To Andr´e Haefliger with admiration Holonomic approximation and Gromov’s h-principle 1 0 0 2 Y. M. Eliashberg N. M. Mishachev n a Stanford University, USA Lipetsk Technical University, Russia J 4 and Stanford University, USA 2 ] G S Abstract . h In1969M.GromovinhisPhDthesis[G1]greatlygeneralizedSmale-Hirsch-Phillipsimmersion- t a m submersiontheory(see[Sm],[Hi],[Ph])byprovingwhatisnowcalledtheh-principleforinvariant [ opendifferentialrelationsoveropenmanifolds. Gromovextractedtheoriginalgeometricideaof 2 Smaleandputittoworkinthemaximalpossiblegenerality. Gromov’sthesiswasbroughttothe v 6 West by A. Phillips and was popularized in his talks. However, most western mathematicians 9 1 firstlearnedaboutGromov’stheoryfromAHaefliger’sarticle[H]. Thecurrentpaperisdevoted 1 to the same subject as the papers of Gromov and Haefliger. It seems to us that we further 0 1 purified Smale-Gromov’s original idea by extracting from it a simple but very general theorem 0 / about holonomic approximationof sections of jet-bundles (see Theorem 1.2.1 below). We show h t a belowthatGromov’stheoremaswellassomeotherresultsintheh-principlespiritareimmediate m 1 corollaries of Theorem 1.2.1. : v i X 1 Holonomic approximation r a 1.1 Jets and holonomy Given a C∞-smooth fibration p :X → V, we denote by X(r) the space of r-jets of smooth sections f : V → X and by Jr : V → X(r) the r-jet of a section f : V → X. When the fibration f X = V × W → V is trivial then the space X(r) is sometimes denoted by Jr(V,W), and called the space of r-jets of maps V → W. A section F : V → X(r) is called holonomic if it has the form Jr for a section f : V → X. The correspondence f 7→ Jr defines the derivation map f f 1The authors are partially supported by theNational Science Foundation 1 Jr : SecX → SecX(r). Its one-to-one image Jr(SecX) coincides with the space HolX(r) ⊂ SecX(r) of holonomic sections, i.e. we have SecX J≃r HolX(r) ֒→ SecX(r). Notice that the C0-topology on SecX(r) induces via Jr the Cr-topology on SecX. Following Gromov’s book [G2] we will denote by OpA an arbitrary small but non-specified (open) neighborhood of a subset A ⊂ V. We will assume that the manifold V and the bundle X(r) are endowedwithRiemannianmetricsanddenotebyU (A)themetricε-neighborhood{x| dist(x,A) < ε ε}. Given an arbitrary subset A ⊂ V a section F : A → X(r) is called holonomic if there exists a holonomic section F : OpA → X(r) such that F| = F. A section F : V → X is called holonomic A e e over A ⊂ V if the restriction F| is holonomic. Given a fibration π : V → B we say that a section A F : V → X(r) is fiberwise holonomic if there exists a continuous family of holonomic sections F : Opπ−1(b) → X(r), b ∈ B, such that for each b ∈ B the sections F and F coincide over the b b fieber π−1(b). Note the following trivial but important fact: e 1.1.1 Any section F : V → X is holonomic over any point v ∈ V. Moreover, it is fiberwise holonomic with respect to the trivial fibration Id :V → V. Indeed, we can take the Taylor polynomial map which corresponds to F(v) with respect to some local coordinate system centered atv as asection F : Opv → X(r). Moreover, the local coordinate v e system can be chosen smoothly depending on its center. ◮ 1.2 Holonomic approximation Question: Is it possible to approximate any section F : V → X(r) by a holonomic section? In other words, given a r-jet section and an arbitrary small neighborhood of the image of this section in the jet space, can one find a holonomic section in this neighborhood? The answer is evidently negative (excluding, of course, the situation when the initial section is already holonomic). For instance, in the case r = q = 1 the question has the following geometrical reformulation: given a function and a n-plane field along the graph of this function, can one C0- perturb this graph to make it almost tangent to the given field? 2 Theproblemoffindingaholonomicapproximation ofasectionofther-jetspacenear a submanifold A ⊂ Rn is also usually unsolvable. The only exception is the zero-dimensional case: as we already stated above in 1.1.1 any section can be approximated near any point by the r-jet of the respective Taylor polynomial map. In contrast, the following theorem says that we always can find a holonomic approximation of a section F : V → X(r) near a slightly deformed submanifold A ⊂ V if the original set A ⊂ V is of e positive codimension. 1.2.1 (Holonomic Approximation Theorem) Let A ⊂ V be a polyhedron of positive codi- mension and F : OpA → X(r) be a section. Then for arbitrary small δ,ε > 0 there exist a diffeomorphism h: V → V with ||h−Id||C0 < δ and a holonomic section F : Oph(A) → X(r) such e that the image h(A) is contained in the domain of the definition of the section F and ||F −F|Oph(A)||C0 < ε. e We use the term polyhedron here in the sense that A is a subcomplex of a certain smooth triangu- lation of the manifold V. As we will see below, the relative and the parametric versions of the theorem are also true. In the relative version the section F is assumed to be already holonomic over OpB, where B is a subpolyhedron of A, while the diffeomorphism h is constructed to be fixed on OpB and F is e required to coincide with F on OpB. Here is the parametric version of 1.2.1. 1.2.2 (Parametric Holonomic Approximation Theorem) Let A ⊂ V be a polyhedron of positive codimension and F : OpA → X(r) be a family of sections parametrized by a cube Im, z m = 0,1,... . Suppose that the sections F are holonomic for z ∈ Op∂Im. Then for arbitrary z small δ,ε > 0 there exist a family of diffeomorphisms h : V → V and a family of holonomic z section F :Oph(A) → X(r), z ∈ Im, such that z e • ||hz −Id||C0 < δ; • h = Id for z ∈ Op∂Im; z • ||Fz −Fz|Ophz(A)||C0 < ε ; e • F = F for z ∈ Op∂Im. z z e 3 Using the induction over skeleta of the polyhedron A, and taking into account that the fibration X → V is trivial over simplices, we reduce Theorem 1.2.1 to its special case for the pair (A,B) = (Ik,∂Ik)⊂ Rn. We consider this special case in the next section. 1.3 Holonomic Approximation over a cube 1.3.1 (Holonomic Approximation over a cube) Let Ik ⊂ Rn, k < n, be the unit cube in the coordinate subspace Rk ⊂ Rn of the first k coordinates. Then for any section F : OpIk → Jr(Rn,Rq) which is holonomic over Op∂Ik and for arbitrary small δ,ε > 0 there exist a diffeomorphism h :Rn → Rn of the form h(x1,...,xn) = (x1,...,xn−1,xn+ϕ(x1,...,xk)), and a holonomic section F : Oph(Ik)→ Jr(Rn,Rq) e such that • |ϕ| <δ; ϕ| = 0; Op∂Ik • the image h(Ik) is contained in the domain of the definition of the section F; • F| = F| and Op∂Ik Op∂Ik e • ||F −F|Oph(Ik)||C0 <ε. e Theorem 1.3.1 will be deduced from Inductional lemma 1.3.2 which we formulate below. Let π :Ik → Ik−l, l = 1,...,k, be the projection l (x1,...,xk)7→ (x1,...,xk−l) whose fibers are l-dimensional cubes Il(y) = y×Il, y = (x1,...,xk−l)∈ Ik−l. Note that for l < k we have Il(y) = Il−1(y,t) where t = xk−l+1. tS∈I 4 1.3.2 (Inductional lemma) Suppose that a section F :OpIk → Jr(Rn,Rq) is holonomic over Op∂Ik and for a positive integer l ≤ k fiberwise holonomic with respect to the fibration πl−1 : Ik → Ik−l+1. Then for arbitrary small δ,ε > 0 there exist a diffeomorphism h :Rn → Rn of the form h(x1,...,xn)= (x1,...,xn−1,xn +ϕ(x1,...,xk)), and a section F : Oph(In)→ Jr(Rn,Rq), e such that • |ϕ| <δ; ϕ| = 0; Op∂Ik • the image h(Ik) is contained in the domain of the definition of the section F; • F| = F| ; Op∂Ik Op∂Ik e • ||F −F|Oph(Ik)||C0 <ε; e • the section F| is fiberwise holonomic with respect to the fibration h(Ik) e π ◦h−1 :h(Ik)→ Ik−l. l Proof of Inductional lemma: We recommend to the reader to keep in mind the simplest case n = 2, k = l =1 while reading the proof for the first time. For (y,t) ∈ Ik−l×I set U (y,t) = U (Il−1(y,t)) and U∂(y,t) = U (∂Il−1(y,t)). δ δ δ δ It follows from the definition of a fiberwise holonomic section and compactness arguments that we can choose δ > 0 so small that there would exist a continuous family of holonomic sections F = Jr : U (y,t) → Jr(Rn,Rq), (y,t) ∈ Ik−l+1 = Ik−l×I, y,t fy,t δ such that for all (y,t) ∈ Ik−l×I: 5 • F is defined on U (y,t); y,t δ • Fy,t|Il−1(y,t) = F|Il−1(y,t). We can further adjust δ and the family F in order to have y,t • F | = F| for all (y,t) ∈Ik−l×I and y,t U∂(y,t) U∂(y,t) δ δ • F = F| for all (y,t) ∈ Op∂(Ik−l×I). y,t Uδ(y,t) For a sufficiently large N, which is determined by the Interpolation Property below, we set i i U (y)= U (y, ), U∂(y)= U∂(y, ), y ∈ Ik−l, i= 0,...,2N , i δ 2N i δ 2N δ top U (y) = U (y)∩{x ≥ }, i i n 2 δ Ubot(y) = U (y)∩{x ≤− }. i i n 2 We will write Fi and fi instead of F and f . y y y, i y, i N N Notice that σ(N) = max ||Fi(x)−Fi−1(x)|| → 0, y y y∈Ik−l,i=1,...,2N,x∈Ui(y)∩Ui−1(y) N→∞ and hence we have the following Interpolation Property. For any ε > 0, a sufficiently large N and all odd integers i = 1,3,...2N −1 there exist continuous families of holonomic sections Gi = Jr : U (y) → Jr(Rn,Rq), y ∈ Ik−l, y gi i y such that • Gi interpolate between Fi−1 and Fi+1 for all y ∈ Ik−l : y y y  Fi+1 on Utop(y)∩Utop, Gi =  y i i+1 y  Fi−1 on Ubot(y)∩Ubot; y i i−1   • ||Giy −Fyi||C0 < ε for all y ∈Ik−l; • Gi| = Fi| = F| for all y ∈ Ik−l; y U∂(y) y U∂(y) U∂(y) i i i • Gi| = Fi| = F| for all y ∈Op∂Ik−l. y Ui(y) y Ui(y) Ui(y) 6 For even values i = 0,2,...,2N we set Gi ≡ Fi. Take a cut-off function θ : R → I, which is y y N equal to 0 on Op(R\I) and is equal to 1 on [ 1 ,1− 1 ], and define a function ϕ : Rk → R by 4N 4N N the formula ϕ (y,t,x) = θ (t)θ (||x||)θ (||y||)cos2Nπt, y ∈ Rk−l,t ∈R,x ∈ Rl−1. N N N N Consider the map h :Rn → Rn, h(x1,...,xn) = (x1,...,xn−1,xn+δ1ϕN(x1,...,xk)) where δ/2 < δ < δ. Viewing the image I˜l(y)= h(Il(y)) as the union 1 2N−1 I˜l(y) = I (y), [ i 0 e where i i+1 I˜i(y) = I˜l(y)∩{ ≤ t = xk−l+1 ≤ } 2N 2N we define a continuous family of holonomic sections F : OpI˜l(y) → Jr(Rn,Rq), y ∈Ik−l y e by the formula F | = Gi| , i = 0,...,2N −1. y OpI˜i y OpI˜i e Then the section F :Oph(Ik) → Jr(Rn,Rq) defined by the formula e F(y,t,x) = F (t,x) for (y,t,x) ∈ Op(Ik−l×I ×In−k−l−1) y e e is holonomic with respect to the fibration π ◦h and is ε-close to F. ◮ l Proof of Theorem 1.3.1: We will prove the theorem by an induction over l. Consider for l = 0,...k the following Inductional hypothesis A(l). Let F : OpIk → Jr(Rn,Rq) be a section which is holonomic over Op∂Ik. For arbitrary small δ,ε > 0 there exist a diffeomorphism h :Rn → Rn of the form h(x1,...,xn)= (x1,...,xn−1,xn +ϕ(x1,...,xk)), and a section Fl :Oph(In) → Jr(Rn,Rq) e as required by the Inductional lemma, i.e 7 • |ϕ| <δ; ϕ| = 0; Op∂Ik • the image h(Ik) is contained in the domain of the definition of the section F; • Fl| = F| ; Op∂Ik Op∂Ik e • ||Fl −F|Oph(Ik)||C0 <ε; e • the section Fl| is fiberwise holonomic with respect to the fibration h(Ik) e π ◦h−1 :h(Ik)→ Ik−l. l Accordingto1.1.1thegivensectionF :In → Jr(Rn,Rq),whichisholonomicnear∂Ik,istautologi- cally fiberwiseholonomicwithrespecttothefibrationbypointsπ :Ik → Ik. ThisimpliesA(0) and 0 thusgivesusthebasefortheinduction. Forl = 1theimplicationA(l−1) ⇒ A(l) followsimmediately from Inductional lemma 1.3.2, but in the general case l > 1 we cannot apply 1.3.2 directly because the section Fl−1 is defined near the deformed cube rather than the original one. Notice however, that the diffeomorphism h : Rn → Rn induces the covering map h∗ : Jr(Rn,Rq) → Jr(Rn,Rq). The section F¯l−1 = (h∗)−1(Fl−1) is defined over OpIk, coincides with F near ∂Ik and fiberwise holonomic with respect to thee fibration πl−1 : Ik → Ik−l+1. Applying Inductional lemma 1.3.2 we can approximate F¯l−1 by a section F′ over a deformed cube h′(Ik). The section F′ coincides with F¯l−1 near ∂Ik and fiberwise holonomeic with respect to the fibration π ◦h′ : h′(Iek) → Ik−l. If F′ l is sufficiently C0-close to F¯l−1, then the section Fl = h∗(F′) is the required approximation of F ein a neighborhood of h′′(Ik), where h′′ = h◦(h′). Tehis provees A(l) and Theorem 1.3.1. ◮ Parametric case. It turns out that Inductional lemma 1.3.2 implies also the parametric version of Theorem 1.3.1. Namely, we have 1.3.3 (Parametric version of Theorem 1.3.1) Let F , z ∈ Im, be a family of sections z OpIk → Jr(Rn,Rq), parameterized by the cube Im. Suppose that k < n and the sections F z are holonomic over Op∂Ik for all z ∈ Im, and holonomic over the whole Ik for z ∈ Op∂Im. Then for arbitrary small δ,ε > 0 there exist a family of diffeomorphisms h :Rn → Rn of the form z hz(x1,...,xn) = (x1,...,xn−1,xn+ϕz(x1,...,xk)), and a family of holonomic sections F : Oph (Ik) → Jr(Rn,Rq) z z e 8 such that • |ϕ |< δ; ϕ | = 0; z z Op∂Ik • the section F is defined in a neighborhood of h (Ik) z ∈Im; z z • F | = F | z Op∂Ik z Op∂Ik e • F = F for z ∈ OpIm and z z e • ||Fz −Fz|Ophz(Ik)||C0 < ε. e Proof: Consider the cube Ik+m = Ik ×Im ⊂ Rn ×Rk = Rn+m. Let Jr(Rn+m|Rn,Rq) be the bundle whose restriction to Rn×z, z ∈ Rm, equals Jr(Rn,Rq). The family of sections F :Ik → Jr(Rn,Rq) z can be viewed as a section F :Ik+m → Jr(Rn+m|Rn,Rq). The section F lifts to a section F : Ik+m → Jr(Rn+m,Rq), so that π◦F = F, where π : Jr(Rn+m,Rq) → Jr(Rn+m|Rn,Rq) is the canonical projection. Moreover, the section F can be chosen holonomic near ∂Ik+m. Hence we can apply Theorem 1.3.1 to get an ε-approximation F of F over a δ-displaced cube h(Ik+m). e e Then the composition F = π ◦F : Ik+m → Jr(Rn+m|Rn,Rq) can be interpreted as the required e e e family {Fz}z∈Im of holonomic ε-approximations of the family {Fz} near {hz(Ik)}. ◮ e In the same way as Theorem 1.3.1 implies Theorem 1.2.1, i.e. via the induction over skeleta, Theorem 1.3.3 implies Theorem 1.2.2. 2 Applications 2.1 Gromov’s h-principle for DiffV-invariant differential relations over open manifolds Differential relations. Adifferential relation (or condition) imposedon sections ϕ: V → X of a fiberbundleX → V is asubsetR⊂ X(r), wherer is called theorder of R. Asection Φ :V → X(r) 9 is called a formal solution of R if Φ(V) ⊂ R. A Cr-section ϕ : V → X is called a solution of the relation R if Jr(V) ⊂ R. We will denote the space of solution of R by SolR, the space of formal ϕ solutions by SecR, and the space of holonomic formal solutions by HolR. The r-jet extension establishes a one-to-one correspondence Jr :SolR→ HolR between the solutions and holonomic formal solutions of R. We will use the term “solution” also for the sections from HolR when the distinction between the solutions of R as sections of X or X(r) is clear from the context, or irrelevant. Diff V -invariant differential relations. Givenafibrationp :X → V wewilldenotebyDiff X V the group of fiber-preserving diffeomorphisms h : X → X, i.e. h ∈ Diff X if and only if there X X V exists a diffeomorphism h : V → V such that p ◦ h = h ◦p. Let π : Diff X → DiffV be V X V V the projection h 7→ h . We are interested when this arrow can be reversed, i.e. when there X V exists a homomorphism j : DiffV → Diff X such that π ◦j = id. We call a fibration X → V, V together with a homomorphism j, natural if such a lift exists. For instance, the trivial fibration X = V ×W → V → V is natural. Here j(h )= h ×id. The tangent bundleT(V)→ V is natural V V as well. If a fibration X → V is natural then any fibration associated with it is natural as well. In particular, if X → V is natural then X(r) → V is natural. The implied lift jr : DiffV → DiffVX(r), h7→ h∗ is defined here by the formula h∗(s)= Jjr(h)◦s¯(h(v)) where s ∈ X(r), v = pr(s)∈ V , and s¯ is a local section near v which represents the r-jet s. Given a natural fibration X → V , a differential relation R ⊂ X(r) is called DiffV-invariant if the action s 7→ h∗s, h ∈ DiffV leaves R invariant. In other words, a differential relation R is DiffV-invariant if it can be defined in V-coordinate free form. Notice that though the definition of a DiffV-invariant relation depends on the choice of the homomorphism j this choice is fairly obvious in most interesting examples, and we will not specify it. The action s7→ h∗s preserves the set of holonomic sections: h∗(Jfr)= Jr j(h)◦f ◦h−1 , (cid:0) (cid:1) 10