LECTURE 7: SARD’S THEOREM

1. Proof of Sard’s Theorem

Now we are ready to prove
Theorem 1.1 (Sard’s Theorem). For any smooth map f : M → N , the set of all
critical values is of measure zero in N .

Note that if n = dim N = 0, then there is no critical points (and thus no critical
values) at all. So in what follows we may assume n > 0.

Proof. Since we may take at most countable many coordinate charts Ui covering M so
that each f (Ui ) ⊂ Xi for some coordinate charts Xi of N , and since a countable union
of measure zero sets is still of measure zero, it suffices to prove the theorem for the
case that M = U ⊂ Rm and N = X ⊂ Rn are Euclidean open subsets.
First observe that if m < n, then the theorem holds trivially. In fact, in this case
we can prove that the whole image f (U ) is of measure zero in Rn . To see this, we
identify U with the subset U × {0} inside U × Rn−m ⊂ Rn . Obviously U × {0} has
measure zero in Rn . Now define a map f˜ : U × Rn−m → Rn via f˜(x, y) = f (x). Then
f (U ) = f˜(U × {0}). (f˜ is smooth since it is the composition of two smooth maps: the
projection map π : Rn → Rm and f : U → Rn .) Since f˜ is a smooth between spaces of
the same dimension, and U × {0} is of measure zero in Rn , we conclude that the image
f˜(U × {0}) is of measure zero in Rn .
We will proceed by induction. The theorem is certainly true for m = 0, since any
countable set has measure zero. We will proceed to prove that the theorem is true for
m assuming that it is true for m − 1. Let C be the set of all critical points of f , then
we need to show that f (C) is of measure zero in N . Denote
Cj = {x ∈ U | ∂ αf (x) = 0 for all |α| ≤ j}.
Obviously for any positive integer k,
f (C) = f (C \C1 ) ∪ f (C1 \C2 ) ∪ · · · ∪ f (Ck−1 \Ck ) ∪ f (Ck ).
Following J. Milnor, we will divide the proof into three steps:
Step 1: f (C \C1 ) has measure zero.
Step 2: f (Ci \Ci+1 ) has measure zero for each i.
m
Step 3: f (Ck ) has measure zero for large k, say for k ≥ n
.
Proof of step 1. For each x ∈ C \ C1 , we will find an open set Ux 3 x such that
f (Ux ∩ C) has measure zero. Since C \C1 can be covered by countably many of such
open sets (by second-countability), this implies f (C \C1 ) is of measure zero.

1
2 LECTURE 7: SARD’S THEOREM

∂f
Note that if n = 1, then a point x is a critical point of f if and only if ∂x i (x) = 0

for all i. It follows C\ C1 = ∅, so f (C\C1 ) must be of measure zero. So we may assume

∂f1
m ≥ n > 1. Since x 6∈ C1 , there is some partial derivative, say ∂x 1 , is not zero at x.

Consider
h : U → Rm , h(x) = (f1 (x), x2 , · · · , xm ).
Then dhx is non-singular. According to the inverse mapping theorem, h maps a
neighborhood Ux of x diffeomorphically onto an open set V in Rm . The composi-
tion g = f ◦ h−1 will then map V into Rn . Moreover, since h−1 is a diffeomorphism on
V , dh−1 is a linear isomorphism everywhere in V . So the set of critical values of g is
exactly f (Ux ∩ C).
Note that the map g we constructed is of the form
g(t, x2 , · · · , xm ) = (t, g2 , · · · , gn ).
So for each t, g induces a map g t : ({t} × Rm−1 ) ∩ V → {t} × Rn−1 . Moreover,
!
1  0
dg = ∗ ∂(gt )i .
∂xj
i,j≥2

It follows that a point in ({t} × Rm−1 ) ∩ V is critical for gt if and only if it is critical
for g. But by the induction hypothesis, Sard’s theorem is true for m − 1, i.e. is true
for each gt . So the set of critical values of gt has measure zero in {t} × Rn−1 . Finally
by applying Fubini’s theorem, we see that the set of critical values of g is of measure
zero.
Proof of step 2. For each x ∈ Ci\Ci+1 , one can find some multi-index α with |α| = i,
so that
• the partial derivative w := ∂ α f vanishes on Ci ,
∂w
• at least one first order partial derivative of w, say ∂x1
, does not vanish at x.
Again by applying the inverse mapping theorem, we conclude that
h : U → Rm , h(x) = w(x), x2 , · · · , xm

maps a neighborhood Ux of x diffeomorphically onto an open set V in Rm . (To get a

better understand and/or to avoid possible mistakes, you may want to think about the
meaning of this map for m = 1.) By construction, h carries Ci ∩ Ux into the hyperplane
{0} × Rm−1 . Again we consider the map g = f ◦ h−1 . Then the critical points of g of
type Ci are all in the hyperplane {0} × Rm−1 . Let
ḡ : ({0} × Rm−1 ) ∩ V → Rn
be the restriction of g. Then the set of critical points of g of type Ci coincides with the
set of critical points of ḡ. By induction, the set of critical values of ḡ is of measure zero
in Rn . It follows that the image of the critical points of g of type Ci is of measure zero.
Therefore, f (Ci ∩ Ux ) is of measure zero. Since Ci \Ci+1 can be covered by countable
many such sets Ux , f (Ci \Ci+1 ) is of measure zero.
 LECTURE 7: SARD’S THEOREM 3

Proof of step 3. Let Q ⊂ U be a cube whose sides are of length δ. We will prove
that for k > m
n
− 1, f (Ck ∩ Q) has measure zero. Since Ck can be covered by countably
many such cubes, this implies that f (Ck ) has measure zero.
From Taylor’s theorem, the compactness of Q and the definition of Ck , we see that
f (x + h) = f (x) + R(x, h),
where |R(x, h)| < a|h|k+1 for x ∈ Ck ∩ Q, x + h ∈ Q, and the constant a depends only
on f and Q. Now we subdivide Q into rm cubes whose sides are of length rδ . Let Q1
be a cube of subdivision that√contains a point x ∈ Ck . Then any point of Q1 can be
written as x + h with |h| < m rδ . It follows that f (Q1 ) lies in a cube with sides of
b √
length rk+1 centered about f (x), where b = 2a( mδ)k+1 is a constant. So f (Ck ∩ Q)
is contained in the union of at most rm cubes having total volume
b
Vol ≤ rm ( )n = bn rm−(k+1)n .
rk+1
m
Since k > n
− 1, we see Vol → 0 as r → ∞. It follows that f (Ck ∩ Q) is of measure
zero. 
Remark. On the other hand, the set of critical values could be a dense subset. For
example, we can list all rational numbers as Q = {r1 , r2 , · · · }. Then we take a smooth
function f0 defined on R such that
supp(f0 ) ⊂ (−1/3, 1/3) and f0 ≡ 1 on (−1/4, 1/4).
Let
∞
X
f (x) = rk f0 (x − k).
k=1
Then each k ∈ N is a critical point of f , and thus the set of critical values of f contains
f (N) = Q.
Remark. Sard’s theorem holds for C r maps between C r manifolds, provided r ≥ 1 +
max(m − n, 0). Moreover, this is sharp. However, the proof above does not work in
this more general setting. (Can you tell why?)

2. An alternative form of Sard’s Theorem

Next we will give an alternative form of Sard’s Theorem. (Reason: In many ap-
plications, one has to handle maps between infinitely dimensional manifolds, called
Banach manifolds. The original form of Sard’s Theorem cannot be extended to this
setting because we don’t have a Lebesgue measure in the infinite dimensional setting.)
Before we state the alternative form, we first give a definition that can be viewed
as the analogue in general topological spaces of the word “of full measure” in measure
spaces:
Definition 2.1. A set A in a topological space is called residual (or “second Baire
category”) if it is a countable intersection of dense open sets.
4 LECTURE 7: SARD’S THEOREM

For example, the set of all irrational numbers, R\Q, is a residual set in R. However,
the set of rational numbers, Q, is not residual in R. Usually when we say some property
is “generic”, we mean that property holds for a residual set.
We now prove
Theorem 2.2 (Sard’s Theorem, an alternative form). Let f : M → N be smooth.
Then the set of regular values of f is residual in N .
[Idea]: Since the set of regular values has full measure, it has to be dense. It remains
to show that the set of regular values is “nearly open”. This is reasonable, since the
set of regular points is open:
x is a regular point ⇔ f is a submersion at x ⇔ f is a submersion near x.
However, this is not enough, since
f ({regular points}) 6= {regular values}.
However, we do have
f ({critical points}) = {critical values}
and we know {critical points} is closed. Unfortunately this is still not enough to deduce
that {critical values} is closed, because in general
f (closed set) 6= closed set.
(See the example at the end of last section.) The solution to this last issue is to apply
a standard trick:
f (compact set) = compact set.

Proof. Again we denote the set of critical points of f by C. We will show that every
x ∈ C admits a neighborhood Ux such that when restricted to U x , the set of regular
values of f , N \ f (C ∩ Ux ), is dense and open in N . Since C can be covered by at most
countable many such open sets {U1 , U2 , · · · }, we conclude that the set of regular values
[ \
N \f (C) = N \ f (C ∩ Ui ) = N \f (C ∩ Ui ),
i i
which is a residual set.
To show the existence of such Ux , we choose a coordinate neighborhood Ux of x so
that Ux is compact. Since C is closed in M (see the explanation before the proof), we
see that C ∩ Ux is compact. So f (C ∩ Ux ) is compact, and thus closed in N . It follows
that N \ f (C ∩ Ux ) is open in N .
On the other hand, according to Sard’s theorem we proved, f (C ∩ Ux ) has measure
zero. So for any y ∈ N and any small neighborhood V of y in Y , one must have
V ∩ (N \f (C ∩ Ux )) 6= ∅,
otherwise V ⊂ f (C ∩ Ux ), which contradicts with the fact that f (C ∩ Ux ) has measure
zero. It follows that N \f (C ∩ Ux ) is dense in N .