Introduction and main results
A geometric condition used in defining a submanifold of a smooth manifold or a connection on a principal bundle over a smooth manifold depends on the ambient structure. Allowing that structure to evolve by a geometric flow therefore changes the equation that the submanifold or connection must satisfy. It is natural to study coupled geometric flows, pairing the ambient evolution with a flow of the submanifold or connection and asking whether a geometrically preferred property of the submanifold (for instance minimality or being calibrated) or of the connection (for instance the connection being an instanton or Yang-Mills) is preserved along the coupled system of flows. Such a study can have consequences in a combination of manners. Depending on what properties are preserved, the coupled system can tell us about the preferred geometric flows for the ambient structures and the sub-objects. It can also tells us that what geometric property is better behaved and worth looking into in detail. This is the motivating philosophy behind the present paper and for us the setting is that of a seven dimensional smooth manifold with a G 2 G_2 G 2 -structure where the underlying geometry is related to the nonassociative algebra of the octonions O \mathbb{O} O . A natural first choice in the coupled system is to use a familiar geometric flow for the ambient structure as well as for the special structure which will be submanifolds or instantons in our setting. However, a priori there is no reason for the two velocities to preserve the equation relating them. We must differentiate that equation and check whether the contributions cancel. This question is motivated by some concrete affirmative answers in literature where such a cancellation do occur: the preservation of Lagrangian submanifolds under mean curvature flow coupled to Kähler–Ricci flow following Smoczyk [34 , 33 ] and the work of Lotay and Pacini in [22 ] where among other things, they prove preservation of the condition of being totally real in an almost Kähler manifold under the coupling of the Maslow flow and the symplectic curvature flow of Streets and Tian [35 ] . Lotay and Pacini also proposed studying analogous couplings for coassociatives and G 2 G_2 G 2 instantons [22 Section 1] . We study this question by first identifying the failure of preservation for the unmodified equations and then constructing elliptic corrections that preserve the required conditions.
Let M M M be a smooth seven dimensional manifold. A positive three form φ \varphi φ is a differential form which is pointwise linearly equivalent to the model in (12) on R 7 \mathbb{R}^7 R 7 . It defines a G 2 G_2 G 2 structure, a metric g = g φ g = g_\varphi g = g φ , and an orientation, with the last two depending on φ \varphi φ in a nonlinear fashion. The existence of such a structure on a smooth manifold is a topological condition. A G 2 G_2 G 2 structure on M M M exists if and only if the structure group of the frame bundle reduces from the group G L ( 7 , R ) \mathrm{GL}(7,\mathbb{R}) GL ( 7 , R ) to the Lie group G 2 G_2 G 2 . This condition is also equivalent to the existence of a positive 3-form φ \varphi φ as above which is equivalent to the vanishing of the first and second Stiefel-Whitney classes which in turn, are equivalent to M M M being orientable and spinable, respectively. We use Karigiannis’ conventions [15 ] and write ψ = ∗ φ \psi= *\varphi ψ = ∗ φ . We fix the orientation and consider positive forms with that orientation. The G 2 G_2 G 2 -structure is closed when d φ = 0 \mathrm{d}\varphi= 0 d φ = 0 , co-closed when d ψ = 0 \mathrm{d}\psi= 0 d ψ = 0 , and torsion-free when ∇ φ φ = 0 \nabla^\varphi\varphi= 0 ∇ φ φ = 0 where ∇ φ \nabla^\varphi ∇ φ is the Levi-Civita connection of the metric induced by φ \varphi φ . Torsion-freeness of φ \varphi φ is equivalent to closedness and co-closedness together. We will keep these assumptions separate, since the two constructions use different ones. See [14 , 31 , 3 , 15 ] for more on manifolds with G 2 G_2 G 2 structures.
Let M M M be a manifold with a G 2 G_2 G 2 structure φ \varphi φ . Let i : C 4 ↪ M 7 i:C^4\hookrightarrow M^7 i : C 4 ↪ M 7 be an embedding, and let A A A be a connection on a principal G G G -bundle P → M P \to M P → M . The curvature F A F_A F A is a two form with values in the adjoint bundle g \mathfrak{g} g and P P P . The defining equations for coassociativity and the G 2 G_2 G 2 instanton condition are, respectively,
i ∗ φ = 0 , F A ∧ ψ = 0. (1) i^*\varphi= 0,\qquad F_A \wedge\psi= 0. \tag*{(1)} i ∗ φ = 0 , F A ∧ ψ = 0. (1)
We orient a coassociative so that i ∗ ψ = vol i ∗ g i^*\psi= \operatorname{vol}_{i^*g} i ∗ ψ = vol i ∗ g .
As is evident from their defining equations, both conditions depend on the ambient G 2 G_2 G 2 structure. A submanifold can be coassociative for one structure and fail to be coassociative for a nearby structure. Similarly, a connection which is an instanton for one structure need not remain an instanton when the structure changes. Thus, when we evolve the ambient structure, we must also choose how the submanifold or connection moves if we want to preserved the special conditions in (1) .
If d ψ = 0 \mathrm{d}\psi= 0 d ψ = 0 , then ψ \psi ψ is a calibration form in the sense of Harvey–Lawson [13 ] and every coassociative is minimal. In the first coupled system which we study this paper, we only assume d φ = 0 \mathrm{d}\varphi= 0 d φ = 0 and not d ψ = 0 \mathrm{d}\psi= 0 d ψ = 0 hence the initial coassociative (or the evolving ones) need not be minimal. If a connection is an instanton and the underlying G 2 G_2 G 2 structure is coclosed then the connection satisfies a version of the Yang-Mills equation with an extra term consisting of the torsion of the G 2 G_2 G 2 structure. However, it need not satisfy the ordinary Yang-Mills equation. More about coassociatives and calibrations can be found in the seminal paper on calibrated geometry by Harvey–Lawson [13 ] and the lecture notes by Lotay [20 ] . In both the coupled systems, our choice of assumptions of d φ = 0 \mathrm{d}\varphi= 0 d φ = 0 and d ψ = 0 \mathrm{d}\psi= 0 d ψ = 0 is motivated by the deformation theory of the two geometric constraints in (1) . When d φ = 0 \mathrm{d}\varphi= 0 d φ = 0 , the normal linearization of the coassociative equation identifies with the exterior derivative on anti-self-dual two-forms, and the nonlinear deformation map is surjective on the space of exact three-forms.
These properties underlie the unobstructed deformation theory of compact coassociatives see [27 ] and [20 Theorem 5.5] . Thus, the natural background in which to study possible evolutions of coassociatives is when the underlying G 2 G_2 G 2 structure is close. Similarly, for G 2 G_2 G 2 -instantons, the condition d ψ = 0 d\psi=0 d ψ = 0 , together with the Bianchi identity places the linearization of the instanton equation F A ∧ ψ = 0 F_A\wedge\psi=0 F A ∧ ψ = 0 in an elliptic deformation complex, whose first cohomology describes infinitesimal deformations modulo gauge see for instance [Wal17] or [20 §8.4] . Although the instanton deformation problem may have obstructions, this provides its standard elliptic framework. This is the reason for working with coclosed G 2 G_2 G 2 structures in the second coupled system in the paper. For the system of coupled flows in this paper, we consider Bryant’s Laplacian flow of closed G 2 G_2 G 2 structures [3 , 5 ] coupled to mean curvature flow of an initial coassociative embedding, and Yang–Mills flow of connections coupled to a prescribed evolution of co-closed G 2 G_2 G 2 structures, a particular case of which is Grigorian’s modified Laplacian coflow [12 ] . In both the cases we first calculate the error in preservation of the coassociative or instanton condition. We then solve a linear elliptic equation to introduce a correction which cancels that error. However, this procedure is not straight-forward and there are three separate issues: 1) the corrected system of PDEs must be defined on a neighborhood of the initial data, 2) it must have solutions for short-time, 3) and those solutions must preserve the required condition. Long-time existence and convergence of solutions is an entirely different issue and require further analytic estimates.
We write δ g = d g ∗ \delta_g=d_g^* δ g = d g ∗ and use the Hodge Laplacian Δ g = d δ g + δ g d \Delta_g=d\delta_g+\delta_gd Δ g = d δ g + δ g d . Subscripts on operators indicate the current metric. For a closed G 2 G_2 G 2 structure, if τ \tau τ denotes the only surviving intrinsic torsion form of φ \varphi φ then τ = δ φ φ \tau=\delta_\varphi\varphi τ = δ φ φ [24 ] and Laplacian flow is the evolution equation ∂ t φ = Δ φ φ = d τ \partial_t\varphi=\Delta_\varphi\varphi=d\tau ∂ t φ = Δ φ φ = d τ . This flow was introduced by Bryant [3 ] as a way to obtain holonomy G 2 G_2 G 2 manifolds using parabolic methods. The short-time existence and uniqueness of solutions of this flow on closed manifolds was proved by Bryant and Xu [5 ] . Further analytic and geometric properties of the Laplacian flow were proved by Lotay and Wei who established derivative estimates, uniqueness and compactness [24 ] , real analyticity [25 ] , and dynamical stability of torsion-free G 2 G_2 G 2 structures [LW19b]. There is an analogous proposal for flows of coclosed G 2 G_2 G 2 -structures called the Laplacian coflow of coclosed G 2 G_2 G 2 structures and that flow and its modification are studied in [17 , 12 ] . For connections on principal bundles, Yang–Mills flow on torsion-free special holonomy manifolds is studied by Oliveira and Waldron [29 ] .
Motivated by the results on the preservation of the Lagrangian condition in an ambient Kähler-Einstein manifold or when the mean curvature flow of an initial Lagrangian in a Kähler manifold is coupled with the Kähler-Ricci flow of the ambient Kähler structures in [34 , 33 ] as well as the results on the Maslov flow of totally real submanifolds in an ambient symplectic curvature flow background in [22 , 21 ] , one might naively want to couple the mean curvature flow of an initial coassociative with the Laplacian flow of the ambient closed G 2 G_2 G 2 -structures and hope to preserve the coassociative condition. Trying to attempt the same, for a coassociative moving by mean curvature in an ordinary Laplacian-flow background, our first calculation in Proposition Proposition 3.2 gives
∂ ∂ t ( i ∗ φ ) = d C ( ∗ C i ∗ τ ) . \frac{\partial}{\partial t}(i^*\varphi)=d_C(*_C i^*\tau). ∂ t ∂ ( i ∗ φ ) = d C ( ∗ C i ∗ τ ) .
A priori , there is no reason for this expression to vanish. In fact, in Proposition Proposition 3.8 we give an explicit example on the compact torus where we show that this term is non-zero even for arbitrarily small exact perturbations of a flat structure and in the same cohomology class. Thus, the naive expectation of coupling the mean curvature flow of coassociatives with the Laplacian flow of closed G 2 G_2 G 2 structures cannot work. Nevertheless, we circumvent this problem as we now describe. We cancel the error introduced by the previous equation by adding a term L X φ \mathcal{L}_X\varphi L X φ to the ambient Laplacian flow. The restriction of the vector field X X X to the submanifold is a vector field in the normal bundle of the coassociative determined by the deformation operator of the coassociative equation. There would be many such normal vector fields and we choose the vector field with the smallest L 2 L^2 L 2 norm. We then extend it to the ambient manifold M M M using fixed data on a tubular neighbourhood of the coassociative and different choices of this extension give pairs which are related by diffeomorphisms of M M M . Equivalently, we may keep the evolution of the G 2 G_2 G 2 structures by the Laplacian flow and in that case we modify the normal velocity of the coassociatives. The Laplacian flow equation determines one part of this velocity. The remaining part lies in the tangent space to the local family of coassociatives and is the negative L 2 L^2 L 2 gradient of volume on that family. Together these two parts give the compatible velocity closest to the mean curvature vector in L 2 L^2 L 2 . Thus, our construction selects an ordinary differential equation on a moving finite-dimensional deformation space of the coassociatives. We explain this interpretation in Proposition Proposition 3.10 .
Having established the coupled system, we look at its analytic properties. The coassociative existence proof uses the closed-structure version of McLean’s deformation theory [27 ] , including the varying ambient results of Kovalev and Lotay [16 Theorem 2.7-2.8] . These give a finite dimensional family of nearby coassociatives. An ordinary differential equation in its harmonic parameters selects the path with the prescribed normal velocity. Quantitative control of this family then allows us to lift a stable ambient flow for all time, and we use the ambient Laplacian flow stability theorem from [LW19b] for this argument. We mention that Lambert and Lotay [19 ] also relate Laplacian flow and mean curvature flow in the setting of semi-flat coassociative fibrations, where mean curvature flow acts on a space-like three-dimensional base. Here, we move a chosen four-dimensional coassociative in a seven-dimensional manifold, without a fibration or symmetry assumptions.
For connections, we prescribe a smooth family of closed positive four forms ψ t \psi_t ψ t . The correction in this case has two parts. A torsion term removes the difference between ordinary Yang–Mills flow and the negative gradient of the instanton defect energy. An elliptic correction then accounts for ∂ t ψ t \partial_t\psi_t ∂ t ψ t . We explain that the theory is most well-behaved near an acyclic instanton (see page 5 or [38 Def. 3.6] ). Near an acyclic instanton, the inverse defining this correction exists even at connections which are not instantons. The energy identity proves that an initially zero instanton defect remains zero, thus proving the preservation of the instanton condition along the prescribed flow of ψ t \psi_t ψ t .
We will see that decay of the instanton defect alone does not give convergence of the connection. To obtain convergence from connections which are not initially instantons, we also need to control the connection itself. Acyclicity provides this control after fixing gauge: the linearized operator has a positive spectral gap, and the nonlinear terms can be absorbed in the corresponding Sobolev energy estimate. The same equation therefore attracts sufficiently close connections to an instanton. We prove this in both the cases, when a coclosed G 2 G_2 G 2 structure is fixed and also along an explicitly controlled flow of coclosed G 2 G_2 G 2 structures. For the ambient flow of coclosed G 2 G_2 G 2 structures we follow our own philosophy of using a well-studied flow of coclosed G 2 G_2 G 2 structures. The flow we look at is the modified Laplacian coclow of coclosed G 2 G_2 G 2 structures which was introduced by Grigorian [12 ] . We obtain global convergence for the coupling system when the constant parameter in the modified coflow is 0 and the evolution takes place near a torsion-free G 2 G_2 G 2 structure. We explain both the settings in detail below.
Throughout the paper, the manifolds and submanifolds are compact and have no boundary, and initial data and solutions are smooth. Sobolev and Hölder norms use fixed smooth reference metrics and connections. We fix 0 < α < 1 0 < \alpha< 1 0 < α < 1 when using C k , α C^{k,\alpha} C k , α norms.
Coupling with coassociatives
Let C 4 ↪ M 7 C^4 \hookrightarrow M^7 C 4 ↪ M 7 be a coassociative submanifold. If g g g is a Riemannian metric induced by the underlying G 2 G_2 G 2 structure, we use on C C C the induced metric g C = i ∗ g g_C=i^*g g C = i ∗ g and we denote its normal bundle by ν \nu ν . With ⊥ \perp ⊥ denoting normal projection, the second fundamental form and the mean curvature vector of C C C in M M M are
I I ( U , V ) = ( ∇ U V ) ⊥ , H = tr g C I I , U , V ∈ T C . \mathrm{II}(U,V)=(\nabla_UV)^\perp,\qquad H=\operatorname{tr}_{g_C}\mathrm{II},\qquad U,V\in TC. II ( U , V ) = ( ∇ U V ) ⊥ , H = tr g C II , U , V ∈ T C .
To describe the change of the coassociative equation under a normal motion, we associate a two form to a normal field V V V and then take its exterior derivative. We set
J i V = i ∗ ( V ⌟ φ ) , D i : Γ ( ν i ) ⟶ d C Ω 2 ( C ) with D i V = d C J i V , θ = i ∗ τ . J_iV=i^*(V\mathbin{\lrcorner}\varphi),\qquad D_i:\Gamma(\nu_i)\longrightarrow d_C\Omega^2(C)\text{ with }D_iV=d_CJ_iV,\qquad\theta=i^*\tau. J i V = i ∗ ( V ┘ φ ) , D i : Γ ( ν i ) ⟶ d C Ω 2 ( C ) with D i V = d C J i V , θ = i ∗ τ .
At a coassociative, J i J_i J i identifies ν i \nu_i ν i with the anti-self-dual two forms on C C C , for the orientation determined by i ∗ ψ = vol g C i^*\psi=\operatorname{vol}_{g_C} i ∗ ψ = vol g C . We use the following notations: π − = ( id − ∗ C ) / 2 \pi_-=(\operatorname{id}-*_{C})/2 π − = ( id − ∗ C ) /2 , δ C = d C ∗ \delta_C=d_C^* δ C = d C ∗ and Δ C = d C δ C + δ C d C \Delta_C=d_C\delta_C+\delta_Cd_C Δ C = d C δ C + δ C d C . The Green’s operator G G G of Δ C \Delta_C Δ C is zero on harmonic forms, and the exact Green’s will be suppressed. The harmonic anti-self-dual two forms are denoted by H − 2 ( C ) \mathcal{H}^2_-(C) H − 2 ( C ) and their dimension by b 2 − ( C ) b_2^-(C) b 2 − ( C ) .
Near a coassociative pair ( φ , C ) (\varphi,C) ( φ , C ) , D i D_i D i has a minimum norm right inverse R i R_i R i on exact three forms, with D i R i q = q D_iR_iq=q D i R i q = q and R i q ⊥ ker D i R_iq\perp\ker D_i R i q ⊥ ker D i . Choose the diffeomorphism-equivariant extension operator E φ , i E_{\varphi,i} E φ , i defined in (23) , fixing its tubular radius and cutoff. It takes a normal vector field along i ( C ) i(C) i ( C ) and extends it to a vector field on M M M . Put (see § 3 for more details)
W φ , i = − R i ( d C θ + D i H ) , X φ , i = E φ , i W φ , i . (2) W_{\varphi,i}=-R_i(d_C\theta+D_iH),\qquad X_{\varphi,i}=E_{\varphi,i}W_{\varphi,i}. \tag*{(2)} W φ , i = − R i ( d C θ + D i H ) , X φ , i = E φ , i W φ , i . (2)
Here W φ , i W_{\varphi,i} W φ , i is a normal vector field and X φ , i X_{\varphi,i} X φ , i is its extension to the manifold M M M . Even though the subscripts φ , i \varphi,i φ , i are cluttering the notations, we still keep them to emphasise the dependence of these objects on the pair ( φ , C ) (\varphi,C) ( φ , C ) . The term d C θ d_C\theta d C θ is caused by the ambient Laplacian flow, and D i H D_iH D i H is caused by mean curvature flow. Thus W W W is chosen to cancel their sum. The equation D i R i q = q D_iR_iq=q D i R i q = q is a right-inverse identity on exact three forms and we remark that it does not say that D i D_i D i has zero kernel. Orthogonality to that kernel chooses one solution. We construct these operators in § 3 . Having discussed the notations, our first main theorem is the following.1
Theorem 1.1. Let M 7 M^7 M 7 be compact with a closed G 2 G_2 G 2 structure φ 0 \varphi_0 φ 0 , and let i 0 : C 4 ↪ M i_0:C^4\hookrightarrow M i 0 : C 4 ↪ M be a compact embedded coassociative. Fix the extension data above in (2) . The coupled system
∂ t φ t = Δ φ t φ t + L X φ t , i t φ t , d φ t = 0 , ∂ t i t = H t (3) \partial_t\varphi_t=\Delta_{\varphi_t}\varphi_t+\mathcal{L}_{X_{\varphi_t,i_t}}\varphi_t,\qquad d\varphi_t=0,\qquad\partial_ti_t=H_t \tag*{(3)} ∂ t φ t = Δ φ t φ t + L X φ t , i t φ t , d φ t = 0 , ∂ t i t = H t (3) has a unique smooth solution from ( φ 0 , i 0 ) (\varphi_0,i_0) ( φ 0 , i 0 ) for a short time. Every smooth solution from these initial data satisfies i t ∗ φ t = 0 i_t^*\varphi_t=0 i t ∗ φ t = 0 and [ φ t ] = [ φ 0 ] [\varphi_t]=[\varphi_0] [ φ t ] = [ φ 0 ] on each interval where the pair remains in a neighborhood on which R i R_i R i and the fixed tubular extension are defined. At a coassociative pair the normal vector field W W W along C C C is explicitly
W = − 2 J i − 1 π − δ C G C d C ( ∗ C θ ) . (4) W=-2J_i^{-1}\pi_{-\delta_C}G_Cd_C(*_C\theta). \tag*{(4)} W = − 2 J i − 1 π − δ C G C d C ( ∗ C θ ) . (4)
Remark 1.2. No condition on b ˉ 2 ( C ) \bar b_2(C) b ˉ 2 ( C ) is required in Theorem 1.1 .
The normal field W W W has the least L 2 L^2 L 2 norm among those cancelling the pullback derivative. A version of the theorem for a fixed embedding is given in Corollary 3.7. The system (3) is a modified coupled system of the Laplacian flow and the mean curvature flow and for comparison, the unmodified system is
∂ t φ t = Δ g t φ t , d φ t = 0 , g t = g φ t , ∂ t i t = H t . (5) \partial_t\varphi_t=\Delta_{g_t}\varphi_t,\qquad d\varphi_t=0,\qquad g_t=g_{\varphi_t},\qquad\partial_ti_t=H_t. \tag*{(5)} ∂ t φ t = Δ g t φ t , d φ t = 0 , g t = g φ t , ∂ t i t = H t . (5)
In this description the coassociative has normal velocity
V = H + W = Π i H − R i d C θ , V=H+W=\Pi_iH-R_id_C\theta, V = H + W = Π i H − R i d C θ ,
where Π i \Pi_i Π i is the L 2 L^2 L 2 orthogonal projection onto ker D i \ker D_i ker D i . The term − R i d C θ -R_id_C\theta − R i d C θ is the compatible normal velocity of least L 2 L^2 L 2 norm and its geometric motivation is explained in Proposition 3.10 . If b ˉ 2 ( C ) = 0 \bar b_2(C)=0 b ˉ 2 ( C ) = 0 , the ambient form determines a unique nearby coassociative image at each sufficiently small time.
For mean curvature flow with the ambient metric fixed, Tsai and Wang proved stability for certain calibrated zero sections [36 ] . They also proved stability for minimal submanifolds satisfying their strong stability condition [37 ] .
We fix the initial parametrizations and measure closeness of embeddings in a normal graph chart at the reference embedding. The assumption [ φ 0 ] = [ φ ˉ ] [\varphi_0]=[\bar\varphi] [ φ 0 ] = [ φ ˉ ] allows us to apply the stability theorem for Laplacian flow.
Theorem 1.3. Let ( M , φ ˉ ) (M,\bar\varphi) ( M , φ ˉ ) be compact with φ ˉ \bar\varphi φ ˉ torsion-free and let i ˉ \bar i i ˉ be a compact coassociative embedding. Fix admissible tubular extension data used in Theorem 1.1 . There exists an integer N N N and a C N , α C^{N,\alpha} C N , α neighborhood of ( φ ˉ , i ˉ ) (\bar\varphi,\bar i) ( φ ˉ , i ˉ ) such that every smooth initial pair in this neighborhood satisfying
d φ 0 = 0 , [ φ 0 ] = [ φ ˉ ] , i 0 ∗ φ 0 = 0 d\varphi_0=0,\qquad[\varphi_0]=[\bar\varphi],\qquad i_0^*\varphi_0=0 d φ 0 = 0 , [ φ 0 ] = [ φ ˉ ] , i 0 ∗ φ 0 = 0 has a unique global solution of (3) . The solution converges smoothly and exponentially to a torsion-free coassociative pair. The limiting ambient G 2 G_2 G 2 structure belongs to Diff 0 ( M ) ⋅ φ ˉ \operatorname{Diff}_0(M)\cdot\bar\varphi Diff 0 ( M ) ⋅ φ ˉ . The limiting coassociative may vary in its local deformation space.
Coupling with instantons
Let P → M P\to M P → M be a principal G G G -bundle with G G G compact. Fix an Ad ( G ) \operatorname{Ad}(G) Ad ( G ) -invariant positive definite inner product on g \mathfrak{g} g , and set E = ad P E=\operatorname{ad}P E = ad P . We write Ω k ( M , E ) \Omega^k(M,E) Ω k ( M , E ) for the E E E -valued k k k -forms and use the fixed inner product in the adjoints and norms of these forms. Let ψ \psi ψ be a closed positive four-form, with associated metric g ψ g_\psi g ψ and dual positive three-form φ = ∗ ψ ψ \varphi=*_\psi\psi φ = ∗ ψ ψ . For a connection A A A on P P P with curvature F A F_A F A and for ( f , a ) ∈ Ω 0 ( M , E ) ⊕ Ω 1 ( M , E ) (f,a)\in\Omega^0(M,E)\oplus\Omega^1(M,E) ( f , a ) ∈ Ω 0 ( M , E ) ⊕ Ω 1 ( M , E ) , set
χ A , ψ = ∗ ( F A ∧ ψ ) , P A , ψ a = ∗ ( d A a ∧ ψ ) , (6) \chi_{A,\psi}=*(F_A\wedge\psi),\qquad\mathcal{P}_{A,\psi}a=*(d_Aa\wedge\psi), \tag*{(6)} χ A , ψ = ∗ ( F A ∧ ψ ) , P A , ψ a = ∗ ( d A a ∧ ψ ) , (6)
D A , ψ ( f , a ) = ( d A ∗ a , d A f + P A , ψ a ) . (7) \mathcal{D}_{A,\psi}(f,a)=\left(d_A^*a,d_Af+\mathcal{P}_{A,\psi}a\right). \tag*{(7)} D A , ψ ( f , a ) = ( d A ∗ a , d A f + P A , ψ a ) . (7)
Definition 1.4. A connection A A A on P → M P\to M P → M is called a G 2 G_2 G 2 -instanton if its curvature F A F_A F A satisfies
F A ∧ ψ = 0. F_A\wedge\psi=0. F A ∧ ψ = 0.
We call χ A , ψ \chi_{A,\psi} χ A , ψ the instanton defect. With ψ \psi ψ fixed, its derivative in the direction a a a is P A , ψ a \mathcal{P}_{A,\psi}a P A , ψ a . At an instanton,
P A , ψ ( d A q ) = ∗ ( [ F A , q ] ∧ ψ ) = 0 for q ∈ Ω 0 ( M , E ) . \mathcal{P}_{A,\psi}(d_Aq)=*( [F_A,q]\wedge\psi)=0 \qquad\text{for }q\in\Omega^0(M,E). P A , ψ ( d A q ) = ∗ ([ F A , q ] ∧ ψ ) = 0 for q ∈ Ω 0 ( M , E ) .
We impose the Coulomb condition d A ∗ a = 0 d_A^*a=0 d A ∗ a = 0 to fix this gauge freedom and add the zero-form f f f to obtain the elliptic operator D A , ψ \mathcal{D}_{A,\psi} D A , ψ . An instanton is acyclic if D A , ψ \mathcal{D}_{A,\psi} D A , ψ is invertible [38 Definition 3.6] . We identify the kernel of the operator D \mathcal{D} D at an instanton with the direct sum of the parallel sections of E E E and the infinitesimal instanton deformations in Coulomb gauge in Lemma 4.2 .
We use the inverse operator D − 1 \mathcal{D}^{-1} D − 1 in the correction below but for instance this does not apply on the full adjoint bundle of a U ( r ) U(r) U ( r ) -bundle as this inverse never exists simply because − 1 Id \sqrt{-1}\operatorname{Id} − 1 Id is parallel for every connection A A A . In particular,
D A , ψ ( − 1 Id , 0 ) = 0. \mathcal{D}_{A,\psi}(\sqrt{-1}\operatorname{Id},0)=0. D A , ψ ( − 1 Id , 0 ) = 0.
Thus, a fixed-determinant version would assume acyclicity on the traceless adjoint bundle. The fixed determinant connection must also satisfy the instanton equation for each ψ t \psi_t ψ t .
For a closed four-form Q Q Q and a pair ( A , ψ ) (A,\psi) ( A , ψ ) for which D A , ψ \mathcal{D}_{A,\psi} D A , ψ is invertible, set
( f , U A , ψ ( Q ) ) = − D A , ψ − 1 ( 0 , ∗ ( F A ∧ Q ) ) . (8) (f,\mathcal{U}_{A,\psi}(Q))=-\mathcal{D}_{A,\psi}^{-1}(0,*(F_A\wedge Q)). \tag*{(8)} ( f , U A , ψ ( Q )) = − D A , ψ − 1 ( 0 , ∗ ( F A ∧ Q )) . (8)
The operator stays invertible near an acyclic instanton, so this formula also applies to nearby connections with nonzero defect. The Bianchi identity and d ψ = 0 d\psi=0 d ψ = 0 give d A ( F A ∧ ψ ) = 0 d_A(F_A\wedge\psi)=0 d A ( F A ∧ ψ ) = 0 , and hence d A ∗ χ A , ψ = 0 d_A^*\chi_{A,\psi}=0 d A ∗ χ A , ψ = 0 . On a compact manifold without boundary, integration by parts gives
⟨ χ A , ψ , d A f ⟩ L 2 = ⟨ d A ∗ χ A , ψ , f ⟩ L 2 = 0. \langle\chi_{A,\psi},d_Af\rangle_{L^2}=\langle d_A^*\chi_{A,\psi},f\rangle_{L^2}=0. ⟨ χ A , ψ , d A f ⟩ L 2 = ⟨ d A ∗ χ A , ψ , f ⟩ L 2 = 0.
Theorem 1.5. Let M 7 M^7 M 7 be compact without boundary and let ψ t \psi_t ψ t be a prescribed smooth family of closed positive four-forms with φ t = ∗ t ψ t \varphi_t=*_{t}\psi_t φ t = ∗ t ψ t and Q t = ∂ t ψ t Q_t=\partial_t\psi_t Q t = ∂ t ψ t . If A 0 A_0 A 0 is an acyclic instanton for ψ 0 \psi_0 ψ 0 , then
∂ t A = − d A ∗ F A + ∗ t ( F A ∧ d φ t ) + U A , ψ t ( Q t ) (9) \partial_tA=-d_A^*F_A+*_{t}(F_A\wedge d\varphi_t)+\mathcal{U}_{A,\psi_t}(Q_t) \tag*{(9)} ∂ t A = − d A ∗ F A + ∗ t ( F A ∧ d φ t ) + U A , ψ t ( Q t ) (9) has a unique smooth solution for short time with initial value A 0 A_0 A 0 . Every smooth solution from A 0 A_0 A 0 satisfies F A t ∧ ψ t = 0 F_{A_t}\wedge\psi_t=0 F A t ∧ ψ t = 0 as long as D A t , ψ t \mathcal{D}_{A_t,\psi_t} D A t , ψ t remains invertible. Along any such solution, the section f f f in (8) vanishes and
d A ∗ ∂ t A = 0 , P A , ψ ∂ t A = − ∗ ( F A ∧ Q ) . (10) d_A^*\partial_tA=0,\qquad\mathcal{P}_{A,\psi}\partial_tA=-*(F_A\wedge Q). \tag*{(10)} d A ∗ ∂ t A = 0 , P A , ψ ∂ t A = − ∗ ( F A ∧ Q ) . (10) On a family of nearby instantons which varies with time, the Yang–Mills and torsion terms in (9) cancel. The remaining equation ∂ t A = U A , ψ t ( Q t ) \partial_tA=\mathcal{U}_{A,\psi_t}(Q_t) ∂ t A = U A , ψ t ( Q t ) gives the Coulomb lift of the branch. § 4.7 gives examples in which the gauge class of instantons is moving, i.e, the velocity in (9) is nonzero.
If the coclosed G 2 G_2 G 2 structure is not moving, we have the following
Theorem 1.6. Let A ⋆ A_\star A ⋆ be an acyclic instanton for a closed positive four form ψ ⋆ \psi_\star ψ ⋆ , and put φ ⋆ = ∗ ψ ⋆ ψ ⋆ \varphi_\star=*_{\psi_\star}\psi_\star φ ⋆ = ∗ ψ ⋆ ψ ⋆ . Fix any integer s ≥ 8 s\geq8 s ≥ 8 , every sufficiently close smooth initial connection in the Sobolev space H s H^s H s has a unique global solution of
∂ t A = − d A ∗ F A + ∗ ( F A ∧ d φ ⋆ ) . \partial_tA=-d_A^*F_A+*(F_A\wedge d\varphi_\star). ∂ t A = − d A ∗ F A + ∗ ( F A ∧ d φ ⋆ ) . It converges smoothly and exponentially to a gauge transform of A ⋆ A_\star A ⋆ . The initial connection need not be an instanton.
On a fixed torsion-free background, the equation in the preceding theorem is the ordinary Yang–Mills flow. Global existence and convergence for this flow near a local minimum of the Yang–Mills energy were proved by Feehan [10 Thm. 6] on closed manifolds of arbitrary dimension. On a fixed coclosed background, the equation is the negative gradient flow of the defect energy described in § 4 . Its relation to the Ω \Omega Ω –Yang–Mills functional, with Ω = − φ ⋆ \Omega=-\varphi_\star Ω = − φ ⋆ , is explained there; see also [7 ] . Acyclicity makes the Hessian positive in Coulomb gauge, so the convergence mechanism is the usual stability mechanism near a nondegenerate minimum modulo gauge. We include this case as a consequence of Theorem 4.16 , which also treats the additional terms produced by a moving ambient form and a moving reference instanton.
We now related our result to a well-studied flow of coclosed G 2 G_2 G 2 -structures called the modified Laplacian coflow of coclosed G 2 G_2 G 2 -structures introduced by Grigorian [12 ] . We recall that the full torsion tensor of φ \varphi φ is defined by ∇ i φ a b c = T i q ψ q a b c \nabla_i\varphi_{abc}=T_i{}^q\psi_{qabc} ∇ i φ ab c = T i q ψ q ab c and tr T = g i j T i j \operatorname{tr}T=g^{ij}T_{ij} tr T = g ij T ij . For a constant c ∈ R c\in\mathbb{R} c ∈ R , which we call the coflow parameter, we take the ambient velocity in Theorem 1.5 to be
Q c ( ψ ) = Δ ψ ψ − 2 d ( ( c + tr T ψ ) φ ) . (11) Q_c(\psi)=\Delta_\psi\psi-2d\bigl((c+\operatorname{tr}T_\psi)\varphi\bigr). \tag*{(11)} Q c ( ψ ) = Δ ψ ψ − 2 d ( ( c + tr T ψ ) φ ) . (11)
This is precisely Grigorian’s modified coflow in our convention; the conversion and the parameter-zero case are discussed in § 4.5 . We have the following
Corollary 1.7. Under the initial hypotheses of Theorem 1.5 , fix c ∈ R c\in\mathbb{R} c ∈ R and couple (9) to ∂ t ψ = Q c ( ψ ) \partial_t\psi=Q_c(\psi) ∂ t ψ = Q c ( ψ ) , with Q = Q c ( ψ ) Q=Q_c(\psi) Q = Q c ( ψ ) and d ψ = 0 d\psi=0 d ψ = 0 . This system has a unique smooth solution from ( A 0 , ψ 0 ) (A_0,\psi_0) ( A 0 , ψ 0 ) for a short time. Every smooth solution from these initial data preserves F A ∧ ψ = 0 F_A\wedge\psi=0 F A ∧ ψ = 0 on each interval where D A t , ψ t D_{A_t,\psi_t} D A t , ψ t is invertible.
Note that the parameter c c c in Grigorian’s modified coflow is a constant. Since torsion-free G 2 G_2 G 2 -structures are stationary points of the flow, one can ask whether they are dynamically stable along the flow. Such a result was proved by Bedulli and Vezzoni in [4 ] for the choice c = 0 c=0 c = 0 which in a way suggests that c = 0 c=0 c = 0 might be a suitable choice in the flow. The global theorem below uses c = 0 c=0 c = 0 . The local result above does not by itself imply global stability for positive c c c . In fact, Proposition 4.19 gives backgrounds with positive linear growth rates when c > 0 c>0 c > 0 which shows that c > 0 c>0 c > 0 can introduce unstable modes. It does not assert instability for every positive parameter on every background.
Theorem 1.8. Let ( M , ψ ˉ ) (M,\bar\psi) ( M , ψ ˉ ) be compact and torsion-free, and let A ⋆ A_\star A ⋆ be an acyclic ψ ˉ \bar\psi ψ ˉ -instanton. Fix 0 < α < 1 0<\alpha<1 0 < α < 1 . There are a finite integer N N N and a constant δ > 0 \delta>0 δ > 0 such that the following holds. Suppose that ψ 0 \psi_0 ψ 0 and A 0 A_0 A 0 are smooth, that ψ 0 \psi_0 ψ 0 is closed and positive, that [ ψ 0 ] = [ ψ ˉ ] [\psi_0]=[\bar\psi] [ ψ 0 ] = [ ψ ˉ ] , and that
∥ ψ 0 − ψ ˉ ∥ C N , α + ∥ A 0 − A ⋆ ∥ H 8 < δ . \lVert\psi_0-\bar\psi\rVert_{C^{N,\alpha}}+\lVert A_0-A_\star\rVert_{H^8}<\delta. ∥ ψ 0 − ψ ˉ ∥ C N , α + ∥ A 0 − A ⋆ ∥ H 8 < δ . These norms use the fixed metric of ψ ˉ \bar\psi ψ ˉ and the reference connection A ⋆ A_\star A ⋆ , and the connection difference is measured in a fixed initial gauge. Then the system
∂ t ψ = Q 0 ( ψ ) , \partial_t\psi=Q_0(\psi), ∂ t ψ = Q 0 ( ψ ) , ∂ t A = − d A ∗ F A + ∗ t ( F A ∧ d ψ t ) + U A , ψ t ( ∂ t ψ t ) \partial_tA=-d_A^\ast F_A+\ast_t(F_A\wedge d\psi_t)+\mathcal{U}_{A,\psi_t}(\partial_t\psi_t) ∂ t A = − d A ∗ F A + ∗ t ( F A ∧ d ψ t ) + U A , ψ t ( ∂ t ψ t ) has a unique global smooth solution. The operator D A t , ψ t D_{A_t,\psi_t} D A t , ψ t remains invertible. The solution converges smoothly and exponentially to a torsion-free structure together with an acyclic instanton. The initial connection may have nonzero instanton defect. If it is initially an instanton, it remains one.
In this theorem, one may take N = r + 8 N=r+8 N = r + 8 , where r ≥ max { r 0 , 10 } r\geq\max\{r_0,10\} r ≥ max { r 0 , 10 } and r 0 r_0 r 0 is an ambient differentiability order sufficient for the transport estimates in the proof of Theorem 4.10 to control the transported connection in C 9 C^9 C 9 . These orders are fixed before δ \delta δ is chosen.
Strategy of proofs
We now explain the strategy of proving the above mentioned theorems.
Preservation and existence. For coassociatives, the error takes values in exact three forms. Hodge theory gives a right inverse of D i D_i D i on this space. Its kernel consists of the harmonic normal deformations; projecting away from that kernel selects the solution of smallest norm. If R 0 R_0 R 0 is the right inverse at the initial coassociative then because D i R 0 D_iR_0 D i R 0 is close to the identity operator on the space of exact 3 3 3 -forms, it is invertible and lets us define a right inverse for D i D_i D i and hence W W W at nearby pairs ( φ , i ) (\varphi,i) ( φ , i ) even when i ⋆ φ ≠ 0 i\star\varphi\neq0 i ⋆ φ = 0 . We therefore define the correction before assuming that the evolving pair is coassociative. The coassociative equation gives ∂ t ( i ∗ φ ) = d C θ + D i H + D i W = 0 \partial_t(i^\ast\varphi)=d_C\theta+D_iH+D_iW=0 ∂ t ( i ∗ φ ) = d C θ + D i H + D i W = 0 for every solution in the chosen neighborhood.
We construct a solution by first solving the unmodified Laplacian flow. The varying coassociative deformation theorem [27 , 16 ] gives a local family of coassociatives for the evolving G 2 G_2 G 2 structures. The parameters in the kernel satisfy an ODEs on a finite-dimensional space. Its solution selects the required normal velocity. A diffeomorphism of M M M then moves the correction into the ambient Laplacian flow equation, leaving the submanifold velocity equal to mean curvature. Reversing these steps proves uniqueness of the solutions.
For instantons, we differentiate the equation for the defect χ \chi χ and use d A ∗ χ = 0 d_A^\ast\chi=0 d A ∗ χ = 0 . When we differentiate χ = ∗ ψ ( F A ∧ ψ ) \chi=\ast_\psi(F_A\wedge\psi) χ = ∗ ψ ( F A ∧ ψ ) , the correction cancels ∗ ψ ( F A ∧ Q ) \ast_\psi(F_A\wedge Q) ∗ ψ ( F A ∧ Q ) and leaves a term − d A F -d_AF − d A F in the defect equation. The Bianchi identity and d ψ = 0 \mathrm{d}\psi= 0 d ψ = 0 give d A ∗ χ = 0 \mathrm{d}^{*}_{A}\chi= 0 d A ∗ χ = 0 , so the L 2 L^{2} L 2 pairing of d A f \mathrm{d}_{A}f d A f with χ \chi χ vanishes by integration by parts. The energy estimate and Gronwall’s inequality then give χ ( t ) = 0 \chi(t) = 0 χ ( t ) = 0 if χ ( 0 ) = 0 \chi(0) = 0 χ ( 0 ) = 0 .
More precisely, to construct nearby instantons in Lemma 4.7, we solve Walpuski’s augmented equation [Wal17, eq. (2.4)] in a fixed Coulomb slice:
*_{\psi}(F_{A} \wedge\psi) + \mathrm{d}_{A}\xi= 0,\qquad\mathrm{d}^{*_{g_{\psi_{0}}}_{A_{0}}}(A-A_{0}) = 0.
Here ξ \xi ξ is an adjoint-valued zero-form. The linearization in ( ξ , A ) (\xi,A) ( ξ , A ) at ( 0 , A 0 , ψ 0 ) (0,A_{0},\psi_{0}) ( 0 , A 0 , ψ 0 ) is D A 0 , ψ 0 D_{A_{0},\psi_{0}} D A 0 , ψ 0 , which is invertible by the acyclicity assumption. The implicit function theorem gives solutions for nearby closed positive forms ψ \psi ψ , by the argument of [Wal17, Proposition 2.11]. Applying d A ∗ \mathrm{d}^{*}_{A} d A ∗ to the first equation gives d A ∗ d A ξ = 0 \mathrm{d}^{*}_{A}\mathrm{d}_{A}\xi= 0 d A ∗ d A ξ = 0 . Integration by parts yields d A ξ = 0 \mathrm{d}_{A}\xi= 0 d A ξ = 0 , so A A A satisfies the instanton equation.
For initial connections with nonzero defect, we must solve the full connection equation. Following [6 Section 2.1] , we first add a gauge term so that its highest-order part becomes a heat operator. We also check that the nonlocal correction U \mathcal{U} U preserves the Sobolev regularity of the connection. Indeed, forming the curvature uses one derivative, while applying D − 1 D^{-1} D − 1 restores that derivative. In particular, U \mathcal{U} U maps the H 6 H^{6} H 6 neighborhood used in the existence proof into H 6 H^{6} H 6 , with local Lipschitz estimates. These estimates allow us to solve the modified equation for a short time. We then apply a time-dependent gauge transformation that cancels the added gauge term and gives a solution of the original equation.
Long-time behavior. For understanding the long time behaviour of the coupled system for coassociatives, we control the graph charts and right inverses uniformly near a fixed pair. We show that the evolving pair ( φ t , i t ) (\varphi_t,i_t) ( φ t , i t ) stays in the neighborhood where the local construction applies. We describe nearby coassociatives as normal graphs over the reference submanifold and obtain uniform estimates for these parametrizations and the right inverses R i R_i R i . Near a torsion-free structure, the stability theorem of [LW19b] keeps the ambient form close to the reference torsion-free G 2 G_{2} G 2 structure, and we extract the required integral bounds from its proof.
For connections, we compare the solution with a nearby instanton which moves along the evolving ambient 4 4 4 -forms. After choosing gauge, we estimate the difference between the two connections. Acyclicity gives a positive spectral gap for the second-order operator in the linearized equation, so the linearized flow decreases this difference. We show that sufficiently small perturbations and the explicit bounds on the moving ψ t \psi_t ψ t allow us to control the additional terms. An estimate in one fixed Sobolev norm then gives exponential decay and keeps the solution in the neighborhood where the correction is defined. Parabolic smoothing gives exponential decay of every higher spatial derivative after positive time. The moving-background assumptions are stated in Theorem 4.16 . If the evolving instanton also converges, the original connection converges to its limit up to a gauge transformation.
For the ordinary Laplacian flow, Lemma 3.17 uses the stability proof of Lotay and Wei [LW19b] to bound the distance between the evolving G 2 G_{2} G 2 structures and the reference torsion-free one as well as the the time integrals of its velocity and torsion. These bounds keep the coassociative in its deformation chart for all time.
For the modified coflow with c = 0 c = 0 c = 0 , Bedulli and Vezzoni [4 Theorem 3.4] prove stability near a torsion-free structure under smallness in every C k C^{k} C k norm. In Lemma 4.18 , we prove the finite-order estimates needed here by a direct energy argument. We show that for each fixed spatial order r r r , it is enough to assume smallness in C r + 8 , α C^{r+8,\alpha} C r + 8 , α .
The same initial norm controls the distance from the reference throughout the flow and gives an exponentially decaying bound for the ambient velocity in C r , α C^{r,\alpha} C r , α . In particular, it bounds
∫ 0 ∞ ∥ ∂ t ψ t ∥ C r , α d t . \int_{0}^{\infty}\lVert\partial_t\psi_t\rVert_{C^{r,\alpha}}\,\mathrm{d}t. ∫ 0 ∞ ∥ ∂ t ψ t ∥ C r , α d t .
These quantitative estimates keep the transported instanton close to the reference connection and verify the moving-background hypotheses of Theorem 4.16 . This gives the global coupled result in Theorem 1.8 .
Organization of the paper. We give basics of G 2 G_{2} G 2 geometry and fix our notations and conventions in § 2 . § 3 treats the coassociative system. We prove our short-time and long-time existence results in that section which includes the proofs of Theorems 1.1 and 1.3 . We also give an explicit example on the flat torus in Proposition 3.8 which shows that a straightforward coupling of the Laplacian flow and the mean curvature flow does not preserve the coassociative condition. We discuss the coupled flow system of connections and the 4 4 4 -form in § 4 . This sections contains the local theory and nonlinear stability of the coupled system, and the application to modified coflow. It also gives examples with nonzero correction and the compatibility condition at a nonacyclic instanton. We finish the paper with some questions which we find interesting but which we do not treat in this paper in § 5 .
Acknowledgments. SP would like to thank Vivek Shende for his immense support and encouragement throughout the project. SP was supported by VILLUMFONDEN, VILLUM Investigator grant 37814 and Danish National Research Foundation grant DNRF157. SD would like to thank Jason Lotay for several discussions on the topics of this paper and for sharing his ideas on coupled flows. He would also like to thank Thomas Walpuski for discussions on gauge theory on G 2 G_2 G 2 manifolds and the Yang-Mills flow. SD acknowledges support by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy - EXC 2121 "Quantum Universe" - 390833306. This project started when SP was a visiting PhD student of SD at the University of Hamburg in 2024-25. SP would like to thank the Department of Mathematics at the University of Hamburg for support and for providing excellent working conditions.
AI usage. AI tools were used for language editing, polishing and for literature review. They were also used as an assistant for proof-checking. The authors take full responsibility for correctness and final contents of the paper.
Basics on G 2 G_2 G 2 structures
We first fix the algebraic signs used throughout the paper. The metric, the Hodge star, and the decomposition of forms all depend on the G 2 G_2 G 2 structure. In particular, differentiating a Hodge star means differentiating the metric as well. We use the model
φ f l a t = d y 123 + d y 1 ∧ ( d x 12 + d x 34 ) + d y 2 ∧ ( d x 13 − d x 24 ) − d y 3 ∧ ( d x 14 + d x 23 ) . (12) \varphi_{\mathrm{flat}}=dy^{123}+dy^1\wedge\left(dx^{12}+dx^{34}\right)+dy^2\wedge\left(dx^{13}-dx^{24}\right)-dy^3\wedge\left(dx^{14}+dx^{23}\right). \tag*{(12)} φ flat = d y 123 + d y 1 ∧ ( d x 12 + d x 34 ) + d y 2 ∧ ( d x 13 − d x 24 ) − d y 3 ∧ ( d x 14 + d x 23 ) . (12)
The ambient orientation is − d y 123 ∧ d x 1234 -dy^{123}\wedge dx^{1234} − d y 123 ∧ d x 1234 , and the coassociative x x x plane is oriented by − d x 1234 -dx^{1234} − d x 1234 . We use the sign conventions in [15 ] . We use the irreducible summands Λ 7 2 = { V ⌟ φ } \Lambda^2_7=\{V\mathbin{\lrcorner}\varphi\} Λ 7 2 = { V ┘ φ } and Λ 14 2 = { β : β ∧ ψ = 0 } \Lambda^2_{14}=\{\beta:\beta\wedge\psi=0\} Λ 14 2 = { β : β ∧ ψ = 0 } of the two form bundle, and write Ω 14 2 ( M ) \Omega^2_{14}(M) Ω 14 2 ( M ) for smooth sections of the latter. The corresponding decomposition of three forms is Λ 3 = Λ 1 3 ⊕ Λ 7 3 ⊕ Λ 27 3 \Lambda^3=\Lambda^3_1\oplus\Lambda^3_7\oplus\Lambda^3_{27} Λ 3 = Λ 1 3 ⊕ Λ 7 3 ⊕ Λ 27 3 , where Λ 1 3 = R φ \Lambda^3_1=\mathbb{R}\varphi Λ 1 3 = R φ , Λ 7 3 = { V ⌟ ψ } \Lambda^3_7=\{V\mathbin{\lrcorner}\psi\} Λ 7 3 = { V ┘ ψ } and Λ 27 3 \Lambda^3_{27} Λ 27 3 is their orthogonal complement. Put Λ j 4 = ∗ Λ j 3 \Lambda^4_j=* \Lambda^3_j Λ j 4 = ∗ Λ j 3 and let π j \pi_j π j denote the corresponding projection on four forms when it is applied to a four form. We will use the algebraic variation formula [15 ]
φ ˙ = ∗ ( 3 4 π 1 ψ ˙ + π 7 ψ ˙ − π 27 ψ ˙ ) . (13) \dot{\varphi}=*\left(\frac{3}{4}\pi_1\dot{\psi}+\pi_7\dot{\psi}-\pi_{27}\dot{\psi}\right). \tag*{(13)} φ ˙ = ∗ ( 4 3 π 1 ψ ˙ + π 7 ψ ˙ − π 27 ψ ˙ ) . (13)
In particular, a variation in Λ 27 4 \Lambda^4_{27} Λ 27 4 gives φ ˙ = − ∗ ψ ˙ \dot{\varphi}=-*\dot{\psi} φ ˙ = − ∗ ψ ˙ . We will use this sign when linearizing the coflow. The map from ψ \psi ψ to its dual three form is nonlinear and (13) is its derivative, including the change of the Hodge star.
For tangent vectors u u u , v v v and a two form β \beta β , our conventions give
( u ⌟ φ ) ∧ ( v ⌟ φ ) ∧ φ = − 6 g ( u , v ) vol g , ∗ ( φ ∧ β ) = { − 2 β , β ∈ Λ 7 2 , β , β ∈ Λ 14 2 . (14) (u\mathbin{\lrcorner}\varphi)\wedge(v\mathbin{\lrcorner}\varphi)\wedge\varphi=-6g(u,v)\operatorname{vol}_g,\qquad*(\varphi\wedge\beta)=\begin{cases}-2\beta,&\beta\in\Lambda^2_7,\\ \beta,&\beta\in\Lambda^2_{14}.\end{cases} \tag*{(14)} ( u ┘ φ ) ∧ ( v ┘ φ ) ∧ φ = − 6 g ( u , v ) vol g , ∗ ( φ ∧ β ) = { − 2 β , β , β ∈ Λ 7 2 , β ∈ Λ 14 2 . (14)
The torsion of a G 2 G_2 G 2 -structure is a 2-tensor T T T which is essentially ∇ φ \nabla\varphi ∇ φ . The explicit expression for T T T in a local orthonormal frame is
T i j = 1 24 ∇ i φ a b c ψ j a b c . T_{ij}=\frac{1}{24}\nabla_i\varphi_{abc}\psi_j^{abc}. T ij = 24 1 ∇ i φ ab c ψ j ab c .
For a closed G 2 G_2 G 2 structure, the torsion form τ \tau τ satisfies
d ψ = − τ ∧ φ , τ ∈ Ω 14 2 ( M ) . (15) d\psi=-\tau\wedge\varphi,\qquad\tau\in\Omega^2_{14}(M). \tag*{(15)} d ψ = − τ ∧ φ , τ ∈ Ω 14 2 ( M ) . (15)
In the notation of [15 ] , τ = − τ 2 \tau=-\tau_2 τ = − τ 2 for a closed structure. The identity d ψ = − τ ∧ φ d\psi=-\tau\wedge\varphi d ψ = − τ ∧ φ will turn the first variation of coassociative volume into a formula involving the restriction of τ \tau τ .
The cross product is defined by g ( U × V , W ) = φ ( U , V , W ) g(U\times V,W)=\varphi(U,V,W) g ( U × V , W ) = φ ( U , V , W ) , and our convention for pointwise norms is ∣ ω ∣ 2 = 1 k ! ω i 1 ⋯ i k ω i 1 ⋯ i k |\omega|^2=\frac{1}{k!}\omega_{i_1\cdots i_k}\omega^{i_1\cdots i_k} ∣ ω ∣ 2 = k ! 1 ω i 1 ⋯ i k ω i 1 ⋯ i k .
Coassociatives and instantons. At a coassociative, the map J i J_i J i introduced in § 1.1 takes values in Λ 2 T ∗ C \Lambda^2T^*C Λ 2 T ∗ C with norm ∣ J i V ∣ 2 = 2 ∣ V ∣ 2 |J_iV|^2=2|V|^2 ∣ J i V ∣ 2 = 2∣ V ∣ 2 , where the anti-self-dual notation follows from our chosen orientation. For nearby embeddings, we retain the orientation of C C C induced by the initial coassociative embedding, and Lemma 3.1 computes J i ( H ) J_i(H) J i ( H ) when d φ = 0 d\varphi=0 d φ = 0 .
For the bundle E E E in § 1.2 , the instanton condition is equivalently F A ∈ Ω 14 2 ( M , E ) F_A\in\Omega^2_{14}(M,E) F A ∈ Ω 14 2 ( M , E ) . Under (14) this implies ∗ ( φ ∧ F A ) = F A *(\varphi\wedge F_A)=F_A ∗ ( φ ∧ F A ) = F A . Thus an instanton can have nonzero curvature and the instanton equation only requires the Λ 7 2 \Lambda^2_7 Λ 7 2 part of the curvature to vanish. This will matter when interpreting the defect energy: it controls that component, not the full curvature.
Differentiating a pullback. A dot or ∂ t \partial_t ∂ t denotes time differentiation on the fixed domain C C C . For smooth φ t \varphi_t φ t and embeddings i t i_t i t with velocity V t = ∂ t i t V_t = \partial_t i_t V t = ∂ t i t , let γ t = i t ∗ φ t \gamma_t = i_t^*\varphi_t γ t = i t ∗ φ t . Then
γ ˙ = i t ∗ φ ˙ + d C i t ∗ ( V ⌟ φ ) + i t ∗ ( V ⌟ d φ ) . (16) \dot{\gamma}=i_t^*\dot{\varphi}+d_C i_t^*(V\mathbin{\lrcorner}\varphi)+i_t^*(V\mathbin{\lrcorner}d\varphi). \tag*{(16)} γ ˙ = i t ∗ φ ˙ + d C i t ∗ ( V ┘ φ ) + i t ∗ ( V ┘ d φ ) . (16)
There are two contributions in this expression. The 3-form changes with time, and the embedding moves as well. To compute the second term, extend V V V locally to an ambient vector field and use Cartan’s magical formula L V = d ( V ⌟ ⋅ ) + V ⌟ d \mathcal{L}_V=d(V\mathbin{\lrcorner}\mathord{\cdot})+V\mathbin{\lrcorner}d L V = d ( V ┘ ⋅ ) + V ┘ d . After pullback, the result depends only on V V V along the embedding, so it is independent of the chosen extension. If d φ = 0 d\varphi=0 d φ = 0 , the last term vanishes. The two remaining terms are the ones we must cancel in order to preserve the coassociative construction.
For local coordinate one forms we use d x i 1 ⋯ i k = d x i 1 ∧ ⋯ ∧ d x i k dx^{i_1\cdots i_k}=dx^{i_1}\wedge\cdots\wedge dx^{i_k} d x i 1 ⋯ i k = d x i 1 ∧ ⋯ ∧ d x i k , and similarly e i 1 ⋯ i k e^{i_1\cdots i_k} e i 1 ⋯ i k for a coframe.
The first coupled system: coassociatives
Correction and preservation
We first calculate how the coassociative equation changes when the ambient G 2 G_2 G 2 structure evolves by the Laplacian flow and the submanifold moves by the mean curvature flow. This calculation will identify the term that we need to cancel. We then construct a right inverse that produces the correction. The existence of a solution to the corrected equation is a separate question, which we address in § 3.3 .
We begin with the relation between mean curvature and torsion. This relation is part of the general theory developed by Ball and Madnick [2 ] . For closed G 2 G_2 G 2 structures, Pacini and Raffero [30 Lemma 2.3] give the first variation of coassociative volume in terms of the ambient torsion. We include the calculation in our conventions because the signs and normalization enter the correction below.
Lemma 3.1. For a compact coassociative i : C 4 ↪ ( M 7 , φ ) i:C^4\hookrightarrow(M^7,\varphi) i : C 4 ↪ ( M 7 , φ ) with d φ = 0 d\varphi=0 d φ = 0 , let θ = i ∗ τ \theta=i^*\tau θ = i ∗ τ where τ ∈ Ω 14 2 \tau\in\Omega^2_{14} τ ∈ Ω 14 2 is the torsion form. The map J i : ν C → Λ − 2 T ∗ C J_i:\nu C\to\Lambda^2_-T^*C J i : ν C → Λ − 2 T ∗ C , J i ( V ) = i ∗ ( V ⌟ φ ) J_i(V)=i^*(V\mathbin{\lrcorner}\varphi) J i ( V ) = i ∗ ( V ┘ φ ) , satisfies
⟨ J i ( V ) , J i ( W ) ⟩ = 2 g ( V , W ) , J i ( H ) = − 2 π − θ . (17) \langle J_i(V),J_i(W)\rangle=2g(V,W),\qquad J_i(H)=-2\pi_-\theta. \tag*{(17)} ⟨ J i ( V ) , J i ( W )⟩ = 2 g ( V , W ) , J i ( H ) = − 2 π − θ . (17) Proof. We first check the pointwise statement. Choose an adapted orthonormal coframe and write φ \varphi φ as in (12) , with the orientations fixed in § 2 . Contraction with the three unit normal vectors gives three mutually orthogonal anti-self-dual two forms, each of squared norm two. By linearity, their pairings give the first identity in (17) . In particular, J i J_i J i is an isomorphism from the normal bundle to the bundle of anti-self-dual two forms and the factor 2 records its normalization.
We next vary the embedding while keeping the ambient forms and metric fixed. Let V V V be any smooth normal field and choose a variation i s i_s i s with initial velocity V V V . The comass inequality says pointwise that i s ∗ ψ ≤ vol i s ∗ g i_s^*\psi\leq\operatorname{vol}_{i_s^*g} i s ∗ ψ ≤ vol i s ∗ g . At the coassociative embedding i 0 = i i_0=i i 0 = i , the two sides agree. Their difference is therefore a nonnegative function of s s s that vanishes at zero, so its derivative at zero is zero. We can thus compute the variation of the volume by differentiating ∫ C i s ∗ ψ \int_C i_s^*\psi ∫ C i s ∗ ψ . The first variation of volume gives the first equality below. Cartan’s formula and Stokes’ theorem give the second with the exact term integrating to zero because C C C is compact and has no boundary:
− ∫ C g ( H , V ) d μ = d d s ∣ 0 ∫ C i s ∗ ψ = ∫ C i ∗ ( V ⌟ d ψ ) . -\int_C g(H,V)\,d\mu=\left.\frac{d}{ds}\right|_0\int_C i_s^*\psi=\int_C i^*(V\mathbin{\lrcorner}d\psi). − ∫ C g ( H , V ) d μ = d s d 0 ∫ C i s ∗ ψ = ∫ C i ∗ ( V ┘ d ψ ) . We now use (15) . Since i ∗ φ = 0 i^*\varphi=0 i ∗ φ = 0 , contraction of d ψ = − τ ∧ φ d\psi=-\tau\wedge\varphi d ψ = − τ ∧ φ gives i ∗ ( V ⌟ d ψ ) = − θ ∧ J i ( V ) i^*(V\mathbin{\lrcorner}d\psi)=-\theta\wedge J_i(V) i ∗ ( V ┘ d ψ ) = − θ ∧ J i ( V ) . For an anti-self-dual two form α \alpha α , we have θ ∧ α = − ⟨ π − θ , α ⟩ d μ \theta\wedge\alpha=-\langle\pi_-\theta,\alpha\rangle d\mu θ ∧ α = − ⟨ π − θ , α ⟩ d μ . Consequently the integrand on the right is ⟨ π − θ , J i ( V ) ⟩ d μ \langle\pi_-\theta,J_i(V)\rangle d\mu ⟨ π − θ , J i ( V )⟩ d μ . We may choose V V V supported in an arbitrarily small neighborhood, so the integral identity gives − g ( H , V ) = ⟨ π − θ , J i ( V ) ⟩ -g(H,V)=\langle\pi_-\theta,J_i(V)\rangle − g ( H , V ) = ⟨ π − θ , J i ( V )⟩ pointwise. As J i ( V ) J_i(V) J i ( V ) ranges over all anti-self-dual two forms, we obtain J i ( H ) = − 2 π − θ J_i(H)=-2\pi_-\theta J i ( H ) = − 2 π − θ . This completes the proof of the lemma.
Proposition 3.2. For a smooth solution of the unmodified coupled system (5) , at every time
∂ t ( i t ∗ φ t ) = d C ( i t ∗ τ t + i t ∗ ( H t ⌟ φ t ) ) . (18) \partial_t(i_t^*\varphi_t)=d_C\bigl(i_t^*\tau_t+i_t^*(H_t\mathbin{\lrcorner}\varphi_t)\bigr). \tag*{(18)} ∂ t ( i t ∗ φ t ) = d C ( i t ∗ τ t + i t ∗ ( H t ┘ φ t ) ) . (18)
At a time when i t i_t i t is coassociative, this becomes
∂ t ( i t ∗ φ t ) = d C ( ∗ C i t ∗ τ t ) . (19) \partial_t(i_t^*\varphi_t)=d_C(*_C i_t^*\tau_t). \tag*{(19)} ∂ t ( i t ∗ φ t ) = d C ( ∗ C i t ∗ τ t ) . (19)
At such a time, the coassociative preservation condition ∂ t ( i t ∗ φ t ) = 0 \partial_t(i_t^*\varphi_t)=0 ∂ t ( i t ∗ φ t ) = 0 is equivalent to δ C ( i t ∗ τ t ) = 0 \delta_C(i_t^*\tau_t)=0 δ C ( i t ∗ τ t ) = 0 .
Proof. We differentiate both the ambient form and the embedding. Since the G 2 G_2 G 2 -structures remain closed along the Laplacian flow, we get that the final contraction term in (16) vanishes, and the Laplacian flow equation ∂ t φ = d τ \partial_t\varphi=d\tau ∂ t φ = d τ gives
∂ t ( i t ∗ φ t ) = d C ( i t ∗ τ t ) + d C ( i t ∗ ( H t ⌟ φ t ) ) . \partial_t(i_t^*\varphi_t)=d_C(i_t^*\tau_t)+d_C(i_t^*(H_t\mathbin{\lrcorner}\varphi_t)). ∂ t ( i t ∗ φ t ) = d C ( i t ∗ τ t ) + d C ( i t ∗ ( H t ┘ φ t )) .
If i t i_t i t is coassociative at the time under consideration, we can use (17) . Since 2 π − = I − ∗ C 2\pi_-=I-*_{C} 2 π − = I − ∗ C , we obtain θ + J i ( H ) = θ − 2 π − θ = ∗ C θ \theta+J_i(H)=\theta-2\pi_-\theta=*_{C}\theta θ + J i ( H ) = θ − 2 π − θ = ∗ C θ . This proves the second formula. Finally, ∗ C *_{C} ∗ C is an isomorphism and δ C = − ∗ C d C ∗ C \delta_C=-*_{C}d_C*_{C} δ C = − ∗ C d C ∗ C on two forms. Thus d C ( ∗ C θ ) = 0 d_C(*_{C}\theta)=0 d C ( ∗ C θ ) = 0 is equivalent to δ C θ = 0 \delta_C\theta=0 δ C θ = 0 . □
The proposition gives conditions for a preservation condition. It is not automatically satisfied, as the compact example below will show. To obtain preservation, we add a term that cancels this exact three form. We will do this by changing the ambient form through a Lie derivative whose vector field is prescribed along the submanifold.
We first specify the spaces on which the cancellation equation is solved. We suppress the dependence on the ambient G 2 G_2 G 2 structure φ \varphi φ in the notation J i J_i J i , D i D_i D i , and R i R_i R i .
Lemma 3.3. If i i i is coassociative, D i D_i D i is surjective onto exact three forms. Its kernel is J i − 1 H − 2 ( C ) J_i^{-1}\mathcal{H}^2_-(C) J i − 1 H − 2 ( C ) , and its minimum norm right inverse is
R i q = 2 J i − 1 π − δ C G C q , q ∈ d C Ω 2 ( C ) , (20) R_iq=2J_i^{-1}\pi_-\delta_CG_Cq,\qquad q\in d_C\Omega^2(C), \tag*{(20)} R i q = 2 J i − 1 π − δ C G C q , q ∈ d C Ω 2 ( C ) , (20) where G C G_C G C is the Green operator on three forms. In particular, D i R i q = q D_iR_iq=q D i R i q = q and R i q ⊥ ker D i R_iq\perp\ker D_i R i q ⊥ ker D i .
Proof. We construct an anti-self-dual primitive of q q q . Set η = δ C G C q \eta=\delta_CG_Cq η = δ C G C q . Since q q q is exact it has zero harmonic projection, and also that d C q = 0 d_Cq=0 d C q = 0 . Since the Green operator commutes with d C d_C d C , we have d C G C q = G C d C q = 0 d_CG_Cq=G_Cd_Cq=0 d C G C q = G C d C q = 0 . The identity Δ C G C q = q \Delta_CG_Cq=q Δ C G C q = q now reduces to d C η = q d_C\eta=q d C η = q . Also δ C η = δ C 2 G C q = 0 \delta_C\eta=\delta_C^2G_Cq=0 δ C η = δ C 2 G C q = 0 , which is equivalent to d C ( ∗ C η ) = 0 d_C(*_C\eta)=0 d C ( ∗ C η ) = 0 . Subtracting ∗ C η *_C\eta ∗ C η therefore does not change the exterior derivative, and gives an anti-self-dual form
d C ( 2 π − η ) = d C ( η − ∗ C η ) = q . d_C(2\pi_-\eta)=d_C(\eta-*_C\eta)=q. d C ( 2 π − η ) = d C ( η − ∗ C η ) = q .
Applying J i − 1 J_i^{-1} J i − 1 proves that the formulas is a right inverse. To determine the kernel, suppose D i V = 0 D_iV=0 D i V = 0 . Then J i V J_iV J i V is closed and anti-self-dual. Anti-self-duality makes a closed two form co-closed as well, so J i V J_iV J i V is harmonic. The converse is immediate.
We next check that the right inverse takes values orthogonal to this kernel. If h h h is harmonic and anti-self-dual, integration by parts gives ⟨ 2 π − δ C G C q , h ⟩ L 2 = 2 ⟨ G C q , d C h ⟩ L 2 = 0 \langle2\pi_-\delta_CG_Cq,h\rangle_{L^2}=2\langle G_Cq,d_Ch\rangle_{L^2}=0 ⟨ 2 π − δ C G C q , h ⟩ L 2 = 2 ⟨ G C q , d C h ⟩ L 2 = 0 . The metric identity in (17) transfers this orthogonality to normal fields. Every other solution of D i V = q D_iV=q D i V = q differs from R i q R_iq R i q by an element k k k of the kernel. Orthogonality gives ∥ R i q + k ∥ 2 2 = ∥ R i q ∥ 2 2 + ∥ k ∥ 2 2 \|R_iq+k\|_2^2=\|R_iq\|_2^2+\|k\|_2^2 ∥ R i q + k ∥ 2 2 = ∥ R i q ∥ 2 2 + ∥ k ∥ 2 2 , proving both minimality and uniqueness of the minimum norm solution. In particular, a nonzero b 2 − ( C ) b_2^-(C) b 2 − ( C ) is the degree of freedom for the solution.
For later use, we also express this right inverse through an elliptic operator. The normalization of J i J_i J i gives J i J i ∗ = 2 π − J_iJ_i^*=2\pi_- J i J i ∗ = 2 π − , so on exact three forms
D i D i ∗ q = 2 d C π − δ C q = d C δ C q = Δ C q . (21) D_iD_i^*q=2d_C\pi_-\delta_Cq=d_C\delta_Cq=\Delta_Cq. \tag*{(21)} D i D i ∗ q = 2 d C π − δ C q = d C δ C q = Δ C q . (21)
Here d C ∗ C δ C q = 0 d_C*_C\delta_Cq=0 d C ∗ C δ C q = 0 follows from δ C 2 q = 0 \delta_C^2q=0 δ C 2 q = 0 , and d C q = 0 d_Cq=0 d C q = 0 gives the last equality. There are no nonzero exact harmonic forms. Thus D i D i ∗ D_iD_i^* D i D i ∗ is invertible on the exact forms, and the same minimum norm inverse is R i = D i ∗ ( D i D i ∗ ) − 1 R_i=D_i^*(D_iD_i^*)^{-1} R i = D i ∗ ( D i D i ∗ ) − 1 . □
We must define the correction before assuming that a solution preserves coassociativity. We therefore extend the inverse R i R_i R i in (20) to an open neighborhood of the initial pair ( φ 0 , i 0 ) (\varphi_0,i_0) ( φ 0 , i 0 ) , including embeddings that are not necessarily coassociative. The next lemma supplies this extension and the estimates needed later. Its proof first constructs some right inverse, then removes its kernel component to obtain the minimum norm inverse.
Lemma 3.4. For smooth pairs ( φ , i ) (\varphi,i) ( φ , i ) sufficiently close to a coassociative pair in the C 2 × C 3 C^2 \times C^3 C 2 × C 3 topology, with d φ = 0 d\varphi=0 d φ = 0 , the operator D i D_i D i remains surjective onto exact three forms and has a finite-dimensional kernel of dimension b 2 ( C ) b_2(C) b 2 ( C ) . Its minimum norm right inverse
R i = D i ∗ ( D i D i ∗ ) − 1 (22) R_i=D_i^*(D_iD_i^*)^{-1} \tag*{(22)} R i = D i ∗ ( D i D i ∗ ) − 1 (22) is defined on exact three forms and depends smoothly on smooth finite-dimensional families of such pairs.
Proof. Fix a reference coassociative pair ( φ 0 , i 0 ) (\varphi_0,i_0) ( φ 0 , i 0 ) , and write ν 0 = ν φ 0 , i 0 \nu_0=\nu_{\varphi_0,i_0} ν 0 = ν φ 0 , i 0 and D 0 = D φ 0 , i 0 D_0=D_{\varphi_0,i_0} D 0 = D φ 0 , i 0 . Let R 0 R_0 R 0 be the right inverse in Lemma 3.3 . We will compare nearby operators on fixed spaces, so that smallness of their difference has a precise meaning.
We first check the symbol. We use Hamilton’s idea of computing the symbol of an operator by replacing the derivatives by the Fourier transform variable. We use the convention ∂ j ↦ − 1 ξ j \partial_j\mapsto\sqrt{-1}\xi_j ∂ j ↦ − 1 ξ j . Thus, the symbol of the operator D i D_i D i is σ ξ ( D i ) V = − 1 ξ ∧ J i V \sigma_\xi(D_i)V=\sqrt{-1}\xi\wedge J_iV σ ξ ( D i ) V = − 1 ξ ∧ J i V at a coassociative pair. Take ξ = ∣ ξ ∣ e 1 ≠ 0 \xi=|\xi|e^1\ne0 ξ = ∣ ξ ∣ e 1 = 0 . If an anti-self-dual two form α \alpha α satisfies ξ ∧ α = 0 \xi\wedge\alpha=0 ξ ∧ α = 0 , then α = e 1 ∧ λ \alpha=e^1\wedge\lambda α = e 1 ∧ λ . But ∗ C α *_C\alpha ∗ C α has no e 1 e^1 e 1 component, whereas α = − ∗ C α \alpha=-*_C\alpha α = − ∗ C α , hence α = 0 \alpha=0 α = 0 . The symbol is therefore injective. Its image is contained in ker ( ξ ∧ : Λ 3 T C ∗ → Λ 4 T C ∗ ) \ker(\xi\wedge:\Lambda^3T_C^*\to\Lambda^4T_C^*) ker ( ξ ∧ : Λ 3 T C ∗ → Λ 4 T C ∗ ) , which has real dimension three, the same as the normal bundle and hence it is an isomorphism onto that space. On the compact unit cosphere, the inverse symbol has a uniform bound. The same property therefore holds for sufficiently nearby pairs. In particular, D i ∗ D i D_i^*D_i D i ∗ D i is elliptic there.
Use reference norms on
X = H 1 ( C , ν 0 ) , Y = d C Ω 2 ( C ) ‾ L 2 ( g C , 0 ) . \mathcal{X}=H^1(C,\nu_0),\qquad Y=\overline{d_C\Omega^2(C)}^{L^2(g_{C,0})}. X = H 1 ( C , ν 0 ) , Y = d C Ω 2 ( C ) L 2 ( g C , 0 ) .
Hodge decomposition shows that Y Y Y is the closed space of exact L 2 L^2 L 2 three forms. We identify the normal bundles as follows. Transport a vector from i 0 ( p ) i_0(p) i 0 ( p ) to i ( p ) i(p) i ( p ) along the short geodesic of a fixed background metric, and then project it onto ν i , p \nu_{i,p} ν i , p . For nearby pairs this gives a bundle isomorphism S i : ν 0 → ν i S_i:\nu_0\to\nu_i S i : ν 0 → ν i , depending smoothly on the pair and satisfying S 0 = I S_0=I S 0 = I . Thus we can regard every nearby operator as an operator with the same domain and target and we have
D ^ i = D i S i : X → Y , ∥ D ^ i − D 0 ∥ H 1 → L 2 ≤ C ∥ J i S i − J 0 ∥ C 1 . \widehat{D}_i=D_iS_i:\mathcal{X}\to Y,\qquad\|\widehat{D}_i-D_0\|_{H^1\to L^2}\le C\|J_iS_i-J_0\|_{C^1}. D i = D i S i : X → Y , ∥ D i − D 0 ∥ H 1 → L 2 ≤ C ∥ J i S i − J 0 ∥ C 1 .
The estimate follows by differentiating J i S i − J 0 J_iS_i-J_0 J i S i − J 0 once, since D ^ i − D 0 = d C ( J i S i − J 0 ) \widehat{D}_i-D_0=d_C(J_iS_i-J_0) D i − D 0 = d C ( J i S i − J 0 ) . Make the pair close enough that ∥ ( D ^ i − D 0 ) R 0 ∥ Y → Y < 1 \|(\widehat{D}_i-D_0)R_0\|_{Y\to Y}<1 ∥ ( D i − D 0 ) R 0 ∥ Y → Y < 1 . We can then define
R ^ i = R 0 [ I Y + ( D ^ i − D 0 ) R 0 ] − 1 , R ~ i = S i R ^ i , D i R ~ i = I Y . \widehat{R}_i=R_0[I_Y+(\widehat{D}_i-D_0)R_0]^{-1},\qquad\widetilde{R}_i=S_i\widehat{R}_i,\qquad D_i\widetilde{R}_i=I_Y. R i = R 0 [ I Y + ( D i − D 0 ) R 0 ] − 1 , R i = S i R i , D i R i = I Y .
The inverse in brackets is the convergent Neumann series for I Y + ( D ^ i − D 0 ) R 0 I_Y+(\widehat{D}_i-D_0)R_0 I Y + ( D i − D 0 ) R 0 . Multiplying by D ^ i \widehat{D}_i D i shows directly that D ^ i R ^ i = I Y \widehat{D}_i\widehat{R}_i=I_Y D i R i = I Y , proving surjectivity.
The inverse found above need not yet have minimum norm. We now do that. The fixed decomposition X = ker D 0 ⊕ im R 0 \mathcal{X}=\ker D_0\oplus\operatorname{im}R_0 X = ker D 0 ⊕ im R 0 also describes the new kernel. The restriction of D ^ i \widehat{D}_i D i to im R 0 \operatorname{im}R_0 im R 0 is an isomorphism onto Y Y Y . Thus, for each k ∈ ker D 0 k\in\ker D_0 k ∈ ker D 0 , there is a unique component in R 0 R_0 R 0 that cancels D ^ i k \widehat{D}_ik D i k , namely − R ^ i D ^ i k -\widehat{R}_i\widehat{D}_ik − R i D i k . Consequently
ker D ^ i = { k − R ^ i D ^ i k : k ∈ ker D 0 } . \ker\widehat{D}_i=\{k-\widehat{R}_i\widehat{D}_ik:k\in\ker D_0\}. ker D i = { k − R i D i k : k ∈ ker D 0 } .
The ker D 0 \ker D_0 ker D 0 component is exactly k k k , so this parametrization is injective. It is also surjective by the fixed decomposition. The kernel dimension therefore remains b 2 ( C ) b_2(C) b 2 ( C ) . Let Π i \Pi_i Π i be the current L 2 L^2 L 2 projection onto ker D i \ker D_i ker D i . Subtracting Π i R ~ i q \Pi_i\widetilde{R}_iq Π i R i q does not change the image under D i D_i D i , and makes the result orthogonal to its kernel. The minimum norm right inverse is therefore R i = ( I − Π i ) R ~ i R_i=(I-\Pi_i)\widetilde{R}_i R i = ( I − Π i ) R i .
To prove the regularity of the right inverse, we next prove the derivative estimates and smooth dependence. The operator D i D i ∗ D_iD_i^* D i D i ∗ is used only on exact forms. To apply the usual elliptic estimates, we place it inside an elliptic operator on all three forms, by adding a term that vanishes on closed forms. Write
Y s = Y ∩ H s ( C , Λ 3 T ∗ C ) , L i = D i D i ∗ + δ C d C . Y_s=Y\cap H^s(C,\Lambda^3T^*C),\qquad L_i=D_iD_i^*+\delta_Cd_C. Y s = Y ∩ H s ( C , Λ 3 T ∗ C ) , L i = D i D i ∗ + δ C d C .
The underlying spaces are fixed, but the adjoints in L i L_i L i use the current metrics. Its symbol has quadratic form ∣ σ ξ ( D i ) ∗ q ∣ 2 + ∣ ξ ∧ q ∣ 2 |\sigma_\xi(D_i)^*q|^2+|\xi\wedge q|^2 ∣ σ ξ ( D i ) ∗ q ∣ 2 + ∣ ξ ∧ q ∣ 2 . If this is zero, then q ∈ ker ( ξ ∧ ) = im σ ξ ( D i ) q\in\ker(\xi\wedge)=\operatorname{im}\sigma_\xi(D_i) q ∈ ker ( ξ ∧ ) = im σ ξ ( D i ) and q q q is orthogonal to that same image. Hence q = 0 q=0 q = 0 , proving ellipticity. The operator is also nonnegative and self-adjoint.
We must check that restriction to exact forms is compatible with this self-adjoint operator. The current L 2 L^2 L 2 orthogonal decomposition is Y ⊕ Y ⊥ Y \oplus Y^\perp Y ⊕ Y ⊥ , where Y ⊥ = ker δ C Y^\perp= \ker\delta_C Y ⊥ = ker δ C , and Hodge theory makes its projections bounded on H 2 H^2 H 2 . If q q q is exact, then d C q = 0 d_Cq = 0 d C q = 0 and L i q = D i D i ∗ q L_iq = D_iD_i^*q L i q = D i D i ∗ q is exact. If q ∈ Y ⊥ q \in Y^\perp q ∈ Y ⊥ , then D i ∗ q = J i ∗ δ C q = 0 D_i^*q = J_i^*\delta_Cq = 0 D i ∗ q = J i ∗ δ C q = 0 , so L i q = δ C d C q ∈ ker δ C L_iq = \delta_Cd_Cq \in\ker\delta_C L i q = δ C d C q ∈ ker δ C . Thus L i L_i L i preserves both orthogonal summands. Its restriction to Y Y Y is consequently self-adjoint as well. All these pairings use the current metrics and near the reference pair their norms are uniformly equivalent to the fixed reference norms.
On Y Y Y , we have L i = D i D i ∗ L_i = D_iD_i^* L i = D i D i ∗ . If L i q = 0 L_iq = 0 L i q = 0 , pairing with q q q gives D i ∗ q = 0 D_i^*q = 0 D i ∗ q = 0 . Such a q q q is orthogonal to the image of D i D_i D i , which is all of Y Y Y , so q = 0 q = 0 q = 0 . The self-adjoint elliptic operator on this closed subspace has compact resolvent and zero kernel. It is therefore invertible, with L i − 1 : Y 0 → Y 2 L_i^{-1}:Y_0 \to Y_2 L i − 1 : Y 0 → Y 2 . Here and below, L i − 1 L_i^{-1} L i − 1 means the inverse of this restriction, not an inverse on all three forms. The elliptic estimate at every integer order s ≥ 0 s \ge0 s ≥ 0 is
∥ q ∥ H s + 2 ≤ C i , s ( ∥ L i q ∥ H s + ∥ q ∥ L 2 ) . \lVert q\rVert_{H^{s+2}} \le C_{i,s}\bigl(\lVert L_iq\rVert_{H^s}+\lVert q\rVert_{L^2}\bigr). ∥ q ∥ H s + 2 ≤ C i , s ( ∥ L i q ∥ H s + ∥ q ∥ L 2 ) .
The low-order inverse bound is uniform in a sufficiently small initial neighborhood. Indeed, the already constructed right inverse satisfies D i R ~ i q = q D_i\widetilde{R}_iq = q D i R i q = q . For q ∈ Y 1 q \in Y_1 q ∈ Y 1 , integration by parts gives ∥ q ∥ 2 2 = ⟨ R ~ i q , D i ∗ q ⟩ \lVert q\rVert_2^2 = \langle\widetilde{R}_iq,D_i^*q\rangle ∥ q ∥ 2 2 = ⟨ R i q , D i ∗ q ⟩ . The uniform bound for R ~ i \widetilde{R}_i R i therefore gives ∥ q ∥ 2 ≤ C ∥ D i ∗ q ∥ 2 \lVert q\rVert_2 \le C\lVert D_i^*q\rVert_2 ∥ q ∥ 2 ≤ C ∥ D i ∗ q ∥ 2 . This bounds the inverse of L i L_i L i in L 2 L^2 L 2 . Applying the preceding elliptic estimate to L i − 1 b L_i^{-1}b L i − 1 b now gives
∥ L i − 1 b ∥ H s + 2 ≤ C i , s ∥ b ∥ H s , R i = D i ∗ L i − 1 : Y s ⟶ H s + 1 ( C , ν i ) . \lVert L_i^{-1}b\rVert_{H^{s+2}} \le C_{i,s}\lVert b\rVert_{H^s}, \qquad R_i = D_i^*L_i^{-1}:Y_s \longrightarrow H^{s+1}(C,\nu_i). ∥ L i − 1 b ∥ H s + 2 ≤ C i , s ∥ b ∥ H s , R i = D i ∗ L i − 1 : Y s ⟶ H s + 1 ( C , ν i ) .
Thus R i R_i R i gains one derivative, and its formula is exactly (22) . These estimates hold on the same low-order neighborhood at every order. Their constants may depend on higher norms of the coefficients.
For a smooth finite-dimensional family of pairs, L i : Y s + 2 → Y s L_i:Y_{s+2}\to Y_s L i : Y s + 2 → Y s depends smoothly on the parameters. Inversion of bounded isomorphisms is smooth. Differentiating L i L i − 1 = I L_iL_i^{-1}=I L i L i − 1 = I gives ∂ L i − 1 = − L i − 1 ( ∂ L i ) L i − 1 \partial L_i^{-1}=-L_i^{-1}(\partial L_i)L_i^{-1} ∂ L i − 1 = − L i − 1 ( ∂ L i ) L i − 1 . The coefficients of S i − 1 D i ∗ S_i^{-1}D_i^* S i − 1 D i ∗ also depend smoothly on the pair. Hence S i − 1 R i : Y s → H s + 1 ( C , ν 0 ) S_i^{-1}R_i:Y_s\to H^{s+1}(C,\nu_0) S i − 1 R i : Y s → H s + 1 ( C , ν 0 ) depends smoothly on the parameters. This is the asserted smooth dependence of R i R_i R i after identifying the moving normal bundles. □
The correction W W W in (2) is a normal field along the submanifold. To form the ambient Lie derivative, we need a vector field on M M M that restricts to W W W . We choose this extension once as part of the equation.
Fix a radius r > 0 r>0 r > 0 smaller than the initial normal tubular radius, and a smooth cutoff ρ \rho ρ equal to one near zero and zero near one. Keep r r r and ρ \rho ρ fixed for nearby pairs. By compactness and the inverse function theorem, after shrinking the neighborhood, every normal exponential map is a diffeomorphism on the normal disc bundle of radius r r r , and its inverse depends smoothly on smooth parameter families. For x = exp p φ ( ν ) x=\exp_p^{\varphi}(\nu) x = exp p φ ( ν ) , with p ∈ i ( C ) p\in i(C) p ∈ i ( C ) and ∣ ν ∣ < r \lvert\nu\rvert<r ∣ ν ∣ < r , transport W ( p ) W(p) W ( p ) along the radial normal geodesic from p p p to x x x . Writing this transport as PT p → x \operatorname{PT}_{p\to x} PT p → x , define
( E ρ , i W ) ( x ) = ρ ( ∣ ν ∣ 2 r 2 ) PT p → x W ( p ) , (23) (E_{\rho,i}W)(x)=\rho\left(\frac{\lvert\nu\rvert^2}{r^2}\right)\operatorname{PT}_{p\to x}W(p), \tag*{(23)} ( E ρ , i W ) ( x ) = ρ ( r 2 ∣ ν ∣ 2 ) PT p → x W ( p ) , (23)
and extend by zero outside the tube. Since the cutoff vanishes near the boundary of the tube, the resulting ambient field is smooth.
We record how this choice behaves under changes of parametrization and ambient coordinates. Replacing i , W i,W i , W by i ∘ a , W ∘ a i\circ a,W\circ a i ∘ a , W ∘ a , where a a a is a diffeomorphism of C C C , leaves the image and its normal field unchanged, hence leaves the extension unchanged. If F F F is an ambient diffeomorphism, then F : ( M , F ∗ g ) → ( M , g ) F:(M,F^*g)\to(M,g) F : ( M , F ∗ g ) → ( M , g ) is an isometry. It preserves normal exponential maps, radial parallel transport, and ∣ ν ∣ \lvert\nu\rvert ∣ ν ∣ . The extension formed from F ∗ φ , F − 1 ∘ i F^*\varphi,F^{-1}\circ i F ∗ φ , F − 1 ∘ i , and d F − 1 W dF^{-1}W d F − 1 W is therefore F ∗ ( E ρ , i W ) F^*(E_{\rho,i}W) F ∗ ( E ρ , i W ) . This is the equivariance needed in § 3.3 . We retain r , ρ r,\rho r , ρ as our fixed data and write X φ , i = E φ , i W φ , i X_{\varphi,i}=E_{\varphi,i}W_{\varphi,i} X φ , i = E φ , i W φ , i .
Proposition 3.5. Every smooth solution of (3) remaining in the neighborhood of Lemma 3.4 , where the fixed tubular extension is defined, satisfies
∂ t ( i t ∗ φ t ) = 0. (24) \partial_t(i_t^*\varphi_t)=0. \tag*{(24)} ∂ t ( i t ∗ φ t ) = 0. (24) The cohomology class [ φ t ] [\varphi_t] [ φ t ] is constant. At a coassociative pair, W W W is given by (4) and is the unique normal field along i ( C ) i(C) i ( C ) of minimum L 2 ( C , g C ) L^2(C,g_C) L 2 ( C , g C ) norm enforcing (24) with submanifold velocity H H H .
Proof. We use the pullback derivative, now including the ambient Lie derivative. Since d φ = 0 d\varphi=0 d φ = 0 , Cartan’s formula gives L X φ = d ( X ⌟ φ ) \mathcal{L}_X\varphi=d(X\mathbin{\lrcorner}\varphi) L X φ = d ( X ┘ φ ) . Since X X X restricts to W W W on the submanifold, its pullback contribution is D i W D_iW D i W . Thus
∂ t ( i t ∗ φ ) = i ∗ ( d τ + L X φ ) + d c i ∗ ( H ⌟ φ ) = d C θ + D i W + D i H = 0. \partial_t(i_t^*\varphi)=i^*(d\tau+\mathcal{L}_X\varphi)+dc\,i^*(H\mathbin{\lrcorner}\varphi)=d_C\theta+D_iW+D_iH=0. ∂ t ( i t ∗ φ ) = i ∗ ( d τ + L X φ ) + d c i ∗ ( H ┘ φ ) = d C θ + D i W + D i H = 0.
The last equality is the defining equation D i W = − d C θ − D i H D_iW=-d_C\theta-D_iH D i W = − d C θ − D i H . Notice that no coassociativity assumption was used in this cancellation, because the inverse was defined on an open neighborhood. Integrating in time proves (24) . Also ∂ t φ = d ( τ + X ⌟ φ ) \partial_t\varphi=d(\tau+X\mathbin{\lrcorner}\varphi) ∂ t φ = d ( τ + X ┘ φ ) is exact, so [ φ t ] = [ φ 0 ] [\varphi_t]=[\varphi_0] [ φ t ] = [ φ 0 ] .
In the same way as in Lemma 3.3 we get that W W W has the smallest L 2 L^2 L 2 -norm.
We can also remove the Lie derivative from the ambient equation. This will be useful for proving existence. Let F t F_t F t solve ∂ t F t = − X t ∘ F t \partial_tF_t=-X_t\circ F_t ∂ t F t = − X t ∘ F t with F 0 = id F_0=\operatorname{id} F 0 = id . The minus sign is chosen so that differentiation of the pullback cancels L X φ \mathcal{L}_{X}\varphi L X φ . Precisely, we have
∂ t ( F t ∗ φ t ) = F t ∗ ( ∂ t φ t − L X t φ t ) = Δ F t ∗ φ t ( F t ∗ φ t ) . \partial_t(F_t^*\varphi_t)=F_t^*(\partial_t\varphi_t-\mathcal{L}_{X_t}\varphi_t)=\Delta_{F_t^*\varphi_t}(F_t^*\varphi_t). ∂ t ( F t ∗ φ t ) = F t ∗ ( ∂ t φ t − L X t φ t ) = Δ F t ∗ φ t ( F t ∗ φ t ) .
Set φ ~ t = F t ∗ φ t \widetilde{\varphi}_t=F_t^*\varphi_t φ t = F t ∗ φ t and ı ~ t = F t − 1 ∘ ı t \widetilde{\imath}_t=F_t^{-1}\circ\imath_t t = F t − 1 ∘ t . Then ı ~ t ∗ φ ~ t = ı t ∗ φ t \widetilde{\imath}_t^*\widetilde{\varphi}_t=\imath_t^*\varphi_t t ∗ φ t = t ∗ φ t . Differentiating F t − 1 ∘ ı t F_t^{-1}\circ\imath_t F t − 1 ∘ t adds W W W to the mean-curvature velocity: the inverse diffeomorphism moves with the opposite sign to F t F_t F t . Since R i D i = I − Π i R_iD_i=I-\Pi_i R i D i = I − Π i on normal fields, the same velocity can be written as H + W = Π i H − R i d C θ H+W=\Pi_iH-R_id_C\theta H + W = Π i H − R i d C θ . Pulling back gives
∂ t ı ~ t = d F t − 1 ( H t + W t ) = Π ı ~ t H ~ t − R ı ~ t d C θ ~ t , θ ~ t = ı ~ t ∗ ( δ φ ~ t φ ~ t ) . (25) \partial_t\widetilde{\imath}_t=dF_t^{-1}(H_t+W_t)=\Pi_{\widetilde{\imath}_t}\widetilde{H}_t-R_{\widetilde{\imath}_t}d_C\widetilde{\theta}_t,\qquad\widetilde{\theta}_t=\widetilde{\imath}_t^*(\delta_{\widetilde{\varphi}_t}\widetilde{\varphi}_t). \tag*{(25)} ∂ t t = d F t − 1 ( H t + W t ) = Π t H t − R t d C θ t , θ t = t ∗ ( δ φ t φ t ) . (25)
Here Π ı ~ t \Pi_{\widetilde{\imath}_t} Π t is the L 2 L^2 L 2 projection onto ker D ı ~ t \ker D_{\widetilde{\imath}_t} ker D t , and H ~ t \widetilde{H}_t H t uses the pulled back metric. The normal operators commute with pullback because F t F_t F t is an isometry between the two metrics. It preserves their defining equations and the L 2 L^2 L 2 orthogonality used to select the inverse.
Thus, when the ambient G 2 G_2 G 2 structure follows ordinary Laplacian flow, the normal velocity has two parts. The equation D i V = − d C θ D_iV=-d_C\theta D i V = − d C θ forces the component − R i d C θ -R_id_C\theta − R i d C θ orthogonal to the kernel. The component Π i H \Pi_iH Π i H lies in the kernel and does not change that equation. Keeping this component selects the admissible velocity closest to H H H in L 2 L^2 L 2 . At a coassociative pair, J i J_i J i identifies Π i \Pi_i Π i with harmonic projection on anti-self-dual two forms. In particular, Π i H = 0 \Pi_iH=0 Π i H = 0 if b 2 − ( C ) = 0 b_2^-(C)=0 b 2 − ( C ) = 0 .
Remark 3.6. At a coassociative pair with torsion-free φ \varphi φ , both τ \tau τ and H H H vanish, so W = 0 W=0 W = 0 .
We can now give a proof of part of our first main result.
Proof of Theorem 1.1 . We have from Proposition 3.5 that ı t ∗ φ t = ı 0 ∗ φ 0 \imath_t^*\varphi_t=\imath_0^*\varphi_0 t ∗ φ t = 0 ∗ φ 0 and [ φ t ] = [ φ 0 ] [\varphi_t]=[\varphi_0] [ φ t ] = [ φ 0 ] for every solution in the stated neighborhood. Since ı 0 ∗ φ 0 = 0 \imath_0^*\varphi_0=0 0 ∗ φ 0 = 0 hence the embeddings remain coassociative. The same proposition gives the formula (1.4) and its minimum norm property. These are the preservation assertions of the theorem. The short-time existence and uniqueness assertion is proved separately in Theorem 3.11 .
The proof of the following corollary is similar to Proposition 3.5 and the existence and uniqueness part is given at the end of § 3.3 .
Corollary 3.7. With the same initial hypotheses and fixed embedding ı = ı 0 \imath=\imath_0 = 0 , let W 0 = − R i d C ( ı ∗ τ ) W^0=-R_id_C(\imath^*\tau) W 0 = − R i d C ( ∗ τ ) and X 0 = E φ , ı W 0 X^0=E_{\varphi,\imath}W^0 X 0 = E φ , W 0 . The equation
∂ t φ = Δ φ φ + L X 0 φ (26) \partial_t\varphi=\Delta_\varphi\varphi+\mathcal{L}_{X^0}\varphi \tag*{(26)} ∂ t φ = Δ φ φ + L X 0 φ (26)
has a unique smooth closed solution from φ 0 \varphi_0 φ 0 for a short time, and every smooth closed solution from φ 0 \varphi_0 φ 0 remaining in the same neighborhood preserves ı ∗ φ = 0 \imath^*\varphi=0 ∗ φ = 0 .
An explicit compact example
Before proving the short-time existence and uniqueness assertions in Theorem 1.1 we give an explicit example which shows the need to modify the coupled system of the Laplacian flow of closed G 2 G_2 G 2 -structures and the mean curvature of submanifolds. Precisely, we show that the error in Proposition 3.2 can be nonzero for compact initial data, even arbitrarily close to a torsion-free structure. The example also lets us compute the correcting normal field directly.
Proposition 3.8. There are closed G 2 G_2 G 2 structures on T 7 T^7 T 7 , arbitrarily close and cohomologous to a flat torsion-free G 2 G_2 G 2 structure, and compact coassociative four-tori for which ordinary Laplacian flow coupled to mean curvature flow fails to preserve the coassociative condition.
Proof. Use coordinates ( y 1 , y 2 , y 3 , x 1 , x 2 , x 3 , x 4 ) (y^1,y^2,y^3,x^1,x^2,x^3,x^4) ( y 1 , y 2 , y 3 , x 1 , x 2 , x 3 , x 4 ) of period 2 π 2\pi 2 π and the model from the earlier mean-curvature calculation. Thus, we have
φ f l a t = d y 123 + d y 1 ∧ ( d x 12 + d x 34 ) + d y 2 ∧ ( d x 13 − d x 24 ) − d y 3 ∧ ( d x 14 + d x 23 ) . \varphi_{\mathrm{flat}}=dy^{123}+dy^1\wedge(dx^{12}+dx^{34})+dy^2\wedge(dx^{13}-dx^{24})-dy^3\wedge(dx^{14}+dx^{23}). φ flat = d y 123 + d y 1 ∧ ( d x 12 + d x 34 ) + d y 2 ∧ ( d x 13 − d x 24 ) − d y 3 ∧ ( d x 14 + d x 23 ) .
Let
φ ε = φ f l a t + ε sin y 1 sin x 1 d y 1 ∧ d x 12 , 0 < ∣ ε ∣ < 1. (27) \varphi_{\varepsilon}=\varphi_{\mathrm{flat}}+\varepsilon\sin y^{1}\sin x^{1}\,dy^{1}\wedge dx^{12},\qquad0<|\varepsilon|<1. \tag*{(27)} φ ε = φ flat + ε sin y 1 sin x 1 d y 1 ∧ d x 12 , 0 < ∣ ε ∣ < 1. (27)
The added term is d ( − ε cos y 1 sin x 1 d x 12 ) d(-\varepsilon\cos y^{1}\sin x^{1}\,dx^{12}) d ( − ε cos y 1 sin x 1 d x 12 ) . It is therefore exact, so the new form is closed and has the same cohomology class as the flat form. To check that it is indeed a G 2 G_{2} G 2 structure and identify the metric, we put a = 1 + ε sin y 1 sin x 1 > 0 a=1+\varepsilon\sin y^{1}\sin x^{1}>0 a = 1 + ε sin y 1 sin x 1 > 0 . Multiply the coordinate coframe by a 1 / 3 a^{1/3} a 1/3 in the y 1 , x 1 , x 2 y^{1},x^{1},x^{2} y 1 , x 1 , x 2 directions and by a − 1 / 6 a^{-1/6} a − 1/6 in the other four directions. Substituting this coframe into the flat model gives (27) . Thus this is an orthonormal coframe for its induced metric, and
g ε = a 2 3 ( ( d y 1 ) 2 + ( d x 1 ) 2 + ( d x 2 ) 2 ) + a − 1 3 ( ( d y 2 ) 2 + ( d y 3 ) 2 + ( d x 3 ) 2 + ( d x 4 ) 2 ) . (28) g_{\varepsilon}=a^{\frac{2}{3}}\left((dy^{1})^{2}+(dx^{1})^{2}+(dx^{2})^{2}\right)+a^{-\frac{1}{3}}\left((dy^{2})^{2}+(dy^{3})^{2}+(dx^{3})^{2}+(dx^{4})^{2}\right). \tag*{(28)} g ε = a 3 2 ( ( d y 1 ) 2 + ( d x 1 ) 2 + ( d x 2 ) 2 ) + a − 3 1 ( ( d y 2 ) 2 + ( d y 3 ) 2 + ( d x 3 ) 2 + ( d x 4 ) 2 ) . (28)
Every summand of φ ε \varphi_{\varepsilon} φ ε contains a d y dy d y factor, so its pullback to C = { y 1 = y 2 = y 3 = 0 } C=\{y^{1}=y^{2}=y^{3}=0\} C = { y 1 = y 2 = y 3 = 0 } vanishes. Hence C C C is coassociative. Since a = 1 a=1 a = 1 on C C C , the induced metric is flat, and its coassociative orientation is − d x 1234 -dx^{1234} − d x 1234 . We remark that flatness of the induced metric does not force H H H to vanish as H H H also depends on normal derivatives of the ambient metric, which need not vanish on C C C .
To calculate θ = ι ∗ τ \theta=\iota^{*}\tau θ = ι ∗ τ , we use τ = − ∗ d ψ \tau=-*\ d\psi τ = − ∗ d ψ . Using the facts that a = 1 a=1 a = 1 along C C C , a short computation gives
θ = ι ∗ τ = ε 3 sin x 1 ( 2 d x 12 − d x 34 ) . (29) \theta=\iota^{*}\tau=\frac{\varepsilon}{3}\sin x^{1}(2dx^{12}-dx^{34}). \tag*{(29)} θ = ι ∗ τ = 3 ε sin x 1 ( 2 d x 12 − d x 34 ) . (29)
We can also compute H H H from the metric. The trace of ∂ y 1 g ε \partial_{y^{1}}g_{\varepsilon} ∂ y 1 g ε over T C TC T C is ( 2 3 + 2 3 − 1 3 − 1 3 ) ε sin x 1 = 2 3 ε sin x 1 \left(\frac{2}{3}+\frac{2}{3}-\frac{1}{3}-\frac{1}{3}\right)\varepsilon\sin x^{1}=\frac{2}{3}\varepsilon\sin x^{1} ( 3 2 + 3 2 − 3 1 − 3 1 ) ε sin x 1 = 3 2 ε sin x 1 . Thus
H = − 1 2 tr T C ( ∂ y 1 g ε ) ∂ y 1 = − ε 3 sin x 1 ∂ y 1 . H=-\frac{1}{2}\operatorname{tr}_{TC}(\partial_{y^{1}}g_{\varepsilon})\partial_{y^{1}}=-\frac{\varepsilon}{3}\sin x^{1}\,\partial_{y^{1}}. H = − 2 1 tr T C ( ∂ y 1 g ε ) ∂ y 1 = − 3 ε sin x 1 ∂ y 1 .
This agrees with J H = − 2 π − θ JH=-2\pi^{-}\theta J H = − 2 π − θ . Since ∗ C d x 12 = − d x 34 *_{C}dx^{12}=-dx^{34} ∗ C d x 12 = − d x 34 and ∗ C d x 34 = − d x 12 *_{C}dx^{34}=-dx^{12} ∗ C d x 34 = − d x 12 ,
d C ( ∗ C θ ) = − 2 ε 3 cos x 1 d x 134 ≠ 0. (30) d_{C}(*_{C}\theta)=-\frac{2\varepsilon}{3}\cos x^{1}\,dx^{134}\ne0. \tag*{(30)} d C ( ∗ C θ ) = − 3 2 ε cos x 1 d x 134 = 0. (30)
Proposition 3.2 now gives a nonzero initial derivative of ι t ∗ φ t \iota_{t}^{*}\varphi_{t} ι t ∗ φ t , although its initial value is zero. The ordinary ambient Laplacian flow exists for a short time. Mean curvature flow in its smooth time-dependent metric also has the usual short-time parabolic existence. Thus these initial data give actual solutions of the unmodified coupling for which coassociativity is not preserved.
Finally,
W = 2 ε 3 sin x 1 ∂ y 1 (31) W=\frac{2\varepsilon}{3}\sin x^{1}\,\partial_{y^{1}} \tag*{(31)} W = 3 2 ε sin x 1 ∂ y 1 (31)
satisfies D ι W = − d C ( ∗ C θ ) D_{\iota}W=-d_{C}(*_{C}\theta) D ι W = − d C ( ∗ C θ ) . Indeed, J ι y 1 = d x 12 + d x 34 J\iota_{y^{1}}=dx^{12}+dx^{34} J ι y 1 = d x 12 + d x 34 , and differentiating the sine coefficient gives D ι W = 2 ε 3 cos x 1 d x 134 D_{\iota}W=\frac{2\varepsilon}{3}\cos x^{1}\,dx^{134} D ι W = 3 2 ε cos x 1 d x 134 , the negative of (30) . Since C C C is flat, its harmonic anti-self-dual forms are constant, so the kernel of D ι D_{\iota} D ι consists of constant normal fields. The sine coefficient in (31) has zero mean. Hence W W W is orthogonal to this kernel and is exactly the minimum norm correction required to make the coupling work.
Existence from a coassociative
We now construct solutions to the coupled systems. Even though we have proved when the coassociatives are preserved this alone does not prove existence, because the correction is part of the unknown coupled equation. We instead solve the ordinary ambient flow of the G 2 G_{2} G 2 structures first and choose a coassociative embedding above each ambient G 2 G_{2} G 2 structure. We then choose the motion within this family to have the required normal velocity for preservation.
The precise family needed for these steps is provided by McLean’s deformation theory [27 ] and by [16 Thm. 2.8] . We use bars below to refer to the prescribed ambient family. Along a graph ȷ ˉ ( t , z ) \bar{\jmath}(t,z) ˉ ( t , z ) , the operators D ˉ \bar{D} D ˉ , R ˉ \bar{R} R ˉ and Π ˉ \bar{\Pi} Π ˉ are formed using ( φ ˉ t , ȷ ˉ ( t , z ) ) (\bar{\varphi}_{t},\bar{\jmath}(t,z)) ( φ ˉ t , ˉ ( t , z )) , with Π ˉ \bar{\Pi} Π ˉ projecting onto ker D ˉ \ker\bar{D} ker D ˉ , and D ˉ 0 \bar{D}_{0} D ˉ 0 denotes the operator at the initial pair ( φ ˉ 0 , ȷ ˉ 0 ) (\bar{\varphi}_{0},\bar{\jmath}_{0}) ( φ ˉ 0 , ˉ 0 ) .
Lemma 3.9. Let φ ˉ t \bar{\varphi}_t φ ˉ t be a smooth family of closed G 2 G_2 G 2 structures with constant cohomology class, and let i 0 i_0 i 0 be coassociative for φ ˉ 0 \bar{\varphi}_0 φ ˉ 0 . Suppose ∂ t φ ˉ t = d τ ˉ t \partial_t\bar{\varphi}_t=d\bar{\tau}_t ∂ t φ ˉ t = d τ ˉ t with τ ˉ t \bar{\tau}_t τ ˉ t smooth. Let B ( t , j ) ∈ Γ ( ν j ) B(t,j)\in\Gamma(\nu_j) B ( t , j ) ∈ Γ ( ν j ) assign a normal field to each pair ( t , j ) (t,j) ( t , j ) in the local finite-dimensional family, smoothly in the family parameters and with B ( t , j ∘ a ) = B ( t , j ) ∘ a B(t,j\circ a)=B(t,j)\circ a B ( t , j ∘ a ) = B ( t , j ) ∘ a for diffeomorphisms a a a of C C C . Then there is a unique short-time path of coassociative embeddings starting at i 0 i_0 i 0 with normal velocity
V ˉ = Π ˉ B − R ˉ d C ( i ˉ ∗ τ ˉ ) . (32) \bar{V}=\bar{\Pi}B-\bar{R}d_C(\bar{i}^{*}\bar{\tau}). \tag*{(32)} V ˉ = Π ˉ B − R ˉ d C ( i ˉ ∗ τ ˉ ) . (32)
Proof. We first describe the nearby coassociatives, without prescribing their velocity. Use the initial metric to write nearby embedded images as normal graphs j v ( p ) = exp i 0 ( p ) v ( p ) j_v(p)=\exp_{i_0(p)}v(p) j v ( p ) = exp i 0 ( p ) v ( p ) , where v ∈ C 2 , α ( C , ν 0 ) v\in C^{2,\alpha}(C,\nu_0) v ∈ C 2 , α ( C , ν 0 ) is small. Their coassociative equation is
F ( t , v ) = j v ∗ φ ˉ t in Y 1 , α : = d C C 2 , α ( C , Λ 2 T ∗ C ) ⊂ C 1 , α ( C , Λ 3 T ∗ C ) . \mathcal{F}(t,v)=j_v^{*}\bar{\varphi}_t\ \text{ in }\ Y^{1,\alpha}:=d_C C^{2,\alpha}(C,\Lambda^2T^{*}C)\subset C^{1,\alpha}(C,\Lambda^3T^{*}C). F ( t , v ) = j v ∗ φ ˉ t in Y 1 , α := d C C 2 , α ( C , Λ 2 T ∗ C ) ⊂ C 1 , α ( C , Λ 3 T ∗ C ) .
We equip Y 1 , α Y^{1,\alpha} Y 1 , α with the C 1 , α C^{1,\alpha} C 1 , α norm. Hodge decomposition makes this space of exact three forms a closed Banach subspace. The map does take values there because the nearby graphs are isotopic, the ambient cohomology class is constant, and therefore [ j v ∗ φ ˉ t ] = [ i 0 ∗ φ ˉ 0 ] = 0 [j_v^{*}\bar{\varphi}_t]=[i_0^{*}\bar{\varphi}_0]=0 [ j v ∗ φ ˉ t ] = [ i 0 ∗ φ ˉ 0 ] = 0 . This is the point where the cohomology hypothesis enters the local construction.
At ( t , v ) = ( 0 , 0 ) (t,v)=(0,0) ( t , v ) = ( 0 , 0 ) , the derivative in v v v is D 0 D_0 D 0 . Its kernel is J 0 − 1 H − 2 ( C ) J_0^{-1}\mathcal{H}^2_-(C) J 0 − 1 H − 2 ( C ) , and the inverse formula 2 J 0 − 1 π − δ C G C 2J_0^{-1}\pi_-\delta_CG_C 2 J 0 − 1 π − δ C G C maps exact C 1 , α C^{1,\alpha} C 1 , α three forms to C 2 , α C^{2,\alpha} C 2 , α normal fields. Hence the restriction of D 0 D_0 D 0 to the fixed L 2 L^2 L 2 complement of its kernel is a bounded isomorphism onto Y 1 , α Y^{1,\alpha} Y 1 , α .
Split the normal field into its kernel component z z z and its orthogonal component w w w . The implicit function theorem solves for w w w in terms of ( t , z ) (t,z) ( t , z ) . This gives all nearby coassociative graphs j ( t , z ) j(t,z) j ( t , z ) , with z z z in an open set of the finite-dimensional space ker D 0 \ker D_0 ker D 0 .
We also need a smooth family on one common time interval, rather than a separate interval for each derivative order. This is essentially the regularity argument in [16 Theorems 2.7 and 2.8] and we give some details. The graph equation is first order elliptic in the normal variables near a coassociative, because its linearization has injective symbol. Differentiating the equation and applying elliptic estimates therefore improves the spatial regularity of a C 2 , α C^{2,\alpha} C 2 , α graph to every order when the ambient data are smooth. At each such graph, the linearization on the fixed kernel complement remains invertible by the original implicit function theorem. Elliptic regularity gives its inverse in the higher spaces as well. We may therefore apply the implicit function theorem at each higher order. The graphs are consequently smooth in space and in ( t , z ) (t,z) ( t , z ) on the same parameter neighbourhood.
We now choose the path in this family. Let V 0 = ( ∂ t j ) ⊥ V_0=(\partial_tj)^\perp V 0 = ( ∂ t j ) ⊥ be the normal velocity obtained by varying t t t while holding z z z fixed, and let Z a = ( ∂ z a j ) ⊥ Z_a=(\partial_{z_a}j)^\perp Z a = ( ∂ z a j ) ⊥ be the normal fields obtained by varying the parameters. Differentiate j ( t , z ) ∗ φ ˉ t = 0 j(t,z)^*\bar{\varphi}_t=0 j ( t , z ) ∗ φ ˉ t = 0 in these two ways to obtain
D ˉ V 0 = − d C ( j ∗ τ ˉ ) , D ˉ Z a = 0. \bar{D}V_0=-d_C(j^*\bar{\tau}),\qquad\bar{D}Z_a=0. D ˉ V 0 = − d C ( j ∗ τ ˉ ) , D ˉ Z a = 0.
Tangential components do not contribute as their pullback variation is the Lie derivative of the zero form j ∗ φ ˉ t j^*\bar{\varphi}_t j ∗ φ ˉ t . The graph parametrization identifies variations in z z z with all normal infinitesimal deformations. Hence the Z a Z_a Z a form a basis of ker D ˉ \ker\bar{D} ker D ˉ . Write
Z t , z = ∑ a v a Z a , Z t , z : R b 2 − ( C ) ⟶ ker D ˉ . (33) Z_{t,z}=\sum_a v^a Z_a,\qquad Z_{t,z}:\mathbb{R}^{b_2^-(C)}\longrightarrow\ker\bar{D}. \tag*{(33)} Z t , z = a ∑ v a Z a , Z t , z : R b 2 − ( C ) ⟶ ker D ˉ . (33)
This map is an isomorphism. The desired velocity V ˉ \bar{V} V ˉ and the fixed-parameter velocity V 0 V_0 V 0 satisfy the same inhomogeneous equation, D ˉ V ˉ = D ˉ V 0 = − d C ( j ∗ τ ˉ ) \bar{D}\bar{V}=\bar{D}V_0=-d_C(j^*\bar{\tau}) D ˉ V ˉ = D ˉ V 0 = − d C ( j ∗ τ ˉ ) . Their difference therefore lies in ker D ˉ \ker\bar{D} ker D ˉ , and there are unique coefficients b a ( t , z ) b^a(t,z) b a ( t , z ) with
V ˉ − V 0 = ∑ a b a ( t , z ) Z a . \bar{V}-V_0=\sum_a b^a(t,z)Z_a. V ˉ − V 0 = a ∑ b a ( t , z ) Z a .
Pair this equation with each Z b Z_b Z b in L 2 L^2 L 2 . The resulting matrix is the Gram matrix of a basis, so it is positive definite. Its inverse and the fields Z a , V 0 , V ˉ Z_a,V_0,\bar{V} Z a , V 0 , V ˉ depend smoothly on ( t , z ) (t,z) ( t , z ) , which makes the coefficients b a b^a b a smooth. We can therefore solve the ODE
z ˙ a = b a ( t , z ) , z ( 0 ) = 0 , (34) \dot{z}^a=b^a(t,z),\qquad z(0)=0, \tag*{(34)} z ˙ a = b a ( t , z ) , z ( 0 ) = 0 , (34)
by the finite-dimensional existence and uniqueness theorem. The resulting graph path has the required normal velocity, but its full velocity may also have a tangential component. Write it as V ˉ + d j ( Z ) \bar{V}+d j(Z) V ˉ + d j ( Z ) and solve ∂ t f t = − Z t ∘ f t \partial_t f_t=-Z_t\circ f_t ∂ t f t = − Z t ∘ f t with f 0 = id C f_0=\operatorname{id}_{C} f 0 = id C . The derivative of j t ∘ f t j_t\circ f_t j t ∘ f t is V ˉ ∘ f t \bar{V}\circ f_t V ˉ ∘ f t : the two terms containing d j ( Z ) d j(Z) d j ( Z ) cancel. Equivariance of B B B and of the minimum norm inverse ensures that this is the prescribed normal field for the reparametrized embedding.
Finally, any other nearby coassociative path with the same initial embedding and prescribed normal velocity has the same equation (34) for its graph parameter. Uniqueness of that ordinary differential equation gives the same moving image. The reparametrization equation then fixes the labeling of its points uniquely. This proves uniqueness of the parametrized path.
In the following proposition, we explain the geometric meaning of the selected velocity.
Proposition 3.10. Fix a closed G 2 G_2 G 2 structure φ \varphi φ and a compact coassociative embedding i i i . Let M φ \mathcal{M}_{\varphi} M φ be its local coassociative deformation space, with embeddings identified when they differ only by reparametrization. Give its tangent space T i M φ = ker D i T_i\mathcal{M}_{\varphi}=\ker D_i T i M φ = ker D i the L 2 L^2 L 2 metric on normal vector fields. If V φ \mathcal{V}_{\varphi} V φ is the volume of the embedded submanifold in the metric g φ g_{\varphi} g φ , then
grad M φ V φ = − Π i H . \operatorname{grad}_{\mathcal{M}_{\varphi}}\mathcal{V}_{\varphi}=-\Pi_i H. grad M φ V φ = − Π i H . Suppose now that the ambient form varies with ∂ t φ t = d τ t \partial_t\varphi_t=d\tau_t ∂ t φ t = d τ t , and set θ = i ∗ τ t \theta=i^*\tau_t θ = i ∗ τ t at a coassociative pair. The normal velocity
V = Π i H − R i d C θ = − R i d C θ − grad M φ t V φ t V=\Pi_iH-R_i d_C\theta=-R_i d_C\theta-\operatorname{grad}_{\mathcal{M}_{\varphi_t}}\mathcal{V}_{\varphi_t} V = Π i H − R i d C θ = − R i d C θ − grad M φ t V φ t is the unique velocity closest to H H H in L 2 L^2 L 2 among the normal velocities satisfying the differentiated coassociative equation. Thus it is the sum of the compatible velocity of minimum norm and the negative volume gradient within the current deformation space.
Proof. First keep the ambient structure fixed. Let i s i_s i s be a path of coassociatives through i i i whose normal velocity at s = 0 s=0 s = 0 is Z Z Z . Differentiating i s ∗ φ = 0 i_s^*\varphi=0 i s ∗ φ = 0 gives D i Z = 0 D_iZ=0 D i Z = 0 . Conversely, the deformation chart identifies every element of ker D i \ker D_i ker D i with a tangent vector to M φ \mathcal{M}_{\varphi} M φ . The first variation of volume gives
d V φ ∣ i ( Z ) = − ∫ C g φ ( H , Z ) d μ i = − ∫ C g φ ( Π i H , Z ) d μ i , d\mathcal{V}_{\varphi}|_i(Z)=-\int_C g_{\varphi}(H,Z)\,d\mu_i=-\int_C g_{\varphi}(\Pi_iH,Z)\,d\mu_i, d V φ ∣ i ( Z ) = − ∫ C g φ ( H , Z ) d μ i = − ∫ C g φ ( Π i H , Z ) d μ i ,
where the second equality follows because Z ∈ ker D i Z\in\ker D_i Z ∈ ker D i and Π i \Pi_i Π i is the orthogonal projection onto that space. This proves the gradient formula.
If the ambient G 2 G_2 G 2 structure is also moving then a compatible normal velocity Y Y Y must satisfy
0 = i ∗ ( ∂ t φ t ) + D i Y = d C θ + D i Y . 0=i^*(\partial_t\varphi_t)+D_iY=d_C\theta+D_iY. 0 = i ∗ ( ∂ t φ t ) + D i Y = d C θ + D i Y .
Since D i R i = I D_iR_i=I D i R i = I on exact three forms, every solution has the form
Y = − R i d C θ + Z , Z ∈ ker D i . Y=-R_i d_C\theta+Z,\qquad Z\in\ker D_i. Y = − R i d C θ + Z , Z ∈ ker D i .
The field R i d C θ R_i d_C\theta R i d C θ is orthogonal to ker D i \ker D_i ker D i . Therefore the orthogonal decomposition of H H H gives
∥ Y − H ∥ L 2 2 = ∥ Z − Π i H ∥ L 2 2 + ∥ R i d C θ + ( I − Π i ) H ∥ L 2 2 . \lVert Y-H\rVert_{L^2}^2=\lVert Z-\Pi_iH\rVert_{L^2}^2+\lVert R_i d_C\theta+(I-\Pi_i)H\rVert_{L^2}^2. ∥ Y − H ∥ L 2 2 = ∥ Z − Π i H ∥ L 2 2 + ∥ R i d C θ + ( I − Π i ) H ∥ L 2 2 .
The second term does not depend on Z Z Z . The first has its unique minimum at Z = Π i H Z=\Pi_iH Z = Π i H , which proves the assertion. The same decomposition gives
∥ Y ∥ L 2 2 = ∥ R i d C θ ∥ L 2 2 + ∥ Z ∥ L 2 2 , \lVert Y\rVert_{L^2}^2=\lVert R_i d_C\theta\rVert_{L^2}^2+\lVert Z\rVert_{L^2}^2, ∥ Y ∥ L 2 2 = ∥ R i d C θ ∥ L 2 2 + ∥ Z ∥ L 2 2 ,
so the compatible velocity of minimum norm is − R i d C θ -R_i d_C\theta − R i d C θ .
We now complete the proof of Theorem 1.1 .
Theorem 3.11. Let φ 0 \varphi_0 φ 0 be a smooth closed G 2 G_2 G 2 structure on compact M M M , and let i 0 : C ↪ M i_0 : C \hookrightarrow M i 0 : C ↪ M be a compact coassociative. Fix the tubular extension data in (2) . Then the coupled system (3) has a unique smooth short-time solution with this initial pair and initial parametrization, remaining in the neighborhood where the correction is defined. It preserves i t ∗ φ t = 0 i_t^*\varphi_t = 0 i t ∗ φ t = 0 and [ φ t ] = [ φ 0 ] [\varphi_t] = [\varphi_0] [ φ t ] = [ φ 0 ] .
Proof. We first use the short-time existence theorem for ordinary Laplacian flow [5 , 24 ] to obtain φ ˉ t \bar{\varphi}_t φ ˉ t from φ 0 \varphi_0 φ 0 . Set τ ˉ = δ φ ˉ φ ˉ \bar{\tau} = \delta_{\bar{\varphi}}\bar{\varphi} τ ˉ = δ φ ˉ φ ˉ . Its velocity is d τ ˉ d\bar{\tau} d τ ˉ , so Lemma 3.9 applies with B = H ˉ B = \bar{H} B = H ˉ . Since R ˉ D ˉ = I − Π ˉ \bar{R}\bar{D} = I - \bar{\Pi} R ˉ D ˉ = I − Π ˉ , the selected normal velocity, as described before, is
V ˉ = Π ˉ H ˉ − R ˉ d C θ ˉ = H ˉ + W ˉ . \bar{V} = \bar{\Pi}\bar{H} - \bar{R}d_C\bar{\theta} = \bar{H} + \bar{W}. V ˉ = Π ˉ H ˉ − R ˉ d C θ ˉ = H ˉ + W ˉ .
We next transfer the additional submanifold velocity W ˉ \bar{W} W ˉ into the ambient equation. Extend it to X ˉ = E φ ˉ , i ˉ W ˉ \bar{X} = E_{\bar{\varphi},\bar{i}}\bar{W} X ˉ = E φ ˉ , i ˉ W ˉ and solve ∂ t F t = X ˉ t ∘ F t \partial_t F_t = \bar{X}_t \circ F_t ∂ t F t = X ˉ t ∘ F t , F 0 = id M F_0 = \operatorname{id}_M F 0 = id M . Define
φ t = F t ∗ φ ˉ t , i t = F t − 1 ∘ i ˉ t . (35) \varphi_t = F_t^*\bar{\varphi}_t,\qquad i_t = F_t^{-1}\circ\bar{i}_t. \tag*{(35)} φ t = F t ∗ φ ˉ t , i t = F t − 1 ∘ i ˉ t . (35)
The normal operators and our fixed extension rule are equivariant under ambient diffeomorphisms, as shown in § 3.1 . Thus X φ , i = F t ∗ X ˉ X_{\varphi,i} = F_t^*\bar{X} X φ , i = F t ∗ X ˉ . Differentiating both expressions in (35) gives
∂ t φ = Δ φ φ + L X φ , ∂ t i = d F t − 1 ( H ˉ + W ˉ − X ˉ ∣ i ˉ ( C ) ) = H . \partial_t\varphi= \Delta_\varphi\varphi+ \mathcal{L}_{X}\varphi,\qquad\partial_t i = dF_t^{-1}\left(\bar{H}+\bar{W}-\bar{X}|_{\bar{i}(C)}\right)=H. ∂ t φ = Δ φ φ + L X φ , ∂ t i = d F t − 1 ( H ˉ + W ˉ − X ˉ ∣ i ˉ ( C ) ) = H .
For the second identity we used X ˉ ∣ i ˉ ( C ) = W ˉ \bar{X}|_{\bar{i}(C)} = \bar{W} X ˉ ∣ i ˉ ( C ) = W ˉ , which cancels the extra normal velocity. After shortening the interval, the pair remains in the neighborhood where both the inverse and the fixed tubular extension are defined. This gives the required solution. Closedness of the G 2 G_2 G 2 structures and the coassociative equation follow directly from pullback, and the cohomology statement also follows from the exact ambient velocity.
To prove uniqueness, take another solution with the same initial data. The preservation identity already proved shows that it remains coassociative. Solve ∂ t q t = − X t ∘ q t \partial_t q_t = -X_t\circ q_t ∂ t q t = − X t ∘ q t and pull back by q t q_t q t . The ambient G 2 G_2 G 2 structure q t ∗ φ t q_t^*\varphi_t q t ∗ φ t then solves ordinary Laplacian flow, while q t − 1 ∘ i t q_t^{-1}\circ i_t q t − 1 ∘ i t has normal velocity H ˉ + W ˉ \bar{H}+\bar{W} H ˉ + W ˉ . Uniqueness of ordinary Laplacian flow fixes the ambient path, and Lemma 3.9 fixes the lifted embedding. Thus the barred pair agrees with the one constructed above. Its reconstruction map q t − 1 q_t^{-1} q t − 1 satisfies ∂ t F t = X ˉ t ∘ F t \partial_tF_t=\bar{X}_t\circ F_t ∂ t F t = X ˉ t ∘ F t with the same initial value. Uniqueness of this last ordinary differential equation recovers the same corrected pair. □
Existence and uniqueness in Corollary 3.7. Apply Lemma 3.9 with B = 0 B=0 B = 0 . □
Choice of extension and continuation
The minimum norm condition determines the normal field W W W uniquely. Its extension away from the submanifold is an additional choice. We now show that changing this extension changes the ambient coordinate description, while leaving the ordinary-flow lift unchanged.
Proposition 3.12. Choose two smooth diffeomorphism-equivariant tubular extensions, both equal to W W W along the submanifold. The resulting solutions with the same coassociative initial pair are equivalent by time-dependent ambient diffeomorphisms for as long as both solutions are defined. In the description with ordinary Laplacian flow they give the same normal-velocity embedding j t j_t j t .
Proof. After removing the ambient Lie derivative, both solutions have the same ordinary ambient flow and the same prescribed normal velocity. The uniqueness argument in Theorem 3.11 therefore gives the same lifted pair ( φ ~ t , j t ) (\tilde{\varphi}_t,j_t) ( φ ~ t , j t ) . Let G t 1 , G t 2 G_t^1,G_t^2 G t 1 , G t 2 be the two maps that reconstruct the corrected solutions from this pair. Setting F t = ( G t 1 ) − 1 ∘ G t 2 F_t=(G_t^1)^{-1}\circ G_t^2 F t = ( G t 1 ) − 1 ∘ G t 2 and using the reconstruction formulas gives
F t ∗ φ t 1 = φ t 2 , F t − 1 ∘ i t 1 = i t 2 , F_t^*\varphi_t^1=\varphi_t^2,\qquad F_t^{-1}\circ i_t^1=i_t^2, F t ∗ φ t 1 = φ t 2 , F t − 1 ∘ i t 1 = i t 2 ,
and hence we get the required equivalence of the ambient structure. □
Proposition 3.13. Suppose an ordinary Laplacian flow φ t \varphi_t φ t extends smoothly across a finite time T T T . A solution j t j_t j t of (25) extends across T T T if its images remain in a compact subset of the space of smooth embedded coassociatives, measured in a fixed sufficiently high finite regularity. More explicitly, it is sufficient that, after reparametrization, the embeddings have uniform C m + 2 , α C^{m+2,\alpha} C m + 2 , α bounds for some m ≥ 4 m\ge4 m ≥ 4 , their induced metrics are uniformly equivalent to a fixed metric on C C C , and their normal tubular radii have a positive lower bound.
Proof. The finite-order bounds give a subsequence of images converging in C m + 1 , β C^{m+1,\beta} C m + 1 , β for 0 < β < α 0 < \beta< \alpha 0 < β < α . The metric lower bound prevents loss of immersion, and the positive lower bound for the tubular radius prevents different parts of the image from meeting in the limit. The limit is therefore an embedded image. Passing to the limit in the pullback equation shows that it is coassociative for φ ~ T \widetilde{\varphi}_T φ T . In a normal graph chart, elliptic regularity for the coassociative equation improves this limit to a smooth image, by the same argument as in Lemma 3.9 .
At every limiting coassociative, the explicit formula in Lemma 3.3 gives a right inverse for D i D_i D i . The compact set of possible limits can therefore be covered by finitely many deformation charts, with uniform inverse bounds after shrinking them. In each chart, the moving image is described by (34) , a smooth finite-dimensional equation. Its coefficients are bounded on a smaller chart, and the local existence time can be chosen uniformly over this finite collection. Starting sufficiently close to T T T therefore continues the image beyond T T T .
The tangential reparametrization fields have the corresponding bounds in these charts, so their ordinary differential equations also continue. This recovers the parametrized embedding j t j_t j t .
Remark 3.14. The argument in the previous proposition gives a sufficient continuation criterion. It does not claim that smoothness of the ambient flow alone prevents degeneration of the submanifold.
Long-time persistence and convergence
We now look at the long-time existence of the coupled system. For long-time existence, we need to keep the lifted embedding inside a region where the preceding construction remains valid. We do this by bounding how far its deformation parameter can travel. A small velocity at each time would not be enough, since a small constant velocity can accumulate a large displacement. We will require a small time integral of the relevant velocities.
Lemma 3.15. Fix a compact coassociative ( φ ˉ , ı ˉ ) (\bar{\varphi},\bar{\imath}) ( φ ˉ , ˉ ) and an integer m ≥ 4 m \ge4 m ≥ 4 . There is a neighborhood of φ ˉ \bar{\varphi} φ ˉ in C m + 3 , α C^{m+3,\alpha} C m + 3 , α such that the construction of Lemma 3.9 is uniform for closed forms φ \varphi φ in that neighborhood with [ ı ˉ ∗ φ ] = 0 [\bar{\imath}^{*}\varphi] = 0 [ ˉ ∗ φ ] = 0 . On a smaller deformation ball, the following estimates hold:
∣ z ˙ ∣ ≤ C ( ∥ φ ˙ ∥ C m + 3 , α + ∥ τ ∥ C m + 1 , α ) , ∥ V ∥ C m , α + ∥ W ∥ C m , α ≤ C ∥ τ ∥ C m + 1 , α , (36) |\dot{z}| \le C\left(\|\dot{\varphi}\|_{C^{m+3,\alpha}}+\|\tau\|_{C^{m+1,\alpha}}\right), \qquad\|V\|_{C^{m,\alpha}}+\|W\|_{C^{m,\alpha}} \le C\|\tau\|_{C^{m+1,\alpha}}, \tag*{(36)} ∣ z ˙ ∣ ≤ C ( ∥ φ ˙ ∥ C m + 3 , α + ∥ τ ∥ C m + 1 , α ) , ∥ V ∥ C m , α + ∥ W ∥ C m , α ≤ C ∥ τ ∥ C m + 1 , α , (36) with V V V and W W W the normal velocities as before.
Proof. We first obtain the graph coordinates in exactly the same way as in Lemma 3.9 . We can then estimate how the graphs move along the ambient motion by differentiating F ( φ , z , w ( φ , z ) ) = 0 \mathcal{F}(\varphi,z,w(\varphi,z))=0 F ( φ , z , w ( φ , z )) = 0 while holding z z z fixed. For an admissible ambient variation η \eta η , the direct variation of F \mathcal{F} F is ı ˉ ∗ η \bar{\imath}^{*}\eta ˉ ∗ η , so
D w F [ D φ w [ η ] ] = − ı ˉ ∗ η . D_w\mathcal{F}[D_\varphi w[\eta]] = -\bar{\imath}^{*}\eta. D w F [ D φ w [ η ]] = − ˉ ∗ η .
Applying the uniformly bounded inverse of the w w w derivative, and then the uniform estimate for the normal graph parametrization, gives
∥ D φ j [ η ] ∥ C m + 2 , α ≤ C ∥ η ∥ C m + 3 , α . (37) \|D_\varphi j[\eta]\|_{C^{m+2,\alpha}} \le C\|\eta\|_{C^{m+3,\alpha}}. \tag*{(37)} ∥ D φ j [ η ] ∥ C m + 2 , α ≤ C ∥ η ∥ C m + 3 , α . (37)
In particular the velocity V 0 = ( ∂ t j ) ⊥ V_0=(\partial_tj)^\perp V 0 = ( ∂ t j ) ⊥ , with z z z held fixed, satisfies this estimate with η = φ ˙ \eta=\dot{\varphi} η = φ ˙ . We next estimate V V V and W W W . In the smaller graph neighborhood, the induced metrics and the maps J i J_i J i , J i − 1 J_i^{-1} J i − 1 have uniform coefficient bounds. Lemma 3.4 gives a uniform low-order inverse bound for the Laplacian on exact three forms. The Schauder estimate upgrades this bound to the derivative orders needed here. Combining it with (20) , we obtain
∥ R i q ∥ C m , α ≤ C ∥ q ∥ C m − 1 , α . \|R_iq\|_{C^{m,\alpha}} \le C\|q\|_{C^{m-1,\alpha}}. ∥ R i q ∥ C m , α ≤ C ∥ q ∥ C m − 1 , α .
The formula J i H = − 2 π − θ J_iH=-2\pi_-\theta J i H = − 2 π − θ controls H H H directly by the torsion restriction. In W = − R i d C ( ∗ C θ ) W=-R_id_C(*_C\theta) W = − R i d C ( ∗ C θ ) , the derivative taken by d C d_C d C is recovered by the right inverse. Thus
∥ H ∥ C m , α + ∥ W ∥ C m , α ≤ C ∥ θ ∥ C m , α ≤ C ∥ τ ∥ C m + 1 , α . \begin{aligned}
\|H\|_{C^{m,\alpha}}+\|W\|_{C^{m,\alpha}} \le C\|\theta\|_{C^{m,\alpha}} \\
&\le C\|\tau\|_{C^{m+1,\alpha}}.
\end{aligned} ∥ H ∥ C m , α + ∥ W ∥ C m , α ≤ C ∥ θ ∥ C m , α ≤ C ∥ τ ∥ C m + 1 , α .
The same bound holds for V = H + W V=H+W V = H + W and we get the second estimates in (36) .
To estimate the deformation parameter, recall that Z a = ( ∂ z a j ) ⊥ Z_a=(\partial_{z_a}j)^\perp Z a = ( ∂ z a j ) ⊥ is a basis of ker D i \ker D_i ker D i . Its Gram matrix G a b = ⟨ Z a , Z b ⟩ L 2 ( C ) G_{ab}=\langle Z_a,Z_b\rangle_{L^2(C)} G ab = ⟨ Z a , Z b ⟩ L 2 ( C ) is positive definite at the reference pair. After shrinking the graph neighborhood, its inverse is uniformly bounded. Pairing V − V 0 = ∑ a z ˙ a Z a V-V_0=\sum_a \dot{z}^a Z_a V − V 0 = ∑ a z ˙ a Z a with the basis gives the explicit equation
z ˙ a = ∑ b ( G − 1 ) a b ⟨ V − V 0 , Z b ⟩ L 2 ( C ) . (38) \dot{z}^a=\sum_b(G^{-1})^{ab}\langle V-V_0,Z_b\rangle_{L^2(C)}. \tag*{(38)} z ˙ a = b ∑ ( G − 1 ) ab ⟨ V − V 0 , Z b ⟩ L 2 ( C ) . (38)
Combining this formula with (37) and the bound for V V V proves the first estimate of (36) .
We can now prove the following.
Theorem 3.16 (An integrable-velocity criterion). Let φ ˉ \bar{\varphi} φ ˉ be torsion-free on compact M M M , and let ı ˉ \bar{\imath} ˉ be a compact coassociative embedding. Fix the tubular extension data with radius smaller than the tubular radius of ı ˉ \bar{\imath} ˉ . There are m ≥ 4 m\geq4 m ≥ 4 and ϵ > 0 \epsilon>0 ϵ > 0 with the following property. Let φ t \varphi_t φ t , t ≥ 0 t\geq0 t ≥ 0 , be a smooth ordinary Laplacian flow such that
sup t ≥ 0 ∥ φ t − φ ˉ ∥ C m + 3 , α < ϵ , [ ı ˉ ∗ φ t ] = 0 , (39) \sup_{t\geq0}\|\varphi_t-\bar{\varphi}\|_{C^{m+3,\alpha}}<\epsilon,\qquad[\bar{\imath}^{*}\varphi_t]=0, \tag*{(39)} t ≥ 0 sup ∥ φ t − φ ˉ ∥ C m + 3 , α < ϵ , [ ˉ ∗ φ t ] = 0 , (39) and
∫ 0 ∞ ( ∥ ∂ t φ t ∥ C m + 3 , α + ∥ τ t ∥ C m + 1 , α ) d t < ϵ . (40) \int_0^\infty\left(\|\partial_t\varphi_t\|_{C^{m+3,\alpha}}+\|\tau_t\|_{C^{m+1,\alpha}}\right)\,dt<\epsilon. \tag*{(40)} ∫ 0 ∞ ( ∥ ∂ t φ t ∥ C m + 3 , α + ∥ τ t ∥ C m + 1 , α ) d t < ϵ . (40) Assume also that the ambient forms are uniformly bounded in every higher norm and that the corresponding higher-norm integrals are finite. For every sufficiently close φ 0 \varphi_0 φ 0 -coassociative i 0 i_0 i 0 , the lift (25) and the reconstructed corrected system exist for all time. Both converge smoothly to coassociative pairs with torsion-free ambient structure. If the ambient velocities and torsion decay exponentially in every norm, the convergence of the pairs and the reconstruction diffeomorphisms is exponential in every norm.
Proof. Choose the uniform chart from Lemma 3.15 with a parameter ball of radius 2 r 2r 2 r , and take the initial graph parameter in the concentric ball of radius r 2 \frac{r}{2} 2 r . On the interval on which the lift stays in this chart we integrate the bound for z ˙ \dot{z} z ˙ in (36) to obtain
∣ z ( t ) − z ( 0 ) ∣ ≤ C ∫ 0 t ( ∥ ∂ s φ s ∥ C m + 3 , α + ∥ τ s ∥ C m + 1 , α ) d s . |z(t)-z(0)|\leq C\int_0^t\left(\|\partial_s\varphi_s\|_{C^{m+3,\alpha}}+\|\tau_s\|_{C^{m+1,\alpha}}\right)\,ds. ∣ z ( t ) − z ( 0 ) ∣ ≤ C ∫ 0 t ( ∥ ∂ s φ s ∥ C m + 3 , α + ∥ τ s ∥ C m + 1 , α ) d s .
We choose ϵ > 0 \epsilon>0 ϵ > 0 so that the bound on the right is less than r 2 \frac{r}{2} 2 r . The parameter then remains within distance r r r of the center. If there were a first exit from this smaller ball, the same estimate up to the exit time would give a contradiction. The finite-dimensional equation therefore stays in a fixed interior region and can be continued for all time. This is where smallness of the total time integral (40) is used.
By Lemma 3.15 , the normal velocities V t V_t V t are integrable in every norm. Integrating from one time to a later time shows that the embeddings are Cauchy in each of these norms. The limit i ∞ i_\infty i ∞ remains embedded, because the low-order smallness keeps the images in a graph neighborhood with a positive tubular radius.
The same estimates give a small value of ∫ 0 ∞ ∥ X ~ t ∥ C 1 d t \int_0^\infty\|\widetilde{X}_t\|_{C^1}\,dt ∫ 0 ∞ ∥ X t ∥ C 1 d t and finite integrals at every higher order. Hence the reconstruction maps F t F_t F t exist globally and converge. We now check that their limit is actually a diffeomorphism, rather than a map with a degenerate derivative. We differentiate their equation
∂ t ( d F t ) = ( ∇ X ~ t ) ∘ F t d F t . \partial_t(dF_t)=(\nabla\widetilde{X}_t)\circ F_t\,dF_t. ∂ t ( d F t ) = ( ∇ X t ) ∘ F t d F t .
Thus, we can again use Gronwall’s inequality to obtain uniform bounds for d F t dF_t d F t and its inverse. The right-hand side is then integrable, so d F t dF_t d F t is Cauchy. The maps F t F_t F t themselves are Cauchy by the integrability of X ~ t \widetilde{X}_t X t . Repeated differentiation gives the same conclusion at higher orders. Thus F t F_t F t converges smoothly to a diffeomorphism F ∞ F_\infty F ∞ . Taking the low-order integral sufficiently small also keeps the reconstructed pair inside the domain of the chosen fixed extension.
The time integral of ∂ t φ t \partial_t\varphi_t ∂ t φ t , together with its higher order versions, gives a smooth limit φ ∞ \varphi_\infty φ ∞ . It is a G 2 G_2 G 2 structure because the 3-forms stay in the chosen small neighborhood of φ ˉ \bar{\varphi} φ ˉ (see [24 Thm. 5.1] and [9 Thm. 5.1] for more details on this argument). Since the torsion τ t \tau_t τ t depends smoothly on the form and its first derivative, so τ t \tau_t τ t also converges. If its limit were nonzero, its norm would be bounded below by a positive constant for all sufficiently large times, contradicting its finite time integral. Hence τ ∞ = 0 \tau_{\infty}=0 τ ∞ = 0 . Passing to the limit in j t ∗ φ t = 0 j^{*}_{t}\varphi_{t}=0 j t ∗ φ t = 0 gives a coassociative limit, and (35) gives the limiting corrected pair.
We can also estimate the distances to these limits. At any fixed order k k k , the construction uses only finitely many derivatives of the ambient G 2 G_{2} G 2 structure. Integrating the corresponding velocity estimates from t t t to infinity therefore gives an integer r k r_{k} r k for which
∥ φ t − φ ∞ ∥ C k + ∥ j t − j ∞ ∥ C k + ∥ F t − F ∞ ∥ C k ≤ C k ∫ t ∞ ( ∥ ∂ q φ q ∥ C r k + ∥ τ q ∥ C r k ) d q . (41) \|\varphi_{t}-\varphi_{\infty}\|_{C^{k}}+\|j_{t}-j_{\infty}\|_{C^{k}}+\|F_{t}-F_{\infty}\|_{C^{k}}\leq C_{k}\int_{t}^{\infty}\left(\|\partial_{q}\varphi_{q}\|_{C^{r_{k}}}+\|\tau_{q}\|_{C^{r_{k}}}\right)\,dq. \tag*{(41)} ∥ φ t − φ ∞ ∥ C k + ∥ j t − j ∞ ∥ C k + ∥ F t − F ∞ ∥ C k ≤ C k ∫ t ∞ ( ∥ ∂ q φ q ∥ C r k + ∥ τ q ∥ C r k ) d q . (41)
If the integrand is bounded by C e − λ q Ce^{-\lambda q} C e − λ q , its tail is bounded by ( C λ ) e − λ t \left(\frac{C}{\lambda}\right)e^{-\lambda t} ( λ C ) e − λ t . Thus the lift and the reconstruction maps converge exponentially. Finally, composing the limits in (35) gives the same conclusion for the corrected pair.
Lemma 3.17. Let φ ˉ \bar{\varphi} φ ˉ be a torsion-free G 2 G_{2} G 2 structure on a compact manifold M M M . Fix an integer r ≥ 0 r\geq0 r ≥ 0 and 0 < α < 1 0<\alpha<1 0 < α < 1 , and put k = max { 5 , r + 3 } k=\max\{5,r+3\} k = max { 5 , r + 3 } and N = k + 4 N=k+4 N = k + 4 . Then there are constants δ , C , σ > 0 \delta,C,\sigma>0 δ , C , σ > 0 with the following property. Suppose that φ 0 \varphi_{0} φ 0 is smooth, closed and positive, that [ φ 0 ] = [ φ ˉ ] [\varphi_{0}]=[\bar{\varphi}] [ φ 0 ] = [ φ ˉ ] , and that ∥ φ 0 − φ ˉ ∥ C N , α < δ \|\varphi_{0}-\bar{\varphi}\|_{C^{N,\alpha}}<\delta ∥ φ 0 − φ ˉ ∥ C N , α < δ . Then the ordinary Laplacian flow from φ 0 \varphi_{0} φ 0 exists for all time and satisfies
sup t ≥ 0 ∥ φ t − φ ˉ ∥ C r , α + ∫ 0 ∞ ( ∥ ∂ t φ t ∥ C r , α + ∥ τ t ∥ C r , α ) d t ≤ C ∥ φ 0 − φ ˉ ∥ C N , α . (42) \sup_{t\geq0}\|\varphi_{t}-\bar{\varphi}\|_{C^{r,\alpha}}+\int_{0}^{\infty}\left(\|\partial_{t}\varphi_{t}\|_{C^{r,\alpha}}+\|\tau_{t}\|_{C^{r,\alpha}}\right)\,dt\leq C\|\varphi_{0}-\bar{\varphi}\|_{C^{N,\alpha}}. \tag*{(42)} t ≥ 0 sup ∥ φ t − φ ˉ ∥ C r , α + ∫ 0 ∞ ( ∥ ∂ t φ t ∥ C r , α + ∥ τ t ∥ C r , α ) d t ≤ C ∥ φ 0 − φ ˉ ∥ C N , α . (42) In fact,
∥ ∂ t φ t ∥ C r , α + ∥ τ t ∥ C r , α ≤ C e − σ t ∥ φ 0 − φ ˉ ∥ C N , α . \|\partial_{t}\varphi_{t}\|_{C^{r,\alpha}}+\|\tau_{t}\|_{C^{r,\alpha}}\leq Ce^{-\sigma t}\|\varphi_{0}-\bar{\varphi}\|_{C^{N,\alpha}}. ∥ ∂ t φ t ∥ C r , α + ∥ τ t ∥ C r , α ≤ C e − σ t ∥ φ 0 − φ ˉ ∥ C N , α . The forms φ t \varphi_{t} φ t are uniformly bounded in every higher spatial norm. For each integer j ≥ 0 j\geq0 j ≥ 0 , there is a finite constant C j C_{j} C j such that
∥ ∂ t φ t ∥ C j , α + ∥ τ t ∥ C j , α ≤ C j e − σ t / 2 . \|\partial_{t}\varphi_{t}\|_{C^{j,\alpha}}+\|\tau_{t}\|_{C^{j,\alpha}}\leq C_{j}e^{-\sigma t/2}. ∥ ∂ t φ t ∥ C j , α + ∥ τ t ∥ C j , α ≤ C j e − σ t /2 . These higher-order constants may depend on higher norms of the smooth initial form. The flow converges smoothly and exponentially to a torsion-free form in Diff 0 ( M ) ⋅ φ ˉ \operatorname{Diff}_{0}(M)\cdot\bar{\varphi} Diff 0 ( M ) ⋅ φ ˉ .
Proof. We first use the Laplacian–DeTurck flow of [LW19b] and write its solution as φ ~ t = φ ˉ + η t \tilde{\varphi}_{t}=\bar{\varphi}+\eta_{t} φ ~ t = φ ˉ + η t . The form η t \eta_{t} η t is exact because the cohomology class is fixed. Fix k = max { 5 , r + 3 } k=\max\{5,r+3\} k = max { 5 , r + 3 } and N = k + 4 N=k+4 N = k + 4 , as in the statement. All norms in the next estimates use the fixed reference metric.
Choose the neighbourhood size ε \varepsilon ε in [LW19b, Theorem 5.1], and let C 1 C_{1} C 1 be the constant in their [LW19b, equation (4.45)]. That theorem requires
∥ η 0 ∥ C k + 4 < δ L W , ∥ η 0 ∥ W k + 4 , 2 ≤ ε 2 C 1 . \|\eta_{0}\|_{C^{k+4}}<\delta_{\mathrm{LW}},\qquad\|\eta_{0}\|_{W^{k+4,2}}\leq\frac{\varepsilon}{2C_{1}}. ∥ η 0 ∥ C k + 4 < δ LW , ∥ η 0 ∥ W k + 4 , 2 ≤ 2 C 1 ε .
Here δ L W \delta_{\mathrm{LW}} δ LW depends on the fixed background, k k k , and ε \varepsilon ε . On the compact reference manifold, both initial norms are bounded by C ∥ η 0 ∥ C N , α C\|\eta_{0}\|_{C^{N,\alpha}} C ∥ η 0 ∥ C N , α . Reducing our δ \delta δ therefore ensures both hypotheses and gives a global DeTurck solution in that fixed neighbourhood.
We obtain from [LW19b, eq. (4.45)]
sup t ≥ 0 ∥ η t ∥ C k ≤ C 1 ∥ η 0 ∥ W k + 4 , 2 ≤ C ∥ η 0 ∥ C N , α . \sup_{t\geq0}\|\eta_{t}\|_{C^{k}}\leq C_{1}\|\eta_{0}\|_{W^{k+4,2}}\leq C\|\eta_{0}\|_{C^{N,\alpha}}. t ≥ 0 sup ∥ η t ∥ C k ≤ C 1 ∥ η 0 ∥ W k + 4 , 2 ≤ C ∥ η 0 ∥ C N , α .
The time-cutoff argument in the proof of [LW19b, Theorem 5.1], ending in their [LW19b, eq. (5.6)], gives, after a fixed time t 0 = t 0 ( k ) > 0 t_{0}=t_{0}(k)>0 t 0 = t 0 ( k ) > 0 ,
∥ ∇ ˉ ℓ η t ∥ L 2 2 ≤ C e − λ t / 2 ∥ η 0 ∥ L 2 2 , 0 ≤ ℓ ≤ k + 4 , \|\bar{\nabla}^{\ell}\eta_{t}\|_{L^{2}}^{2}\leq Ce^{-\lambda t/2}\|\eta_{0}\|_{L^{2}}^{2},\qquad0\leq\ell\leq k+4, ∥ ∇ ˉ ℓ η t ∥ L 2 2 ≤ C e − λ t /2 ∥ η 0 ∥ L 2 2 , 0 ≤ ℓ ≤ k + 4 ,
where λ 1 > 0 \lambda_{1} > 0 λ 1 > 0 is the reference spectral gap used there. Their derivative condition is ⌊ ℓ / 2 ⌋ ≤ k − 1 \lfloor\ell/2 \rfloor\le k - 1 ⌊ ℓ /2 ⌋ ≤ k − 1 and it includes ℓ ≤ k + 4 \ell\le k + 4 ℓ ≤ k + 4 because k ≥ 5 k \ge5 k ≥ 5 . The constants at these fixed orders depend on the background, k k k , and the chosen neighbourhood, and not on higher initial derivatives. Summing these inequalities, taking square roots, and using W k + 4 , 2 ↪ C k W^{k+4,2} \hookrightarrow C^k W k + 4 , 2 ↪ C k in dimension seven gives
∥ η t ∥ C k ≤ C e − λ 1 t / 4 ∥ η 0 ∥ C N , α , t ≥ t 0 . \|\eta_t\|_{C^k} \le Ce^{-\lambda_1 t/4}\|\eta_0\|_{C^{N,\alpha}}, \qquad t \ge t_0. ∥ η t ∥ C k ≤ C e − λ 1 t /4 ∥ η 0 ∥ C N , α , t ≥ t 0 .
On 0 ≤ t ≤ t 0 0 \le t \le t_0 0 ≤ t ≤ t 0 , the uniform estimate above gives the same bound after increasing C C C by a fixed factor. Thus the estimate holds for all t ≥ 0 t \ge0 t ≥ 0 with σ = λ 1 4 \sigma= \frac{\lambda_1}{4} σ = 4 λ 1 . In particular, the DeTurck limit is φ ˉ \bar{\varphi} φ ˉ itself.
Let V t V_t V t be the DeTurck vector field. It depends smoothly on the form and its first derivative, and it vanishes at φ ˉ \bar{\varphi} φ ˉ . Consequently,
∥ V t ∥ C k − 1 ≤ C ∥ η t ∥ C k ≤ C e − σ t ∥ η 0 ∥ C N , α . \|V_t\|_{C^{k-1}} \le C\|\eta_t\|_{C^k} \le Ce^{-\sigma t}\|\eta_0\|_{C^{N,\alpha}}. ∥ V t ∥ C k − 1 ≤ C ∥ η t ∥ C k ≤ C e − σ t ∥ η 0 ∥ C N , α .
Let F t F_t F t solve ∂ t F t = − V t ∘ F t \partial_t F_t=-V_t\circ F_t ∂ t F t = − V t ∘ F t , with F 0 F_0 F 0 the identity. Computations similar to the ones in the proof of Theorem 3.16 gives
sup t ≥ 0 ∥ F t − id ∥ C k − 1 ≤ C ∫ 0 ∞ ∥ V t ∥ C k − 1 d t ≤ C ∥ η 0 ∥ C N , α . \sup_{t\ge0}\|F_t-\operatorname{id}\|_{C^{k-1}} \le C\int_0^\infty\|V_t\|_{C^{k-1}}\,dt \le C\|\eta_0\|_{C^{N,\alpha}}. t ≥ 0 sup ∥ F t − id ∥ C k − 1 ≤ C ∫ 0 ∞ ∥ V t ∥ C k − 1 d t ≤ C ∥ η 0 ∥ C N , α .
Here differences of maps are measured in fixed charts near the identity. These estimates give uniform bounds for pullback in the norms used below.
The ordinary solution is φ t = F t ∗ φ ~ t \varphi_t=F_t^*\widetilde{\varphi}_t φ t = F t ∗ φ t . Naturality of the Laplacian and of the torsion form gives
∂ t φ t = F t ∗ ( Δ φ ~ t φ ~ t ) , τ t = F t ∗ τ ( φ ~ t ) . \partial_t\varphi_t=F_t^*(\Delta_{\widetilde{\varphi}_t}\widetilde{\varphi}_t), \qquad\tau_t=F_t^*\tau(\widetilde{\varphi}_t). ∂ t φ t = F t ∗ ( Δ φ t φ t ) , τ t = F t ∗ τ ( φ t ) .
The first expression inside the pullback is second order in η t \eta_t η t , and the second is first order and they both vanish when η t = 0 \eta_t=0 η t = 0 . On the fixed small neighbourhood,
∥ Δ φ ~ t φ ~ t ∥ C r , α + ∥ τ ( φ ~ t ) ∥ C r , α ≤ C ∥ η t ∥ C r + 2 , α ≤ C ∥ η t ∥ C k . \|\Delta_{\widetilde{\varphi}_t}\widetilde{\varphi}_t\|_{C^{r,\alpha}}+\|\tau(\widetilde{\varphi}_t)\|_{C^{r,\alpha}} \le C\|\eta_t\|_{C^{r+2,\alpha}} \le C\|\eta_t\|_{C^k}. ∥ Δ φ t φ t ∥ C r , α + ∥ τ ( φ t ) ∥ C r , α ≤ C ∥ η t ∥ C r + 2 , α ≤ C ∥ η t ∥ C k .
The last inequality follows from k ≥ r + 3 k\ge r+3 k ≥ r + 3 . Pullback of these forms uses C r + 1 , α C^{r+1,\alpha} C r + 1 , α bounds for F t F_t F t , which are supplied by the C k − 1 C^{k-1} C k − 1 bounds because k − 1 ≥ r + 2 k-1\ge r+2 k − 1 ≥ r + 2 . We obtain the stated exponential estimate for ∂ t φ t \partial_t\varphi_t ∂ t φ t and τ t \tau_t τ t . Furthermore,
∥ φ t − φ ˉ ∥ C r , α ≤ ∥ F t ∗ η t ∥ C r , α + ∥ F t ∗ φ ˉ − φ ˉ ∥ C r , α ≤ C ( ∥ η t ∥ C k + ∥ F t − id ∥ C k − 1 ) ≤ C ∥ η 0 ∥ C N , α . \begin{aligned}
\|\varphi_t-\bar{\varphi}\|_{C^{r,\alpha}} &\le\|F_t^*\eta_t\|_{C^{r,\alpha}}+\|F_t^*\bar{\varphi}-\bar{\varphi}\|_{C^{r,\alpha}} \\
&\le C(\|\eta_t\|_{C^k}+\|F_t-\operatorname{id}\|_{C^{k-1}})\le C\|\eta_0\|_{C^{N,\alpha}}.
\end{aligned} ∥ φ t − φ ˉ ∥ C r , α ≤ ∥ F t ∗ η t ∥ C r , α + ∥ F t ∗ φ ˉ − φ ˉ ∥ C r , α ≤ C ( ∥ η t ∥ C k + ∥ F t − id ∥ C k − 1 ) ≤ C ∥ η 0 ∥ C N , α .
Integration of the velocity and torsion bounds proves (42) .
It remains to explain the higher-order assertion without assuming smallness of every initial derivative. Theorem 1.3 of [LW19b], whose smooth-convergence conclusion is proved after Theorem 5.1, gives smooth convergence of this ordinary solution from the finite-order neighbourhood just used. Hence φ t \varphi_t φ t has uniform bounds in every fixed spatial norm. Since its velocity is a smooth second-order expression in φ t \varphi_t φ t and its torsion is a smooth first-order expression, both also have uniform bounds in every norm. These bounds are finite on the initial time interval because the initial form is smooth.
We combine these bounds with the exponential decay already proved in C 0 C^0 C 0 . For a tensor u u u and any integer L > j + α L>j+\alpha L > j + α , the interpolation inequality in the fixed reference metric gives
∥ u ∥ C j , α ≤ C ∥ u ∥ C 0 1 − ( j + α ) / L ∥ u ∥ C L ( j + α ) / L . \|u\|_{C^{j,\alpha}}\le C\|u\|_{C^0}^{1-(j+\alpha)/L}\|u\|_{C^L}^{(j+\alpha)/L}. ∥ u ∥ C j , α ≤ C ∥ u ∥ C 0 1 − ( j + α ) / L ∥ u ∥ C L ( j + α ) / L .
Take L = 2 j + 2 L=2j+2 L = 2 j + 2 and apply this to u = ∂ t φ t u=\partial_t\varphi_t u = ∂ t φ t and u = τ t u=\tau_t u = τ t . The exponent of the decaying C 0 C^0 C 0 norm is greater than 1 / 2 1/2 1/2 , so both tensors decay at least as fast as C j e − σ t / 2 C_j e^{-\sigma t/2} C j e − σ t /2 in C j , α C^{j,\alpha} C j , α . Their time integrals are therefore finite at every order. Integrating the velocity from t t t to infinity proves exponential convergence of φ t \varphi_t φ t in every norm. The limit is torsion-free and belongs to Diff 0 ( M ) ⋅ φ ˉ \operatorname{Diff}_0(M)\cdot\bar{\varphi} Diff 0 ( M ) ⋅ φ ˉ . The ordinary limit can differ from φ ˉ \bar{\varphi} φ ˉ , which is why the first term in (42) is a uniform smallness bound rather than a decay estimate towards φ ˉ \bar{\varphi} φ ˉ .
We can now give a proof of another of our main theorems.
Proof of Theorem 1.3 . Let m m m and ϵ \epsilon ϵ be supplied by Theorem 3.16 , with the fixed tubular extension data in that theorem. Apply Lemma 3.17 with r = m + 3 r=m+3 r = m + 3 so its finite initial regularity can be taken to be N = m + 10 N=m+10 N = m + 10 . We reduce the initial neighbourhood so that the right-hand side of (42) is smaller than ϵ \epsilon ϵ , after including the fixed norm-comparison constants. Then
sup t ≥ 0 ∥ φ t − φ ˉ ∥ C m + 3 , α < ϵ , \sup_{t\geq0}\|\varphi_t-\bar{\varphi}\|_{C^{m+3,\alpha}}<\epsilon, t ≥ 0 sup ∥ φ t − φ ˉ ∥ C m + 3 , α < ϵ ,
∫ 0 ∞ ( ∥ ∂ t φ t ∥ C m + 3 , α + ∥ τ t ∥ C m + 1 , α ) d t < ϵ . \int_0^\infty\left(\|\partial_t\varphi_t\|_{C^{m+3,\alpha}}+\|\tau_t\|_{C^{m+1,\alpha}}\right)\,dt<\epsilon. ∫ 0 ∞ ( ∥ ∂ t φ t ∥ C m + 3 , α + ∥ τ t ∥ C m + 1 , α ) d t < ϵ .
The second estimate uses the continuous inclusion C m + 3 , α ⊂ C m + 1 , α C^{m+3,\alpha}\subset C^{m+1,\alpha} C m + 3 , α ⊂ C m + 1 , α for the torsion term. Since the Laplacian flow preserves the cohomology class and ı ˉ \bar{\imath} ˉ is φ ˉ \bar{\varphi} φ ˉ -coassociative, we also have
[ ı ˉ ∗ φ t ] = ı ˉ ∗ [ φ ˉ ] = 0. [\bar{\imath}^{*}\varphi_t]=\bar{\imath}^{*}[\bar{\varphi}]=0. [ ˉ ∗ φ t ] = ˉ ∗ [ φ ˉ ] = 0.
These are exactly (39) –(40) . The higher-order bounds and integrals required by Theorem 3.16 follow from the same lemma. Taking the initial embedding in the neighbourhood specified in that theorem therefore gives the global lift and the global corrected solution. The exponential bounds in every spatial norm give their smooth exponential convergence.
The ordinary ambient limit belongs to Diff 0 ( M ) ⋅ φ ˉ \operatorname{Diff}_{0}(M)\cdot\bar{\varphi} Diff 0 ( M ) ⋅ φ ˉ by Lemma 3.17. The reconstruction diffeomorphisms supplied by Theorem 3.16 start at the identity and converge smoothly to a diffeomorphism. Their limit is also isotopic to the identity as it is C 1 C^{1} C 1 -close to one of these diffeomorphisms for sufficiently large time, and sufficiently close diffeomorphisms are isotopic. Thus the corrected ambient limit is again in Diff 0 ( M ) ⋅ φ ˉ \operatorname{Diff}_{0}(M)\cdot\bar{\varphi} Diff 0 ( M ) ⋅ φ ˉ . Finally, the short-time uniqueness in Theorem 3.11 applies at every time and gives uniqueness on the whole interval.
Remark 3.18. Theorem 1.3 concerns initially coassociative pairs. For the correction used in this paper, the pullback ı t ∗ φ t \imath_t^{*}\varphi_t t ∗ φ t is constant even when it is initially nonzero. Thus this equation does not make arbitrary nearby embeddings coassociative. The convergence assertion above also allows the neutral directions coming from harmonic anti-self-dual two forms but it does not assert convergence to the initial embedding.
The second coupled system: instantons
We now study our second coupled system of flows in G 2 G_{2} G 2 geometry in this section. We are interested in the Yang–Mills flow of connections on a principal G G G -bundle over a manifold M M M with a G 2 G_{2} G 2 structure. We will proceed in the same way as in § 3 and along the way, we will figure out the correct flow of coclosed G 2 G_{2} G 2 structures which preserves the condition of a family of connections being a G 2 G_{2} G 2 instanton. More on gauge theory on manifolds with G 2 G_{2} G 2 structures can be found in [38 Wal17] .
The correction and its energy identity
We now choose the connection velocity so that an instanton remains an instanton when the ambient four form ψ \psi ψ evolves along a geometric flow. We first calculate the error caused by the ambient motion. We then construct an elliptic correction and prove an energy identity. As was the case with coassociatives in § 3 the existence of the resulting flow is a separate question which is addressed in § 4.2 .
Throughout this section, d ψ = 0 d\psi=0 d ψ = 0 . All adjoints, Hodge stars and inner products use g ψ g_{\psi} g ψ which is the metric which is obtained from g φ g_{\varphi} g φ , φ = ∗ ψ ψ \varphi=*_{\psi}\psi φ = ∗ ψ ψ and the fixed fibre metric on E = ad P E=\operatorname{ad}P E = ad P . A difference of connections, and hence a connection velocity, is an E E E valued one form. Our curvature convention is F A = d A + 1 2 [ A ∧ A ] F_A=dA+\frac{1}{2}[A\wedge A] F A = d A + 2 1 [ A ∧ A ] . The bracket combines the wedge product with the Lie bracket on g \mathfrak{g} g . Thus ∂ t F A = d A ∂ t A \partial_tF_A=d_A\partial_tA ∂ t F A = d A ∂ t A and d A 2 = [ F A , ⋅ ] d_A^2=[F_A,\cdot] d A 2 = [ F A , ⋅ ] where d A d_A d A is the exterior covariant derivative. We will also use the Bianchi identity d A F A = 0 d_AF_A=0 d A F A = 0 .
With our convention, a gauge transformation acts by k ∗ A = k − 1 A k + k − 1 d k k^{*}A=k^{-1}Ak+k^{-1}dk k ∗ A = k − 1 A k + k − 1 d k , and its infinitesimal action is d A q d_Aq d A q for q ∈ Ω 0 ( E ) q\in\Omega^{0}(E) q ∈ Ω 0 ( E ) . The Coulomb condition d A ∗ a = 0 d_A^{*}a=0 d A ∗ a = 0 therefore means that a a a is L 2 L^{2} L 2 orthogonal to all infinitesimal gauge directions: ⟨ a , d A q ⟩ L 2 = ⟨ d A ∗ a , q ⟩ L 2 = 0 \langle a,d_Aq\rangle_{L^{2}}=\langle d_A^{*}a,q\rangle_{L^{2}}=0 ⟨ a , d A q ⟩ L 2 = ⟨ d A ∗ a , q ⟩ L 2 = 0 . This is how we remove changes that only alter the gauge representative.
Recall that A A A is a G 2 G_{2} G 2 instanton if its curvature F A ∈ Ω 14 2 F_A\in\Omega^{2}_{14} F A ∈ Ω 14 2 which means that F A ∧ ψ = 0 F_A\wedge\psi=0 F A ∧ ψ = 0 . Using (14) the instanton condition can we expressed as the vanishing of a one form. Define L ψ : Ω 2 ( E ) → Ω 1 ( E ) L_{\psi}:\Omega^{2}(E)\to\Omega^{1}(E) L ψ : Ω 2 ( E ) → Ω 1 ( E ) by L ψ β = ∗ ( β ∧ ψ ) L_{\psi}\beta= *(\beta\wedge\psi) L ψ β = ∗ ( β ∧ ψ ) . Let π 7 \pi_{7} π 7 be orthogonal projection to the Λ 7 2 \Lambda^{2}_{7} Λ 7 2 component, acting on each E E E component. The adjoint and the two compositions of L ψ L_{\psi} L ψ are
L ψ ∗ a = ∗ ( a ∧ ψ ) = a ♯ ⌟ φ , L ψ L ψ ∗ = 3 id , L ψ ∗ L ψ = 3 π 7 . (43) L_{\psi}^{*}a = *(a \wedge\psi) = a^{\sharp}\mathbin{\lrcorner}\varphi,\qquad L_{\psi}L_{\psi}^{*}=3\operatorname{id},\qquad L_{\psi}^{*}L_{\psi}=3\pi_{7}. \tag*{(43)} L ψ ∗ a = ∗ ( a ∧ ψ ) = a ♯ ┘ φ , L ψ L ψ ∗ = 3 id , L ψ ∗ L ψ = 3 π 7 . (43)
Write χ = L ψ F A \chi=L_{\psi}F_{A} χ = L ψ F A and P = L ψ d A \mathcal{P}=L_{\psi}d_{A} P = L ψ d A , as in (6) . Thus χ = 0 \chi=0 χ = 0 is the instanton equation, and ∣ χ ∣ 2 = 3 ∣ π 7 F A ∣ 2 |\chi|^{2}=3|\pi_{7}F_{A}|^{2} ∣ χ ∣ 2 = 3∣ π 7 F A ∣ 2 . In particular, the defect measures one component of the curvature, rather than the whole curvature. For fixed ψ \psi ψ , its linearization in a connection direction a a a is P a \mathcal{P}a P a . We also note, as before that
d A ∗ χ = − ∗ d A ( F A ∧ ψ ) = 0. (44) d_{A}^{*}\chi=-*d_{A}(F_{A}\wedge\psi)=0. \tag*{(44)} d A ∗ χ = − ∗ d A ( F A ∧ ψ ) = 0. (44)
For bundle valued forms, ⟨ α ∧ β ⟩ \langle\alpha\wedge\beta\rangle ⟨ α ∧ β ⟩ denotes the form obtained by pairing the E E E components: in an orthonormal frame of E E E , it is ∑ λ α λ ∧ β λ \sum_{\lambda}\alpha^{\lambda}\wedge\beta^{\lambda} ∑ λ α λ ∧ β λ . Second, P \mathcal{P} P is formally self-adjoint. For a , b ∈ Ω 1 ( E ) a,b\in\Omega^{1}(E) a , b ∈ Ω 1 ( E ) , integration by parts, using d ψ = 0 d\psi=0 d ψ = 0 , gives
∫ M ⟨ P a , b ⟩ d μ = ∫ M ⟨ d A a ∧ b ⟩ ∧ ψ = ∫ M ⟨ a ∧ d A b ⟩ ∧ ψ = ∫ M ⟨ a , P b ⟩ d μ . \int_{M}\langle\mathcal{P}a,b\rangle\,d\mu=\int_{M}\langle d_{A}a\wedge b\rangle\wedge\psi=\int_{M}\langle a\wedge d_{A}b\rangle\wedge\psi=\int_{M}\langle a,\mathcal{P}b\rangle\,d\mu. ∫ M ⟨ P a , b ⟩ d μ = ∫ M ⟨ d A a ∧ b ⟩ ∧ ψ = ∫ M ⟨ a ∧ d A b ⟩ ∧ ψ = ∫ M ⟨ a , P b ⟩ d μ .
This proves self-adjointness. In particular
P χ = P ∗ χ = d A ∗ L ψ ∗ L ψ F A = 3 d A ∗ π 7 F A . (45) \mathcal{P}\chi=\mathcal{P}^{*}\chi=d_{A}^{*}L_{\psi}^{*}L_{\psi}F_{A}=3d_{A}^{*}\pi_{7}F_{A}. \tag*{(45)} P χ = P ∗ χ = d A ∗ L ψ ∗ L ψ F A = 3 d A ∗ π 7 F A . (45)
For fixed ψ \psi ψ , the derivative of 1 2 ∫ M ∣ χ ∣ 2 d μ \frac{1}{2}\int_{M}|\chi|^{2}d\mu 2 1 ∫ M ∣ χ ∣ 2 d μ in the direction a a a is ⟨ χ , P a ⟩ L 2 \langle\chi,\mathcal{P}a\rangle_{L^{2}} ⟨ χ , P a ⟩ L 2 . Self-adjointness moves P \mathcal{P} P onto χ \chi χ , so the gradient is P χ \mathcal{P}\chi P χ . This explains the term − P χ -\mathcal{P}\chi − P χ in the flow below.
We next compare this gradient with the Yang–Mills gradient. From (14) we have
∗ F A = φ ∧ F A + 3 ∗ π 7 F A . (46) *F_{A}=\varphi\wedge F_{A}+3*\pi_{7}F_{A}. \tag*{(46)} ∗ F A = φ ∧ F A + 3 ∗ π 7 F A . (46)
On two forms in dimension seven, d A ∗ = ∗ d A ∗ d_{A}^{*}=*d_{A}* d A ∗ = ∗ d A ∗ . Apply ∗ d A *d_{A} ∗ d A to (46) . The Bianchi identity removes the derivative of F A F_{A} F A in the first term, leaving ∗ ( d φ ∧ F A ) = ∗ ( F A ∧ d φ ) *(d\varphi\wedge F_{A})=*(F_{A}\wedge d\varphi) ∗ ( d φ ∧ F A ) = ∗ ( F A ∧ d φ ) . The last term is 3 d A ∗ π 7 F A = P χ 3d_{A}^{*}\pi_{7}F_{A}=\mathcal{P}\chi 3 d A ∗ π 7 F A = P χ by (45) . We, thus obtain
d A ∗ F A = ∗ ( F A ∧ d φ ) + P χ . (47) d_{A}^{*}F_{A}=*(F_{A}\wedge d\varphi)+\mathcal{P}\chi. \tag*{(47)} d A ∗ F A = ∗ ( F A ∧ d φ ) + P χ . (47)
Therefore the torsion corrected Yang–Mills velocity
− d A ∗ F A + ∗ ( F A ∧ d φ ) = − P χ (48) -d_{A}^{*}F_{A}+*(F_{A}\wedge d\varphi)=-\mathcal{P}\chi \tag*{(48)} − d A ∗ F A + ∗ ( F A ∧ d φ ) = − P χ (48)
is the negative gradient of the instanton defect energy give by
1 2 ∫ M ∣ χ ∣ 2 d μ = 1 2 ( ∫ M ∣ F A ∣ 2 d μ − ∫ M ⟨ F A ∧ F A ⟩ ∧ φ ) . \frac{1}{2}\int_{M}|\chi|^{2}\,d\mu=\frac{1}{2}\left(\int_{M}|F_{A}|^{2}\,d\mu-\int_{M}\langle F_{A}\wedge F_{A}\rangle\wedge\varphi\right). 2 1 ∫ M ∣ χ ∣ 2 d μ = 2 1 ( ∫ M ∣ F A ∣ 2 d μ − ∫ M ⟨ F A ∧ F A ⟩ ∧ φ ) .
Remark 4.1. For a fixed ambient structure, this energy is one half of the Ω \Omega Ω -Yang–Mills functional considered by Chen and Wentworth [7 Section 1.1] , with Ω = − φ \Omega=-\varphi Ω = − φ in our conventions and with their trace pairing expressed using our fixed invariant fibre inner product.
Now we also allow ψ t \psi_{t} ψ t to vary, and set Ω 4 ( M ) ∋ Q = ∂ t ψ \Omega^{4}(M)\ni Q=\partial_{t}\psi Ω 4 ( M ) ∋ Q = ∂ t ψ and b = ∗ ( F A ∧ Q ) b=*(F_{A}\wedge Q) b = ∗ ( F A ∧ Q ) . When we differentiate χ = ∗ ( F A ∧ ψ ) \chi=*(F_{A}\wedge\psi) χ = ∗ ( F A ∧ ψ ) , the Hodge star, the curvature, and the final factor ψ \psi ψ each contribute one term. Keeping these contributions separate gives
∂ t χ = K t χ + P ∂ t A + b , K t = ( ∂ t ∗ ) ∗ − 1 ∣ Ω 1 . (49) \partial_{t}\chi=K_{t}\chi+\mathcal{P}\partial_{t}A+b,\qquad K_{t}=(\partial_{t}*)*^{-1}\big|_{\Omega^{1}}. \tag*{(49)} ∂ t χ = K t χ + P ∂ t A + b , K t = ( ∂ t ∗ ) ∗ − 1 Ω 1 . (49)
The differentiated star acts on six forms. Thus K t K_{t} K t is an algebraic operator determined by g t − 1 ∂ t g t g_{t}^{-1}\partial_{t}g_{t} g t − 1 ∂ t g t , with no derivatives of χ \chi χ . For ordinary Yang–Mills flow, substitute ∂ t A = − d A ∗ F A \partial_{t}A=-d_{A}^{*}F_{A} ∂ t A = − d A ∗ F A and use (47) to get
∂ t χ = K t χ − P 2 χ − P ∗ ( F A ∧ d φ ) + b . (50) \partial_{t}\chi=K_{t}\chi-\mathcal{P}^{2}\chi-\mathcal{P}*(F_{A}\wedge d\varphi)+b. \tag*{(50)} ∂ t χ = K t χ − P 2 χ − P ∗ ( F A ∧ d φ ) + b . (50)
At an instanton, the terms containing χ \chi χ vanish. The remaining terms come from torsion and the variation of ψ \psi ψ . Thus we see that the ordinary Yang–Mills flow would preserve the instanton equation to first order only if ∗ ( F A ∧ Q ) = P ( ∗ ( F A ∧ d ϕ ) ) \ast(F_A\wedge Q)=P(\ast(F_A\wedge d\phi)) ∗ ( F A ∧ Q ) = P ( ∗ ( F A ∧ d ϕ )) . Adding the torsion term in (48) removes the torsion error, but leaves the ambient forcing b b b .
At an instanton, the compatibility equation obtained by differentiating the instanton condition also appears in the study of simultaneous deformations of the ambient structure and the connection by de la Ossa, Larfors and Svanes [28 Section 5.1] . Their ambient deformations are in the setting of torsion-free G 2 G_2 G 2 structures. Here we allow coclosed backgrounds with torsion and use the compatibility equation to choose a connection velocity for the prescribed ambient variation Q Q Q .
We would therefore like to solve P u = − b \mathcal{P}u=-b P u = − b . We cannot invert P \mathcal{P} P on all one forms because at an instanton, it vanishes on every gauge direction d A q d_Aq d A q . Instead, we impose d A ∗ u = 0 d_A^\ast u=0 d A ∗ u = 0 and introduce a function (or a 0 0 0 -form) f f f . The enlarged operator is precisely D \mathcal{D} D from (7) and its first component imposes the gauge condition, and its second component is d A f + P u d_Af+\mathcal{P}u d A f + P u . The first step is to check ellipticity of the operator. We do this below.
Lemma 4.2. For d ψ = 0 d\psi=0 d ψ = 0 , the operator D \mathcal{D} D in (7) is formally self-adjoint and elliptic on Ω 0 ( E ) ⊕ Ω 1 ( E ) \Omega^0(E)\oplus\Omega^1(E) Ω 0 ( E ) ⊕ Ω 1 ( E ) . At an instanton its kernel is
ker D = ker d A ∣ Ω 0 ( E ) ⊕ H A , ψ 1 , where H A , ψ 1 = { a : d A ∗ a = 0 , P a = 0 } . (51) \ker\mathcal{D}=\ker d_A|_{\Omega^0(E)}\oplus\mathcal{H}^1_{A,\psi},\qquad\text{where}\qquad\mathcal{H}^1_{A,\psi}=\{a:d_A^\ast a=0,\ \mathcal{P}a=0\}. \tag*{(51)} ker D = ker d A ∣ Ω 0 ( E ) ⊕ H A , ψ 1 , where H A , ψ 1 = { a : d A ∗ a = 0 , P a = 0 } . (51)
Proof. Self adjointness follows from that of P \mathcal{P} P and the adjoint relation between d A d_A d A and d A ∗ d_A^\ast d A ∗ . With the symbol convention of Lemma 3.4, the principal symbol at a real covector ξ \xi ξ is
σ ξ ( D ) ( f , a ) = − 1 ( − ⟨ ξ , a ⟩ , ξ f + ξ × a ) . (52) \sigma_\xi(\mathcal{D})(f,a)=\sqrt{-1}\,(-\langle\xi,a\rangle,\ \xi f+\xi\times a). \tag*{(52)} σ ξ ( D ) ( f , a ) = − 1 ( − ⟨ ξ , a ⟩ , ξ f + ξ × a ) . (52)
The cross product on covectors is the metric dual of the cross product on vectors and acts trivially on the E E E factor. For ordinary one forms a , b a,b a , b , the adjoint identity (43) gives ⟨ ∗ ( ξ ∧ a ∧ ψ ) , b ⟩ = ⟨ ξ ∧ a , b ♯ ⌟ ϕ ⟩ = φ ( ξ ♯ , a ♯ , b ♯ ) \langle\ast(\xi\wedge a\wedge\psi),b\rangle=\langle\xi\wedge a,b^\sharp\mathbin{\lrcorner}\phi\rangle=\varphi(\xi^\sharp,a^\sharp,b^\sharp) ⟨ ∗ ( ξ ∧ a ∧ ψ ) , b ⟩ = ⟨ ξ ∧ a , b ♯ ┘ ϕ ⟩ = φ ( ξ ♯ , a ♯ , b ♯ ) , and applying the same identity componentwise to E E E valued forms proves the cross product expression in (52) . Since ξ × ( ξ × a ) = − ∣ ξ ∣ 2 a + ⟨ ξ , a ⟩ ξ \xi\times(\xi\times a)=-|\xi|^2a+\langle\xi,a\rangle\xi ξ × ( ξ × a ) = − ∣ ξ ∣ 2 a + ⟨ ξ , a ⟩ ξ , the square of this symbol is ∣ ξ ∣ 2 id |\xi|^2\operatorname{id} ∣ ξ ∣ 2 id , which proves ellipticity.
At an instanton the compositions
P d A f = ∗ ( [ F A , f ] ∧ ψ ) = 0 , d A ∗ P a = − ∗ ( [ F A , a ] ∧ ψ ) = 0 (53) \mathcal{P}d_Af=\ast([F_A,f]\wedge\psi)=0,\qquad d_A^\ast\mathcal{P}a=-\ast([F_A,a]\wedge\psi)=0 \tag*{(53)} P d A f = ∗ ([ F A , f ] ∧ ψ ) = 0 , d A ∗ P a = − ∗ ([ F A , a ] ∧ ψ ) = 0 (53)
vanish. The latter identity follows from [ F A , a ] ∧ ψ = [ F A ∧ ψ , a ] [F_A,a]\wedge\psi=[F_A\wedge\psi,a] [ F A , a ] ∧ ψ = [ F A ∧ ψ , a ] . Applying d A ∗ d_A^\ast d A ∗ to the second equation in D ( f , a ) = 0 \mathcal{D}(f,a)=0 D ( f , a ) = 0 therefore yields d A ∗ d A f = 0 d_A^\ast d_Af=0 d A ∗ d A f = 0 , and integration gives d A f = 0 d_Af=0 d A f = 0 and then P a = 0 \mathcal{P}a=0 P a = 0 , proving (51) .
Note that the two kernel spaces have different meanings. The space ker d A ∣ Ω 0 ( E ) \ker d_A|_{\Omega^0(E)} ker d A ∣ Ω 0 ( E ) consists of infinitesimal gauge transformations fixing A A A . The space H A , ψ 1 \mathcal{H}^1_{A,\psi} H A , ψ 1 consists of infinitesimal instanton deformations orthogonal to gauge directions. The notion of acyclicity means that both spaces vanish. This is the usual nondegeneracy condition for isolated instantons which is discussed in detail in the work of Walpuski [38 ] .
For a self-adjoint elliptic operator on a compact manifold, the cokernel is identified with the kernel. Thus zero kernel gives both existence and uniqueness for the linear equation. The inverse gains one derivative, with estimates
∥ D − 1 v ∥ H s + 1 ≤ C s ∥ v ∥ H s , s ≥ 0. (54) \|\mathcal{D}^{-1}v\|_{H^{s+1}}\leq C_s\|v\|_{H^s},\qquad s\geq0. \tag*{(54)} ∥ D − 1 v ∥ H s + 1 ≤ C s ∥ v ∥ H s , s ≥ 0. (54)
To compare nearby operators, fix the bundle B = E ⊕ ( T ∗ M ⊗ E ) \mathcal{B}=E\oplus(T^\ast M\otimes E) B = E ⊕ ( T ∗ M ⊗ E ) and reference norms, and write D 0 = D A 0 , ψ 0 : H 1 ( B ) → L 2 ( B ) \mathcal{D}_0=\mathcal{D}_{A_0,\psi_0}:H^1(\mathcal{B})\to L^2(\mathcal{B}) D 0 = D A 0 , ψ 0 : H 1 ( B ) → L 2 ( B ) for an invertible reference operator, whose coefficients give the estimate
∥ D A , ψ − D 0 ∥ H 1 → L 2 ≤ C ( ∥ A − A 0 ∥ C 0 + ∥ ψ − ψ 0 ∥ C 1 ) . \|\mathcal{D}_{A,\psi}-\mathcal{D}_0\|_{H^1\to L^2}\leq C(\|A-A_0\|_{C^0}+\|\psi-\psi_0\|_{C^1}). ∥ D A , ψ − D 0 ∥ H 1 → L 2 ≤ C ( ∥ A − A 0 ∥ C 0 + ∥ ψ − ψ 0 ∥ C 1 ) .
Although the comparison uses fixed norms, the adjoints still use the current metric. Its first derivatives enter the zeroth-order coefficients, which explains the C 1 C^1 C 1 norm of ψ − ψ 0 \psi-\psi_0 ψ − ψ 0 . If the perturbation is small enough that ∥ ( D A , ψ − D 0 ) D 0 − 1 ∥ L 2 → L 2 < 1 \|(\mathcal{D}_{A,\psi}-\mathcal{D}_0)\mathcal{D}_0^{-1}\|_{L^2\to L^2}<1 ∥ ( D A , ψ − D 0 ) D 0 − 1 ∥ L 2 → L 2 < 1 ,
D A , ψ − 1 = D 0 − 1 [ I + ( D A , ψ − D 0 ) D 0 − 1 ] − 1 : L 2 ( B ) ⟶ H 1 ( B ) , \mathcal{D}_{A,\psi}^{-1}=\mathcal{D}_0^{-1}\left[I+(\mathcal{D}_{A,\psi}-\mathcal{D}_0)\mathcal{D}_0^{-1}\right]^{-1}:L^2(\mathcal{B})\longrightarrow H^1(\mathcal{B}), D A , ψ − 1 = D 0 − 1 [ I + ( D A , ψ − D 0 ) D 0 − 1 ] − 1 : L 2 ( B ) ⟶ H 1 ( B ) ,
where the bracketed inverse is a norm convergent Neumann series. For smooth coefficients, the elliptic estimate ∥ v ∥ H s + 1 ≤ C s ( ∥ D v ∥ H s + ∥ v ∥ L 2 ) \|v\|_{H^{s+1}}\leq C_s(\|\mathcal{D}v\|_{H^s}+\|v\|_{L^2}) ∥ v ∥ H s + 1 ≤ C s ( ∥ D v ∥ H s + ∥ v ∥ L 2 ) and this base inverse prove (54) at every integer s ≥ 0 s\geq0 s ≥ 0 on the same neighborhood. Smooth inversion between the fixed spaces H s + 1 ( B ) H^{s+1}(\mathcal{B}) H s + 1 ( B ) and H s ( B ) H^s(\mathcal{B}) H s ( B ) , with the inverse differentiation formula used in Lemma 3.4 , gives smooth dependence on smooth parameter families at each order. The estimate constants may depend on higher coefficient norms.
For the calculations below, abbreviate u = U A , ψ ( Q ) u = \mathcal{U}_{A,\psi}(Q) u = U A , ψ ( Q ) . The expression described in (8) is then
D A , ψ ( f , u ) = ( 0 , − b ) . (55) \mathcal{D}_{A,\psi}(f,u)=(0,-b). \tag*{(55)} D A , ψ ( f , u ) = ( 0 , − b ) . (55)
Thus
d A ∗ u = 0 , P u + d A f = − b . (56) d_A^*u=0,\qquad\mathcal{P}u+d_Af=-b. \tag*{(56)} d A ∗ u = 0 , P u + d A f = − b . (56)
These equations define ( f , u ) (f,u) ( f , u ) on an open neighborhood, including connections with possibly nonzero defect. Invertibility makes the solution unique. Every operation commutes with gauge transformations, so the correction is gauge-equivariant. It is nonlocal because computing it requires an elliptic solution over M M M . We retain the scalar f f f in this open neighborhood as it need not vanish away from instantons but as we have already see, it vanishes on an instanton locus.
Proposition 4.3. At an acyclic instanton, the auxiliary section f f f in (55) is zero, and u u u is the unique solution of
d A ∗ u = 0 , ∗ ( d A u ∧ ψ ) = − ∗ ( F A ∧ Q ) . (57) d_A^*u=0,\qquad* (d_Au\wedge\psi)=-*(F_A\wedge Q). \tag*{(57)} d A ∗ u = 0 , ∗ ( d A u ∧ ψ ) = − ∗ ( F A ∧ Q ) . (57) Among all velocities v v v satisfying d A v ∧ ψ + F A ∧ Q = 0 d_Av\wedge\psi+F_A\wedge Q=0 d A v ∧ ψ + F A ∧ Q = 0 , it has minimal L 2 L^2 L 2 norm.
Proof. We have already proved f = 0 f=0 f = 0 . If v v v is another compatible velocity, we can solve d A ∗ d A q = d A ∗ ( v − u ) d_A^*d_Aq=d_A^*(v-u) d A ∗ d A q = d A ∗ ( v − u ) because the covariant Laplacian d A ∗ d A d_A^*d_A d A ∗ d A on Ω 0 ( M , E ) \Omega^0(M,E) Ω 0 ( M , E ) is invertible, its kernel being ker d A = { 0 } \ker d_A=\{0\} ker d A = { 0 } . By the choice of q q q , the form v − u − d A q v-u-d_Aq v − u − d A q is Coulomb. It is also killed by P \mathcal{P} P which can be seen from the facts that the velocities v v v and u u u satisfy the same compatibility equation, and P d A q = 0 \mathcal{P}d_Aq=0 P d A q = 0 . Hence it belongs to H A , ψ 1 \mathcal{H}^1_{A,\psi} H A , ψ 1 and vanishes by acyclicity. Thus, v = u + d A q v=u+d_Aq v = u + d A q . Since d A ∗ u = 0 d_A^*u=0 d A ∗ u = 0 , these summands are orthogonal, giving ∥ v ∥ 2 2 = ∥ u ∥ 2 2 + ∥ d A q ∥ 2 2 \lVert v\rVert_2^2=\lVert u\rVert_2^2+\lVert d_Aq\rVert_2^2 ∥ v ∥ 2 2 = ∥ u ∥ 2 2 + ∥ d A q ∥ 2 2 and the minimum norm claim. □
The Coulomb condition therefore has a concrete role here: by removing the component of the velocity tangent to the gauge orbit, it selects the minimum norm solution of the cancellation equation, just as orthogonality does in the coassociative construction in § 3 .
Consider a smooth solution of (9) on an interval where the correction is defined. By (48) , its equation is
∂ t A = − P χ + u . (58) \partial_tA=-\mathcal{P}\chi+u. \tag*{(58)} ∂ t A = − P χ + u . (58)
Combining (49) and (56) gives the defect equation
∂ t χ = − P 2 χ − d A f + K t χ . (59) \partial_t\chi=-\mathcal{P}^2\chi-d_Af+K_t\chi. \tag*{(59)} ∂ t χ = − P 2 χ − d A f + K t χ . (59)
Notice that the ambient forcing b b b from (50) has been replaced by − d A f -d_Af − d A f . Although this term need not vanish pointwise, it disappears from the energy identity when we do integration by parts and pair f f f with d A ∗ χ = 0 d_A^*\chi=0 d A ∗ χ = 0 . This uses (44) for arbitrary connections, not just instantons. We now calculate all contributions from the changing metric.
Theorem 4.4. Let M M M be compact without boundary, let ψ t \psi_t ψ t be a smooth family of closed positive four forms, and let A t A_t A t be a smooth solution of (9) . Suppose D A t , ψ t \mathcal{D}_{A_t,\psi_t} D A t , ψ t is invertible throughout the interval. Let h = ∂ t g t h=\partial_tg_t h = ∂ t g t and define E ( t ) = 1 2 ∫ M ∣ χ t ∣ g t 2 d μ t \mathcal{E}(t)=\frac{1}{2}\int_M\lvert\chi_t\rvert_{g_t}^2\,d\mu_t E ( t ) = 2 1 ∫ M ∣ χ t ∣ g t 2 d μ t . Then
E ′ ( t ) = − ∥ P χ ∥ L 2 ( g t ) 2 + 1 2 ∫ M h i j ⟨ χ i , χ j ⟩ d μ t − 1 4 ∫ M ( tr g t h ) ∣ χ ∣ 2 d μ t . (60) \mathcal{E}'(t)=-\lVert\mathcal{P}\chi\rVert_{L^2(g_t)}^2+\frac{1}{2}\int_M h^{ij}\langle\chi_i,\chi_j\rangle\,d\mu_t-\frac{1}{4}\int_M(\operatorname{tr}_{g_t}h)\lvert\chi\rvert^2\,d\mu_t. \tag*{(60)} E ′ ( t ) = − ∥ P χ ∥ L 2 ( g t ) 2 + 2 1 ∫ M h ij ⟨ χ i , χ j ⟩ d μ t − 4 1 ∫ M ( tr g t h ) ∣ χ ∣ 2 d μ t . (60) Let ∥ h ∥ ∞ , ∞ \lVert h\rVert_{\infty,\infty} ∥ h ∥ ∞ , ∞ denote the supremum of the pointwise operator norm of g − 1 h g^{-1}h g − 1 h , and let ∥ tr h ∥ ∞ \lVert\operatorname{tr}h\rVert_\infty ∥ tr h ∥ ∞ be the supremum of the absolute metric trace. Then
E ( t ) ≤ E ( 0 ) exp ( ∫ 0 t ( ∥ h ∥ ∞ , ∞ + 1 2 ∥ tr h ∥ ∞ ) d s ) . (61) \mathcal{E}(t)\leq\mathcal{E}(0)\exp\left(\int_0^t\left(\lVert h\rVert_{\infty,\infty}+\frac{1}{2}\lVert\operatorname{tr}h\rVert_\infty\right)\,ds\right). \tag*{(61)} E ( t ) ≤ E ( 0 ) exp ( ∫ 0 t ( ∥ h ∥ ∞ , ∞ + 2 1 ∥ tr h ∥ ∞ ) d s ) . (61)
Proof. This is a straightforward computation.
If the metric is stationary, the defect energy is nonincreasing, including when the positive four form varies within that fixed metric. This gives the following
Corollary 4.5. Along any smooth corrected connection solution, suppose
∥ D A t , ψ t v ∥ L 2 ≥ λ ( t ) ∥ v ∥ L 2 \lVert D_{A_t,\psi_t}v\rVert_{L^2} \ge\lambda(t)\lVert v\rVert_{L^2} ∥ D A t , ψ t v ∥ L 2 ≥ λ ( t ) ∥ v ∥ L 2 with λ ( t ) > 0 \lambda(t)>0 λ ( t ) > 0 . Let c h = ∥ g − 1 ∂ t g ∥ o p , ∞ + 1 2 ∥ tr g ∂ t g ∥ ∞ c_h=\lVert g^{-1}\partial_tg\rVert_{\mathrm{op},\infty}+\frac{1}{2}\lVert\operatorname{tr}_g\partial_tg\rVert_\infty c h = ∥ g − 1 ∂ t g ∥ op , ∞ + 2 1 ∥ tr g ∂ t g ∥ ∞ . Then
E ( t ) ≤ E ( 0 ) exp ( ∫ 0 t ( c h ( q ) − 2 λ ( q ) 2 ) d q ) . \mathcal{E}(t)\le\mathcal{E}(0)\exp\left(\int_0^t\left(c_h(q)-2\lambda(q)^2\right)\,dq\right). E ( t ) ≤ E ( 0 ) exp ( ∫ 0 t ( c h ( q ) − 2 λ ( q ) 2 ) d q ) . In particular, c h ≤ 2 λ 2 − η c_h\le2\lambda^2-\eta c h ≤ 2 λ 2 − η , η > 0 \eta>0 η > 0 , implies E ( t ) ≤ E ( 0 ) e − η t \mathcal{E}(t)\le\mathcal{E}(0)e^{-\eta t} E ( t ) ≤ E ( 0 ) e − η t .
Proof. The Bianchi identity gives d A ∗ χ = 0 d_A^*\chi=0 d A ∗ χ = 0 , and hence D ( 0 , χ ) = ( 0 , P χ ) D(0,\chi)=(0,P\chi) D ( 0 , χ ) = ( 0 , P χ ) . Therefore ∥ P χ ∥ 2 ≥ λ 2 ∥ χ ∥ 2 = 2 λ 2 E \lVert P\chi\rVert^2\ge\lambda^2\lVert\chi\rVert^2=2\lambda^2\mathcal{E} ∥ P χ ∥ 2 ≥ λ 2 ∥ χ ∥ 2 = 2 λ 2 E . We insert this in the defect-energy identity (61) above and apply Gronwall’s inequality.
We now give the proof of Theorem 1.5.
Proof of Theorem 1.5. Let A t A_t A t be a smooth solution of (1.9) on an interval where the correction is defined. The initial defect vanishes. On each compact time subinterval, Theorem 4.4 bounds the nonnegative energy by its zero initial value times a finite factor. Thus E ( t ) = 0 \mathcal{E}(t)=0 E ( t ) = 0 . The defect is smooth, so its zero L 2 L^2 L 2 norm makes it vanish everywhere and hence we have F A t ∧ ψ t = 0 F_{A_t}\wedge\psi_t=0 F A t ∧ ψ t = 0 . By (48) the connection equation then reduces to ∂ t A = U A , ψ ( Q ) \partial_tA=\mathcal{U}_{A,\psi}(Q) ∂ t A = U A , ψ ( Q ) , and Proposition 4.3 gives f = 0 f=0 f = 0 together with the two equations in (1.10), while existence and uniqueness for a short time are proved in Theorem 4.8.
Existence for the connection equation
As in the case of coassociatives, we now construct the solutions used in the preservation argument. The basic strategy is the same as in § 3 . We first prove that the correction has the required mapping properties. We then use the nearby instanton branch to construct the flow from an instanton. Finally, a gauge-fixed parabolic equation gives existence from nearby connections whose defect may be nonzero.
Let ψ t \psi_t ψ t , 0 ≤ t ≤ T 0 0\le t\le T_0 0 ≤ t ≤ T 0 , be a prescribed smooth family of closed positive four forms, and set Q t = ∂ t ψ t Q_t=\partial_t\psi_t Q t = ∂ t ψ t . Pairs of zero and one forms are sections of the fixed bundle B \mathcal{B} B introduced after Lemma 4.2. Choose a smooth reference connection A ⋆ A_\star A ⋆ for which D A ⋆ , ψ 0 D_{A_\star,\psi_0} D A ⋆ , ψ 0 is invertible, hold A ⋆ A_\star A ⋆ fixed, and write nearby connections as A = A ⋆ + a A=A_\star+a A = A ⋆ + a , with a ∈ Ω 1 ( M , E ) a\in\Omega^1(M,E) a ∈ Ω 1 ( M , E ) . The initial connection A i n A_{\mathrm{in}} A in in Theorem 4.9 is chosen near this reference connection, and neither connection is required to be an instanton. Writing D ⋆ = D A ⋆ , ψ 0 D_\star=D_{A_\star,\psi_0} D ⋆ = D A ⋆ , ψ 0 , we use the Sobolev norms from § 2 with fixed metric g 0 = g ψ 0 g_0=g_{\psi_0} g 0 = g ψ 0 and the connection induced by A ⋆ A_\star A ⋆ . We write L ( H m , H n ) \mathcal{L}(H^m,H^n) L ( H m , H n ) for bounded linear maps with the operator norm. We use H 6 H^6 H 6 for connections and H 5 H^5 H 5 for curvature, accounting for the one derivative in F A F_A F A . In dimension seven, H 5 H^5 H 5 is an algebra and H 6 H^6 H 6 controls two classical derivatives of the coefficients. In particular, multiplication maps H 6 × H 5 H^6\times H^5 H 6 × H 5 continuously into H 5 H^5 H 5 . The inverse of D D D will return the derivative used in forming the curvature.
Lemma 4.6. On a sufficiently small H 6 H^6 H 6 neighborhood of A ⋆ A_\star A ⋆ and a sufficiently short ambient interval, the map
H 6 ( M , T ∗ M ⊗ E ) ∋ a ⟼ U A ⋆ + a , ψ t ( Q t ) ∈ H 6 ( M , T ∗ M ⊗ E ) H^6(M,T^*M\otimes E)\ni a\longmapsto\mathcal{U}_{A_\star+a,\psi_t}(Q_t)\in H^6(M,T^*M\otimes E) H 6 ( M , T ∗ M ⊗ E ) ∋ a ⟼ U A ⋆ + a , ψ t ( Q t ) ∈ H 6 ( M , T ∗ M ⊗ E ) is smooth, with local Lipschitz constants uniform in t t t . It is smooth at every higher Sobolev order on the intersection with that same neighborhood. For smooth a = A − A ⋆ a=A-A_\star a = A − A ⋆ with ∥ a ∥ H 6 ≤ K \lVert a\rVert_{H^6}\le K ∥ a ∥ H 6 ≤ K in this neighborhood, and every integer r ≥ 6 r\ge6 r ≥ 6 , we have
∥ ( f , u ) ∥ H r ≤ C r , K ( 1 + ∥ a ∥ H r ) . (62) \lVert(f,u)\rVert_{H^r}\le C_{r,K}(1+\lVert a\rVert_{H^r}). \tag*{(62)} ∥( f , u ) ∥ H r ≤ C r , K ( 1 + ∥ a ∥ H r ) . (62) On bounded subsets at any order s ≥ 6 s\ge6 s ≥ 6 one also has
∥ U A 1 , ψ ( Q ) − U A 2 , ψ ( Q ) ∥ H s ≤ C ∥ Q ∥ C s + 1 ∥ A 1 − A 2 ∥ H s . (63) \lVert\mathcal{U}_{A_1,\psi}(Q)-\mathcal{U}_{A_2,\psi}(Q)\rVert_{H^s}
\le C\lVert Q\rVert_{C^{s+1}}\lVert A_1-A_2\rVert_{H^s}. \tag*{(63)} ∥ U A 1 , ψ ( Q ) − U A 2 , ψ ( Q ) ∥ H s ≤ C ∥ Q ∥ C s + 1 ∥ A 1 − A 2 ∥ H s . (63) For smooth connections with uniform coefficient bounds and a uniform low-order inverse bound, the correction is locally Lipschitz in L 2 L^2 L 2 .
Proof. We first work at the low order. At the reference pair, elliptic invertibility gives D ⋆ : H 6 ( B ) → H 5 ( B ) \mathcal{D}_{\star}: H^{6}(B) \to H^{5}(B) D ⋆ : H 6 ( B ) → H 5 ( B ) as an isomorphism. For fixed ψ t \psi_t ψ t , changing A A A changes only the zeroth-order coefficients of D \mathcal{D} D : the difference of two covariant derivatives is multiplication by the difference of the connections. The product estimates therefore give
∥ D A , ψ t − D ⋆ ∥ L ( H 6 , H 5 ) ≤ C ( ∥ A − A ⋆ ∥ H 6 + ∥ ψ t − ψ 0 ∥ C 6 ) . \|\mathcal{D}_{A,\psi_t}-\mathcal{D}_{\star}\|_{\mathcal{L}(H^{6},H^{5})}\leq C\bigl(\|A-A_{\star}\|_{H^{6}}+\|\psi_t-\psi_0\|_{C^{6}}\bigr). ∥ D A , ψ t − D ⋆ ∥ L ( H 6 , H 5 ) ≤ C ( ∥ A − A ⋆ ∥ H 6 + ∥ ψ t − ψ 0 ∥ C 6 ) .
Choose the neighborhood and time interval so that ∥ ( D A , ψ t − D ⋆ ) D ⋆ − 1 ∥ L ( H 5 , H 5 ) < 1 2 \|(\mathcal{D}_{A,\psi_t}-\mathcal{D}_{\star})\mathcal{D}_{\star}^{-1}\|_{\mathcal{L}(H^{5},H^{5})}<\frac{1}{2} ∥ ( D A , ψ t − D ⋆ ) D ⋆ − 1 ∥ L ( H 5 , H 5 ) < 2 1 . Then
D A , ψ t − 1 = D ⋆ − 1 [ I + ( D A , ψ t − D ⋆ ) D ⋆ − 1 ] − 1 : H 5 ⟶ H 6 \mathcal{D}_{A,\psi_t}^{-1}=\mathcal{D}_{\star}^{-1}[I+(\mathcal{D}_{A,\psi_t}-\mathcal{D}_{\star})\mathcal{D}_{\star}^{-1}]^{-1}:H^{5}\longrightarrow H^{6} D A , ψ t − 1 = D ⋆ − 1 [ I + ( D A , ψ t − D ⋆ ) D ⋆ − 1 ] − 1 : H 5 ⟶ H 6
is bounded uniformly and depends smoothly on A A A , while the curvature
F A ⋆ + a = F A ⋆ + d A ⋆ a + 1 2 [ a ∧ a ] F_{A_{\star}+a}=F_{A_{\star}}+d_{A_{\star}}a+\frac{1}{2}[a\wedge a] F A ⋆ + a = F A ⋆ + d A ⋆ a + 2 1 [ a ∧ a ]
is a smooth map H 6 → H 5 H^{6}\to H^{5} H 6 → H 5 . Multiplying the curvature by the prescribed smooth Q t Q_t Q t preserves H 5 H^{5} H 5 , and applying the inverse returns it to H 6 H^{6} H 6 . This proves the first assertion of the lemma and the local Lipschitz estimate. It also explains why the nonlocal correction does not lose a derivative of the connection.
Writing w = ( f , u ) w=(f,u) w = ( f , u ) , we observe that the difference D A , ψ t − D A ⋆ , ψ t \mathcal{D}_{A,\psi_t}-\mathcal{D}_{A_{\star},\psi_t} D A , ψ t − D A ⋆ , ψ t is multiplication by a a a followed by fixed algebraic contractions, so for r > 6 r>6 r > 6 we can combine the elliptic estimate for the smooth reference operator with the product inequality
∥ ( D A , ψ t − D A ⋆ , ψ t ) w ∥ H r − 1 ≤ C r ( ∥ a ∥ L ∞ ∥ w ∥ H r − 1 + ∥ a ∥ H r − 1 ∥ w ∥ L ∞ ) . \|(\mathcal{D}_{A,\psi_t}-\mathcal{D}_{A_{\star},\psi_t})w\|_{H^{r-1}}\leq C_r\bigl(\|a\|_{L^{\infty}}\|w\|_{H^{r-1}}+\|a\|_{H^{r-1}}\|w\|_{L^{\infty}}\bigr). ∥ ( D A , ψ t − D A ⋆ , ψ t ) w ∥ H r − 1 ≤ C r ( ∥ a ∥ L ∞ ∥ w ∥ H r − 1 + ∥ a ∥ H r − 1 ∥ w ∥ L ∞ ) .
to give, after interpolation,
∥ w ∥ H r ≤ C r , K ( ∥ D A , ψ t w ∥ H r − 1 + ( 1 + ∥ a ∥ H r − 1 ) ∥ w ∥ H 6 ) . \|w\|_{H^r}\leq C_{r,K}\bigl(\|\mathcal{D}_{A,\psi_t}w\|_{H^{r-1}}+(1+\|a\|_{H^{r-1}})\|w\|_{H^6}\bigr). ∥ w ∥ H r ≤ C r , K ( ∥ D A , ψ t w ∥ H r − 1 + ( 1 + ∥ a ∥ H r − 1 ) ∥ w ∥ H 6 ) .
The interpolation step uses ∥ w ∥ H r − 1 ≤ ε ∥ w ∥ H r + C r , ε ∥ w ∥ H 6 \|w\|_{H^{r-1}}\leq\varepsilon\|w\|_{H^r}+C_{r,\varepsilon}\|w\|_{H^6} ∥ w ∥ H r − 1 ≤ ε ∥ w ∥ H r + C r , ε ∥ w ∥ H 6 for r > 6 r>6 r > 6 . The bound on ∥ a ∥ H 6 \|a\|_{H^6} ∥ a ∥ H 6 controls ∥ a ∥ L ∞ \|a\|_{L^{\infty}} ∥ a ∥ L ∞ , so we can choose ε \varepsilon ε small enough to move the resulting multiple of ∥ w ∥ H r \|w\|_{H^r} ∥ w ∥ H r to the left. Also H 6 H^6 H 6 controls ∥ w ∥ L ∞ \|w\|_{L^{\infty}} ∥ w ∥ L ∞ . For our solution the base inverse estimate bounds ∥ w ∥ H 6 \|w\|_{H^6} ∥ w ∥ H 6 in terms of K K K , while ∗ ( F A ∧ Q ) \ast(F_A\wedge Q) ∗ ( F A ∧ Q ) has H r − 1 H^{r-1} H r − 1 norm at most C r , K ( 1 + ∥ a ∥ H r ) C_{r,K}(1+\|a\|_{H^r}) C r , K ( 1 + ∥ a ∥ H r ) , giving (62) , with the case r = 6 r=6 r = 6 following directly from the base estimate.
For an arbitrary source in H r − 1 H^{r-1} H r − 1 , we start with its H 6 H^6 H 6 solution and apply the reference elliptic estimate successively to D A ⋆ , ψ t w = D A , ψ t w − ( D A , ψ t − D A ⋆ , ψ t ) w \mathcal{D}_{A_{\star},\psi_t}w=\mathcal{D}_{A,\psi_t}w-(\mathcal{D}_{A,\psi_t}-\mathcal{D}_{A_{\star},\psi_t})w D A ⋆ , ψ t w = D A , ψ t w − ( D A , ψ t − D A ⋆ , ψ t ) w . Beginning with the inverse already constructed at the low order, this argument successively gives one more derivative whenever the source and coefficients have that derivative. It yields the inverse H r − 1 → H r H^{r-1}\to H^r H r − 1 → H r on the same neighborhood, with bounds uniform on sets where ∥ a ∥ H r \|a\|_{H^r} ∥ a ∥ H r is bounded. Smooth inversion between these fixed Sobolev spaces proves smooth dependence at each order. We remark that we are not imposing a new smallness condition on every high derivative. For proving (63) , we differentiate D A , ψ w = − ( 0 , ∗ ( F A ∧ Q ) ) \mathcal{D}_{A,\psi}w=-(0,\ast(F_A\wedge Q)) D A , ψ w = − ( 0 , ∗ ( F A ∧ Q )) in a connection direction b b b to get
D A , ψ ∂ A w [ b ] = − ( ∂ A D [ b ] ) w − ( 0 , ∗ ( d A b ∧ Q ) ) . \mathcal{D}_{A,\psi}\,\partial_Aw[b]=-(\partial_A\mathcal{D}[b])w-(0,\ast(d_Ab\wedge Q)). D A , ψ ∂ A w [ b ] = − ( ∂ A D [ b ]) w − ( 0 , ∗ ( d A b ∧ Q )) .
and similar self-adjointness and elliptic regularity argument proves (63) .
Lemma 4.7. Let A 0 A_0 A 0 be an acyclic instanton for a closed positive ψ 0 \psi_0 ψ 0 . Every smooth closed positive ψ \psi ψ sufficiently close to ψ 0 \psi_0 ψ 0 in C 2 , α C^{2,\alpha} C 2 , α has a unique nearby instanton in the slice d A 0 ∗ g 0 ( A − A 0 ) = 0 d_{A_0}^{\ast_{g_0}}(A-A_0)=0 d A 0 ∗ g 0 ( A − A 0 ) = 0 . It is acyclic and depends smoothly on smooth finite-dimensional parameter families of ψ \psi ψ .
Proof. We use the implicit function theorem argument following [Wall17, Proposition 2.17]. The enlarged equation allows us to work with co-closed structures. We keep the gauge condition fixed at ( A 0 , g 0 ) (A_0,g_0) ( A 0 , g 0 ) , so that the domain and target are fixed Banach spaces. Consider
F ( f , a , ψ ) = ( d A 0 ∗ g 0 a , ∗ ψ ( F A 0 + a ∧ ψ ) + d A 0 + a f ) . \mathcal{F}(f,a,\psi)=\left(d_{A_0}^{\ast_{g_0}}a,\ast_{\psi}(F_{A_0+a}\wedge\psi)+d_{A_0+a}f\right). F ( f , a , ψ ) = ( d A 0 ∗ g 0 a , ∗ ψ ( F A 0 + a ∧ ψ ) + d A 0 + a f ) .
The unknowns lie in C 2 , α ( B ) C^{2,\alpha}(B) C 2 , α ( B ) , the target is C 1 , α ( B ) C^{1,\alpha}(B) C 1 , α ( B ) , and the ambient parameter ranges over positive forms in the closed subspace ker d ⊂ C 2 , α ( Λ 4 T ∗ M ) \ker d\subset C^{2,\alpha}(\Lambda^4T^{\ast}M) ker d ⊂ C 2 , α ( Λ 4 T ∗ M ) . This map is smooth, with derivative in ( f , a ) (f,a) ( f , a ) at ( 0 , 0 , ψ 0 ) (0,0,\psi_0) ( 0 , 0 , ψ 0 ) equal to D A 0 , ψ 0 \mathcal{D}_{A_0,\psi_0} D A 0 , ψ 0 , which is an isomorphism by acyclicity and elliptic regularity, so the implicit function theorem supplies ( f ( ψ ) , a ( ψ ) ) (f(\psi),a(\psi)) ( f ( ψ ) , a ( ψ )) .
We must check that a solution of the enlarged equation is an actual instanton. Write A = A 0 + a A=A_0+a A = A 0 + a and apply d A ∗ d_A^* d A ∗ to its second equation. By d A ∗ χ = 0 d_A^*\chi=0 d A ∗ χ = 0 , we get d A ∗ d A f = 0 d_A^*d_Af=0 d A ∗ d A f = 0 and hence f = 0 f=0 f = 0 and we get an acyclic instanton.
For smooth ψ \psi ψ , the slice and instanton equations have smooth principal coefficients and terms polynomial in a a a with no derivatives of a a a beyond the linear first order part, so elliptic regularity makes a a a smooth. At the solution, the derivative in the unknowns is
B ψ ( q , b ) = ( d A 0 ∗ b , d A 0 q + P A , ψ b ) . B_\psi(q,b)=\left(d_{A_0}^*b,d_{A_0}q+P_{A,\psi}b\right). B ψ ( q , b ) = ( d A 0 ∗ b , d A 0 q + P A , ψ b ) .
The implicit function theorem makes this derivative invertible at the original low regularity, and its symbol remains elliptic nearby. Elliptic regularity therefore upgrades its inverse to C r , α ( B ) → C r + 1 , α ( B ) C^{r,\alpha}(B)\to C^{r+1,\alpha}(B) C r , α ( B ) → C r + 1 , α ( B ) for every integer r ≥ 1 r\geq1 r ≥ 1 .
We also need smooth dependence on one common parameter neighborhood. Fix the neighborhood on which the C 2 , α C^{2,\alpha} C 2 , α branch was constructed. At each parameter value, the higher-order implicit function theorem gives a local branch of higher regularity. That branch solves the same equation in the original neighborhood, so uniqueness at the low order identifies it with the original branch. The original branch is therefore smooth at every order throughout the fixed neighborhood. For a path, one time interval suffices; we do not take an intersection of shrinking intervals as the regularity order increases.
We now prove the short-time existence and uniqueness for the corrected system of flows.
Theorem 4.8. Let ψ t \psi_t ψ t be a prescribed smooth closed positive family and let A 0 A_0 A 0 be an acyclic ψ 0 \psi_0 ψ 0 instanton. The coupled Yang-Mills system in equation (9) has a unique smooth short-time solution with initial value A 0 A_0 A 0 . It remains an instanton and satisfies
d A t ∗ ∂ t A t = 0 , P A , ψ ∂ t A = − ∗ ( F A ∧ Q ) . d_{A_t}^*\partial_tA_t=0,\qquad P_{A,\psi}\partial_tA=-*(F_A\wedge Q). d A t ∗ ∂ t A t = 0 , P A , ψ ∂ t A = − ∗ ( F A ∧ Q ) . Proof. We obtain a path of instantons A ^ t \widehat{A}_t A t in a fixed reference slice by Lemma 4.7. The slice fixes the representatives, but its condition is different from requiring the velocity to be Coulomb for the current connection and current metric. We impose the latter condition by a time-dependent gauge transformation.
First, differentiating the instanton equation gives
P A ^ , ψ ∂ t A ^ t = − ∗ ( F A ^ ∧ Q ) . P_{\widehat{A},\psi}\partial_t\widehat{A}_t=-*(F_{\widehat{A}}\wedge Q). P A , ψ ∂ t A t = − ∗ ( F A ∧ Q ) . The Hodge star variation term vanishes because the instanton defect is zero. Using the current metric, we solve
d A ∗ d A q = − d A ∗ ∂ t A ^ . d_A^*d_Aq=-d_A^*\partial_t\widehat{A}. d A ∗ d A q = − d A ∗ ∂ t A . Acyclicity makes this Laplacian invertible on E E E valued zero forms, so the equation determines a smooth q t q_t q t , which we use to generate a family of gauge transformations k t k_t k t . With gauge action k ∗ A = k − 1 A k + k − 1 d k k^*A=k^{-1}Ak+k^{-1}dk k ∗ A = k − 1 A k + k − 1 d k , we solve k t k ˙ t − 1 = q t k_t\dot{k}_t^{-1}=q_t k t k ˙ t − 1 = q t , k 0 = i d k_0=\mathrm{id} k 0 = id , and set A t = k t ∗ A ^ t A_t=k_t^*\widehat{A}_t A t = k t ∗ A t . Then
∂ t A t = k t − 1 ( ∂ t A ^ t + d A ^ t q t ) k t . \partial_tA_t=k_t^{-1}\left(\partial_t\widehat{A}_t+d_{\widehat{A}_t}q_t\right)k_t. ∂ t A t = k t − 1 ( ∂ t A t + d A t q t ) k t . This velocity satisfies d A t ∗ ∂ t A t = 0 d_{A_t}^*\partial_tA_t=0 d A t ∗ ∂ t A t = 0 and the displayed linearized instanton equation, since P A d A Q = 0 P_{A}d_AQ=0 P A d A Q = 0 at an instanton, so Proposition 4.3 identifies it with the unique solution U A , ψ ( Q ) \mathcal{U}_{A,\psi}(Q) U A , ψ ( Q ) . Together with χ = 0 \chi=0 χ = 0 and (48) , this shows that the path solves (9) .
Every other smooth solution with the same initial data remains an instanton by (60) and hence solves A ˙ = U A , ψ ( Q t ) \dot{A}=\mathcal{U}_{A,\psi}(Q_t) A ˙ = U A , ψ ( Q t ) , for which Lemma 4.6 gives local Lipschitz dependence in H 6 H^6 H 6 , so uniqueness follows from the Banach space ODE theorem.
Theorem 4.9. Let M M M be compact without boundary and let ψ t \psi_t ψ t , 0 ≤ t ≤ T 0 0\leq t\leq T_0 0 ≤ t ≤ T 0 , be a prescribed smooth family of closed positive four forms. Suppose A ∗ A_* A ∗ is smooth and D A ∗ , ψ 0 \mathcal{D}_{A_*,\psi_0} D A ∗ , ψ 0 is invertible. There are δ > 0 \delta>0 δ > 0 and 0 < T ≤ T 0 0<T\leq T_0 0 < T ≤ T 0 such that every smooth connection A i n A_{\mathrm{in}} A in with ∥ A i n − A ∗ ∥ H 6 < δ \lVert A_{\mathrm{in}}-A_*\rVert_{H^6}<\delta ∥ A in − A ∗ ∥ H 6 < δ admits a unique smooth solution of (9) on [ 0 , T ] [0,T] [ 0 , T ] with A ( 0 ) = A i n A(0)=A_{\mathrm{in}} A ( 0 ) = A in . The operator D A t , ψ t \mathcal{D}_{A_t,\psi_t} D A t , ψ t remains invertible. The initial connection need not be an instanton.
Proof. We first solve an equation with an additional gauge term, and then remove that term by a time-dependent gauge transformation. This is the construction of [6 Section 2.1] , with Lemma 4.6 controlling the nonlocal correction. We give the existence and uniqueness argument for our equation below. The additional point to check is that the elliptic correction has the required mapping and difference estimates and these are supplied by Lemma 4.6.
Write V t ( A ) = − d A ∗ t F A + ∗ t ( F A ∧ d φ t ) + U A , ψ t ( Q t ) V_t(A)=-d_{A}^{*_{t}}F_A+*_{t}(F_A\wedge d\varphi_t)+\mathcal{U}_{A,\psi_t}(Q_t) V t ( A ) = − d A ∗ t F A + ∗ t ( F A ∧ d φ t ) + U A , ψ t ( Q t ) for the right-hand side of (9) . Since each operation defining U \mathcal{U} U commutes with a gauge transformation, the convention k ∗ A = k − 1 A k + k − 1 d k k*A=k^{-1}Ak+k^{-1}dk k ∗ A = k − 1 A k + k − 1 d k gives
U k ∗ A , ψ ( Q ) = k − 1 U A , ψ ( Q ) k , V t ( k ∗ A ) = k − 1 V t ( A ) k . \mathcal{U}_{k*A,\psi}(Q)=k^{-1}\mathcal{U}_{A,\psi}(Q)k,\qquad V_t(k*A)=k^{-1}V_t(A)k. U k ∗ A , ψ ( Q ) = k − 1 U A , ψ ( Q ) k , V t ( k ∗ A ) = k − 1 V t ( A ) k .
Let B = A ⋆ + a B=A_{\star}+a B = A ⋆ + a be the connection in the gauge fixed equation and set
ζ t ( B ) = d A ⋆ ∗ t ( B − A ⋆ ) . \zeta_t(B)=d_{A_{\star}}^{*_{t}}(B-A_{\star}). ζ t ( B ) = d A ⋆ ∗ t ( B − A ⋆ ) .
We solve
∂ t B = V t ( B ) − d B ζ t ( B ) , B ( 0 ) = A i n . (64) \partial_tB=V_t(B)-d_B\zeta_t(B),\qquad B(0)=A_{\mathrm{in}}. \tag*{(64)} ∂ t B = V t ( B ) − d B ζ t ( B ) , B ( 0 ) = A in . (64)
The original equation is degenerate in gauge directions. The added term supplies the missing second-order part in those directions and makes the equation for B B B strictly parabolic. We will remove it when reconstructing A A A below.
Write K a ω = [ a ∧ ω ] K_a\omega=[a\wedge\omega] K a ω = [ a ∧ ω ] , so d B = d A ⋆ + K a d_B=d_{A_{\star}}+K_a d B = d A ⋆ + K a . All adjoints here still use g t g_t g t . We denote by
Δ A ⋆ , t = d A ⋆ ∗ g t d A ⋆ + d A ⋆ d A ⋆ ∗ g t , F B = F A ⋆ + d A ⋆ a + 1 2 [ a ∧ a ] . \Delta_{A_{\star},t}=d_{A_{\star}}^{*_{g_t}}d_{A_{\star}}+d_{A_{\star}}d_{A_{\star}}^{*_{g_t}},\qquad F_B=F_{A_{\star}}+d_{A_{\star}}a+\frac{1}{2}[a\wedge a]. Δ A ⋆ , t = d A ⋆ ∗ g t d A ⋆ + d A ⋆ d A ⋆ ∗ g t , F B = F A ⋆ + d A ⋆ a + 2 1 [ a ∧ a ] .
Direct substitution in (64) gives
∂ t a + Δ A ⋆ , t a = N t ( a ) + U A ⋆ + a , ψ t ( Q t ) , (65) \partial_ta+\Delta_{A_{\star},t}a=N_t(a)+\mathcal{U}_{A_{\star}+a,\psi_t}(Q_t), \tag*{(65)} ∂ t a + Δ A ⋆ , t a = N t ( a ) + U A ⋆ + a , ψ t ( Q t ) , (65)
where
N t ( a ) = − d A ⋆ ∗ g t F A ⋆ − 1 2 d A ⋆ ∗ g t [ a ∧ a ] − K a ∗ g t F B − [ a , d A ⋆ ∗ g t a ] + ∗ t ( F B ∧ d φ t ) . (66) N_t(a)=-d_{A_{\star}}^{*_{g_t}}F_{A_{\star}}-\frac{1}{2}d_{A_{\star}}^{*_{g_t}}[a\wedge a]-K_a^{*_{g_t}}F_B-[a,d_{A_{\star}}^{*_{g_t}}a]+*_{t}(F_B\wedge d\varphi_t). \tag*{(66)} N t ( a ) = − d A ⋆ ∗ g t F A ⋆ − 2 1 d A ⋆ ∗ g t [ a ∧ a ] − K a ∗ g t F B − [ a , d A ⋆ ∗ g t a ] + ∗ t ( F B ∧ d φ t ) . (66)
In the bracket with a zero form, the remaining one-form index is retained. The curvature expansion shows that every term in N t ( a ) N_t(a) N t ( a ) contains at most one derivative of a a a . Lemma 4.6 shows that the correction preserves the Sobolev order of a a a . Thus all second derivatives of the unknown are on the left. The coefficients of that operator depend only on the prescribed metric and the fixed reference connection, and its principal symbol is ∣ ξ ∣ g t 2 I |\xi|_{g_t}^{2}I ∣ ξ ∣ g t 2 I .
Let H ( t , s ) H(t,s) H ( t , s ) be the evolution family for ∂ t + Δ A ⋆ , t \partial_t+\Delta_{A_{\star},t} ∂ t + Δ A ⋆ , t . Its coefficients are smooth and uniformly elliptic on a fixed compact time interval. Linear nonautonomous parabolic theory gives this family [23 sections 6.1-6.2] . In fixed reference Sobolev norms it satisfies
∥ H ( t , s ) v ∥ H 6 ≤ C ∥ v ∥ H 6 , ∥ H ( t , s ) v ∥ H 6 ≤ C ( t − s ) − 1 / 2 ∥ v ∥ H 5 , 0 ≤ s < t ≤ T . (67) \lVert H(t,s)v\rVert_{H^6}\leq C\lVert v\rVert_{H^6},\qquad\lVert H(t,s)v\rVert_{H^6}\leq C(t-s)^{-1/2}\lVert v\rVert_{H^5},\qquad0\leq s<t\leq T. \tag*{(67)} ∥ H ( t , s ) v ∥ H 6 ≤ C ∥ v ∥ H 6 , ∥ H ( t , s ) v ∥ H 6 ≤ C ( t − s ) − 1/2 ∥ v ∥ H 5 , 0 ≤ s < t ≤ T . (67)
For completeness, these bounds directly follow from energy estimates. For a smooth v v v , put u ( q ) = H ( q , s ) v u(q)=H(q,s)v u ( q ) = H ( q , s ) v and y j ( q ) = ∥ u ( q ) ∥ H j 2 y_j(q)=\lVert u(q)\rVert_{H^j}^2 y j ( q ) = ∥ u ( q ) ∥ H j 2 . Uniform ellipticity, commutation with fixed reference derivatives, and integration by parts gives
y j ′ ( q ) + c j y j + 1 ( q ) ≤ C j y j ( q ) , j = 5 , 6 y_j'(q)+c_jy_{j+1}(q)\leq C_jy_j(q),\qquad j=5,6 y j ′ ( q ) + c j y j + 1 ( q ) ≤ C j y j ( q ) , j = 5 , 6
with constants uniform for 0 ≤ s ≤ q ≤ T 0\leq s\leq q\leq T 0 ≤ s ≤ q ≤ T . Gronwall’s inequality at order six proves the first bound above. At order five it gives
y 5 ( t ) + c ∫ s t y 6 ( q ) d q ≤ C y 5 ( s ) y_5(t)+c\int_s^t y_6(q)\,dq\leq Cy_5(s) y 5 ( t ) + c ∫ s t y 6 ( q ) d q ≤ C y 5 ( s )
Multiplying the order six differential inequality by q − s q-s q − s and dropping its negative coercive term yields d d q ( ( q − s ) y 6 ( q ) ) ≤ ( 1 + C 6 ( q − s ) ) y 6 ( q ) \frac{d}{dq}\bigl((q-s)y_6(q)\bigr) \le(1+C_6(q-s))y_6(q) d q d ( ( q − s ) y 6 ( q ) ) ≤ ( 1 + C 6 ( q − s )) y 6 ( q ) . After integration and the preceding order five estimate, ( t − s ) y 6 ( t ) ≤ C y 5 ( s ) (t-s)y_6(t) \le Cy_5(s) ( t − s ) y 6 ( t ) ≤ C y 5 ( s ) . Taking square roots proves the second bound, and density extends it to v ∈ H 5 v \in H^5 v ∈ H 5 . The second estimate gains the one derivative needed for the local nonlinearity. Its time singularity is integrable because ∫ 0 t ( t − s ) − 1 / 2 d s = 2 t \int_0^t(t-s)^{-1/2}\,ds=2\sqrt{t} ∫ 0 t ( t − s ) − 1/2 d s = 2 t . This is the reason that the integral equation below is a contraction for small time. The constants are uniform on a fixed short interval. Write N t ( a ) = N t ( a ) + U A ⋆ + a , ψ t ( Q t ) \mathcal{N}_t(a)=N_t(a)+\mathcal{U}_{A_\star+a,\psi_t}(Q_t) N t ( a ) = N t ( a ) + U A ⋆ + a , ψ t ( Q t ) . On a sufficiently small H 6 H^6 H 6 ball, products and Lemma 4.6 give
∥ N t ( a ) − N t ( b ) ∥ H 5 ≤ L ∥ a − b ∥ H 6 , sup t ∥ N t ( a ) ∥ H 5 ≤ M . (68) \|\mathcal{N}_t(a)-\mathcal{N}_t(b)\|_{H^5}\le L\|a-b\|_{H^6},\qquad\sup_t\|\mathcal{N}_t(a)\|_{H^5}\le M. \tag*{(68)} ∥ N t ( a ) − N t ( b ) ∥ H 5 ≤ L ∥ a − b ∥ H 6 , t sup ∥ N t ( a ) ∥ H 5 ≤ M . (68)
Choose a closed ball of radius ρ \rho ρ strictly inside the inverse neighborhood. For a i n = A i n − A ⋆ a_{\mathrm{in}}=A_{\mathrm{in}}-A_\star a in = A in − A ⋆ , the integral equation is a = T a a=\mathcal{T}a a = T a , where
( T a ) ( t ) = H ( t , 0 ) a i n + ∫ 0 t H ( t , s ) N s ( a ( s ) ) d s . (69) (\mathcal{T}a)(t)=H(t,0)a_{\mathrm{in}}+\int_0^t H(t,s)\mathcal{N}_s(a(s))\,ds. \tag*{(69)} ( T a ) ( t ) = H ( t , 0 ) a in + ∫ 0 t H ( t , s ) N s ( a ( s )) d s . (69)
On C ( [ 0 , T ] , H 6 ) C([0,T],H^6) C ([ 0 , T ] , H 6 ) , the integral is bounded by 2 C M T 2CM\sqrt{T} 2 C M T , and
sup 0 ≤ t ≤ T ∥ ( T a ) ( t ) − ( T b ) ( t ) ∥ H 6 ≤ 2 C L T sup 0 ≤ t ≤ T ∥ a ( t ) − b ( t ) ∥ H 6 . \sup_{0\le t\le T}\|(\mathcal{T}a)(t)-(\mathcal{T}b)(t)\|_{H^6}\le2CL\sqrt{T}\sup_{0\le t\le T}\|a(t)-b(t)\|_{H^6}. 0 ≤ t ≤ T sup ∥ ( T a ) ( t ) − ( T b ) ( t ) ∥ H 6 ≤ 2 C L T 0 ≤ t ≤ T sup ∥ a ( t ) − b ( t ) ∥ H 6 .
First choose δ \delta δ so that C δ < ρ / 4 C\delta<\rho/4 C δ < ρ /4 , and then choose T T T so that 2 C M T < ρ / 4 2CM\sqrt{T}<\rho/4 2 C M T < ρ /4 and 2 C L T < 1 / 2 2CL\sqrt{T}<1/2 2 C L T < 1/2 . The map then sends the closed ball of radius ρ \rho ρ in C ( [ 0 , T ] , H 6 ) C([0,T],H^6) C ([ 0 , T ] , H 6 ) into the ball of radius ρ / 2 \rho/2 ρ /2 and is a contraction there. Its unique fixed point is a solution of (65) , continuous into H 6 H^6 H 6 and differentiable into H 4 H^4 H 4 .
To prove smoothness on this same interval, use the bound ∥ a ∥ H 6 ≤ K \|a\|_{H^6}\le K ∥ a ∥ H 6 ≤ K , product estimates and (62) to obtain, for r ≥ 6 r\ge6 r ≥ 6 ,
∥ N t ( a ) ∥ H r − 1 ≤ C r , K ( 1 + ∥ a ∥ H r ) . \|\mathcal{N}_t(a)\|_{H^{r-1}}\le C_{r,K}(1+\|a\|_{H^r}). ∥ N t ( a ) ∥ H r − 1 ≤ C r , K ( 1 + ∥ a ∥ H r ) .
We differentiate (65) using the fixed reference connection and integrate by parts in the second-order term. Uniform ellipticity controls ∥ a ∥ H r + 1 \|a\|_{H^{r+1}} ∥ a ∥ H r + 1 . We pair the remaining term in H r − 1 H^{r-1} H r − 1 with a a a in H r + 1 H^{r+1} H r + 1 and use Young’s inequality. This gives an arbitrarily small multiple of ∥ a ∥ H r + 1 \|a\|_{H^{r+1}} ∥ a ∥ H r + 1 plus C r , K ( 1 + ∥ a ∥ H r ) C_{r,K}(1+\|a\|_{H^r}) C r , K ( 1 + ∥ a ∥ H r ) . Absorbing the small multiple yields
d d t ∥ a ∥ H r 2 + c ∥ a ∥ H r + 1 2 ≤ C r , K ( 1 + ∥ a ∥ H r 2 ) . (70) \frac{d}{dt}\|a\|_{H^r}^2+c\|a\|_{H^{r+1}}^2\le C_{r,K}(1+\|a\|_{H^r}^2). \tag*{(70)} d t d ∥ a ∥ H r 2 + c ∥ a ∥ H r + 1 2 ≤ C r , K ( 1 + ∥ a ∥ H r 2 ) . (70)
The commutators with the fixed reference derivatives have smooth coefficients and are absorbed in the same estimate. For each integer r > 6 r>6 r > 6 , we first solve locally in H r H^r H r by the contraction argument with heat estimates at that order, and uniqueness in H 6 H^6 H 6 identifies this solution with the one already constructed wherever both are defined. Applying Gronwall’s inequality to (70) then bounds its H r H^r H r norm in terms of the initial H r H^r H r norm and C r , K C_{r,K} C r , K up to any time before T T T .
The low-order solution stays in the smaller H 6 H^6 H 6 ball chosen above, so it remains a positive distance from the boundary where the inverse was defined. On bounded H r H^r H r subsets of this neighborhood, Lemma 4.6 and the inverse differentiation formula following Lemma 4.2 give uniform local Lipschitz bounds for N t : H r → H r − 1 \mathcal{N}_t:H^r\to H^{r-1} N t : H r → H r − 1 . These bounds and the distance to the boundary give a positive restart time for the contraction.
If the H r H^r H r solution stopped before T T T , its norm would still be bounded by (70) . Restarting sufficiently close to the stopping time would extend it, a contradiction. Consequently a ∈ C ( [ 0 , T ] , H r ) a\in C([0,T],H^r) a ∈ C ([ 0 , T ] , H r ) for every r r r . This proves higher regularity on the same interval as the low-order solution. The equation then gives a ∈ C 1 ( [ 0 , T ] , H r − 2 ) a\in C^1([0,T],H^{r-2}) a ∈ C 1 ([ 0 , T ] , H r − 2 ) at each order, and repeated differentiation in time supplies all time derivatives, including at t = 0 t=0 t = 0 .
To recover the original equation from B t B_t B t , solve the pointwise ODE for gauge transformations
k ˙ t k t − 1 = ζ t ( B t ) , k 0 = id , A t = k t ∗ B t . (71) \dot{k}_t k_t^{-1}=\zeta_t(B_t),\qquad k_0=\operatorname{id},\qquad A_t=k_t^*B_t. \tag*{(71)} k ˙ t k t − 1 = ζ t ( B t ) , k 0 = id , A t = k t ∗ B t . (71)
The right-hand side is a smooth section of the adjoint bundle, so the ODE defines smooth gauge transformations on [ 0 , T ] [0,T] [ 0 , T ] . Using the chosen gauge action convention gives
∂ t A t = k t − 1 ( ∂ t B t + d B t ( k ˙ t k t − 1 ) ) k t = k t − 1 V t ( B t ) k t = V t ( A t ) . \partial_t A_t=k_t^{-1}\left(\partial_t B_t+d_{B_t}\left(\dot{k}_t k_t^{-1}\right)\right)k_t
=k_t^{-1}\mathcal{V}_t(B_t)k_t=\mathcal{V}_t(A_t). ∂ t A t = k t − 1 ( ∂ t B t + d B t ( k ˙ t k t − 1 ) ) k t = k t − 1 V t ( B t ) k t = V t ( A t ) .
This gives a solution of the original equation with A ( 0 ) = A i n A(0)=A_{\mathrm{in}} A ( 0 ) = A in , and since D B t , ψ t \mathcal{D}_{B_t,\psi_t} D B t , ψ t is invertible by the choice of fixed point neighborhood, its gauge conjugate D A t , ψ t \mathcal{D}_{A_t,\psi_t} D A t , ψ t remains invertible as well and hence we get the existence of solutions.
We now prove uniqueness. The contraction proves uniqueness of the gauge-fixed equation. We still need uniqueness for the original equation, because two original solutions need not come with the same gauge-fixing construction. Let A 1 , A 2 A_1,A_2 A 1 , A 2 be smooth solutions with the same initial value and put v = A 1 − A 2 v=A_1-A_2 v = A 1 − A 2 . Abbreviate χ j = χ A j , ψ t \chi_j=\chi_{A_j,\psi_t} χ j = χ A j , ψ t and P j = P A j , ψ t P_j=P_{A_j,\psi_t} P j = P A j , ψ t . The curvature expansion gives
χ 1 − χ 2 = P 2 v + 1 2 ∗ t ( [ v ∧ v ] ∧ ψ t ) , ( P 1 − P 2 ) w = ∗ t ( [ v ∧ w ] ∧ ψ t ) . \chi_1-\chi_2=P_2v+\frac{1}{2}*_{t}([v\wedge v]\wedge\psi_t),\qquad(P_1-P_2)w=*_{t}([v\wedge w]\wedge\psi_t). χ 1 − χ 2 = P 2 v + 2 1 ∗ t ([ v ∧ v ] ∧ ψ t ) , ( P 1 − P 2 ) w = ∗ t ([ v ∧ w ] ∧ ψ t ) .
Put n ( v ) = 1 2 ∗ t ( [ v ∧ v ] ∧ ψ t ) n(v)=\frac{1}{2}*_{t}([v\wedge v]\wedge\psi_t) n ( v ) = 2 1 ∗ t ([ v ∧ v ] ∧ ψ t ) and C ( v , w ) = ∗ t ( [ v ∧ w ] ∧ ψ t ) C(v,w)=*_{t}([v\wedge w]\wedge\psi_t) C ( v , w ) = ∗ t ([ v ∧ w ] ∧ ψ t ) . Substituting the two expansions into (48) gives the exact difference equation
∂ t v = − P 2 2 v − P 2 n ( v ) − C ( v , P 2 v ) − C ( v , n ( v ) ) − C ( v , χ 2 ) + U A 1 , ψ t ( Q t ) − U A 2 , ψ t ( Q t ) . \partial_t v=-P_2^2v-P_2n(v)-C(v,P_2v)-C(v,n(v))-C(v,\chi_2)+U_{A_1,\psi_t}(Q_t)-U_{A_2,\psi_t}(Q_t). ∂ t v = − P 2 2 v − P 2 n ( v ) − C ( v , P 2 v ) − C ( v , n ( v )) − C ( v , χ 2 ) + U A 1 , ψ t ( Q t ) − U A 2 , ψ t ( Q t ) .
Pair this equation with v v v in the current L 2 L^2 L 2 inner product. Self-adjointness gives − ∥ P 2 v ∥ 2 2 -\lVert P_2v\rVert_2^2 − ∥ P 2 v ∥ 2 2 from the first term and − ⟨ n ( v ) , P 2 v ⟩ L 2 -\langle n(v),P_2v\rangle_{L^2} − ⟨ n ( v ) , P 2 v ⟩ L 2 from the second. On a compact time interval both solutions are smooth, so v v v and χ 2 \chi_2 χ 2 have bounded pointwise norms. The algebraic estimates ∣ n ( v ) ∣ ≤ C ∣ v ∣ 2 |n(v)|\le C|v|^2 ∣ n ( v ) ∣ ≤ C ∣ v ∣ 2 and ∣ C ( v , w ) ∣ ≤ C ∣ v ∣ ∣ w ∣ |C(v,w)|\le C|v||w| ∣ C ( v , w ) ∣ ≤ C ∣ v ∣∣ w ∣ , followed by Young’s inequality, bound the next four terms by 1 2 ∥ P 2 v ∥ 2 2 + C ∥ v ∥ 2 2 \frac{1}{2}\lVert P_2v\rVert_2^2+C\lVert v\rVert_2^2 2 1 ∥ P 2 v ∥ 2 2 + C ∥ v ∥ 2 2 .
For the nonlocal term, one may subtract its two defining elliptic equations. If w j = ( f j , u j ) w_j=(f_j,u_j) w j = ( f j , u j ) , their difference satisfies
D 1 ( w 1 − w 2 ) = − ( D 1 − D 2 ) w 2 − ( 0 , ∗ t ( ( F A 1 − F A 2 ) ∧ Q t ) ) . \mathcal{D}_1(w_1-w_2)=-(\mathcal{D}_1-\mathcal{D}_2)w_2-\left(0,*_{t}((F_{A_1}-F_{A_2})\wedge Q_t)\right). D 1 ( w 1 − w 2 ) = − ( D 1 − D 2 ) w 2 − ( 0 , ∗ t (( F A 1 − F A 2 ) ∧ Q t ) ) .
We estimate this source in H − 1 H^{-1} H − 1 rather than in L 2 L^2 L 2 , so that we do not need a derivative of v v v in the final bound. The first term is algebraic in v v v , with bounded smooth coefficients. In the curvature difference, d A 2 v d_{A_2}v d A 2 v has H − 1 H^{-1} H − 1 norm at most C ∥ v ∥ 2 C\lVert v\rVert_2 C ∥ v ∥ 2 , and the quadratic term is bounded by C ∥ v ∥ ∞ ∥ v ∥ 2 C\lVert v\rVert_\infty\lVert v\rVert_2 C ∥ v ∥ ∞ ∥ v ∥ 2 . Smoothness on the compact time interval bounds ∥ v ∥ ∞ \lVert v\rVert_\infty ∥ v ∥ ∞ . Thus the whole source has H − 1 H^{-1} H − 1 norm at most C ∥ v ∥ 2 C\lVert v\rVert_2 C ∥ v ∥ 2 . Applying D 1 − 1 : H − 1 → L 2 \mathcal{D}_1^{-1}:H^{-1}\to L^2 D 1 − 1 : H − 1 → L 2 gives ∥ u 1 − u 2 ∥ 2 ≤ C ∥ v ∥ 2 \lVert u_1-u_2\rVert_2\le C\lVert v\rVert_2 ∥ u 1 − u 2 ∥ 2 ≤ C ∥ v ∥ 2 . This also shows that no assumption about invertibility along the line segment between the two connections is needed. Finally, differentiating the metric and volume in ∥ v ∥ 2 2 \lVert v\rVert_2^2 ∥ v ∥ 2 2 contributes at most C ∥ v ∥ 2 2 C\lVert v\rVert_2^2 C ∥ v ∥ 2 2 . Thus
d d t ∥ v ∥ 2 2 ≤ C ∥ v ∥ 2 2 . \frac{d}{dt}\lVert v\rVert_2^2\le C\lVert v\rVert_2^2. d t d ∥ v ∥ 2 2 ≤ C ∥ v ∥ 2 2 .
and Gronwall’s inequality proves uniqueness for the original equation.
Continuation and transport along ambient paths
There are two questions about longer time intervals. We first ask whether an instanton can be transported along a controlled ambient ψ t \psi_t ψ t evolution for all time. We then give sufficient bounds for a general connection solution to continue past a finite time.
For transport, staying in the reference neighborhood gives an instanton at each time through the local branch. We also need a time-dependent gauge to make its velocity Coulomb. The size of this gauge is controlled by the total amount of ambient motion, measured by the time integral of the velocity.
Fix an acyclic reference pair ( A ∗ , ψ ∗ ) (A_*,\psi_*) ( A ∗ , ψ ∗ ) and write I ( ψ ) \mathcal{I}(\psi) I ( ψ ) for the instanton in the fixed reference Coulomb slice supplied by Lemma 4.7. We have the following result which tells us about the global transport along a small ambient path.
Theorem 4.10. Let ( A ⋆ , ψ ⋆ ) (A_\star,\psi_\star) ( A ⋆ , ψ ⋆ ) be as in Lemma 4.7. There are a finite integer r r r and ϵ > 0 \epsilon>0 ϵ > 0 such that a smooth closed positive family ψ t \psi_t ψ t , t ≥ 0 t\geq0 t ≥ 0 , satisfying
sup t ≥ 0 ∥ ψ t − ψ ⋆ ∥ C r < ϵ , ∫ 0 ∞ ∥ Q t ∥ C r d t < ϵ (72) \sup_{t\geq0}\|\psi_t-\psi_\star\|_{C^r}<\epsilon,\qquad\int_0^\infty\|Q_t\|_{C^r}\,\mathrm{d}t<\epsilon \tag*{(72)} t ≥ 0 sup ∥ ψ t − ψ ⋆ ∥ C r < ϵ , ∫ 0 ∞ ∥ Q t ∥ C r d t < ϵ (72) has a global instanton lift satisfying (10) , starting at I ( ψ 0 ) \mathcal{I}(\psi_0) I ( ψ 0 ) . Assume the analogous higher-norm bounds are finite. If ψ t \psi_t ψ t converges smoothly and Q t Q_t Q t is integrable in every norm, the lift converges smoothly to an acyclic instanton for the limiting four form. Exponential bounds for the ambient tail give exponential convergence of the lift. The lift remains uniformly close to A ⋆ A_\star A ⋆ at any prescribed fixed finite order if the corresponding bounds in (72) are small.
Proof. The proof is similar in spirit to the proof of Theorem 3.16 . Let B ‾ t = I ( ψ t ) \overline{B}_t=\mathcal{I}(\psi_t) B t = I ( ψ t ) . The local branch is defined for all t t t because the whole ambient path stays in its fixed neighborhood. It is acyclic, with a uniform inverse bound. Differentiating its instanton equation gives P B ‾ , t ∂ t B ‾ = − ∗ t ( F B ‾ ∧ Q t ) \mathcal{P}_{\overline{B},t}\partial_t\overline{B}=-*_{t}(F_{\overline{B}}\wedge Q_t) P B , t ∂ t B = − ∗ t ( F B ∧ Q t ) . We solve the equation
d B ‾ ∗ t d B ‾ q t = d B ‾ ∗ t ∂ t B ‾ . d_{\overline{B}}^{*_{t}}d_{\overline{B}}q_t=d_{\overline{B}}^{*_{t}}\partial_t\overline{B}. d B ∗ t d B q t = d B ∗ t ∂ t B .
The scalar Laplacian is uniformly invertible because its kernel is zero at the reference pair and this persists in the fixed neighborhood. We record why its solution is controlled by the ambient velocity. The derivative of the slice condition is d A ⋆ ∗ ∂ t B ‾ = 0 d_{A_\star}^{*}\partial_t\overline{B}=0 d A ⋆ ∗ ∂ t B = 0 , so the invertible linearization B ψ B_{\psi} B ψ in Lemma 4.7 gives
B ψ t ( 0 , ∂ t B ‾ t ) = ( 0 , − ∗ t ( F B ‾ t ∧ Q t ) ) . B_{\psi_t}(0,\partial_t\overline{B}_t)=(0,-*_{t}(F_{\overline{B}_t}\wedge Q_t)). B ψ t ( 0 , ∂ t B t ) = ( 0 , − ∗ t ( F B t ∧ Q t )) .
The inverse of B ψ t B_{\psi_t} B ψ t gains one derivative and bounds ∂ t B ‾ t \partial_t\overline{B}_t ∂ t B t by Q t Q_t Q t . The scalar equation for q t q_t q t has one derivative of ∂ t B ‾ t \partial_t\overline{B}_t ∂ t B t on the right, and its second-order inverse gains two derivatives. Combining the estimates, for each fixed k ≥ 0 k\geq0 k ≥ 0 we obtain
∥ ∂ t B ‾ t ∥ C k + 1 , α + ∥ q t ∥ C k + 2 , α ≤ C k ∥ Q t ∥ C k + 2 , α . (73) \|\partial_t\overline{B}_t\|_{C^{k+1,\alpha}}+\|q_t\|_{C^{k+2,\alpha}}\leq C_k\|Q_t\|_{C^{k+2,\alpha}}. \tag*{(73)} ∥ ∂ t B t ∥ C k + 1 , α + ∥ q t ∥ C k + 2 , α ≤ C k ∥ Q t ∥ C k + 2 , α . (73)
The constants use bounded ambient and branch coefficients at the corresponding higher orders and the uniform low-order inverse bounds. These are available by the regularity argument in Lemma 4.7. The derivative counts in (73) are deliberately more than is needed at some orders as they allow us to use one fixed finite norm when choosing the small neighborhood. Put v t = ∂ t B ‾ t − d B ‾ t q t v_t=\partial_t\overline{B}_t-d_{\overline{B}_t}q_t v t = ∂ t B t − d B t q t . Then
d B ‾ t ∗ t v t = 0 , P B ‾ , t v t = − ∗ t ( F B ‾ t ∧ Q t ) . d_{\overline{B}_t}^{*_{t}}v_t=0,\qquad\mathcal{P}_{\overline{B},t}v_t=-*_{t}(F_{\overline{B}_t}\wedge Q_t). d B t ∗ t v t = 0 , P B , t v t = − ∗ t ( F B t ∧ Q t ) .
Here P B ‾ , t d B ‾ q = 0 \mathcal{P}_{\overline{B},t}d_{\overline{B}}q=0 P B , t d B q = 0 because B ‾ \overline{B} B is an instanton. Thus v t = U B ‾ t , ψ t ( Q t ) v_t=\mathcal{U}_{\overline{B}_t,\psi_t}(Q_t) v t = U B t , ψ t ( Q t ) . Choose a gauge g t g_t g t with ( ∂ t g t ) g t − 1 = − q t (\partial_tg_t)g_t^{-1}=-q_t ( ∂ t g t ) g t − 1 = − q t and g 0 = 1 g_0=1 g 0 = 1 . Then B t = g t ∗ B ‾ t B_t=g_t^*\overline{B}_t B t = g t ∗ B t has velocity U B t , ψ t ( Q t ) \mathcal{U}_{B_t,\psi_t}(Q_t) U B t , ψ t ( Q t ) . At each point of M M M , the gauge equation is an ordinary differential equation in the compact structure group, so it exists for every finite time. To bound spatial derivatives, we differentiate in local bundle trivializations. At each order, the highest derivative of g t g_t g t satisfies a linear equation. Its coefficients involve q t q_t q t and the remaining terms involve derivatives of q t q_t q t and derivatives of g t g_t g t already controlled at lower orders. The time-integral bounds, (73) , and Gronwall’s inequality therefore bound all derivatives successively.
Integrating each differentiated equation between two later times bounds the difference by the tails of the corresponding time integrals. These tails tend to zero, so g t g_t g t is Cauchy in every C k C^k C k norm. The pointwise limit lies in the compact structure group and its smoothness follows from the derivative convergence. Thus the limit is a gauge transformation. The same estimates give
∥ g t − 1 ∥ C k + 1 ≤ C k ∫ 0 t ∥ Q q ∥ C k + 3 d q . \|g_t-1\|_{C^{k+1}}\leq C_k\int_0^t\|Q_q\|_{C^{k+3}}\,\mathrm{d}q. ∥ g t − 1 ∥ C k + 1 ≤ C k ∫ 0 t ∥ Q q ∥ C k + 3 d q .
when the relevant integral is small, in local charts about the identity. The formula for the gauge action then keeps B t B_t B t close to A ⋆ A_\star A ⋆ in C k C^k C k . Taking the finite integer r r r large enough gives the first assertion. The higher-norm integrability gives smooth convergence, and exponential bounds for the integrals from t t t to infinity give exponential convergence. The limiting instanton is acyclic because the uniform reference neighborhood has a uniform inverse bound, and conjugating by a gauge transformation preserves the spectrum. Uniqueness follows from Theorem 4.9 . □ \square □
We now give an analogue of Proposition 3.13 for the coupled system of the flow of 4-forms and the Yang–Mills flow.
Proposition 4.11. Suppose ψ t \psi_t ψ t extends smoothly across a finite T T T and remains positive and closed. A solution of (9) extends smoothly across T T T , modulo a time-dependent gauge, if in such a gauge
sup t < T ∥ A t − A r e f ∥ H s < ∞ , sup t < T ∥ D A t , ψ t − 1 ∥ L 2 → L 2 < ∞ \sup_{t<T}\|A_t-A_{\mathrm{ref}}\|_{H^s}<\infty,\qquad\sup_{t<T}\|\mathcal{D}_{A_t,\psi_t}^{-1}\|_{L^2\to L^2}<\infty t < T sup ∥ A t − A ref ∥ H s < ∞ , t < T sup ∥ D A t , ψ t − 1 ∥ L 2 → L 2 < ∞ for one integer s ≥ 8 s\geq8 s ≥ 8 . Here A r e f A_{\mathrm{ref}} A ref is fixed and smooth. Equivalently, the displayed bounds may be imposed on a gauge-fixed representative with its gauge term included in the equation.
Remark 4.12. This is a sufficient finite-order criterion using a connection bound in a gauge and the inverse bound.
Proof. Since the proof if similar to the proof of Proposition 3.13, we omit some details. We prove continuation by finding a local existence time that is uniform as the starting time approaches T T T . Use the same fixed smooth reference connection to express the gauge-fixed local equations. The assumed H s H^s H s bound controls the coefficients of D \mathcal{D} D at low order and in the higher Sobolev norms. The uniform L 2 L^2 L 2 inverse bound and the elliptic estimate successively give uniform bounds D − 1 : H s − 1 → H s \mathcal{D}^{-1}:H^{s-1}\to H^s D − 1 : H s − 1 → H s . To see that no neighborhood at every order is required, first use the L 2 L^2 L 2 bound in the order-one elliptic estimate and then apply the higher estimates inductively. Sobolev multiplication bounds all lower-order coefficients by the given H s H^s H s bound. The resolvent identity then gives a uniform radius on which the inverse remains bounded. The curvature expansion and the estimates for U \mathcal{U} U give uniform bounds and Lipschitz constants for the nonlinearity H s → H s − 1 H^s\to H^{s-1} H s → H s − 1 on that radius. The heat estimates used in Theorem 4.9 are uniform because the ambient metric extends smoothly across T T T . Consequently the contraction argument has an existence time bounded below by one positive number, independently of its starting time t 0 < T t_0<T t 0 < T .
If the assumed bound is obtained in a time-dependent gauge, fix its value at t 0 t_0 t 0 and use that single gauge transformation on the initial connection for the new local solution. Gauge covariance and uniqueness identify the new solution with the original one on their common interval, modulo this fixed transformation. Choose t 0 t_0 t 0 closer to T T T than the uniform existence time. The new solution then continues the geometric solution across T T T . Parabolic and elliptic regularity give smoothness there. This proves continuation of the geometric solution. If the gauge transformations themselves extend smoothly, the original representative extends as well.
Nonlinear stability
In this section, we first show that, near an acyclic instanton, the defect controls the connection after a Coulomb gauge has been chosen. We then compare a solution with a nearby moving instanton. A fixed positive elliptic operator gives decay of the difference in one Sobolev norm. We use this decay to keep the solution in the neighborhood where the correction is defined, and then use the parabolic equation to obtain decay of every higher derivative. The initial connection in this argument need not be an instanton.
Fix a closed positive ψ ∗ \psi_* ψ ∗ and an acyclic ψ ∗ \psi_* ψ ∗ -instanton A ∗ A_* A ∗ . We use an integer s ≥ 8 s\geq8 s ≥ 8 throughout this subsection. This is a sufficient regularity assumption, and we do not claim that it is optimal. In dimension seven, H s − 1 H^{s-1} H s − 1 is a Banach algebra and H s H^s H s embeds continuously into C 4 C^4 C 4 . We also remain above the H 6 H^6 H 6 regularity used in Theorem 4.9 . The higher coefficient bounds in (78) are separate assumptions on the prescribed background and they do not follow from this Sobolev embedding. Use the notation
L ∗ = P ∗ 2 + d A ∗ d A ∗ ∗ , P ∗ = P A ∗ , ψ ∗ . (74) \mathcal{L}_*=\mathcal{P}_*^2+d_{A_*}d_{A_*}^*,\qquad\mathcal{P}_*=\mathcal{P}_{A_*,\psi_*}. \tag*{(74)} L ∗ = P ∗ 2 + d A ∗ d A ∗ ∗ , P ∗ = P A ∗ , ψ ∗ . (74)
At an instanton P ∗ d A ∗ = 0 \mathcal{P}_*d_{A_*}=0 P ∗ d A ∗ = 0 and d A ∗ ∗ P ∗ = 0 d_{A_*}^*\mathcal{P}_*=0 d A ∗ ∗ P ∗ = 0 , so the mixed terms in D ∗ 2 \mathcal{D}_*^2 D ∗ 2 vanish. Its one-form block is therefore L ∗ \mathcal{L}_* L ∗ . For a one form a a a , integration by parts gives
⟨ L ∗ a , a ⟩ L 2 = ∥ P ∗ a ∥ L 2 2 + ∥ d A ∗ ∗ a ∥ L 2 2 . \langle\mathcal{L}_*a,a\rangle_{L^2}=\|\mathcal{P}_*a\|_{L^2}^2+\|d_{A_*}^*a\|_{L^2}^2. ⟨ L ∗ a , a ⟩ L 2 = ∥ P ∗ a ∥ L 2 2 + ∥ d A ∗ ∗ a ∥ L 2 2 .
The first term detects a change in the instanton equation, while the second detects the component in the gauge directions. If both vanish, acyclicity gives a = 0 a=0 a = 0 . Thus L ∗ \mathcal{L}_* L ∗ is positive, self-adjoint and strongly elliptic, with no kernel. Since the manifold is compact, we get that there exists a lowest eigenvalue ν > 0 \nu>0 ν > 0 of L \mathcal{L} L . For integers k ≥ 0 k\geq0 k ≥ 0 , we use the following fixed Sobolev norms
∣ a ∣ k = ∥ L ∗ k / 2 a ∥ L 2 ( g ∗ ) . (75) |a|_k=\|\mathcal{L}_*^{k/2}a\|_{L^2(g_*)}. \tag*{(75)} ∣ a ∣ k = ∥ L ∗ k /2 a ∥ L 2 ( g ∗ ) . (75)
Elliptic regularity and positivity make these norms equivalent to the ordinary H k H^k H k norms. To see the estimate that we will use, expand a a a in an orthonormal eigenbasis of L ⋆ \mathcal{L}_\star L ⋆ . If its eigenvalues are μ j ≥ ν \mu_j \ge\nu μ j ≥ ν and the coefficients of a a a are a j a_j a j , then ∣ a ∣ k 2 = ∑ j μ j k ∣ a j ∣ 2 \lvert a\rvert_k^2 = \sum_j \mu_j^k\lvert a_j\rvert^2 ∣ a ∣ k 2 = ∑ j μ j k ∣ a j ∣ 2 . Comparing these sums term by term gives
∣ a ∣ k + 1 2 ≥ ν ∣ a ∣ k 2 . \lvert a\rvert_{k+1}^2 \ge\nu\lvert a\rvert_k^2. ∣ a ∣ k + 1 2 ≥ ν ∣ a ∣ k 2 .
Lemma 4.13. Every sufficiently close H s H^s H s connection is gauge equivalent, by a gauge transformation close to the identity, to A ⋆ + a A_\star+a A ⋆ + a with d A ⋆ ⋆ a = 0 d_{A_\star}^\star a=0 d A ⋆ ⋆ a = 0 . In that slice,
∥ a ∥ H s ≤ C ∥ χ A ⋆ + a , ψ ⋆ ∥ H s − 1 . \lVert a\rVert_{H^s} \le C\lVert\chi_{A_\star+a,\psi_\star}\rVert_{H^{s-1}}. ∥ a ∥ H s ≤ C ∥ χ A ⋆ + a , ψ ⋆ ∥ H s − 1 . If a a a is sufficiently small in C 0 C^0 C 0 , one also has ∥ a ∥ H 1 ≤ C ∥ χ A ⋆ + a , ψ ⋆ ∥ L 2 \lVert a\rVert_{H^1} \le C\lVert\chi_{A_\star+a,\psi_\star}\rVert_{L^2} ∥ a ∥ H 1 ≤ C ∥ χ A ⋆ + a , ψ ⋆ ∥ L 2 .
Proof. We first choose the gauge. Consider d A ⋆ ⋆ ( ( exp q ) ⋆ ( A ⋆ + a ) − A ⋆ ) d_{A_\star}^\star\bigl((\exp q)^\star(A_\star+a)-A_\star\bigr) d A ⋆ ⋆ ( ( exp q ) ⋆ ( A ⋆ + a ) − A ⋆ ) as a function of the scalar gauge parameter q q q and the one form a a a . Its derivative with respect to q q q at zero is d A ⋆ ⋆ d A ⋆ d_{A_\star}^\star d_{A_\star} d A ⋆ ⋆ d A ⋆ . Acyclicity gives ker d A ⋆ ∣ Ω 0 ( E ) = 0 \ker d_{A_\star}|_{\Omega^0(E)}=0 ker d A ⋆ ∣ Ω 0 ( E ) = 0 , so this scalar Laplacian is an isomorphism H s + 1 → H s − 1 H^{s+1}\to H^{s-1} H s + 1 → H s − 1 . The gauge action is smooth between these spaces for s > 7 2 s>\frac{7}{2} s > 2 7 . The implicit function theorem therefore provides a unique sufficiently small q q q for which the transformed connection satisfies the stated Coulomb condition.
In the slice, D ⋆ ( 0 , a ) = ( 0 , P ⋆ a ) D_\star(0,a)=(0,P_\star a) D ⋆ ( 0 , a ) = ( 0 , P ⋆ a ) , so the elliptic inverse gives ∥ a ∥ H s ≤ C ∥ P ⋆ a ∥ H s − 1 \lVert a\rVert_{H^s}\le C\lVert P_\star a\rVert_{H^{s-1}} ∥ a ∥ H s ≤ C ∥ P ⋆ a ∥ H s − 1 . Since the instanton defect satisfies χ A ⋆ + a = P ⋆ a + 1 2 ∗ ( [ a ∧ a ] ∧ ψ ⋆ ) \chi_{A_\star+a}=P_\star a+\frac{1}{2}\ast([a\wedge a]\wedge\psi_\star) χ A ⋆ + a = P ⋆ a + 2 1 ∗ ([ a ∧ a ] ∧ ψ ⋆ ) , Sobolev multiplication gives
∥ a ∥ H s ≤ C ∥ χ A ⋆ + a ∥ H s − 1 + C ∥ a ∥ H s 2 . \lVert a\rVert_{H^s}\le C\lVert\chi_{A_\star+a}\rVert_{H^{s-1}}+C\lVert a\rVert_{H^s}^2. ∥ a ∥ H s ≤ C ∥ χ A ⋆ + a ∥ H s − 1 + C ∥ a ∥ H s 2 .
Choose the ball in the Sobolev space H s H^s H s so that C ∥ a ∥ H s ≤ 1 2 C\lVert a\rVert_{H^s}\le\frac{1}{2} C ∥ a ∥ H s ≤ 2 1 . The last term is then at most 1 2 ∥ a ∥ H s \frac{1}{2}\lVert a\rVert_{H^s} 2 1 ∥ a ∥ H s and can be moved to the left-hand side. This proves the first estimate. At order one, the quadratic term is bounded in L 2 L^2 L 2 by C ∥ a ∥ C 0 ∥ a ∥ L 2 C\lVert a\rVert_{C^0}\lVert a\rVert_{L^2} C ∥ a ∥ C 0 ∥ a ∥ L 2 . Since ∥ a ∥ L 2 ≤ ∥ a ∥ H 1 \lVert a\rVert_{L^2}\le\lVert a\rVert_{H^1} ∥ a ∥ L 2 ≤ ∥ a ∥ H 1 , sufficiently small C 0 C^0 C 0 norm gives the same absorption and proves the second estimate.
Remark 4.14. The gauge condition is essential in the proof of Lemma 4.13 . A nontrivial pure gauge change can have zero defect and a nonzero connection difference. Lemma 4.13 gives a local comparison between the instanton defect and the distance from the reference instanton after gauge fixing. We record it for this geometric interpretation. The dynamical proof below instead estimates the connection difference directly and does not use this lemma.
For the evolution argument, the reference instanton is allowed to move with the ambient 4-form. We subtract its equation before estimating the perturbation. This removes the forcing that would otherwise be present even when the difference of the two connections is zero.
Lemma 4.15. Suppose B t B_t B t is a smooth family of ψ t \psi_t ψ t -instantons satisfying ∂ t B t = U B t , ψ t ( Q t ) \partial_tB_t=U_{B_t,\psi_t}(Q_t) ∂ t B t = U B t , ψ t ( Q t ) . For A ^ = B t + a \widehat A=B_t+a A = B t + a , add the gauge term − d A ^ d B t ⋆ a -d_{\widehat A}d_{B_t}^\star a − d A d B t ⋆ a . Then a a a satisfies
∂ t a = − L t a + N t ( a ) + U B t + a , ψ t ( Q t ) − U B t , ψ t ( Q t ) , where L t = P B t , t 2 + d B t d B t ⋆ . (76) \partial_ta=-\mathcal{L}_ta+N_t(a)+U_{B_t+a,\psi_t}(Q_t)-U_{B_t,\psi_t}(Q_t),\qquad\text{where }\mathcal{L}_t=\mathcal{P}_{B_t,t}^2+d_{B_t}d_{B_t}^\star. \tag*{(76)} ∂ t a = − L t a + N t ( a ) + U B t + a , ψ t ( Q t ) − U B t , ψ t ( Q t ) , where L t = P B t , t 2 + d B t d B t ⋆ . (76) On a small H s H^s H s ball, s ≥ 8 s\ge8 s ≥ 8 , we have
∥ N t ( a ) ∥ H s − 1 ≤ C ∥ a ∥ H s 2 . (77) \lVert N_t(a)\rVert_{H^{s-1}}\le C\lVert a\rVert_{H^s}^2. \tag*{(77)} ∥ N t ( a ) ∥ H s − 1 ≤ C ∥ a ∥ H s 2 . (77) Moreover, if
∥ ψ t − ψ ⋆ ∥ C s + 2 + ∥ B t − A ⋆ ∥ C s + 1 + ∥ Q t ∥ C s + 1 ≤ ϵ , (78) \lVert\psi_t-\psi_\star\rVert_{C^{s+2}}+\lVert B_t-A_\star\rVert_{C^{s+1}}+\lVert Q_t\rVert_{C^{s+1}}\le\epsilon, \tag*{(78)} ∥ ψ t − ψ ⋆ ∥ C s + 2 + ∥ B t − A ⋆ ∥ C s + 1 + ∥ Q t ∥ C s + 1 ≤ ϵ , (78) then
∥ ( L t − L ⋆ ) a ∥ H s − 1 ≤ C ϵ ∥ a ∥ H s + 1 , ∥ U B t + a , ψ t ( Q t ) − U B t , ψ t ( Q t ) ∥ H s − 1 ≤ C ϵ ∥ a ∥ H s . (79) \begin{aligned}
\lVert(\mathcal{L}_t-\mathcal{L}_\star)a\rVert_{H^{s-1}}&\le C\epsilon\lVert a\rVert_{H^{s+1}},\\
\lVert U_{B_t+a,\psi_t}(Q_t)-U_{B_t,\psi_t}(Q_t)\rVert_{H^{s-1}}&\le C\epsilon\lVert a\rVert_{H^s}. \tag*{(79)}
\end{aligned} ∥( L t − L ⋆ ) a ∥ H s − 1 ∥ U B t + a , ψ t ( Q t ) − U B t , ψ t ( Q t ) ∥ H s − 1 ≤ C ϵ ∥ a ∥ H s + 1 , ≤ C ϵ ∥ a ∥ H s . (79)
Proof. We first expand the equation and identify the terms that are at least quadratic in a a a . Put C t ( a , v ) = ∗ t ( [ a ∧ v ] ∧ ψ t ) C_t(a,v)=*_{t}([a\wedge v]\wedge\psi_t) C t ( a , v ) = ∗ t ([ a ∧ v ] ∧ ψ t ) and n t ( a ) = 1 2 ∗ t ( [ a ∧ a ] ∧ ψ t ) n_t(a)=\frac{1}{2}*_{t}([a\wedge a]\wedge\psi_t) n t ( a ) = 2 1 ∗ t ([ a ∧ a ] ∧ ψ t ) . Since B t B_t B t is an instanton, the curvature expansion gives χ B t + a , ψ t = P B t , t a + n t ( a ) \chi_{B_t+a,\psi_t}=P_{B_t,t}a+n_t(a) χ B t + a , ψ t = P B t , t a + n t ( a ) , and P B t + a , t = P B t , t + C t ( a , ⋅ ) P_{B_t+a,t}=P_{B_t,t}+C_t(a,\mathord\cdot) P B t + a , t = P B t , t + C t ( a , ⋅ ) . The gauge term is − d B t d B t ∗ a − [ a , d B t ∗ a ] -d_{B_t}d_{B_t}^{*}a-[a,d_{B_t}^{*}a] − d B t d B t ∗ a − [ a , d B t ∗ a ] . The linear terms combine to − L t a -\mathcal{L}_t a − L t a . Subtracting ∂ t B t = U B t , ψ t ( Q t ) \partial_t B_t=\mathcal{U}_{B_t,\psi_t}(Q_t) ∂ t B t = U B t , ψ t ( Q t ) leaves the difference of the two corrections in (76) . The remaining local terms are exactly
N t ( a ) = − P B t , t n t ( a ) − C t ( a , P B t , t a ) − C t ( a , n t ( a ) ) − [ a , d B t ∗ a ] . N_t(a)=-P_{B_t,t}n_t(a)-C_t(a,P_{B_t,t}a)-C_t(a,n_t(a))-[a,d_{B_t}^{*}a]. N t ( a ) = − P B t , t n t ( a ) − C t ( a , P B t , t a ) − C t ( a , n t ( a )) − [ a , d B t ∗ a ] .
Every summand has either two factors of a a a and at most one derivative, or three factors of a a a and no derivatives. Because H s − 1 H^{s-1} H s − 1 is an algebra, the quadratic terms are bounded by C ∥ a ∥ H s 2 C\lVert a\rVert_{H^s}^2 C ∥ a ∥ H s 2 . On the chosen small ball the cubic term obeys the same bound. This proves (77) .
We next compare the second-order operators. In fixed coordinate and bundle charts, write
( L t − L ∗ ) a = ∑ ∣ β ∣ ≤ 2 c β ( t , x ) ∂ β a . (\mathcal{L}_t-\mathcal{L}_*)a=\sum_{|\beta|\leq2}c_\beta(t,x)\partial^\beta a. ( L t − L ∗ ) a = ∣ β ∣ ≤ 2 ∑ c β ( t , x ) ∂ β a .
The coefficients are smooth functions of the background forms, connections, and their derivatives. Each coefficient vanishes when ( ψ t , B t ) = ( ψ ∗ , A ∗ ) (\psi_t,B_t)=(\psi_*,A_*) ( ψ t , B t ) = ( ψ ∗ , A ∗ ) . The derivative orders in (78) therefore bound their C s − 1 C^{s-1} C s − 1 norms by C ϵ C\epsilon C ϵ . When we apply up to s − 1 s-1 s − 1 derivatives to c β ∂ β a c_\beta\partial^\beta a c β ∂ β a , Leibniz’s rule puts at most s − 1 s-1 s − 1 derivatives on the coefficient and at most s + 1 s+1 s + 1 derivatives on a a a . Summing these product bounds gives C ϵ ∥ a ∥ H s + 1 C\epsilon\lVert a\rVert_{H^{s+1}} C ϵ ∥ a ∥ H s + 1 , which is the first estimate in (79) .
For the correction, put w A = − D A , ψ − 1 ( 0 , ∗ ( F A ∧ Q ) ) w_A=-D_{A,\psi}^{-1}(0,* (F_A\wedge Q)) w A = − D A , ψ − 1 ( 0 , ∗ ( F A ∧ Q )) . Subtracting the equations at B + a B+a B + a and B B B gives
D B + a , ψ ( w B + a − w B ) = − ( D B + a , ψ − D B , ψ ) w B − ( 0 , ∗ ( ( d B a + 1 2 [ a ∧ a ] ) ∧ Q ) ) , (80) D_{B+a,\psi}(w_{B+a}-w_B)=-(D_{B+a,\psi}-D_{B,\psi})w_B-(0,*((d_Ba+\tfrac{1}{2}[a\wedge a])\wedge Q)), \tag*{(80)} D B + a , ψ ( w B + a − w B ) = − ( D B + a , ψ − D B , ψ ) w B − ( 0 , ∗ (( d B a + 2 1 [ a ∧ a ]) ∧ Q )) , (80)
and estimates of Lemma 4.6 give the second line of (79) .
We will also need estimates at orders above s s s without assuming smallness of the higher background norms. For every fixed r ≥ s r\geq s r ≥ s , the same product estimates and (80) give
∥ N t ( a ) ∥ H r − 1 + ∥ U B t + a , ψ t ( Q t ) − U B t , ψ t ( Q t ) ∥ H r − 1 ≤ C r ∥ a ∥ H r (81) \lVert N_t(a)\rVert_{H^{r-1}}+\lVert\mathcal{U}_{B_t+a,\psi_t}(Q_t)-\mathcal{U}_{B_t,\psi_t}(Q_t)\rVert_{H^{r-1}}\leq C_r\lVert a\rVert_{H^r} \tag*{(81)} ∥ N t ( a ) ∥ H r − 1 + ∥ U B t + a , ψ t ( Q t ) − U B t , ψ t ( Q t ) ∥ H r − 1 ≤ C r ∥ a ∥ H r (81)
on the fixed small H s H^s H s ball. Here the constants may depend on bounded higher derivatives of the reference family; these derivatives need not be small. We explain why the highest norm of a a a occurs only linearly. The low-order inverse first bounds w B + a − w B w_{B+a}-w_B w B + a − w B in terms of ∥ a ∥ H s \lVert a\rVert_{H^s} ∥ a ∥ H s . Moving the zeroth-order difference D B + a − D B D_{B+a}-D_B D B + a − D B to the right lets us apply higher elliptic estimates successively, as in the proof of (62) . In the product terms, a highest derivative of a a a is multiplied by a controlled low-order factor. The other terms have their highest derivative on the smooth reference coefficients or on a lower-order unknown already controlled by the induction. The same product estimates apply to N t ( a ) N_t(a) N t ( a ) . This proves (81) with constants depending on the fixed low-order ball and the higher coefficient bounds. Every term still contains a factor of a a a , so no constant forcing term appears in this estimate.
Theorem 4.16 (Attraction to a moving acyclic instanton). Let ψ t \psi_t ψ t and B t B_t B t be smooth for all t ≥ 0 t\geq0 t ≥ 0 , with ψ t \psi_t ψ t closed and positive, B t B_t B t an instanton, and ∂ t B t = U B t , ψ t ( Q t ) \partial_tB_t=\mathcal{U}_{B_t,\psi_t}(Q_t) ∂ t B t = U B t , ψ t ( Q t ) . Assume that ψ t \psi_t ψ t , B t − A ∗ B_t-A_* B t − A ∗ and Q t Q_t Q t are uniformly bounded in every spatial C k C^k C k norm. For some sufficiently small ϵ > 0 \epsilon>0 ϵ > 0 , suppose (78) holds for all t t t . If the smooth initial connection A 0 A_0 A 0 is sufficiently close to B 0 B_0 B 0 in H s H^s H s , s ≥ 8 s\geq8 s ≥ 8 , then (9) has a unique global smooth solution. Moreover, there are smooth gauge transformations g t g_t g t , with g 0 = 1 g_0=1 g 0 = 1 , such that
∥ g t − 1 ∗ A t − B t ∥ H s ≤ C e − γ t ∥ A 0 − B 0 ∥ H s (82) \lVert g_t^{-1*}A_t-B_t\rVert_{H^s}\leq Ce^{-\gamma t}\lVert A_0-B_0\rVert_{H^s} \tag*{(82)} ∥ g t − 1 ∗ A t − B t ∥ H s ≤ C e − γ t ∥ A 0 − B 0 ∥ H s (82) for some γ > 0 \gamma>0 γ > 0 . The difference decays exponentially in every C k C^k C k norm after positive time, and g t g_t g t converges smoothly to a gauge transformation g ∞ g_\infty g ∞ . If B t B_t B t converges smoothly and exponentially to B ∞ B_\infty B ∞ , then A t A_t A t converges smoothly and exponentially to g ∞ ∗ B ∞ g_\infty^*B_\infty g ∞ ∗ B ∞ .
Proof. We first solve (76) for a short time with a ( 0 ) = A 0 − B 0 a(0)=A_0-B_0 a ( 0 ) = A 0 − B 0 . The parabolic construction in Theorem 4.9 applies as replacing a fixed reference connection by the prescribed smooth B t B_t B t changes only the coefficients. We now estimate this solution while it remains in a small H s H^s H s ball, and then prove that it cannot leave that ball. We use the fixed norm (75) , so differentiating the norm produces no metric-variation terms. The useful derivative pairing is
∣ ⟨ a , b ⟩ s ∣ = ∣ ⟨ L ⋆ ( s + 1 ) / 2 a , L ⋆ ( s − 1 ) / 2 b ⟩ L 2 ∣ ≤ ∣ a ∣ s + 1 ∣ b ∣ s − 1 \left|\langle a,b\rangle_s\right|=\left|\left\langle\mathcal{L}_\star^{(s+1)/2}a,\mathcal{L}_\star^{(s-1)/2}b\right\rangle_{L^2}\right|\leq|a|_{s+1}|b|_{s-1} ∣ ⟨ a , b ⟩ s ∣ = ⟨ L ⋆ ( s + 1 ) /2 a , L ⋆ ( s − 1 ) /2 b ⟩ L 2 ≤ ∣ a ∣ s + 1 ∣ b ∣ s − 1
This identity follows by moving powers of the fixed self-adjoint operator L ⋆ \mathcal{L}_\star L ⋆ between the two factors. It allows us to pair a term controlled in H s − 1 H^{s-1} H s − 1 with the dissipative quantity ∣ a ∣ s + 1 |a|_{s+1} ∣ a ∣ s + 1 . In particular, the second-order operator difference is bounded by C ϵ ∣ a ∣ s + 1 C\epsilon|a|_{s+1} C ϵ ∣ a ∣ s + 1 , the local nonlinearity by C ∣ a ∣ s ∣ a ∣ s + 1 C|a|_s|a|_{s+1} C ∣ a ∣ s ∣ a ∣ s + 1 , and the correction difference by C ϵ ∣ a ∣ s ∣ a ∣ s + 1 C\epsilon|a|_s|a|_{s+1} C ϵ ∣ a ∣ s ∣ a ∣ s + 1 . Thus, we obtain from Lemma 4.15
1 2 d d t ∣ a ∣ s 2 ≤ − ∣ a ∣ s + 1 2 + C ϵ ∣ a ∣ s + 1 2 + C ∣ a ∣ s 2 ∣ a ∣ s + 1 + C ϵ ∣ a ∣ s ∣ a ∣ s + 1 ≤ − ( 1 − C 1 ϵ − C 2 ∣ a ∣ s ) ∣ a ∣ s + 1 2 . (83) \begin{aligned}
\frac{1}{2}\frac{d}{dt}|a|_s^2\leq-|a|_{s+1}^2+C\epsilon|a|_{s+1}^2+C|a|_s^2|a|_{s+1}+C\epsilon|a|_s|a|_{s+1} \\
&\leq-\left(1-C_1\epsilon-C_2|a|_s\right)|a|_{s+1}^2. \tag*{(83)}
\end{aligned} 2 1 d t d ∣ a ∣ s 2 ≤ − ∣ a ∣ s + 1 2 + C ϵ ∣ a ∣ s + 1 2 + C ∣ a ∣ s 2 ∣ a ∣ s + 1 + C ϵ ∣ a ∣ s ∣ a ∣ s + 1 ≤ − ( 1 − C 1 ϵ − C 2 ∣ a ∣ s ) ∣ a ∣ s + 1 2 . (83)
where in the last line we used ∣ a ∣ s ≤ ν − 1 / 2 ∣ a ∣ s + 1 |a|_s\leq\nu^{-1/2}|a|_{s+1} ∣ a ∣ s ≤ ν − 1/2 ∣ a ∣ s + 1 . Choose a radius ρ \rho ρ inside the neighborhood of definition of the correction and choose ϵ \epsilon ϵ so that C 1 ϵ + C 2 ρ < 1 2 C_1\epsilon+C_2\rho<\frac{1}{2} C 1 ϵ + C 2 ρ < 2 1 . Take ∣ a ( 0 ) ∣ s < ρ |a(0)|_s<\rho ∣ a ( 0 ) ∣ s < ρ . As long as ∣ a ( t ) ∣ s < ρ |a(t)|_s<\rho ∣ a ( t ) ∣ s < ρ , the preceding inequality implies
1 2 d d t ∣ a ∣ s 2 ≤ − 1 2 ∣ a ∣ s + 1 2 ≤ − ν 2 ∣ a ∣ s 2 . \frac{1}{2}\frac{d}{dt}|a|_s^2\leq-\frac{1}{2}|a|_{s+1}^2\leq-\frac{\nu}{2}|a|_s^2. 2 1 d t d ∣ a ∣ s 2 ≤ − 2 1 ∣ a ∣ s + 1 2 ≤ − 2 ν ∣ a ∣ s 2 .
Integration then gives ∣ a ( t ) ∣ s ≤ e − ν t / 2 ∣ a ( 0 ) ∣ s < ρ |a(t)|_s\leq e^{-\nu t/2}|a(0)|_s<\rho ∣ a ( t ) ∣ s ≤ e − ν t /2 ∣ a ( 0 ) ∣ s < ρ . If a first exit time existed, continuity would give ∣ a ∣ s = ρ |a|_s=\rho ∣ a ∣ s = ρ there, which contradicts the bound ∣ a ( t ) ∣ s < ρ |a(t)|_s<\rho ∣ a ( t ) ∣ s < ρ . Thus the small ball is preserved throughout the existence interval. This argument proves the required smallness.
Together with the assumed bound on B t − A ⋆ B_t-A_\star B t − A ⋆ , this keeps the gauge-fixed connection in a fixed neighborhood of A ⋆ A_\star A ⋆ where D D D has a uniformly bounded inverse. The local parabolic argument then has a uniform positive restart time on each finite time interval. If the maximal time were finite, we could restart sufficiently close to it and continue beyond it. This contradiction proves global existence. The same energy estimate gives exponential decay in H s H^s H s .
The higher-order estimates are obtained in the same as we obtained (70) and we get
y r ′ ( t ) + c r y r + 1 ( t ) ≤ C r y r ( t ) , r ≥ s , (84) y_r'(t)+c_ry_{r+1}(t)\leq C_ry_r(t),\qquad r\geq s, \tag*{(84)} y r ′ ( t ) + c r y r + 1 ( t ) ≤ C r y r ( t ) , r ≥ s , (84)
where c r > 0 c_r>0 c r > 0 and the constants are independent of t t t .
The right-hand side of this inequality is not negative, so applying Gronwall at order r r r alone does not prove decay. We instead use the negative term to gain one derivative at a time on shorter time intervals, starting from the decay already proved in H s H^s H s .
Fix an interval [ u , u + 1 ] [u,u+1] [ u , u + 1 ] and let M u = sup u ≤ q ≤ u + 1 y s ( q ) M_u=\sup_{u\leq q\leq u+1}y_s(q) M u = sup u ≤ q ≤ u + 1 y s ( q ) . Integration of (84) at r = s r=s r = s gives
∫ u u + 1 y s + 1 ( q ) d q ≤ C M u . \int_u^{u+1}y_{s+1}(q)\,dq\leq CM_u. ∫ u u + 1 y s + 1 ( q ) d q ≤ C M u .
Choose a smooth nondecreasing time cutoff η \eta η which vanishes at u u u and is one on [ u + 1 / 2 , u + 1 ] [u+1/2,u+1] [ u + 1/2 , u + 1 ] . Multiplication of the inequality at order s + 1 s+1 s + 1 by η \eta η gives
( η y s + 1 ) ′ + c s + 1 η y s + 2 ≤ ( C s + 1 η + η ′ ) y s + 1 . (\eta y_{s+1})'+c_{s+1}\eta y_{s+2}\leq(C_{s+1}\eta+\eta')y_{s+1}. ( η y s + 1 ) ′ + c s + 1 η y s + 2 ≤ ( C s + 1 η + η ′ ) y s + 1 .
Integrating this inequality first from u u u to any time in [ u + 1 / 2 , u + 1 ] [u+1/2,u+1] [ u + 1/2 , u + 1 ] allows us to bound y s + 1 y_{s+1} y s + 1 there, and then we can integrate to u + 1 u+1 u + 1 to bound the integral of y s + 2 y_{s+2} y s + 2 on that final half interval. The initial term vanishes because η ( u ) = 0 \eta(u)=0 η ( u ) = 0 . Both terms on the right are bounded by the preceding integral estimate, so
sup u + 1 / 2 ≤ q ≤ u + 1 y s + 1 ( q ) + ∫ u + 1 / 2 u + 1 y s + 2 ( q ) d q ≤ C M u . \sup_{u+1/2\leq q\leq u+1}y_{s+1}(q)+\int_{u+1/2}^{u+1}y_{s+2}(q)\,dq\leq CM_u. u + 1/2 ≤ q ≤ u + 1 sup y s + 1 ( q ) + ∫ u + 1/2 u + 1 y s + 2 ( q ) d q ≤ C M u .
For the next derivative order, choose a cutoff supported in the interval where we have just controlled the integral of y s + 2 y_{s+2} y s + 2 . Continue with cutoffs which become one at u + 1 − 2 − j u+1-2^{-j} u + 1 − 2 − j , for j = 2 , 3 , … j=2,3,\ldots j = 2 , 3 , … . Their derivatives have larger bounds as j j j increases, but only finitely many steps are needed for any fixed r r r . At each step the integral obtained at the preceding order bounds the new cutoff term. We therefore obtain control of the r r r th norm at the final endpoint from the supremum of the s s s th norm on the full interval. Taking $u=t-1 gives
∥ a ( t ) ∥ H r ≤ C r sup t − 1 ≤ q ≤ t ∥ a ( q ) ∥ H s ≤ C r e − γ t ∥ a ( 0 ) ∥ H s , t ≥ 1. (85) \|a(t)\|_{H^r}\leq C_r\sup_{t-1\leq q\leq t}\|a(q)\|_{H^s}\leq C_re^{-\gamma t}\|a(0)\|_{H^s},\qquad t\geq1. \tag*{(85)} ∥ a ( t ) ∥ H r ≤ C r t − 1 ≤ q ≤ t sup ∥ a ( q ) ∥ H s ≤ C r e − γ t ∥ a ( 0 ) ∥ H s , t ≥ 1. (85)
Because the energy norms are fixed, this argument does not differentiate the high-order background coefficients in time. Sobolev embedding now gives exponential decay in every spatial C k C^k C k norm. On the initial bounded time interval, smooth initial data and (84) give finite higher norms. Thus we have one smooth solution for all time, even though only the initial H s H^s H s norm was required to be small.
It remains to remove the gauge term and identify the limit in the original equation. Set A ^ = B t + a \widehat{A}=B_t+a A = B t + a and ξ t = d B t ∗ a \xi_t=d_{B_t}^*a ξ t = d B t ∗ a . Let g t g_t g t solve g ˙ t g t − 1 = ξ t \dot{g}_t g_t^{-1}=\xi_t g ˙ t g t − 1 = ξ t , g 0 = 1 g_0=1 g 0 = 1 , and put A t = g t ∗ A ^ t A_t=g_t^*\widehat{A}_t A t = g t ∗ A t . The derivative of this time-dependent gauge transformation adds precisely the term that was subtracted to make the equation parabolic
∂ t A t = g t − 1 ( ∂ t A ^ t + d A ^ t ξ t ) g t = V t ( A t ) . \partial_t A_t=g_t^{-1}\left(\partial_t\widehat{A}_t+d_{\widehat{A}_t}\xi_t\right)g_t=\mathcal{V}_t(A_t). ∂ t A t = g t − 1 ( ∂ t A t + d A t ξ t ) g t = V t ( A t ) .
so this recovers the original equation with the same initial value. By (85) , ξ t = d B t ∗ a \xi_t=d_{B_t}^*a ξ t = d B t ∗ a decays exponentially in every spatial norm. The gauge equation is global
∥ g t − g ∞ ∥ C k ≤ C k ∫ t ∞ ∥ ξ q ∥ C k d q ≤ C k e − γ t , t ≥ 1. \lVert g_t-g_\infty\rVert_{C^k}\leq C_k\int_t^\infty\lVert\xi_q\rVert_{C^k}\,dq\leq C_ke^{-\gamma t},\qquad t\geq1. ∥ g t − g ∞ ∥ C k ≤ C k ∫ t ∞ ∥ ξ q ∥ C k d q ≤ C k e − γ t , t ≥ 1.
giving exponential convergence. Since g t − 1 ∗ A t = B t + a t g_t^{-1*}A_t=B_t+a_t g t − 1 ∗ A t = B t + a t , the estimate for a t a_t a t proves (4.40). If B t B_t B t converges exponentially to B ∞ B_\infty B ∞ , continuity of the gauge action in the corresponding higher norms gives A t → g ∞ ∗ B ∞ A_t\to g_\infty^*B_\infty A t → g ∞ ∗ B ∞ with the same type of decay. The limiting representative need not be B ∞ B_\infty B ∞ itself. Uniqueness follows from the smooth uniqueness for the original equation already proved.
We can finally give a proof of our Theorem 1.6.
Proof of Theorem 1.6. We apply Theorem 4.16 to the constant background
ψ t = ψ ⋆ , B t = A ⋆ , for all t ≥ 0. \psi_t=\psi_\star,\qquad B_t=A_\star,\qquad\text{for all }t\geq0. ψ t = ψ ⋆ , B t = A ⋆ , for all t ≥ 0.
Then Q t = ∂ t ψ t = 0 Q_t=\partial_t\psi_t=0 Q t = ∂ t ψ t = 0 . Since A ⋆ A_\star A ⋆ is an acyclic ψ ⋆ \psi_\star ψ ⋆ -instanton, each B t B_t B t is an acyclic ψ t \psi_t ψ t -instanton. Moreover,
∂ t B t = 0 = U A ⋆ , ψ ⋆ ( 0 ) , \partial_tB_t=0=\mathcal{U}_{A_\star,\psi_\star}(0), ∂ t B t = 0 = U A ⋆ , ψ ⋆ ( 0 ) ,
so the required evolution equation for the reference connection is satisfied. The left-hand side of (78) is zero, and all the required spatial bounds are uniform because the background is smooth and independent of time. Since U A , ψ ⋆ ( 0 ) = 0 \mathcal{U}_{A,\psi_\star}(0)=0 U A , ψ ⋆ ( 0 ) = 0 , equation (9) reduces to
∂ t A = − d A ∗ F A + ∗ ( F A ∧ d φ ⋆ ) . \partial_tA=-d_A^*F_A+*(F_A\wedge d\varphi_\star). ∂ t A = − d A ∗ F A + ∗ ( F A ∧ d φ ⋆ ) .
Theorem 4.16 therefore gives a unique global smooth solution for every sufficiently small smooth H s H^s H s perturbation of A ⋆ A_\star A ⋆ . Since B t = A ⋆ B_t=A_\star B t = A ⋆ for all time, its limiting connection is B ∞ = A ⋆ B_\infty=A_\star B ∞ = A ⋆ . The convergence conclusion of that theorem gives smooth exponential convergence of A t A_t A t to g ∞ ∗ A ⋆ g_\infty^*A_\star g ∞ ∗ A ⋆ .
Remark 4.17. Only d ψ ⋆ = 0 d\psi_\star=0 d ψ ⋆ = 0 is required in the above proof we have not assumed d φ ⋆ = 0 d\varphi_\star=0 d φ ⋆ = 0 . Thus the background need not be torsion-free.
Modified Laplacian coflow
We now take the ambient family to solve Grigorian’s modified Laplacian coflow [12 ] . Since its equation is independent of the connection, we first solve the ambient problem and then solve the connection equation along that solution. We explain the conversion of conventions and the ambient estimates used below.
We first note that [12 ] , eq. (2.2) uses ( u ⌟ φ ) ∧ ( v ⌟ φ ) = 6 g ( u , v ) vol + (u\lrcorner\varphi)\wedge(v\lrcorner\varphi)=6g(u,v)\operatorname{vol}_+ ( u ┘ φ ) ∧ ( v ┘ φ ) = 6 g ( u , v ) vol + , whereas (14) has the opposite sign. Keeping φ \varphi φ and g g g fixed therefore gives vol + = − vol − \operatorname{vol}_+=-\operatorname{vol}_- vol + = − vol − , ∗ + = − ∗ − *_+=-*- ∗ + = − ∗ − and ψ + = − ψ \psi_+=-\psi ψ + = − ψ . Equation (2.10) of [12 ] defines the full torsion by ∇ a φ b c d = ( T + ) e a ( ψ + ) e b c d \nabla_a\varphi_{bcd}=(T_+)^e{}_a(\psi_+)_{ebcd} ∇ a φ b c d = ( T + ) e a ( ψ + ) e b c d , with no extra numerical factor. Our definition consequently gives T + = − T T_+=-T T + = − T , and hence tr T + = − tr T \operatorname{tr}T_+=-\operatorname{tr}T tr T + = − tr T . Also, [12 ] , eq. (1.4)
∂ t ψ + = Δ ψ + + 2 d ( ( c − tr T + ) φ ) , c > 0 , \partial_t\psi_+=\Delta\psi_++2d\bigl((c-\operatorname{tr}T_+)\varphi\bigr),\qquad c>0, ∂ t ψ + = Δ ψ + + 2 d ( ( c − tr T + ) φ ) , c > 0 ,
where c c c is his positive constant A A A . We negate the equation and substitute ψ + = − ψ \psi_+=-\psi ψ + = − ψ and tr T + = − tr T \operatorname{tr}T_+=-\operatorname{tr}T tr T + = − tr T . The Hodge Laplacian is unchanged by orientation reversal, since the two changes of sign in its star operators cancel. We therefore obtain exactly (11) . The same algebraic conversion applies when $c=0`.
Proof of Corollary 1.7. Grigorian introduces a positive parameter A A A (which we denote by c c c here) in the modification of the Laplacian coflow, but the local existence argument also applies to every fixed c ∈ R c \in\mathbb{R} c ∈ R . Indeed, the parameter occurs only in the first-order term − 2 c d φ -2c\,d\varphi − 2 c d φ , so it does not change the second-order symbol of the gauge-fixed equation, see also the gauge-fixing argument in [12 Theorems 5.1 and 6.9] and [BV20, Introduction]. Thus the short-time theory used here includes both c = 0 c=0 c = 0 and negative values of c c c .
The velocity of the flow is Q = Q c ( ψ ) Q=Q_c(\psi) Q = Q c ( ψ ) , which is exact, since both terms defining Q c Q_c Q c are exterior derivatives on closed forms. In particular, d Q = 0 dQ=0 d Q = 0 . Theorem 1.5 therefore gives preservation of the instanton equation for every existing smooth coupled solution while the correction is defined.
To construct the coupled solution, first solve the ambient equation and then apply Theorem 4.8 to this prescribed family, starting from the given acyclic instanton. We shorten the common interval if necessary so that D \mathcal{D} D remains invertible. Uniqueness of the ambient solution and uniqueness of the connection evolution give uniqueness of the coupled solution.
For a prescribed smooth solution of unmodified coflow, the same preservation calculation applies with Q = Δ ψ ψ Q=\Delta_\psi\psi Q = Δ ψ ψ . However, the well-posedness of the unmodified Laplacian coflow is still an open problem. For the global result below, we use modified coflow with c = 0 c=0 c = 0 . The next lemma gives the quantitative estimates needed for the global coupled argument. It uses the gauge-fixed equation of [4 ] and controls the solution and its total motion using one finite-order initial norm.
Lemma 4.18. Let ψ ˉ \bar{\psi} ψ ˉ be a torsion-free structure on compact M M M . Fix an integer r ≥ 0 r \ge0 r ≥ 0 and 0 < α < 1 0 < \alpha< 1 0 < α < 1 , and put N = r + 8 N=r+8 N = r + 8 . There are constants δ , C , σ > 0 \delta,C,\sigma>0 δ , C , σ > 0 with the following property. If ψ 0 \psi_0 ψ 0 is smooth, closed and positive, [ ψ 0 ] = [ ψ ˉ ] [\psi_0]=[\bar{\psi}] [ ψ 0 ] = [ ψ ˉ ] , and ∥ ψ 0 − ψ ˉ ∥ C N , α < δ \|\psi_0-\bar{\psi}\|_{C^{N,\alpha}}<\delta ∥ ψ 0 − ψ ˉ ∥ C N , α < δ , then parameter-zero modified coflow has a unique global smooth solution and
sup t ≥ 0 ∥ ψ t − ψ ˉ ∥ C r , α + sup t ≥ 0 e σ t ∥ ∂ t ψ t ∥ C r , α + ∫ 0 ∞ ∥ ∂ t ψ t ∥ C r , α d t ≤ C ∥ ψ 0 − ψ ˉ ∥ C N , α . (86) \sup_{t\geq0}\|\psi_t-\bar{\psi}\|_{C^{r,\alpha}}+\sup_{t\geq0}e^{\sigma t}\|\partial_t\psi_t\|_{C^{r,\alpha}}+\int_0^\infty\|\partial_t\psi_t\|_{C^{r,\alpha}}\,dt\leq C\|\psi_0-\bar{\psi}\|_{C^{N,\alpha}}. \tag*{(86)} t ≥ 0 sup ∥ ψ t − ψ ˉ ∥ C r , α + t ≥ 0 sup e σ t ∥ ∂ t ψ t ∥ C r , α + ∫ 0 ∞ ∥ ∂ t ψ t ∥ C r , α d t ≤ C ∥ ψ 0 − ψ ˉ ∥ C N , α . (86) The solution converges smoothly and exponentially to F ∞ ∗ ψ ˉ F_\infty^*\bar{\psi} F ∞ ∗ ψ ˉ for a diffeomorphism F ∞ F_\infty F ∞ isotopic to the identity. Every higher spatial norm of the solution is uniformly bounded, and every higher spatial norm of its velocity has an exponentially decaying bound.
Proof. We use the gauge fixing in [4 Lemma 3.3] , translated to our orientation convention as above, and take the background connection in its DeTurck vector field to be the Levi–Civita connection of ψ ˉ \bar{\psi} ψ ˉ . Write ∇ ˉ \bar{\nabla} ∇ ˉ and Δ ˉ \bar{\Delta} Δ ˉ for this connection and its nonnegative Hodge Laplacian. If ψ ˉ = ψ ˉ + h \bar{\psi}=\bar{\psi}+h ψ ˉ = ψ ˉ + h is closed, the gauge-fixed equation is
∂ t h = − Δ ˉ h + d ( R ( h ) ∇ ˉ h ) , R ( 0 ) = 0. \partial_t h=-\bar{\Delta}h+d\bigl(\mathcal{R}(h)\bar{\nabla}h\bigr), \qquad\mathcal{R}(0)=0. ∂ t h = − Δ ˉ h + d ( R ( h ) ∇ ˉ h ) , R ( 0 ) = 0.
Here R ( h ) \mathcal{R}(h) R ( h ) is a smooth bundle homomorphism depending pointwise on h h h . We explain why the remainder has this form, since a pointwise quadratic bound alone would not suffice for the estimates below. On closed four forms, the gauge-fixed velocity is the exterior derivative of
∗ ψ d ψ − 2 ( tr T ψ ) φ + V ( φ ) ψ , φ = ∗ ψ ψ . *_{\psi}d\psi-2(\operatorname{tr}T_{\psi})\varphi+V(\varphi)\,\psi, \qquad\varphi=*_{\psi}\psi. ∗ ψ d ψ − 2 ( tr T ψ ) φ + V ( φ ) ψ , φ = ∗ ψ ψ .
Each summand is linear in ∇ ˉ h \bar{\nabla}h ∇ ˉ h , with coefficients depending smoothly on h h h : the reference structure is parallel, the Hodge star and φ \varphi φ depend algebraically on ψ \psi ψ , and V V V is linear in the difference of the two Levi–Civita connections. Subtract the linearization of this first-order expression at h = 0 h=0 h = 0 . The remainder is R ( h ) ∇ ˉ h \mathcal{R}(h)\bar{\nabla}h R ( h ) ∇ ˉ h , with R ( 0 ) = 0 \mathcal{R}(0)=0 R ( 0 ) = 0 . The exterior derivative of its linear part is − Δ ˉ h -\bar{\Delta}h − Δ ˉ h on closed variations by the linearization formula. This also explains the nonlinear expansion used in [4 proof of Theorem 3.4, Step 2] .
For local existence we may regard the above equation as an equation for all four forms. If h h h is small in C 0 C^0 C 0 , its principal symbol is a small perturbation of that of − Δ ˉ -\bar{\Delta} − Δ ˉ , so this extension is uniformly parabolic. The usual local existence theorem for smooth quasilinear parabolic systems gives a smooth solution. It stays closed because
∂ t ( d h ) = − Δ ˉ ( d h ) , d h ( 0 ) = 0. \partial_t(dh)=-\bar{\Delta}(dh), \qquad dh(0)=0. ∂ t ( d h ) = − Δ ˉ ( d h ) , d h ( 0 ) = 0.
Its harmonic part is constant, since both Δ ˉ h \bar{\Delta}h Δ ˉ h and the exact remainder are orthogonal to harmonic forms. The initial perturbation is exact, so the solution remains exact. Consequently this extended equation agrees with the gauge-fixed coflow along the solution we have constructed.
Choose m = r + 8 m=r+8 m = r + 8 . We first use the fixed Sobolev norms
∥ u ∥ H j = ∥ ( 1 + Δ ˉ ) j / 2 u ∥ L 2 . \|u\|_{H^j}=\|(1+\bar{\Delta})^{j/2}u\|_{L^2}. ∥ u ∥ H j = ∥ ( 1 + Δ ˉ ) j /2 u ∥ L 2 .
The same notation is used on three and four forms, with the Hodge Laplacian of the appropriate degree. These norms are equivalent to the usual Sobolev norms. Let λ > 0 \lambda> 0 λ > 0 be the first eigenvalue of Δ ˉ \bar{\Delta} Δ ˉ on exact four forms. Since d d d commutes with the Hodge Laplacian, exactness gives
∥ d ∗ ( 1 + Δ ˉ ) m / 2 h ∥ L 2 2 ≥ λ ∥ h ∥ H m 2 , \lVert d^*(1+\bar{\Delta})^{m/2}h\rVert_{L^2}^2 \geq\lambda\lVert h\rVert_{H^m}^2, ∥ d ∗ ( 1 + Δ ˉ ) m /2 h ∥ L 2 2 ≥ λ ∥ h ∥ H m 2 ,
∥ h ∥ H m + 1 2 = ∥ h ∥ H m 2 + ∥ d ∗ ( 1 + Δ ˉ ) m / 2 h ∥ L 2 2 . \lVert h\rVert_{H^{m+1}}^2 = \lVert h\rVert_{H^m}^2 + \lVert d^*(1+\bar{\Delta})^{m/2}h\rVert_{L^2}^2. ∥ h ∥ H m + 1 2 = ∥ h ∥ H m 2 + ∥ d ∗ ( 1 + Δ ˉ ) m /2 h ∥ L 2 2 .
All adjoints in this proof use the fixed background metric. Because m > 7 / 2 + 1 m > 7/2 + 1 m > 7/2 + 1 , Sobolev multiplication and the chain rule give, on a fixed small H m H^m H m ball,
∥ R ( h ) ∇ ˉ h ∥ H m ≤ C ( ∥ h ∥ C 0 ∥ h ∥ H m + 1 + ∥ h ∥ H m ∥ ∇ ˉ h ∥ C 0 ) ≤ C ∥ h ∥ H m ∥ h ∥ H m + 1 . \begin{aligned}
\lVert\mathcal{R}(h)\bar{\nabla}h\rVert_{H^m} \leq C\left(\lVert h\rVert_{C^0}\lVert h\rVert_{H^{m+1}} + \lVert h\rVert_{H^m}\lVert\bar{\nabla}h\rVert_{C^0}\right) \\
&\leq C\lVert h\rVert_{H^m}\lVert h\rVert_{H^{m+1}}.
\end{aligned} ∥ R ( h ) ∇ ˉ h ∥ H m ≤ C ( ∥ h ∥ C 0 ∥ h ∥ H m + 1 + ∥ h ∥ H m ∥ ∇ ˉ h ∥ C 0 ) ≤ C ∥ h ∥ H m ∥ h ∥ H m + 1 .
Pair the equation with ( 1 + Δ ˉ ) m h (1+\bar{\Delta})^m h ( 1 + Δ ˉ ) m h and integrate d d d by parts. The preceding estimates and exactly the same argument as in the beginning of the proof of Theorem 4.16 proves exponential decay of h ( t ) h(t) h ( t ) and thus the solution cannot leave the ball during its existence interval.
We next establish higher estimates using only this one smallness assumption. For this calculation define the fixed derivative norms
∥ h ∥ j , ∇ ˉ 2 = ∑ ℓ = 0 j ∥ ∇ ˉ ℓ h ∥ L 2 2 . \lVert h\rVert_{j,\bar{\nabla}}^2 = \sum_{\ell=0}^{j}\lVert\bar{\nabla}^{\ell}h\rVert_{L^2}^2. ∥ h ∥ j , ∇ ˉ 2 = ℓ = 0 ∑ j ∥ ∇ ˉ ℓ h ∥ L 2 2 .
They are equivalent to the spectral Sobolev norms used above. We use these derivative norms to keep track of the coefficient of the highest derivatives in the energy estimate. In differentiating the remainder, separate the term with all derivatives on ∇ ˉ h \bar{\nabla}h ∇ ˉ h :
∇ ˉ j ( R ( h ) ∇ ˉ h ) = R ( h ) ∇ ˉ j + 1 h + [ ∇ ˉ j , R ( h ) ] ∇ ˉ h . \bar{\nabla}^{j}\left(\mathcal{R}(h)\bar{\nabla}h\right) = \mathcal{R}(h)\bar{\nabla}^{j+1}h + [\bar{\nabla}^{j},\mathcal{R}(h)]\bar{\nabla}h. ∇ ˉ j ( R ( h ) ∇ ˉ h ) = R ( h ) ∇ ˉ j + 1 h + [ ∇ ˉ j , R ( h )] ∇ ˉ h .
After integration by parts, the first term and its lower-order counterparts contribute at most
C ∥ R ( h ) ∥ C 0 ∑ ℓ = 0 j ∥ ∇ ˉ ℓ + 1 h ∥ L 2 2 . C\lVert\mathcal{R}(h)\rVert_{C^0}\sum_{\ell=0}^{j}\lVert\bar{\nabla}^{\ell+1}h\rVert_{L^2}^2. C ∥ R ( h ) ∥ C 0 ℓ = 0 ∑ j ∥ ∇ ˉ ℓ + 1 h ∥ L 2 2 .
The constant here is independent of j j j . We absorb this sum directly into the same derivative sum supplied by the leading Laplacian term, before using any equivalence with spectral Sobolev norms. Thus the absorption uses the single fixed C 0 C^0 C 0 smallness chosen above. Leibniz’s rule and the Sobolev product estimates give
∥ [ ∇ ˉ j , R ( h ) ] ∇ ˉ h ∥ L 2 ≤ C j ( ∥ ∇ ˉ R ( h ) ∥ C 0 ∥ ∇ ˉ h ∥ H j − 1 + ∥ R ( h ) ∥ H j ∥ ∇ ˉ h ∥ C 0 ) ≤ C j ∥ h ∥ C 1 ∥ h ∥ j , ∇ ˉ . \begin{aligned}
\lVert[\bar{\nabla}^{j},\mathcal{R}(h)]\bar{\nabla}h\rVert_{L^2} \leq C_j\left(\lVert\bar{\nabla}\mathcal{R}(h)\rVert_{C^0}\lVert\bar{\nabla}h\rVert_{H^{j-1}} + \lVert\mathcal{R}(h)\rVert_{H^j}\lVert\bar{\nabla}h\rVert_{C^0}\right) \\
&\leq C_j\lVert h\rVert_{C^1}\lVert h\rVert_{j,\bar{\nabla}}.
\end{aligned} ∥[ ∇ ˉ j , R ( h )] ∇ ˉ h ∥ L 2 ≤ C j ( ∥ ∇ ˉ R ( h ) ∥ C 0 ∥ ∇ ˉ h ∥ H j − 1 + ∥ R ( h ) ∥ H j ∥ ∇ ˉ h ∥ C 0 ) ≤ C j ∥ h ∥ C 1 ∥ h ∥ j , ∇ ˉ .
Again the same proof as for (70) gives
d d t ∥ h ∥ j , ∇ ˉ 2 + c j ∥ h ∥ j + 1 , ∇ ˉ 2 ≤ C j ∥ h ∥ j , ∇ ˉ 2 , j ≥ m , \frac{d}{dt}\lVert h\rVert_{j,\bar{\nabla}}^2 + c_j\lVert h\rVert_{j+1,\bar{\nabla}}^2 \leq C_j\lVert h\rVert_{j,\bar{\nabla}}^2,\qquad j \geq m, d t d ∥ h ∥ j , ∇ ˉ 2 + c j ∥ h ∥ j + 1 , ∇ ˉ 2 ≤ C j ∥ h ∥ j , ∇ ˉ 2 , j ≥ m ,
where c j > 0 c_j > 0 c j > 0 .
On every finite time interval this inequality bounds each higher norm by its initial value and Gronwall’s inequality. If a finite maximal time existed, the equation would also bound every spatial norm of ∂ t h \partial_t h ∂ t h on that interval and hence h h h would have a smooth limit at the maximal time. Positivity and uniform parabolicity persist by the low-order bound, so local existence from this limit would extend the solution. This proves global existence.
The same inequalities give exponential higher-order decay in the same manner as such proof before and we get
∥ h ( t ) ∥ j , ∇ ˉ 2 ≤ e − γ t ( ∥ h ( 0 ) ∥ j , ∇ ˉ 2 + C j ′ λ − γ ∥ h ( 0 ) ∥ H m 2 ) . \lVert h(t)\rVert_{j,\bar{\nabla}}^2 \leq e^{-\gamma t}\left(\lVert h(0)\rVert_{j,\bar{\nabla}}^2 + \frac{C'_j}{\lambda-\gamma}\lVert h(0)\rVert_{H^m}^2\right). ∥ h ( t ) ∥ j , ∇ ˉ 2 ≤ e − γ t ( ∥ h ( 0 ) ∥ j , ∇ ˉ 2 + λ − γ C j ′ ∥ h ( 0 ) ∥ H m 2 ) .
Thus every higher norm decays, although its initial size need not be small. We finally remove the DeTurck term. Let F t F_t F t solve
∂ t F t = − V ( ψ ~ t ) ∘ F t , F 0 = id , ψ ~ t = ∗ ψ ~ t ψ ~ t . \partial_t F_t=-V(\widetilde{\psi}_t)\circ F_t,\qquad F_0=\operatorname{id},\qquad\widetilde{\psi}_t=*_{\widetilde{\psi}_t}\widetilde{\psi}_t. ∂ t F t = − V ( ψ t ) ∘ F t , F 0 = id , ψ t = ∗ ψ t ψ t .
The vector field vanishes at the reference structure and uses only one derivative of h h h . Since m = r + 8 m=r+8 m = r + 8 , Sobolev embedding and the low-order decay imply
∥ h ( t ) ∥ C r + 3 , α + ∥ V ( ψ ~ t ) ∥ C r + 2 , α ≤ C e − λ t / 2 ∥ h ( 0 ) ∥ H m . \lVert h(t)\rVert_{C^{r+3,\alpha}}+\lVert V(\widetilde{\psi}_t)\rVert_{C^{r+2,\alpha}}\leq Ce^{-\lambda t/2}\lVert h(0)\rVert_{H^m}. ∥ h ( t ) ∥ C r + 3 , α + ∥ V ( ψ t ) ∥ C r + 2 , α ≤ C e − λ t /2 ∥ h ( 0 ) ∥ H m .
The vector field is therefore integrable in these norms, with a small integral. The same idea of differentiating the equation for F t F_t F t and using Gronwall’s inequality and then integrating it proves that it is a diffeomorphism and is isotopic to the identity. Put ψ t = F t ∗ ψ ~ t \psi_t=F_t^*\widetilde{\psi}_t ψ t = F t ∗ ψ t . The pullback derivative formula cancels the DeTurck term, so
∂ t ψ t = F t ∗ Q 0 ( ψ ~ t ) . \partial_t\psi_t=F_t^*Q_0(\widetilde{\psi}_t). ∂ t ψ t = F t ∗ Q 0 ( ψ t ) .
The operator Q 0 Q_0 Q 0 uses two derivatives and vanishes at ψ ‾ \overline{\psi} ψ . The preceding estimates and the uniform pullback bounds therefore give
∥ ∂ t ψ t ∥ C r , α ≤ C e − λ t / 2 ∥ h ( 0 ) ∥ H m , ∥ ψ t − ψ ‾ ∥ C r , α ≤ C ∥ h ( 0 ) ∥ H m . \lVert\partial_t\psi_t\rVert_{C^{r,\alpha}}\leq Ce^{-\lambda t/2}\lVert h(0)\rVert_{H^m},
\qquad
\lVert\psi_t-\overline{\psi}\rVert_{C^{r,\alpha}}\leq C\lVert h(0)\rVert_{H^m}. ∥ ∂ t ψ t ∥ C r , α ≤ C e − λ t /2 ∥ h ( 0 ) ∥ H m , ∥ ψ t − ψ ∥ C r , α ≤ C ∥ h ( 0 ) ∥ H m .
The second inequality includes both F t ∗ h ( t ) F_t^*h(t) F t ∗ h ( t ) and F t ∗ ψ ‾ − ψ ‾ F_t^*\overline{\psi}-\overline{\psi} F t ∗ ψ − ψ . Integrating the first inequality proves the required time-integral bound. Taking N = m N=m N = m and reducing δ \delta δ supplies the initial H m H^m H m smallness and proves (86) . The higher estimates give the remaining assertions, with a smaller common exponential rate if necessary. Smooth uniqueness for the ordinary modified coflow is proved by Grigorian. □
Proof of Theorem 1.8. We use Theorem 4.16 with s = 8 s=8 s = 8 and ψ ⋆ = ψ ‾ \psi_\star=\overline{\psi} ψ ⋆ = ψ . We first choose the finite order at which the ambient path must be controlled. Apply the estimates in the proof of Theorem 4.10 at an order which controls the transported connection in C 9 C^9 C 9 . Let r 0 r_0 r 0 be a sufficient ambient order for these estimates, and fix r ≥ max { r 0 , 10 } r\geq\max\{r_0,10\} r ≥ max { r 0 , 10 } . All these choices are made before the initial neighbourhood is chosen. Set N = r + 8 N=r+8 N = r + 8 . Lemma 4.18 , at the order r r r chosen above, gives a global ordinary modified coflow ψ t \psi_t ψ t with parameter zero whenever ∥ ψ 0 − ψ ‾ ∥ C N , α \lVert\psi_0-\overline{\psi}\rVert_{C^{N,\alpha}} ∥ ψ 0 − ψ ∥ C N , α is sufficiently small. Put Q t = ∂ t ψ t Q_t=\partial_t\psi_t Q t = ∂ t ψ t . Its estimates give
sup t ≥ 0 ∥ ψ t − ψ ‾ ∥ C r + sup t ≥ 0 ∥ Q t ∥ C r + ∫ 0 ∞ ∥ Q t ∥ C r d t ≤ C ∥ ψ 0 − ψ ‾ ∥ C N , α . \sup_{t\geq0}\lVert\psi_t-\overline{\psi}\rVert_{C^r}+\sup_{t\geq0}\lVert Q_t\rVert_{C^r}+\int_0^\infty\lVert Q_t\rVert_{C^r}\,\mathrm{d}t\leq C\lVert\psi_0-\overline{\psi}\rVert_{C^{N,\alpha}}. t ≥ 0 sup ∥ ψ t − ψ ∥ C r + t ≥ 0 sup ∥ Q t ∥ C r + ∫ 0 ∞ ∥ Q t ∥ C r d t ≤ C ∥ ψ 0 − ψ ∥ C N , α .
The higher spatial norms of ψ t \psi_t ψ t are uniformly bounded, and Q t Q_t Q t decays exponentially in every spatial norm. In particular, the higher-order time integrals are finite. For sufficiently small initial distance, these estimates imply (72) . Thus Theorem 4.10 supplies a global instanton family B t B_t B t , with
B 0 = I ( ψ 0 ) , ∂ t B t = U B t , ψ t ( Q t ) . B_0=\mathcal{I}(\psi_0),\qquad\partial_tB_t=U_{B_t,\psi_t}(Q_t). B 0 = I ( ψ 0 ) , ∂ t B t = U B t , ψ t ( Q t ) .
We spell out the quantitative bound needed for the next theorem. The local branch from Lemma 4.7 satisfies I ( ψ ‾ ) = A ⋆ \mathcal{I}(\overline{\psi})=A_\star I ( ψ ) = A ⋆ . Its smooth dependence at the chosen finite order bounds I ( ψ t ) − A ⋆ \mathcal{I}(\psi_t)-A_\star I ( ψ t ) − A ⋆ by ψ t − ψ ‾ \psi_t-\overline{\psi} ψ t − ψ . The estimate (73) and the differentiated gauge equation in the proof of Theorem 4.10 bound the additional change by the time integral of Q t Q_t Q t . Our choice of r r r therefore gives
sup t ≥ 0 ∥ B t − A ⋆ ∥ C 9 ≤ C ( sup t ≥ 0 ∥ ψ t − ψ ‾ ∥ C r + ∫ 0 ∞ ∥ Q t ∥ C r d t ) ≤ C ∥ ψ 0 − ψ ‾ ∥ C N , α . \sup_{t\geq0}\lVert B_t-A_\star\rVert_{C^9}\leq C\left(\sup_{t\geq0}\lVert\psi_t-\overline{\psi}\rVert_{C^r}+\int_0^\infty\lVert Q_t\rVert_{C^r}\,\mathrm{d}t\right)
\leq C\lVert\psi_0-\overline{\psi}\rVert_{C^{N,\alpha}}. t ≥ 0 sup ∥ B t − A ⋆ ∥ C 9 ≤ C ( t ≥ 0 sup ∥ ψ t − ψ ∥ C r + ∫ 0 ∞ ∥ Q t ∥ C r d t ) ≤ C ∥ ψ 0 − ψ ∥ C N , α .
The constants are uniform on the fixed finite-order neighbourhood. The higher-order versions of the transport estimates give uniform higher bounds for B t B_t B t and exponential convergence to an instanton B ∞ B_\infty B ∞ for the limiting form ψ ∞ \psi_\infty ψ ∞ .
Since r ≥ 10 r \ge10 r ≥ 10 , the preceding estimates imply
sup t ≥ 0 ( ∥ ψ t − ψ ˉ ∥ C 10 + ∥ B t − A ⋆ ∥ C 9 + ∥ Q t ∥ C 9 ) ≤ C ∥ ψ 0 − ψ ˉ ∥ C N , α . \sup_{t \ge0}\left(\|\psi_t-\bar{\psi}\|_{C^{10}}+\|B_t-A_\star\|_{C^9}+\|Q_t\|_{C^9}\right)\le C\|\psi_0-\bar{\psi}\|_{C^{N,\alpha}}. t ≥ 0 sup ( ∥ ψ t − ψ ˉ ∥ C 10 + ∥ B t − A ⋆ ∥ C 9 + ∥ Q t ∥ C 9 ) ≤ C ∥ ψ 0 − ψ ˉ ∥ C N , α .
Reduce the initial neighbourhood so that this is below the ϵ \epsilon ϵ in (78) . This checks that condition with s = 8 s=8 s = 8 . All the uniform higher spatial bounds in Theorem 4.16 have also been verified.
Finally,
∥ A 0 − B 0 ∥ H 8 ≤ ∥ A 0 − A ⋆ ∥ H 8 + ∥ I ( ψ 0 ) − A ⋆ ∥ H 8 ≤ ∥ A 0 − A ⋆ ∥ H 8 + C ∥ ψ 0 − ψ ˉ ∥ C N , α . \|A_0-B_0\|_{H^8}\le\|A_0-A_\star\|_{H^8}+\|\mathcal{I}(\psi_0)-A_\star\|_{H^8}
\le\|A_0-A_\star\|_{H^8}+C\|\psi_0-\bar{\psi}\|_{C^{N,\alpha}}. ∥ A 0 − B 0 ∥ H 8 ≤ ∥ A 0 − A ⋆ ∥ H 8 + ∥ I ( ψ 0 ) − A ⋆ ∥ H 8 ≤ ∥ A 0 − A ⋆ ∥ H 8 + C ∥ ψ 0 − ψ ˉ ∥ C N , α .
We choose the initial neighbourhood small enough that this is below the initial radius in Theorem 4.16 . That theorem now gives the global smooth connection evolution and smooth exponential convergence to k ∞ ∗ B ∞ k_\infty^*B_\infty k ∞ ∗ B ∞ . Here k ∞ k_\infty k ∞ is the limiting gauge transformation in that theorem. Lemma 4.18 gives smooth exponential convergence of ψ t \psi_t ψ t to the torsion-free form ψ ∞ \psi_\infty ψ ∞ .
The gauge-fixed connection remains in the neighbourhood where the deformation operator has a uniformly bounded inverse. Gauge conjugation preserves invertibility, so the correction remains defined for the original connection throughout the solution. The limiting instanton is acyclic for the same reason. If A 0 A_0 A 0 is initially an instanton, Theorem 4.4 preserves its zero defect. Uniqueness follows from ambient uniqueness and Theorem 4.9 .
We relate our results to another family of flows in G 2 G_2 G 2 geometry which concern the heterotic G 2 G_2 G 2 system. Geometric flows related to the heterotic G 2 G_2 G 2 system have been studied by Ashmore, Minasian and Proto [1 ] . Garcia-Fernandez, Moreno, Payne and Streets [11 ] introduce a parabolic flow for the large volume heterotic G 2 G_2 G 2 system and prove short-time existence and Shi-type estimates. Related stationary coupled equations for the geometry and a bundle connection are studied by da Silva Jr., Garcia-Fernandez, Lotay and Sá Earp [32 ] .
Karigiannis, Picard and Suan ask in [18 Remark 4.13] for a natural connection flow towards a G 2 G_2 G 2 instanton in the anomaly-flow setting. Our Theorem 4.16 gives a partial result in this direction. Near an acyclic instanton, it gives exponential attraction for the corrected connection flow along prescribed coclosed backgrounds satisfying its hypotheses. The correction is nonlocal, since it is obtained by an elliptic inverse. The ambient evolution used here does not depend on the evolving connection. The full α ′ \alpha' α ′ -coupled anomaly flow, where the curvature of the connection also enters the ambient equation, is not covered by this argument.
The coflow parameter
The parameter c c c multiplies a first-order term. It does not change the principal symbol, but it can change the sign of a low-frequency eigenvalue. The following example shows that, for each nonzero parameter, stability can fail at the linear level on a suitable flat background.
Proposition 4.19. For every c ∈ R ∖ { 0 } c\in\mathbb{R}\setminus\{0\} c ∈ R ∖ { 0 } , there is a flat torsion-free G 2 G_2 G 2 structure on a flat seven-torus for which the linearization of ∂ t ψ t = Q c ( ψ t ) \partial_t\psi_t=Q_c(\psi_t) ∂ t ψ t = Q c ( ψ t ) has a positive eigenvalue on exact closed four forms. Its eigenmode is L 2 L^2 L 2 orthogonal to all infinitesimal diffeomorphism directions.
Proof. Write the torus as a product of a circle of length L L L and a flat six-torus, with an adapted S U ( 3 ) SU(3) S U ( 3 ) structure on the six-plane. Let s s s be arclength on the circle. On real primitive three forms of type ( 2 , 1 ) + ( 1 , 2 ) (2,1)+(1,2) ( 2 , 1 ) + ( 1 , 2 ) on the six-plane, set
J η = ∗ ( d s ∧ η ) . J\eta=*(d s\wedge\eta). J η = ∗ ( d s ∧ η ) .
On this space J 2 = − I J^2=-I J 2 = − I . Choose a nonzero constant form η 0 \eta_0 η 0 in this space, take k = 2 π / L k=2\pi/L k = 2 π / L , and let ϵ ∈ { 1 , − 1 } \epsilon\in\{1,-1\} ϵ ∈ { 1 , − 1 } . Define
η ϵ ( s ) = cos ( k s ) η 0 − ϵ sin ( k s ) J η 0 , ζ ϵ = ∗ η ϵ . \eta_\epsilon(s)=\cos(ks)\eta_0-\epsilon\sin(ks)J\eta_0,\qquad
\zeta_\epsilon=*\eta_\epsilon. η ϵ ( s ) = cos ( k s ) η 0 − ϵ sin ( k s ) J η 0 , ζ ϵ = ∗ η ϵ .
Differentiation and J 2 = − I J^2=-I J 2 = − I give
∂ s η ϵ = − ϵ k J η ϵ , ∗ d η ϵ = J ∂ s η ϵ = ϵ k η ϵ , d η ϵ = ϵ k ζ ϵ , Δ ζ ϵ = k 2 ζ ϵ . \begin{aligned}
\partial_s\eta_\epsilon&=-\epsilon kJ\eta_\epsilon,\\
*d\eta_\epsilon&=J\partial_s\eta_\epsilon=\epsilon k\eta_\epsilon,\\
d\eta_\epsilon&=\epsilon k\zeta_\epsilon,\qquad\Delta\zeta_\epsilon=k^2\zeta_\epsilon.
\end{aligned} ∂ s η ϵ ∗ d η ϵ d η ϵ = − ϵ k J η ϵ , = J ∂ s η ϵ = ϵ k η ϵ , = ϵ k ζ ϵ , Δ ζ ϵ = k 2 ζ ϵ .
Thus ζ ϵ = ( ϵ / k ) d η ϵ \zeta_\epsilon=(\epsilon/k)d\eta_\epsilon ζ ϵ = ( ϵ / k ) d η ϵ is exact and closed. The S U ( 3 ) SU(3) S U ( 3 ) decomposition gives η ϵ ∈ Ω 27 3 \eta_\epsilon\in\Omega^3_{27} η ϵ ∈ Ω 27 3 and ζ ϵ ∈ Ω 27 4 \zeta_\epsilon\in\Omega^4_{27} ζ ϵ ∈ Ω 27 4 . We compute the derivative of each term in Q c Q_c Q c at the flat torsion-free form ψ ˉ \bar\psi ψ ˉ . By (13) , the variation of the associated three form in the direction ζ ϵ \zeta_\epsilon ζ ϵ is φ ˙ = − η ϵ \dot\varphi=-\eta_\epsilon φ ˙ = − η ϵ . For closed four forms, Δ ψ ψ = d ∗ d φ \Delta_\psi\psi=d*d\varphi Δ ψ ψ = d ∗ d φ . Since d φ ˉ = 0 d\bar\varphi=0 d φ ˉ = 0 , differentiating the Hodge star in this expression contributes zero. Hence the Laplacian term gives
d ∗ d φ ˙ = − d ∗ d η ϵ = − k 2 ζ ϵ , d*d\dot\varphi=-d*d\eta_\epsilon=-k^2\zeta_\epsilon, d ∗ d φ ˙ = − d ∗ d η ϵ = − k 2 ζ ϵ ,
and tr T = 1 4 ⟨ d φ , ψ ⟩ \operatorname{tr}T=\frac{1}{4}\langle d\varphi,\psi\rangle tr T = 4 1 ⟨ d φ , ψ ⟩ . At the torsion-free reference form,
D ( tr T ) [ ζ ϵ ] = 1 4 ⟨ d φ ˙ , ψ ˉ ⟩ = − ϵ k 4 ⟨ ζ ϵ , ψ ˉ ⟩ = 0. \begin{aligned}
D(\operatorname{tr}T)[\zeta_\epsilon]&=\frac{1}{4}\langle d\dot\varphi,\bar\psi\rangle\\
&=-\frac{\epsilon k}{4}\langle\zeta_\epsilon,\bar\psi\rangle=0.
\end{aligned} D ( tr T ) [ ζ ϵ ] = 4 1 ⟨ d φ ˙ , ψ ˉ ⟩ = − 4 ϵ k ⟨ ζ ϵ , ψ ˉ ⟩ = 0.
The last equality uses Ω 27 4 ⊥ Ω 1 4 \Omega^4_{27}\perp\Omega^4_1 Ω 27 4 ⊥ Ω 1 4 . The derivative of the torsion-trace term therefore vanishes. Finally, the derivative of − 2 c d φ -2c\,d\varphi − 2 c d φ is 2 c d η ϵ 2c\,d\eta_\epsilon 2 c d η ϵ . Combining these terms gives
D Q c ∣ ψ ˉ ζ ϵ = ( − k 2 + 2 ϵ c k ) ζ ϵ . DQ_c|_{\bar\psi}\zeta_\epsilon=(-k^2+2\epsilon ck)\zeta_\epsilon. D Q c ∣ ψ ˉ ζ ϵ = ( − k 2 + 2 ϵc k ) ζ ϵ .
Choose ϵ = 1 \epsilon=1 ϵ = 1 when c > 0 c>0 c > 0 and ϵ = − 1 \epsilon=-1 ϵ = − 1 when c < 0 c<0 c < 0 . The eigenvalue is then − k 2 + 2 ∣ c ∣ k -k^2+2|c|k − k 2 + 2∣ c ∣ k , which is positive if 0 < k < 2 ∣ c ∣ 0<k<2|c| 0 < k < 2∣ c ∣ . It is enough to take L > π / ∣ c ∣ L>\pi/|c| L > π /∣ c ∣ .
We finally check that this mode is not an infinitesimal diffeomorphism. Since d ψ ˉ = 0 d\bar\psi=0 d ψ ˉ = 0 , every such direction is L Y ψ ˉ = d ( Y ⌟ ψ ˉ ) \mathcal{L}_Y\bar\psi=d(Y\mathbin{\lrcorner}\bar\psi) L Y ψ ˉ = d ( Y ┘ ψ ˉ ) . On four forms in dimension seven, δ = ∗ d ∗ \delta=*d* δ = ∗ d ∗ , so
δ ζ ϵ = ∗ d η ϵ = ϵ k η ϵ ∈ Ω 27 3 . \delta\zeta_\epsilon=*d\eta_\epsilon=\epsilon k\eta_\epsilon\in\Omega^3_{27}. δ ζ ϵ = ∗ d η ϵ = ϵ k η ϵ ∈ Ω 27 3 .
On the other hand, Y ⌟ ψ ˉ ∈ Ω 7 3 Y\mathbin{\lrcorner}\bar\psi\in\Omega^3_7 Y ┘ ψ ˉ ∈ Ω 7 3 . Integration by parts and orthogonality of these summands give
⟨ ζ ϵ , L Y ψ ˉ ⟩ L 2 = ⟨ δ ζ ϵ , Y ⌟ ψ ˉ ⟩ L 2 = 0. \langle\zeta_\epsilon,\mathcal{L}_Y\bar\psi\rangle_{L^2}
=\langle\delta\zeta_\epsilon,Y\mathbin{\lrcorner}\bar\psi\rangle_{L^2}=0. ⟨ ζ ϵ , L Y ψ ˉ ⟩ L 2 = ⟨ δ ζ ϵ , Y ┘ ψ ˉ ⟩ L 2 = 0.
Since ζ ϵ ≠ 0 \zeta_\epsilon\ne0 ζ ϵ = 0 , it cannot itself be an infinitesimal diffeomorphism direction. □
Remark 4.20. For each nonzero parameter the proposition produces a suitable background with a growing linear mode. The background, in particular the circle length, is allowed to depend on the parameter. This does not prove instability for every nonzero parameter on a fixed background, and it is not a nonlinear instability theorem. On a fixed torus the displayed modes have eigenvalues − k 2 ± 2 c k -k^2\pm2ck − k 2 ± 2 c k ; their signs depend on the allowed frequencies as well as on c c c . The example therefore explains why an extension of the global stability theorem to nonzero parameters requires information about the spectrum of the chosen background.
Examples with nonzero correction
In this section, we give examples in which the correction U A , ψ ( Q ) U_{A,\psi}(Q) U A , ψ ( Q ) described in previous sections is nonzero, so the connection must move to remain an instanton as the ambient ψ t \psi_t ψ t evolves. We first compute the correction for an ambient family obtained by diffeomorphisms, and then construct examples with nonzero initial correction along the modified Laplacian coflow.
Proposition 4.21. Let A A A be an acyclic instanton for a closed positive ψ \psi ψ and let Y Y Y be an ambient vector field. For Q = L Y ψ Q=\mathcal{L}_Y\psi Q = L Y ψ , put v = Y ⌟ F A v=Y\mathbin{\lrcorner}F_A v = Y ┘ F A . Then
U A , ψ ( Q ) = v − d A ( d A ∗ d A ) − 1 d A ∗ v . (87) U_{A,\psi}(Q)=v-d_A(d_A^*d_A)^{-1}d_A^*v. \tag*{(87)} U A , ψ ( Q ) = v − d A ( d A ∗ d A ) − 1 d A ∗ v . (87)
If F A F_A F A is not identically zero, there is a compactly supported Y Y Y in a coordinate ball for which it is nonzero. Proof. We move both the ambient 4-form and the connection by the flow of the vector field Y Y Y . The instanton equation is preserved by this simultaneous pullback. In a local trivialization, the covariant Cartan formula writes the connection velocity as Y ⌟ F A Y \mathbin{\lrcorner} F_A Y ┘ F A plus an infinitesimal gauge term. At an instanton P A d A = 0 \mathcal{P}_A d_A=0 P A d A = 0 , so that gauge term does not affect the linearized instanton equation. Differentiating the equation therefore gives
P A ( Y ⌟ F A ) = − ∗ ( F A ∧ L Y ψ ) . \mathcal{P}_A(Y \mathbin{\lrcorner} F_A)=-*(F_A\wedge\mathcal{L}_Y\psi). P A ( Y ┘ F A ) = − ∗ ( F A ∧ L Y ψ ) .
The velocity v = Y ⌟ F A v=Y\mathbin{\lrcorner} F_A v = Y ┘ F A need not be Coulomb. We solve d A ∗ d A q = d A ∗ v d_A^*d_Aq=d_A^*v d A ∗ d A q = d A ∗ v and replace v v v by v − d A q v-d_Aq v − d A q . Acyclicity makes the scalar Laplacian invertible, so q q q is determined uniquely. The new velocity satisfies d A ∗ ( v − d A q ) = 0 d_A^*(v-d_Aq)=0 d A ∗ ( v − d A q ) = 0 , and its image under P A \mathcal{P}_A P A is unchanged because P A d A q = 0 \mathcal{P}_Ad_Aq=0 P A d A q = 0 . The uniqueness in Proposition 4.3 now gives (4.45). If the forcing in the differentiated equation is nonzero, its solution cannot be zero.
To arrange nonzero forcing, choose a point p p p with F A ( p ) ≠ 0 F_A(p)\ne0 F A ( p ) = 0 and a nonzero ordinary two-form component β \beta β of the curvature in a frame of E E E . Choose a nonzero coefficient of β \beta β in a coframe and wedge with four of the other five basis covectors. The resulting four form q q q has β ∧ q ≠ 0 \beta\wedge q\ne0 β ∧ q = 0 . Because positive four forms form an open GL ( 7 ) \operatorname{GL}(7) GL ( 7 ) orbit (or are stable forms), the derivative of the orbit map is onto the space of four forms. We can therefore prescribe Y ( p ) = 0 Y(p)=0 Y ( p ) = 0 and the first derivative of Y Y Y at p p p so that ( L Y ψ ) ( p ) = q (\mathcal{L}_Y\psi)(p)=q ( L Y ψ ) ( p ) = q . Multiplying it by a cutoff equal to one near p p p preserves these data and gives a smooth vector field supported in a coordinate ball. The forcing has a nonzero component at p p p by construction.
The ambient evolution of the 4-form obtained by pulling back ψ \psi ψ remains closed and positive. Thus the example fits the prescribed-family coupling as well as the instantaneous compatibility equation.
Proposition 4.22. Suppose ( M , ψ ˉ ) (M,\bar{\psi}) ( M , ψ ˉ ) is compact and torsion-free and carries a nonflat acyclic instanton A ⋆ A_\star A ⋆ . Fix a constant c ∈ R c\in\mathbb{R} c ∈ R and set
Q c ( ψ ) = Δ ψ ψ − 2 d ( ( c + tr T ψ ) φ ) . Q_c(\psi)=\Delta_\psi\psi-2d\bigl((c+\operatorname{tr}T_\psi)\varphi\bigr). Q c ( ψ ) = Δ ψ ψ − 2 d ( ( c + tr T ψ ) φ ) . Then there are closed positive forms ψ ε \psi_\varepsilon ψ ε , cohomologous and arbitrarily close to ψ ˉ \bar{\psi} ψ ˉ , and acyclic ψ ε \psi_\varepsilon ψ ε -instantons A ε A_\varepsilon A ε for which
∗ ψ ε ( F A ε ∧ Q c ( ψ ε ) ) ≠ 0. *_{\psi_\varepsilon}(F_{A_\varepsilon}\wedge Q_c(\psi_\varepsilon))\ne0. ∗ ψ ε ( F A ε ∧ Q c ( ψ ε )) = 0. Consequently their correction is initially nonzero along the actual modified-coflow coupling. When c = 0 c=0 c = 0 and the perturbation is sufficiently small, these examples belong to the global convergent solutions of Theorem 1.8 .
Proof. Choose p p p where a scalar component β \beta β of F A ⋆ F_{A_\star} F A ⋆ is nonzero. It is an element of Λ 14 2 = g 2 \Lambda^2_{14}=\mathfrak{g}_2 Λ 14 2 = g 2 . the Lie algebra of the Lie group G 2 G_2 G 2 . We conjugate inside G 2 G_2 G 2 and put it in a maximal torus, so that in an adapted coframe on the six-plane perpendicular to d s ds d s ,
ω = e 12 + e 34 + e 56 , β = a e 12 + b e 34 + d e 56 , a + b + d = 0. \omega=e^{12}+e^{34}+e^{56},\qquad\beta=ae^{12}+be^{34}+de^{56},\qquad a+b+d=0. ω = e 12 + e 34 + e 56 , β = a e 12 + b e 34 + d e 56 , a + b + d = 0.
Not all a , b , d a,b,d a , b , d vanish. If, for example, a ≠ b a\ne b a = b , take θ = ( e 12 − e 34 ) ∧ e 5 \theta=(e^{12}-e^{34})\wedge e^5 θ = ( e 12 − e 34 ) ∧ e 5 . Then θ \theta θ is a real primitive form of type ( 2 , 1 ) + ( 1 , 2 ) (2,1)+(1,2) ( 2 , 1 ) + ( 1 , 2 ) and β ∧ θ = ( b − a ) e 12345 ≠ 0 \beta\wedge\theta=(b-a)e^{12345}\ne0 β ∧ θ = ( b − a ) e 12345 = 0 . Put ζ ^ = d s ∧ θ \widehat{\zeta}=ds\wedge\theta ζ = d s ∧ θ . The SU ( 3 ) \operatorname{SU}(3) SU ( 3 ) decomposition gives ζ ^ ∈ Λ 27 4 \widehat{\zeta}\in\Lambda^4_{27} ζ ∈ Λ 27 4 . Thus F A ⋆ ( p ) ∧ ζ ^ ≠ 0 F_{A_\star}(p)\wedge\widehat{\zeta}\ne0 F A ⋆ ( p ) ∧ ζ = 0 .
We next choose an exact perturbation whose coflow velocity at p p p contains this four-form direction. Since coflow is second order, we can do this by prescribing the second derivatives of the perturbation while making its value and first derivatives vanish. In fact, the same computation as in Proposition 4.19 shows that the symbol satisfies
σ d s ( D Q c ∣ ψ ˉ ) ζ ^ = − ζ ^ . (88) \sigma_{ds}(D Q_c|_{\bar{\psi}})\widehat{\zeta}=-\widehat{\zeta}. \tag*{(88)} σ d s ( D Q c ∣ ψ ˉ ) ζ = − ζ . (88)
Now we use normal coordinates at p p p and extend θ \theta θ with constant coordinate coefficients. Let ρ \rho ρ be a cutoff equal to one near p p p , and define
ζ = d ( ρ s 3 θ 6 ) . \zeta=d\left(\frac{\rho s^3\theta}{6}\right). ζ = d ( 6 ρ s 3 θ ) .
Near p p p , the cutoff is one and θ \theta θ has constant coordinate coefficients, so ζ = 1 2 s 2 d s ∧ θ \zeta= \frac{1}{2}s^2\,ds\wedge\theta ζ = 2 1 s 2 d s ∧ θ . At p p p , where s = 0 s=0 s = 0 , both ζ \zeta ζ and all its first derivatives vanish. Its only nonzero second coordinate derivative is ∂ s 2 ζ ( p ) = ζ ^ \partial_s^2\zeta(p)=\hat{\zeta} ∂ s 2 ζ ( p ) = ζ ^ . Interpreting (88) as the coefficient of this coordinate derivative gives
( D Q c ∣ ψ ˉ ) ζ ( p ) = ζ ^ . (DQ_{c| \bar{\psi}})\zeta(p)=\hat{\zeta}. ( D Q c ∣ ψ ˉ ) ζ ( p ) = ζ ^ .
The form ζ \zeta ζ need not lie in Ω 27 4 \Omega^4_{27} Ω 27 4 away from p p p . Only the algebraic projection at p p p enters the second-order coefficient applied to its nonzero second derivative ∂ s 2 ζ ( p ) = ζ ^ \partial_s^2\zeta(p)=\hat{\zeta} ∂ s 2 ζ ( p ) = ζ ^ . Every term in which a derivative hits a projection or another coefficient contains ζ ( p ) \zeta(p) ζ ( p ) or a first derivative of ζ \zeta ζ at p p p . These vanish by construction. All lower-order terms vanish for the same reason, including the first-order term containing c c c .
Set ψ ε = ψ ˉ + ε ζ \psi_\varepsilon=\bar{\psi}+\varepsilon\zeta ψ ε = ψ ˉ + εζ . It is closed, cohomologous to ψ ˉ \bar{\psi} ψ ˉ , and positive for sufficiently small ε \varepsilon ε . By Lemma 4.7 there is a smooth acyclic instanton family A ε A_\varepsilon A ε with A 0 = A ⋆ A_0=A_\star A 0 = A ⋆ . Since Q c ( ψ ˉ ) = 0 Q_c(\bar{\psi})=0 Q c ( ψ ˉ ) = 0 ,
∗ ψ ε ( F A ε ∧ Q c ( ψ ε ) ) ( p ) = ε ∗ ψ ˉ ( F A ⋆ ∧ ζ ^ ) ( p ) + O ( ε 2 ) . *_{\psi_\varepsilon}(F_{A_\varepsilon}\wedge Q_c(\psi_\varepsilon))(p)=\varepsilon*_{\bar{\psi}}(F_{A_\star}\wedge\hat{\zeta})(p)+O(\varepsilon^2). ∗ ψ ε ( F A ε ∧ Q c ( ψ ε )) ( p ) = ε ∗ ψ ˉ ( F A ⋆ ∧ ζ ^ ) ( p ) + O ( ε 2 ) .
When differentiating this expression at ε = 0 \varepsilon=0 ε = 0 , the variations of the Hodge star and of the curvature multiply Q c ( ψ ˉ ) = 0 Q_c(\bar{\psi})=0 Q c ( ψ ˉ ) = 0 and hence disappear. The remaining first-order term is the one displayed above, and it is nonzero by our choice of ζ ^ \hat{\zeta} ζ ^ . Thus the forcing is nonzero for all sufficiently small nonzero ε \varepsilon ε . At an instanton the defining equations give P U = − ∗ ( F A ∧ Q c ) PU=-*(F_A\wedge Q_c) P U = − ∗ ( F A ∧ Q c ) , so the correction cannot vanish.
These are initial data for actual solutions. The modified coflow has smooth short-time existence for them [Gri13; BV20], and Theorem 4.8 supplies the connection lift. For c = 0 c=0 c = 0 , we may make the perturbation small enough to apply Theorem 1.8 , which proves the global assertion. □ \square □
Remark 4.23. The hypothesis in Proposition 4.22 has compact examples. The gluing theorem of [Wal13] produces acyclic instantons on compact torsion-free G 2 G_2 G 2 manifolds, and choosing nontrivial instanton gluing data gives nonflat curvature. The proposition uses those examples as initial backgrounds; it does not require a new compact instanton gluing construction.
The obstruction without acyclicity
In the preceding sections, acyclicity made the augmented operator invertible and supplied the spectral gap used in the stability proof. When its kernel is nonzero, a compatible connection velocity may fail to exist. The relation between ambient deformations and instanton deformations was studied by de la Ossa, Larfors and Svanes [[28 Sections 5.3–5.4] ], using a curvature map which determines which infinitesimal ambient deformations preserve the instanton condition. Here we state the solvability condition for a prescribed closed variation Q Q Q by applying the Fredholm alternative to the self-adjoint operator D A , ψ \mathcal{D}_{A,\psi} D A , ψ . Thus the source must be orthogonal to its kernel. This is a linear compatibility statement and it does not give nonlinear continuation through a change in the kernel dimension.
Proposition 4.24. Let A A A be any instanton for a closed positive ψ \psi ψ on compact M M M , and let Q Q Q be a closed four form. Put b = ∗ ( F A ∧ Q ) b=*(F_A\wedge Q) b = ∗ ( F A ∧ Q ) and
H A , ψ 1 = { a : d A ∗ a = 0 , P A a = 0 } . \mathcal{H}^{1}_{A,\psi}=\{a:d_A^*a=0,\ \mathcal{P}_Aa=0\}. H A , ψ 1 = { a : d A ∗ a = 0 , P A a = 0 } . There is a compatible velocity v v v satisfying
d A ∗ v = 0 , P A v = − b (89) d_A^*v=0,\qquad\mathcal{P}_Av=-b \tag*{(89)} d A ∗ v = 0 , P A v = − b (89) if and only if
⟨ b , a ⟩ L 2 = 0 for every a ∈ H A , ψ 1 . (90) \langle b,a\rangle_{L^2}=0\qquad\text{for every }a\in\mathcal{H}^{1}_{A,\psi}. \tag*{(90)} ⟨ b , a ⟩ L 2 = 0 for every a ∈ H A , ψ 1 . (90) When this holds, there is a unique such velocity orthogonal to H A , ψ 1 \mathcal{H}^{1}_{A,\psi} H A , ψ 1 . It has minimum L 2 L^2 L 2 norm among all velocities satisfying the differentiated instanton equation, including those not in Coulomb gauge. On smooth families with constant kernel dimension, these minimum norm solutions depend smoothly provided (90) holds.
Proof. At an instanton the kernel of the self-adjoint elliptic operator D D D is ker d A ∣ Ω 0 ⊕ H A , ψ 1 \ker d_A|_{\Omega^0} \oplus\mathcal{H}^1_{A,\psi} ker d A ∣ Ω 0 ⊕ H A , ψ 1 . The range is its L 2 L^2 L 2 orthogonal complement. The source ( 0 , − b ) (0,-b) ( 0 , − b ) is already orthogonal to the scalar part of the kernel, and its orthogonality to the one-form part is exactly (90) . This proves that the augmented equation D ( f , v ) = ( 0 , − b ) D(f,v)=(0,-b) D ( f , v ) = ( 0 , − b ) is solvable if and only if that condition holds.
When it holds, choose the unique solution orthogonal to ker D \ker D ker D . Apply d A ∗ d_A^* d A ∗ to its second equation and as before we get f = 0 f=0 f = 0 . The remaining component v v v satisfies (89) and is orthogonal to H A , ψ 1 \mathcal{H}^1_{A,\psi} H A , ψ 1 .
To compare with any other compatible velocity w w w , including one that is not Coulomb, we follow the same strategy as in the proof of Proposition 4.3 and obtain the minimum norm property and its uniqueness.
For a smooth family with constant kernel dimension, identify the varying L 2 L^2 L 2 spaces by the positive square roots of their metric and density changes. On this fixed Hilbert space choose, at any reference parameter, a circle ∣ z ∣ = ρ |z|=\rho ∣ z ∣ = ρ which encloses zero but no other eigenvalue. The resolvent remains defined on that circle for nearby parameters, and the projection is
Π = 1 2 π − 1 ∮ ∣ z ∣ = ρ ( z − D ) − 1 d z . \Pi= \frac{1}{2\pi\sqrt{-1}}\oint_{|z|=\rho}(z-D)^{-1}\,dz. Π = 2 π − 1 1 ∮ ∣ z ∣ = ρ ( z − D ) − 1 d z .
Smooth dependence of the resolvent gives smooth dependence of this projection. Its rank is locally constant. Since the kernel dimension is also constant by hypothesis, the enclosed spectral subspace cannot contain any additional nonzero eigenvectors nearby and hence it is exactly ker D \ker D ker D . On that kernel, D + Π D+\Pi D + Π acts as the identity. On the orthogonal complement it acts as the invertible restriction of D D D . Therefore
D † = ( D + Π ) − 1 ( I − Π ) D^\dagger=(D+\Pi)^{-1}(I-\Pi) D † = ( D + Π ) − 1 ( I − Π )
vanishes on the kernel and inverts D D D on its orthogonal complement. This formula shows its smooth dependence on the parameters, and elliptic regularity gives the corresponding maps between higher Sobolev spaces. Applying it to ( 0 , − b ) (0,-b) ( 0 , − b ) gives the minimum norm solution smoothly. The construction is local in the parameter. The locally constructed solutions agree wherever their domains overlap, by uniqueness, so they define smooth solutions along the given family.
Remark 4.25. In the proof above, uniform bounds over the whole family would require uniform estimates in addition to the local argument provided.
Remark 4.26. This proposition gives the exact first-order obstruction to removing acyclicity. Constant kernel dimension on a chosen family is an additional hypothesis, and is generally not an open condition on all connections. Thus the proposition does not define a smooth correction on an arbitrary open neighborhood of a nonacyclic instanton. If an eigenvalue tends to zero, the inverse on its eigenspace can diverge unless the corresponding forcing component vanishes at a sufficient rate. A general nonlinear transport theorem across such a crossing would need further information about the moduli space and these obstruction components.
Further questions
We saw many unifying themes in constructing coupled system of flows of the ambient G 2 G_2 G 2 structures (via φ \varphi φ and ψ \psi ψ ) and two naturally related objects in the form of coassociative submanifolds moving by the mean curvature flow and instantons moving by the Yang-Mills flow. There were many unifying themes between the ideas behind these objects. Both constructions choose a velocity by an L 2 L^2 L 2 minimization condition, but the quantities being minimized are different. For coassociatives, we minimize the correction to mean curvature. With ordinary Laplacian flow in the ambient space, this gives the normal velocity Π i H − R i d C ( i ∗ τ ) \Pi_iH-R_i d_C(i^*\tau) Π i H − R i d C ( i ∗ τ ) . It is the compatible velocity closest to H H H , and its kernel component is Π i H \Pi_iH Π i H . It need not be the compatible velocity of smallest norm. For instantons, we remove the gauge component and choose the compatible velocity of smallest L 2 L^2 L 2 norm. In both cases the correction is nonlocal: an elliptic inverse is applied over the whole submanifold or ambient manifold. The preceding theorems show that these choices give actual evolutions and convergence under the stated neighborhood and stability assumptions. All these points give rise to several interesting questions for further research. We mention some of them below.
Can the embedding flow start from an embedding that is not coassociative? Our correction is defined for embeddings close to a coassociative embedding, even when their pullback is not zero. Does the corrected embedding equation have a smooth solution for a short time from every such initial embedding?
Our existence proof does not answer this question, since it uses the finite-dimensional family of embeddings whose pullback is zero. An existence proof for initial embeddings with nonzero pullback would therefore require a different argument. A separate question is whether the correction can be changed so that a small nonzero pullback tends to zero.
2. Can the coassociative embeddings develop a singularity while the ambient flow remains smooth? More precisely, can the tubular radius tend to zero or the second fundamental form become unbounded in finite time, even though the ambient structure remains smooth? If this can happen, what does the degeneration look like? If it cannot happen under suitable additional assumptions, what estimates prevent it?
Proposition 3.13 gives continuation under its compactness and tubular bounds. These assumptions allow us to keep writing nearby embeddings as normal graphs. The proposition does not explain what happens when these bounds fail. Answering the questions above would help us understand whether the embedding can break down before the ambient flow does.
3. Can the connection flow continue when the kernel dimension changes? Suppose that the deformation operator develops a larger kernel along a family of instantons. Under what additional conditions can the connection flow be continued smoothly through this point? See also Remark 4.26.
4. For which positive coflow parameters does stability hold? Fix a torsion-free background. Is there a range of nonzero parameters for which sufficiently small perturbations remain close to the background and converge? Can such a result give an analogue of Theorem 1.8?
Our global coupled theorem uses parameter zero. Proposition 4.19 produces, for every nonzero parameter, a flat background with a growing linear mode. The background is allowed to vary with the parameter. The proposition therefore does not answer the stability question for a fixed background. One would first need to identify parameter values for which the linearized operator has the required decay on the fixed cohomology class, modulo diffeomorphisms, and then control the nonlinear terms.
5. Can attraction to an instanton be proved for larger initial perturbations? What additional geometric bounds would allow us to prove global existence and convergence when the initial connection is not sufficiently close to an acyclic instanton?
Corollary 4.5 gives decay of the defect under its spectral and metric assumptions but we need further estimates to prevent curvature concentration, where the curvature becomes large in increasingly small regions. We would also need to control the ambient structure and address any loss of invertibility of the deformation operator. The question is which additional assumptions provide these bounds and allow the attraction theorem to extend beyond small initial perturbations.
For ordinary Yang–Mills flow on a fixed compact torsion-free G 2 G_{2} G 2 manifold, Oliveira and Waldron [29 ] show that a uniform pointwise bound on the Λ 7 2 \Lambda^{2}_{7} Λ 7 2 component of the curvature prevents a finite-time singularity. The L 2 L^{2} L 2 defect estimate in Corollary 4.5 does not provide this pointwise bound. Can a corresponding continuation criterion be proved for our corrected connection flow along a moving coclosed background?
6. Are there similar coupled flows in Spin(7)-geometry? There is a closely related exceptional geometry on 8-dimensional manifolds where the special geometric structure is described by an admissible 4-form Φ \Phi Φ . The submanifolds corresponding to coassociatives are called Cayley submanifolds and there is an analogous notion of Spin(7)-instantons. So can we define similar coupled flows as in this paper for those objects? Natural flows of Spin(7)-structures are discussed in [9 ] and [8 ] . Probably the first step would be use the results in [2 ] and do a computation similar to Lemma 3.1 and Proposition 3.2 by using particulars of Spin(7)-geometry. (SD:) Fachbereich Mathematik, Universität Hamburg, Bundesstraße 55, 20146 Hamburg, Germany. shubham.dwivedi@uni-hamburg.de
(SP:) siva17aug@gmail.com