Introduction

Arnold’s fixed-point conjecture compares Hamiltonian dynamics with the critical-point theory of smooth functions. For a sufficiently short autonomous Hamiltonian flow on a closed symplectic manifold, its fixed points are exactly the critical points of its Hamiltonian. The conjecture asks whether an arbitrary Hamiltonian diffeomorphism must retain the minimum number of critical points forced by the topology of the manifold. Arnold’s formulation explicitly includes the count of geometrically distinct points [1], Section 2. We construct a smooth counterexample on the complex quadric threefold, which also violates the weaker rational cup-length bound.

The two degenerate bounds. For a closed smooth manifold MM, write

crit⁡(h)={q∈M:dhq=0},Crit⁡(M)=min⁡h∈C∞(M,R)#crit⁡(h).\operatorname{crit}(h)=\{q\in M:dh_q=0\},\qquad\operatorname{Crit}(M)=\min_{h\in C^\infty(M,\mathbb{R})}\#\operatorname{crit}(h).

The minimum here permits degenerate critical points, whereas the Morse number minimizes only over Morse functions. Let cuplength⁡(M;Q)\operatorname{cuplength}(M;\mathbb{Q}) be one plus the maximum number of positive-degree rational cohomology classes with nonzero cup product. Thus our cup-length convention includes the unit. The Lusternik–Schnirelmann inequality gives

cuplength⁡(M;Q)≤Crit⁡(M);\operatorname{cuplength}(M;\mathbb{Q})\leq\operatorname{Crit}(M);

Section 5 recalls the cup-product argument in a form that applies to arbitrary smooth functions.

A Hamiltonian diffeomorphism of a symplectic manifold (M,ω)(M,\omega) is the time-one map of a smooth time-dependent Hamiltonian flow. We write Ham⁡(M,ω)\operatorname{Ham}(M,\omega) for these maps, use the convention ιXHtω=dHt\iota_{X_{H_t}}\omega=dH_t, and set Fix⁡(ϕ)={q∈M:ϕ(q)=q}\operatorname{Fix}(\phi)=\{q\in M:\phi(q)=q\}. The critical-number Arnold bound asserts that

#Fix⁡(ϕ)≥Crit⁡(M)(1)\#\operatorname{Fix}(\phi)\geq\operatorname{Crit}(M) \tag*{(1)}

for every ϕ∈Ham⁡(M,ω)\phi\in\operatorname{Ham}(M,\omega) on every closed connected symplectic manifold. Its weaker rational cup-length form replaces Crit⁡(M)\operatorname{Crit}(M) by cuplength⁡(M;Q)\operatorname{cuplength}(M;\mathbb{Q}). Neither formulation assumes nondegeneracy; see Golovko [7], Conjectures 1.4 and 1.6. A fixed point qq is nondegenerate when 11 is not an eigenvalue of dϕqd\phi_q.

Let

Q3={[z0:z1:z2:z3:z4]∈CP4:z02+z12+z22+z32+z42=0},(2)Q^{3}=\{[z_{0}:z_{1}:z_{2}:z_{3}:z_{4}]\in\mathbb{CP}^{4}:z_{0}^{2}+z_{1}^{2}+z_{2}^{2}+z_{3}^{2}+z_{4}^{2}=0\}, \tag*{(2)}

with the restricted Fubini–Study symplectic form, normalized in Section 3.

Theorem 1.1. There is a smooth Hamiltonian diffeomorphism ϕ\phi of the closed connected symplectic six-manifold Q3Q^{3} such that

#Fix⁡(ϕ)=3<4=Crit⁡(Q3)=cuplength⁡(Q3;Q).(3)\#\operatorname{Fix}(\phi)=3<4=\operatorname{Crit}(Q^{3})=\operatorname{cuplength}(Q^{3};\mathbb{Q}). \tag*{(3)}

At least one fixed point of ϕ\phi is degenerate.

This disproves both unrestricted smooth bounds. The count includes all fixed points, so imposing a contractibility condition on their Hamiltonian trajectories cannot restore either inequality. The nondegenerate homological Arnold inequalities concern a different hypothesis and remain compatible with this example.

The count of three is sharp. Gong [8], Section 1.1.1 obtains at least nn fixed points for every Hamiltonian diffeomorphism of the standard complex quadric of complex dimension n≥2n\ge2. His normalization of the Fubini–Study form differs from ours by a positive scalar, which does not change the Hamiltonian group: rescaling the form and a generating Hamiltonian by the same scalar preserves its vector field. Thus our example attains his bound for n=3n=3.

Historical context and related work. The positive cases explain why the unrestricted comparison is natural. Conley and Zehnder proved that every Hamiltonian diffeomorphism of the standard symplectic torus T2nT^{2n} has at least 2n+12n+1 fixed points [3], and Fortune obtained the bound n+1n+1 on standard CPn\mathbb{CP}^{n} [5]. Hofer obtained cup-length bounds without nondegeneracy when π2(M)=0\pi_{2}(M)=0 [9], Theorem 4. Rudyak and Oprea proved the full critical-number bound when both the symplectic class and the first Chern class vanish on π2(M)\pi_{2}(M) [17], Corollary 4.3. The quadric has spheres of positive symplectic area and lies outside these latter hypotheses.

Floer’s pseudoholomorphic-curve methods established homological Morse inequalities for nondegenerate Hamiltonian fixed points, including the monotone setting [4]. The lower bound by the sum of rational Betti numbers was subsequently established for nondegenerate Hamiltonian diffeomorphisms on every closed symplectic manifold by Fukaya and Ono, Liu and Tian, and Ruan [6, 11, 16]. The distinction between these bounds and the degenerate formulations above is essential; see [7], Theorem 1.3 and Conjectures 1.4 and 1.6. Ma [13], Theorem 1.5 asserted that the minimum number of fixed points over all Hamiltonian diffeomorphisms always equals Crit⁡(M)\operatorname{Crit}(M). Theorem 1.1 contradicts that assertion.

Regularity also matters. Buhovsky, Humilière, and Seyfaddini constructed Hamiltonian homeomorphisms with a single fixed point on every closed connected symplectic manifold of real dimension at least four [2]. Our construction stays in the smooth Hamiltonian group. In a preprint posted September 27, 2026, Jiao [10], Theorem 1.1 reports a smooth Hamiltonian diffeomorphism of the equal-area product S2×S2S^{2}\times S^{2} with two fixed points, using local saddle merging and mixed generating functions. The construction here instead uses a finite-order map on a six-dimensional quadric. None of these fixed-point results is an input to our construction.

Proof mechanism. The key is to transfer a critical-point count from a fixed submanifold to the fixed-point count of a nearby Hamiltonian map. Let AA be a finite-order Hamiltonian map with fixed submanifold FF, whose fixed tangent vectors are exactly those tangent to FF. Extend a function ff on FF to an AA-invariant function KK on the ambient manifold, and let BsB_s be its Hamiltonian flow. If Am=idA^m = \mathrm{id}, invariance gives

(A∘Bε)m=Bmε.(A \circ B_\varepsilon)^m = B_{m\varepsilon}.

For sufficiently small ε>0\varepsilon> 0, the right side has only stationary fixed points. Any fixed point of A∘BεA \circ B_\varepsilon is therefore a critical point of KK fixed by AA. Finite-group averaging makes ambient criticality on FF equivalent to criticality of ff, giving

Fix⁡(A∘Bε)=crit⁡(f).\operatorname{Fix}(A \circ B_\varepsilon) = \operatorname{crit}(f).

This criticality step is the finite-group case of Palais’s principle of symmetric criticality [15], Section 5; the power identity excludes additional fixed points away from FF.

For the quadric, take the Hamiltonian involution

A[z0:z1:z2:z3:z4]=[z0:−z1:−z2:−z3:−z4].A[z_0:z_1:z_2:z_3:z_4] = [z_0:-z_1:-z_2:-z_3:-z_4].

Its fixed locus is the quadric surface Q2≅S2×S2Q^2 \cong S^2 \times S^2. This product admits functions with only three critical points, as recorded by Mare [14], Proposition 3.2.5 through an extension of Takens’s sphere-product construction [18]. We give an explicit formula and compute its entire critical set. The isolated degenerate critical point becomes a degenerate Hamiltonian fixed point. In contrast, symplectic volume forces every smooth function on the ambient Q3Q^3 to have at least four critical points, and a Hamiltonian rotation realizes that minimum.

Section 2 proves the finite-order transfer, including the short-period estimate and invariant extension. Section 3 constructs the Hamiltonian involution and its tangent splitting. Section 4 gives the three-critical-point function. Section 5 proves the critical-number and rational cup-length equalities on Q3Q^3, and Section 6 assembles the construction and verifies degeneracy.

Perturbing a finite-order Hamiltonian map

A short Hamiltonian perturbation of a finite-order map can replace its fixed submanifold by the critical set of a function on that submanifold. The tangent-space hypothesis below is the usual clean fixed-set condition; we state it explicitly to identify the geometric input.

Proposition 2.1. Let (M,ω)(M,\omega) be a closed symplectic manifold, and let A∈Ham⁡(M,ω)A \in\operatorname{Ham}(M,\omega) satisfy Am=idA^m = \mathrm{id} for some positive integer mm. Suppose that F=Fix⁡(A)F = \operatorname{Fix}(A) is a closed embedded submanifold and that

TqF=ker⁡(dAq−id)for every q∈F.T_qF = \ker(\mathrm{d}A_q - \mathrm{id}) \qquad\text{for every } q \in F.

For every smooth function f:F→Rf:F \to\mathbb{R}, there is a smooth AA-invariant extension K:M→RK:M \to\mathbb{R} such that, if BsB_s is its Hamiltonian flow, then for every sufficiently small ε>0\varepsilon> 0 the map

ϕε=A∘Bε\phi_\varepsilon= A \circ B_\varepsilon

is Hamiltonian and satisfies

Fix⁡(ϕε)=crit⁡(f),\operatorname{Fix}(\phi_\varepsilon) = \operatorname{crit}(f),

where the critical set on the right is viewed as a subset of MM.

The proof combines finite-group averaging with the exclusion of short nonconstant periodic trajectories. We establish the latter fact first. Lipschitz control of a vector field gives a lower bound for its nonconstant periods, a classical principle for which Yorke proved the sharp Euclidean estimate [20]. The following elementary local argument suffices here.

Lemma 2.2. Let XX be a smooth vector field on a closed manifold MM, and let Ψt\Psi_t be its flow. There is a number δ>0\delta> 0 such that

Fix⁡(Ψt)={q∈M:X(q)=0}for every 0<t<δ.\operatorname{Fix}(\Psi_t)=\{q\in M:X(q)=0\}\qquad\text{for every }0<t<\delta.

Proof. The flow is defined for all time because MM is compact. Suppose that there were points qjq_j and times tj>0t_j>0 with

tj⟶0,Ψtj(qj)=qj,X(qj)≠0.t_j\longrightarrow0,\qquad\Psi_{t_j}(q_j)=q_j,\qquad X(q_j)\ne0.

After passing to a subsequence, assume that qj→qq_j\to q. Choose coordinates near qq, and a Euclidean ball about its coordinate image whose closure lies inside the coordinate domain. By continuity of the flow, all sufficiently late trajectory segments {Ψs(qj):0≤s≤tj}\{\Psi_s(q_j):0\le s\le t_j\} lie in this ball. The coordinate vector field VV is Lipschitz there with a constant L≥0L\ge0, since its derivative is bounded on the closed ball.

Write uj(s)u_j(s) for the coordinate trajectory and set

Cj=max⁡0≤s≤tj∣u˙j(s)∣>0.C_j=\max_{0\le s\le t_j}|\dot{u}_j(s)|>0.

For any s,r∈[0,tj]s,r\in[0,t_j], convexity of the ball and the Lipschitz estimate give

∣u˙j(s)−u˙j(r)∣=∣V(uj(s))−V(uj(r))∣≤L∣uj(s)−uj(r)∣≤LCj∣s−r∣≤LCjtj.\begin{aligned} |\dot{u}_j(s)-\dot{u}_j(r)| &=|V(u_j(s))-V(u_j(r))|\\ &\le L|u_j(s)-u_j(r)|\le LC_j|s-r|\le LC_jt_j. \end{aligned}

The closed trajectory has zero mean coordinate velocity: ∫0tju˙j(r) dr=0\int_0^{t_j}\dot{u}_j(r)\,dr=0. At a time sjs_j of maximum speed, it follows that

Cj=∣1tj∫0tj(u˙j(sj)−u˙j(r)) dr∣≤LCjtj.C_j=\left|\frac{1}{t_j}\int_0^{t_j}(\dot{u}_j(s_j)-\dot{u}_j(r))\,dr\right|\le LC_jt_j.

Dividing by CjC_j contradicts tj→0t_j\to0. Thus all fixed points of a sufficiently short positive-time map are zeros of XX. Conversely, every zero of XX is stationary and hence fixed by every Ψt\Psi_t. □\square

Proof of Proposition 2.1. The first step is to extend ff so that criticality along FF is equivalent to ambient criticality at points of FF. This is the finite-group case of symmetric criticality [15], which we verify directly. Extend ff smoothly to a function f~\widetilde{f} on MM: in submanifold charts, extend independently of the normal coordinates, and combine these local extensions using a partition of unity. Define

K=1m∑j=0m−1f~∘Aj.K=\frac{1}{m}\sum_{j=0}^{m-1}\widetilde{f}\circ A^j.

Then K∘A=KK\circ A=K and K∣F=fK|_F=f.

Fix q∈Fq\in F, put D=dAqD=dA_q, and consider

Pq=1m∑j=0m−1Dj.P_q=\frac{1}{m}\sum_{j=0}^{m-1}D^j.

Since Dm=id⁡D^m=\operatorname{id}, we have DPq=PqDP_q=P_q. Moreover, Pqv=vP_qv=v whenever Dv=vDv=v. Thus PqP_q is a projection onto ker⁡(D−id⁡)=TqF\ker(D-\operatorname{id})=T_qF. Differentiating the averaged extension gives, for every v∈TqMv\in T_qM,

dKq(v)=df~q(Pqv)=dfq(Pqv).dK_q(v)=d\widetilde{f}_q(P_qv)=df_q(P_qv).

Consequently,

crit⁡(K)∩F=crit⁡(f).(4)\operatorname{crit}(K) \cap F = \operatorname{crit}(f). \tag*{(4)}

Because AA is symplectic and K∘A=KK \circ A = K, its differential carries the Hamiltonian vector field XKX_K to itself. Uniqueness of solutions therefore gives A∘Bs=Bs∘AA \circ B_s = B_s \circ A for every ss. Apply Lemma 2.2 to XKX_K, and choose 0<ε<δ/m0 < \varepsilon< \delta/m. We then have

ϕεm=Bmε,Fix⁡(Bmε)=crit⁡(K).\phi_{\varepsilon}^{m} = B_{m\varepsilon}, \qquad\operatorname{Fix}(B_{m\varepsilon}) = \operatorname{crit}(K).

If ϕε(q)=q\phi_\varepsilon(q) = q, then qq is fixed by BmεB_{m\varepsilon}, so dKq=0\mathrm{d}K_q = 0. It follows that Bε(q)=qB_\varepsilon(q) = q, and hence A(q)=qA(q) = q. Conversely, every point of crit⁡(K)∩F\operatorname{crit}(K) \cap F is fixed by ϕε\phi_\varepsilon. Equation 2.1 proves the asserted equality of fixed and critical sets.

For completeness, the composition is generated by a smooth Hamiltonian on the unit time interval. Choose a Hamiltonian path AtA_t from id⁡\operatorname{id} to AA, generated by HtH_t, and a smooth nondecreasing function a:[0,1]→[0,1]a : [0,1] \to[0,1] equal to 00 near 00 and to 11 near 11. The time-dependent Hamiltonian

Lt(x)={2εa′(2t)K(x),0≤t≤12,2a′(2t−1)Ha(2t−1)(x),12≤t≤1L_t(x) = \begin{cases} 2\varepsilon a'(2t)K(x), & 0 \leq t \leq\frac{1}{2},\\ 2a'(2t-1)H_{a(2t-1)}(x), & \frac{1}{2} \leq t \leq1 \end{cases}

is smooth, since both pieces vanish near the joining time. Its flow first follows Bεa(2t)B_{\varepsilon a(2t)}, and then Aa(2t−1)∘BεA_{a(2t-1)} \circ B_\varepsilon, so its time-one map is ϕε\phi_\varepsilon.

The proposition imposes no bound on the total number of critical points of KK, and requires invariance under AA, not under an isotopy from id⁡\operatorname{id} to AA. A critical point q∉Fq \notin F satisfies ϕε(q)=A(q)≠q\phi_\varepsilon(q) = A(q) \ne q, so it contributes no fixed point.

A Hamiltonian involution of a quadric

We now construct the manifold and involution to which Proposition 2.1 will apply. Let M⊂CP4M \subset\mathbb{C}\mathbb{P}^{4} be the quadric defined in (1.2).

Proposition 3.1. The manifold MM is closed, connected, and of real dimension six. For the restricted Fubini–Study form ω\omega, normalized below, the map

A[z0:z1:z2:z3:z4]=[z0:−z1:−z2:−z3:−z4]A[z_0 : z_1 : z_2 : z_3 : z_4] = [z_0 : -z_1 : -z_2 : -z_3 : -z_4]

is a Hamiltonian involution. Its fixed set is the smooth submanifold

F=M∩{z0=0}≅S2×S2.F = M \cap\{z_0 = 0\} \cong S^2 \times S^2.

For every q∈Fq \in F, the differential dAq\mathrm{d}A_q acts as +id⁡+\operatorname{id} on TqFT_qF and as −id⁡-\operatorname{id} on a complementary real two-dimensional subspace of TqMT_qM.

Proof. In any standard affine chart of CP4\mathbb{C}\mathbb{P}^{4}, the equation defining MM takes the form 1+∑jwj2=01 + \sum_j w_j^2 = 0. Its differential cannot vanish on its zero set: that would force all the wjw_j to be zero. Thus MM is a smooth complex hypersurface, of complex dimension three. As a closed subset of projective space it is compact, and it has no boundary.

To see connectedness, write a nonzero isotropic vector as z=x+iyz = x + iy with x,y∈R5x,y \in\mathbb{R}^{5}. Its real and imaginary parts satisfy

∣x∣=∣y∣>0,x⋅y=0.\lvert x\rvert= \lvert y\rvert> 0, \qquad x \cdot y = 0.

After rescaling, they form an ordered orthonormal pair. Every such pair extends to a positively oriented orthonormal basis, so these pairs form one orbit of the connected group SO⁡(5)\operatorname{SO}(5). Their map (x,y)↦[x+iy](x,y) \mapsto[x+iy] onto MM proves that MM is connected.

We construct the symplectic form with its normalization explicit. On the unit sphere S9⊂C5S^9 \subset\mathbb{C}^5, put

α=∑j=04(xj dyj−yj dxj).\alpha= \sum_{j=0}^{4}(x_j\,\mathrm{d}y_j-y_j\,\mathrm{d}x_j).

The form dα\mathrm{d}\alpha is invariant under scalar phase rotations. If v∈TzS9v \in T_zS^9, then

dα(iz,v)=−2Re⁡∑j=04zj‾vj=0.\mathrm{d}\alpha(iz,v)=-2\operatorname{Re}\sum_{j=0}^{4}\overline{z_j}v_j=0.

Consequently it descends under the quotient map p:S9→CP4p:S^9 \to\mathbb{CP}^4 to a two-form σ\sigma, characterized by p∗σ=dαp^*\sigma=\mathrm{d}\alpha. Local sections show that σ\sigma is smooth and closed. A projective tangent vector has a horizontal lift v∈C5v \in\mathbb{C}^5 satisfying ∑jzj‾vj=0\sum_j \overline{z_j}v_j=0; multiplication by ii on these lifts induces the complex structure. Since

dα(v,iv)=2∣v∣2,\mathrm{d}\alpha(v,iv)=2\lvert v\rvert^2,

σ\sigma is positive on every complex tangent line. Its restriction ω=σ∣M\omega=\left.\sigma\right|_M is therefore closed and nondegenerate.

We use a family of Hamiltonian rotations: equal speeds will give the involution, and unequal speeds will later give a function with four critical points on MM. For a,b,t∈Ra,b,t\in\mathbb{R}, let Φta,b\Phi_t^{a,b} keep z0z_0 fixed and act on (z1,z2)T(z_1,z_2)^{\mathsf{T}} and (z3,z4)T(z_3,z_4)^{\mathsf{T}} by RatR_{at} and RbtR_{bt}, respectively, where

Rθ=(cos⁡θ−sin⁡θsin⁡θcos⁡θ).R_\theta= \begin{pmatrix} \cos\theta& -\sin\theta\\ \sin\theta& \cos\theta \end{pmatrix}.

These real orthogonal, complex unitary transformations preserve the quadric and α\alpha. If Za,bZ_{a,b} is their generating vector field on S9S^9, Cartan's formula gives

ιZa,bdα=−d(α(Za,b)).\iota_{Z_{a,b}}\mathrm{d}\alpha=-\mathrm{d}\bigl(\alpha(Z_{a,b})\bigr).

The function α(Za,b)\alpha(Z_{a,b}) is phase invariant, because the rotations commute with scalar multiplication. It thus descends to a smooth function ga,bg_{a,b} on projective space. The displayed identity descends as well, so Ga,b=−ga,b∣MG_{a,b}=-\left.g_{a,b}\right|_M generates the projected rotations on MM with our convention ιXGa,bω=dGa,b\iota_{X_{G_{a,b}}}\omega=\mathrm{d}G_{a,b}. Explicitly,

Ga,b([z])=2aIm⁡(z1‾z2)+2bIm⁡(z3‾z4)∑j=04∣zj∣2.(5)G_{a,b}([z])=\frac{2a\operatorname{Im}(\overline{z_1}z_2)+2b\operatorname{Im}(\overline{z_3}z_4)}{\sum_{j=0}^{4}\lvert z_j\rvert^2}. \tag*{(5)}

In particular, A=Φ1π,πA=\Phi_1^{\pi,\pi} is Hamiltonian and preserves ω\omega.

A projective line fixed by AA lies in one of the two eigenspaces of diag⁡(1,−1,−1,−1,−1)\operatorname{diag}(1,-1,-1,-1,-1). The positive eigenspace gives the single point [1:0:0:0:0][1:0:0:0:0], which does not lie in MM. The negative eigenspace gives precisely F=M∩{z0=0}F=M\cap\{z_0=0\}. To identify FF, use the invertible complex-linear map

(z1,z2,z3,z4)⟼(z1+iz2z3+iz4−z3+iz4z1−iz2).(z_1,z_2,z_3,z_4)\longmapsto \begin{pmatrix} z_1+iz_2 & z_3+iz_4\\ -z_3+iz_4 & z_1-iz_2 \end{pmatrix}.

Its determinant is z12+z22+z32+z42z_1^2+z_2^2+z_3^2+z_4^2. It therefore identifies FF with the projectivized nonzero rank-one matrices. The map

CP1×CP1⟶F,([u],[v])⟼[uvT]\mathbb{CP}^1\times\mathbb{CP}^1\longrightarrow F,\qquad([u],[v])\longmapsto[uv^{\mathsf{T}}]

is a diffeomorphism: a nonzero column and row recover the two projective factors, smoothly on the corresponding matrix charts. Finally, stereographic coordinates identify each CP1\mathbb{CP}^{1} with S2S^{2}. Only this smooth identification is needed below.

For the tangent-space assertion, fix q∈Fq \in F and choose an affine chart zj=1z_j = 1 with j>0j > 0. In its normalized coordinates, AA sends w0w_0 to −w0-w_0 and leaves all other coordinates fixed. The equation of MM is

w02+1+∑k>0k≠jwk2=0.w_0^{2} + 1 + \sum_{\substack{k>0\\k\ne j}} w_k^{2} = 0.

At qq, where w0=0w_0 = 0, its linearization places no constraint on the w0w_0 direction. The remaining differential is nonzero, so its kernel in the other coordinates is exactly TqFT_qF. Thus

TqM=TqF⊕C∂∂w0,T_qM = T_qF \oplus\mathbb{C}\frac{\partial}{\partial w_0},

with the asserted +1+1 and −1-1 actions. This also verifies directly that FF is a smooth complex surface. The positivity established above therefore makes ω∣F\omega|_F symplectic.

Three critical points on the fixed submanifold

We next construct a function on the fixed submanifold with three critical points. The comparison with functions on the ambient quadric will follow in Section 5. Three-critical-point functions on products of spheres belong to the classical study of critical-point minima and category by Takens [18]; the S2×S2S^2 \times S^2 case is recorded by Mare [14] (Proposition 3.2.5). We use the explicit formula below and determine its entire critical set directly, without importing an existence theorem or assuming that its critical points are Morse.

Proposition 4.1. Let p=(0,0,1)p = (0,0,1) and e=(1,0,0)e = (1,0,0) in R3\mathbb{R}^3. The smooth function

f:S2×S2⟶R,f(x,y)=e⋅x+(x−p)⋅y,(6)f:S^2 \times S^2 \longrightarrow\mathbb{R}, \qquad f(x,y) = e \cdot x + (x-p) \cdot y, \tag*{(6)}

has exactly three critical points.

Proof. For a fixed x≠px \ne p, the yy-dependence is a height function on the second sphere, with critical points y=±(x−p)/∣x−p∣y = \pm(x-p)/|x-p|. At x=px = p that height function is constant, but criticality of the full function still requires its derivative in the xx directions to vanish. We therefore separate this exceptional fiber from the two height branches.

The constrained critical-point equations are

e+y=λx,x−p=μy,∣x∣=∣y∣=1,e+y=\lambda x,\qquad x-p=\mu y,\qquad|x|=|y|=1,

for real multipliers λ,μ\lambda,\mu. If x=px=p, the first equation gives y=−e+λpy=-e+\lambda p. Its unit norm forces λ=0\lambda=0, yielding the unique point

q0=(p,−e),f(q0)=0.q_0=(p,-e),\qquad f(q_0)=0.

Suppose now that x≠px \ne p and set t=∣x−p∣>0t = |x-p| > 0. The second equation in (4.2) gives y=s(x−p)/ty = s(x-p)/t for a sign s∈{1,−1}s \in\{1,-1\}. The first then becomes

te−sp=(λt−s)x.te-sp=(\lambda t-s)x.

Since the left side is nonzero, λt−s≠0\lambda t-s\ne0 and xx lies in the plane spanned by ee and pp. Write x=(u,0,w)x=(u,0,w). The definition of tt gives w=1−t2/2w=1-t^2/2. Taking the first equation in (4.2) against the tangent vector (w,0,−u)(w,0,-u) gives

w+sut=0,sou=−stw.w+\frac{su}{t}=0,\qquad\text{so}\qquad u=-stw.

Consequently the unit-norm condition reduces to

0=(1−t2/2)2(1+t2)−1=t4(t2−3)4.(7)0=(1-t^2/2)^2(1+t^2)-1=\frac{t^4(t^2-3)}{4}. \tag*{(7)}

Because t>0t>0, this forces t=3t=\sqrt{3}, w=−1/2w=-1/2, and u=s3/2u=s\sqrt{3}/2. In particular, the other pole x=−px=-p, for which t=2t=2, is excluded. The two resulting candidates are

qs=((s32,0,−12),(12,0,−s32)),s∈{1,−1}.(8)q_s=\left(\left(\frac{s\sqrt{3}}{2},0,-\frac{1}{2}\right),\left(\frac{1}{2},0,-\frac{s\sqrt{3}}{2}\right)\right),\qquad s\in\{1,-1\}. \tag*{(8)}

Both satisfy (4.2) with λ=μ=s3\lambda=\mu=s\sqrt{3}, and

f(qs)=s332.f(q_s)=s\frac{3\sqrt{3}}{2}.

The cases x=px=p and x≠px\ne p exhaust the domain, proving the count.

Remark 4.2. The critical point q0q_0 is degenerate. Indeed, in local coordinates

x=(u,v,1−u2−v2),y=(−1−a2−b2,a,b),x=(u,v,\sqrt{1-u^2-v^2}),\qquad y=(-\sqrt{1-a^2-b^2},a,b),

the function has expansion

f=va+O(∥(u,v,a,b)∥3).f=va+O(\lVert(u,v,a,b)\rVert^3).

Its Hessian therefore has eigenvalues 1,−1,0,01,-1,0,0, so its rank is two. The complete critical-point calculation proves that this degenerate point is isolated. This is why a count restricted to Morse functions would not capture the example.

The critical number and cup length of the quadric

We now show that four is both the critical number and the rational cup length of the ambient quadric. The lower bound is the classical cup-product obstruction from Lusternik–Schnirelmann theory [12]; see Weber [19] for a modern treatment valid without a Morse assumption. We give the differential-form proof, then obtain the matching upper bound from the unequal-speed rotations of Section 3.

Lemma 5.1. Let MM be a closed smooth manifold and r≥1r\ge1 an integer. Suppose there are classes aj∈Hdj(M;R)a_j\in H^{d_j}(M;\mathbb{R}), with dj>0d_j>0 for 1≤j≤r1\le j\le r, such that a1⌣⋯⌣ar≠0a_1\smile\cdots\smile a_r\ne0. Then every smooth function on MM has at least r+1r+1 critical points.

Proof. Suppose that h:M→Rh:M\to\mathbb{R} has at most rr critical points. List its distinct critical values as c1<⋯<ckc_1<\cdots<c_k, where 1≤k≤r1\le k\le r; the absolute minimum and maximum are among them. For a fixed Riemannian metric let gtg_t be the negative-gradient flow of hh, which is complete by compactness, and put Pb={h≤b}P_b=\{h\le b\}. We first cover MM by kk open sets, each diffeomorphic to a finite disjoint union of coordinate balls. Every closed positive-degree form will be exact on each of these sets; the cover will therefore force every product of kk positive-degree cohomology classes to vanish.

Around the critical points at cic_i, choose an open set UiU_i that is a union of disjoint coordinate balls, and a smaller neighborhood ViV_i with Vi‾⊂Ui\overline{V_i} \subset U_i. A bound on the speed of the flow and the positive distance between M∖UiM \setminus U_i and ViV_i give Ti>0T_i > 0 such that trajectories starting outside UiU_i stay outside ViV_i for 0≤t≤Ti0 \le t \le T_i.

Choose a closed level band {∣h−ci∣≤δi}\{|h-c_i| \le\delta_i\} narrow enough that it contains no critical points outside ViV_i. On the part outside ViV_i, compactness gives a positive lower bound bib_i for ∣∇h∣2|\nabla h|^2; if that set is empty, any positive bib_i will do. Choose 0<ϵi<δi0 < \epsilon_i < \delta_i so that the intervals [ci−ϵi,ci+ϵi][c_i-\epsilon_i,c_i+\epsilon_i] are separated and 2ϵi<Tibi2\epsilon_i < T_i b_i. Along the negative-gradient flow,

ddth(gt(q))=−∣∇h(gt(q))∣2.\frac{d}{dt}h(g_t(q))=-|\nabla h(g_t(q))|^2.

A point of Pci+ϵi∖UiP_{c_i+\epsilon_i}\setminus U_i either already lies below ci−ϵic_i-\epsilon_i or decreases at rate at least bib_i until it reaches that level. It follows that

gTi(Pci+ϵi∖Ui)⊂Pci−ϵi.g_{T_i}(P_{c_i+\epsilon_i}\setminus U_i)\subset P_{c_i-\epsilon_i}.

For i>1i>1, a further fixed flow time Si>0S_i>0 carries Pci−ϵiP_{c_i-\epsilon_i} into Pci−1+ϵi−1P_{c_{i-1}+\epsilon_{i-1}}: the compact band between these levels contains no critical points and has a positive gradient-norm lower bound.

Since c1c_1 is the absolute minimum, Pc1−ϵ1P_{c_1-\epsilon_1} is empty, and (5.1) shows that Pc1+ϵ1⊂U1P_{c_1+\epsilon_1}\subset U_1. Inductively, cover Pci+ϵiP_{c_i+\epsilon_i} by UiU_i and the preimages under gTi+Sig_{T_i+S_i} of the i−1i-1 sets covering the preceding upper sublevel. Equation (5.1) proves that these sets cover. Each gtg_t is a diffeomorphism of MM, so every member remains diffeomorphic to a finite disjoint union of coordinate balls. The last upper sublevel is all of MM, since ckc_k is the absolute maximum.

Write the resulting cover as O1,…,OkO_1,\ldots,O_k. Choose closed forms αj\alpha_j representing aja_j for 1≤j≤k1\le j\le k. Since these forms have positive degree, the Poincaré lemma on each ball gives αj=dλj\alpha_j=d\lambda_j on OjO_j. Choose a smooth partition of unity ψ1,…,ψk\psi_1,\ldots,\psi_k subordinate to the cover. Let χ:[0,1]→[0,1]\chi:[0,1]\to[0,1] be smooth, zero on [0,1/(4k)][0,1/(4k)] and one on [1/(2k),1][1/(2k),1], and put ρj=χ(ψj)\rho_j=\chi(\psi_j). At every point some ψj≥1/k\psi_j\ge1/k, so some ρj\rho_j is identically one nearby. Each ρjλj\rho_j\lambda_j extends smoothly by zero outside OjO_j. The closed forms

αj′=αj−d(ρjλj)\alpha'_j=\alpha_j-d(\rho_j\lambda_j)

represent the same classes as αj\alpha_j. Near every point, at least one of these forms vanishes, and hence α1′∧⋯∧αk′=0\alpha'_1\wedge\cdots\wedge\alpha'_k=0. Thus a1⌣⋯⌣ak=0a_1\smile\cdots\smile a_k=0, contradicting the nonvanishing of a1⌣⋯⌣ara_1\smile\cdots\smile a_r because k≤rk\le r.

Extension of coefficients H∗(M;Q)→H∗(M;R)H^*(M;\mathbb{Q})\to H^*(M;\mathbb{R}) is injective for a closed manifold and respects cup products. The lemma therefore also gives

cuplength⁡(M;Q)≤Crit⁡(M),(9)\operatorname{cuplength}(M;\mathbb{Q})\le\operatorname{Crit}(M), \tag*{(9)}

with the unit-inclusive convention of Section 1. It also recovers the symplectic-volume obstruction used in our example.

Corollary 5.2. Every smooth function on a closed connected symplectic manifold of real dimension 2n2n, with n≥1n\ge1, has at least n+1n+1 critical points.

Proof. The class [ω]n[\omega]^n is nonzero, since ∫Mωn>0\int_M\omega^n>0 in the symplectic orientation. Apply Lemma 5.1 with each aj=[ω]a_j=[\omega]. ∎

Corollary 5.3. For the quadric threefold in (1.2),

Crit⁡(Q3)=cuplength⁡(Q3;Q)=4.\operatorname{Crit}(Q^3)=\operatorname{cuplength}(Q^3;\mathbb{Q})=4.

Proof. Let h=c1(OQ3(1))∈H2(Q3;Z)h=c_1(\mathcal{O}_{Q^3}(1))\in H^2(Q^3;\mathbb{Z}) be the hyperplane class. Since Q3Q^3 has degree two in CP4\mathbb{CP}^4,

h3=2[pt],h^3=2[\mathrm{pt}],

where [pt][\mathrm{pt}] evaluates to one on the fundamental class. This product remains nonzero over Q\mathbb{Q}, giving cuplength⁡(Q3;Q)≥4\operatorname{cuplength}(Q^3;\mathbb{Q})\ge4.

For the matching upper bound, take the Hamiltonian G1,2G_{1,2} of (3.1). Its critical points are precisely the zeros of its Hamiltonian vector field. That field is induced on Q3Q^3 by the complex-linear generator

T=diag⁡(0,J,2J),J=(0−110).T=\operatorname{diag}(0,J,2J),\qquad J=\begin{pmatrix}0&-1\\1&0\end{pmatrix}.

The induced projective vector field vanishes at [z][z] exactly when Tz∈CzTz\in\mathbb{C}z. The eigenvalues 0,±i,±2i0,\pm i,\pm2i of TT are distinct, so its projective zeros are its five eigenlines. The zero eigenline [1:0:0:0:0][1:0:0:0:0] misses Q3Q^3, whereas the other four lines

[0:1:±i:0:0],[0:0:0:1:±i][0:1:\pm i:0:0],\qquad[0:0:0:1:\pm i]

all satisfy the quadric equation. Thus G1,2G_{1,2} has exactly four critical points on Q3Q^3. Together with (5.2), this gives

4≤cuplength⁡(Q3;Q)≤Crit⁡(Q3)≤4.4\le\operatorname{cuplength}(Q^3;\mathbb{Q})\le\operatorname{Crit}(Q^3)\le4.

The Hamiltonian counterexample

We now apply the finite-order reduction to the function on the fixed surface and compare the resulting fixed-point count with the ambient critical number and cup length.

Proof of Theorem 1.1. Take the symplectic quadric (M,ω)(M,\omega) and Hamiltonian involution AA from Proposition 3.1. Its fixed submanifold FF is diffeomorphic to S2×S2S^2\times S^2. Transport the function of Proposition 4.1 to FF; it has exactly three critical points.

Proposition 2.1 provides a smooth AA-invariant extension K:M→RK:M\to\mathbb{R} with K∣F=fK|_F=f. Let BsB_s denote its Hamiltonian flow. For all sufficiently small ε>0\varepsilon>0, the map

ϕ=A∘Bε\phi=A\circ B_\varepsilon

is a Hamiltonian diffeomorphism and satisfies Fix⁡(ϕ)=crit⁡(f)\operatorname{Fix}(\phi)=\operatorname{crit}(f), with the critical points viewed as points of FF. The exact comparison in Corollary 5.3 now gives

#Fix⁡(ϕ)=3<4=Crit⁡(M)=cuplength⁡(M;Q).\#\operatorname{Fix}(\phi)=3<4=\operatorname{Crit}(M)=\operatorname{cuplength}(M;\mathbb{Q}).

It remains to verify the asserted degeneracy. Write ωF=ω∣F\omega_F=\omega|_F, which is symplectic, and let XfX_f be the Hamiltonian vector field of ff on (F,ωF)(F,\omega_F). Invariance of KK makes XKX_K tangent to FF, and restricting its defining equation to TFTF gives XK∣F=XfX_K|_F=X_f. Indeed, at q∈Fq\in F invariance gives dAqXK(q)=XK(q)dA_qX_K(q)=X_K(q), and the fixed tangent space is TqFT_qF. Thus BsB_s preserves FF, and its restriction is the Hamiltonian flow of ff on FF. At the critical point q0q_0 of Remark 4.2, differentiating ιXfωF=df\iota_{X_f}\omega_F=df gives

ωF(DXf(q0)v,w)=Hess⁡q0f(v,w).(10)\omega_F(DX_f(q_0)v,w)=\operatorname{Hess}_{q_0}f(v,w). \tag*{(10)}

Thus any nonzero vector vv in the Hessian kernel satisfies DXf(q0)v=0DX_f(q_0)v = 0. Since A∣F=idA|_F = \mathrm{id} and q0q_0 is an equilibrium,

dϕq0v=exp⁡(εDXf(q0))v=v.d\phi_{q_0}v = \exp(\varepsilon DX_f(q_0))v = v.

The fixed point q0q_0 therefore has eigenvalue 1 in its linearization. This argument takes place entirely in the invariant subspace Tq0FT_{q_0}F; no assumption on the normal Hessian of the extension KK is needed. □\square

References

  1. [1]V. I. Arnol’d, First steps in symplectic topology, Russian Mathematical Surveys 41 (1986), no. 6, 1–21. doi:10.1070/RM1986v041n06ABEH004221.DOI
  2. [2]L. Buhovsky, V. Humilière, and S. Seyfaddini, A C⁰ counterexample to the Arnold conjecture, Inventiones Mathematicae 213 (2018), no. 2, 759–809. doi:10.1007/s00222-018-0797-x.DOI
  3. [3]C. C. Conley and E. Zehnder, The Birkhoff–Lewis fixed point theorem and a conjecture of V. I. Arnold, Inventiones Mathematicae 73 (1983), 33–49. doi:10.1007/BF01393824.DOI
  4. [4]A. Floer, Symplectic fixed points and holomorphic spheres, Communications in Mathematical Physics 120 (1989), no. 4, 575–611. doi:10.1007/BF01260388.DOI
  5. [5]B. Fortune, A symplectic fixed point theorem for CPⁿ, Inventiones Mathematicae 81 (1985), no. 1, 29–46. doi:10.1007/BF01388770.DOI
  6. [6]K. Fukaya and K. Ono, Arnold conjecture and Gromov–Witten invariant, Topology 38 (1999), no. 5, 933–1048. doi:10.1016/S0040-9383(98)00042-1.DOI
  7. [7]R. Golovko, On variants of Arnold conjecture, Archivum Mathematicum 56 (2020), no. 5, 277–286. doi:10.5817/AM2020-5-277.DOI
  8. [8]W. Gong, Degenerate symplectic fixed points and Gromov–Witten invariants, arXiv:2507.04191v5, December 21, 2025.arxiv.org/abs/2507.04191
  9. [9]H. Hofer, Lusternik–Schnirelman-theory for Lagrangian intersections, Annales de l’Institut Henri Poincaré C, Analyse non linéaire 5 (1988), no. 5, 465–499. doi:10.1016/S0294-1449(16)30339-0.DOI
  10. [10]H. Jiao, A smooth Hamiltonian diffeomorphism with two fixed points on S² × S², arXiv:2609.33626v1, September 27, 2026.arxiv.org/abs/2609.33626
  11. [11]G. Liu and G. Tian, Floer homology and Arnold conjecture, Journal of Differential Geometry 49 (1998), no. 1, 1–74. doi:10.4310/jdg/1214460936.DOI
  12. [12]L. Lusternik and L. Schnirelmann, Méthodes topologiques dans les problèmes variationnels. I. Espaces à un nombre fini de dimensions, Actualités scientifiques et industrielles 188, Hermann, Paris, 1934.
  13. [13]R. Ma, Proofs on Arnold conjectures, arXiv:0808.0613v7, July 5, 2013.
  14. [14]A.-L. Mare, Topology of isotropy orbits, doctoral dissertation, Universität Augsburg, 1997. Author-hosted text.
  15. [15]R. S. Palais, The principle of symmetric criticality, Communications in Mathematical Physics 69 (1979), no. 1, 19–30. doi:10.1007/BF01941322.DOI
  16. [16]Y. Ruan, Virtual neighbourhoods and pseudo-holomorphic curves, Turkish Journal of Mathematics 23 (1999), no. 1, 161–232. Publisher record; arXiv:alg-geom/9611021v2.arxiv.org/abs/alg-geom/9611021
  17. [17]Y. B. Rudyak and J. Oprea, On the Lusternik–Schnirelmann category of symplectic manifolds and the Arnold conjecture, Mathematische Zeitschrift 230 (1999), no. 4, 673–678. doi:10.1007/PL00004709.DOI
  18. [18]F. Takens, The maximal number of critical points of a function on a compact manifold and the Lusternik–Schnirelman category, Inventiones Mathematicae 6 (1968), no. 3, 197–244. doi:10.1007/BF01404825.DOI
  19. [19]J. Weber, Conley pairs in geometry—Lusternik-Schnirelmann theory and more, Expositiones Mathematicae 37 (2019), no. 1, 25–47. doi:10.1016/j.exmath.2017.10.003.
  20. [20]J. A. Yorke, Periods of periodic solutions and the Lipschitz constant, Proceedings of the American Mathematical Society 22 (1969), no. 2, 509–512. doi:10.1090/S0002-9939-1969-0245916-7.DOI

Paper details

Contents