For Any Assignment related queries, Call us at : - +1 678 648 4277

You can mail us at : - [email protected] or

reach us at : - https://www.mathhomeworksolver.com/

Problems
Problem 1 : Modified inner products for column vectors
Consider the inner product (x, y) = x∗By from class (lecture 5.5
notes), where the vectors are in CN and B is an N × N Hermitian
positive-definite matrix.
(a) Show that this inner product satisfies the required properties
of inner products from class: (x, y) = (y, x), (x, x) > 0 except
for x = 0. (Linearity (x, αy+z) = α(x, y)+(x, z) is obvious from
linearity the of matrix operations; you need not show it.)

(b) If M is an arbitrary (possibly complex) N × N matrix, define

the adjoint M† by (x, My) = (M† x, y) (for all x, y). (In this
problem, we use † instead of ∗ for the adjoint in order to
avoid confusion with the conjugate transpose: for this inner
product, the adjoint MT is not the conjugate transpose M∗ =
MT .) Give an explicit formula for M† in terms of M and B .

(c) Using your formula from above, show that M† = M (i.e., M is

self-adjoint/Hermitian for this inner product) if M = B−1A for
some A = A∗ .
Problem 2: Finite-difference approximations
For this question you may find it helpful to refer to the notes and
readings from lecture 3. Suppose that we want to compute the
operation
for some smooth function c(x) (you can assume c has a
convergent Taylor series everywhere). Now, we want to
construct a finite-difference approximation for Aˆ with u(x) on
Ω = [0, L] and Dirichlet boundary conditions u(0) = u(L) = 0,
similar to class, approximating u(mΔx) ≈ um for M equally
spaced points m = 1, 2, . . . , M, u0 = uM+1 = 0, and Δx =

(a) Using center-difference operations, construct a finite-

difference approximation for Au ˆ evaluated at mΔx. (Hint:
use a centered first-derivative evaluated at grid points m +
0.5, as in class, followed by multiplication by c, followed by
another centered first derivative. Do not ′ ′ ′′ separate Au ˆ
by the product rule into c’ u’ + cu” first, as that will make
the factorization in part (d) more difficult.)

(b) Show that your finite-difference expressions correspond to

approximating Au ˆ by Au where u is the column vector of
the M points um and A is a real-symmetric matrix of the
form A = −DTCD (give C, and show that D is the same as
the 1st-derivative matrix from lecture).
(c) In Julia, the diagm(c) command will create a diagonal matrix
from a vector c. The function diff1(M) = [ [1.0 zeros(1,M-1)];
diagm(ones(M-1),1) - eye(M) ] will allow you to create the (M +
1)×M matrix D from class via D = diff1(M) for any given value
of M. Using these two commands, construct the matrix A from
part (d) for M = 100 and L = 1 and c(x) = e3x via

L=1
M = 100
D = diff1(M)
dx = L / (M+1)
x = dx*0.5:dx:L # sequence of x values from 0.5*dx to <= L in
steps of dx
C = ....something from c(x)...hint: use diagm...
A = -D’ * C * D / dx^2

You can now get the eigenvalues and eigenvectors by λ, U =

eig(A), where λ is an array of eigenvalues and U is a matrix
whose columns are the corresponding eigenvectors (notice that
all the λ are < 0 since A is negative-definite).

using PyPlot plot

(dx:dx:L-dx, U[:,end-3:end])
xlabel("x"); ylabel("eigenfunctions")
legend(["fourth", "third", "second", "first"])
(ii) Verify that the first two eigenfunctions are indeed
orthogonal with dot(U[:,end], U[:,end-1]) in Julia, which
should be zero up to roundoff errors
(iii) Verify that you are getting second-order convergence of the
eigenvalues: compute the smallest-magnitude eigenvalue λM
[end] for M = 100, 200, 400, 800 and check that the differences
are decreasing by roughly a factor of 4 (i.e. |λ100 − λ200|
should be about 4 times larger than |λ200 − λ400|, and so on),
since doubling the resolution should multiply errors by 1/4.

(d) For c(x) = 1, we saw in class that the eigenfunctions are

sin(nπx/L). How do these compare to the eigenvectors you
plotted in the previous part? Try changing c(x) to some other
function (note: still needs to be real and > 0), and see how
different you can make the eigenfunctions from sin(nπx/L). Is
there some feature that always remains similar, no matter how
much you change c?

Problem 3: Discrete diffusion

In this problem, you will examine thermal conduction in a

system of a finite number N of pieces, and then take the N → ∞
limit to recover the heat equation. In particular:
• You have a metal bar of length L and cross-sectional area a
(hence a volume La), with a varying temperature T along the
rod. We conceptually subdivide the rod into N (touching)
pieces of length Δx = L/N.

• If Δx is small, we can approximate each piece as having a

uniform temperature Tn within the piece (n = 1, 2, . . . , N),
giving a vector T of N temperatures.

• Suppose that the rate q (in units of W) at which heat flows

across the boundary from piece n to piece n + 1 is given by q =
where κ is the metal’s thermal conductivity
(in units of W/m·K). That is, piece n loses energy at a rate q,
and piece n + 1 gains energy at the same rate, and the heat
flows faster across bigger areas, over shorter distances, or for
larger temperature differences. Note that q > 0 if Tn > Tn+1
and q < 0 if Tn < Tn+1: heat flows from the hotter piece to the
cooler piece.

• If an amount of heat ΔQ (in J) flows into a piece, its

temperature changes by ΔT = ΔQ/(cρaΔx), where c is the
specific heat capacity (in J/kg·K) and ρ is the density (kg/m3)
of the metal.

• The rod is insulated: no heat flows out the sides or through the
ends.
Given these assumptions, you should be able to answer the
following:
(a) “Newton’s law of cooling” says that that the temperature of
an object changes at a rate (K/s) proportional to the
temperature difference with its surroundings. Derive the
equivalent here: show that our assumptions above imply that
for some constant α, for 1 < n < N.
Also give the
(slightly different) equations for n = 1
and n = N.
(b) Write your equation from the previous part in matrix form:
for some matrix A.
(c) Let T (x, t) be the temperature along the rod, and suppose
Tn(t) = T ([n − 0.5]Δx, t) (the temperature at the center of the
n-th piece). Take the limit N → ∞ (with L fixed, so that
and derive a partial differential equation

What is Aˆ? (Don’t worry about the x = 0, L ends until the next part.)

(d) What are the boundary conditions on T (x, t) at x = 0 and

L? Check that if you go backwards, and form a center-
difference approximation of Aˆ with these boundary
conditions, that you recover the matrix A from above.

(e) How does your Aˆ change in the N → ∞ limit if the

conductivity is a function κ(x) of x?
(f) Suppose that instead of a thin metal bar (1d), you have an L
× L thin metal plate (2d), with a temperature T (x, y, t) and a
constant conductivity κ. If you go through the steps above
dividing it into N × N little squares of size Δx × Δy, what PDE
do you get for T in the limit N → ∞? (Many of the steps should
be similar to above.)
 Solutions
Problem 1
(a) We have (x, x) = x∗Bx > 0 for x = 0 by definition of
positive-definiteness. We have (x, y) = ∗ ∗ x∗By = (B∗x)
y = (Bx) y = y∗(Bx) = (y, x) by B = B∗.

(b) (x, My) = x∗BMy = (M†x, y) = x∗M†∗By for all x, y, and

hence we must have BM = M†∗B ∗ , or M†∗ = BMB−1
=⇒ M† = (BMB−1) = (B−1)∗M∗B∗ . Using the fact that

(c) If M = B−1A where A = A∗, then M† = B−1AB−1B =

B−1A = M. Q.E.D.
Define cm+0.5
Problem 2:
= c([m
(a) As in class, let u' ([m + 0.5]Δx) ≈ u’m+0.5 =

+ 0.5]Δx). Now we want to take the derivative of

in order to approximate Au at m by a center .5
difference:

There are other ways to solve this problem of course, that are
also second-order accurate.
(b) In order to approximate Au, we did three things: compute u'
by a center-difference as in class, multiply by cm+0.5 at each
point m + 0.5, then compute the derivative by another center-
difference. The first and last steps are exactly the same center-
difference steps as in class, so they correspond as in class to
multiplying by D and −DT , respectively, where D is the (M +
1) × M matrix

The middle step, multiplying the (M + 1)-component vector u'

by cm+0.5 at each point is just multiplication by a diagonal (M
+ 1) × (M + 1) matrix
 Putting these steps together in sequence, from right to left,
means that A = −DT CD
(c) In Julia, the diagm(c) command will create a diagonal
matrix from a vector c. The function diff1(M) = [ [1.0
zeros(1,M-1)]; diagm(ones(M-1),1) - eye(M) ] will allow you
to create the (M + 1) × M matrix D from class via D = diff1(M)
for any given value of M. Using these two commands, we
construct the matrix A from part (d) for
M = 100 3x and L = 1 and c(x) = e3x via
L=1
M = 100
D = diff1(M)
dx = L / (M+1)
x = dx*0.5:dx:L # sequence of x values from 0.5*dx to <= L in
steps of dx
c(x) = exp(3x)
C = diagm(c(x))
A = -D’ *C*D / dx^2

You can now get the eigenvalues and eigenvectors by λ, U =

eig(A), where λ is an array of eigenvalues and U is a matrix
whose columns are the corresponding eigenvectors (notice that
all the λ are < 0 since A is negative-definite).

(i) The plot is shown in Figure 1. The eigenfunctions look

vaguely “sine-like”—they have the same number of
oscillations as sin(nπx/L) for n = 1, 2, 3, 4—but are
“squeezed” to the left-hand side.

(ii) We find that the dot product is ≈ 4.3 × 10−16, which is zero
up to roundoff errors (your exact value may differ, but
should be of the same order of magnitude).

(iii) In the posted IJulia notebook for the solutions, we show a

plot of |λ2M −λM| as a function of M on a log–log scale,
and verify that it indeed decreases ∼ 1/M2.
You can also just look at the numbers instead of plotting, and
we find that this difference decreases by a factor of ≈ 3.95 from
M = 100 to M = 200 and by a factor of ≈ 3.98 from M = 200 to
M = 400, almost exactly the expected factor of 4. (For fun, in
the solutions I went to M = 1600, but you only needed to go to
M = 800.)

(d) In general, the eigenfunctions have the same number of

nodes (sign oscillations) as sin(nπx/L), but the oscillations
pushed towards the region of high c(x). This is even more
dramatic if we increase the c(x) contrast. In Figure xxx, we
show two examples. First, c(x) = e20x, in which all of the
functions are squished to the left where c is small. Second
c(x)=1 for x < 0.3 and 100 otherwise—in this case, the
oscillations are at the left 1/3 where c is small, but the function
is not zero in the right 2/3. Instead, the function is nearly
constant where c is large. The reason for this has to do with the
continuity of u: it is easy to see from the operator that ' ' )' ' cu
must be continuous for (cu to exist, and hence the slope u must
decrease by a factor of
Figure 2: First four eigenfunctions of Au = (cu for two
different choices of c(x).
100 for x > 0.3, leading to a u that is nearly constant. (We will
explore some of these issues further later in the semester.)

Problem 3:
(a) The heat capacity equation tells us that
where dQn/dt is the rate
of change of the heat in the n-th piece. The thermal
conductivity equation tells us that dQn/dt, in turn, is equal to
the sum of the rates q at which heat flows from n + 1 and n −
1 into n:

where The only difference for T1 and TN is that they

have no heat flow
n − 1 and n + 1, respectively, since the ends are insulated:
α(TN−1 − TN ).

(b) We can obtain A in two ways. First, we can simply look

directly at our equations above, Which give
corresponding rows of the matrix
Alternatively, we can write each of the above steps—
differentiating T to get the rate of heat flow q to the left at each
of the N −1 interfaces between the pieces, then taking the
difference of the q’s to get dT /dt, in matrix form, to write:

in terms of the D matrix from class (except with N reduced by

1), which gives the same A as above. As we will see in the parts
below, this is indeed a second-derivative approximation, but
with different boundary conditions—Neumann conditions—
than the Dirichlet conditions in class.
By the way, it is interesting to consider −DDT , compared to the
−DT D we had in class. Clearly, −DDT is real-symmetric and
negative semidefinite. It is not, however, negative definite, since
DT does not (and cannot) have full column rank (its rank must
be ≤ the number of rows N − 1, and in fact in class we showed
that it has rank N − 1).

(c) Ignoring the ends for the moment, for all the interior points
we have
which is exactly our familiar center-
difference
approximation for at the point n (x = [n − 0.5]Δx). Hence,

everywhere in the interior our equations converge to and

thus

(d) The boundary conditions are The easiest way to see this is to
observe that our heat flow q is really a first derivative, and zero
heat flow at the ends
Means zero derivatives. That is, qn+0.5 =
is really an approximate
derivative: while the flows q0.5 and
qN+0.5 to/from n = 0 and n = N + 1 is zero, and hence q0.5 =
qN+0.5 = 0 ≈
Working backwards, (setting 1 for convenience)
consider
with these boundary conditions and center-difference
approximations. We are given Tn = T([n − 0.5]Δx,t) for n =
1,...,N. First, we compute
for n = 1,...,N − 1 (−DT T using the D above). Unlike the
Dirichlet case in class, we don’t com­

which are zero by the boundary conditions. Then, we

compute our approximate 2nd for
derivatives
n = 1,...,N, where we let

using the D from above). This gives

at the endpoints, and

for 1 <n<N,

which are precisely the rows of our A matrix above.

(e) If κ(x), then we get a different κ and α factor for each

Tn+1 − Tn difference:
Where

we differentiated,
In the N → ∞ limit, this gives
multiplied

by κ, differentiated again, and then divided by cρ. (You weren’t

asked to handle the case where cρ is not a constant, so it’s okay
if you commuted cρ with the derivatives.)

(f) If we discretize to Tm,n = T(mΔx, nΔy), the steps are

basically the same except that we have to consider the heat flow
in both the x and y directions, and hence we have to take
differences in both x and y. In particular, suppose the thickness
of the block is h. In this case, heat will flow from Tm,n to
Tm+1,n at a rate

where hΔy is the area of the interface between the two blocks.
Then, to convert into a rate of temperature change, we will
divide by cρhΔxΔy, where hΔxΔy is the volume of the block.
Putting this all together, we obtain:
where the thing in [··· ] is precisely the five-point stencil
approximation for ∇2 from class. Hence, we obtain

where for fun I have put the κ in the middle, which is the right
place if κ is not a constant (you were not required to do this).