Introduction and main result
The prescribed minimal period problem asks whether every T > 0 T>0 T > 0 occurs as the least positive period of a nonconstant solution of an autonomous Hamiltonian system.
Consider the following second-order mechanical system
q ¨ ( t ) + ∇ V ( q ( t ) ) = 0 , q ( t ) ∈ R n , (1) \ddot{q}(t)+\nabla V(q(t))=0,\qquad q(t)\in\mathbb{R}^n, \tag*{(1)} q ¨ ( t ) + ∇ V ( q ( t )) = 0 , q ( t ) ∈ R n , (1)
where q ˙ ( t ) = d d t q ( t ) \dot{q}(t)=\frac{d}{dt}q(t) q ˙ ( t ) = d t d q ( t ) , V : R n → R V:\mathbb{R}^n\to\mathbb{R} V : R n → R is the potential energy, and ∇ V ( q ) \nabla V(q) ∇ V ( q ) denotes the gradient of V V V with respect to q ∈ R n q\in\mathbb{R}^n q ∈ R n .
Throughout this paper, the following assumptions about the potential V V V will be used:
(V1) V ∈ C 1 ( R n , R ) V\in C^1(\mathbb{R}^n,\mathbb{R}) V ∈ C 1 ( R n , R ) ,
(V1′) V ∈ C 2 ( R n , R ) V\in C^2(\mathbb{R}^n,\mathbb{R}) V ∈ C 2 ( R n , R ) ,
(V2) there exist μ > 2 \mu>2 μ > 2 , r 0 > 0 r_0>0 r 0 > 0 such that 0 < μ V ( q ) ≤ ⟨ ∇ V ( q ) , q ⟩ 0<\mu V(q)\le\langle\nabla V(q),q\rangle 0 < μ V ( q ) ≤ ⟨ ∇ V ( q ) , q ⟩ for ∣ q ∣ ≥ r 0 |q|\ge r_0 ∣ q ∣ ≥ r 0 ,
(V3) V ( q ) = o ( ∣ q ∣ 2 ) V(q)=o(|q|^2) V ( q ) = o ( ∣ q ∣ 2 ) as q → 0 q\to0 q → 0 , i.e. lim ∣ q ∣ → 0 V ( q ) ∣ q ∣ 2 = 0 \lim_{|q|\to0}\frac{V(q)}{|q|^2}=0 lim ∣ q ∣ → 0 ∣ q ∣ 2 V ( q ) = 0 ,
(V4) V ( q ) ≥ 0 V(q)\ge0 V ( q ) ≥ 0 for every q ∈ R n q\in\mathbb{R}^n q ∈ R n ,
(V5) V ( − q ) = V ( q ) V(-q)=V(q) V ( − q ) = V ( q ) for every q ∈ R n q\in\mathbb{R}^n q ∈ R n .
In 1978, [10 ] proved the existence of nonconstant T T T -periodic solutions if the potential satisfies (V1)–(V4). Note that a T T T -periodic solution can also be an iterate of an orbit with a shorter period, so this existence statement alone does not determine its minimal period. Rabinowitz also asked whether, under (V1)–(V4), the system admits a nonconstant solution with any prescribed minimal period.
The system (1) can be written as
( p ˙ ( t ) q ˙ ( t ) ) = ( − ∇ V ( q ( t ) ) p ( t ) ) = J ∇ H ( x ) , \begin{pmatrix} \dot{p}(t) \\ \dot{q}(t) \end{pmatrix} = \begin{pmatrix} -\nabla V(q(t)) \\ p(t) \end{pmatrix} = J\nabla H(x), ( p ˙ ( t ) q ˙ ( t ) ) = ( − ∇ V ( q ( t )) p ( t ) ) = J ∇ H ( x ) ,
where
J = ( 0 − I I 0 ) , x = ( p , q ) ∈ R 2 n , H ( p , q ) = 1 2 ∣ p ∣ 2 + V ( q ) , J=\begin{pmatrix} 0 & -I \\ I & 0 \end{pmatrix}, \qquad x=(p,q)\in\mathbb{R}^{2n}, \qquad H(p,q)=\frac{1}{2}|p|^2+V(q), J = ( 0 I − I 0 ) , x = ( p , q ) ∈ R 2 n , H ( p , q ) = 2 1 ∣ p ∣ 2 + V ( q ) ,
and I I I is the n × n n\times n n × n identity matrix. Here ∇ H ( x ) = ( p , ∇ V ( q ) ) \nabla H(x)=(p,\nabla V(q)) ∇ H ( x ) = ( p , ∇ V ( q )) . Thus (1) is equivalent to the first-order Hamiltonian system
x ˙ = J ∇ H ( x ) , x ∈ R 2 n . (2) \dot{x}=J\nabla H(x), \qquad x\in\mathbb{R}^{2n}. \tag*{(2)} x ˙ = J ∇ H ( x ) , x ∈ R 2 n . (2)
For convex first-order Hamiltonian systems, Ekeland and Hofer [3 ] obtained prescribed minimal-period solutions by relating the variational characterization of a critical point to its index and local geometry. For the second-order problem with the potential V V V satisfying (V1′) and (V2)–(V5), Long [8 ] worked in a space of time-even, half-antiperiodic loops and proved that a suitable symmetric mountain pass solution has minimal period either T T T or T / 3 T/3 T /3 (also cf. Theorem 2.1 of [9 ] ). The possibility of a triple iterate is the specific issue addressed here.
In addition, there are also prescribed minimal-period results under additional radial hypotheses. Xiao [12 ] treated even potentials satisfying a global radial Hessian inequality. Xiao and Zhu [13 ] obtained related conclusions under a weak Nehari condition; the global differential inequality used there should be distinguished from an Ambrosetti–Rabinowitz condition imposed only outside a ball. Further Nehari-type results appear in Xiao and Chen [11 ] . Discrete variational approximation for even systems was developed by Kuang, Chen and Guo [6 ] , and related results for nonsymmetric potentials were studied by Kuang and Guo [7 ] .
For T > 0 T>0 T > 0 , define
E T = W 1 , 2 ( R / T Z , R n ) , (3) E_T=W^{1,2}(\mathbb{R}/T\mathbb{Z},\mathbb{R}^n), \tag*{(3)} E T = W 1 , 2 ( R / T Z , R n ) , (3)
S T = { u ∈ E T : u ( − t ) = u ( t ) , u ( t + T / 2 ) = − u ( t ) } , (4) S_T=\{u\in E_T:u(-t)=u(t),\ u(t+T/2)=-u(t)\}, \tag*{(4)} S T = { u ∈ E T : u ( − t ) = u ( t ) , u ( t + T /2 ) = − u ( t )} , (4)
ψ T ( u ) = ∫ 0 T ( 1 2 ∣ u ′ ( t ) ∣ 2 − V ( u ( t ) ) ) d t . (5) \psi_T(u)=\int_0^T\left(\frac{1}{2}|u'(t)|^2-V(u(t))\right)\,\mathrm{d}t. \tag*{(5)} ψ T ( u ) = ∫ 0 T ( 2 1 ∣ u ′ ( t ) ∣ 2 − V ( u ( t )) ) d t . (5)
We define the path class with endpoint of negative functional value and its minimax value:
Γ T = { γ ∈ C ( [ 0 , 1 ] , S T ) : γ ( 0 ) = 0 , ψ T ( γ ( 1 ) ) < 0 } , c T = inf γ ∈ Γ T max s ∈ [ 0 , 1 ] ψ T ( γ ( s ) ) . (6) \Gamma_T=\{\gamma\in C([0,1],S_T):\gamma(0)=0,\ \psi_T(\gamma(1))<0\},\qquad c_T=\inf_{\gamma\in\Gamma_T}\max_{s\in[0,1]}\psi_T(\gamma(s)). \tag*{(6)} Γ T = { γ ∈ C ([ 0 , 1 ] , S T ) : γ ( 0 ) = 0 , ψ T ( γ ( 1 )) < 0 } , c T = γ ∈ Γ T inf s ∈ [ 0 , 1 ] max ψ T ( γ ( s )) . (6)
Here continuity of paths is always understood in the W 1 , 2 W^{1,2} W 1 , 2 topology. Whenever c T > 0 c_T>0 c T > 0 , let C T C_T C T denote the path component containing zero in
{ u ∈ S T : ψ T ( u ) < c T } . (7) \{u\in S_T:\psi_T(u)<c_T\}. \tag*{(7)} { u ∈ S T : ψ T ( u ) < c T } . (7)
Motivated by these above results, especially the first problem proposed by Long in Section 5 of [9 ] , we prove the following theorem.
Theorem 1.1. Let n ≥ 1 n\geq1 n ≥ 1 be an integer and let V : R n → R V:\mathbb{R}^n\to\mathbb{R} V : R n → R satisfy (V1′) and (V2)–(V5). Then, for every T > 0 T>0 T > 0 , the system (1) possesses a nonconstant classical solution
q ∈ C 2 ( R , R n ) ∩ S T q\in C^2(\mathbb{R},\mathbb{R}^n)\cap S_T q ∈ C 2 ( R , R n ) ∩ S T
whose minimal positive period is exactly T T T . More precisely, the solution satisfies
ψ T ( q ) = c T > 0 , D ( ψ T ∣ S T ) ( q ) = 0 , q ∈ C T ‾ . (8) \psi_T(q)=c_T>0,\qquad D(\psi_T|_{S_T})(q)=0,\qquad q\in\overline{C_T}. \tag*{(8)} ψ T ( q ) = c T > 0 , D ( ψ T ∣ S T ) ( q ) = 0 , q ∈ C T . (8)
Our proof uses global information carried by a mountain pass selection within the variational framework of Ambrosetti and Rabinowitz [1 ] . The local geometry and location of mountain pass critical points have been studied by Hofer [5 ] and Ghoussoub and Preiss [4 ] ; component-based selection also occurs in the study of partially convex reversible systems by Zhou [14 ] . We give a self-contained version for a mountain pass class whose terminal point may be any point of negative action. This version selects a critical point in the closure of the component containing zero in the strict sublevel below the mountain pass value.
Sketch of proof of Theorem 1.1
Fix T > 0 T>0 T > 0 . The symmetric space S T S_T S T excludes nonzero constant loops. The assumptions at the origin and at infinity give mountain pass geometry and the Palais–Smale condition for ψ T ∣ S T \psi_T|_{S_T} ψ T ∣ S T (cf. Lemmas 2.1 and 2.2). A deformation argument for the path class (1.6), whose negative action endpoint is free, selects a critical point q q q with
ψ T ( q ) = c T > 0 , q ∈ C T ‾ \psi_T(q)=c_T>0,\qquad q\in\overline{C_T} ψ T ( q ) = c T > 0 , q ∈ C T
(cf. Corollary 2.5). The component C T C_T C T contains no point of negative action: a path from zero to such a point inside { ψ T < c T } \{\psi_T<c_T\} { ψ T < c T } would contradict the definition of c T c_T c T .
Let τ \tau τ be the minimal period of q q q . Its reflection symmetry and half-antiperiodicity imply T = k τ T=k\tau T = k τ with k k k odd (cf. Lemma 2.3 ). If τ < T \tau<T τ < T , write k = 2 m + 1 ≥ 3 k=2m+1\geq3 k = 2 m + 1 ≥ 3 and restrict the orbit to [ 0 , L ] [0,L] [ 0 , L ] , where L = T / 4 L=T/4 L = T /4 . The reflection isomorphism gives ψ T = 4 I L \psi_T=4I_L ψ T = 4 I L . Cutoff velocity fields yield an isotropic vector h h h outside the radical of D 2 I L ( q ) D^2I_L(q) D 2 I L ( q ) and a negative definite subspace of dimension m m m (cf. Lemma 2.7 ).
On this quarter-period interval, we construct paths by deleting part of the motion and inserting a waiting segment at a turning point. Their action losses are explicit integrals of − ∣ q ′ ∣ 2 -|q'|^2 − ∣ q ′ ∣ 2 , and their terminal actions are negative. For k = 3 k=3 k = 3 , two such paths have opposite first-order tangents h h h and − 2 h -2h − 2 h . A local coordinate in which the functional has the form I L ( q ) + g ( y ) − ξ 2 I_L(q)+g(y)-\xi^2 I L ( q ) + g ( y ) − ξ 2 connects each nearby point of smaller action to one of these paths. For k ≥ 5 k\geq5 k ≥ 5 , the negative subspace contains a negative two-plane, which connects the two local exit directions, so one surgery path suffices. Therefore, Lemma 2.6 gives the strict sublevel exit property in Proposition 3.1 .
Since q ∈ C T ‾ q\in\overline{C_T} q ∈ C T , this exit property applies to a point of C T C_T C T sufficiently close to q q q . Joining zero to that point within C T C_T C T , and then following its exit path to negative action, produces an admissible path whose compact image has maximum strictly below c T c_T c T . This contradiction excludes every odd iterate k ≥ 3 k\geq3 k ≥ 3 and proves τ = T \tau=T τ = T .
Section 2 establishes the variational framework, compactness, boundary selection, local sublevel geometry, and the required negative directions. Section 3 proves Theorem 1.1 using the orbit-surgery exclusion. The conclusion concerns an appropriately selected symmetric solution. It does not assert that every critical point obtained from every fixed negative endpoint has minimal period T T T .
Some preliminaries
The variational framework
We use the notation H 1 = W 1 , 2 H^1=W^{1,2} H 1 = W 1 , 2 . More explicitly, E T E_T E T consists of functions u ∈ L 2 ( 0 , T ; R n ) u\in L^2(0,T;\mathbb{R}^n) u ∈ L 2 ( 0 , T ; R n ) whose weak derivative u ′ u' u ′ belongs to L 2 ( 0 , T ; R n ) L^2(0,T;\mathbb{R}^n) L 2 ( 0 , T ; R n ) and satisfy u ( 0 ) = u ( T ) u(0)=u(T) u ( 0 ) = u ( T ) . The weak derivative is characterized by
∫ 0 T ⟨ u , ϕ ′ ⟩ d t = − ∫ 0 T ⟨ u ′ , ϕ ⟩ d t , ∀ ϕ ∈ C c ∞ ( ( 0 , T ) , R n ) . \int_0^T \langle u,\phi'\rangle\,\mathrm{d}t=-\int_0^T \langle u',\phi\rangle\,\mathrm{d}t,\qquad\forall\phi\in C_c^\infty((0,T),\mathbb{R}^n). ∫ 0 T ⟨ u , ϕ ′ ⟩ d t = − ∫ 0 T ⟨ u ′ , ϕ ⟩ d t , ∀ ϕ ∈ C c ∞ (( 0 , T ) , R n ) .
The inner product
( u , v ) H 1 = ∫ 0 T ( ⟨ u , v ⟩ + ⟨ u ′ , v ′ ⟩ ) d t (u,v)_{H^1}=\int_0^T\left(\langle u,v\rangle+\langle u',v'\rangle\right)\,\mathrm{d}t ( u , v ) H 1 = ∫ 0 T ( ⟨ u , v ⟩ + ⟨ u ′ , v ′ ⟩ ) d t
makes E T E_T E T to be a Hilbert space. This H 1 H^1 H 1 -norm satisfies ∥ u ∥ H 1 2 = ∥ u ∥ 2 2 + ∥ u ′ ∥ 2 2 \|u\|_{H^1}^2=\|u\|_2^2+\|u'\|_2^2 ∥ u ∥ H 1 2 = ∥ u ∥ 2 2 + ∥ u ′ ∥ 2 2 .
On a bounded interval, the fundamental theorem for Sobolev functions states that w ∈ W 1 , p ( a , b ) w\in W^{1,p}(a,b) w ∈ W 1 , p ( a , b ) , 1 ≤ p ≤ ∞ 1\le p\le\infty 1 ≤ p ≤ ∞ , has an absolutely continuous representative such that
w ( t ) − w ( s ) = ∫ s t w ′ ( r ) d r , ∀ s , t ∈ [ a , b ] ; w(t)-w(s)=\int_s^t w'(r)\,\mathrm{d}r,\qquad\forall s,t\in[a,b]; w ( t ) − w ( s ) = ∫ s t w ′ ( r ) d r , ∀ s , t ∈ [ a , b ] ;
see [2 Theorem 8.2] . In fact, the primitive of the weak derivative has that derivative, and its difference from w w w has zero distributional derivative and is constant. We always use this representative and extend every loop periodically to R \mathbb{R} R . The assumptions also imply that V ( 0 ) = 0 V(0)=0 V ( 0 ) = 0 and ∇ V ( 0 ) = 0 \nabla V(0)=0 ∇ V ( 0 ) = 0 .
For c T > 0 c_T>0 c T > 0 , put A T = { u ∈ S T : ψ T ( u ) < c T } A_T=\{u\in S_T:\psi_T(u)<c_T\} A T = { u ∈ S T : ψ T ( u ) < c T } . Then
C T = { u ∈ A T : ∃ η ∈ C ( [ 0 , 1 ] , A T ) , η ( 0 ) = 0 , η ( 1 ) = u } . C_T=\{u\in A_T:\exists\eta\in C([0,1],A_T),\ \eta(0)=0,\ \eta(1)=u\}. C T = { u ∈ A T : ∃ η ∈ C ([ 0 , 1 ] , A T ) , η ( 0 ) = 0 , η ( 1 ) = u } .
Thus q ∈ C T ‾ q\in\overline{C_T} q ∈ C T means that every H 1 H^1 H 1 -ball about q q q in S T S_T S T contains a point of C T C_T C T .
Lemma 2.1. Assume that V V V satisfies (V1′) and (V2)–(V5). Then S T S_T S T is a closed Hilbert subspace of E T E_T E T , and
∥ u ∥ T : = ∥ u ′ ∥ L 2 ( 0 , T ) \|u\|_T:=\|u'\|_{L^2(0,T)} ∥ u ∥ T := ∥ u ′ ∥ L 2 ( 0 , T ) is a Hilbert norm equivalent to the H 1 H^1 H 1 -norm on S T S_T S T . The restriction ψ T ∣ S T \psi_T|_{S_T} ψ T ∣ S T is of class C 2 C^2 C 2 , the class Γ T \Gamma_T Γ T is nonempty, and 0 < c T < ∞ 0<c_T<\infty 0 < c T < ∞ .
If L = T / 4 L=T/4 L = T /4 , the restriction to [ 0 , L ] [0,L] [ 0 , L ] is a bounded linear isomorphism
S T ≃ X L : = { u ∈ H 1 ( [ 0 , L ] , R n ) : u ( L ) = 0 } . (9) S_T\simeq X_L:=\{u\in H^1([0,L],\mathbb{R}^n):u(L)=0\}. \tag*{(9)} S T ≃ X L := { u ∈ H 1 ([ 0 , L ] , R n ) : u ( L ) = 0 } . (9) Let R L u = u ∣ [ 0 , L ] R_Lu=u|_{[0,L]} R L u = u ∣ [ 0 , L ] , we have
ψ T ( u ) = 4 I L ( R L u ) , I L ( w ) = ∫ 0 L ( 1 2 ∣ w ′ ( t ) ∣ 2 − V ( w ( t ) ) ) d t . (10) \psi_T(u)=4I_L(R_Lu),\qquad I_L(w)=\int_0^L\left(\frac{1}{2}|w'(t)|^2-V(w(t))\right)\,\mathrm{d}t. \tag*{(10)} ψ T ( u ) = 4 I L ( R L u ) , I L ( w ) = ∫ 0 L ( 2 1 ∣ w ′ ( t ) ∣ 2 − V ( w ( t )) ) d t . (10) Here w ∈ X L w\in X_L w ∈ X L . We identify a symmetric loop with its restriction through (9) ; thus ψ T ( u ) = 4 I L ( u ) \psi_T(u)=4I_L(u) ψ T ( u ) = 4 I L ( u ) will also be used when the domain of the argument is clear.
Proof. On E T E_T E T , define the reflection and half-period translation by
( R u ) ( t ) = u ( − t ) , ( A u ) ( t ) = u ( t + T / 2 ) . (\mathcal{R}u)(t)=u(-t),\qquad(\mathcal{A}u)(t)=u(t+T/2). ( R u ) ( t ) = u ( − t ) , ( A u ) ( t ) = u ( t + T /2 ) .
These two maps are linear isometries for the H 1 H^1 H 1 -norm. Hence S T = ker ( R − Id ) ∩ ker ( A + Id ) S_T=\ker(\mathcal{R}-\operatorname{Id})\cap\ker(\mathcal{A}+\operatorname{Id}) S T = ker ( R − Id ) ∩ ker ( A + Id ) is closed. For u ∈ S T u\in S_T u ∈ S T , its two symmetries give
u ( T / 4 ) = u ( − T / 4 ) = − u ( T / 4 ) , u(T/4)=u(-T/4)=-u(T/4), u ( T /4 ) = u ( − T /4 ) = − u ( T /4 ) ,
so u ( T / 4 ) = 0 u(T/4)=0 u ( T /4 ) = 0 . Then the fundamental theorem gives
u ( t ) = ∫ T / 4 t u ′ ( r ) d r , ∀ 0 ≤ t ≤ T . u(t)=\int_{T/4}^{t}u'(r)\,\mathrm{d}r,\qquad\forall0\le t\le T. u ( t ) = ∫ T /4 t u ′ ( r ) d r , ∀0 ≤ t ≤ T .
Let I t I_t I t be the interval with endpoints T / 4 T/4 T /4 and t t t , the Cauchy–Schwarz inequality yields
∣ u ( t ) ∣ ≤ ∣ t − T / 4 ∣ 1 / 2 ( ∫ I t ∣ u ′ ( r ) ∣ 2 d r ) 1 / 2 ≤ T ∥ u ′ ∥ L 2 ( 0 , T ) . \lvert u(t)\rvert\le\lvert t-T/4\rvert^{1/2}\left(\int_{I_t}\lvert u'(r)\rvert^2\,\mathrm{d}r\right)^{1/2}\le\sqrt{T}\lVert u'\rVert_{L^2(0,T)}. ∣ u ( t )∣ ≤ ∣ t − T /4 ∣ 1/2 ( ∫ I t ∣ u ′ ( r ) ∣ 2 d r ) 1/2 ≤ T ∥ u ′ ∥ L 2 ( 0 , T ) .
Taking the supremum and then integrating its square proves
∥ u ∥ ∞ ≤ T ∥ u ′ ∥ 2 , ∥ u ∥ 2 2 ≤ T ∥ u ∥ ∞ 2 ≤ T 2 ∥ u ′ ∥ 2 2 . (11) \lVert u\rVert_\infty\le\sqrt{T}\lVert u'\rVert_2,\qquad\lVert u\rVert_2^2\le T\lVert u\rVert_\infty^2\le T^2\lVert u'\rVert_2^2. \tag*{(11)} ∥ u ∥ ∞ ≤ T ∥ u ′ ∥ 2 , ∥ u ∥ 2 2 ≤ T ∥ u ∥ ∞ 2 ≤ T 2 ∥ u ′ ∥ 2 2 . (11)
In particular,
∥ u ′ ∥ 2 ≤ ∥ u ∥ H 1 ≤ 1 + T 2 ∥ u ′ ∥ 2 . \lVert u'\rVert_2\le\lVert u\rVert_{H^1}\le\sqrt{1+T^2}\lVert u'\rVert_2. ∥ u ′ ∥ 2 ≤ ∥ u ∥ H 1 ≤ 1 + T 2 ∥ u ′ ∥ 2 .
Thus this proves the norm equivalence.
For w ∈ X L w\in X_L w ∈ X L , we define its extension on [ 0 , 4 L ] [0,4L] [ 0 , 4 L ] by
( E L w ) ( t ) = { w ( t ) , 0 ≤ t ≤ L , − w ( 2 L − t ) , L ≤ t ≤ 2 L , − w ( t − 2 L ) , 2 L ≤ t ≤ 3 L , w ( 4 L − t ) , 3 L ≤ t ≤ 4 L . (\mathcal{E}_Lw)(t)=
\begin{cases}
w(t), & 0\le t\le L,\\
-w(2L-t), & L\le t\le2L,\\
-w(t-2L), & 2L\le t\le3L,\\
w(4L-t), & 3L\le t\le4L.
\end{cases} ( E L w ) ( t ) = ⎩ ⎨ ⎧ w ( t ) , − w ( 2 L − t ) , − w ( t − 2 L ) , w ( 4 L − t ) , 0 ≤ t ≤ L , L ≤ t ≤ 2 L , 2 L ≤ t ≤ 3 L , 3 L ≤ t ≤ 4 L .
The adjacent values match because w ( L ) = 0 w(L)=0 w ( L ) = 0 ; the values at 0 0 0 and 4 L 4L 4 L also match. The extension is absolutely continuous on each piece and, by matching, on the whole period. Its piecewise derivative belongs to L 2 L^2 L 2 . So It belongs to H 1 H^1 H 1 . The formulas imply E L w ( − t ) = E L w ( t ) \mathcal{E}_Lw(-t)=\mathcal{E}_Lw(t) E L w ( − t ) = E L w ( t ) and E L w ( t + 2 L ) = − E L w ( t ) \mathcal{E}_Lw(t+2L)=-\mathcal{E}_Lw(t) E L w ( t + 2 L ) = − E L w ( t ) after periodic continuation. Conversely these identities uniquely determine every quarter from the first. Furthermore,
∥ E L w ∥ H 1 ( 0 , 4 L ) 2 = 4 ∥ w ∥ H 1 ( 0 , L ) 2 . \lVert\mathcal{E}_Lw\rVert_{H^1(0,4L)}^2=4\lVert w\rVert_{H^1(0,L)}^2. ∥ E L w ∥ H 1 ( 0 , 4 L ) 2 = 4 ∥ w ∥ H 1 ( 0 , L ) 2 .
Thus the extension and restriction are bounded inverse linear maps.
On each quarter, change variables to r ∈ [ 0 , L ] r\in[0,L] r ∈ [ 0 , L ] . Reflection or a sign change leaves ∣ w ′ ( r ) ∣ 2 \lvert w'(r)\rvert^2 ∣ w ′ ( r ) ∣ 2 invariant, and (V5) gives V ( − w ( r ) ) = V ( w ( r ) ) V(-w(r))=V(w(r)) V ( − w ( r )) = V ( w ( r )) . Hence, for j = 0 , 1 , 2 , 3 j=0,1,2,3 j = 0 , 1 , 2 , 3 ,
∫ j L ( j + 1 ) L ( 1 2 ∣ ( E L w ) ′ ( t ) ∣ 2 − V ( E L w ( t ) ) ) d t = ∫ 0 L ( 1 2 ∣ w ′ ( r ) ∣ 2 − V ( w ( r ) ) ) d r . \int_{jL}^{(j+1)L}\left(\frac{1}{2}\lvert(\mathcal{E}_Lw)'(t)\rvert^2-V(\mathcal{E}_Lw(t))\right)\,\mathrm{d}t
=\int_0^L\left(\frac{1}{2}\lvert w'(r)\rvert^2-V(w(r))\right)\,\mathrm{d}r. ∫ j L ( j + 1 ) L ( 2 1 ∣( E L w ) ′ ( t ) ∣ 2 − V ( E L w ( t )) ) d t = ∫ 0 L ( 2 1 ∣ w ′ ( r ) ∣ 2 − V ( w ( r )) ) d r .
Summing the four equal integrals gets ψ T ( E L w ) = 4 I L ( w ) \psi_T(\mathcal{E}_Lw)=4I_L(w) ψ T ( E L w ) = 4 I L ( w ) , as claimed in (10) .
The continuous embedding H 1 ( [ 0 , T ] ) ↪ C 0 ( [ 0 , T ] ) H^1([0,T])\hookrightarrow C^0([0,T]) H 1 ([ 0 , T ]) ↪ C 0 ([ 0 , T ]) places the ranges of all loops in a bounded H 1 H^1 H 1 neighborhood in one compact ball of R n \mathbb{R}^n R n . Taylor’s formula, together with uniform continuity of D 2 V D^{2}V D 2 V on that ball, gives continuous Fréchet derivatives
D ψ T ( u ) [ v ] = ∫ 0 T ⟨ u ′ , v ′ ⟩ − ⟨ ∇ V ( u ) , v ⟩ d t , (12) D\psi_T(u)[v]=\int_{0}^{T}\langle u',v'\rangle-\langle\nabla V(u),v\rangle\,dt, \tag*{(12)} D ψ T ( u ) [ v ] = ∫ 0 T ⟨ u ′ , v ′ ⟩ − ⟨ ∇ V ( u ) , v ⟩ d t , (12)
D 2 ψ T ( u ) [ v , w ] = ∫ 0 T ⟨ v ′ , w ′ ⟩ − ⟨ D 2 V ( u ) v , w ⟩ d t . (13) D^{2}\psi_T(u)[v,w]=\int_{0}^{T}\langle v',w'\rangle-\langle D^{2}V(u)v,w\rangle\,dt. \tag*{(13)} D 2 ψ T ( u ) [ v , w ] = ∫ 0 T ⟨ v ′ , w ′ ⟩ − ⟨ D 2 V ( u ) v , w ⟩ d t . (13)
It follows from (V2) that the potential is positive on the sphere ∣ x ∣ = r 0 |x|=r_0 ∣ x ∣ = r 0 . Hence b 0 = min ∣ x ∣ = r 0 V ( x ) > 0 b_0=\min_{|x|=r_0}V(x)>0 b 0 = min ∣ x ∣ = r 0 V ( x ) > 0 . For a unit vector e e e and r ≥ r 0 r\ge r_0 r ≥ r 0 ,
d d r ( r − μ V ( r e ) ) = ⟨ ∇ V ( r e ) , r e ⟩ − μ V ( r e ) r μ + 1 ≥ 0. \frac{\mathrm{d}}{\mathrm{d}r}\left(r^{-\mu}V(re)\right)=\frac{\langle\nabla V(re),re\rangle-\mu V(re)}{r^{\mu+1}}\ge0. d r d ( r − μ V ( r e ) ) = r μ + 1 ⟨ ∇ V ( r e ) , r e ⟩ − μ V ( r e ) ≥ 0.
Integrating from r 0 r_0 r 0 to r r r yields V ( r e ) ≥ r μ r 0 − μ V ( r 0 e ) ≥ b r μ V(re)\ge r^{\mu}r_0^{-\mu}V(r_0e)\ge br^{\mu} V ( r e ) ≥ r μ r 0 − μ V ( r 0 e ) ≥ b r μ , where b = b 0 r 0 − μ b=b_0r_0^{-\mu} b = b 0 r 0 − μ . On ∣ x ∣ ≤ r 0 |x|\le r_0 ∣ x ∣ ≤ r 0 , the nonnegativity gives V ( x ) ≥ 0 ≥ b ∣ x ∣ μ − b 0 V(x)\ge0\ge b|x|^{\mu}-b_0 V ( x ) ≥ 0 ≥ b ∣ x ∣ μ − b 0 . Thus,
V ( x ) ≥ b ∣ x ∣ μ − b 0 , ∀ x ∈ R n . (14) V(x)\ge b|x|^{\mu}-b_0,\qquad\forall x\in\mathbb{R}^{n}. \tag*{(14)} V ( x ) ≥ b ∣ x ∣ μ − b 0 , ∀ x ∈ R n . (14)
Choose ε = 1 / ( 4 T 2 ) \varepsilon=1/(4T^2) ε = 1/ ( 4 T 2 ) . By (V3), there is δ > 0 \delta>0 δ > 0 such that V ( x ) ≤ ε ∣ x ∣ 2 V(x)\le\varepsilon|x|^2 V ( x ) ≤ ε ∣ x ∣ 2 for ∣ x ∣ ≤ δ |x|\le\delta ∣ x ∣ ≤ δ . Choose ρ > 0 \rho>0 ρ > 0 with T ρ ≤ δ \sqrt{T}\rho\le\delta T ρ ≤ δ . If ∥ u ∥ T ≤ ρ \|u\|_T\le\rho ∥ u ∥ T ≤ ρ , (2.3) gives ∥ u ∥ ∞ ≤ δ \|u\|_\infty\le\delta ∥ u ∥ ∞ ≤ δ , and hence
ψ T ( u ) ≥ 1 2 ∥ u ′ ∥ 2 2 − ε ∥ u ∥ 2 2 ≥ ( 1 2 − ε T 2 ) ∥ u ∥ T 2 = 1 4 ∥ u ∥ T 2 . (15) \psi_T(u)\ge\frac{1}{2}\|u'\|_2^2-\varepsilon\|u\|_2^2\ge\left(\frac{1}{2}-\varepsilon T^2\right)\|u\|_T^2=\frac{1}{4}\|u\|_T^2. \tag*{(15)} ψ T ( u ) ≥ 2 1 ∥ u ′ ∥ 2 2 − ε ∥ u ∥ 2 2 ≥ ( 2 1 − ε T 2 ) ∥ u ∥ T 2 = 4 1 ∥ u ∥ T 2 . (15)
For every γ ∈ Γ T \gamma\in\Gamma_T γ ∈ Γ T , its negative action endpoint satisfies ∥ γ ( 1 ) ∥ T > ρ \|\gamma(1)\|_T>\rho ∥ γ ( 1 ) ∥ T > ρ , while ∥ γ ( 0 ) ∥ T = 0 \|\gamma(0)\|_T=0 ∥ γ ( 0 ) ∥ T = 0 . Then the continuity and the intermediate value theorem give some s ∗ ∈ ( 0 , 1 ) s_*\in(0,1) s ∗ ∈ ( 0 , 1 ) with ∥ γ ( s ∗ ) ∥ T = ρ \|\gamma(s_*)\|_T=\rho ∥ γ ( s ∗ ) ∥ T = ρ . So it yields max s ψ T ( γ ( s ) ) ≥ ρ 2 / 4 \max_s\psi_T(\gamma(s))\ge\rho^2/4 max s ψ T ( γ ( s )) ≥ ρ 2 /4 .
For a fixed nonzero u ∈ S T u\in S_T u ∈ S T , set A = ∫ 0 T ∣ u ( t ) ∣ μ d t > 0 A=\int_0^T|u(t)|^\mu\,dt>0 A = ∫ 0 T ∣ u ( t ) ∣ μ d t > 0 . The estimate (2.6) gives
ψ T ( r u ) ≤ 1 2 r 2 ∥ u ∥ T 2 − b r μ A + b 0 T ≡ η ( r ) ⟶ − ∞ ( r → ∞ ) , \psi_T(ru)\le\frac{1}{2}r^2\|u\|_T^2-br^\mu A+b_0T\equiv\eta(r)\longrightarrow-\infty\qquad(r\to\infty), ψ T ( r u ) ≤ 2 1 r 2 ∥ u ∥ T 2 − b r μ A + b 0 T ≡ η ( r ) ⟶ − ∞ ( r → ∞ ) ,
due to lim r → ∞ η ( r ) r − μ = − b A < 0 \lim_{r\to\infty}\eta(r)r^{-\mu}=-bA<0 lim r → ∞ η ( r ) r − μ = − b A < 0 . A nonzero vector times cos ( 2 π t / T ) \cos(2\pi t/T) cos ( 2 π t / T ) supplies such a loop. Choose R > 0 R>0 R > 0 with ψ T ( R u ) < 0 \psi_T(Ru)<0 ψ T ( R u ) < 0 and set γ ( s ) = s R u \gamma(s)=sRu γ ( s ) = s R u . Then γ ∈ Γ T \gamma\in\Gamma_T γ ∈ Γ T , and ψ T ∘ γ \psi_T\circ\gamma ψ T ∘ γ has a finite maximum on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Together with (2.7), this proves Γ T ≠ ∅ \Gamma_T\ne\varnothing Γ T = ∅ and ρ 2 / 4 ≤ c T < ∞ \rho^2/4\le c_T<\infty ρ 2 /4 ≤ c T < ∞ .
Compactness, critical points and primitive periods
Lemma 2.2. Assume that V V V satisfies (V1′) and (V2)–(V5). Then the functional ψ T ∣ S T \psi_T|_{S_T} ψ T ∣ S T satisfies the Palais–Smale condition, that is, every sequence ( u j ) (u_j) ( u j ) for which ( ψ T ( u j ) ) (\psi_T(u_j)) ( ψ T ( u j )) is bounded and ∥ D ( ψ T ∣ S T ) ( u j ) ∥ → 0 \|D(\psi_T|_{S_T})(u_j)\|\to0 ∥ D ( ψ T ∣ S T ) ( u j ) ∥ → 0 has a strongly convergent subsequence in H 1 H^1 H 1 . Every positive action critical point is a nonconstant classical solution of (1.1), and belongs to C 3 ( R , R n ) C^3(\mathbb{R},\mathbb{R}^n) C 3 ( R , R n ) .
Proof. Set F ( x ) = ⟨ ∇ V ( x ) , x ⟩ − μ V ( x ) F(x)=\langle\nabla V(x),x\rangle-\mu V(x) F ( x ) = ⟨ ∇ V ( x ) , x ⟩ − μ V ( x ) and
M = max { 0 , − min ∣ x ∣ ≤ r 0 F ( x ) } . M=\max\left\{0,-\min_{|x|\le r_0}F(x)\right\}. M = max { 0 , − ∣ x ∣ ≤ r 0 min F ( x ) } .
The minimum exists by the continuity of F F F . It follows from (V2) that F ≥ 0 F\ge0 F ≥ 0 outside the ball, so F ≥ − M F\ge-M F ≥ − M on R n \mathbb{R}^n R n . By (2.4) with v = u v=u v = u , we obtain
μ ψ T ( u ) − D ψ T ( u ) [ u ] = ( μ 2 − 1 ) ∥ u ′ ∥ 2 2 + ∫ 0 T F ( u ( t ) ) d t ≥ ( μ 2 − 1 ) ∥ u ∥ T 2 − M T , ∀ u ∈ S T . (16) \begin{aligned}
\mu\psi_T(u)-D\psi_T(u)[u]=\left(\frac{\mu}{2}-1\right)\|u'\|_2^2+\int_0^T F(u(t))\,dt \\
&\ge\left(\frac{\mu}{2}-1\right)\|u\|_T^2-MT,\qquad\forall u\in S_T. \tag*{(16)}
\end{aligned} μ ψ T ( u ) − D ψ T ( u ) [ u ] = ( 2 μ − 1 ) ∥ u ′ ∥ 2 2 + ∫ 0 T F ( u ( t )) d t ≥ ( 2 μ − 1 ) ∥ u ∥ T 2 − M T , ∀ u ∈ S T . (16)
For a P.S. sequence, choose C ≥ 0 C \ge0 C ≥ 0 with ∣ ψ T ( u j ) ∣ ≤ C |\psi_T(u_j)| \le C ∣ ψ T ( u j ) ∣ ≤ C and put
ϵ j = ∥ D ( ψ T ∣ S T ) ( u j ) ∥ ( S T ) ∗ ⟶ 0 , \epsilon_j=\left\|D(\psi_T|_{S_T})(u_j)\right\|_{(S_T)^*}\longrightarrow0, ϵ j = ∥ D ( ψ T ∣ S T ) ( u j ) ∥ ( S T ) ∗ ⟶ 0 ,
where the dual norm corresponds to ∥ ⋅ ∥ T \|\cdot\|_T ∥ ⋅ ∥ T . Let A = μ / 2 − 1 > 0 A=\mu/2-1>0 A = μ /2 − 1 > 0 , then
A ∥ u j ∥ T 2 ≤ μ C + M T + ϵ j ∥ u j ∥ T . A\|u_j\|_T^2\leq\mu C+MT+\epsilon_j\|u_j\|_T. A ∥ u j ∥ T 2 ≤ μ C + M T + ϵ j ∥ u j ∥ T .
If ∥ u j ∥ T → ∞ \|u_j\|_T\to\infty ∥ u j ∥ T → ∞ , then, after being divided by ∥ u j ∥ T 2 \|u_j\|_T^2 ∥ u j ∥ T 2 , the above inequality gives A ≤ 0 A\leq0 A ≤ 0 . Thus ( u j ) (u_j) ( u j ) is bounded, and the norm equivalence gives boundedness in H 1 H^1 H 1 .
The estimates (11) and
∣ u j ( t ) − u j ( s ) ∣ ≤ ∥ u j ′ ∥ 2 ∣ t − s ∣ 1 / 2 |u_j(t)-u_j(s)|\leq\|u_j'\|_2|t-s|^{1/2} ∣ u j ( t ) − u j ( s ) ∣ ≤ ∥ u j ′ ∥ 2 ∣ t − s ∣ 1/2
give the uniform boundedness and equicontinuity on a period interval. Passing to a subsequence, u j ⇀ u u_j\rightharpoonup u u j ⇀ u in H 1 H^1 H 1 and u j → u u_j\to u u j → u uniformly, which implies that u ∈ S T u\in S_T u ∈ S T .
Now we have
∫ 0 T ⟨ u j ′ , u j ′ − u ′ ⟩ d t = D ψ T ( u j ) ( u j − u ) + ∫ 0 T ⟨ ∇ V ( u j ) , u j − u ⟩ d t ⟶ 0 , \int_0^T\langle u_j',u_j'-u'\rangle\,dt=D\psi_T(u_j)(u_j-u)+\int_0^T\langle\nabla V(u_j),u_j-u\rangle\,dt\longrightarrow0, ∫ 0 T ⟨ u j ′ , u j ′ − u ′ ⟩ d t = D ψ T ( u j ) ( u j − u ) + ∫ 0 T ⟨ ∇ V ( u j ) , u j − u ⟩ d t ⟶ 0 ,
where the first term tends to zero by the boundedness and the P.S. hypothesis; the second one is due to the uniform convergence and boundedness of ∇ V \nabla V ∇ V . Subtracting the term that tends to zero by weak convergence yields
∥ u j ′ − u ′ ∥ 2 2 ⟶ 0. \|u_j'-u'\|_2^2\longrightarrow0. ∥ u j ′ − u ′ ∥ 2 2 ⟶ 0.
Then the norm equivalence proves the strong convergence in H 1 H^1 H 1 .
Let q q q be a critical point and L = T / 4 L=T/4 L = T /4 . The restriction and (2.2) imply D I L ( q ) [ v ] = 0 D I_L(q)[v]=0 D I L ( q ) [ v ] = 0 , ∀ v ∈ X L \forall v\in X_L ∀ v ∈ X L . For a test function ϕ ∈ C c ∞ ( ( 0 , L ) , R n ) \phi\in C_c^\infty((0,L),\mathbb{R}^n) ϕ ∈ C c ∞ (( 0 , L ) , R n ) , extending it by zero to the endpoints gives ϕ ∈ X L \phi\in X_L ϕ ∈ X L , and q + ϵ ϕ ∈ X L q+\epsilon\phi\in X_L q + ϵ ϕ ∈ X L for every ϵ ∈ R \epsilon\in\mathbb{R} ϵ ∈ R . Thus
0 = d d ϵ I L ( q + ϵ ϕ ) ∣ ϵ = 0 = ∫ 0 L ( ⟨ q ′ , ϕ ′ ⟩ − ⟨ ∇ V ( q ) , ϕ ⟩ ) d t . 0=\left.\frac{d}{d\epsilon}I_L(q+\epsilon\phi)\right|_{\epsilon=0}
=\int_0^L\left(\langle q',\phi'\rangle-\langle\nabla V(q),\phi\rangle\right)\,dt. 0 = d ϵ d I L ( q + ϵ ϕ ) ϵ = 0 = ∫ 0 L ( ⟨ q ′ , ϕ ′ ⟩ − ⟨ ∇ V ( q ) , ϕ ⟩ ) d t .
This yields q ¨ = − ∇ V ( q ) \ddot q=-\nabla V(q) q ¨ = − ∇ V ( q ) on ( 0 , L ) (0,L) ( 0 , L ) . Since q q q and ∇ V ( q ) \nabla V(q) ∇ V ( q ) are continuous, this proves q ∈ C 2 ( [ 0 , L ] , R n ) q\in C^2([0,L],\mathbb{R}^n) q ∈ C 2 ([ 0 , L ] , R n ) .
For every v ∈ X L v\in X_L v ∈ X L , noting that v ( L ) = 0 v(L)=0 v ( L ) = 0 , by the integration by parts, we have
0 = D I L ( q ) [ v ] = [ ⟨ q ′ ( t ) , v ( t ) ⟩ ] 0 L − ∫ 0 L ⟨ q ¨ + ∇ V ( q ) , v ⟩ d t = − ⟨ q ′ ( 0 ) , v ( 0 ) ⟩ . 0=D I_L(q)[v]=[\langle q'(t),v(t)\rangle]_0^L-\int_0^L\langle\ddot q+\nabla V(q),v\rangle\,dt=-\langle q'(0),v(0)\rangle. 0 = D I L ( q ) [ v ] = [⟨ q ′ ( t ) , v ( t )⟩ ] 0 L − ∫ 0 L ⟨ q ¨ + ∇ V ( q ) , v ⟩ d t = − ⟨ q ′ ( 0 ) , v ( 0 )⟩ .
For any ζ ∈ R n \zeta\in\mathbb{R}^n ζ ∈ R n , the variation v ( t ) = ( 1 − t / L ) ζ ∈ X L v(t)=(1-t/L)\zeta\in X_L v ( t ) = ( 1 − t / L ) ζ ∈ X L and v ( 0 ) = ζ v(0)=\zeta v ( 0 ) = ζ . Taking ζ = q ′ ( 0 ) \zeta=q'(0) ζ = q ′ ( 0 ) gives q ′ ( 0 ) = 0 q'(0)=0 q ′ ( 0 ) = 0 . So we have
q ′ ( 0 ) = 0 , q ( L ) = 0. (17) q'(0)=0,\qquad q(L)=0. \tag*{(17)} q ′ ( 0 ) = 0 , q ( L ) = 0. (17)
Even reflection at 0 matches the first derivative because q ′ ( 0 ) = 0 q'(0)=0 q ′ ( 0 ) = 0 . At L L L , odd reflection matches the value because q ( L ) = 0 q(L)=0 q ( L ) = 0 , matches the first derivative automatically, and matches the second derivative because q ¨ ( L ) = − ∇ V ( 0 ) = 0 \ddot q(L)=-\nabla V(0)=0 q ¨ ( L ) = − ∇ V ( 0 ) = 0 . Since ∇ V \nabla V ∇ V is odd, both reflections preserve the same equation. They give a globally C 2 C^{2} C 2 solution. As V ∈ C 2 V \in C^{2} V ∈ C 2 , the composition ∇ V ( q ( t ) ) \nabla V(q(t)) ∇ V ( q ( t )) is C 1 C^{1} C 1 , and
d d t q ¨ ( t ) = − D 2 V ( q ( t ) ) q ′ ( t ) \frac{\mathrm{d}}{\mathrm{d}t}\ddot{q}(t)=-D^{2}V(q(t))q'(t) d t d q ¨ ( t ) = − D 2 V ( q ( t )) q ′ ( t )
is continuous. Thus the claimed C 3 C^{3} C 3 regularity of q q q is proved. □
Lemma 2.3. If q : R → R n q:\mathbb{R}\to\mathbb{R}^{n} q : R → R n is nonconstant, continuous and periodic , its set of periods is τ Z \tau\mathbb{Z} τ Z for some τ > 0 \tau>0 τ > 0 . Moreover, if q ∈ S T q\in S_{T} q ∈ S T , then T = k τ T=k\tau T = k τ for a positive odd integer k k k , and q ∈ S τ q\in S_{\tau} q ∈ S τ .
Proof. The period set
G = { h ∈ R : q ( t + h ) = q ( t ) , ∀ t ∈ R } G=\{h\in\mathbb{R}:q(t+h)=q(t),\ \forall t\in\mathbb{R}\} G = { h ∈ R : q ( t + h ) = q ( t ) , ∀ t ∈ R } is an additive subgroup of R \mathbb{R} R . If g j ∈ G g_{j}\in G g j ∈ G and g j → g g_{j}\to g g j → g , the continuity gives q ( t + g ) = lim j q ( t + g j ) = q ( t ) q(t+g)=\lim_{j}q(t+g_{j})=q(t) q ( t + g ) = lim j q ( t + g j ) = q ( t ) , ∀ t \forall t ∀ t . So It is closed.
If the positive period sequence h j → 0 h_{j}\to0 h j → 0 , then for any s ∈ R s\in\mathbb{R} s ∈ R , choose m j = ⌊ s / h j ⌋ m_{j}=\lfloor s/h_{j}\rfloor m j = ⌊ s / h j ⌋ , it holds
0 ≤ s − m j h j < h j , 0\leq s-m_{j}h_{j}<h_{j}, 0 ≤ s − m j h j < h j , so m j h j → s m_{j}h_{j}\to s m j h j → s . Since each integer multiple of a period is a period,
q ( t + s ) = lim j → ∞ q ( t + m j h j ) = q ( t ) , ∀ t , s ∈ R , q(t+s)=\lim_{j\to\infty}q(t+m_{j}h_{j})=q(t),\qquad\forall t,s\in\mathbb{R}, q ( t + s ) = j → ∞ lim q ( t + m j h j ) = q ( t ) , ∀ t , s ∈ R , which makes q q q to be constant. Therefore τ = inf ( G ∩ ( 0 , ∞ ) ) > 0 \tau=\inf(G\cap(0,\infty))>0 τ = inf ( G ∩ ( 0 , ∞ )) > 0 .
Choose g j ∈ G g_{j}\in G g j ∈ G with τ ≤ g j < τ + 1 / j \tau\leq g_{j}<\tau+1/j τ ≤ g j < τ + 1/ j , the closedness of G G G gives τ ∈ G \tau\in G τ ∈ G . For any g ∈ G g\in G g ∈ G , write
g = p τ + r , p = ⌊ g / τ ⌋ ∈ Z , 0 ≤ r < τ . g=p\tau+r,\qquad p=\lfloor g/\tau\rfloor\in\mathbb{Z},\qquad0\leq r<\tau. g = p τ + r , p = ⌊ g / τ ⌋ ∈ Z , 0 ≤ r < τ . The subgroup property implies r = g − p τ ∈ G r=g-p\tau\in G r = g − p τ ∈ G . A positive r r r would contradict the definition of τ \tau τ , so r = 0 r=0 r = 0 . Hence G = τ Z G=\tau\mathbb{Z} G = τ Z . For q ∈ S T q\in S_{T} q ∈ S T , one has T ∈ G T\in G T ∈ G and therefore T = k τ T=k\tau T = k τ with k ∈ N k\in\mathbb{N} k ∈ N .
If k k k were even, T / 2 T/2 T /2 would be a period. Then half-antiperiodicity give q ( t ) = − q ( t ) q(t)=-q(t) q ( t ) = − q ( t ) , ∀ t \forall t ∀ t , contradiction! Thus k = 2 j + 1 k=2j+1 k = 2 j + 1 for an integer j ≥ 0 j\geq0 j ≥ 0 . Since T / 2 = j τ + τ / 2 T/2=j\tau+\tau/2 T /2 = j τ + τ /2 , we have
q ( t + τ / 2 ) = − q ( t ) . q(t+\tau/2)=-q(t). q ( t + τ /2 ) = − q ( t ) . Hence q ∈ S τ q\in S_{\tau} q ∈ S τ . □
A boundary-selected mountain pass point
We next isolate the precise selection property used in the proof. For a subset A A A of a Hilbert space, A ‾ \overline{A} A denotes its closure.
Lemma 2.4. Let H H H be a real Hilbert space , let f ∈ C 2 ( H , R ) f\in C^{2}(H,\mathbb{R}) f ∈ C 2 ( H , R ) with f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , and define
Γ = { γ ∈ C ( [ 0 , 1 ] , H ) : γ ( 0 ) = 0 , f ( γ ( 1 ) ) < 0 } , c = inf γ ∈ Γ max s ∈ [ 0 , 1 ] f ( γ ( s ) ) . \Gamma=\{\gamma\in C([0,1],H):\gamma(0)=0,\ f(\gamma(1))<0\},\qquad c=\inf_{\gamma\in\Gamma}\max_{s\in[0,1]}f(\gamma(s)). Γ = { γ ∈ C ([ 0 , 1 ] , H ) : γ ( 0 ) = 0 , f ( γ ( 1 )) < 0 } , c = γ ∈ Γ inf s ∈ [ 0 , 1 ] max f ( γ ( s )) . Assume Γ ≠ ∅ \Gamma\ne\varnothing Γ = ∅ , 0 < c < ∞ 0<c<\infty 0 < c < ∞ , and f f f satisfies the P.S. condition at c c c : every sequence x j x_{j} x j with f ( x j ) → c f(x_{j})\to c f ( x j ) → c and ∥ D f ( x j ) ∥ → 0 \lVert Df(x_{j})\rVert\to0 ∥ D f ( x j )∥ → 0 has a strongly convergent subsequence. Let C C C be the path component of { f < c } \{f<c\} { f < c } containing zero, and set
K c = { x ∈ H : f ( x ) = c , D f ( x ) = 0 } . K_{c}=\{x\in H:f(x)=c,\ Df(x)=0\}. K c = { x ∈ H : f ( x ) = c , D f ( x ) = 0 } . Then K c K_{c} K c is nonempty and compact, and
K c ∩ C ‾ ≠ ∅ . (18) K_{c}\cap\overline{C}\ne\varnothing. \tag*{(18)} K c ∩ C = ∅ . (18)
Proof. We identify the derivative with the gradient through
D f ( x ) [ v ] = ( ∇ f ( x ) , v ) H , ∥ D f ( x ) ∥ H ∗ = ∥ ∇ f ( x ) ∥ H . Df(x)[v] = (\nabla f(x),v)_H,\qquad\lVert Df(x)\rVert_{H^*} = \lVert\nabla f(x)\rVert_H. D f ( x ) [ v ] = ( ∇ f ( x ) , v ) H , ∥ D f ( x ) ∥ H ∗ = ∥ ∇ f ( x ) ∥ H .
Thus the P.S. condition is unchanged when using ∇ f \nabla f ∇ f . The gradient is C 1 C^1 C 1 , since f f f is C 2 C^2 C 2 and the Riesz identification is a bounded linear map. Its derivative is bounded on a sufficiently small ball at each point. So integration along line segments shows that ∇ f \nabla f ∇ f is locally Lipschitz.
Step 1: K c ≠ ∅ K_c \ne\emptyset K c = ∅ and K c K_c K c is compact.
Suppose first that K c = ∅ K_c = \emptyset K c = ∅ . The P.S. condition gives η , σ > 0 \eta,\sigma> 0 η , σ > 0 such that
∥ ∇ f ( x ) ∥ ≥ σ when ∣ f ( x ) − c ∣ ≤ 2 η . (19) \lVert\nabla f(x)\rVert\ge\sigma\qquad\text{when } \lvert f(x)-c\rvert\le2\eta. \tag*{(19)} ∥ ∇ f ( x )∥ ≥ σ when ∣ f ( x ) − c ∣ ≤ 2 η . (19)
Otherwise, for every η , σ > 0 \eta,\sigma> 0 η , σ > 0 , there are x j ∈ H x_j \in H x j ∈ H with ∣ f ( x j ) − c ∣ ≤ 1 / j \lvert f(x_j)-c\rvert\le1/j ∣ f ( x j ) − c ∣ ≤ 1/ j and ∥ ∇ f ( x j ) ∥ < 1 / j \lVert\nabla f(x_j)\rVert< 1/j ∥ ∇ f ( x j )∥ < 1/ j . A strongly convergent P.S. subsequence would have a limit x ∗ x_* x ∗ with f ( x ∗ ) = c f(x_*)=c f ( x ∗ ) = c and ∇ f ( x ∗ ) = 0 \nabla f(x_*)=0 ∇ f ( x ∗ ) = 0 , contradicting K c = ∅ K_c = \emptyset K c = ∅ . Choose η \eta η so that 2 η < c 2\eta< c 2 η < c .
Define the Lipschitz cutoff function
α ( λ ) = { 1 , ∣ λ − c ∣ ≤ η , 2 − ∣ λ − c ∣ / η , η < ∣ λ − c ∣ < 2 η , 0 , ∣ λ − c ∣ ≥ 2 η . \alpha(\lambda) =
\begin{cases}
1, & \lvert\lambda-c\rvert\le\eta,\\
2-\lvert\lambda-c\rvert/\eta, & \eta< \lvert\lambda-c\rvert< 2\eta,\\
0, & \lvert\lambda-c\rvert\ge2\eta.
\end{cases} α ( λ ) = ⎩ ⎨ ⎧ 1 , 2 − ∣ λ − c ∣ / η , 0 , ∣ λ − c ∣ ≤ η , η < ∣ λ − c ∣ < 2 η , ∣ λ − c ∣ ≥ 2 η .
It takes values in [ 0 , 1 ] [0,1] [ 0 , 1 ] and has Lipschitz constant 1 / η 1/\eta 1/ η . Consider the vector field
X ( x ) = − α ( f ( x ) ) β ( x ) ∇ f ( x ) max { ∥ ∇ f ( x ) ∥ 2 , σ 2 } . (20) X(x)=-\alpha(f(x))\beta(x)\frac{\nabla f(x)}{\max\{\lVert\nabla f(x)\rVert^2,\sigma^2\}}. \tag*{(20)} X ( x ) = − α ( f ( x )) β ( x ) max {∥ ∇ f ( x ) ∥ 2 , σ 2 } ∇ f ( x ) . (20)
In this step, take β ≡ 1 \beta\equiv1 β ≡ 1 . Let r x = ∥ ∇ f ( x ) ∥ r_x=\lVert\nabla f(x)\rVert r x = ∥ ∇ f ( x )∥ . According to the definitions, 0 ≤ α , β ≤ 1 0\le\alpha,\beta\le1 0 ≤ α , β ≤ 1 and
∥ X ( x ) ∥ ≤ r x max { r x 2 , σ 2 } ≤ 1 σ . \lVert X(x)\rVert\le\frac{r_x}{\max\{r_x^2,\sigma^2\}} \le\frac{1}{\sigma}. ∥ X ( x )∥ ≤ max { r x 2 , σ 2 } r x ≤ σ 1 .
Since the denominator is bounded away from zero, so X X X is locally Lipschitz. The local existence and uniqueness of its flow follow by applying the contraction principle to x ( t ) = x ( 0 ) + ∫ 0 t X ( x ( r ) ) d r x(t)=x(0)+\int_0^t X(x(r))\,\mathrm{d}r x ( t ) = x ( 0 ) + ∫ 0 t X ( x ( r )) d r on a small time interval.
Let [ 0 , b ∗ ) [0,b_*) [ 0 , b ∗ ) be the maximal forward interval of x ( t ) = Φ t ( x 0 ) x(t)=\Phi_t(x_0) x ( t ) = Φ t ( x 0 ) and suppose b ∗ < ∞ b_*<\infty b ∗ < ∞ . For 0 ≤ s < t < b ∗ 0\le s<t<b_* 0 ≤ s < t < b ∗ ,
∥ x ( t ) − x ( s ) ∥ ≤ ∫ s t ∥ X ( x ( r ) ) ∥ d r ≤ ∣ t − s ∣ σ . \lVert x(t)-x(s)\rVert\le\int_s^t \lVert X(x(r))\rVert\,\mathrm{d}r \le\frac{\lvert t-s\rvert}{\sigma}. ∥ x ( t ) − x ( s )∥ ≤ ∫ s t ∥ X ( x ( r ))∥ d r ≤ σ ∣ t − s ∣ .
Thus for every ϵ > 0 \epsilon>0 ϵ > 0 , the values at any two times in ( b ∗ − σ ϵ , b ∗ ) (b_*-\sigma\epsilon,b_*) ( b ∗ − σ ϵ , b ∗ ) differ by less than ϵ \epsilon ϵ . The completeness of H H H gives a unique limit x ∗ = lim t → b ∗ x ( t ) x_*=\lim_{t\to b_*}x(t) x ∗ = lim t → b ∗ x ( t ) . Extend x x x continuously by x ( b ∗ ) = x ∗ x(b_*)=x_* x ( b ∗ ) = x ∗ . The integral equation passes to t = b ∗ t=b_* t = b ∗ , since X X X is bounded and continuous. Local existence at x ∗ x_* x ∗ supplies a solution for times greater than b ∗ b_* b ∗ , and joining it to x x x attains a solution across b ∗ b_* b ∗ , with matching derivative X ( x ∗ ) X(x_*) X ( x ∗ ) . The uniqueness makes it to be an extension of the original trajectory, contradicting the maximality. So b ∗ = ∞ b_*=\infty b ∗ = ∞ .
The local continuous dependence follows from the same contraction construction and Gronwall’s inequality. Repeating it over finitely many local time intervals covering any fixed compact trajectory segment proves continuity of x 0 ↦ Φ t ( x 0 ) x_0\mapsto\Phi_t(x_0) x 0 ↦ Φ t ( x 0 ) .
For x ( t ) = Φ t ( x ) x(t)=\Phi_t(x) x ( t ) = Φ t ( x ) , the chain rule gives
d d t f ( x ( t ) ) = D f ( x ( t ) ) [ X ( x ( t ) ) ] = − α ( f ( x ( t ) ) ) β ( x ( t ) ) ∥ ∇ f ( x ( t ) ) ∥ 2 max { ∥ ∇ f ( x ( t ) ) ∥ 2 , σ 2 } . \frac{\mathrm{d}}{\mathrm{d}t}f(x(t))=Df(x(t))[X(x(t))]=-\alpha(f(x(t)))\beta(x(t))\frac{\lVert\nabla f(x(t))\rVert^2}{\max\{\lVert\nabla f(x(t))\rVert^2,\sigma^2\}}. d t d f ( x ( t )) = D f ( x ( t )) [ X ( x ( t ))] = − α ( f ( x ( t ))) β ( x ( t )) max {∥ ∇ f ( x ( t )) ∥ 2 , σ 2 } ∥ ∇ f ( x ( t )) ∥ 2 .
When β ≡ 1 \beta\equiv1 β ≡ 1 , in the central band ∣ f ( x ( t ) ) − c ∣ ≤ η \lvert f(x(t))-c\rvert\leq\eta ∣ f ( x ( t )) − c ∣ ≤ η , one has α = 1 \alpha=1 α = 1 and ∥ ∇ f ∥ ≥ σ \lVert\nabla f\rVert\geq\sigma ∥ ∇ f ∥ ≥ σ by (19) . Therefore, the fraction is exactly 1 1 1 , and the derivative equals − 1 -1 − 1 . For any points the bounds are
− 1 ≤ d d t f ( x ( t ) ) ≤ 0. -1\leq\frac{\mathrm{d}}{\mathrm{d}t}f(x(t))\leq0. − 1 ≤ d t d f ( x ( t )) ≤ 0.
The derivative is zero where a cutoff vanishes. The unit descent rate is used only on the region where both cutoffs are one and the gradient lower bound holds.
Choose 0 < ε < min { η , c / 4 } 0<\varepsilon<\min\{\eta,c/4\} 0 < ε < min { η , c /4 } and γ ∈ Γ \gamma\in\Gamma γ ∈ Γ with max s f ( γ ( s ) ) ≤ c + ε \max_s f(\gamma(s))\leq c+\varepsilon max s f ( γ ( s )) ≤ c + ε . Define γ ~ ( s ) = Φ 2 ε ( γ ( s ) ) \widetilde{\gamma}(s)=\Phi_{2\varepsilon}(\gamma(s)) γ ( s ) = Φ 2 ε ( γ ( s )) . The endpoint values are 0 0 0 and a negative number, both below c − 2 η > 0 c-2\eta>0 c − 2 η > 0 . Thus X = 0 X=0 X = 0 at both endpoints, and uniqueness makes them to be fixed by the flow. Consequently γ ~ ∈ Γ \widetilde{\gamma}\in\Gamma γ ∈ Γ .
For fixed s s s , let x s ( t ) = Φ t ( γ ( s ) ) x_s(t)=\Phi_t(\gamma(s)) x s ( t ) = Φ t ( γ ( s )) , 0 ≤ t ≤ 2 ε 0\leq t\leq2\varepsilon 0 ≤ t ≤ 2 ε . If f ( x s ( 2 ε ) ) ≥ c f(x_s(2\varepsilon))\geq c f ( x s ( 2 ε )) ≥ c , the monotonicity gives
c ≤ f ( x s ( 2 ε ) ) ≤ f ( x s ( t ) ) ≤ f ( x s ( 0 ) ) ≤ c + ε < c + η . c\leq f(x_s(2\varepsilon))\leq f(x_s(t))\leq f(x_s(0))\leq c+\varepsilon<c+\eta. c ≤ f ( x s ( 2 ε )) ≤ f ( x s ( t )) ≤ f ( x s ( 0 )) ≤ c + ε < c + η .
The entire segment lies in [ c , c + η ] [c,c+\eta] [ c , c + η ] , and hence
f ( x s ( 2 ε ) ) = f ( x s ( 0 ) ) + ∫ 0 2 ε ( − 1 ) d t = f ( x s ( 0 ) ) − 2 ε ≤ c − ε < c , f(x_s(2\varepsilon))=f(x_s(0))+\int_0^{2\varepsilon}(-1)\,\mathrm{d}t=f(x_s(0))-2\varepsilon\leq c-\varepsilon<c, f ( x s ( 2 ε )) = f ( x s ( 0 )) + ∫ 0 2 ε ( − 1 ) d t = f ( x s ( 0 )) − 2 ε ≤ c − ε < c ,
a contradiction. Therefore f ( γ ~ ( s ) ) < c f(\widetilde{\gamma}(s))<c f ( γ ( s )) < c for every s s s . The continuous function f ∘ γ ~ f\circ\widetilde{\gamma} f ∘ γ attains a maximum strictly below c c c on [ 0 , 1 ] [0,1] [ 0 , 1 ] , contradicting the definition of c c c . This proves K c ≠ ∅ \mathcal{K}_c\neq\varnothing K c = ∅ .
Every sequence in K c \mathcal{K}_c K c is a P.S. sequence at c c c , and its strongly convergent subsequence has limit in K c \mathcal{K}_c K c by continuity. Thus K c \mathcal{K}_c K c is sequentially compact and hence compact in the Hilbert norm metric.
Step 2: K c ∩ C ‾ ≠ ∅ \mathcal{K}_c\cap\overline{C}\neq\varnothing K c ∩ C = ∅ .
Suppose that K c ∩ C ‾ = ∅ \mathcal{K}_c\cap\overline{C}=\varnothing K c ∩ C = ∅ . The distance between the compact set K c \mathcal{K}_c K c and the closed set C ‾ \overline{C} C is positive. Choose r > 0 r>0 r > 0 such that
dist ( K c , C ‾ ) > 4 r . (21) \operatorname{dist}(\mathcal{K}_c,\overline{C})>4r. \tag*{(21)} dist ( K c , C ) > 4 r . (21)
The P.S. condition again gives η , σ > 0 \eta,\sigma>0 η , σ > 0 , with 2 η < c 2\eta<c 2 η < c , such that
∥ ∇ f ( x ) ∥ ≥ σ if ∣ f ( x ) − c ∣ ≤ 2 η and dist ( x , K c ) ≥ r . (22) \lVert\nabla f(x)\rVert\geq\sigma\quad\text{if}\quad\lvert f(x)-c\rvert\leq2\eta\quad\text{and}\quad\operatorname{dist}(x,\mathcal{K}_c)\geq r. \tag*{(22)} ∥ ∇ f ( x )∥ ≥ σ if ∣ f ( x ) − c ∣ ≤ 2 η and dist ( x , K c ) ≥ r . (22)
Otherwise, it would yield x j ∈ H x_j\in H x j ∈ H with
∣ f ( x j ) − c ∣ ≤ 1 / j , ∥ ∇ f ( x j ) ∥ < 1 / j , dist ( x j , K c ) ≥ r . \lvert f(x_j)-c\rvert\leq1/j,\qquad\lVert\nabla f(x_j)\rVert<1/j,\qquad\operatorname{dist}(x_j,\mathcal{K}_c)\geq r. ∣ f ( x j ) − c ∣ ≤ 1/ j , ∥ ∇ f ( x j )∥ < 1/ j , dist ( x j , K c ) ≥ r .
A P.S. subsequence converges strongly to x ∗ ∈ K c x_*\in\mathcal{K}_c x ∗ ∈ K c , whereas
r ≤ dist ( x j , K c ) ≤ ∥ x j − x ∗ ∥ ⟶ 0 r\leq\operatorname{dist}(x_j,\mathcal{K}_c)\leq\lVert x_j-x_*\rVert\longrightarrow0 r ≤ dist ( x j , K c ) ≤ ∥ x j − x ∗ ∥ ⟶ 0
along that subsequence, a contradiction.
Choose α \alpha α as above, with the present η \eta η , and set d ( x ) = dist ( x , K c ) d(x) = \operatorname{dist}(x, K_c) d ( x ) = dist ( x , K c ) . Define
β ( x ) = { 0 , d ( x ) ≤ r , ( d ( x ) − r ) / r , r < d ( x ) < 2 r , 1 , d ( x ) ≥ 2 r . \beta(x) =
\begin{cases}
0, & d(x) \le r, \\
(d(x)-r)/r, & r < d(x) < 2r, \\
1, & d(x) \ge2r.
\end{cases} β ( x ) = ⎩ ⎨ ⎧ 0 , ( d ( x ) − r ) / r , 1 , d ( x ) ≤ r , r < d ( x ) < 2 r , d ( x ) ≥ 2 r .
Since ∣ d ( x ) − d ( y ) ∣ ≤ ∥ x − y ∥ |d(x)-d(y)| \le\|x-y\| ∣ d ( x ) − d ( y ) ∣ ≤ ∥ x − y ∥ , the function β \beta β has Lipschitz constant 1 / r 1/r 1/ r . In this step, we use (20) with these cutoffs and σ \sigma σ here. Its local Lipschitz regularity, global flow and speed bound 1 / σ 1/\sigma 1/ σ can be established as in Step 1. If α ( f ( x ) ) = β ( x ) = 1 \alpha(f(x)) = \beta(x) = 1 α ( f ( x )) = β ( x ) = 1 , then ∣ f ( x ) − c ∣ ≤ η |f(x)-c| \le\eta ∣ f ( x ) − c ∣ ≤ η and d ( x ) ≥ 2 r d(x) \ge2r d ( x ) ≥ 2 r . Therefore, it follows from (22) that f f f decreases at exactly unit speed.
Choose 0 < ε < min { η , c / 4 } 0 < \varepsilon< \min\{\eta, c/4\} 0 < ε < min { η , c /4 } small enough such that
2 ε / σ < r . (23) 2\varepsilon/\sigma< r. \tag*{(23)} 2 ε / σ < r . (23)
Deform an path γ ∈ Γ \gamma\in\Gamma γ ∈ Γ of height at most c + ε c+\varepsilon c + ε by the flow γ ~ ( s ) = Φ 2 ε ( γ ( s ) ) \widetilde{\gamma}(s) = \Phi_{2\varepsilon}(\gamma(s)) γ ( s ) = Φ 2 ε ( γ ( s )) of vector field X X X for time 2 ε 2\varepsilon 2 ε . Its endpoints will be kept unchanged. We claim
f ( γ ~ ( s ) ) ≥ c ⟹ dist ( γ ~ ( s ) , K c ) < 3 r . (24) f(\widetilde{\gamma}(s)) \ge c \quad\Longrightarrow\quad\operatorname{dist}(\widetilde{\gamma}(s), K_c) < 3r. \tag*{(24)} f ( γ ( s )) ≥ c ⟹ dist ( γ ( s ) , K c ) < 3 r . (24)
Fix s ∈ [ 0 , 1 ] s \in[0,1] s ∈ [ 0 , 1 ] and write
x ( t ) = Φ t ( γ ( s ) ) , ∀ 0 ≤ t ≤ 2 ε , z = x ( 2 ε ) = γ ~ ( s ) . x(t) = \Phi_t(\gamma(s)), \qquad\forall\,0 \le t \le2\varepsilon,\qquad z = x(2\varepsilon) = \widetilde{\gamma}(s). x ( t ) = Φ t ( γ ( s )) , ∀ 0 ≤ t ≤ 2 ε , z = x ( 2 ε ) = γ ( s ) .
In fact, suppose that f ( z ) ≥ c f(z) \ge c f ( z ) ≥ c and dist ( z , K c ) ≥ 3 r \operatorname{dist}(z,K_c) \ge3r dist ( z , K c ) ≥ 3 r . The speed bound implies that, for every t ∈ [ 0 , 2 ε ] t \in[0,2\varepsilon] t ∈ [ 0 , 2 ε ] , by (23) it yields
∥ x ( t ) − z ∥ ≤ ∫ t 2 ε ∥ X ( x ( θ ) ) ∥ d θ ≤ 2 ε − t σ ≤ 2 ε σ < r . \|x(t)-z\| \le\int_t^{2\varepsilon} \|X(x(\theta))\|\,\mathrm{d}\theta\le\frac{2\varepsilon-t}{\sigma} \le\frac{2\varepsilon}{\sigma} < r. ∥ x ( t ) − z ∥ ≤ ∫ t 2 ε ∥ X ( x ( θ )) ∥ d θ ≤ σ 2 ε − t ≤ σ 2 ε < r .
For every p ∈ K c p \in K_c p ∈ K c , the triangle inequality gives ∥ x ( t ) − p ∥ ≥ ∥ z − p ∥ − ∥ x ( t ) − z ∥ \|x(t)-p\| \ge\|z-p\|-\|x(t)-z\| ∥ x ( t ) − p ∥ ≥ ∥ z − p ∥ − ∥ x ( t ) − z ∥ . Taking the infimum over p p p yields
dist ( x ( t ) , K c ) ≥ dist ( z , K c ) − ∥ x ( t ) − z ∥ > 3 r − r = 2 r . \operatorname{dist}(x(t),K_c) \ge\operatorname{dist}(z,K_c)-\|x(t)-z\| > 3r-r = 2r. dist ( x ( t ) , K c ) ≥ dist ( z , K c ) − ∥ x ( t ) − z ∥ > 3 r − r = 2 r .
Therefore β ( x ( t ) ) = 1 \beta(x(t)) = 1 β ( x ( t )) = 1 . The monotonicity also gives
c ≤ f ( z ) ≤ f ( x ( t ) ) ≤ f ( x ( 0 ) ) ≤ c + ε < c + η , c \le f(z) \le f(x(t)) \le f(x(0)) \le c+\varepsilon< c+\eta, c ≤ f ( z ) ≤ f ( x ( t )) ≤ f ( x ( 0 )) ≤ c + ε < c + η ,
so α ( f ( x ( t ) ) ) = 1 \alpha(f(x(t))) = 1 α ( f ( x ( t ))) = 1 and (22) gives ∥ ∇ f ( x ( t ) ) ∥ ≥ σ \|\nabla f(x(t))\| \ge\sigma ∥∇ f ( x ( t )) ∥ ≥ σ . Therefore, on the entire segment, we have
d d t f ( x ( t ) ) = − 1. \frac{\mathrm{d}}{\mathrm{d}t}f(x(t)) = -1. d t d f ( x ( t )) = − 1.
Integrating yields
f ( z ) = f ( γ ( s ) ) − 2 ε ≤ c + ε − 2 ε = c − ε < c , f(z) = f(\gamma(s))-2\varepsilon\le c+\varepsilon-2\varepsilon= c-\varepsilon< c, f ( z ) = f ( γ ( s )) − 2 ε ≤ c + ε − 2 ε = c − ε < c ,
which contradicts to f ( z ) ≥ c f(z) \ge c f ( z ) ≥ c . Thus the claim (24) is proved.
Since γ ~ ∈ Γ \widetilde{\gamma} \in\Gamma γ ∈ Γ , its maximum is at least c c c . As f ( γ ~ ( 0 ) ) = 0 < c f(\widetilde{\gamma}(0)) = 0 < c f ( γ ( 0 )) = 0 < c , the continuity implies that
S = { s ∈ [ 0 , 1 ] : f ( γ ~ ( s ) ) = c } ≠ ∅ . S = \{s \in[0,1] : f(\widetilde{\gamma}(s)) = c\} \ne\varnothing. S = { s ∈ [ 0 , 1 ] : f ( γ ( s )) = c } = ∅ .
It is compact, so s 0 = min S s_0 = \min S s 0 = min S exists, positive and satisfies f ( γ ~ ( s 0 ) ) = c f(\widetilde{\gamma}(s_0)) = c f ( γ ( s 0 )) = c . For s < s 0 s < s_0 s < s 0 one has f ( γ ~ ( s ) ) < c f(\widetilde{\gamma}(s)) < c f ( γ ( s )) < c . For such s s s , we consider the following parametrized deformation
η s ( θ ) = γ ~ ( s θ ) , 0 ≤ θ ≤ 1. \eta_s(\theta) = \widetilde{\gamma}(s\theta), \qquad0 \le\theta\le1. η s ( θ ) = γ ( s θ ) , 0 ≤ θ ≤ 1.
It starts at 0 0 0 , ends at γ ~ ( s ) \tilde{\gamma}(s) γ ~ ( s ) , and stays in { f < c } \{f<c\} { f < c } . Therefore γ ~ ( s ) ∈ C \tilde{\gamma}(s)\in C γ ~ ( s ) ∈ C . Taking s j ↑ s 0 s_j\uparrow s_0 s j ↑ s 0 gives γ ~ ( s 0 ) ∈ C ‾ \tilde{\gamma}(s_0)\in\overline{C} γ ~ ( s 0 ) ∈ C . Estimate (2.16) gives distance less than 3 r 3r 3 r from this point to K c K_c K c , whereas (2.13) makes that distance greater than 4 r 4r 4 r , a contradiction. Thus K c ∩ C ‾ ≠ ∅ K_c\cap\overline{C}\ne\varnothing K c ∩ C = ∅ .
Corollary 2.5. Assume that V V V satisfies ( V 1 ′ ) (V1') ( V 1 ′ ) and ( V 2 ) (V2) ( V 2 ) –( V 5 ) (V5) ( V 5 ) . Then, for every T > 0 T>0 T > 0 , there is a nonconstant classical solution q ∈ S T q\in S_T q ∈ S T of (1.1) such that
D ( ψ T ∣ S T ) ( q ) = 0 , ψ T ( q ) = c T > 0 , q ∈ C T ‾ . (25) D(\psi_T|_{S_T})(q)=0,\qquad\psi_T(q)=c_T>0,\qquad q\in\overline{C_T}. \tag*{(25)} D ( ψ T ∣ S T ) ( q ) = 0 , ψ T ( q ) = c T > 0 , q ∈ C T . (25) Moreover,
C T ∩ { ψ T < 0 } = ∅ . (26) C_T\cap\{\psi_T<0\}=\varnothing. \tag*{(26)} C T ∩ { ψ T < 0 } = ∅ . (26) Proof. Apply Lemma 2.4 to H = S T H=S_T H = S T with its derivative inner product and f = ψ T ∣ S T f=\psi_T|_{S_T} f = ψ T ∣ S T . Lemmas 2.1 and 2.2 verify all assumptions and obtain a nonconstant classical solution. If a negative action point belongs to C T C_T C T , a path within this component from zero to that point would have compact image and maximum strictly below c T c_T c T . It would be an element of Γ T \Gamma_T Γ T , contradicting (1.6).
Local geometry of strict sublevels
For a continuous symmetric bilinear form B B B on a Hilbert space, its radical is defined by
rad ( B ) = { h : B ( h , w ) = 0 for every w } . \operatorname{rad}(B)=\{h:B(h,w)=0\text{ for every }w\}. rad ( B ) = { h : B ( h , w ) = 0 for every w } .
An isotropic vector is one satisfying B ( h , h ) = 0 B(h,h)=0 B ( h , h ) = 0 . Such a vector need not lie in the radical. This distinction will be important for the cutoff velocity fields.
Lemma 2.6. Let H H H be a real Hilbert space, f ∈ C 2 ( H , R ) f\in C^2(H,\mathbb{R}) f ∈ C 2 ( H , R ) , D f ( x 0 ) = 0 Df(x_0)=0 D f ( x 0 ) = 0 , and f ( x 0 ) = c > 0 f(x_0)=c>0 f ( x 0 ) = c > 0 . Write B = D 2 f ( x 0 ) B=D^2f(x_0) B = D 2 f ( x 0 ) .
If B ( v , v ) < 0 B(v,v)<0 B ( v , v ) < 0 for some v ∈ H v\in H v ∈ H , let Y = ker B ( v , ⋅ ) Y=\ker B(v,\cdot) Y = ker B ( v , ⋅ ) . Then there exist open neighborhoods U ⊂ H U\subset H U ⊂ H of x 0 x_0 x 0 and V ⊂ R × Y V\subset\mathbb{R}\times Y V ⊂ R × Y of ( 0 , 0 ) (0,0) ( 0 , 0 ) , and a homeomorphism Θ : U → V \Theta:U\to V Θ : U → V with Θ ( x 0 ) = ( 0 , 0 ) \Theta(x_0)=(0,0) Θ ( x 0 ) = ( 0 , 0 ) , such that
f ~ ( ξ , y ) : = f ( Θ − 1 ( ξ , y ) ) = c + g ( y ) − ξ 2 , (27) \tilde{f}(\xi,y):=f\bigl(\Theta^{-1}(\xi,y)\bigr)=c+g(y)-\xi^2, \tag*{(27)} f ~ ( ξ , y ) := f ( Θ − 1 ( ξ , y ) ) = c + g ( y ) − ξ 2 , (27) where g g g is a continuous function defined near zero in Y Y Y and g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 . There are two fixed local axis points P + , P − P_+,P_- P + , P − and an open neighborhood O O O of x 0 x_0 x 0 such that every point of O ∩ { f < c } O\cap\{f<c\} O ∩ { f < c } can be joined to at least one of P + , P − P_+,P_- P + , P − by a path in { f < c } \{f<c\} { f < c } .
Suppose in addition that B ( h , h ) = 0 B(h,h)=0 B ( h , h ) = 0 and h ∉ rad ( B ) h\notin\operatorname{rad}(B) h ∈ / rad ( B ) . Then every sufficiently nearby point of { f < c } \{f<c\} { f < c } can be joined in { f < c } \{f<c\} { f < c } to a point of negative value if either of the following conditions holds:
(i) There are continuous paths U : [ 0 , s 1 ] → H U:[0,s_1]\to H U : [ 0 , s 1 ] → H and W : [ 0 , s 2 ] → H W:[0,s_2]\to H W : [ 0 , s 2 ] → H , with s 1 , s 2 > 0 s_1,s_2>0 s 1 , s 2 > 0 , starting at x 0 x_0 x 0 , lying in { f < c } \{f<c\} { f < c } for every positive parameter, and ending in { f < 0 } \{f<0\} { f < 0 } , such that
U s = x 0 + s h + o ( s ) , W s = x 0 − d s h + o ( s ) ( s ↓ 0 ) U_s=x_0+sh+o(s),\qquad W_s=x_0-dsh+o(s)\qquad(s\downarrow0) U s = x 0 + s h + o ( s ) , W s = x 0 − d s h + o ( s ) ( s ↓ 0 ) for some $d>0.
(ii) The form B B B has a two-dimensional negative definite subspace, and there is a path U U U with the properties above and tangent h h h .
Proof. Since B ( v , v ) < 0 B(v,v)<0 B ( v , v ) < 0 , for z ∈ H z\in H z ∈ H , we define
p ( z ) = B ( v , z ) B ( v , v ) , y ( z ) = z − p ( z ) v . p(z)=\frac{B(v,z)}{B(v,v)},\qquad y(z)=z-p(z)v. p ( z ) = B ( v , v ) B ( v , z ) , y ( z ) = z − p ( z ) v .
Then B ( v , y ( z ) ) = 0 B(v,y(z))=0 B ( v , y ( z )) = 0 , so y ( z ) ∈ Y y(z)\in Y y ( z ) ∈ Y and z = p ( z ) v + y ( z ) z=p(z)v+y(z) z = p ( z ) v + y ( z ) . If a v ∈ Y av\in Y a v ∈ Y , then 0 = a B ( v , v ) 0=aB(v,v) 0 = a B ( v , v ) , hence a = 0 a=0 a = 0 , and the decomposition is unique. Continuity of B B B makes p p p and y y y to be bounded linear maps, and Y Y Y is closed. Thus H = R v ⊕ Y H=\mathbb{R}v\oplus Y H = R v ⊕ Y is a topological direct sum.
Define the scalar C 1 C^1 C 1 map
F ( t , y ) = D f ( x 0 + t v + y ) [ v ] , ∀ ( t , y ) ∈ R × Y . F(t,y)=Df(x_0+tv+y)[v],\qquad\forall(t,y)\in\mathbb{R}\times Y. F ( t , y ) = D f ( x 0 + t v + y ) [ v ] , ∀ ( t , y ) ∈ R × Y .
At the origin,
F ( 0 , 0 ) = 0 , ∂ t F ( 0 , 0 ) = B ( v , v ) < 0 , D y F ( 0 , 0 ) [ z ] = B ( z , v ) = 0 ( z ∈ Y ) . F(0,0)=0,\qquad\partial_tF(0,0)=B(v,v)<0,\qquad D_yF(0,0)[z]=B(z,v)=0\quad(z\in Y). F ( 0 , 0 ) = 0 , ∂ t F ( 0 , 0 ) = B ( v , v ) < 0 , D y F ( 0 , 0 ) [ z ] = B ( z , v ) = 0 ( z ∈ Y ) .
Therefore, the implicit function theorem gives a C 1 C^1 C 1 function t = φ ( y ) t=\varphi(y) t = φ ( y ) near 0 0 0 in Y Y Y satisfying F ( φ ( y ) , y ) = 0 F(\varphi(y),y)=0 F ( φ ( y ) , y ) = 0 and φ ( 0 ) = 0 \varphi(0)=0 φ ( 0 ) = 0 . Differentiating this identity in an arbitrary direction z ∈ Y z\in Y z ∈ Y at zero, it yields
B ( v , v ) D φ ( 0 ) [ z ] + B ( z , v ) = 0. B(v,v)D\varphi(0)[z]+B(z,v)=0. B ( v , v ) D φ ( 0 ) [ z ] + B ( z , v ) = 0.
Thus D φ ( 0 ) = 0 D\varphi(0)=0 D φ ( 0 ) = 0 since B ( z , v ) = 0 B(z,v)=0 B ( z , v ) = 0 for z ∈ Y z\in Y z ∈ Y . In particular, the Fréchet differentiability gives φ ( y ) = o ( ∥ y ∥ ) \varphi(y)=o(\lVert y\rVert) φ ( y ) = o (∥ y ∥) as y → 0 y\to0 y → 0 in Y Y Y .
For y y y near zero, set
F y ( t ) = f ( x 0 + ( φ ( y ) + t ) v + y ) , g ( y ) = F y ( 0 ) − c . F_y(t)=f(x_0+(\varphi(y)+t)v+y),\qquad g(y)=F_y(0)-c. F y ( t ) = f ( x 0 + ( φ ( y ) + t ) v + y ) , g ( y ) = F y ( 0 ) − c .
The implicit equation gives F y ′ ( 0 ) = 0 F_y'(0)=0 F y ′ ( 0 ) = 0 . By continuity of D 2 f D^2f D 2 f and B ( v , v ) < 0 B(v,v)<0 B ( v , v ) < 0 , there exist λ , Λ > 0 \lambda,\Lambda>0 λ , Λ > 0 and a product neighborhood in ( t , y ) (t,y) ( t , y ) on which λ ≤ − F y ′ ′ ( t ) ≤ Λ \lambda\leq-F_y''(t)\leq\Lambda λ ≤ − F y ′′ ( t ) ≤ Λ . The Taylor’s formula with Cauchy integral remainder, for both positive and negative t t t , gives
F y ( t ) = c + g ( y ) − a ( t , y ) t 2 (28) F_y(t)=c+g(y)-a(t,y)t^2 \tag*{(28)} F y ( t ) = c + g ( y ) − a ( t , y ) t 2 (28)
where
a ( t , y ) = − ∫ 0 1 ( 1 − θ ) D 2 f ( x 0 + ( φ ( y ) + θ t ) v + y ) [ v , v ] d θ . a(t,y)=-\int_0^1(1-\theta)D^2f(x_0+(\varphi(y)+\theta t)v+y)[v,v]\,d\theta. a ( t , y ) = − ∫ 0 1 ( 1 − θ ) D 2 f ( x 0 + ( φ ( y ) + θ t ) v + y ) [ v , v ] d θ .
The function a a a is continuous and satisfies λ / 2 ≤ a ( t , y ) ≤ Λ / 2 \lambda/2\leq a(t,y)\leq\Lambda/2 λ /2 ≤ a ( t , y ) ≤ Λ/2 . Define
χ ( t , y ) = t a ( t , y ) , ξ = χ ( t , y ) . \chi(t,y)=t\sqrt{a(t,y)},\qquad\xi=\chi(t,y). χ ( t , y ) = t a ( t , y ) , ξ = χ ( t , y ) .
For fixed y y y , the strict negativity of F y ′ ′ F_y'' F y ′′ and F y ′ ( 0 ) = 0 F_y'(0)=0 F y ′ ( 0 ) = 0 show that F y ′ ( t ) > 0 F_y'(t)>0 F y ′ ( t ) > 0 for t < 0 t<0 t < 0 and F y ′ ( t ) < 0 F_y'(t)<0 F y ′ ( t ) < 0 for t > 0 t>0 t > 0 . So the two branches − F y ( 0 ) − F y ( t ) -\sqrt{F_y(0)-F_y(t)} − F y ( 0 ) − F y ( t ) for t < 0 t<0 t < 0 and F y ( 0 ) − F y ( t ) \sqrt{F_y(0)-F_y(t)} F y ( 0 ) − F y ( t ) for t > 0 t>0 t > 0 are strictly increasing. They meet at χ ( 0 , y ) = 0 \chi(0,y)=0 χ ( 0 , y ) = 0 , so t ↦ χ ( t , y ) t\mapsto\chi(t,y) t ↦ χ ( t , y ) is strictly increasing on the whole interval.
For fixed y y y , next we verify the uniform bound on ∂ t χ ( t , y ) \partial_t\chi(t,y) ∂ t χ ( t , y ) . Here t t t is a scalar displacement from ( x 0 + φ ( y ) v + y ) (x_0+\varphi(y)v+y) ( x 0 + φ ( y ) v + y ) in the direction v v v . For this derivative,
∂ t χ ( 0 , y ) = lim t → 0 χ ( t , y ) − χ ( 0 , y ) t = lim t → 0 a ( t , y ) = a ( 0 , y ) . \partial_t\chi(0,y)=\lim_{t\to0}\frac{\chi(t,y)-\chi(0,y)}{t}=\lim_{t\to0}\sqrt{a(t,y)}=\sqrt{a(0,y)}. ∂ t χ ( 0 , y ) = t → 0 lim t χ ( t , y ) − χ ( 0 , y ) = t → 0 lim a ( t , y ) = a ( 0 , y ) .
a ( 0 , y ) = − D 2 f ( x 0 + φ ( y ) v + y ) [ v , v ] ∫ 0 1 ( 1 − θ ) d θ = − 1 2 D 2 f ( x 0 + φ ( y ) v + y ) [ v , v ] . \begin{aligned}
a(0,y)=-D^2f(x_0+\varphi(y)v+y)[v,v]\int_0^1(1-\theta)\,d\theta\\
&=-\frac{1}{2}D^2f(x_0+\varphi(y)v+y)[v,v].
\end{aligned} a ( 0 , y ) = − D 2 f ( x 0 + φ ( y ) v + y ) [ v , v ] ∫ 0 1 ( 1 − θ ) d θ = − 2 1 D 2 f ( x 0 + φ ( y ) v + y ) [ v , v ] .
This proves the positive value without differentiating the merely continuous function a a a . The derivative F y ′ ( 0 ) F_y'(0) F y ′ ( 0 ) in the v v v direction at ( x 0 + φ ( y ) v + y ) (x_0+\varphi(y)v+y) ( x 0 + φ ( y ) v + y ) is zero.
Let b ( t , y ) = − F y ′ ′ ( t ) b(t,y)=-F_y''(t) b ( t , y ) = − F y ′′ ( t ) , and then λ ≤ b ≤ Λ \lambda\leq b\leq\Lambda λ ≤ b ≤ Λ . Note that χ ( t , y ) 2 = g ( y ) − F y ( t ) + c \chi(t,y)^2=g(y)-F_y(t)+c χ ( t , y ) 2 = g ( y ) − F y ( t ) + c , we have
F y ′ ( t ) = − t ∫ 0 1 b ( θ t , y ) d θ , ∂ t χ ( t , y ) = ∫ 0 1 b ( θ t , y ) d θ 2 a ( t , y ) , t ≠ 0 , F_y'(t)=-t\int_0^1 b(\theta t,y)\,d\theta,\qquad\partial_t\chi(t,y)=\frac{\int_0^1 b(\theta t,y)\,d\theta}{2\sqrt{a(t,y)}},\qquad t\ne0, F y ′ ( t ) = − t ∫ 0 1 b ( θ t , y ) d θ , ∂ t χ ( t , y ) = 2 a ( t , y ) ∫ 0 1 b ( θ t , y ) d θ , t = 0 ,
which, together with the value at zero, these formulas imply, throughout a small coordinate neighborhood,
0 < λ 2 Λ / 2 ≤ ∂ t χ ( t , y ) ≤ Λ 2 λ / 2 < ∞ . 0<\frac{\lambda}{2\sqrt{\Lambda/2}}\leq\partial_t\chi(t,y)\leq\frac{\Lambda}{2\sqrt{\lambda/2}}<\infty. 0 < 2 Λ/2 λ ≤ ∂ t χ ( t , y ) ≤ 2 λ /2 Λ < ∞.
The lower bound supplies the uniform Lipschitz bound for the scalar inverse.
Choose δ > 0 \delta>0 δ > 0 and r ∗ > 0 r_* >0 r ∗ > 0 such that the preceding estimates hold for ∣ t ∣ ≤ δ |t|\leq\delta ∣ t ∣ ≤ δ and y ∈ B Y ( 0 , r ∗ ) y\in B_Y(0,r_*) y ∈ B Y ( 0 , r ∗ ) . Since a ( t , y ) ≥ λ / 2 > 0 a(t,y)\geq\lambda/2>0 a ( t , y ) ≥ λ /2 > 0 , we obtain
χ ( δ , y ) ≥ δ λ / 2 , χ ( − δ , y ) ≤ − δ λ / 2 . \chi(\delta,y)\geq\delta\sqrt{\lambda/2},\qquad\chi(-\delta,y)\leq-\delta\sqrt{\lambda/2}. χ ( δ , y ) ≥ δ λ /2 , χ ( − δ , y ) ≤ − δ λ /2 .
Fix 0 < ρ ∗ < δ λ / 2 0<\rho_*<\delta\sqrt{\lambda/2} 0 < ρ ∗ < δ λ /2 . Strict monotonicity and the intermediate value theorem give, for each ( ξ , y ) ∈ ( − ρ ∗ , ρ ∗ ) × B Y ( 0 , r ∗ ) (\xi,y)\in(-\rho_*,\rho_*)\times B_Y(0,r_*) ( ξ , y ) ∈ ( − ρ ∗ , ρ ∗ ) × B Y ( 0 , r ∗ ) , a unique t = ϑ ( ξ , y ) ∈ ( − δ , δ ) t=\vartheta(\xi,y)\in(-\delta,\delta) t = ϑ ( ξ , y ) ∈ ( − δ , δ ) with χ ( t , y ) = ξ \chi(t,y)=\xi χ ( t , y ) = ξ .
A product neighborhood here is exactly the Cartesian product
V 0 = ( − ρ ∗ , ρ ∗ ) × B Y ( 0 , r ∗ ) . V_0=(-\rho_*,\rho_*)\times B_Y(0,r_*). V 0 = ( − ρ ∗ , ρ ∗ ) × B Y ( 0 , r ∗ ) .
It is open in R × Y \mathbb{R}\times Y R × Y , and the allowed interval of ξ \xi ξ is the same for every y y y in the indicated ball.
Let m ∗ > 0 m_*>0 m ∗ > 0 be the uniform lower bound for ∂ t χ \partial_t\chi ∂ t χ . The mean value theorem gives
∣ ϑ ( ξ 1 , y ) − ϑ ( ξ 2 , y ) ∣ ≤ m ∗ − 1 ∣ ξ 1 − ξ 2 ∣ . |\vartheta(\xi_1,y)-\vartheta(\xi_2,y)|\leq m_*^{-1}|\xi_1-\xi_2|. ∣ ϑ ( ξ 1 , y ) − ϑ ( ξ 2 , y ) ∣ ≤ m ∗ − 1 ∣ ξ 1 − ξ 2 ∣.
If ( ξ j , y j ) → ( ξ , y ) (\xi_j,y_j)\to(\xi,y) ( ξ j , y j ) → ( ξ , y ) , set t ∗ = ϑ ( ξ , y ) t_*=\vartheta(\xi,y) t ∗ = ϑ ( ξ , y ) . For all large j j j ,
∣ ϑ ( ξ j , y j ) − t ∗ ∣ ≤ m ∗ − 1 ∣ ξ j − χ ( t ∗ , y j ) ∣ ⟶ 0 , |\vartheta(\xi_j,y_j)-t_*|\leq m_*^{-1}|\xi_j-\chi(t_*,y_j)|\longrightarrow0, ∣ ϑ ( ξ j , y j ) − t ∗ ∣ ≤ m ∗ − 1 ∣ ξ j − χ ( t ∗ , y j ) ∣ ⟶ 0 ,
by continuity of χ \chi χ . This proves the continuity of ϑ \vartheta ϑ . The inverse chart in the Hilbert space is
Θ − 1 ( ξ , y ) = x 0 + ( φ ( y ) + ϑ ( ξ , y ) ) v + y . \Theta^{-1}(\xi,y)=x_0+(\varphi(y)+\vartheta(\xi,y))v+y. Θ − 1 ( ξ , y ) = x 0 + ( φ ( y ) + ϑ ( ξ , y )) v + y .
Its inverse uses the bounded projection p p p , the map y = x − x 0 − p ( x − x 0 ) v y=x-x_0-p(x-x_0)v y = x − x 0 − p ( x − x 0 ) v , and ξ = χ ( p ( x − x 0 ) − φ ( y ) , y ) \xi=\chi(p(x-x_0)-\varphi(y),y) ξ = χ ( p ( x − x 0 ) − φ ( y ) , y ) , and these are continuous. Thus it gives a homeomorphic chart on the open neighborhood U 0 = Θ − 1 ( V 0 ) U_0=\Theta^{-1}(V_0) U 0 = Θ − 1 ( V 0 ) of x 0 x_0 x 0 .
Choose 0 < 2 d 0 < ρ ∗ 0<2d_0<\rho_* 0 < 2 d 0 < ρ ∗ and then 0 < r 1 < r ∗ 0<r_1<r_* 0 < r 1 < r ∗ so small that ∣ g ( y ) ∣ < d 0 2 / 4 |g(y)|<d_0^2/4 ∣ g ( y ) ∣ < d 0 2 /4 for ∥ y ∥ < r 1 \|y\|<r_1 ∥ y ∥ < r 1 . Define
O = Θ − 1 ( ( − d 0 / 2 , d 0 / 2 ) × B Y ( 0 , r 1 ) ) , P ± = Θ − 1 ( ± d 0 , 0 ) . O=\Theta^{-1}\left((-d_0/2,d_0/2)\times B_Y(0,r_1)\right),\qquad P_\pm=\Theta^{-1}(\pm d_0,0). O = Θ − 1 ( ( − d 0 /2 , d 0 /2 ) × B Y ( 0 , r 1 ) ) , P ± = Θ − 1 ( ± d 0 , 0 ) .
Then O O O is an open neighborhood of x 0 x_0 x 0 in the larger coordinate neighborhood U 0 U_0 U 0 . The connecting paths below are constructed in U 0 U_0 U 0 and may leave O O O .
Let x ∈ O x \in O x ∈ O have coordinates ( ξ , y ) (\xi,y) ( ξ , y ) and satisfy g ( y ) − ξ 2 < 0 g(y)-\xi^2<0 g ( y ) − ξ 2 < 0 . Choose e ∈ { 1 , − 1 } e \in\{1,-1\} e ∈ { 1 , − 1 } to agree with the sign of ξ \xi ξ ; when ξ = 0 \xi=0 ξ = 0 choose either sign, since then g ( y ) < 0 g(y)<0 g ( y ) < 0 . The first path in coordinates is
( ( 1 − θ ) ξ + θ e d 0 , y ) , 0 ≤ θ ≤ 1. ((1-\theta)\xi+\theta e d_0,y),\qquad0\leq\theta\leq1. (( 1 − θ ) ξ + θ e d 0 , y ) , 0 ≤ θ ≤ 1.
Its scalar coordinate has nondecreasing absolute value. Hence its value of f − c f-c f − c is at most g ( y ) − ξ 2 < 0 g(y)-\xi^2<0 g ( y ) − ξ 2 < 0 . The second path is
( e d 0 , ( 1 − θ ) y ) , 0 ≤ θ ≤ 1. (e d_0,(1-\theta)y),\qquad0\leq\theta\leq1. ( e d 0 , ( 1 − θ ) y ) , 0 ≤ θ ≤ 1.
The ball in Y Y Y is convex, so this path has f − c = g ( ( 1 − θ ) y ) − d 0 2 < d 0 2 / 4 − d 0 2 < 0 f-c=g((1-\theta)y)-d_0^2<d_0^2/4-d_0^2<0 f − c = g (( 1 − θ ) y ) − d 0 2 < d 0 2 /4 − d 0 2 < 0 . Composing both paths with Θ − 1 \Theta^{-1} Θ − 1 , and the concatenation joins x x x to P e P_e P e within { f < c } \{f<c\} { f < c } . Since φ ( 0 ) = 0 \varphi(0)=0 φ ( 0 ) = 0 and ϑ ( ξ , 0 ) \vartheta(\xi,0) ϑ ( ξ , 0 ) has the sign of ξ \xi ξ , the points P + P_+ P + and P − P_- P − lie on opposite v v v rays in H H H .
Since h ∉ rad ( B ) h\notin\operatorname{rad}(B) h ∈ / rad ( B ) , choose w w w with B ( h , w ) ≠ 0 B(h,w)\ne0 B ( h , w ) = 0 ; changing its sign makes B ( h , w ) > 0 B(h,w)>0 B ( h , w ) > 0 . Put v = h − ϵ w v=h-\epsilon w v = h − ϵ w . For sufficiently small ϵ > 0 \epsilon>0 ϵ > 0 , isotropy of h h h gives
B ( v , v ) = − 2 ϵ B ( h , w ) + ϵ 2 B ( w , w ) < 0 , B ( v , h ) = − ϵ B ( w , h ) ≠ 0. (29) \begin{aligned}
B(v,v)&=-2\epsilon B(h,w)+\epsilon^2B(w,w)<0,\\
B(v,h)&=-\epsilon B(w,h)\ne0. \tag*{(29)}
\end{aligned} B ( v , v ) B ( v , h ) = − 2 ϵ B ( h , w ) + ϵ 2 B ( w , w ) < 0 , = − ϵ B ( w , h ) = 0. (29)
Use the above chart for this v v v . Let x s = x 0 + s h + r s x_s=x_0+sh+r_s x s = x 0 + s h + r s with ∥ r s ∥ = o ( s ) \lVert r_s\rVert=o(s) ∥ r s ∥ = o ( s ) , we define
p s = B ( v , x s − x 0 ) B ( v , v ) , y s = x s − x 0 − p s v , t s = p s − φ ( y s ) . p_s=\frac{B(v,x_s-x_0)}{B(v,v)},\qquad y_s=x_s-x_0-p_sv,\qquad t_s=p_s-\varphi(y_s). p s = B ( v , v ) B ( v , x s − x 0 ) , y s = x s − x 0 − p s v , t s = p s − φ ( y s ) .
The bounded projections give
p s = B ( v , h ) B ( v , v ) s + o ( s ) , ∥ y s ∥ = O ( s ) . p_s=\frac{B(v,h)}{B(v,v)}s+o(s),\qquad\lVert y_s\rVert=O(s). p s = B ( v , v ) B ( v , h ) s + o ( s ) , ∥ y s ∥ = O ( s ) .
Since φ ( 0 ) = 0 \varphi(0)=0 φ ( 0 ) = 0 and D φ ( 0 ) = 0 D\varphi(0)=0 D φ ( 0 ) = 0 , one has φ ( y s ) = o ( s ) \varphi(y_s)=o(s) φ ( y s ) = o ( s ) . Therefore t s = s B ( v , h ) / B ( v , v ) + o ( s ) t_s=sB(v,h)/B(v,v)+o(s) t s = s B ( v , h ) / B ( v , v ) + o ( s ) . Also a ( t s , y s ) → a ( 0 , 0 ) = − B ( v , v ) / 2 > 0 a(t_s,y_s)\to a(0,0)=-B(v,v)/2>0 a ( t s , y s ) → a ( 0 , 0 ) = − B ( v , v ) /2 > 0 . It follows that
ξ ( s ) = χ ( t s , y s ) = t s a ( t s , y s ) = − B ( v , v ) 2 B ( v , h ) B ( v , v ) s + o ( s ) . (30) \xi(s)=\chi(t_s,y_s)=t_s\sqrt{a(t_s,y_s)}=\sqrt{\frac{-B(v,v)}{2}}\frac{B(v,h)}{B(v,v)}s+o(s). \tag*{(30)} ξ ( s ) = χ ( t s , y s ) = t s a ( t s , y s ) = 2 − B ( v , v ) B ( v , v ) B ( v , h ) s + o ( s ) . (30)
It follows from (29) that the above coefficient is nonzero. For a path tangent to − d h -dh − d h , the coefficient is multiplied by − d -d − d .
In case (i), (30) shows that the ξ \xi ξ coordinates of U s U_s U s and W s W_s W s have opposite signs for all sufficiently small positive parameters. Choose s U , s W > 0 s_U,s_W>0 s U , s W > 0 so small that U s U , W s W ∈ O U_{s_U},W_{s_W}\in O U s U , W s W ∈ O and their scalar coordinates have opposite signs. The constructed axis paths join U s U U_{s_U} U s U to one of P + P_+ P + and W s W W_{s_W} W s W to the other. Starting at an axis, reverse the appropriate connection and follow U ∣ [ s U , s 1 ] U|_{[s_U,s_1]} U ∣ [ s U , s 1 ] or W ∣ [ s W , s 2 ] W|_{[s_W,s_2]} W ∣ [ s W , s 2 ] to its negative value endpoint. Every parameter used on these paths is positive, so the concatenations always stay in { f < c } \{f<c\} { f < c } and do not pass through x 0 x_0 x 0 . So each point of O ∩ { f < c } O\cap\{f<c\} O ∩ { f < c } connects to a point of negative value below c c c .
In case (ii), let N N N be a two-dimensional negative definite subspace. The linear map B ( v , ⋅ ) : N → R B(v,\mathord\cdot):N\to\mathbb{R} B ( v , ⋅ ) : N → R has a nonzero kernel vector z z z . Thus z ∈ Y z\in Y z ∈ Y , B ( z , z ) < 0 B(z,z)<0 B ( z , z ) < 0 , and B ( v , z ) = 0 B(v,z)=0 B ( v , z ) = 0 . Let w θ = cos θ v + sin θ z w_\theta=\cos\theta\,v+\sin\theta\,z w θ = cos θ v + sin θ z , it yields
B ( w θ , w θ ) = cos 2 θ B ( v , v ) + sin 2 θ B ( z , z ) ≤ − κ , B(w_\theta,w_\theta)=\cos^2\theta B(v,v)+\sin^2\theta B(z,z)\leq-\kappa, B ( w θ , w θ ) = cos 2 θ B ( v , v ) + sin 2 θ B ( z , z ) ≤ − κ ,
where κ = min { − B ( v , v ) , − B ( z , z ) } > 0 \kappa=\min\{-B(v,v),-B(z,z)\}>0 κ = min { − B ( v , v ) , − B ( z , z )} > 0 . The set { w θ : 0 ≤ θ ≤ π } \{w_\theta:0\leq\theta\leq\pi\} { w θ : 0 ≤ θ ≤ π } is compact and bounded. By the continuity of D 2 f D^2f D 2 f at x 0 x_0 x 0 and D f ( x 0 ) = 0 Df(x_0)=0 D f ( x 0 ) = 0 , we obtain
f ( x 0 + r w θ ) = c + 1 2 r 2 B ( w θ , w θ ) + o ( r 2 ) , f(x_0+rw_\theta)=c+\frac{1}{2}r^2B(w_\theta,w_\theta)+o(r^2), f ( x 0 + r w θ ) = c + 2 1 r 2 B ( w θ , w θ ) + o ( r 2 ) ,
where the remainder is uniform in θ \theta θ . Therefore, for small r > 0 r>0 r > 0 , the whole curve satisfies f ≤ c − κ r 2 / 4 < c f\leq c-\kappa r^2/4<c f ≤ c − κ r 2 /4 < c and joins the two v v v rays. It follow from the strict negativeness of B ( v , v ) B(v,v) B ( v , v ) that small nonzero segments of these rays also lie below c c c . Hence P + P_+ P + and P − P_- P − are connected below c c c . A small positive point of U U U connects to one axis, and its positive-parameter path ends at negative value. Thus both axes, and hence every sufficiently nearby strict sublevel point, connect to negative value below c c c . This proves case (ii) and the lemma. □
Negative directions along odd iterates
The next lemma gives the part of the cutoff Jacobi field construction needed here. It is consistent with the odd iteration estimates of Long [8 ] , but the required negative directions are constructed directly.
Lemma 2.7. Let V ∈ C 2 ( R n , R ) V\in C^2(\mathbb{R}^n,\mathbb{R}) V ∈ C 2 ( R n , R ) be even with V ( 0 ) = 0 V(0)=0 V ( 0 ) = 0 , and let q ∈ S T q\in S_T q ∈ S T be a nonconstant classical solution of (1.1). Suppose its minimal period is τ = T / k \tau=T/k τ = T / k , where k = 2 m + 1 ≥ 3 k=2m+1\geq3 k = 2 m + 1 ≥ 3 . Set a = τ / 4 a=\tau/4 a = τ /4 and L = k a L=ka L = k a . On X L X_L X L , let
B ( u , w ) = ∫ 0 L ( ⟨ u ′ , w ′ ⟩ − ⟨ D 2 V ( q ( t ) ) u , w ⟩ ) d t . (31) B(u,w)=\int_0^L\left(\langle u',w'\rangle-\langle D^2V(q(t))u,w\rangle\right)\,dt. \tag*{(31)} B ( u , w ) = ∫ 0 L ( ⟨ u ′ , w ′ ⟩ − ⟨ D 2 V ( q ( t )) u , w ⟩ ) d t . (31) Then B B B has a negative definite subspace of dimension m m m . In particular it has a negative two-plane if k ≥ 5 k\geq5 k ≥ 5 . The parameters satisfy
a = T 4 k , L = T 4 , b = 2 m a = m T 2 k = L − a ∈ ( 0 , L ) . a=\frac{T}{4k},\qquad L=\frac{T}{4},\qquad b=2ma=\frac{mT}{2k}=L-a\in(0,L). a = 4 k T , L = 4 T , b = 2 ma = 2 k m T = L − a ∈ ( 0 , L ) . The field
h ( t ) = { q ′ ( t ) , 0 ≤ t ≤ b , 0 , b ≤ t ≤ L (32) h(t)=
\begin{cases}
q'(t), & 0\leq t\leq b,\\
0, & b\leq t\leq L
\end{cases} \tag*{(32)} h ( t ) = { q ′ ( t ) , 0 , 0 ≤ t ≤ b , b ≤ t ≤ L (32) belongs to X L X_L X L , is isotropic, and is not in rad ( B ) \operatorname{rad}(B) rad ( B ) .
Proof. By Lemma 2.3, it follows q ∈ S τ q\in S_\tau q ∈ S τ . Set t j = 2 j a t_j=2ja t j = 2 j a for 0 ≤ j ≤ m 0\leq j\leq m 0 ≤ j ≤ m . Then
q ( t j ) = ( − 1 ) j q ( 0 ) , q ′ ( t j ) = 0 , q ¨ ( t j ) ≠ 0. (33) q(t_j)=(-1)^jq(0),\qquad q'(t_j)=0,\qquad\ddot q(t_j)\ne0. \tag*{(33)} q ( t j ) = ( − 1 ) j q ( 0 ) , q ′ ( t j ) = 0 , q ¨ ( t j ) = 0. (33)
Indeed, if both velocity and acceleration vanished at a turning point, its position would be a critical point of V V V . The uniqueness would make the solution to be constant. Since q q q is C 3 C^3 C 3 from Lemma 2.2, its velocity y = q ′ y=q' y = q ′ satisfies
y ′ ′ + D 2 V ( q ) y = 0. (34) y''+D^2V(q)y=0. \tag*{(34)} y ′′ + D 2 V ( q ) y = 0. (34)
For 1 ≤ j ≤ m 1 \le j \le m 1 ≤ j ≤ m , let h j h_j h j equal q ′ q' q ′ on [ t j − 1 , t j ] [t_{j-1},t_j] [ t j − 1 , t j ] and zero elsewhere on [ 0 , L ] [0,L] [ 0 , L ] . Since q ′ ( t j − 1 ) = q ′ ( t j ) = 0 q'(t_{j-1})=q'(t_j)=0 q ′ ( t j − 1 ) = q ′ ( t j ) = 0 , its values match at both endpoints and h j ∈ X L h_j \in X_L h j ∈ X L . Its weak derivative equals q ¨ \ddot{q} q ¨ on ( t j − 1 , t j ) (t_{j-1},t_j) ( t j − 1 , t j ) and zero elsewhere. Then the integration by parts and (34) yield
B ( h j , w ) = ∫ t j − 1 t j ( ⟨ q ¨ , w ′ ⟩ − ⟨ D 2 V ( q ) q ′ , w ⟩ ) d t = [ ⟨ q ¨ ( t ) , w ( t ) ⟩ ] t j − 1 t j − ∫ t j − 1 t j ⟨ d d t q ¨ + D 2 V ( q ) q ′ , w ⟩ d t = [ ⟨ q ¨ ( t ) , w ( t ) ⟩ ] t j − 1 t j , ∀ w ∈ X L . \begin{aligned}
B(h_j,w) &= \int_{t_{j-1}}^{t_j}\left(\langle\ddot{q},w'\rangle-\langle D^2V(q)q',w\rangle\right)\,\mathrm{d}t \\
&= \left[\langle\ddot{q}(t),w(t)\rangle\right]_{t_{j-1}}^{t_j}-\int_{t_{j-1}}^{t_j}\left\langle\frac{\mathrm{d}}{\mathrm{d}t}\ddot{q}+D^2V(q)q',w\right\rangle\,\mathrm{d}t \\
&= \left[\langle\ddot{q}(t),w(t)\rangle\right]_{t_{j-1}}^{t_j},\qquad\forall w\in X_L.
\end{aligned} B ( h j , w ) = ∫ t j − 1 t j ( ⟨ q ¨ , w ′ ⟩ − ⟨ D 2 V ( q ) q ′ , w ⟩ ) d t = [ ⟨ q ¨ ( t ) , w ( t )⟩ ] t j − 1 t j − ∫ t j − 1 t j ⟨ d t d q ¨ + D 2 V ( q ) q ′ , w ⟩ d t = [ ⟨ q ¨ ( t ) , w ( t )⟩ ] t j − 1 t j , ∀ w ∈ X L .
Taking w = h j w=h_j w = h j gives zero because h j h_j h j vanishes at the two nodes. For i ≠ j i\ne j i = j , the supports of h i , h j h_i,h_j h i , h j and their derivatives have disjoint interiors, so we obtain
B ( h i , h j ) = 0 ( 1 ≤ i , j ≤ m ) , (35) B(h_i,h_j)=0\qquad(1\le i,j\le m), \tag*{(35)} B ( h i , h j ) = 0 ( 1 ≤ i , j ≤ m ) , (35)
B ( h j , w ) = ⟨ q ¨ ( t j ) , w ( t j ) ⟩ − ⟨ q ¨ ( t j − 1 ) , w ( t j − 1 ) ⟩ . (36) B(h_j,w)=\langle\ddot{q}(t_j),w(t_j)\rangle-\langle\ddot{q}(t_{j-1}),w(t_{j-1})\rangle. \tag*{(36)} B ( h j , w ) = ⟨ q ¨ ( t j ) , w ( t j )⟩ − ⟨ q ¨ ( t j − 1 ) , w ( t j − 1 )⟩ . (36)
We claim that the functionals { B ( h j , ⋅ ) } j = 1 m \{B(h_j,\cdot)\}_{j=1}^m { B ( h j , ⋅ ) } j = 1 m are linearly independent. In fact, suppose ∑ j = 1 m λ j B ( h j , w ) = 0 \sum_{j=1}^m\lambda_jB(h_j,w)=0 ∑ j = 1 m λ j B ( h j , w ) = 0 for every w ∈ X L w\in X_L w ∈ X L and some nonzero coefficient. Let J J J be the largest index with λ J ≠ 0 \lambda_J\ne0 λ J = 0 . Choose a smooth scalar function χ \chi χ supported in a sufficiently small interval about the interior point t J t_J t J , with χ ( t J ) = 1 \chi(t_J)=1 χ ( t J ) = 1 , whose support contains no other node and does not meet L L L . Set w ( t ) = χ ( t ) q ¨ ( t J ) w(t)=\chi(t)\ddot{q}(t_J) w ( t ) = χ ( t ) q ¨ ( t J ) . Then w ∈ X L w\in X_L w ∈ X L and
w ( t J ) = q ¨ ( t J ) , w ( t i ) = 0 , ∀ i ≠ J . w(t_J)=\ddot{q}(t_J),\qquad w(t_i)=0,\ \forall i\ne J. w ( t J ) = q ¨ ( t J ) , w ( t i ) = 0 , ∀ i = J .
Then the terms with j < J j<J j < J vanish by their endpoint pairings, and coefficients with j > J j>J j > J are zero. Therefore, formula (36) gives λ J ∣ q ¨ ( t J ) ∣ 2 = 0 \lambda_J|\ddot{q}(t_J)|^2=0 λ J ∣ q ¨ ( t J ) ∣ 2 = 0 . This contradicts λ J ≠ 0 \lambda_J\ne0 λ J = 0 and (33) .
Now we define
A : X L ⟶ R m , A ( w ) = ( B ( h 1 , w ) , … , B ( h m , w ) ) . A:X_L\longrightarrow\mathbb{R}^m,\qquad A(w)=\bigl(B(h_1,w),\ldots,B(h_m,w)\bigr). A : X L ⟶ R m , A ( w ) = ( B ( h 1 , w ) , … , B ( h m , w ) ) .
and let M = A ( X L ) M=A(X_L) M = A ( X L ) . If M M M is a proper subspace of R m \mathbb{R}^m R m , there exists a nonzero vector λ = ( λ 1 , … , λ m ) ∈ M ⊥ \lambda=(\lambda_1,\ldots,\lambda_m)\in M^\perp λ = ( λ 1 , … , λ m ) ∈ M ⊥ . Then the functional ℓ λ ( z ) = ∑ i = 1 m λ i z i \ell_\lambda(z)=\sum_{i=1}^m\lambda_i z_i ℓ λ ( z ) = ∑ i = 1 m λ i z i ( ∀ z ∈ R m ) (\forall z\in\mathbb{R}^m) ( ∀ z ∈ R m ) vanishes on M M M . In particular, ℓ λ ( A ( w ) ) = ∑ i = 1 m λ i B ( h i , w ) = 0 \ell_\lambda(A(w))=\sum_{i=1}^m\lambda_iB(h_i,w)=0 ℓ λ ( A ( w )) = ∑ i = 1 m λ i B ( h i , w ) = 0 for every w ∈ X L w\in X_L w ∈ X L , which implies that { B ( h j , ⋅ ) } j = 1 m \{B(h_j,\cdot)\}_{j=1}^m { B ( h j , ⋅ ) } j = 1 m are linearly dependent. This contradiction shows that M = R m M=\mathbb{R}^m M = R m and A A A is surjective.
For the standard basis { e j } j = 1 m \{e_j\}_{j=1}^m { e j } j = 1 m of R m \mathbb{R}^m R m , choose w j ∈ X L w_j\in X_L w j ∈ X L with A ( w j ) = e j A(w_j)=e_j A ( w j ) = e j ; this gives B ( h i , w j ) = δ i j B(h_i,w_j)=\delta_{ij} B ( h i , w j ) = δ ij exactly. Let G = ( G i j ) G=(G_{ij}) G = ( G ij ) with G i j = B ( w i , w j ) G_{ij}=B(w_i,w_j) G ij = B ( w i , w j ) . Then it follows from (35) that
B ( ∑ j α j ( h j − ϵ w j ) , ∑ j α j ( h j − ϵ w j ) ) = − 2 ϵ ∣ α ∣ 2 + ϵ 2 α T G α . B\left(\sum_j\alpha_j(h_j-\epsilon w_j),\sum_j\alpha_j(h_j-\epsilon w_j)\right)=-2\epsilon|\alpha|^2+\epsilon^2\alpha^{T}G\alpha. B ( j ∑ α j ( h j − ϵ w j ) , j ∑ α j ( h j − ϵ w j ) ) = − 2 ϵ ∣ α ∣ 2 + ϵ 2 α T G α .
For sufficiently small ϵ > 0 \epsilon>0 ϵ > 0 , this is strictly negative if α ≠ 0 \alpha\ne0 α = 0 . Therefore, the vectors ( h j − ϵ w j ) (h_j-\epsilon w_j) ( h j − ϵ w j ) are linearly independent, and their span is a negative definite subspace of dimension m m m .
So we have h = ∑ j = 1 m h j ∈ X L h=\sum_{j=1}^m h_j\in X_L h = ∑ j = 1 m h j ∈ X L . Then it follows from (35) and (36) that
B ( h , h ) = 0 , B ( h , w ) = ⟨ q ¨ ( b ) , w ( b ) ⟩ − ⟨ q ¨ ( 0 ) , w ( 0 ) ⟩ . (37) B(h,h)=0,\qquad B(h,w)=\langle\ddot{q}(b),w(b)\rangle-\langle\ddot{q}(0),w(0)\rangle. \tag*{(37)} B ( h , h ) = 0 , B ( h , w ) = ⟨ q ¨ ( b ) , w ( b )⟩ − ⟨ q ¨ ( 0 ) , w ( 0 )⟩ . (37)
Choose w w w smooth and supported near the interior point b ∈ ( 0 , L ) b\in(0,L) b ∈ ( 0 , L ) , with w ( b ) = q ¨ ( b ) w(b)=\ddot{q}(b) w ( b ) = q ¨ ( b ) . Then B ( h , w ) = q ¨ ( b ) ∣ 2 > 0 B(h,w)=\ddot{q}(b)|^2>0 B ( h , w ) = q ¨ ( b ) ∣ 2 > 0 . This proves h ∉ rad ( B ) h\notin\operatorname{rad}(B) h ∈ / rad ( B ) . □
Proof of Theorem 1.1
Firstly, Corollary 2.5 gives a classical solution satisfying (1.8). It remains to exclude a shorter minimal period. The exclusion uses only part of the assumptions in § 1 .
Proposition 3.1. Let V ∈ C 2 ( R n , R ) V \in C^{2}(\mathbb{R}^{n},\mathbb{R}) V ∈ C 2 ( R n , R ) be even and nonnegative with V ( 0 ) = 0 V(0)=0 V ( 0 ) = 0 . Suppose q ∈ S T q \in S_{T} q ∈ S T is a nonconstant classical solution of (1.1), c = ψ T ( q ) > 0 c=\psi_{T}(q)>0 c = ψ T ( q ) > 0 , and its minimal period is less than T T T . Then there is an open neighborhood O O O of q q q in S T S_{T} S T such that every u ∈ O u \in O u ∈ O with ψ T ( u ) < c \psi_{T}(u)<c ψ T ( u ) < c can be joined to a negative action point by a continuous path in { ψ T < c } \{\psi_{T}<c\} { ψ T < c } .
We prove this proposition in the next three subsections. By Lemma 2.3, write
τ = T / k , k = 2 m + 1 ≥ 3 , a = τ / 4 = T 4 k , \tau= T/k,\qquad k=2m+1\ge3,\qquad a=\tau/4=\frac{T}{4k}, τ = T / k , k = 2 m + 1 ≥ 3 , a = τ /4 = 4 k T ,
L = k a = T 4 , b = 2 m a = ( k − 1 ) a = L − a . (38) L=ka=\frac{T}{4},\qquad b=2ma=(k-1)a=L-a. \tag*{(38)} L = k a = 4 T , b = 2 ma = ( k − 1 ) a = L − a . (38)
Thus [ 0 , L ] [0,L] [ 0 , L ] contains k k k primitive quarter-periods, and [ b , L ] [b,L] [ b , L ] is its last primitive quarter-period. Regard q q q as its restriction to X L X_{L} X L . It is a critical point of I L I_{L} I L with critical value c / 4 c/4 c /4 . Set
A 0 = q ( 0 ) , B 0 = q ( b ) = ( − 1 ) m A 0 , E = 1 2 ∣ q ′ ( t ) ∣ 2 + V ( q ( t ) ) . (39) A_{0}=q(0),\qquad B_{0}=q(b)=(-1)^{m}A_{0},\qquad E=\frac{1}{2}|q'(t)|^{2}+V(q(t)). \tag*{(39)} A 0 = q ( 0 ) , B 0 = q ( b ) = ( − 1 ) m A 0 , E = 2 1 ∣ q ′ ( t ) ∣ 2 + V ( q ( t )) . (39)
The energy conservation and the symmetries imply
q ′ ( 0 ) = q ′ ( b ) = 0 , q ( L ) = 0 , q ( b + s ) = q ( b − s ) , V ( A 0 ) = V ( B 0 ) = E . (40) q'(0)=q'(b)=0,\qquad q(L)=0,\qquad q(b+s)=q(b-s),\qquad V(A_{0})=V(B_{0})=E. \tag*{(40)} q ′ ( 0 ) = q ′ ( b ) = 0 , q ( L ) = 0 , q ( b + s ) = q ( b − s ) , V ( A 0 ) = V ( B 0 ) = E . (40)
Since E = ∣ q ′ ( L ) ∣ 2 / 2 ≥ 0 E=|q'(L)|^{2}/2\ge0 E = ∣ q ′ ( L ) ∣ 2 /2 ≥ 0 . If E = 0 E=0 E = 0 , then q ( L ) = q ′ ( L ) = 0 q(L)=q'(L)=0 q ( L ) = q ′ ( L ) = 0 , and the uniqueness, together with ∇ V ( 0 ) = 0 \nabla V(0)=0 ∇ V ( 0 ) = 0 , gives the zero solution. So the energy E E E is positive.
Write
ℓ ( t ) = 1 2 ∣ q ′ ( t ) ∣ 2 − V ( q ( t ) ) = ∣ q ′ ( t ) ∣ 2 − E , J = ∫ 0 a ℓ ( t ) d t . (41) \ell(t)=\frac{1}{2}|q'(t)|^{2}-V(q(t))=|q'(t)|^{2}-E,\qquad J=\int_{0}^{a}\ell(t)\,\mathrm{d}t. \tag*{(41)} ℓ ( t ) = 2 1 ∣ q ′ ( t ) ∣ 2 − V ( q ( t )) = ∣ q ′ ( t ) ∣ 2 − E , J = ∫ 0 a ℓ ( t ) d t . (41)
Time reflection gives ℓ ( − t ) = ℓ ( t ) \ell(-t)=\ell(t) ℓ ( − t ) = ℓ ( t ) . Then, by half-antiperiodicity with primitive half-period τ / 2 = 2 a \tau/2=2a τ /2 = 2 a and the evenness of V V V , we have
ℓ ( 2 a − t ) = ℓ ( t ) , ∫ a 2 a ℓ ( t ) d t = ∫ 0 a ℓ ( t ) d t = J . \ell(2a-t)=\ell(t),\qquad\int_{a}^{2a}\ell(t)\,\mathrm{d}t=\int_{0}^{a}\ell(t)\,\mathrm{d}t=J. ℓ ( 2 a − t ) = ℓ ( t ) , ∫ a 2 a ℓ ( t ) d t = ∫ 0 a ℓ ( t ) d t = J .
Thus each successive interval of length a a a has action J J J , including [ b , L ] [b,L] [ b , L ] . Therefore I L ( q ) = k J I_{L}(q)=kJ I L ( q ) = k J . Since V ≥ 0 V\ge0 V ≥ 0 , it yields 1 2 ∣ q ′ ( t ) ∣ 2 = E − V ( q ( t ) ) ≤ E \frac{1}{2}|q'(t)|^{2}=E-V(q(t))\le E 2 1 ∣ q ′ ( t ) ∣ 2 = E − V ( q ( t )) ≤ E . Integrating over [ 0 , a ] [0,a] [ 0 , a ] gives
J = ∫ 0 a ∣ q ′ ( t ) ∣ 2 d t − a E ≤ 2 a E − a E = a E . (42) J=\int_{0}^{a}|q'(t)|^{2}\,\mathrm{d}t-aE\le2aE-aE=aE. \tag*{(42)} J = ∫ 0 a ∣ q ′ ( t ) ∣ 2 d t − a E ≤ 2 a E − a E = a E . (42)
The first action-decreasing surgery
For 0 ≤ s ≤ b 0\le s\le b 0 ≤ s ≤ b , define
U s ( t ) = { q ( t + s ) , 0 ≤ t ≤ b − s , B 0 , b − s ≤ t ≤ b , q ( t ) , b ≤ t ≤ L . (43) U_{s}(t)=
\begin{cases}
q(t+s), & 0\le t\le b-s,\\
B_{0}, & b-s\le t\le b,\\
q(t), & b\le t\le L.
\end{cases} \tag*{(43)} U s ( t ) = ⎩ ⎨ ⎧ q ( t + s ) , B 0 , q ( t ) , 0 ≤ t ≤ b − s , b − s ≤ t ≤ b , b ≤ t ≤ L . (43)
One can easily check that U s U_{s} U s is well-defined, U s ∈ X L U_{s}\in X_{L} U s ∈ X L and U 0 = q U_{0}=q U 0 = q .
The path s ↦ U s s \mapsto U_s s ↦ U s is continuous in H 1 H^1 H 1 . Indeed, if s j → s s_j \to s s j → s , both U s j ( t ) → U s ( t ) U_{s_j}(t) \to U_s(t) U s j ( t ) → U s ( t ) and d s j ( t ) → d s ( t ) d_{s_j}(t) \to d_s(t) d s j ( t ) → d s ( t ) hold except at the finitely many limiting junctions, where d s ( t ) d_s(t) d s ( t ) denotes the piecewise weak derivative of U s ( t ) U_s(t) U s ( t ) . Note that on [ 0 , L ] [0,L] [ 0 , L ] they satisfy the uniform bounds
∣ U s j ( t ) ∣ , ∣ U s ( t ) ∣ ≤ max [ 0 , L ] ∣ q ∣ , ∣ d s j ( t ) ∣ , ∣ d s ( t ) ∣ ≤ max [ 0 , L ] ∣ q ′ ∣ . |U_{s_j}(t)|, |U_s(t)| \le\max_{[0,L]} |q|,\qquad|d_{s_j}(t)|, |d_s(t)| \le\max_{[0,L]} |q'|. ∣ U s j ( t ) ∣ , ∣ U s ( t ) ∣ ≤ [ 0 , L ] max ∣ q ∣ , ∣ d s j ( t ) ∣ , ∣ d s ( t ) ∣ ≤ [ 0 , L ] max ∣ q ′ ∣.
Then the dominated convergence theorem gives
∥ U s j − U s ∥ H 1 2 = ∫ 0 L ∣ U s j − U s ∣ 2 d t + ∫ 0 L ∣ d s j − d s ∣ 2 d t ⟶ 0. \|U_{s_j}-U_s\|_{H^1}^2=\int_0^L |U_{s_j}-U_s|^2\,\mathrm{d}t+\int_0^L |d_{s_j}-d_s|^2\,\mathrm{d}t\longrightarrow0. ∥ U s j − U s ∥ H 1 2 = ∫ 0 L ∣ U s j − U s ∣ 2 d t + ∫ 0 L ∣ d s j − d s ∣ 2 d t ⟶ 0.
By the definition (3.6) of U s U_s U s , V ( B 0 ) = E V(B_0)=E V ( B 0 ) = E and ℓ ( t ) = ∣ q ′ ( t ) ∣ 2 − E \ell(t)=|q'(t)|^2-E ℓ ( t ) = ∣ q ′ ( t ) ∣ 2 − E , we obtain
I L ( U s ) − I L ( q ) = ∫ 0 b − s ℓ ( t + s ) d t − s V ( B 0 ) + ∫ b L ℓ ( t ) d t − ∫ 0 L ℓ ( t ) d t = ∫ s b ℓ ( r ) d r − s E + ∫ b 0 ℓ ( t ) d t = − ∫ 0 s ∣ q ′ ( t ) ∣ 2 d t < 0 , ∀ s > 0. (44) \begin{aligned}
I_L(U_s)-I_L(q)=\int_0^{b-s}\ell(t+s)\,\mathrm{d}t-sV(B_0)+\int_b^L\ell(t)\,\mathrm{d}t-\int_0^L\ell(t)\,\mathrm{d}t \\
&=\int_s^b\ell(r)\,\mathrm{d}r-sE+\int_b^0\ell(t)\,\mathrm{d}t=-\int_0^s|q'(t)|^2\,\mathrm{d}t<0,\qquad\forall s>0. \tag*{(44)}
\end{aligned} I L ( U s ) − I L ( q ) = ∫ 0 b − s ℓ ( t + s ) d t − s V ( B 0 ) + ∫ b L ℓ ( t ) d t − ∫ 0 L ℓ ( t ) d t = ∫ s b ℓ ( r ) d r − s E + ∫ b 0 ℓ ( t ) d t = − ∫ 0 s ∣ q ′ ( t ) ∣ 2 d t < 0 , ∀ s > 0. (44)
where the last strict inequality holds because the solution q ( t ) q(t) q ( t ) is nonconstant.
When s = b s=b s = b , by (3.5), it yields
I L ( U b ) = J − b E ≤ a E − ( k − 1 ) a E = − ( k − 2 ) a E < 0. (45) I_L(U_b)=J-bE\le aE-(k-1)aE=-(k-2)aE<0. \tag*{(45)} I L ( U b ) = J − b E ≤ a E − ( k − 1 ) a E = − ( k − 2 ) a E < 0. (45)
The first-order tangent of this path is the field h h h in (2.24):
U s = q + s h + o ( s ) in H 1 ( [ 0 , L ] , R n ) ( s ↓ 0 ) . (46) U_s=q+sh+o(s)\qquad\text{in }H^1([0,L],\mathbb{R}^n)\quad(s\downarrow0). \tag*{(46)} U s = q + s h + o ( s ) in H 1 ([ 0 , L ] , R n ) ( s ↓ 0 ) . (46)
Set R s ≡ U s − q − s h R_s\equiv U_s-q-sh R s ≡ U s − q − s h . Let M = max [ 0 , L ] ∣ q ¨ ∣ M=\max_{[0,L]}|\ddot q| M = max [ 0 , L ] ∣ q ¨ ∣ and ω ( δ ) = sup { ∣ q ¨ ( r ) − q ¨ ( t ) ∣ : r , t ∈ [ 0 , L ] , ∣ r − t ∣ ≤ δ } \omega(\delta)=\sup\{|\ddot q(r)-\ddot q(t)|:r,t\in[0,L],|r-t|\le\delta\} ω ( δ ) = sup { ∣ q ¨ ( r ) − q ¨ ( t ) ∣ : r , t ∈ [ 0 , L ] , ∣ r − t ∣ ≤ δ } . Obviously, there holds ω ( δ ) → 0 \omega(\delta)\to0 ω ( δ ) → 0 . Since q ′ ( b ) = 0 q'(b)=0 q ′ ( b ) = 0 , the weak derivative of h h h equals q ¨ \ddot q q ¨ on ( 0 , b ) (0,b) ( 0 , b ) and 0 0 0 on ( b , L ) (b,L) ( b , L ) . Next we will estimate R s R_s R s on three different intervals.
i) On [ 0 , b − s ] [0,b-s] [ 0 , b − s ] . Here h = q ′ h=q' h = q ′ , and the fundamental theorem gives
R s ( t ) = ∫ 0 s ( q ′ ( t + r ) − q ′ ( t ) ) d r , ∣ R s ( t ) ∣ ≤ 1 2 M s 2 , R_s(t)=\int_0^s(q'(t+r)-q'(t))\,\mathrm{d}r,\qquad|R_s(t)|\le\tfrac12 Ms^2, R s ( t ) = ∫ 0 s ( q ′ ( t + r ) − q ′ ( t )) d r , ∣ R s ( t ) ∣ ≤ 2 1 M s 2 ,
R s ′ ( t ) = ∫ 0 s ( q ¨ ( t + r ) − q ¨ ( t ) ) d r , ∣ R s ′ ( t ) ∣ ≤ s ω ( s ) . R_s'(t)=\int_0^s(\ddot q(t+r)-\ddot q(t))\,\mathrm{d}r,\qquad|R_s'(t)|\le s\omega(s). R s ′ ( t ) = ∫ 0 s ( q ¨ ( t + r ) − q ¨ ( t )) d r , ∣ R s ′ ( t ) ∣ ≤ s ω ( s ) .
This part contributes at most b ( M s 2 / 2 + s ω ( s ) ) \sqrt b\,(Ms^2/2+s\omega(s)) b ( M s 2 /2 + s ω ( s )) to the H 1 H^1 H 1 -norm.
ii) On [ b − s , b ] [b-s,b] [ b − s , b ] . First,
q ′ ( t ) = − ∫ t b q ¨ ( r ) d r , ∣ q ′ ( t ) ∣ ≤ M ( b − t ) ≤ M s . q'(t)=-\int_t^b\ddot q(r)\,\mathrm{d}r,\qquad|q'(t)|\le M(b-t)\le Ms. q ′ ( t ) = − ∫ t b q ¨ ( r ) d r , ∣ q ′ ( t ) ∣ ≤ M ( b − t ) ≤ M s .
Integrating this estimate on [ t , b ] [t,b] [ t , b ] again gives
∣ q ( b ) − q ( t ) ∣ ≤ ∫ t b M ( b − r ) d r = 1 2 M ( b − t ) 2 ≤ 1 2 M s 2 . |q(b)-q(t)|\le\int_t^b M(b-r)\,\mathrm{d}r=\tfrac12M(b-t)^2\le\tfrac12Ms^2. ∣ q ( b ) − q ( t ) ∣ ≤ ∫ t b M ( b − r ) d r = 2 1 M ( b − t ) 2 ≤ 2 1 M s 2 .
In this interval, almost everywhere there holds
R s ( t ) = q ( b ) − q ( t ) − s q ′ ( t ) , R s ′ ( t ) = − q ′ ( t ) − s q ¨ ( t ) , R_s(t)=q(b)-q(t)-sq'(t),\qquad R_s'(t)=-q'(t)-s\ddot q(t), R s ( t ) = q ( b ) − q ( t ) − s q ′ ( t ) , R s ′ ( t ) = − q ′ ( t ) − s q ¨ ( t ) ,
which implies that ∣ R s ( t ) ∣ ≤ 3 2 M s 2 |R_s(t)|\le\frac32Ms^2 ∣ R s ( t ) ∣ ≤ 2 3 M s 2 and ∣ R s ′ ( t ) ∣ ≤ 2 M s |R_s'(t)|\le2Ms ∣ R s ′ ( t ) ∣ ≤ 2 M s . Hence
∥ R s ∥ L 2 ( b − s , b ) ≤ 3 2 M s 5 / 2 , ∥ R s ′ ∥ L 2 ( b − s , b ) ≤ 2 M s 3 / 2 . \|R_s\|_{L^2(b-s,b)}\le\tfrac32Ms^{5/2},\qquad\|R_s'\|_{L^2(b-s,b)}\le2Ms^{3/2}. ∥ R s ∥ L 2 ( b − s , b ) ≤ 2 3 M s 5/2 , ∥ R s ′ ∥ L 2 ( b − s , b ) ≤ 2 M s 3/2 .
The total contribution of this part to the H 1 H^1 H 1 -norm is O ( s 3 / 2 ) O(s^{3/2}) O ( s 3/2 ) .
iii) On [ b , L ] [b,L] [ b , L ] . One has U s = q U_s=q U s = q and h = 0 h=0 h = 0 , so R s = R s ′ = 0 R_s=R'_s=0 R s = R s ′ = 0 almost everywhere.
Combining the three intervals, for a constant C C C independent of small s > 0 s>0 s > 0 ,
∥ U s − q − s h ∥ H 1 ( 0 , L ) ≤ C ( s 2 + s ω ( s ) + s 3 / 2 ) = o ( s ) . \lVert U_s-q-sh\rVert_{H^1(0,L)}\le C\left(s^2+s\omega(s)+s^{3/2}\right)=o(s). ∥ U s − q − s h ∥ H 1 ( 0 , L ) ≤ C ( s 2 + s ω ( s ) + s 3/2 ) = o ( s ) .
This establishes (46) in the Hilbert norm.
The opposite surgery for a triple iterate
Suppose k = 3 k=3 k = 3 , so b = 2 a b=2a b = 2 a and L = 3 a L=3a L = 3 a . For 0 ≤ s ≤ a 0\le s\le a 0 ≤ s ≤ a , define
W s ( t ) = { A 0 , 0 ≤ t ≤ 2 s , q ( t − 2 s ) , 2 s ≤ t ≤ 2 a + s , q ( t ) , 2 a + s ≤ t ≤ 3 a . (47) W_s(t)=
\begin{cases}
A_0, & 0\le t\le2s,\\
q(t-2s), & 2s\le t\le2a+s,\\
q(t), & 2a+s\le t\le3a.
\end{cases} \tag*{(47)} W s ( t ) = ⎩ ⎨ ⎧ A 0 , q ( t − 2 s ) , q ( t ) , 0 ≤ t ≤ 2 s , 2 s ≤ t ≤ 2 a + s , 2 a + s ≤ t ≤ 3 a . (47)
At the first junction the values match because q ( 0 ) = A 0 q(0)=A_0 q ( 0 ) = A 0 . At the second one they match because q ( 2 a − s ) = q ( 2 a + s ) q(2a-s)=q(2a+s) q ( 2 a − s ) = q ( 2 a + s ) by the reflection about 2 a 2a 2 a . Thus W s ∈ X 3 a W_s\in X_{3a} W s ∈ X 3 a and W 0 = q W_0=q W 0 = q . The same argument as for U s U_s U s proves H 1 H^1 H 1 continuity.
By the definition (3.10) of W s W_s W s and V ( A 0 ) = E V(A_0)=E V ( A 0 ) = E , we obtain
I 3 a ( W s ) − I 3 a ( q ) = − 2 s E + ∫ 0 2 a − s ℓ ( r ) d r + ∫ 2 a + s 3 a ℓ ( t ) d t − ∫ 0 3 a ℓ ( t ) d t = − 2 s E − ∫ 2 a − s 2 a + s ℓ ( t ) d t = − ∫ 2 a − s 2 a + s ∣ q ′ ( t ) ∣ 2 d t < 0 , ∀ s > 0. (48) \begin{aligned}
I_{3a}(W_s)-I_{3a}(q)
&=-2sE+\int_0^{2a-s}\ell(r)\,\mathrm{d}r+\int_{2a+s}^{3a}\ell(t)\,\mathrm{d}t-\int_0^{3a}\ell(t)\,\mathrm{d}t\\
&=-2sE-\int_{2a-s}^{2a+s}\ell(t)\,\mathrm{d}t
=-\int_{2a-s}^{2a+s}\lvert q'(t)\rvert^2\,\mathrm{d}t<0,\qquad\forall s>0. \tag*{(48)}
\end{aligned} I 3 a ( W s ) − I 3 a ( q ) = − 2 s E + ∫ 0 2 a − s ℓ ( r ) d r + ∫ 2 a + s 3 a ℓ ( t ) d t − ∫ 0 3 a ℓ ( t ) d t = − 2 s E − ∫ 2 a − s 2 a + s ℓ ( t ) d t = − ∫ 2 a − s 2 a + s ∣ q ′ ( t ) ∣ 2 d t < 0 , ∀ s > 0. (48)
When s = a s=a s = a , it is easy to see that
I 3 a ( W a ) = J − 2 a E ≤ − a E < 0. (49) I_{3a}(W_a)=J-2aE\le-aE<0. \tag*{(49)} I 3 a ( W a ) = J − 2 a E ≤ − a E < 0. (49)
The first-order tangent of this path is the field − 2 h -2h − 2 h in (32) :
W s = q − 2 s h + o ( s ) in H 1 ( [ 0 , 3 a ] , R n ) ( s ↓ 0 ) . (50) W_s=q-2sh+o(s)\quad\text{in }H^1([0,3a],\mathbb{R}^n)\quad(s\downarrow0). \tag*{(50)} W s = q − 2 s h + o ( s ) in H 1 ([ 0 , 3 a ] , R n ) ( s ↓ 0 ) . (50)
Let R s W ≡ W s − q + 2 s h R_s^W\equiv W_s-q+2sh R s W ≡ W s − q + 2 s h , b = 2 a b=2a b = 2 a , and take 0 < s < a / 2 0<s<a/2 0 < s < a /2 . Use the constant M M M and ω ( δ ) \omega(\delta) ω ( δ ) from the proof for U U U , but being replaced by the interval [ 0 , 3 a ] [0,3a] [ 0 , 3 a ] .
On [ 2 s , b ] [2s,b] [ 2 s , b ] , by the fundamental theorem, we obtain
R s W ( t ) = q ( t − 2 s ) − q ( t ) + 2 s q ′ ( t ) , ∣ R s W ( t ) ∣ ≤ 2 M s 2 , R_s^W(t)=q(t-2s)-q(t)+2sq'(t),\qquad\lvert R_s^W(t)\rvert\le2Ms^2, R s W ( t ) = q ( t − 2 s ) − q ( t ) + 2 s q ′ ( t ) , ∣ R s W ( t )∣ ≤ 2 M s 2 ,
( R s W ) ′ ( t ) = q ′ ( t − 2 s ) − q ′ ( t ) + 2 s q ¨ ( t ) , ∣ ( R s W ) ′ ( t ) ∣ ≤ 2 s ω ( 2 s ) . (R_s^W)'(t)=q'(t-2s)-q'(t)+2s\ddot q(t),\qquad\lvert(R_s^W)'(t)\rvert\le2s\omega(2s). ( R s W ) ′ ( t ) = q ′ ( t − 2 s ) − q ′ ( t ) + 2 s q ¨ ( t ) , ∣( R s W ) ′ ( t )∣ ≤ 2 s ω ( 2 s ) .
On [ 0 , 2 s ] [0,2s] [ 0 , 2 s ] , one has q ′ ( 0 ) = 0 q'(0)=0 q ′ ( 0 ) = 0 , ∣ q ′ ( t ) ∣ ≤ M t \lvert q'(t)\rvert\le Mt ∣ q ′ ( t )∣ ≤ M t , and ∣ q ( t ) − q ( 0 ) ∣ ≤ M t 2 / 2 \lvert q(t)-q(0)\rvert\le Mt^2/2 ∣ q ( t ) − q ( 0 )∣ ≤ M t 2 /2 . Hence
R s W = q ( 0 ) − q + 2 s q ′ , ( R s W ) ′ = − q ′ + 2 s q ¨ , R_s^W=q(0)-q+2sq',\qquad(R_s^W)'=-q'+2s\ddot q, R s W = q ( 0 ) − q + 2 s q ′ , ( R s W ) ′ = − q ′ + 2 s q ¨ ,
which establish the bounds ∣ R s W ∣ ≤ 6 M s 2 \lvert R_s^W\rvert\le6Ms^2 ∣ R s W ∣ ≤ 6 M s 2 and ∣ ( R s W ) ′ ∣ ≤ 4 M s \lvert(R_s^W)'\rvert\le4Ms ∣( R s W ) ′ ∣ ≤ 4 M s .
On [ b , b + s ] [b,b+s] [ b , b + s ] , the cutoff h h h is zero. Both t t t and t − 2 s t-2s t − 2 s lie within 2 s 2s 2 s of b b b , where q ′ ( b ) = 0 q'(b)=0 q ′ ( b ) = 0 . Thus
R s W ( t ) = q ( t − 2 s ) − q ( t ) , ∣ R s W ( t ) ∣ ≤ 4 M s 2 , ∣ ( R s W ) ′ ( t ) ∣ ≤ 2 M s . R_s^W(t)=q(t-2s)-q(t),\qquad\lvert R_s^W(t)\rvert\le4Ms^2,\qquad\lvert(R_s^W)'(t)\rvert\le2Ms. R s W ( t ) = q ( t − 2 s ) − q ( t ) , ∣ R s W ( t )∣ ≤ 4 M s 2 , ∣( R s W ) ′ ( t )∣ ≤ 2 M s .
Note that the error is zero for t ≥ b + s t\ge b+s t ≥ b + s . Since two intervals [ 0 , 2 s ] [0,2s] [ 0 , 2 s ] and [ b , b + s ] [b,b+s] [ b , b + s ] have total length 3 s 3s 3 s , their H 1 H^1 H 1 contribution is O ( s 3 / 2 ) O(s^{3/2}) O ( s 3/2 ) . The contribution on [ 2 s , b ] [2s,b] [ 2 s , b ] is O ( s 2 + s ω ( 2 s ) ) O(s^2+s\omega(2s)) O ( s 2 + s ω ( 2 s )) . Thus
∥ W s − q + 2 s h ∥ H 1 ( 0 , 3 a ) ≤ C ( s 2 + s ω ( 2 s ) + s 3 / 2 ) = o ( s ) , (51) \lVert W_s-q+2sh\rVert_{H^1(0,3a)}\le C\left(s^2+s\omega(2s)+s^{3/2}\right)=o(s), \tag*{(51)} ∥ W s − q + 2 s h ∥ H 1 ( 0 , 3 a ) ≤ C ( s 2 + s ω ( 2 s ) + s 3/2 ) = o ( s ) , (51)
This establishes (3.13) in the Hilbert norm.
Exclusion and completion of the proof
Proof of Proposition 3.1. Consider the functional I L I_L I L defined on X L X_L X L with positive critical value I L ( q ) = c / 4 I_L(q)=c/4 I L ( q ) = c /4 . Its Hessian is the form B B B in (2.23). By Lemma 2.7, the field h h h is isotropic and not in rad ( B ) \operatorname{rad}(B) rad ( B ) .
When k = 3 k=3 k = 3 , (3.6)–(3.13) give two paths that lie strictly below the critical value for positive parameter, end at negative values, and have tangents h h h and − 2 h -2h − 2 h . By the case (i) of Lemma 2.6 with d = 2 d=2 d = 2 , every sufficiently nearby point of smaller action can be joined, below I L ( q ) I_L(q) I L ( q ) , to a point of negative action.
When k ≥ 5 k\ge5 k ≥ 5 , Lemma 2.7 supplies a negative two-plane. By (3.7), (3.8) and (3.9), the path U U U satisfies all conditions in case (ii) of Lemma 2.6. So the same conclusion follows.
Finally transport the neighborhood and all connecting paths through the bounded reflection isomorphism (2.1). The identity ψ T = 4 I L \psi_T=4I_L ψ T = 4 I L preserves strict inequalities and the sign of the endpoint action. This proves the proposition on S T S_T S T . □
Proof of Theorem 1.1. Fix an integer n ≥ 1 n\ge1 n ≥ 1 , a potential satisfying (V1′) and (V2)–(V5), and T > 0 T>0 T > 0 . Corollary 2.5 gives a nonconstant classical solution q ∈ S T q\in S_T q ∈ S T satisfying (1.8). Let τ \tau τ be its minimal positive period, whose existence follows from Lemma 2.3.
Suppose τ < T \tau<T τ < T . Proposition 3.1 gives an open neighborhood O O O of q q q with the stated exit property. Choose δ > 0 \delta>0 δ > 0 such that B H 1 ( q , δ ) ∩ S T ⊂ O B_{H^1}(q,\delta)\cap S_T\subset O B H 1 ( q , δ ) ∩ S T ⊂ O . Since q ∈ C T q\in C_T q ∈ C T , there is
u ∈ C T ∩ B H 1 ( q , δ ) ⊂ O ∩ C T . u\in C_T\cap B_{H^1}(q,\delta)\subset O\cap C_T. u ∈ C T ∩ B H 1 ( q , δ ) ⊂ O ∩ C T .
In particular ψ T ( u ) < c T \psi_T(u)<c_T ψ T ( u ) < c T . By the definition of C T C_T C T , there is a continuous path η 0 : [ 0 , 1 ] → { ψ T < c T } \eta_0:[0,1]\to\{\psi_T<c_T\} η 0 : [ 0 , 1 ] → { ψ T < c T } from 0 0 0 to u u u . Proposition 3.1 gives a continuous path η 1 : [ 0 , 1 ] → { ψ T < c T } \eta_1:[0,1]\to\{\psi_T<c_T\} η 1 : [ 0 , 1 ] → { ψ T < c T } from u u u to a point e e e with ψ T ( e ) < 0 \psi_T(e)<0 ψ T ( e ) < 0 . Define
γ ( s ) = { η 0 ( 2 s ) , 0 ≤ s ≤ 1 / 2 , η 1 ( 2 s − 1 ) , 1 / 2 ≤ s ≤ 1. \gamma(s)=
\begin{cases}
\eta_0(2s), & 0\le s\le1/2,\\
\eta_1(2s-1), & 1/2\le s\le1.
\end{cases} γ ( s ) = { η 0 ( 2 s ) , η 1 ( 2 s − 1 ) , 0 ≤ s ≤ 1/2 , 1/2 ≤ s ≤ 1.
Obviously, γ \gamma γ is H 1 H^1 H 1 -continuous. It starts at zero, ends at e e e , and belongs to Γ T \Gamma_T Γ T . Every point on this path has action less than c T c_T c T . The continuous function ψ T ∘ γ \psi_T\circ\gamma ψ T ∘ γ attains its maximum at some s ∗ ∈ [ 0 , 1 ] s_*\in[0,1] s ∗ ∈ [ 0 , 1 ] . Thus, by the definition (1.6) of c T c_T c T , we obtain
c T ≤ max s ∈ [ 0 , 1 ] ψ T ( γ ( s ) ) = ψ T ( γ ( s ∗ ) ) < c T . c_T\le\max_{s\in[0,1]}\psi_T(\gamma(s))=\psi_T(\gamma(s_*))<c_T. c T ≤ s ∈ [ 0 , 1 ] max ψ T ( γ ( s )) = ψ T ( γ ( s ∗ )) < c T .
a contradiction. Therefore τ = T \tau=T τ = T . The time-even and half-antiperiodic identities follow from q ∈ S T q\in S_T q ∈ S T , and the positive action and classical regularity follow from the selection and compactness lemmas. □
Remark 3.2. The growth outside some ball and the assumptions near the origin are used for mountain pass geometry and Palais–Smale compactness. Evenness of the potential gives the reflection space and the matching identities in the surgeries. Nonnegativity gives the negative action endpoint estimates (3.8) and (3.12). The local argument uses the C 2 C^2 C 2 regularity.
AI disclosure. ChatGPT was used to assist with developing the draft arguments, some proof strategies and language refinement. The authors take responsibility for the human verification and exposition.
References
Huagui Duan, School of Mathematics, Jilin University, Changchun, Jilin, 130012, The People’s Republic of China
Email address: duanhg@nankai.edu.cn
Yan Wu, School of Mathematical Sciences, Nankai University, Tianjin 300071, The People’s Republic of China