From a vector field to a flow
In Guide 2 you met a vector field X as a smooth choice of tangent vector X_p in T_p M at every point p. So far it has been a static field of arrows. Now make it dynamic: imagine dropping a speck of dust at p and letting the arrows push it along, always moving in the direction X says, at the speed |X| says. The path it traces is the integral curve through p — a curve gamma(t) with gamma(0) = p and gamma'(t) = X_{gamma(t)} for all t. Collecting one such curve through every point at once gives the flow of X.
Concretely the flow is a map phi: (-epsilon, epsilon) x M -> M written phi_t(p), where phi_0 = identity and phi_{s+t} = phi_s composed with phi_t. That last equation is the heart of the matter: flowing for time s, then time t, is the same as flowing for time s+t. So {phi_t} is a one-parameter group of diffeomorphisms — for each fixed t, phi_t is a diffeomorphism of M with inverse phi_{-t}. The existence and smoothness of phi_t is exactly the existence-and-uniqueness theorem for ODEs, dressed up on a manifold.
The Lie bracket: do two flows commute?
Given two vector fields X and Y, run their flows in turn. Flow along X for time t, then along Y for time t, then back along X for time t, then back along Y for time t. On R^n with straight-line flows you return exactly to where you started. On a curved manifold — or when X and Y genuinely interact — you do not: there is a tiny residual displacement. To first order it vanishes, but the second-order leftover is a genuine tangent vector, and that vector is the Lie bracket [X, Y].
There is a cleaner algebraic face of the same object. Think of a vector field as a first-order differential operator acting on functions: X(f) is the directional derivative of f along X. Then XY and YX are both second-order operators, but their second-order parts cancel, and what survives is again first-order — a vector field. That is the bracket: [X, Y](f) = X(Y(f)) - Y(X(f)). The two pictures agree, and the bracket measures the failure of the flows of X and Y to commute.
[X, Y](f) = X(Y(f)) - Y(X(f)) (operator definition) In coordinates, X = a^i d/dx^i, Y = b^j d/dx^j: [X, Y]^k = a^i (db^k/dx^i) - b^i (da^k/dx^i) Key identities: [X, Y] = -[Y, X] (antisymmetry) [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y]] = 0 (Jacobi) flows commute <=> [X, Y] = 0
Antisymmetry and the Jacobi identity say the vector fields on M form an (infinite-dimensional) Lie algebra. This is no accident: when M is a Lie group, the left-invariant vector fields form a finite-dimensional Lie algebra under exactly this bracket, and the exponential map sends a left-invariant field to its flow at time 1. The bracket you computed by chasing dust specks is the same bracket that organizes SU(2) and every other symmetry group.
Frobenius: when do brackets close up?
Now suppose at each point you are handed not one direction but a whole k-dimensional plane D_p inside T_p M, varying smoothly — a distribution of rank k. Ask the integrability question: can you thread these planes together into k-dimensional submanifolds tangent to D everywhere, the way streamlines thread a single vector field into curves? Sometimes yes (the planes 'foliate' M into leaves), sometimes flatly no.
The Frobenius theorem gives the exact, checkable condition: D is integrable if and only if it is involutive — whenever X and Y are vector fields lying in D, their bracket [X, Y] also lies in D. The Lie bracket is precisely the obstruction. If the planes 'twist' so that bracketing two tangent fields pokes out of the distribution, no integral submanifold can exist; if the bracket always stays inside, the leaves are guaranteed. This is why the bracket was worth defining so carefully.
A vivid non-example lives in contact geometry. On R^3 take the distribution killed by the 1-form dz - y dx; these planes twist so violently that any two points can be joined by a path staying tangent to them, so there is no surface tangent to the distribution at all. Maximal non-involutivity, the opposite extreme from a foliation, is exactly a contact structure — the bracket pokes out as hard as it possibly can.
Reading shape from one function: Morse theory
Now flip the whole subject around. Instead of building structure on M, take a single smooth function f: M -> R and ask what it tells you about M. Its critical points are where the differential df vanishes — the points where the gradient flow of f stalls. By Sard's theorem almost every value is regular, so critical points are sparse; the interesting information sits exactly at them. The function f is a Morse function if every critical point is nondegenerate: the Hessian (the matrix of second derivatives in any chart) is invertible there.
At a nondegenerate critical point the Morse lemma says you can choose coordinates in which f looks exactly like a sum of plus-or-minus squares: f = -x_1^2 - ... - x_k^2 + x_{k+1}^2 + ... + x_n^2. The number k of minus signs is the Morse index — the dimension of the directions in which f goes down. Picture the height function on a torus standing upright: a minimum (index 0) at the bottom, a maximum (index n) at the top, and two saddles (index 1) at the inner waist. Index counts how 'saddle-like' the point is.
- Pick a Morse function f on M (height functions for embedded M almost always work, by Sard).
- Sweep the sublevel set M_a = { f <= a } upward as a increases.
- Crossing a regular value changes nothing topologically — M_a only deforms smoothly.
- Crossing a critical value of index k attaches one k-cell (a 'handle') to M_a.
- When the sweep finishes, M is built as a CW complex with one cell per critical point.
This is the punchline: a single function reconstructs the homotopy type of the whole manifold, cell by cell. Counting critical points of each index gives the Morse inequalities — the number of index-k critical points is at least the k-th Betti number b_k, so any Morse function must have at least sum of b_k critical points. The alternating sum of critical counts equals the Euler characteristic. Topology that you would otherwise compute through de Rham cohomology falls out of watching where one function levels off.
How the rung fits together
Step back and the five guides form one arc. Charts gave us M; the tangent bundle gave us vectors; the regular value theorem and transversality told us when preimages and intersections are manifolds; the degree of a map counted preimages with sign. This guide added the verbs: flows move points, the bracket measures their non-commuting, Frobenius decides when a field of planes integrates, and Morse functions turn analysis on M back into its topology.
Be honest about where this stops. We stated the Morse inequalities and the CW structure but did not prove the handle-attachment lemma carefully; that needs the gradient flow of f and a transversality argument, a chapter in Milnor, not a paragraph. The deeper refinement — the Morse index theorem relating the index of a geodesic to conjugate points — belongs to the Riemannian track and is genuinely a course. Treat this guide as a true map of the terrain, not a substitute for walking it.