d 16 η 4 c 4 = 0 .
At
( η , c ) = 2 λ , 0 ,
we have
Ψ 2 λ , 0 = 0
and
∂ η Ψ 2 λ , 0 = λ 2 2 > 0 .
Thus the implicit function theorem gives a unique smooth first-contact curve
η = η sc ( c )
for sufficiently small c.

6.7. Exact Verification of the Separation Law

Write
η sc ( c ) = 2 λ + K 2 c 2 + K 4 c 4 + O ( c 6 ) .
Substituting this expansion into (6.45) and comparing powers of c gives
K 2 = 2 b λ 4 ,
and
K 4 = 6 b 2 λ 7 − 2 d λ 6 .
Therefore
η sc ( c ) = 2 λ + 2 b λ 4 c 2 + 6 b 2 λ 7 − 2 d λ 6 c 4 + O ( c 6 ) .
The leading coefficient
2 b λ 4
coincides exactly with
Γ 3 λ * 4 = 6 b 3 λ 4 ,
as predicted by Theorem 4.9.
This model therefore verifies not only the quadratic exponent but also the absence of a cubic correction.

6.8. The Fully Clipped Symmetric Branch

We now turn to the post-contact regime. Define
a c ( η ) : = η c 2 ,
and consider the pair
Y ± ( η , c ) = ( ± a c ( η ) , 0 ) .
Its gradient magnitude is
g ( a c ) = a c λ − b a c 2 + d a c 4 ,
and a direct calculation gives
g ( a c ) − c = η c 2 Ψ ( η , c ) ,
with Ψ as in (6.45). Hence g ( a c ) = c at η = η sc ( c ) , and since ∂ η Ψ ( η sc ( c ) , c ) > 0 for all sufficiently small c,
g ( a c ) > c for η > η sc ( c )
sufficiently close to the contact value. Both points Y ± therefore lie in the fully clipped region. On the invariant axis, radial clipping acts as
F η , c ( x , 0 ) = ( x − η c , 0 ) ( g ( x ) > c ) , F η , c ( x , 0 ) = ( x + η c , 0 ) ( | g ( x ) | > c , x < 0 ) ,
and since 2 a c = η c ,
F η , c ( a c , 0 ) = ( − a c , 0 ) , F η , c ( − a c , 0 ) = ( a c , 0 ) .
Thus:
Proposition 
(Explicit fully clipped period-two branch). For every sufficiently small c > 0 and every η > η sc ( c ) sufficiently close to η sc ( c ) ,
P 2 , clip ( η , c ) = − η c 2 , 0 , η c 2 , 0
is a fully clipped period-two orbit.
Its amplitude satisfies exactly
2 a c = η c .

6.9. A Continuum of Clipped Period-Two Cycles

The symmetric clipped cycle is not isolated.
For sufficiently small c > 0 and η > η sc ( c ) sufficiently close to contact, we have
g ( a c ) > c .
By continuity, there exists δ > 0 such that
g ( a c + s ) > c , g ( a c − s ) > c
for every
| s | < δ .
Define
Y + ( s ) = ( a c + s , 0 ) ,
Y − ( s ) = ( − a c + s , 0 ) .
After reducing δ if necessary,
a c + s > 0 , − a c + s < 0 .
Since
Y + ( s ) − Y − ( s ) = ( η c , 0 ) ,
radial clipping gives
F η , c ( Y + ( s ) ) = Y − ( s ) ,
F η , c ( Y − ( s ) ) = Y + ( s ) .
We therefore obtain the following stronger result.
Theorem 
(A local family of clipped period-two cycles). For every sufficiently small c > 0 and η > η sc ( c ) sufficiently close to contact, there exists δ > 0 such that
P 2 ( s ) = ( − a c + s , 0 ) , ( a c + s , 0 ) , | s | < δ ,
is a one-parameter family of fully clipped period-two cycles.
Equivalently,
F η , c 2 ( x , 0 ) = ( x , 0 )
on a nontrivial interval of the positive clipped component and on its image under F η , c .
Proof. 
For | s | < δ , both points lie in the clipped regime by (6.59).
At Y + ( s ) , the gradient points in the positive x direction, so the clipped step is exactly
− η c e 1 .
Therefore
F η , c ( Y + ( s ) ) = ( a c + s − η c , 0 ) = ( − a c + s , 0 ) = Y − ( s ) .
At Y − ( s ) , the gradient points in the negative x direction, so the clipped step is
+ η c e 1 .
Thus
F η , c ( Y − ( s ) ) = ( − a c + s + η c , 0 ) = ( a c + s , 0 ) = Y + ( s ) .
□
Remark (Geometric meaning of the unit multiplier)
The continuum in Theorem 6.2 gives a concrete nonlinear interpretation of the universal unit multiplier proved in Theorem 5.4.
The neutral multiplier is not merely an artifact of linearization: in this model the two-step map is exactly the identity along the one-dimensional clipped family.

6.10. Exact Post-Contact Multipliers

Consider the period-two cycle
P 2 ( s ) = { Y − ( s ) , Y + ( s ) } .
Set
r + ( s ) = g ( a c + s ) , r − ( s ) = g ( a c − s ) .
At either point, the Euclidean clipping direction is parallel to e 1 . Therefore
I − e 1 e 1 ⊤ = 0 0 0 1 .
Since
D 2 V ( x , 0 ) = λ − 3 b x 2 + 5 d x 4 0 0 ν ,
the one-step Jacobians along the clipped cycle are
D F η , c ( Y + ( s ) ) = 1 0 0 1 − η ν c r + ( s ) ,
D F η , c ( Y − ( s ) ) = 1 0 0 1 − η ν c r − ( s ) .
Hence the two-step monodromy matrix is
M clip ( s ) = 1 0 0 1 − η ν c r + ( s ) 1 − η ν c r − ( s ) .
Thus the two multipliers are
ρ x clip = 1
and
ρ y clip ( s ) = 1 − η ν c r + ( s ) 1 − η ν c r − ( s ) .
For the symmetric member s = 0 ,
ρ y clip ( 0 ) = 1 − η ν c g ( a c ) 2 .
At the switching contact,
g ( a c ) = c ,
so the outer limiting transverse multiplier becomes
ρ y clip = ( 1 − η sc ν ) 2 .
Since
η sc ( c ) → 2 λ
as c → 0 + and
0 < ν < λ ,
one has
1 − η sc ( c ) ν < 1
for all sufficiently small c.
By continuity, after reducing the post-contact parameter interval and the family interval | s | < δ , both transverse one-step multipliers have modulus less than one. Hence the local clipped family is transversely attracting while remaining exactly neutral along its one-dimensional period-two direction.

6.11. The Local Transition in the Solvable Model

Figure 6 displays the continuum of clipped two-cycles, the identity of the two-step map along this family, and the exact longitudinal and transverse multipliers.
The planar model provides an explicit realization of the entire smooth-to-clipped scenario. For 0 < η < η f = 2 / λ the origin is locally asymptotically stable; for η f < η < η sc ( c ) there is a unique small stable smooth two-cycle { ( − a sm , 0 ) , ( a sm , 0 ) } whose amplitude grows as
a sm ∼ λ 2 b η − η f .
At η = η sc ( c ) both points reach the clipping boundary simultaneously with
a sm = η sc c 2 , η sc ( c ) = 2 λ + 2 b λ 4 c 2 + O ( c 4 ) .
For η > η sc ( c ) sufficiently close to contact, the clipped map possesses a one-parameter family of period-two cycles with exact step constraint x + − x − = η c , and the smooth critical multiplier ρ x sm < 1 is replaced by the exact neutral multiplier ρ x clip = 1 . The model thus exhibits the transition from a stable smooth period-two branch, through a simultaneous clipping contact, to a neutral family of clipped two-cycles.

6.12. A Concrete Parameter Choice

To make the preceding formulas explicit, consider the parameter choice
λ = 2 , ν = 1 2 , b = 1 , d = 1 .
Then
η f = 1 , Γ = 6 , ℓ = 1 ,
and therefore
A = 2 , B = 2 2 , K 2 = 1 8 .
Hence the smooth period-two branch satisfies x ± ( μ ) = x * ± 2 μ v * + O ( μ ) , while the first switching threshold obeys
η sc ( c ) = 1 + 1 8 c 2 + O ( c 4 ) ,
with the next correction computable explicitly:
η sc ( c ) = 1 + 1 8 c 2 + 1 64 c 4 + O ( c 6 ) .
The example thus provides an explicit realization of the general quadratic separation law of Section 4.
Figure 7. Exact planar-model formulas for λ = 2 , ν = 0.5 , b = d = 1 and c = 0.3 , satisfying b 2 < 4 d λ . Here r is the small positive root of g ( r ) = c , and η sc = 2 r / c . (a) The smooth symmetric cycle has amplitude a 2 = ( b − b 2 − 4 d ( λ − 2 / η ) ) / ( 2 d ) . After contact, the shaded open interval r < p < η c − r parametrizes clipped cycles { p , p − η c } on the invariant axis; its boundaries are excluded. The solid orange line selects p = η c / 2 . (b) The smooth longitudinal multiplier ( 1 − η g ′ ( a ) ) 2 remains below one up to contact, while strictly clipped cycles have multiplier one. The open marker excludes a classical clipped derivative at the boundary. Curves are evaluations of analytic formulas, not trajectory simulations.
Figure 7. Exact planar-model formulas for λ = 2 , ν = 0.5 , b = d = 1 and c = 0.3 , satisfying b 2 < 4 d λ . Here r is the small positive root of g ( r ) = c , and η sc = 2 r / c . (a) The smooth symmetric cycle has amplitude a 2 = ( b − b 2 − 4 d ( λ − 2 / η ) ) / ( 2 d ) . After contact, the shaded open interval r < p < η c − r parametrizes clipped cycles { p , p − η c } on the invariant axis; its boundaries are excluded. The solid orange line selects p = η c / 2 . (b) The smooth longitudinal multiplier ( 1 − η g ′ ( a ) ) 2 remains below one up to contact, while strictly clipped cycles have multiplier one. The open marker excludes a classical clipped derivative at the boundary. Curves are evaluations of analytic formulas, not trajectory simulations.
Preprints 237619 g007

7. Conclusion and Discussion

We studied the first interaction between a smooth flip bifurcation and the endogenous switching geometry generated by radial gradient clipping. The mechanism operates on two sharply separated scales: clipping is locally invisible near the critical point and becomes dynamically relevant only at finite amplitude.

7.0.0.1. Local scale.

For every fixed c > 0 the clipped map F η , c coincides with the ordinary gradient map G η on a neighborhood of a nondegenerate critical point; hence the fixed point, its linearization, its local invariant structures, its center-manifold reduction, and its smooth flip coefficient are independent of the clipping threshold. In particular, for a strict local minimum x * with simple maximal Hessian eigenvalue λ * , the first stability threshold remains
η f = 2 λ * .
Under the supercriticality condition
Γ = 3 T [ v * , v * , H − 1 g * ] − Q [ v * 4 ] > 0 ,
the loss of stability creates a stable smooth period-two branch
x ± ( μ ) = x * ± A μ v * + O ( μ ) , μ = η − η f , A 2 = 3 λ * 2 Γ .

7.0.0.2. Finite-amplitude scale.

Clipping becomes active only when this orbit reaches the switching hypersurface
Σ c = { x : ∥ ∇ V ( x ) ∥ = c } .
The gradient structure makes the interaction rigid: every smooth period-two orbit satisfies the exact identity
∇ V ( x + ) = − ∇ V ( x − ) ,
so both points have identical clipping exposure and reach Σ c simultaneously. Matching the square-root amplitude scale with the threshold produces the principal asymptotic law
η sc ( c ) − η f = Γ 3 λ * 4 ∥ v * ∥ 2 c 2 + O ( c 4 ) ,
which for Euclidean clipping becomes
η sc ( c ) = 2 λ * + Γ 3 λ * 4 c 2 + O ( c 4 ) .
The even-power expansion reflects an additional symmetry of the gradient problem: in the signed square-root parameter ε = η − η f the half-difference of the two-cycle is odd while its midpoint is even, which together with (7.4) removes the cubic correction from the contact curve.

7.0.0.3. Post-contact scale.

A clipped period-two orbit cannot mix strictly smooth and strictly clipped points. This restriction alone does not guarantee a post-contact continuation or exclude boundary itineraries. Every strictly fully clipped period-two orbit obeys the exact step-length constraint
∥ y + − y − ∥ = η c .
Most importantly, radial clipping creates an unavoidable neutral direction: the monodromy matrix
M c = D F η , c ( y − ) D F η , c ( y + )
of a fully clipped period-two orbit satisfies
1 ∈ σ ( M c ) .
Every fully clipped two-cycle is therefore nonhyperbolic: the radial normalization removes first-order information in the gradient-magnitude direction. The solvable planar model of Section 6 shows that this unit multiplier is not a formal artifact — it corresponds to an actual loss of isolation of the periodic orbit.

7.1. Dynamical Interpretation

The results separate two effects that are often conflated. The first is the loss of local stability of the minimum: a smooth gradient-descent phenomenon governed by D 2 V ( x * ) , D 3 V ( x * ) , D 4 V ( x * ) , and unaffected by clipping. The second is the modification of an already finite-amplitude trajectory once it reaches ∥ ∇ V ∥ = c . Clipping does not postpone the first local flip threshold; it introduces a second dynamical scale beyond it, the switching-contact scale η sc ( c ) − η f = O ( c 2 ) . This distinction is useful when interpreting large-step optimization dynamics: linear stability determines when the critical point loses stability, while the clipping threshold determines when the resulting nonlinear oscillation begins to experience a different piecewise-smooth evolution.

7.2. Limitations

The theory focuses on an isolated nondegenerate minimum with a simple maximal Hessian eigenvalue, which isolates a one-dimensional flip mode but excludes the additional neutral directions that arise when the set of minimizers forms a manifold. Second, the post-contact unit multiplier prevents a universal hyperbolic continuation theorem: under the present assumptions alone, the existence, uniqueness, and nonlinear stability of a fully clipped continuation are not determined. Third, the paper treats radial clipping; coordinatewise clipping, adaptive normalization, momentum, and stochastic mini-batch updates generate different switching geometries. Finally, the analysis is local in the bifurcation parameter and does not classify the remote global dynamics after the orbit enters the clipped region.

7.3. Directions for Further Work

Several extensions appear natural: reducing the nonhyperbolic post-contact dynamics to a higher-order scalar or center-manifold normal form, which the exact unit multiplier suggests may form a new class of degenerate border interactions specific to radial gradient maps; replacing the isolated minimum by a Morse–Bott manifold of minima, combining the − 1 flip direction with + 1 tangent directions; degenerate smooth flips with Γ = 0 , where a matching principle r ( μ ) ≍ μ γ , d ( c ) ≍ c p suggests the general contact scale μ sc ≍ c p / γ ; and determining which structures persist for momentum, stochastic gradients, and nonradial clipping.

7.4. Final Perspective

Near a nondegenerate minimum, radial gradient clipping is locally invisible: F η , c = G η , so the smooth map determines when the first oscillatory instability is born. Once the oscillation reaches finite amplitude, the same potential generates an endogenous switching boundary at the explicit scale
η sc ( c ) − η f ∼ Γ 3 λ * 4 ∥ v * ∥ 2 c 2 .
At that point the dynamics cease to be governed solely by the smooth flip normal form: radial clipping removes the gradient-magnitude degree of freedom and forces any strictly fully clipped period-two cycle to carry an exact neutral multiplier. The clipped gradient map thus exhibits a sharp transition from smooth bifurcation theory to intrinsically nonhyperbolic switching dynamics, with the transition scale determined explicitly by the local geometry of the underlying potential.

Author Contributions

Both authors contributed to the conception, analysis, and writing of the manuscript.

Funding

The authors declare that no specific funding was received for this work.

Data Availability Statement

Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.

Acknowledgments

During the preparation of this work, the authors used the DeepSeek V4 Harness (deepseek-v4-flash model) for typesetting, compilation, and manuscript-preparation assistance, and OpenAI ChatGPT (GPT-5.6 Sol) for mathematical discussion, language editing, and manuscript organization. All mathematical arguments, references, and conclusions were independently checked by the authors, who take full responsibility for the manuscript.

Conflicts of Interest

The authors declare that they have no conflict of interest.

References

  1. Robbins, H.; Monro, S. A Stochastic Approximation Method. Ann. Math. Stat. 1951, 22, 400–407. [Google Scholar] [CrossRef]
  2. Polyak, B.T. Introduction to Optimization. In Translations Series in Mathematics and Engineering; Optimization Software: New York, 1987. [Google Scholar]
  3. Nesterov, Y. Lectures on Convex Optimization. In Springer Optimization and Its Applications, 2 ed.; Springer: Cham, 2018; Vol. 137. [Google Scholar]
  4. Bottou, L.; Curtis, F.E.; Nocedal, J. Optimization Methods for Large-Scale Machine Learning. SIAM Rev. 2018, 60, 223–311. [Google Scholar] [CrossRef]
  5. Kuznetsov, Y.A. Elements of Applied Bifurcation Theory. In Applied Mathematical Sciences, 3 ed.; Springer: New York, 2004; Vol. 112. [Google Scholar]
  6. Guckenheimer, J.; Holmes, P. Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields. In Applied Mathematical Sciences; Springer: New York, 1983; Vol. 42. [Google Scholar]
  7. Hale, J.K.; Koçak, H. Dynamics and Bifurcations. In Texts in Applied Mathematics; Springer: New York, 1991; Vol. 3. [Google Scholar]
  8. Cohen, J.M.; Kaur, S.; Li, Y.; Kolter, J.Z.; Talwalkar, A. Gradient Descent on Neural Networks Typically Occurs at the Edge of Stability. In Proceedings of the International Conference on Learning Representations, 2021. [Google Scholar]
  9. Mulayoff, R.; Stich, S.U. On the Stability of Nonlinear Dynamics in GD and SGD: Beyond Quadratic Potentials. Proc. Proc. 39th Annu. Conf. Learn. Theory. PMLR, Preliminary version; 2026, arXiv:2602.14789Vol. 336, Proceedings of Machine Learning Research, 5210–5243. [Google Scholar]
  10. Gan, E. A Bifurcation Theory Framework for Gradient Descent on the Edge of Stability. arXiv 2026, arXiv:2606.15551. [Google Scholar]
  11. Litman, E. The Origin of Edge of Stability. arXiv 2026, arXiv:2604.20446. [Google Scholar]
  12. Pascanu, R.; Mikolov, T.; Bengio, Y. On the Difficulty of Training Recurrent Neural Networks. In Proceedings of the Proceedings of the 30th International Conference on Machine Learning, 2013; Proceedings of Machine Learning Research; Vol. 28. [Google Scholar]
  13. Zhang, J.; He, T.; Sra, S.; Jadbabaie, A. Why Gradient Clipping Accelerates Training: A Theoretical Justification for Adaptivity. In Proceedings of the International Conference on Learning Representations, 2020. [Google Scholar]
  14. Koloskova, A.; Hendrikx, H.; Stich, S.U. Revisiting Gradient Clipping: Stochastic Bias and Tight Convergence Guarantees. Proc. Proc. 40th Int. Conf. Mach. Learn. PMLR 2023, Vol. 202, 17343–17363. [Google Scholar]
  15. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, 2016. [Google Scholar]
  16. Chow, S.N.; Hale, J.K. Methods of Bifurcation Theory. In Grundlehren der mathematischen Wissenschaften; Springer: New York, 1982; Vol. 251. [Google Scholar]
  17. Wiggins, S. Introduction to Applied Nonlinear Dynamical Systems and Chaos. In Texts in Applied Mathematics, 2 ed.; Springer: New York, 2003; Vol. 2. [Google Scholar]
  18. Arnold, V.I. Geometrical Methods in the Theory of Ordinary Differential Equations. In Grundlehren der mathematischen Wissenschaften; Springer: New York, 1983; Vol. 250. [Google Scholar]
  19. Sotomayor, J. Generic Bifurcations of Dynamical Systems. In Dynamical Systems; Peixoto, M.M., Ed.; Academic Press: New York, 1973; pp. 561–582. [Google Scholar]
  20. Palis, J.; de Melo, W. Geometric Theory of Dynamical Systems: An Introduction; Springer: New York, 1982. [Google Scholar]
  21. Nusse, H.E.; Yorke, J.A. Border-Collision Bifurcations Including “Period Two to Period Three” for Piecewise Smooth Systems. Physica D. Nonlinear Phenom. 1992, 57, 39–57. [Google Scholar] [CrossRef]
  22. Nusse, H.E.; Yorke, J.A. Border-Collision Bifurcations for Piecewise Smooth One-Dimensional Maps. Int. J. Bifurc. Chaos 1995, 5, 189–207. [Google Scholar] [CrossRef]
  23. Banerjee, S.; Grebogi, C. Border Collision Bifurcations in Two-Dimensional Piecewise Smooth Maps. Phys. Rev. E 1999, 59, 4052–4061. [Google Scholar] [CrossRef]
  24. Feigin, M.I. Doubling of the Oscillation Period with C-Bifurcations in Piecewise Continuous Systems. J. Appl. Math. Mech. (PMM) 1970, 34, 861–869. [Google Scholar] [CrossRef]
  25. di Bernardo, M.; Budd, C.J.; Champneys, A.R.; Kowalczyk, P. Piecewise-Smooth Dynamical Systems: Theory and Applications. In Applied Mathematical Sciences; Springer: London, 2008; Vol. 163. [Google Scholar] [CrossRef]
  26. Simpson, D.J.W. Bifurcations in Piecewise-Smooth Continuous Systems; World Scientific: Singapore, 2010. [Google Scholar]
  27. Simpson, D.J.W. Border-Collision Bifurcations in RN. SIAM Rev. 2016, 58, 177–226. [Google Scholar] [CrossRef]
  28. Filippov, A.F. Differential Equations with Discontinuous Righthand Sides. In Mathematics and its Applications; Kluwer Academic Publishers: Dordrecht, 1988; Vol. 18. [Google Scholar]
  29. di Bernardo, M.; Budd, C.J.; Champneys, A.R.; Kowalczyk, P.; Nordmark, A.B.; Olivar Tost, G.; Piiroinen, P.T. Bifurcations in Nonsmooth Dynamical Systems. SIAM Rev. 2008, 50, 629–701. [Google Scholar] [CrossRef]
  30. di Bernardo, M.; Hogan, S.J. Discontinuity-Induced Bifurcations of Piecewise Smooth Dynamical Systems. Philos. Trans. R. Soc. A 2010, 368, 4915–4935. [Google Scholar] [CrossRef]
  31. Leine, R.I.; Nijmeijer, H. Dynamics and Bifurcations of Non-Smooth Mechanical Systems. In Lecture Notes in Control and Information Sciences; Springer: Berlin, 2004; Vol. 392. [Google Scholar]
  32. Simpson, D.J.W.; Meiss, J.D. Neimark–Sacker Bifurcations in Planar, Piecewise-Smooth, Continuous Maps. SIAM J. Appl. Dyn. Syst. 2008, 7, 795–824. [Google Scholar] [CrossRef]
  33. Simpson, D.J.W.; Meiss, J.D. Simultaneous Border-Collision and Period-Doubling Bifurcations. Chaos 2009, 19, 033146. [Google Scholar] [CrossRef]
  34. Simpson, D.J.W.; Meiss, J.D. Shrinking Point Bifurcations of Resonance Tongues for Piecewise-Smooth, Continuous Maps. Nonlinearity 2009, 22, 1123–1144. [Google Scholar] [CrossRef]
  35. Simpson, D.J.W.; Meiss, J.D. Resonance Near Border-Collision Bifurcations in Piecewise-Smooth, Continuous Maps. Nonlinearity 2010, 23, 3091–3118. [Google Scholar] [CrossRef]
  36. Simpson, D.J.W.; Meiss, J.D. Aspects of Bifurcation Theory for Piecewise-Smooth, Continuous Systems. Physica D. Nonlinear Phenom. 2012, 241, 1861–1868. [Google Scholar] [CrossRef]
  37. Carr, J. Applications of Centre Manifold Theory. In Applied Mathematical Sciences; Springer: New York, 1981; Vol. 35. [Google Scholar]
  38. Hirsch, M.W.; Pugh, C.C.; Shub, M. Invariant Manifolds. In Lecture Notes in Mathematics; Springer: Berlin, 1977; Vol. 583. [Google Scholar]
Figure 1. Exact bifurcation branches of the planar model with λ = 2 , ν = 1 / 2 , b = d = 1 , and c = 0.3 . (a) The upper and lower branches are the two endpoints of a single period-two orbit, exchanged by one iteration; they are not two distinct cycles. Solid green curves are the attracting smooth cycle before η sc ≈ 1.01137028 . Dotted gray curves continue only the smooth reference map. Solid orange curves select the symmetric member of the clipped family; they are nonhyperbolic. (b) The exact smooth amplitude agrees to leading order with A μ , where μ = η − 1 and A = 2 .
Figure 1. Exact bifurcation branches of the planar model with λ = 2 , ν = 1 / 2 , b = d = 1 , and c = 0.3 . (a) The upper and lower branches are the two endpoints of a single period-two orbit, exchanged by one iteration; they are not two distinct cycles. Solid green curves are the attracting smooth cycle before η sc ≈ 1.01137028 . Dotted gray curves continue only the smooth reference map. Solid orange curves select the symmetric member of the clipped family; they are nonhyperbolic. (b) The exact smooth amplitude agrees to leading order with A μ , where μ = η − 1 and A = 2 .
Preprints 237619 g001
Figure 2. Conceptual illustration of the flip-to-switching mechanism. (a) The ellipse represents the leading Euclidean approximation ∥ H ( x − x * ) ∥ 2 = c ; the two-cycle approaches opposite boundary points together. Higher-order midpoint drift is suppressed in this schematic. (b) The contact locus has quadratic separation from the flip locus. The scales are normalized and the plotted parabola represents the leading law, not a numerical continuation. The region above contact does not assert existence of clipped two-cycles.
Figure 2. Conceptual illustration of the flip-to-switching mechanism. (a) The ellipse represents the leading Euclidean approximation ∥ H ( x − x * ) ∥ 2 = c ; the two-cycle approaches opposite boundary points together. Higher-order midpoint drift is suppressed in this schematic. (b) The contact locus has quadratic separation from the flip locus. The scales are normalized and the plotted parabola represents the leading law, not a numerical continuation. The region above contact does not assert existence of clipped two-cycles.
Preprints 237619 g002
Figure 3. Exact switching boundary in the planar model, g ( x ) 2 + ν 2 y 2 = c 2 , for the parameters of Figure 1. The horizontal and vertical axes have different scales. (a) The smooth cycle at η = 1.006 lies inside the boundary. (b) At η sc , both endpoints ( ± r , 0 ) reach the boundary simultaneously. (c) At η = 1.06 , the symmetric clipped cycle lies outside. Curved arrows denote discrete one-step exchange; they are not trajectory paths in the intervening region. Unlike Figure 2, the boundary and endpoints here are computed from exact model formulas.
Figure 3. Exact switching boundary in the planar model, g ( x ) 2 + ν 2 y 2 = c 2 , for the parameters of Figure 1. The horizontal and vertical axes have different scales. (a) The smooth cycle at η = 1.006 lies inside the boundary. (b) At η sc , both endpoints ( ± r , 0 ) reach the boundary simultaneously. (c) At η = 1.06 , the symmetric clipped cycle lies outside. Curved arrows denote discrete one-step exchange; they are not trajectory paths in the intervening region. Unlike Figure 2, the boundary and endpoints here are computed from exact model formulas.
Preprints 237619 g003
Figure 4. Quantitative contact law in the solvable model. (a) The exact separation 2 r ( c ) / c − 1 , with g ( r ( c ) ) = c , is compared with K c 2 = c 2 / 8 for 0.02 ≤ c ≤ 0.4 . (b) At fixed c = 0.3 , the two smooth cycle points have the same exposure E ( μ ) = g ( a ( 1 + μ ) ) . Its intersection with E = c determines the unique simultaneous contact. All curves use analytic formulas and scalar root solving, without fitted coefficients.
Figure 4. Quantitative contact law in the solvable model. (a) The exact separation 2 r ( c ) / c − 1 , with g ( r ( c ) ) = c , is compared with K c 2 = c 2 / 8 for 0.02 ≤ c ≤ 0.4 . (b) At fixed c = 0.3 , the two smooth cycle points have the same exposure E ( μ ) = g ( a ( 1 + μ ) ) . Its intersection with E = c determines the unique simultaneous contact. All curves use analytic formulas and scalar root solving, without fitted coefficients.
Preprints 237619 g004
Figure 5. Iterates of the actual clipped scalar map on the invariant axis, f η , c ( x ) = x − η clip ( g ( x ) , − c , c ) , with the parameters of Figure 1. (a,b) Cobweb rectangles show the one-step exchange of the two endpoints for a smooth cycle at η = 1.006 and an asymmetric clipped cycle at η = 1.06 . (c) Both trajectories alternate between two values; the smooth trajectory uses 1752 burn-in steps, while the clipped trajectory starts at x 0 = 0.16 and is exactly periodic. (d) From x 0 = 0.04 , | x n | decays to zero below the flip ( η = 0.96 ) and approaches the smooth cycle amplitude above it. Connecting segments in (c) show iteration order, not a continuous-time flow.
Figure 5. Iterates of the actual clipped scalar map on the invariant axis, f η , c ( x ) = x − η clip ( g ( x ) , − c , c ) , with the parameters of Figure 1. (a,b) Cobweb rectangles show the one-step exchange of the two endpoints for a smooth cycle at η = 1.006 and an asymmetric clipped cycle at η = 1.06 . (c) Both trajectories alternate between two values; the smooth trajectory uses 1752 burn-in steps, while the clipped trajectory starts at x 0 = 0.16 and is exactly periodic. (d) From x 0 = 0.04 , | x n | decays to zero below the flip ( η = 0.96 ) and approaches the smooth cycle amplitude above it. Connecting segments in (c) show iteration order, not a continuous-time flow.
Preprints 237619 g005
Figure 6. Fully clipped family for η = 1.06 and c = 0.3 . Write a c = η c / 2 and | s | < a c − r ; the interval is open. (a) Each value of s gives one cycle with endpoints ( a c + s , 0 ) and ( − a c + s , 0 ) . (b) The two-step map is exactly the identity on the positive-endpoint interval. (c) The longitudinal multiplier is one throughout the family, while the exact transverse multiplier ( 1 − η ν c / g ( a c + s ) ) ( 1 − η ν c / g ( a c − s ) ) lies inside the unit disk for these parameters. Neutrality along the family therefore coexists with transverse contraction; no individual member is linearly exponentially attracting.
Figure 6. Fully clipped family for η = 1.06 and c = 0.3 . Write a c = η c / 2 and | s | < a c − r ; the interval is open. (a) Each value of s gives one cycle with endpoints ( a c + s , 0 ) and ( − a c + s , 0 ) . (b) The two-step map is exactly the identity on the positive-endpoint interval. (c) The longitudinal multiplier is one throughout the family, while the exact transverse multiplier ( 1 − η ν c / g ( a c + s ) ) ( 1 − η ν c / g ( a c − s ) ) lies inside the unit disk for these parameters. Neutrality along the family therefore coexists with transverse contraction; no individual member is linearly exponentially attracting.
Preprints 237619 g006
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.