doceamus . . .

let us teach

The Core Ideas in Our

Teaching
Gilbert Strang

What will our students remember? One answer 1. Its nullspace N(A) (the kernel) dimension n−r
2. Its column space C(A) (the range) r
comes quickly but it is a counsel of despair: nothing
3. Its row space, which is C(AT ) r
at all. At the other extreme is an impossible hope 4. The nullspace N(AT ) of the transpose m−r
that we all cherish: everything we say. Let me
look for an intermediate answer, closer to reality, These are the spaces that we want students to
possibly by changing the question. remember. I draw them as often as possible (two
in R n and two in R m ). I count their basis vectors to
I have come to believe that each course has a
find their dimension: the first big theorems in linear
central core. We may not see it ourselves, when we
algebra. The rank r determines all dimensions. I
teach a new topic every day. For the calculus course,
propose multiple choices of A—the beauty of this
I won’t even venture an answer—at least not here.
subject is in the wonderful variety of matrices. And
My examples will be differential equations and
I connect the four subspaces to factorizations of
linear algebra, because writing a textbook forced
A, which are really choices of bases that lie at the
me to uncover (painfully slowly!) the underlying
absolute center of pure and applied linear algebra.
structure of the course.
The bases in U and Q and S and V become
May I begin with linear algebra. The ideas increasingly perfect.
of a vector space and a basis for that space
A = LU Elimination gives an echelon basis for the row space
are central. It is a serious job to help students
A = QR Gram-Schmidt gives an orthonormal basis for C(A)
understand these words. The building blocks are A = SΛS−1 Eigenvectors give a basis in which A is diagonal
“linear combinations” and “linear independence.” A = UΣVT Orthonormal bases in the columns of U and V .
We certainly need good examples, and good bases
We are constantly constructing bases for the
for them. I think it is here that the course becomes
fundamental subspaces. Elimination and Gram-
coherent—or it can scatter into unconnected
Schmidt orthogonalization end after finitely many
examples of isolated ideas.
steps. Diagonalization by eigenvectors is deeper
I will start with a matrix A. A more abstract
and better, but A must be square and nondefec-
person would start from a linear transformation. tive. The Singular Value Decomposition produces
But we are aiming for a basis; we are choosing perfect bases vi and ui for all four subspaces—
coordinates; they bring us to a matrix. There are orthonormal and also diagonalizing for every
four fundamental subspaces associated with that matrix A:
matrix:
Avi = σi ui (i ≤ r ) Avi = 0 and AT ui = 0 (i > r )
Gilbert Strang is professor of mathematics at M.I.T. His The success of the SVD comes from the spectral
email address is [email protected] theorem for symmetric matrices: AT A has a full
Members of the Editorial Board for Doceamus are: David set of orthonormal eigenvectors vi . Beautifully, the
Bressoud, Roger Howe, Karen King, William McCallum, and ui turn out to be orthonormal eigenvectors of AAT .
Mark Saul. This can be a highlight for the last days of a linear
DOI: http://dx.doi.org/10.1090/noti1174 algebra course.

November 2014 Notices of the AMS 1243

 For an earlier day, one idea is to ask students to 1. Combine exponentials est with weights F (s)
“read” a few matrices: to get f (t). By linearity, the solution y(t) will
combine the exponentials F(s)G(s)est .
" # " # " #
cos θ − sin θ 1 0 −1 1 0
sin θ cos θ 0 0 0 −1 1 2. Combine impulses δ(t − s) with weights f (s)
The rotation is familiar, the projection is almost to get f (t). By linearity, the solution y(t) will
too easy. The difference matrix is also the incidence combine the impulse responses f (s)g(t − s).
matrix for a simple graph (three nodes in a line). Where est is localized at frequency s, the delta
Incidence matrices of a larger graph are terrific function δ(t − s) is completely localized at time s.
examples—all four subspaces have a meaning. Method 1 uses the Laplace transform. The
May I turn from subspaces to the basic course transform of f (t) gives the right weights F (s):
on differential equations. Part of this course is a
collection of methods to solve separable equations, F(s) = transform of f (t)
exact equations, logistic equations y 0 = ay − by 2 , y(t) = inverse transform of F(s)G(s).
and more. We go forward to systems of equations, The transform F(s) might be easy. The hard part
and test nonlinear equations for stability. But the
is the inverse Laplace transform, to combine the
coherent part (the central problem) is to solve
solutions F(s)G(s)est into y(t).
linear equations with constant coefficients. How
Realistically, we know a very limited number
can we present their solutions?
of transform pairs. Method 1 almost limits us
I believe we have to answer this question. It is the
to the same short list as before: f can combine
ODE equivalent of solving Ax = 0 and Ax = b and
e(a+iω)t , cos ωt, sin ωt, t, and their products. This
Ax = λx. It certainly rests on the most important
is a space of functions whose derivatives stay in
functions in this course: exponentials est and eλt .
the space. You can guess that I am advocating
By working with exponentials, we (almost) turn the
Method 2, which begins with an impulse δ(t):
differential equation into algebra.
Start with the simplest right-hand sides f (t) = 0
(1)
and est .
Ag 00 +Bg 0 +Cg = δ(t) with g(0) = 0 and g 0 (0) = 0.
Ay 00 + By 0 + Cy = 0 Ay 00 + By 0 + Cy = est
Introducing that delta function is a good thing!
The key idea is to expect solutions y = Gest : We are finding the fundamental solution g(t)—the
G(As 2 +Bs +C)est = 0 G(As 2 +Bs +C)est = est . Green’s function, the growth factor, the impulse
On the left, two values of s are allowed: the roots response. This is a high point in the course. And
s1 and s2 of As 2 + Bs + C = 0. On the right, any s it is easy to do, because this same g(t) also solves
is allowed (and the possibilities s1 = s2 and s = s1 the homogeneous equation:
and s = s1 = s2 need special attention). Normally (2)
we have Ag 00 +Bg 0 +Cg = 0 with g(0) = 0 and g 0 (0) = 1/A.
The solution must have the form g(t) = c1 es1 t +
yn = ynullspace = c1 es1 t + c2 es2 t
c2 es2 t . The two initial conditions give c1 and c2 and
1
yp = yparticular = G(s)est = est . a neat formula for g(t):
As 2 + Bs + C
Those two parts of y(t) connect linear differential (3) !
equations to linear algebra. The complete solution es1 t − es2 t tes1 t
g(t) = or g(t) = when s1 = s2 .
combines all yn with one yp . Linearity is in control A(s1 − s2 ) A
and the consequence is y = yn + yp . Then the original equation, with any right side
I apologize for asking you to read what you f (t), is solved by
know so well. The simplicity of y = Gest has to Zt
be recognized and remembered. This is where (4) y particular (t) = g(t − s) f (s) ds.
calculus meets algebra. G is the prime example of 0

an undetermined coefficient (determined by the Discussion. In coming quickly to the formula

equation). An elementary course could continue as for y(t), I have left multiple loose ends. Let me
far as f (t) = eiωt and cos ωt and sin ωt and stop. go backwards more slowly, as we would certainly
The serious question is to solve the differential do in a classroom. Methods 1 and 2 are closely
equation for all f (t). connected. The Laplace transform of δ(t) is 1.
I see two instructive ways to reach y(t). Both Then equation (1) transforms to
begin with special right-hand sides, and combine
(5) (As 2 + Bs + C) G(s) = 1.
the solutions. The combination has to be an integral
and not just a finite sum: calculus is needed now. The transfer function G(s) = 1/(As 2 + Bs + C) is
Here are the good options: the Laplace transform of the impulse response g(t).

1244 Notices of the AMS Volume 61, Number 10

These functions can be written in terms of A, B, C yn + yp for any right-hand side f (t) and initial
or s1 and s2 . A lot of effort has gone into√choosing condition y(0) is
good parameters! The damping ratio B/ 4AC and Zt
(6) y(t) = y(0)eat + ea(t−s) f (s) ds.
p
the natural frequency C/A are two of the best.
0
We must also explain why equations (1) and
The input f (s) at time s grows in the remaining time
(2) have the same solution g(t). Mechanically, this
t − s by the factor ea(t−s) . The solution y(t) (the
comes from partial fractions:
integral) combines all of these outputs ea(t−s) f (s).
1 1 That single paragraph translates into weeks of
=
As 2 + Bs + C A(s − s1 )(s − s2 ) teaching, even without δ(t). Perhaps first order
1 1 1 equations with constant coefficients might be the
 
= − .
A(s1 − s2 ) s − s1 s − s2 one topic that is understood and remembered?
I don’t like to think so, because a teacher has
The inverse Laplace transform confirms that es1 t
to remain an optimist.
and es2 t go into g(t).
I plan to prepare video lectures going at a normal
Here is a truly “mechanical” explanation of
pace, and linked to http://math.mit.edu/dela.
(1) = (2). A bat hits a ball at t = 0. The velocity
That website has much more about differential
jumps instantly to g 0 (0) = 1/A. This comes from
equations and linear algebra and a new textbook
integrating Ag 00 + Bg 0 + Cg = δ(t) from t = 0 to
for those courses.
t = h. The left side produces the jump in Ag 0 and
the integral of δ(t) is 1. The other terms disappear
as h → 0, leaving Ag 0 (0) = 1.
In working with δ(t), some faith is needed. It
is worth developing and it is not misplaced. A
delta function is an extremely useful model. So
is its integral the step function, which turns on a
switch at t = 0. By linearity, the step response is
the integral of g(t).
Finally, let me connect Method 1 directly to
Method 2. In the first method, the Laplace transform
of y(t) is F(s) G(s). In the second method, y(t) is
the convolution of f (t) with g(t). The connection
is the Convolution Rule: The transform of a
convolution f (t)∗g(t) is a multiplication F (s) G(s).
In the language of signal processing, any con-
stant coefficient linear equation can be solved in
the “s-domain” or the “t-domain.” The poles s1 , s2
of the transfer function G(s) = 1/(As 2 + Bs + C)
control the behavior of y(t): oscillation, decay, or
instability. The whole course develops out of the
quadratic formula for those roots s1 and s2 .
Note. The actual course would start with first
order equations:
y 0 − ay = 0 y 0 − ay = est
The null solutions are yn = ceat . The particular
solution is yp = est /(s − a). The transfer function
is G(s) = 1/(s − a). The fundamental solution
(impulse response, growth factor, Green’s function)
solves
g 0 − ag = δ(t) with g(0) = 0
g 0 − ag = 0 with g(0) = 1
This function is simply g = eat . At this early
point it doesn’t need all those names! We recognize
it as 1/(integrating factor). Its Laplace transform
is G(s) = 1/(s − a). For systems y 0 = Ay, we
have the matrix exponential g = eAt . The solution

November 2014 Notices of the AMS 1245