MTH 868 — Lecture 07

Date: 2026-01-28
Topic: Regular points, critical points, submanifolds, and regular level sets


1. Setup and Recall

Let

  • ( F : M \to N ) be a smooth map between smooth manifolds.

Critical vs Regular Points and Values

  • A critical point ( p \in M ) is one where the differential [ dF_p : T_p M \to T_{F(p)} N ] is not surjective.

  • A regular point is a point where ( dF_p ) is surjective.

  • A critical value is a point in ( N ) that is the image of some critical point.
  • A regular value is a point in ( N ) that is not a critical value.

Stats intuition:
This is rank deficiency. A critical point is where your Jacobian drops rank.
Regular values are outputs where the Jacobian behaves nicely everywhere in the preimage.


2. Characterizing Critical Points for ( F : M \to \mathbb{R} )

Proposition

Let ( F : M \to \mathbb{R} ) be smooth.
Then ( p \in M ) is a critical point iff:

Equivalent conditions

  1. In any chart ( (U, \varphi = (x^1, \dots, x^n)) ) around ( p ), [ \frac{\partial F}{\partial x^i}(p) = 0 \quad \forall i ]

  2. There exists a chart where all partial derivatives vanish.


Why this works (lecture explanation)

  • The differential ( dF_p ) is a linear map [ dF_p : T_p M \to T_{F(p)} \mathbb{R} ]

  • Since ( \mathbb{R} ) is 1-dimensional, surjectivity means:
    • the linear map is non-zero
  • In coordinates, using the basis [ \left{ \frac{\partial}{\partial x^i} \bigg|_p \right} ] the differential is represented by the row vector [ [ \partial_1 F(p) \;\; \cdots \;\; \partial_n F(p) ] ]

  • This has rank 0 iff all entries are zero

Key invariant:
“All partial derivatives vanish” is a coordinate-independent statement, even though it looks coordinate-dependent.


3. Why Partial Derivatives Vanishing Is Chart-Independent

Suppose ( (U, x^i) ) and ( (V, y^j) ) are two charts near ( p ).

By the chain rule: [ \frac{\partial F}{\partial y^j} = \sum_i \frac{\partial x^i}{\partial y^j} \frac{\partial F}{\partial x^i} ]

In matrix form: [ [\partial_{y} F] = [\partial_x F] \cdot J(\varphi \circ \psi^{-1}) ]

  • The Jacobian of the transition map is invertible
  • Therefore: [ [\partial_x F] = 0 \iff [\partial_y F] = 0 ]

Stats analogy:
Gradient = 0 is invariant under reparameterization
Just like score functions transform covariantly.


4. Regular Submanifolds (Definition)

Let ( S \subset N ).

Definition

( S ) is a regular ( k )-dimensional submanifold of ( N ) if:

For every ( p \in S ), there exists a chart
( (U, \varphi = (x^1, \dots, x^n)) ) of ( N ) about ( p ) such that: [ S \cap U = { q \in U \mid x^{k+1}(q) = \cdots = x^n(q) = 0 } ]

These charts are called adapted charts


Geometric meaning

  • Locally, ( S ) looks like ( \mathbb{R}^k )
  • The remaining ( n-k ) coordinates are “normal directions”

Important:
Not every chart does this.
The definition is about existence, not universality.


Examples

  1. Linear subspaces
    • ( xy )-plane in ( \mathbb{R}^3 )
  2. Graphs of smooth functions

Let ( f \in C^\infty(\mathbb{R}) ).
Define: [ S = { (x, f(x)) } \subset \mathbb{R}^2 ]

Define the chart: [ \varphi(x,y) = (x, y - f(x)) ]

Then:

  • ( S ) corresponds to ( { y = 0 } )
  • Inverse chart: [ \varphi^{-1}(u,v) = (u, v + f(u)) ]

Thus graphs are smooth 1-manifolds.


5. Codimension

Definition

[ \operatorname{codim}(S) = \dim(N) - \dim(S) ]

  • Often easier to reason about constraints than dimensions
  • Each independent equation increases codimension by 1

6. Level Sets

Definition

Let ( F : N \to M ) be smooth, ( c \in M ).

The ( c )-level set is: [ F^{-1}(c) = { x \in N \mid F(x) = c } ]


Example: Sphere

Let: [ F(x,y,z) = x^2 + y^2 + z^2 ]

Then: [ F^{-1}(1) = S^2 \subset \mathbb{R}^3 ]

Compute differential: [ dF_{(x,y,z)} = [2x \;\; 2y \;\; 2z] ]

  • Surjective everywhere except at ( (0,0,0) )
  • Since ( (0,0,0) \notin F^{-1}(1) ), 1 is a regular value

7. Implicit Function Theorem View

Let: [ G = F - 1 \quad \Rightarrow \quad F^{-1}(1) = G^{-1}(0) ]

At the north pole ( (0,0,1) ): [ \frac{\partial G}{\partial z} = 2 \neq 0 ]

By the Implicit Function Theorem:

  • There exists a neighborhood where [ z = h(x,y) ]
  • The level set is locally the graph of a smooth function

Conclusion:
Regular level sets are smooth submanifolds.


8. Regular Level Set Theorem (Main Result)

Theorem

Let ( F : N \to M ) be smooth.

If ( c \in M ) is a regular value, then:

  • ( F^{-1}(c) ) is a smooth submanifold
  • Dimension: [ \dim F^{-1}(c) = \dim N - \dim M ]
  • Codimension: [ \operatorname{codim}(F^{-1}(c)) = \dim M ]

Proof sketch (what actually happened in lecture)

  1. Reduce to ( c = 0 ) via ( G = F - c )
  2. Use surjectivity of ( dF_p ) to find a nonzero partial derivative
  3. Reorder coordinates if needed
  4. Apply inverse / implicit function theorem
  5. Show local model is coordinate hyperplane

Stats translation:
Regular value = full-rank Jacobian everywhere on the constraint set
⇒ constraint surface is smooth, dimension reduced by number of constraints


9. Why This Matters (Big Picture)

  • This theorem underlies:
    • constraint optimization
    • Lagrange multipliers
    • manifolds defined by equations
    • likelihood surfaces under constraints
    • parameter identifiability

If you remember one thing:

Regular = full rank Jacobian ⇒ smooth geometry

Everything else is machinery to make that precise.


Comments