The vanishing condition: what a Lagrangian submanifold is
From Guide 1 you have a symplectic manifold (M, omega): a smooth manifold of even dimension 2n carrying a closed (d omega = 0) and nondegenerate 2-form omega. Nondegeneracy means omega(v, w) = 0 for all w forces v = 0 — at each point omega is an antisymmetric, invertible pairing on T_p M, the geometric ghost of a symplectic inner product. Now ask a natural sub-question: on which submanifolds L of M does omega become as degenerate as possible? A Lagrangian submanifold is the extreme answer — a submanifold L on which omega restricts to exactly zero (omega|_L = 0), and which is as large as that condition permits, namely of dimension n, precisely half the dimension of M.
Why does 'omega vanishes' cap the dimension at n? At a point p, a subspace V of T_p M on which omega is identically zero is called isotropic. Linear symplectic algebra (the same algebra behind Darboux from Guide 1) shows every isotropic subspace has dimension at most n, and the maximal ones — the Lagrangian subspaces — have dimension exactly n. A Lagrangian submanifold is just a submanifold whose tangent space T_p L is a Lagrangian subspace of T_p M at every point. So the slogan 'half-dimensional and omega-killing' is forced: it is the largest you can be while staying isotropic, not an arbitrary choice of size.
The model examples: where Lagrangians actually come from
Abstract definitions earn their keep through examples, so lead with the canonical source. Take any manifold X and form its cotangent bundle T*X, which from Guide 1 you know is the universal symplectic manifold: it carries a tautological 1-form lambda (the 'pdq' of mechanics) and the symplectic form omega = -d lambda = dp ^ dq. The zero section — the copy of X sitting inside T*X as the points with zero covector — is the cleanest Lagrangian there is. Along it lambda vanishes by construction, so omega = -d lambda restricts to zero, and its dimension is dim X = n. The base manifold reappears as a Lagrangian inside its own cotangent bundle.
Now the example that explains why physicists and geometers care. Let alpha be a 1-form on X — that is, a section of T*X — and look at its graph, the set { (x, alpha_x) : x in X } sitting inside T*X. A short computation shows the tautological form pulls back along the graph to alpha itself, so the graph is Lagrangian if and only if d alpha = 0, i.e. exactly when alpha is a closed 1-form. The zero section is the special case alpha = 0. So Lagrangians-near-the-zero-section are nothing but closed 1-forms, and the EXACT ones (alpha = df) are graphs of differentials of functions — the 'generating functions' that run all of Hamilton-Jacobi theory. This is the bridge from Guide 2's Hamiltonian flows to geometry: a Lagrangian encodes a closed form, and its cohomology class in de Rham cohomology H^1(X) measures how far it is from being a graph of a single function.
Two more examples cement the picture, and the second carries a famous slogan. First: the diagonal. Given any symplectic (M, omega), the product M x M with the form omega (-) omega = pr_1*omega - pr_2*omega is symplectic, and the diagonal { (m, m) } is Lagrangian — the sign flip is exactly what makes omega vanish on it. Second, the consequence: a diffeomorphism f: M -> M is a symplectomorphism (preserves omega, Guide 1's structure-preserving maps) if and only if its graph is a Lagrangian in (M x M, omega (-) omega). This is Weinstein's creed, half-joking and half-deep: 'everything is a Lagrangian submanifold.' Maps, symmetries, correspondences — recast them as Lagrangians and the symplectic machinery does the work.
The Lagrangian Grassmannian and the birth of the Maslov index
To do topology with Lagrangians we first need the space of all of them at a single point. Fix the standard symplectic vector space (R^{2n}, omega_0). The set of all Lagrangian subspaces of it is the Lagrangian Grassmannian, written L(n) or Lambda(n). Concretely it is a smooth compact manifold: the unitary group U(n) acts transitively on Lagrangian subspaces (think of a Lagrangian as the real span of an orthonormal complex basis), and the subgroup fixing the standard real Lagrangian R^n is the orthogonal group O(n), so L(n) = U(n)/O(n), a manifold of dimension n(n+1)/2. The simplest case is vivid: L(1) is the set of lines through the origin in the plane (every line is Lagrangian in dimension 2), which is the real projective line RP^1, a circle.
Here is the topological miracle that makes an index possible: the fundamental group of the Lagrangian Grassmannian is the integers, pi_1(L(n)) = Z, for every n. (The determinant-squared map det^2: U(n)/O(n) -> S^1 is the culprit — it detects exactly one circle's worth of winding.) Because pi_1 is Z, a loop of Lagrangian subspaces carries a well-defined integer: how many times it winds around that one essential circle. That integer is the Maslov index. For n = 1 you can see it bare-handed: a loop of lines through the origin in R^2 is a path in RP^1, and its Maslov index counts how many half-turns the line makes (the doubling appears because a line returns to itself after a 180-degree, not 360-degree, rotation).
Why care about a winding number of subspaces? Because along a Lagrangian submanifold L inside M, or along the orbit of a Hamiltonian flow from Guide 2, the tangent Lagrangian T_p L sweeps out exactly such a loop in L(n) as you go around a closed path. The Maslov index of that loop is a topological invariant of the Lagrangian (more precisely of the loop with respect to a chosen reference), and it is the integer that appears as a phase correction in the WKB / semiclassical approximation of quantum mechanics — the 'Maslov correction' that fixes the naive Bohr-Sommerfeld quantization. It also feeds directly into Floer theory and the grading of intersection points, the modern frontier glimpsed in Guide 5. The index turns soft pictures of vanishing forms into hard integers you can compute and add.
Weinstein's theorem: a Lagrangian's neighborhood is its cotangent bundle
Guide 1 gave you the local rigidity of symplectic geometry: Darboux's theorem says all symplectic manifolds look identical near a point — there are no local invariants, every (M, omega) is locally (R^{2n}, omega_0). Weinstein's theorem is the relative version, upgrading 'near a point' to 'near a Lagrangian'. The Weinstein Lagrangian neighborhood theorem states: if L is a (compact) Lagrangian submanifold of (M, omega), then a neighborhood of L in M is symplectomorphic to a neighborhood of the zero section in the cotangent bundle (T*L, omega_canonical), by a symplectomorphism carrying L to the zero section. In words: up to symplectomorphism, a Lagrangian does not know which symplectic manifold it lives in — its surroundings are always just its own cotangent bundle.
The proof is a beautiful application of the Moser trick you met proving Darboux in Guide 1, so the method should feel familiar rather than new. You first build a diffeomorphism from a neighborhood of L to a neighborhood of the zero section that matches the symplectic forms to first order along L (a linear-algebra step using that T_p L is Lagrangian, so its symplectic-orthogonal complement is canonically T*_p L). Then you have two symplectic forms agreeing on L, and Moser's homotopy argument — interpolate omega_t = (1 - t) omega_0 + t omega_1, solve for a flow killing the difference — deforms one into the other while fixing L pointwise. The compactness of L is what keeps the Moser flow defined for all the time you need; drop compactness and the statement needs care.
Darboux (Guide 1): near a POINT, every (M, omega) ~ (R^{2n}, omega_0)
Weinstein (this guide): near a LAGRANGIAN L,
(nbhd of L in M, omega) ~ (nbhd of zero section in T*L, -d lambda)
with L mapped to the zero section
Moser trick (shared engine):
omega_t = (1-t) omega_0 + t omega_1, d/dt omega_t = -d alpha
choose X_t with iota_{X_t} omega_t = alpha, flow it -> pulls omega_1 back to omega_0Why Weinstein matters, and honest caveats
The payoff is that questions about Lagrangians become questions about cotangent bundles, where you have full calculus. Recall from the model examples that Lagrangians C^1-close to the zero section of T*L correspond to closed 1-forms on L. Combining this with Weinstein: any Lagrangian L' that is a small deformation of L can be described by a closed 1-form on L, and if H^1(L) = 0 (for instance L a sphere) every nearby Lagrangian is the graph of df, hence intersects L wherever df = 0 — and df must vanish at least at a max and a min. This is the seed of the Arnold conjecture and Lagrangian intersection theory: counting forced intersections via the topology (Morse theory, Betti numbers) of L. Fixed points of a Hamiltonian symplectomorphism become intersection points of its graph with the diagonal, both Lagrangians — Weinstein's 'everything is a Lagrangian' made quantitative.
- Check a candidate L is Lagrangian: confirm dim L = n (half of dim M) AND omega|_L = 0; for a graph in T*X, this reduces to checking the corresponding 1-form is closed.
- To compute a Maslov index, follow the loop of tangent Lagrangians T_p L in the Grassmannian L(n) and count its winding via the det^2 map into S^1, relative to your chosen reference Lagrangian.
- To understand a Lagrangian's neighborhood, invoke Weinstein: replace the ambient M near L by T*L with its canonical form, then translate the geometric question into closed/exact 1-forms on L.
- To force intersections, read off H^1(L) and the Morse theory of L: vanishing H^1 plus Morse inequalities lower-bound how many points a nearby Lagrangian must share with L.