Bifurcation Theory

Outline. The first three sections treat problems in dimension 1. A final section shows that for
N > 1 a reduction to the scalar case is possible. The Hopf-Andreev bifurcation is special to
N > 1. The main tool is the Implicit Function Theorem.

1 The single bifurcation curve case

This section title looks like a nonsequitor since bifurcation is about more than one solution. That
is what the "bi" indicates. However, a curve in (a, x)-space which doubles back on itself can
have more than one point for a given parameter value a.

Study equilibria of x0 = f (a, x). The emphasis is on the dependence on a. Equivalently, one
studies the zero set {(a, x) : f (a, x) = 0}. Viewed that way the differential equation has
disappeared. One of the key properties of the equilibria are their stability. That reintroduces
dynamics. Suppose that (a, x) is an equilibrium, that is

f (a, x) = 0 . (1.1)

The Implicit Function Theorem guarantees that if

∇a,x f (a, x) 6= 0 (1.2)

then near (a, x) the level set is a smooth curve with normal vectors parallel to ∇a,x f (a, x).
The curve is a graph x = k(a) near a when fx (a, x) 6= 0. It is locally a graph a = h(x) when
fa (a, x) 6= 0. The derivatives of all orders of h or k can be computed by implicit differentiation.

1.1 The no bifurcation case

If
∂f (a, x)
6= 0 (1.3)
∂x
the Implicit Function Theorem implies that {f = 0} is locally a graph

x = k(a) , k(a) = x .

1
 x

x=k(a)

No bifurcation

With (1.3) there is no bifurcation. For a ≈ a, there is a nearby equilbrium, k(a) ≈ k(a) = x.
Differentiate f (a, k(a)) = 0 with respect to a to find fa + fx k 0 = 0. The slope is given by

−fa (a, k(a))

k 0 (a) = . (1.4)
fx (a, k(a))

Next compute the stability of the equilibria so long as (1.3) holds, The linearized equation at
the equilibrium x = k(a) is y 0 = fx (a, k(a))y. The coefficient, fx (a, k(a)) is not equal to zero.
The equilibrium x of x0 = f (a, x) is asymptotically stable when fx (a, x) < 0. In the opposite
case fx (a, x) > 0, orbits converge to x as t → −∞. Since fx (a, x) 6= 0 its sign does not change
if a only changes a little. The nearby equilibria remain either attracting or repelling. They can
change stability only if fx vanishes.

1.2 The bifurcation case

A more interesting case is when (1.2) holds and

fx (a, x) = 0 . (1.5)

Then fa (a, x) 6= 0 so the Implicit Function Theorem implies that there is a curve of equilibrium
given by
−fxx (a, x)
a = h(x), with h(x) = a, h0 (x) = 0, h00 = . (1.6)
fa (a, x)
To compute the last formulas, differentiate f (h(x), x) = 0 with respect to x to find

fa (h(x), x) h0 (x) + fx (h(x), x) = 0 .

implying h0 (x) = 0. Differentiate again with respect to x to find,

 
faa h0 + fax h0 + fa h00 + fxa h0 + fxx = 0 .

At x = x, h0 (x) = 0 eliminates three of the five terms. This yields the formula for h00 (x).
The figure at the start of §1 is such a curve with h0 = 0 at the point with the vertical tangent
and h00 < 0 at that point so the curve lies (locally) to the left of this vertical tangent. This is a
subcritical bifurcation since the equilibria exists for parameter values a smaller than a.

Exercise 1.1 Suppose that the hypotheses of this section hold and that h00 (x) 6= 0. Show that for
x ≈ x the equilibria on {f = 0} have opposite stabilities for x > x and x < x. Hint. Show that
the leading order of the Taylor expansion is fx (h(x), x) ≈ fxx (a, x)(x − x). Discussion. At the
bifurcation point a stable and unstable branch meet and annihilate.

2
2 Two curves crossing
In order to have a more complicated zero set near a, x, one must have ∇a,x f (a, x) = 0. Then
the leading order Taylor expansion of f equilibrium is
    
faa (a, x) fax (a, x) a−a a−a
f (a, x) = , + h.o.t. .
fax (a, x) fxx (a, x) x−x x−x

The matrix of second derivatives of f is real and symmetric. It therefore has two real eigenvalues,
possibly equal. Since singular matrices are rare it is expected that neither of the two eigenvalues
is equal to zero.
If both are positive then f has a strict local minimum at a, x so f > 0 on a punctured neighbor-
hood of a, x and the equilibrium is isolated. Similarly the equilibrium is isolated if the matrix
has two negative eigenvalues.
The remaining expected case is when the matrix of second derivatives has one positive and one
negative real eigenvalue. Then the graph z = f (a, x) near a, x is saddle shaped and the level set
at height z = 0 consists of two curves crossing transversally at a, x. The zero set is a curvy X
shaped figure. This shows that after a single smooth curve the next expected behavior for the
set {f = 0} is an X shaped crossing. These are naturally called bifurcation points.

3 Transcritical and pitchfork

The analysis of the X shaped crossings is easiest when one of the branches is known. There are
many applications, where x = x is an equilibrium for all values of a. The study of the X shape in
this case reduces to studying the second branch near (a, x). This is the problem that we analyze.
Translating x coordinates we may suppose that x = 0. Seek a second curve of equilibria through
(a, x) = (a, 0). We are given that

f (a, 0) = 0 for all a . (3.1)

In order for there to be two branches intersecting at (a, x) is it necessary that (1.2) be violated,
that is
∇a,x f (a, x) = 0 . (3.2)
That fa = 0 follows from (3.1). The second condition

fx (a, x) = 0 (3.3)

is a necessary condition for a to be the intersection point of two curves of equilibria.

To analyze further the strategy is to separate out the root x = 0 using the identity
Z 1
∂f (a, θx)
f (a, x) = x g(a, x), g(a, x) := dθ . (3.4)
0 ∂x

This is a special case of Taylor’s Theorem and follows easily from the Fundamental Theorem of
Calculus applied to
dκ
κ(θ) := f (a, θx) , with = fx (a, θx) x .
dθ

3
Then Z 1 Z 1
dκ ∂f (a, θx)
f (a, x) = κ(1) = κ(1) − κ(0) = dθ = x dθ ,
0 dθ 0 ∂x
proving (3.4).
If f ∈ C k with k ≥ 1, then g ∈ C k−1 . The set {f = 0} is the union of {x = 0} and the set
{g = 0}.
Differentiating with respect to x yields fx = g + xgx . Setting x = 0 yields g. The necessary
condition fx (a, x) = 0 for bifurcation is equivalent to g(a, x) = 0.

3.1 The general case

To investigate {g = 0} using the Implicit Function Theorem, we need to find the partial deriva-
tives of g. Suppose that f ∈ C 3 with k ≥ 3. Differentiating the identity f = xg yields

fx = g + xgx , fa = xga , faa = xgaa , fax = xgax + ga , fxx = 2gx + xgxx .

Evaluating at (a, 0) using (3.1) yields the values

g, 0, 0, ga , 2gx .

The preceding computation yields

ga (a, 0) = fax (a, 0) , gx (a, 0) = fxx (a, 0)/2 . (3.5)

The Implicit Function Theorem implies the set {g = 0} near (a, 0) is a C k−1 curve a = h(x)
provided that ga (a, x) 6= 0. For the original problem this yields the following result.

Theorem 3.1 If in addition to (3.1) and (3.3) one has fax (a, 0) 6= 0 then near (a, 0) the zero
set of f consists of {x = 0} and a C k−1 curve a = h(x) intersecting x = 0 transversally at (a, 0).
Precisely, h(0) = a and h0 (0) = −fxx (a, 0)/2fax (a, 0).

Exercise 3.1 . Derive the formula for h0 (0). Hint. Differentiate g(h(x), x) = 0 with respect
to x then use (3.5).

3.2 The transcritical case

The name derives from the fact that {g = 0} crosses the line a = a transversely with nonzero
and finite slope. There are subbranches on each side of {a = a}. The analysis is a series of
exercises.

Exercise 3.2 Suppose that fax (a, 0) > 0 and fxx (a, x) < 0. The curve g = 0 then has positive
slope so near (a, 0) the signs of x and a − a agree as in the figure.

4
 x

a=h(x)

Transcritical Bifurcation

Show that the equilibrium x = 0 is asymptotically stable for a on the left of and near a. Show
that it is unstable for a ≈ a to the right. Hint. The stability is determined by considering the
linearized equation y 0 = fx (h(x), x) y. One needs the sign of fx (a, 0). Expand about a = a.
Discussion. This equilibrium losses its stability as a passes through a from left to right. You
should think of fx (a, 0) as a 1 × 1 matrix with a negative eigenvalue for a < a that crosses to
positive when a increases past a.

Exercise 3.3 With the hypotheses of the preceding exercise, show that the equilibria on {g = 0}
are asymptotically stable for x small positive and unstable for x small negative. Draw sketches of
the branches indicating with an s or u stable and unstable branches. Hint. Expand fx (h(x), x)
about x = 0. Discussion. The stabilities of the two halves of {x = 0} (resp. {g = 0}) are
opposed on the opposite sides of the equilibrium. The {x = 0} branch is stable to the left and the
{g = 0} branch is stable to the right. This is called exchange of stability. There is an analogous
exchange result when {x = 0} is unstable to the left of (a, 0)

Remark. One has analogous results whenever fax (a, x) 6= 0 and fxx (a, x) 6= 0. This guarantees
that g = 0 has finite nonzero slope at a. The rest of the analysis is the same with care taken for
all sign possibilities. A model is Exercise 3.6.

3.3 The pitchfork

The pitchfork occurs when ∇a,x g(a, 0) 6= 0 so {g = 0} is locally smooth, and in addition {g = 0}
has vertical slope and nonvanishing curvature at (a, 0) The vertical slope holds if and only if the
normal to {g = 0} at this point is horizontal, if and only if gx (a, 0) = 0. Using (3.5) this holds
if and only if fxx (a, 0) = 0. Then thanks to (3.5), ∇a,x g 6= 0 holds if and only if fax (a, 0) 6= 0.
These conditions are assumed for the remainder of this section.
Then, {g = 0} is given locally as a graph a = h(x) with h(0) = a and h0 (0) = 0. The sign of
h00 (0) predicts which way {g = 0} breaks in the generic case of nonvanishing curvature. The
supercritical case h00 > 0 is sketched.

5
 x

a=h(x)

a
h’=0
h’’>0

Supercitical Pitchfork

Exactly as in the derivation of (1.6),

−gxx (a, 0)
h00 (0) = . (3.6)
ga (a, 0)

Exercise 3.4 Continue the computation at the beginning of §3.1 to find a formula for fxxx .
Evaluate at x = 0 to show that fxxx (a, 0) = 3 gxx (a, 0). Using (3.5) and (3.6) find the formula
−fxxx (a, 0)
h00 (0) = .
3 fax (a, 0)

The next exercise examines the stability of the equilibria on {g = 0}. The stability is determined
by the sign of fx .

Exercise 3.5 i. Compute the derivatives of fx (h(x), x) with respect to x. Evaluate at x = 0. ii.
Find the leading Taylor expansions of fx (a, 0) about a = a and fx (h(x), x) about x = 0. Ans. i.

dfx (h(x), x) d2 fx (h(x), x) 2

= 0, = fxxx (a, 0) .
dx x=0 dx2 x=0 3
2
ii. fx (a, 0) ≈ fxa (a, 0)(a − a), fx (h(x), x) ≈ 3 fxxx (a, 0) x2 /2 .

Exercise 3.6 i. Use the preceding exercises together with formula (3.5) to show that the equilib-
ria on {x = 0} near and on opposite sides of (a, 0) have opposite asymptotic stabilities. ii. Show
that the equilibria near (a, 0) on {g = 0} have the same asymptotic stability as the equilibria
(a, 0) on the opposite side of {a = a}. Hint. The pitchfork can face left or right. The equilibria
on the handle can be stable or unstable. Thus there are four possibilities. Sketch some of the
possibilities indicating with an s or u the stable and unstable branches. Discussion. Replacing
f by −f does not change the set of equilibria and corresponds to reversing the direction of time.
It changes stable equilibria to unstable and vice versa. This remark reduces the set of possibilities
from four to two.

4 The case N > 1

Surprisingly, a large part of the multidimensional case can be reduced to the scalar analysis just
performed. Consider a system of N nonlinear equations

X 0 = F (a, X), F (a, X) = 0 .

6
Here a is a real parameter and X takes values in RN .
Equilibria satisfy
F (a, X) = 0 . (4.1)
Equation (4.1) is a vector so equivalent to N scalar equations. There are N equations for the N +1
unknowns a, X. The solutions set should be a one dimensional object, a curve in (N + 1)-space.
The first two sections below are dedicated to the study of a single curve of equilibria. The final
section describes Hopf-Andreev bifurcation which is a phenomenon not present in the scalar case
N = 1.

4.1 The no bifurcation case

Continuing the equation count started above, for each fixed a equation (4.1) is N equations for
the N unknowns. In unexceptional cases one expects to have unique solutions X(a) ≈ X(a)
for a ≈ a. If the resulting curve of solutions (a, X(a)) were differentiable, then differentiating
F (a, X(a)) = 0 with respect to a using the chain rule yields
dX dX
Fa + FX = 0, = −(FX )−1 Fa ,
da da
involving the N × N matrix valued function FX (a, X(a)). If the N × N matrix FX (a, X) is
invertible, then the Implicit Function Theorem asserts that the set {F = 0} of equilibria is given
in a neighborhood of (a, X) by a smooth curve
dK
X = K(a) , = −(FX (a, K(a))−1 Fa (a, K(a)) ,
da
This is the vector version of the scalar no bifurcation result, with analogous formulas.

4.2 One curve of sub critical or super critical solutions

There is a surprising reduction to the scalar case.
When FX is not invertible, the zero set can still be a smooth curve of equilibria. The implicit
function theorem assures that this is so when

rank Fa,X (a, X) = N . (4.2)

On the other hand one always has

rank FX (a, X) ≤ rank Fa,X (a, X) ≤ rank FX (a, X) + 1

When FX is not invertible it has rank < N so for (4.2) to hold one must have

rank FX (a, X) = N − 1 (4.3)

Hypothesis (4.2)-(4.3) is equivalent to the assumption that there is a subset of N − 1 coordinates

X̂ := x1 , . . . , xj−1 , xj+1 , . . . , xN

and a set of N − 1 equations

F̂ := F1 , . . . , Fm−1 , Fm+1 , . . . , FN

7
so that the N − 1 × N − 1 matrix F̂X̂ (a, X) is invertible. In that case the Implicit Function
Theorem asserts that the set of points F̂ (a, X) = 0 is locally two dimensional parametrized by
(a, xj ) ≈ (a, xj ),
 
X̂ = H(a, xj ) := h1 (a, xj ), . . . , hj−1 (a, xj ) , hj+1 (a, xj ), . . . , hN (a, xj ) .

The equation F (a, X) = 0 is equivalent to F̂ (a, X̂) = 0 together with the additional equation
Fm (a, X) = 0. Define
 
γ(a, xj ) := Fm a, h1 (a, xj ), . . . , hm−1 (a, xj ) , xj , hm+1 (a, xj ), . . . , hN (a, xj ) .

Then F (a, X) = 0 is equivalent to

γ(a, xj ) = 0 . (4.4)
Equation (4.4) is a scalar equation in two variables that is analysed exactly as in §3. When the
solution set is of the form a = h(xj ) with h0 (xj ) = 0 and h00 (xj ) > 0 one finds supercritical
bifurcation. The subcritical case is h00 (xj ) < 0.
The analysis of stability of the equilibria is somewhat more difficult. The matrix of the lin-
earization is FX (a, X). By hypothesis, 0 is an eigenvalue of multiplicity one at a, X. Using
perturbation theory of eigenvalues one can show that a real eigenvalue crosses the imaginary
axis as a increases through a. If the equilibrium is stable for a < a, a ≈ a one concludes instabil-
ity for a just to the right of a. The interested reader is referred to texts on Bifurcation Theory
for this perturbation theory computation.

4.3 Transcritical and pitchfork bifurcations

Continuing the reduction from the last subsection, one can consider the case where X = 0 is an
equilibrium for all a. Then one must have H(a, 0) = 0 and γ(a, 0) = 0 for all a. One can then
factor γ(a, xj ) = xj g(a, xj ) and any second branch of equilibria is defined by g(a, xj ) = 0. This
is analyzed as in the one dimensional case to reveal transcritical and pitchfork bifurcations.
Using perturbation theory of eigenvalues one can show that the N > 1 analogue of exchange of
stability in the one dimensional case is that FX has a real eigenvalue that changes sign at a. For
example a saddle can turn to a sink or a source. That is called a saddle-node bifurcation.

4.4 The Hopf-Andreev Bifurcation

There is a phenomenon that can occur for N > 1 and not for N = 1.
The dimension of the stable manifold is the number of eigenvalues in Re z < 0. This can change
by an eigenvalue passing through 0. This is analogous to the dimension 1 case and is discussed
in the preceding subsections. There is a second mechanism when N > 1. A pair of complex
conjugate eigenvalues can cross the imaginary axis while FX stays invertible. In this case no new
equilibria are born but the dimension of the stable manifold changes by two. If the eigenvalues
cross from left to right one expects a periodic orbit to be born. This is called Hopf-Andreev
Bifurcation. A clear simple example is given in HSD page 181-182. A longer important example
is given in HSD §12.4.
An interesting experiment is shown in the youtube.com video entitled "Sweet Hopf Bifurcation".
The interpretation of the video is as follows. If the upper plate were close to the lower one with

8
honey would fall in a straight stream. When the stream is stretched beyond a critical length a
bifurcation to a stable periodic in time motion occurs. It is that motion that is shown in the
youtube video.
If a complex conjugate pair crosses from right to left, a periodic orbit is expected to shrink and
disappear. It is the preceding case played in reverse.