LECTURE 4

Example 29. Let us consider a more abstract example without obvious visualisation. Let
O(n) be the set of orthogonal n × n real matrices. The set M (n) of all n × n real matrices is
2
the Euclidean vector space Rn so trivially a smooth n2 -manifold. The set S(n) of all n × n
real symmetric matrices identifies with the Euclidean vector space Rn(n+1)/2 so also trivially
a smooth (n(n + 1)/2)-manifold. Then we define f : M (n) → S(n) by f (A) = AT A. This
function is polynomial in each entry so smooth. Its differential is DfA (H) = H T A + AT H.
Let us prove that Id, which is trivially in the image f (M (n)), is a regular value. Indeed
let any A ∈ M (n) so that AT A = Id and we want to solve DfA (H) = H T A + AT H = B
for an arbitrary B ∈ S(n) (surjectivity). The symmetry of B implies B = (B T + B)/2
and it is enough to solve AT H = B/2. Left-multiplying by A we get H = AB/2 which
solves the problem. We deduce finally that O(n) is a submanifold of M (n) of dimension
n2 − n(n + 1)/2 = n(n − 1)/2. (Note that O(n) is both a manifold and a group, and the
group operation is compatible with the manifold structure i.e. smooth: this is called a Lie
group.)

The pre-image theorem raises the question: how abundant are regular values? The Sard
theorem is a deep result that says intuitively that because critical values are the result of
“local folding” (lower dimension on the infinitesimal variations) they will be very rare indeed.
The proper tool to assess how rare they are is measure theory so let us start with a revision.

Definition 30 (Revision from Probability & Measure). Given n ∈ N∗ , a set A ⊂ Rn has

measure zero if for any ε > 0 it can covered by a countable collection of rectangular
subsets
Q whose volumes sum to less than ε. ThisP means: for P > 0 there are (Ri )i≥1 with
any ε Q
Ri = nj=1 [ai,j , bi,j ] such that A ⊂ ∪i≥1 Ri and i≥1 |Ri | = i≥1 nj=1 (bi,j − ai,j ) ≤ ε.

We now extend this definition to manifolds:

Definition 31. Given X n-manifold, a set A ⊂ X has measure zero in X if for every
parametrisation ϕ : U ∈ Rn → U ⊂ X, the set ϕ−1 (U ∩ A) has measure zero in Rn .

Remark 32. Note in particular that if A ⊂ Rn has measure zero, A has empty interior, i.e.
it cannot contain any non-empty open subset of Rn . And the same is true for A ⊂ X of
measure zero in a manifold X by using the parametrisations.

Example 33. If X n-manifold, and Y ⊂ X submanifold of dimension m < n, then Y has

measure zero in X (prove it with parametrisations).

Theorem 34 (Sard’s theorem). Given X and Y manifolds and f : X → Y a smooth map,

then the set of critical values f (C) is a set of measure zero in Y . Therefore the set of regular
values is dense in Y .

The proof is strictly non-examinable. If there is some interest I can add notes on Moodle.

8
2.5 Transversality
We have a good concept of those y ∈ Y so that the solutions to the equation f (x) = y
between two manifolds X and Y will form a manifold: the concept of regular values. Can
we find a more general notion when we replace y ∈ Y by a submanifold Z ⊂ Y and want to
know whether f −1 (Z) is a submanifold of X?

Definition 35. Given X and Y manifolds, f : X → Y smooth map and Z ⊂ Y submanifold,

we say that f is transversal (or transverse) to Z (written f ⋔ Z) if

∀ x ∈ f −1 (Z), ImDfx + Tf (x) Z = Tf (x) Y.

Remark 36. • In this definition the sum does not have to be direct.

• If f −1 (Z) = ∅, the statement is trivially true and f ⋔ Z.

• If Z = {y} is a singleton, this definition coincides with y being a regular value with the
convention Ty {y} = {0} (manifold of dimension zero).

• If X ⊂ Y and f is simply the canonical inclusion, when f ⋔ Z we say that X is

transversal to Z (X ⋔ Z), and we have Tx X + Tx Z = Tx Y at every x ∈ X ∩ Z.

• (Non-examinable) The notion of transversality is very powerful, partly because it is

“stable” (open), in the sense that if we smoothly perturb a transversal map it will
still be smooth and transversal. It also “generic” (dense) in the sense that the lack
of transversality corresponds to “rare” configurations where tangent vector spaces are
both contained in a lower-dimensional subspace, so in many situations we can “wiggle”
f or Z to restore transversality.

Example 37. Consider X = R, Y = R2 and f : X → Y defined by f (t) = (0, t), and

Z = R × {0} ⊂ Y . Then f ⋔ Z since f −1 (Z) = {0} and at the unique pre-image 0 ∈ X of Z,
we have ImDf0 + T(0,0) Z = {0} × R + R × {0} = R2 = T0 Y . It also means that (confirming
the intuitive idea associated with the word “transversal”) {0} × R ⋔ R × {0}. The previous
reasoning would still work with f (t) = (t, αt) with α ̸= 0 (check it), but f (t) = (t, t2 ) is not
transversal to Z: indeed f −1 (Z) = 0 again but ImDf0 = R × {0}.

With this concept we can now prove a more general version of the pre-image theorem:

Theorem 38 (Generalised pre-image theorem). Given X and Y manifold, Z submanifold

of Y , f : X → Y smooth map transversal to Z, then f −1 (Z) is a submanifold of X and the
codimension of f −1 (Z) in X is equal to the codimension of Z in Y .

Proof. We start by realising locally Z as the solution to a smooth implicit equation on Y .

Consider z ∈ Z, then there is V open in Y around z and ϕ : V ⊂ Rm → V diffeomorphism
with m = dim Y and V open around 0 in Rm . Then we use that Z is also a manifold, with say
dimension k = dim Z: there is U open in Z around z and ψ : U ⊂ Rk → U diffeomorphism
with U open around 0 in Rk . By taking U smaller if necessary we can assume U ⊂ V. Then
if ι denotes the canonical inclusion Z → Y , then we define F = ϕ−1 ◦ ι ◦ ψ : U ⊂ Rk →
V ⊂ Rm . Then F is smooth by composition and DF0 = (Dϕ0 )−1 ◦ Dι ◦ Dψ0 where Dι
is simply a canonical inclusion Rk → Rm between the tangent spaces; hence DF0 has rank

9
k by composition, and reordering if necessary we assume the upper-left k × k entries form
an invertible matrix. Then we define G : U × Rm−k → Rm by G(u, v) = F (u) + (0, v). 
Then, writing F = (F1 , F2 ) for the k first and m − k last components, DG0 = ∂∂uu FF12 Id 0

is surjective so invertible and the inverse function theorem for manifold implies that G is
a local diffeomorphism around zero and if we call h1 , . . . , hm−k : U → R the last m − k
coordinates of G−1 ◦ ϕ−1 : V ⊂ Y → U ⊂ Rm , these are smooth functions and for y ∈ V, the
fact that y ∈ Z is equivalent to h1 (y) = · · · = hm−k (y) = 0: indeed G−1 ◦ ϕ−1 (y) = (u, 0)
equivalent by left-composing with G to ϕ−1 (y) = G(u, 0) = F (u) = ϕ−1 ◦ ι ◦ ψ(u) equivalent
by left-composing with ϕ to y = ψ(u) so y ∈ Z. Let us denote H = (h1 , . . . , hm−k ). Then
f −1 (Z) ∩ f −1 (V) is the zeroes of H ◦ f in f −1 (V), and DHy : Ty Y → Rm−k has kernel Ty Z,
for y ∈ U , and is therefore surjective by dimensionality. We deduce that 0 ∈ Rm−k is a
regular value H ◦ f : f −1 (V) → Rm−k : its differential at x ∈ f −1 (V) so that H(f (x)) = 0
(hence f (x) ∈ Z ∩ V) is D(H ◦ f )x = DHf (x) ◦ Dfx , which is surjective if Dfx hits all the
directions not in the kernel of DHf (x) , i.e. ImDfx + Tf (x) Z = Tf (x) Y , which is precisely
the transversality condition. Hence the pre-image theorem implies that (H ◦ f )−1 (0) is a
submanifold in X of dimension dim X − (m − k), hence its codimension in X is m − k which
is the codimension of Z in Y .

The following corollary is immediate by particularising the previous theorem to the case
X ⊂ Y and f inclusion, but is important enough to deserve a separate statement.

Corollary 39. The intersection of two transversal submanifolds X and Z of Y is a subman-

ifold of codimension codim(X ∩ Z) = codimX + codimZ.

10