## Motivation.

The notation I prefer for relativistic geometric algebra uses Hestenes’ space time algebra (STA) [2], where the basis is a four dimensional space $$\setlr{ \gamma_\mu }$$, subject to Dirac matrix like relations $$\gamma_\mu \cdot \gamma_\nu = \eta_{\mu \nu}$$.

In this formalism a four vector is just the sum of the products of coordinates and basis vectors, for example, using summation convention

\label{eqn:boostToParavector:160}
x = x^\mu \gamma_\mu.

The invariant for a four-vector in STA is just the square of that vector

\label{eqn:boostToParavector:180}
\begin{aligned}
x^2
&= (x^\mu \gamma_\mu) \cdot (x^\nu \gamma_\nu) \\
&= \sum_\mu (x^\mu)^2 (\gamma_\mu)^2 \\
&= (x^0)^2 – \sum_{k = 1}^3 (x^k)^2 \\
&= (ct)^2 – \Bx^2.
\end{aligned}

Recall that a four-vector is time-like if this squared-length is positive, spacelike if negative, and light-like when zero.

Time-like projections are possible by dotting with the “lab-frame” time like basis vector $$\gamma_0$$

\label{eqn:boostToParavector:200}
ct = x \cdot \gamma_0 = x^0,

and space-like projections are wedges with the same

\label{eqn:boostToParavector:220}
\Bx = x \cdot \gamma_0 = x^k \sigma_k,

where sums over Latin indexes $$k \in \setlr{1,2,3}$$ are implied, and where the elements $$\sigma_k$$

\label{eqn:boostToParavector:80}
\sigma_k = \gamma_k \gamma_0.

which are bivectors in STA, can be viewed as an Euclidean vector basis $$\setlr{ \sigma_k }$$.

Rotations in STA involve exponentials of space like bivectors $$\theta = a_{ij} \gamma_i \wedge \gamma_j$$

\label{eqn:boostToParavector:240}
x’ = e^{ \theta/2 } x e^{ -\theta/2 }.

Boosts, on the other hand, have exactly the same form, but the exponentials are with respect to space-time bivectors arguments, such as $$\theta = a \wedge \gamma_0$$, where $$a$$ is any four-vector.

Observe that both boosts and rotations necessarily conserve the space-time length of a four vector (or any multivector with a scalar square).

\label{eqn:boostToParavector:260}
\begin{aligned}
\lr{x’}^2
&=
\lr{ e^{ \theta/2 } x e^{ -\theta/2 } } \lr{ e^{ \theta/2 } x e^{ -\theta/2 } } \\
&=
e^{ \theta/2 } x \lr{ e^{ -\theta/2 } e^{ \theta/2 } } x e^{ -\theta/2 } \\
&=
e^{ \theta/2 } x^2 e^{ -\theta/2 } \\
&=
x^2 e^{ \theta/2 } e^{ -\theta/2 } \\
&=
x^2.
\end{aligned}

## Paravectors.

Paravectors, as used by Baylis [1], represent four-vectors using a Euclidean multivector basis $$\setlr{ \Be_\mu }$$, where $$\Be_0 = 1$$. The conversion between STA and paravector notation requires only multiplication with the timelike basis vector for the lab frame $$\gamma_0$$

\label{eqn:boostToParavector:40}
\begin{aligned}
X
&= x \gamma_0 \\
&= \lr{ x^0 \gamma_0 + x^k \gamma_k } \gamma_0 \\
&= x^0 + x^k \gamma_k \gamma_0 \\
&= x^0 + \Bx \\
&= c t + \Bx,
\end{aligned}

We need a different structure for the invariant length in paravector form. That invariant length is
\label{eqn:boostToParavector:280}
\begin{aligned}
x^2
&=
\lr{ \lr{ ct + \Bx } \gamma_0 }
\lr{ \lr{ ct + \Bx } \gamma_0 } \\
&=
\lr{ \lr{ ct + \Bx } \gamma_0 }
\lr{ \gamma_0 \lr{ ct – \Bx } } \\
&=
\lr{ ct + \Bx }
\lr{ ct – \Bx }.
\end{aligned}

Baylis introduces an involution operator $$\overline{{M}}$$ which toggles the sign of any vector or bivector grades of a multivector. For example, if $$M = a + \Ba + I \Bb + I c$$, where $$a,c \in \mathbb{R}$$ and $$\Ba, \Bb \in \mathbb{R}^3$$ is a multivector with all grades $$0,1,2,3$$, then the involution of $$M$$ is

\label{eqn:boostToParavector:300}
\overline{{M}} = a – \Ba – I \Bb + I c.

Utilizing this operator, the invariant length for a paravector $$X$$ is $$X \overline{{X}}$$.

Let’s consider how boosts and rotations can be expressed in the paravector form. The half angle operator for a boost along the spacelike $$\Bv = v \vcap$$ direction has the form

\label{eqn:boostToParavector:120}
L = e^{ -\vcap \phi/2 },

\label{eqn:boostToParavector:140}
\begin{aligned}
X’
&=
c t’ + \Bx’ \\
&=
x’ \gamma_0 \\
&=
L x L^\dagger \\
&=
e^{ -\vcap \phi/2 } x^\mu \gamma_\mu
e^{ \vcap \phi/2 } \gamma_0 \\
&=
e^{ -\vcap \phi/2 } x^\mu \gamma_\mu \gamma_0
e^{ -\vcap \phi/2 } \\
&=
e^{ -\vcap \phi/2 } \lr{ x^0 + \Bx } e^{ -\vcap \phi/2 } \\
&=
L X L.
\end{aligned}

Because the involution operator toggles the sign of vector grades, it is easy to see that the required invariance is maintained

\label{eqn:boostToParavector:320}
\begin{aligned}
X’ \overline{{X’}}
&=
L X L
\overline{{ L X L }} \\
&=
L X L
\overline{{ L }} \overline{{ X }} \overline{{ L }} \\
&=
L X \overline{{ X }} \overline{{ L }} \\
&=
X \overline{{ X }} L \overline{{ L }} \\
&=
X \overline{{ X }}.
\end{aligned}

Let’s explicitly expand the transformation of \ref{eqn:boostToParavector:140}, so we can relate the rapidity angle $$\phi$$ to the magnitude of the velocity. This is most easily done by splitting the spacelike component $$\Bx$$ of the four vector into its projective and rejective components

\label{eqn:boostToParavector:340}
\begin{aligned}
\Bx
&= \vcap \vcap \Bx \\
&= \vcap \lr{ \vcap \cdot \Bx + \vcap \wedge \Bx } \\
&= \vcap \lr{ \vcap \cdot \Bx } + \vcap \lr{ \vcap \wedge \Bx } \\
&= \Bx_\parallel + \Bx_\perp.
\end{aligned}

The exponential

\label{eqn:boostToParavector:360}
e^{-\vcap \phi/2}
=
\cosh\lr{ \phi/2 }
– \vcap \sinh\lr{ \phi/2 },

commutes with any scalar grades and with $$\Bx_\parallel$$, but anticommutes with $$\Bx_\perp$$, so

\label{eqn:boostToParavector:380}
\begin{aligned}
X’
&=
\lr{ c t + \Bx_\parallel } e^{ -\vcap \phi/2 } e^{ -\vcap \phi/2 }
+
\Bx_\perp e^{ \vcap \phi/2 } e^{ -\vcap \phi/2 } \\
&=
\lr{ c t + \Bx_\parallel } e^{ -\vcap \phi }
+
\Bx_\perp \\
&=
\lr{ c t + \vcap \lr{ \vcap \cdot \Bx } } \lr{ \cosh \phi – \vcap \sinh \phi }
+
\Bx_\perp \\
&=
\Bx_\perp
+
\lr{ c t \cosh\phi – \lr{ \vcap \cdot \Bx} \sinh \phi }
+
\vcap \lr{ \lr{ \vcap \cdot \Bx } \cosh\phi – c t \sinh \phi } \\
&=
\Bx_\perp
+
\cosh\phi \lr{ c t – \lr{ \vcap \cdot \Bx} \tanh \phi }
+
\vcap \cosh\phi \lr{ \vcap \cdot \Bx – c t \tanh \phi }.
\end{aligned}

Employing the argument from [3],
we want $$\phi$$ defined so that this has structure of a Galilean transformation in the limit where $$\phi \rightarrow 0$$. This means we equate

\label{eqn:boostToParavector:400}
\tanh \phi = \frac{v}{c},

so that for small $$\phi$$

\label{eqn:boostToParavector:420}
\Bx’ = \Bx – \Bv t.

We can solving for $$\sinh^2 \phi$$ and $$\cosh^2 \phi$$ in terms of $$v/c$$ using

\label{eqn:boostToParavector:440}
\tanh^2 \phi
= \frac{v^2}{c^2}
=
\frac{ \sinh^2 \phi }{1 + \sinh^2 \phi}
=
\frac{ \cosh^2 \phi – 1 }{\cosh^2 \phi}.

which after picking the positive root required for Galilean equivalence gives
\label{eqn:boostToParavector:460}
\begin{aligned}
\cosh \phi &= \frac{1}{\sqrt{1 – (\Bv/c)^2}} \equiv \gamma \\
\sinh \phi &= \frac{v/c}{\sqrt{1 – (\Bv/c)^2}} = \gamma v/c.
\end{aligned}

The Lorentz boost, written out in full is

\label{eqn:boostToParavector:480}
ct’ + \Bx’
=
\Bx_\perp
+
\gamma \lr{ c t – \frac{\Bv}{c} \cdot \Bx }
+
\gamma \lr{ \vcap \lr{ \vcap \cdot \Bx } – \Bv t }
.

Authors like Chappelle, et al., that also use paravectors [4], specify the form of the Lorentz transformation for the electromagnetic field, but for that transformation reversion is used instead of involution.
I plan to explore that in a later post, starting from the STA formalism that I already understand, and see if I can make sense
of the underlying rationale.

# References

[1] William Baylis. Electrodynamics: a modern geometric approach, volume 17. Springer Science \& Business Media, 2004.

[2] C. Doran and A.N. Lasenby. Geometric algebra for physicists. Cambridge University Press New York, Cambridge, UK, 1st edition, 2003.

[3] L. Landau and E. Lifshitz. The Classical theory of fields. Addison-Wesley, 1951.

[4] James M Chappell, Samuel P Drake, Cameron L Seidel, Lachlan J Gunn, and Derek Abbott. Geometric algebra for electrical and electronic engineers. Proceedings of the IEEE, 102 0(9), 2014.

## Electric field of a spherical shell. Ka-Tex rendered

This is a test of KaTex, the latex rendering engine used for Khan academy. They advertise themselves as much faster than mathjax, but it looks like the reason for that is because they generate images that look crappy unless the browser resolution is matched to the images just right.

Here’s a rerendering of an old post, with the latex rendered with WP-KaTeX instead of MathJax-LaTeX.

# The post

### Problem:

Calculate the field due to a spherical shell. The field is

$\mathbf{E} = \frac{\sigma}{4 \pi \epsilon_0} \int \frac{(\mathbf{r} - \mathbf{r}')}{{{\left\lvert{{\mathbf{r} - \mathbf{r}'}}\right\rvert}}^3} da',$

where $\mathbf{r}'$ is the position to the area element on the shell. For the test position, let $\mathbf{r} = z \mathbf{e}_3$.

### Solution:

We need to parameterize the area integral. A complex-number like geometric algebra representation works nicely.

\begin{aligned}\mathbf{r}' &= R \left( \sin\theta \cos\phi, \sin\theta \sin\phi, \cos\theta \right) \\ &= R \left( \mathbf{e}_1 \sin\theta \left( \cos\phi + \mathbf{e}_1 \mathbf{e}_2 \sin\phi \right) + \mathbf{e}_3 \cos\theta \right) \\ &= R \left( \mathbf{e}_1 \sin\theta e^{i\phi} + \mathbf{e}_3 \cos\theta \right).\end{aligned}

Here $i = \mathbf{e}_1 \mathbf{e}_2$ has been used to represent to horizontal rotation plane.

The difference in position between the test vector and area-element is

$\mathbf{r} - \mathbf{r}' = \mathbf{e}_3 {\left({ z - R \cos\theta }\right)} - R \mathbf{e}_1 \sin\theta e^{i \phi},$

with an absolute squared length of

\begin{aligned}{{\left\lvert{{\mathbf{r} - \mathbf{r}' }}\right\rvert}}^2 &= {\left({ z - R \cos\theta }\right)}^2 + R^2 \sin^2\theta \\ &= z^2 + R^2 - 2 z R \cos\theta.\end{aligned}

As a side note, this is a kind of fun way to prove the old “cosine-law” identity. With that done, the field integral can now be expressed explicitly

\begin{aligned} \mathbf{E} &= \frac{\sigma}{4 \pi \epsilon_0} \int_{\phi = 0}^{2\pi} \int_{\theta = 0}^\pi R^2 \sin\theta d\theta d\phi \frac{\mathbf{e}_3 {\left({ z - R \cos\theta }\right)} - R \mathbf{e}_1 \sin\theta e^{i \phi}} { {\left({z^2 + R^2 - 2 z R \cos\theta}\right)}^{3/2} } \\ &= \frac{2 \pi R^2 \sigma \mathbf{e}_3}{4 \pi \epsilon_0} \int_{\theta = 0}^\pi \sin\theta d\theta \frac{z - R \cos\theta} { {\left({z^2 + R^2 - 2 z R \cos\theta}\right)}^{3/2} } \\ &= \frac{2 \pi R^2 \sigma \mathbf{e}_3}{4 \pi \epsilon_0} \int_{\theta = 0}^\pi \sin\theta d\theta \frac{ R( z/R - \cos\theta) } { (R^2)^{3/2} {\left({ (z/R)^2 + 1 - 2 (z/R) \cos\theta}\right)}^{3/2} } \\ &= \frac{\sigma \mathbf{e}_3}{2 \epsilon_0} \int_{u = -1}^{1} du \frac{ z/R - u} { {\left({1 + (z/R)^2 - 2 (z/R) u}\right)}^{3/2} }. \end{aligned}

Observe that all the azimuthal contributions get killed. We expect that due to the symmetry of the problem. We are left with an integral that submits to Mathematica, but doesn’t look fun to attempt manually. Specifically

$\int_{-1}^1 \frac{a-u}{{\left({1 + a^2 - 2 a u}\right)}^{3/2}} du = \frac{2}{a^2},$

if $a > 1$, and zero otherwise, so

$\boxed{ \mathbf{E} = \frac{\sigma (R/z)^2 \mathbf{e}_3}{\epsilon_0} }$

for $z > R$, and zero otherwise.

In the problem, it is pointed out to be careful of the sign when evaluating $\sqrt{ R^2 + z^2 - 2 R z }$, however, I don’t see where that is even useful?

# KaTex commentary

1. Conditional patterns, such as:
\left\{
\begin{array}{l l}
\frac{\sigma (R/z)^2 \mathbf{e}_3}{\epsilon_0}
& \quad \mbox{if $$z > R$$ } \\
0 & \quad \mbox{if $$z < R$$ }
\end{array}
\right.


messed up KaTex, resulting in render errors like:

2. The latex has to be all in one line, or else KaTex renders the newlines explicitly. Example:
Having to condense all my latex onto a single line is one of the reasons I switched from the default wordpress latex engine to mathjax. It was annoying enough that I started paying for my wordpress hosting, and stopped posting on my old free peeterjoot.wordpress.com blog. Using KaTex and having to go back to single line latex would suck!
3. The rendering looks like crap, unless you match your resolution to exactly those used to create the images. The mathjax rendering may be slower, but looks much better!
4. The Mathjax-Latex wordpress plugin has some support for equation labeling and references. I don’t see a way to do those with the WP-KaTex plugin.
5. I can have a large set of macros installed in my default.js matching a subset of what I have in my .sty files. I don’t see a way to do that with the WP-KaTex plugin, but perhaps there is just no documented mechanism. KaTex itself does have a macro mechanism.
6. Left justified display mode is hard to read. The mathjax rendered centered display mode looks much better.

## EDIT.

I’m not sure I was getting the katex plugin when I used the [ latex ] … [ /latex ] tags.  I see some comments that indicate that there is built in handling of these tags in the Jetpack plugin.  If I change frontend.php in the katex plugin to use [ katex ] … [ /katex ] tags instead, then I see much different results.

## second experiment in screen recording

Here’s a second attempt at recording a blackboard style screen recording:

To handle the screen transitions, equivalent to clearing my small blackboard, I switched to using a black background and just moved the text as I filled things up.  This worked much better.  I still drew with mischief, and recorded with OBS, but then did a small post production edit in iMovie to remove a little bit of dead air and to edit out one particularly bad flub.

This talk covers the product of two vectors, defines the dot and wedge products, and shows how the 3D wedge product is related to the cross product.  I recorded some additional discussion of duality that I left out of this video, which was long enough without it.

## experiment in screen recording: An introduction to Geometric Algebra.

I’ve been curious for a while what it would take to create lesson style screen recordings, and finally got around to trying one myself:

This is an introduction to Geometric (Clifford) algebra.  I briefly outline a geometrical interpretation of various products of unit vectors, rules for reducing products of unit vectors, and the axioms that justify those rules.

I made this recording using the OBS screen recorder, using a Mac, drawing with a Wacom tablet and using Mischief as my drawing application.  I have to find a way to do the screen clearing transitions more smoothly, as there are sizable dead time delays while I do the ‘File -> Import -> Recent -> …  ; Don’t save’ sequence in mischief to reload.  I also um and ah more than I like, something I think I could avoid if presenting this to a real live person.

## Notes compilation for ECE1505, Convex Optimization

March 18, 2017 ece1505 No comments

I’ve now posted a notes compilation for the subset of the Convex Optimization (ECE1505H) course I was taking in the winter 2017 session.

This course was taught by Prof. S. Draper.

These convex optimization notes are incomplet, covering only the first 9 lectures. The unredacted notes include my solution to problem set 1 (149 pages, vs. 131 pages).

I initially enrolled on this optimization course because I needed a specific quota of ECE courses to satisfy the M.Eng graduation requirements, and the electromagnetics group wasn’t offering enough courses.  I remembered liking linear programming in high school, and always wanted to understand the rational for some of the assumptions that was based on that were never proven in class.  Specifically, I recall that it was stated, but not proved in that high school class, that the extreme values were always found at the vertices of the optimization region.  So, my thought was, I’ll have fun learning the basis for those assumptions, and also learn about optimization theory in general.

It turns out that optimization theory, at least as presented in this course, is very very dry.  It was an endless seeming sequence of definition and proof, with the end goal so far away that it was very difficult to see the big picture.  I worked through the a number of weeks of this particular course before I had enough and bailed.  Work is too fun right now to torture myself and spend the time on an academic course that I am not enjoying, so I dropped it and am back to full time work at LzLabs (from 80%) until the next session at UofT starts again.

The reason I enrolled on the M.Eng in the first place was to study material that I was interested in.  Ideally I would have done that in a part time physics grad context, but that was not available, so I found that the M.Eng allowed me to take an interesting (but constrained) mix of physics and engineering electromagnetism courses.  However, when I enrolled, the electromagnetism course selection was a lot better, and now unfortunately it is sparse and includes only courses that I’d already taken.  I don’t want the M.Eng degree paper badly enough to torture myself with a course that I’m not actually interested in.

I now actually have a plan to satisfy both the degree requirements and my interests (using a project “course”).  That will involve independent study on Geometric Algebra applications to engineering electromagnetism.  I am irked that I have to pay a part time engineering program fee next year to self study, but it does seem worthwhile to come out of the M.Eng study with an actual degree as a side effect, so I am going to go ahead and do it anyways.

## Disclaimer

Peeter’s lecture notes from class. These may be incoherent and rough.

These are notes for the UofT course ECE1505H, Convex Optimization, taught by Prof. Stark Draper, from [1].

## Today

• Finish local vs global.
• Compositions of functions.
• Introduction to convex optimization problems.

## Continuing proof:

We want to prove that if

\begin{equation*}
\begin{aligned}
\end{aligned},
\end{equation*}

then $$\Bx^\conj$$ is a local optimum.

Proof:

Again, using Taylor approximation

\label{eqn:convexOptimizationLecture8:20}
F(\Bx^\conj + \Bv) = F(\Bx^\conj) + \lr{ \spacegrad F(\Bx^\conj)}^\T \Bv + \inv{2} \Bv^\T \spacegrad^2 F(\Bx^\conj) \Bv + o(\Norm{\Bv}^2)

The linear term is zero by assumption, whereas the Hessian term is given as $$> 0$$. Any direction that you move in, if your move is small enough, this is going uphill at a local optimum.

## Summarize:

For twice continuously differentiable functions, at a local optimum $$\Bx^\conj$$, then

\label{eqn:convexOptimizationLecture8:40}
\begin{aligned}
\end{aligned}

If, in addition, $$F$$ is convex, then $$\spacegrad F(\Bx^\conj) = 0$$ implies that $$\Bx^\conj$$ is a global optimum. i.e. for (unconstrained) convex functions, local and global optimums are equivalent.

• It is possible that a convex function does not have a global optimum. Examples are $$F(x) = e^x$$
(fig. 1)
, which has an $$\inf$$, but no lowest point.

fig. 1. Exponential has no global optimum.

• Our discussion has been for unconstrained functions. For constrained problems (next topic) is not not necessarily true that $$\spacegrad F(\Bx) = 0$$ implies that $$\Bx$$ is a global optimum, even for $$F$$ convex.

As an example of a constrained problem consider

\label{eqn:convexOptimizationLecture8:n}
\begin{aligned}
\min &2 x^2 + y^2 \\
x &\ge 3 \\
y &\ge 5.
\end{aligned}

The level sets of this objective function are plotted in fig. 2. The optimal point is at $$\Bx^\conj = (3,5)$$, where $$\spacegrad F \ne 0$$.

fig. 2. Constrained problem with optimum not at the zero gradient point.

## Projection

Given $$\Bx \in \mathbb{R}^n, \By \in \mathbb{R}^p$$, if $$h(\Bx,\By)$$ is convex in $$\Bx, \By$$, then

\label{eqn:convexOptimizationLecture8:60}
F(\Bx_0) = \inf_\By h(\Bx_0,\By)

is convex in $$\Bx$$, as sketched in fig. 3.

fig. 3. Epigraph of $$h$$ is a filled bowl.

The intuition here is that shining light on the (filled) “bowl”. That is, the image of $$\textrm{epi} h$$ on the $$\By = 0$$ screen which we will show is a convex set.

Proof:

Since $$h$$ is convex in $$\begin{bmatrix} \Bx \\ \By \end{bmatrix} \in \textrm{dom} h$$, then

\label{eqn:convexOptimizationLecture8:80}
\textrm{epi} h = \setlr{ (\Bx,\By,t) | t \ge h(\Bx,\By), \begin{bmatrix} \Bx \\ \By \end{bmatrix} \in \textrm{dom} h },

is a convex set.

We also have to show that the domain of $$F$$ is a convex set. To show this note that

\label{eqn:convexOptimizationLecture8:100}
\begin{aligned}
\textrm{dom} F
&= \setlr{ \Bx | \exists \By s.t. \begin{bmatrix} \Bx \\ \By \end{bmatrix} \in \textrm{dom} h } \\
&= \setlr{
\begin{bmatrix}
I_{n\times n} & 0_{n \times p}
\end{bmatrix}
\begin{bmatrix}
\Bx \\
\By
\end{bmatrix}
| \begin{bmatrix} \Bx \\ \By \end{bmatrix} \in \textrm{dom} h
}.
\end{aligned}

This is an affine map of a convex set. Therefore $$\textrm{dom} F$$ is a convex set.

\label{eqn:convexOptimizationLecture8:120}
\begin{aligned}
\textrm{epi} F
&=
\setlr{ \begin{bmatrix} \Bx \\ \By \end{bmatrix} | t \ge \inf h(\Bx,\By), \Bx \in \textrm{dom} F, \By: \begin{bmatrix} \Bx \\ \By \end{bmatrix} \in \textrm{dom} h } \\
&=
\setlr{
\begin{bmatrix}
I & 0 & 0 \\
0 & 0 & 1
\end{bmatrix}
\begin{bmatrix}
\Bx \\
\By \\
t
\end{bmatrix}
|
t \ge h(\Bx,\By), \begin{bmatrix} \Bx \\ \By \end{bmatrix} \in \textrm{dom} h
}.
\end{aligned}

### Example:

The function

\label{eqn:convexOptimizationLecture8:140}
F(\Bx) = \inf_{\By \in C} \Norm{ \Bx – \By },

over $$\Bx \in \mathbb{R}^n, \By \in C$$, ,is convex if $$C$$ is a convex set. Reason:

• $$\Bx – \By$$ is linear in $$(\Bx, \By)$$.
• $$\Norm{ \Bx – \By }$$ is a convex function if the domain is a convex set
• The domain is $$\mathbb{R}^n \times C$$. This will be a convex set if $$C$$ is.
• $$h(\Bx, \By) = \Norm{\Bx -\By}$$ is a convex function if $$\textrm{dom} h$$ is a convex set. By setting $$\textrm{dom} h = \mathbb{R}^n \times C$$, if $$C$$ is convex, $$\textrm{dom} h$$ is a convex set.
• $$F()$$

## Composition of functions

Consider

\label{eqn:convexOptimizationLecture8:160}
\begin{aligned}
F(\Bx) &= h(g(\Bx)) \\
\textrm{dom} F &= \setlr{ \Bx \in \textrm{dom} g | g(\Bx) \in \textrm{dom} h } \\
F &: \mathbb{R}^n \rightarrow \mathbb{R} \\
g &: \mathbb{R}^n \rightarrow \mathbb{R} \\
h &: \mathbb{R} \rightarrow \mathbb{R}.
\end{aligned}

Cases:

1. $$g$$ is convex, $$h$$ is convex and non-decreasing.
2. $$g$$ is convex, $$h$$ is convex and non-increasing.

Show for 1D case ( $$n = 1$$). Get to $$n > 1$$ by applying to all lines.

1. \label{eqn:convexOptimizationLecture8:180}
\begin{aligned}
F'(x) &= h'(g(x)) g'(x) \\
F”(x) &=
h”(g(x)) g'(x) g'(x)
+
h'(g(x)) g”(x) \\
&=
h”(g(x)) (g'(x))^2
+
h'(g(x)) g”(x) \\
&=
\lr{ \ge 0 } \cdot \lr{ \ge 0 }^2 + \lr{ \ge 0 } \cdot \lr{ \ge 0 },
\end{aligned}

since $$h$$ is respectively convex, and non-decreasing.

2. \label{eqn:convexOptimizationLecture8:180b}
\begin{aligned}
F'(x) =
\lr{ \ge 0 } \cdot \lr{ \ge 0 }^2 + \lr{ \le 0 } \cdot \lr{ \le 0 },
\end{aligned}

since $$h$$ is respectively convex, and non-increasing, and g is concave.

## Extending to multiple dimensions

\label{eqn:convexOptimizationLecture8:200}
\begin{aligned}
F(\Bx)
&= h(g(\Bx)) = h( g_1(\Bx), g_2(\Bx), \cdots g_k(\Bx) ) \\
g &: \mathbb{R}^n \rightarrow \mathbb{R} \\
h &: \mathbb{R}^k \rightarrow \mathbb{R}.
\end{aligned}

is convex if $$g_i$$ is convex for each $$i \in [1,k]$$ and $$h$$ is convex and non-decreasing in each argument.

Proof:

again assume $$n = 1$$, without loss of generality,

\label{eqn:convexOptimizationLecture8:220}
\begin{aligned}
g &: \mathbb{R} \rightarrow \mathbb{R}^k \\
h &: \mathbb{R}^k \rightarrow \mathbb{R} \\
\end{aligned}

\label{eqn:convexOptimizationLecture8:240}
F”(\Bx)
=
\begin{bmatrix}
g_1(\Bx) & g_2(\Bx) & \cdots & g_k(\Bx)
\end{bmatrix}
\begin{bmatrix}
g_1′(\Bx) \\ g_2′(\Bx) \\ \vdots \\ g_k'(\Bx)
\end{bmatrix}
+
\begin{bmatrix}
g_1”(\Bx) \\ g_2”(\Bx) \\ \vdots \\ g_k”(\Bx)
\end{bmatrix}

The Hessian is PSD.

### Example:

\label{eqn:convexOptimizationLecture8:260}
F(x) = \exp( g(x) ) = h( g(x) ),

where $$g$$ is convex is convex, and $$h(y) = e^y$$. This implies that $$F$$ is a convex function.

### Example:

\label{eqn:convexOptimizationLecture8:280}
F(x) = \inv{g(x)},

is convex if $$g(x)$$ is concave and positive. The most simple such example of such a function is $$h(x) = 1/x, \textrm{dom} h = \mathbb{R}_{++}$$, which is plotted in fig. 4.

fig. 4. Inverse function is convex over positive domain.

### Example:

\label{eqn:convexOptimizationLecture8:300}
F(x) = – \sum_{i = 1}^n \log( -F_i(x) )

is convex on $$\setlr{ x | F_i(x) < 0 \forall i }$$ if all $$F_i$$ are convex.

• Due to $$\textrm{dom} F$$, $$-F_i(x) > 0 \,\forall x \in \textrm{dom} F$$
• $$\log(x)$$ concave on $$\mathbb{R}_{++}$$ so $$-\log$$ convex also non-increasing (fig. 5).

fig. 5. Negative logarithm convex over positive domain.

\label{eqn:convexOptimizationLecture8:320}
F(x) = \sum h_i(x)

but
\label{eqn:convexOptimizationLecture8:340}
h_i(x) = -\log(-F_i(x)),

which is a convex and non-increasing function ($$-\log$$), of a convex function $$-F_i(x)$$. Each
$$h_i$$ is convex, so this is a sum of convex functions, and is therefore convex.

### Example:

Over $$\textrm{dom} F = S^n_{++}$$

\label{eqn:convexOptimizationLecture8:360}
F(X) = \log \det X^{-1}

To show that this is convex, check all lines in domain. A line in $$S^n_{++}$$ is a 1D family of matrices

\label{eqn:convexOptimizationLecture8:380}
\tilde{F}(t) = \log \det( \lr{X_0 + t H}^{-1} ),

where $$X_0 \in S^n_{++}, t \in \mathbb{R}, H \in S^n$$.

F9

For $$t$$ small enough,

\label{eqn:convexOptimizationLecture8:400}
X_0 + t H \in S^n_{++}

\label{eqn:convexOptimizationLecture8:420}
\begin{aligned}
\tilde{F}(t)
&= \log \det( \lr{X_0 + t H}^{-1} ) \\
&= \log \det\lr{ X_0^{-1/2} \lr{I + t X_0^{-1/2} H X_0^{-1/2} }^{-1} X_0^{-1/2} } \\
&= \log \det\lr{ X_0^{-1} \lr{I + t X_0^{-1/2} H X_0^{-1/2} }^{-1} } \\
&= \log \det X_0^{-1} + \log\det \lr{I + t X_0^{-1/2} H X_0^{-1/2} }^{-1} \\
&= \log \det X_0^{-1} – \log\det \lr{I + t X_0^{-1/2} H X_0^{-1/2} } \\
&= \log \det X_0^{-1} – \log\det \lr{I + t M }.
\end{aligned}

If $$\lambda_i$$ are eigenvalues of $$M$$, then $$1 + t \lambda_i$$ are eigenvalues of $$I + t M$$. i.e.:

\label{eqn:convexOptimizationLecture8:440}
\begin{aligned}
(I + t M) \Bv
&=
I \Bv + t \lambda_i \Bv \\
&=
(1 + t \lambda_i) \Bv.
\end{aligned}

This gives

\label{eqn:convexOptimizationLecture8:460}
\begin{aligned}
\tilde{F}(t)
&= \log \det X_0^{-1} – \log \prod_{i = 1}^n (1 + t \lambda_i) \\
&= \log \det X_0^{-1} – \sum_{i = 1}^n \log (1 + t \lambda_i)
\end{aligned}

• $$1 + t \lambda_i$$ is linear in $$t$$.
• $$-\log$$ is convex in its argument.
• sum of convex function is convex.

### Example:

\label{eqn:convexOptimizationLecture8:480}
F(X) = \lambda_\max(X),

is convex on $$\textrm{dom} F \in S^n$$

(a)
\label{eqn:convexOptimizationLecture8:500}
\lambda_{\max} (X) = \sup_{\Norm{\Bv}_2 \le 1} \Bv^\T X \Bv,

\label{eqn:convexOptimizationLecture8:520}
\begin{bmatrix}
\lambda_1 & & & \\
& \lambda_2 & & \\
& & \ddots & \\
& & & \lambda_n
\end{bmatrix}

Recall that a decomposition

\label{eqn:convexOptimizationLecture8:540}
\begin{aligned}
X &= Q \Lambda Q^\T \\
Q^\T Q = Q Q^\T = I
\end{aligned}

can be used for any $$X \in S^n$$.

(b)

Note that $$\Bv^\T X \Bv$$ is linear in $$X$$. This is a max of a number of linear (and convex) functions, so it is convex.

Last example:

(non-symmetric matrices)

\label{eqn:convexOptimizationLecture8:560}
F(X) = \sigma_\max(X),

is convex on $$\textrm{dom} F = \mathbb{R}^{m \times n}$$. Here

\label{eqn:convexOptimizationLecture8:580}
\sigma_\max(X) = \sup_{\Norm{\Bv}_2 = 1} \Norm{X \Bv}_2

This is called an operator norm of $$X$$. Using the SVD

\label{eqn:convexOptimizationLecture8:600}
\begin{aligned}
X &= U sectionigma V^\T \\
U &= \mathbb{R}^{m \times r} \\
sectionigma &\in \mathrm{diag} \in \mathbb{R}{ r \times r } \\
V^T &\in \mathbb{R}^{r \times n}.
\end{aligned}

Have

\label{eqn:convexOptimizationLecture8:620}
\Norm{X \Bv}_2^2
=
\Norm{ U sectionigma V^\T \Bv }_2^2
=
\Bv^\T V sectionigma U^\T U sectionigma V^\T \Bv
=
\Bv^\T V sectionigma sectionigma V^\T \Bv
=
\Bv^\T V sectionigma^2 V^\T \Bv
=
\tilde{\Bv}^\T sectionigma^2 \tilde{\Bv},

where $$\tilde{\Bv} = \Bv^\T V$$, so

\label{eqn:convexOptimizationLecture8:640}
\Norm{X \Bv}_2^2
=
\sum_{i = 1}^r \sigma_i^2 \Norm{\tilde{\Bv}}
\le \sigma_\max^2 \Norm{\tilde{\Bv}}^2,

or
\label{eqn:convexOptimizationLecture8:660}
\Norm{X \Bv}_2
\le \sqrt{ \sigma_\max^2 } \Norm{\tilde{\Bv}}
\le
\sigma_\max.

Set $$\Bv$$ to the right singular value of $$X$$ to get equality.

# References

[1] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004.

## Posted notes for Electromagnetic Theory (ECE1228H), taught by Prof. M. Mojahedi, fall 2016

February 3, 2017 math and physics play No comments ,

I’ve now posted redacted notes for the Electromagnetic Theory (ECE1228H) course I took last fall, taught by Prof. M. Mojahedi.  This course covered a subset of the following:

• Maxwell’s equations
• constitutive relations and boundary conditions
• wave polarization.
• Field representations: potentials
• Green’s functions and integral equations.
• Theorems and concepts: duality, uniqueness, images, equivalence, reciprocity and Babinet’s principles.
• Plane cylindrical and spherical waves and waveguides.

These notes are fairly compact, only 183 pages, with the full version weighing in at 256 pages.

As always, feel free to contact me for the complete version (i.e. including my problem set solutions) if you interested, but not asking because you are taking or planning to take this course.

## ECE1505H Convex Optimization. Lecture 6: First and second order conditions. Taught by Prof.\ Stark Draper

### Disclaimer

Peeter’s lecture notes from class. These may be incoherent and rough.

These are notes for the UofT course ECE1505H, Convex Optimization, taught by Prof. Stark Draper, from [1].

### Today

• First and second order conditions for convexity of differentiable functions.
• Consequences of convexity: local and global optimality.
• Properties.

### Quasi-convex

$$F_1$$ and $$F_2$$ convex implies $$\max( F_1, F_2)$$ convex.

fig. 1. Min and Max

Note that $$\min(F_1, F_2)$$ is NOT convex.

If $$F : \mathbb{R}^n \rightarrow \mathbb{R}$$ is convex, then $$F( \Bx_0 + t \Bv )$$ is convex in $$t\,\forall t \in \mathbb{R}, \Bx_0 \in \mathbb{R}^n, \Bv \in \mathbb{R}^n$$, provided $$\Bx_0 + t \Bv \in \textrm{dom} F$$.

Idea: Restrict to a line (line segment) in $$\textrm{dom} F$$. Take a cross section or slice through $$F$$ alone the line. If the result is a 1D convex function for all slices, then $$F$$ is convex.

This is nice since it allows for checking for convexity, and is also nice numerically. Attempting to test a given data set for non-convexity with some random lines can help disprove convexity. However, to show that $$F$$ is convex it is required to test all possible slices (which isn’t possible numerically, but is in some circumstances possible analytically).

### Differentiable (convex) functions

Definition: First order condition.

If

\begin{equation*}
F : \mathbb{R}^n \rightarrow \mathbb{R}
\end{equation*}

is differentiable, then $$F$$ is convex iff $$\textrm{dom} F$$ is a convex set and $$\forall \Bx, \Bx_0 \in \textrm{dom} F$$

\begin{equation*}
F(\Bx) \ge F(\Bx_0) + \lr{\spacegrad F(\Bx_0)}^\T (\Bx – \Bx_0).
\end{equation*}

This is the first order Taylor expansion. If $$n = 1$$, this is $$F(x) \ge F(x_0) + F'(x_0) ( x – x_0)$$.

The first order condition says a convex function \underline{always} lies above its first order approximation, as sketched in fig. 3.

fig. 2. First order approximation lies below convex function

When differentiable, the supporting plane is the tangent plane.

Definition: Second order condition

If $$F : \mathbb{R}^n \rightarrow \mathbb{R}$$ is twice differentiable, then $$F$$ is convex iff $$\textrm{dom} F$$ is a convex set and $$\spacegrad^2 F(\Bx) \ge 0 \,\forall \Bx \in \textrm{dom} F$$.

The Hessian is always symmetric, but is not necessarily positive. Recall that the Hessian is the matrix of the second order partials $$(\spacegrad F)_{ij} = \partial^2 F/(\partial x_i \partial x_j)$$.

The scalar case is $$F”(x) \ge 0 \, \forall x \in \textrm{dom} F$$.

An implication is that if $$F$$ is convex, then $$F(x) \ge F(x_0) + F'(x_0) (x – x_0) \,\forall x, x_0 \in \textrm{dom} F$$

Since $$F$$ is convex, $$\textrm{dom} F$$ is convex.

Consider any 2 points $$x, y \in \textrm{dom} F$$, and $$\theta \in [0,1]$$. Define

\label{eqn:convexOptimizationLecture6:60}
z = (1-\theta) x + \theta y \in \textrm{dom} F,

then since $$\textrm{dom} F$$ is convex

\label{eqn:convexOptimizationLecture6:80}
F(z) =
F( (1-\theta) x + \theta y )
\le
(1-\theta) F(x) + \theta F(y )

Reordering

\label{eqn:convexOptimizationLecture6:220}
\theta F(x) \ge
\theta F(x) + F(z) – F(x),

or
\label{eqn:convexOptimizationLecture6:100}
F(y) \ge
F(x) + \frac{F(x + \theta(y-x)) – F(x)}{\theta},

which is, in the limit,

\label{eqn:convexOptimizationLecture6:120}
F(y) \ge
F(x) + F'(x) (y – x),

completing one direction of the proof.

To prove the other direction, showing that

\label{eqn:convexOptimizationLecture6:140}
F(x) \ge F(x_0) + F'(x_0) (x – x_0),

implies that $$F$$ is convex. Take any $$x, y \in \textrm{dom} F$$ and any $$\theta \in [0,1]$$. Define

\label{eqn:convexOptimizationLecture6:160}
z = \theta x + (1 -\theta) y,

which is in $$\textrm{dom} F$$ by assumption. We want to show that

\label{eqn:convexOptimizationLecture6:180}
F(z) \le \theta F(x) + (1-\theta) F(y).

By assumption

1. $$F(x) \ge F(z) + F'(z) (x – z)$$
2. $$F(y) \ge F(z) + F'(z) (y – z)$$

Compute

\label{eqn:convexOptimizationLecture6:200}
\begin{aligned}
\theta F(x) + (1-\theta) F(y)
&\ge
\theta \lr{ F(z) + F'(z) (x – z) }
+ (1-\theta) \lr{ F(z) + F'(z) (y – z) } \\
&=
F(z) + F'(z) \lr{ \theta( x – z) + (1-\theta) (y-z) } \\
&=
F(z) + F'(z) \lr{ \theta x + (1-\theta) y – \theta z – (1 -\theta) z } \\
&=
F(z) + F'(z) \lr{ \theta x + (1-\theta) y – z} \\
&=
F(z) + F'(z) \lr{ z – z} \\
&= F(z).
\end{aligned}

### Proof of the 2nd order case for $$n = 1$$

Want to prove that if

\label{eqn:convexOptimizationLecture6:240}
F : \mathbb{R} \rightarrow \mathbb{R}

is a convex function, then $$F”(x) \ge 0 \,\forall x \in \textrm{dom} F$$.

By the first order conditions $$\forall x \ne y \in \textrm{dom} F$$

\label{eqn:convexOptimizationLecture6:260}
\begin{aligned}
F(y) &\ge F(x) + F'(x) (y – x)
F(x) &\ge F(y) + F'(y) (x – y)
\end{aligned}

Can combine and get

\label{eqn:convexOptimizationLecture6:280}
F'(x) (y-x) \le F(y) – F(x) \le F'(y)(y-x)

Subtract the two derivative terms for

\label{eqn:convexOptimizationLecture6:340}
\frac{(F'(y) – F'(x))(y – x)}{(y – x)^2} \ge 0,

or
\label{eqn:convexOptimizationLecture6:300}
\frac{F'(y) – F'(x)}{y – x} \ge 0.

In the limit as $$y \rightarrow x$$, this is
\label{eqn:convexOptimizationLecture6:320}
\boxed{
F”(x) \ge 0 \,\forall x \in \textrm{dom} F.
}

Now prove the reverse condition:

If $$F”(x) \ge 0 \,\forall x \in \textrm{dom} F \subseteq \mathbb{R}$$, implies that $$F : \mathbb{R} \rightarrow \mathbb{R}$$ is convex.

Note that if $$F”(x) \ge 0$$, then $$F'(x)$$ is non-decreasing in $$x$$.

i.e. If $$x < y$$, where $$x, y \in \textrm{dom} F$$, then

\label{eqn:convexOptimizationLecture6:360}
F'(x) \le F'(y).

Consider any $$x,y \in \textrm{dom} F$$ such that $$x < y$$, where

\label{eqn:convexOptimizationLecture6:380}
F(y) – F(x) = \int_x^y F'(t) dt \ge F'(x) \int_x^y 1 dt = F'(x) (y-x).

This tells us that

\label{eqn:convexOptimizationLecture6:400}
F(y) \ge F(x) + F'(x)(y – x),

which is the first order condition. Similarly consider any $$x,y \in \textrm{dom} F$$ such that $$x < y$$, where

\label{eqn:convexOptimizationLecture6:420}
F(y) – F(x) = \int_x^y F'(t) dt \le F'(y) \int_x^y 1 dt = F'(y) (y-x).

This tells us that

\label{eqn:convexOptimizationLecture6:440}
F(x) \ge F(y) + F'(y)(x – y).

### Vector proof:

$$F$$ is convex iff $$F(\Bx + t \Bv)$$ is convex $$\forall \Bx,\Bv \in \mathbb{R}^n, t \in \mathbb{R}$$, keeping $$\Bx + t \Bv \in \textrm{dom} F$$.

Let
\label{eqn:convexOptimizationLecture6:460}
h(t ; \Bx, \Bv) = F(\Bx + t \Bv)

then $$h(t)$$ satisfies scalar first and second order conditions for all $$\Bx, \Bv$$.

\label{eqn:convexOptimizationLecture6:480}
h(t) = F(\Bx + t \Bv) = F(g(t)),

where $$g(t) = \Bx + t \Bv$$, where

\label{eqn:convexOptimizationLecture6:500}
\begin{aligned}
F &: \mathbb{R}^n \rightarrow \mathbb{R} \\
g &: \mathbb{R} \rightarrow \mathbb{R}^n.
\end{aligned}

This is expressing $$h(t)$$ as a composition of two functions. By the first order condition for scalar functions we know that

\label{eqn:convexOptimizationLecture6:520}
h(t) \ge h(0) + h'(0) t.

Note that

\label{eqn:convexOptimizationLecture6:540}
h(0) = \evalbar{F(\Bx + t \Bv)}{t = 0} = F(\Bx).

Let’s figure out what $$h'(0)$$ is. Recall hat for any $$\tilde{F} : \mathbb{R}^n \rightarrow \mathbb{R}^m$$

\label{eqn:convexOptimizationLecture6:560}
D \tilde{F} \in \mathbb{R}^{m \times n},

and
\label{eqn:convexOptimizationLecture6:580}
{D \tilde{F}(\Bx)}_{ij} = \PD{x_j}{\tilde{F_i}(\Bx)}

This is one function per row, for $$i \in [1,m], j \in [1,n]$$. This gives

\label{eqn:convexOptimizationLecture6:600}
\begin{aligned}
\frac{d}{dt} F(\Bx + \Bv t)
&=
\frac{d}{dt} F( g(t) ) \\
&=
\frac{d}{dt} h(t) \\
&= D h(t) \\
&= D F(g(t)) \cdot D g(t)
\end{aligned}

The first matrix is in $$\mathbb{R}^{1\times n}$$ whereas the second is in $$\mathbb{R}^{n\times 1}$$, since $$F : \mathbb{R}^n \rightarrow \mathbb{R}$$ and $$g : \mathbb{R} \rightarrow \mathbb{R}^n$$. This gives

\label{eqn:convexOptimizationLecture6:620}
\frac{d}{dt} F(\Bx + \Bv t)
= \evalbar{D F(\tilde{\Bx})}{\tilde{\Bx} = g(t)} \cdot D g(t).

That first matrix is

\label{eqn:convexOptimizationLecture6:640}
\begin{aligned}
\evalbar{D F(\tilde{\Bx})}{\tilde{\Bx} = g(t)}
&=
\evalbar{
\lr{\begin{bmatrix}
\PD{\tilde{x}_1}{ F(\tilde{\Bx})} &
\PD{\tilde{x}_2}{ F(\tilde{\Bx})} & \cdots
\PD{\tilde{x}_n}{ F(\tilde{\Bx})}
\end{bmatrix}
}}{ \tilde{\Bx} = g(t) = \Bx + t \Bv } \\
&=
\evalbar{
}{
\tilde{\Bx} = g(t)
} \\
=
\end{aligned}

The second Jacobian is

\label{eqn:convexOptimizationLecture6:660}
D g(t)
=
D
\begin{bmatrix}
g_1(t) \\
g_2(t) \\
\vdots \\
g_n(t) \\
\end{bmatrix}
=
D
\begin{bmatrix}
x_1 + t v_1 \\
x_2 + t v_2 \\
\vdots \\
x_n + t v_n \\
\end{bmatrix}
=
\begin{bmatrix}
v_1 \\
v_1 \\
\vdots \\
v_n \\
\end{bmatrix}
=
\Bv.

so

\label{eqn:convexOptimizationLecture6:680}
h'(t) = D h(t) = \lr{ \spacegrad F(g(t))}^\T \Bv,

and
\label{eqn:convexOptimizationLecture6:700}
h'(0) = \lr{ \spacegrad F(g(0))}^\T \Bv
=

Finally

\label{eqn:convexOptimizationLecture6:720}
\begin{aligned}
F(\Bx + t \Bv)
&\ge h(0) + h'(0) t \\
&= F(\Bx) + \lr{ \spacegrad F(\Bx) }^\T (t \Bv) \\
&= F(\Bx) + \innerprod{ \spacegrad F(\Bx) }{ t \Bv}.
\end{aligned}

Which is true for all $$\Bx, \Bx + t \Bv \in \textrm{dom} F$$. Note that the quantity $$t \Bv$$ is a shift.

### Epigraph

Recall that if $$(\Bx, t) \in \textrm{epi} F$$ then $$t \ge F(\Bx)$$.

\label{eqn:convexOptimizationLecture6:740}
t \ge F(\Bx) \ge F(\Bx_0) + \lr{\spacegrad F(\Bx_0) }^\T (\Bx – \Bx_0),

or

\label{eqn:convexOptimizationLecture6:760}
0 \ge
-(t – F(\Bx_0)) + \lr{\spacegrad F(\Bx_0) }^\T (\Bx – \Bx_0),

In block matrix form

\label{eqn:convexOptimizationLecture6:780}
0 \ge
\begin{bmatrix}
\lr{ \spacegrad F(\Bx_0) }^\T & -1
\end{bmatrix}
\begin{bmatrix}
\Bx – \Bx_0 \\
t – F(\Bx_0)
\end{bmatrix}

With $$\Bw = \begin{bmatrix} \lr{ \spacegrad F(\Bx_0) }^\T & -1 \end{bmatrix}$$, the geometry of the epigraph relation to the half plane is sketched in fig. 3.

fig. 3. Half planes and epigraph.

# References

[1] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004.

## Disclaimer

Peeter’s lecture notes from class. These may be incoherent and rough.

These are notes for the UofT course ECE1505H, Convex Optimization, taught by Prof. Stark Draper, from [1].

## Last time

• examples of sets: planes, half spaces, balls, ellipses, cone of positive semi-definite matrices
• generalized inequalities
• examples of convexity preserving operations

## Today

• more examples of convexity preserving operations
• separating and supporting hyperplanes
• basic definitions of convex functions
• epigraphs, quasi-convexity, sublevel sets
• first and second order conditions for convexity of differentiable functions.

## Operations that preserve convexity

If $$S_\alpha$$ is convex $$\forall \alpha \in A$$, then

\label{eqn:convexOptimizationLecture5:40}
\cup_{\alpha \in A} S_\alpha,

is convex.

Example:

\label{eqn:convexOptimizationLecture5:60}
F(\Bx) = A \Bx + \Bb

\label{eqn:convexOptimizationLecture5:80}
\begin{aligned}
\Bx &\in \mathbb{R}^n \\
A &\in \mathbb{R}^{m \times n} \\
F &: \mathbb{R}^{n} \rightarrow \mathbb{R}^m \\
\Bb &\in \mathbb{R}^m
\end{aligned}

1. If $$S \in \mathbb{R}^n$$ is convex, then\label{eqn:convexOptimizationLecture5:100}
F(S) = \setlr{ F(\Bx) | \Bx \in S }
is convex if $$F$$ is affine.
2. If $$S \in \mathbb{R}^m$$ is convex, then\label{eqn:convexOptimizationLecture5:120}
F^{-1}(S) = \setlr{ \Bx | F(\Bx) \in S }

is convex.

Example:

\label{eqn:convexOptimizationLecture5:140}
\setlr{ \By | \By = A \Bx + \Bb, \Norm{\Bx} \le 1}

is convex. Here $$A \Bx + \Bb$$ is an affine function ($$F(\Bx)$$. This is the image of a (convex) unit ball, through an affine map.

Earlier saw when defining ellipses

\label{eqn:convexOptimizationLecture5:160}
\By = P^{1/2} \Bx + \Bx_c

Example :

\label{eqn:convexOptimizationLecture5:180}
\setlr{ \Bx | \Norm{ A \Bx + \Bb } \le 1 },

is convex. This can be seen by writing

\label{eqn:convexOptimizationLecture5:200}
\begin{aligned}
\setlr{ \Bx | \Norm{ A \Bx + \Bb } \le 1 }
&=
\setlr{ \Bx | \Norm{ F(\Bx) } \le 1 } \\
&=
\setlr{ \Bx | F(\Bx) \in \mathcal{B} },
\end{aligned}

where $$\mathcal{B} = \setlr{ \By | \Norm{\By} \le 1 }$$. This is the pre-image (under $$F()$$) of a unit norm ball.

Example:

\label{eqn:convexOptimizationLecture5:220}
\setlr{ \Bx \in \mathbb{R}^n | x_1 A_1 + x_2 A_2 + \cdots x_n A_n \le \mathcal{B} }

where $$A_i \in S^m$$ and $$\mathcal{B} \in S^m$$, and the inequality is a matrix inequality. This is a convex set. The constraint is a “linear matrix inequality” (LMI).

This has to do with an affine map:

\label{eqn:convexOptimizationLecture5:240}
F(\Bx) = B – 1 x_1 A_1 – x_2 A_2 – \cdots x_n A_n \ge 0

(positive semi-definite inequality). This is a mapping

\label{eqn:convexOptimizationLecture5:480}
F : \mathbb{R}^n \rightarrow S^m,

since all $$A_i$$ and $$B$$ are in $$S^m$$.

This $$F(\Bx) = B – A(\Bx)$$ is a constant and a factor linear in x, so is affine. Can be written

\label{eqn:convexOptimizationLecture5:260}
\setlr{ \Bx | B – A(\Bx) \ge 0 }
=
\setlr{ \Bx | B – A(\Bx) \in S^m_{+} }

This is a pre-image of a cone of PSD matrices, which is convex. Therefore, this is a convex set.

## Separating hyperplanes

Theorem: Separating hyperplanes

If $$S, T \subseteq \mathbb{R}^n$$ are convex and disjoint
i.e. $$S \cup T = 0$$, then
there exists on $$\Ba \in \mathbb{R}^n$$ $$\Ba \ne 0$$ and a $$\Bb \in \mathbb{R}^n$$ such that

\begin{equation*}
\Ba^\T \Bx \ge \Bb \, \forall \Bx \in S
\end{equation*}

and
\begin{equation*}
\Ba^\T \Bx < \Bb \,\forall \Bx \in T.
\end{equation*}

An example of a hyperplanes that separates two sets and two sets that are not separable is sketched in fig 1.1

Proof in the book.

Theorem: Supporting hyperplane
If $$S$$ is convex then $$\forall x_0 \in \partial S = \textrm{cl}(S) \ \textrm{int}(S)$$, where
$$\partial S$$ is the boundary of $$S$$, then $$\exists$$ an $$\Ba \ne 0 \in \mathbb{R}^n$$ such that $$\Ba^\T \Bx \le \Ba^\T x_0 \, \forall \Bx \in S$$.

Here $$\$$ denotes “without”.

An example is sketched in fig. 3, for which

fig. 3. Supporting hyperplane.

• The vector $$\Ba$$ perpendicular to tangent plane.
• inner product $$\Ba^\T (\Bx – \Bx_0) \le 0$$.

A set with a supporting hyperplane is sketched in fig 4a whereas fig 4b shows that there is not necessarily a unique supporting hyperplane at any given point, even if $$S$$ is convex.

fig 4a. Set with supporting hyperplane.

fig 4b. No unique supporting hyperplane possible.

## basic definitions of convex functions

Theorem: Convex functions
If $$F : \mathbb{R}^n \rightarrow \mathbb{R}$$ is defined on a convex domain (i.e. $$\textrm{dom} F \subseteq \mathbb{R}^n$$ is a convex set), then $$F$$ is convex if $$\forall \Bx, \By \in \textrm{dom} F$$, $$\forall \theta \in [0,1] \in \mathbb{R}$$

\label{eqn:convexOptimizationLecture5:340}
F( \theta \Bx + (1-\theta) \By \le \theta F(\Bx) + (1-\theta) F(\By)

An example is sketched in fig. 5.

fig. 5. Example of convex function.

Remarks

• Require $$\textrm{dom} F$$ to be a convex set. This is required so that the function at the point $$\theta u + (1-\theta) v$$ can be evaluated. i.e. so that $$F(\theta u + (1-\theta) v)$$ is well defined. Example: $$\textrm{dom} F = (-\infty, 0] \cup [1, \infty)$$ is not okay, because a linear combination in $$(0,1)$$ would be undesirable.
• Parameter $$\theta$$ is “how much up” the line segment connecting $$(u, F(u)$$ and $$(v, F(v)$$. This line segment never below the bottom of the bowl.
The function is \underlineAndIndex{concave}, if $$-F$$ is convex.
i.e. If the convex function is flipped upside down. That is\label{eqn:convexOptimizationLecture5:360}
F(\theta \Bx + (1-\theta) \By ) \ge \theta F(\Bx) + (1-\theta) F(\By) \,\forall \Bx,\By \in \textrm{dom} F, \theta \in [0,1].
• a “strictly” convex function means $$\forall \theta \in [0,1]$$\label{eqn:convexOptimizationLecture5:380}
F(\theta \Bx + (1-\theta) \By ) < \theta F(\Bx) + (1-theta) F(\By).
• Strictly concave function $$F$$ means $$-F$$ is strictly convex.
• Examples:\imageFigure{../figures/ece1505-convex-optimization/l5Fig6a}{}{fig:l5:l5Fig6a}{0.2}

fig 6a. Not convex or concave.

fig 6b. Not strictly convex

Definition: Epigraph of a function

The epigraph $$\textrm{epi} F$$ of a function $$F : \mathbb{R}^n \rightarrow \mathbb{R}$$ is

\begin{equation*}
\textrm{epi} F = \setlr{ (\Bx,t) \in \mathbb{R}^{n +1} | \Bx \in \textrm{dom} F, t \ge F(\Bx) },
\end{equation*}

where $$\Bx \in \mathbb{R}^n, t \in \mathbb{R}$$.

fig. 7. Epigraph.

Theorem: Convexity and epigraph.
If $$F$$ is convex implies $$\textrm{epi} F$$ is a convex set.

Proof:

For convex function, a line segment connecting any 2 points on function is above the function. i.e. it is $$\textrm{epi} F$$.

Many authors will go the other way around, showing \ref{dfn:convexOptimizationLecture5:400} from \ref{thm:convexOptimizationLecture5:420}. That is:

Pick any 2 points in $$\textrm{epi} F$$, $$(\Bx,\mu) \in \textrm{epi} F$$ and $$(\By, \nu) \in \textrm{epi} F$$. Consider convex combination

\label{eqn:convexOptimizationLecture5:420}
\theta( \Bx, \mu ) + (1-\theta) (\By, \nu) =
(\theta \Bx (1-\theta) \By, \theta \mu (1-\theta) \nu )
\in \textrm{epi} F,

since $$\textrm{epi} F$$ is a convex set.

By definition of $$\textrm{epi} F$$

\label{eqn:convexOptimizationLecture5:440}
F( \theta \Bx (1-\theta) \By ) \le \theta \mu (1-\theta) \nu.

Picking $$\mu = F(\Bx), \nu = F(\By)$$ gives
\label{eqn:convexOptimizationLecture5:460}
F( \theta \Bx (1-\theta) \By ) \le \theta F(\Bx) (1-\theta) F(\By).

## Extended value function

Sometimes convenient to work with “extended value function”

\label{eqn:convexOptimizationLecture5:500}
\tilde{F}(\Bx) =
\left\{
\begin{array}{l l}
F(\Bx) & \quad \mbox{If $$\Bx \in \textrm{dom} F$$} \\
\end{array}
\right.

Examples:

• Linear (affine) functions (fig. 8) are both convex and concave.

fig. 8. Linear functions.

• $$x^2$$ is convex, sketched in fig. 9.

• $$\log x, \textrm{dom} F = \mathbb{R}_{+}$$ concave, sketched in fig. 10.

fig. 10. Concave (logarithm.)

• $$\Norm{\Bx}$$ is convex. $$\Norm{ \theta \Bx + (1-\theta) \By } \le \theta \Norm{ \Bx } + (1-\theta) \Norm{\By }$$.
• $$1/x$$ is convex on $$\setlr{ x | x > 0 } = \textrm{dom} F$$, and concave on $$\setlr{ x | x < 0 } = \textrm{dom} F$$. \label{eqn:convexOptimizationLecture5:520} \tilde{F}(x) = \left\{ \begin{array}{l l} \inv{x} & \quad \mbox{If $$x > 0$$} \\
\end{array}
\right.

Definition: Sublevel

The sublevel set of a function $$F : \mathbb{R}^n \rightarrow \mathbb{R}$$ is

\begin{equation*}
C(\alpha) = \setlr{ \Bx \in \textrm{dom} F | F(\Bx) \le \alpha }
\end{equation*}

Convex sublevel

Non-convex sublevel.

Theorem:
If $$F$$ is convex then $$C(\alpha)$$ is a convex set $$\forall \alpha$$.

This is not an if and only if condition, as illustrated in fig. 12.

fig. 12. Convex sublevel does not imply convexity.

There $$C(\alpha)$$ is convex, but the function itself is not.

Proof:

Since $$F$$ is convex, then $$\textrm{epi} F$$ is a convex set.

• Let\label{eqn:convexOptimizationLecture5:580}
\mathcal{A} = \setlr{ (\Bx,t) | t = \alpha }
is a convex set.
• $$\mathcal{A} \cap \textrm{epi} F$$is a convex set since it is the intersection of convex sets.
• Project $$\mathcal{A} \cap \textrm{epi} F$$ onto \R{n} (i.e. domain of $$F$$ ). The projection is an affine mapping. Image of a convex set through affine mapping is a convex set.

Definition: Quasi-convex.

A function is quasi-convex if \underline{all} of its sublevel sets are convex.

## Composing convex functions

Properties of convex functions:

• If $$F$$ is convex, then $$\alpha F$$ is convex $$\forall \alpha > 0$$.
• If $$F_1, F_2$$ are convex, then the sum $$F_1 + F_2$$ is convex.
• If $$F$$ is convex, then $$g(\Bx) = F(A \Bx + \Bb)$$ is convex $$\forall \Bx \in \setlr{ \Bx | A \Bx + \Bb \in \textrm{dom} F }$$.

Note: for the last

\label{eqn:convexOptimizationLecture5:620}
\begin{aligned}
g &: \mathbb{R}^m \rightarrow \mathbb{R} \\
F &: \mathbb{R}^n \rightarrow \mathbb{R} \\
\Bx &\in \mathbb{R}^m \\
A &\in \mathbb{R}^{n \times m} \\
\Bb &\in \mathbb{R}^n
\end{aligned}

Proof (of last):

\label{eqn:convexOptimizationLecture5:640}
\begin{aligned}
g( \theta \Bx + (1-\theta) \By )
&=
F( \theta (A \Bx + \Bb) + (1-\theta) (A \By + \Bb) ) \\
&\le
\theta F( A \Bx + \Bb) + (1-\theta) F (A \By + \Bb) \\
&= \theta g(\Bx) + (1-\theta) g(\By).
\end{aligned}

# References

[1] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004.

## ECE1505H Convex Optimization. Lecture 4: Sets and convexity. Taught by Prof. Stark Draper

January 25, 2017 ece1505 No comments , , , , , , ,

ECE1505H Convex Optimization. Lecture 4: Sets and convexity. Taught by Prof. Stark Draper

### Disclaimer

Peeter’s lecture notes from class. These may be incoherent and rough.

These are notes for the UofT course ECE1505H, Convex Optimization, taught by Prof. Stark Draper, covering [1] content.

### Today

• more on various sets: hyperplanes, half-spaces, polyhedra, balls, ellipses, norm balls, cone of PSD
• generalize inequalities
• operations that preserve convexity
• separating and supporting hyperplanes.

## Hyperplanes

Find some $$\Bx_0 \in \mathbb{R}^n$$ such that $$\Ba^\T \Bx_0 = \Bb$$, so

\label{eqn:convexOptimizationLecture4:20}
\begin{aligned}
\setlr{ \Bx | \Ba^\T \Bx = \Bb }
&=
\setlr{ \Bx | \Ba^\T \Bx = \Ba^\T \Bx_0 } \\
&=
\setlr{ \Bx | \Ba^\T (\Bx – \Bx_0) } \\
&=
\Bx_0 + \Ba^\perp,
\end{aligned}

where

\label{eqn:convexOptimizationLecture4:40}
\Ba^\perp = \setlr{ \Bv | \Ba^\T \Bv = 0 }.

fig. 1. Parallel hyperplanes.

Recall

\label{eqn:convexOptimizationLecture4:60}
\Norm{\Bz}_\conj = \sup_\Bx \setlr{ \Bz^\T \Bx | \Norm{\Bx} \le 1 }

Denote the optimizer of above as $$\Bx^\conj$$. By definition

\label{eqn:convexOptimizationLecture4:80}
\Bz^\T \Bx^\conj \ge \Bz^\T \Bx \quad \forall \Bx, \Norm{\Bx} \le 1

This defines a half space in which the unit ball

\label{eqn:convexOptimizationLecture4:100}
\setlr{ \Bx | \Bz^\T (\Bx – \Bx^\conj \le 0 }

Start with the $$l_1$$ norm, duals of $$l_1$$ is $$l_\infty$$

fig. 2. Half space containing unit ball.

Similar pic for $$l_\infty$$, for which the dual is the $$l_1$$ norm, as sketched in fig. 3.  Here the optimizer point is at $$(1,1)$$

fig. 3. Half space containing the unit ball for l_infinity

and a similar pic for $$l_2$$, which is sketched in fig. 4.

fig. 4. Half space containing for l_2 unit ball.

## Polyhedra

\label{eqn:convexOptimizationLecture4:120}
\begin{aligned}
\mathcal{P}
&= \setlr{ \Bx |
\Ba_j^\T \Bx \le \Bb_j, j \in [1,m],
\Bc_i^\T \Bx = \Bd_i, i \in [1,p]
} \\
&=
\setlr{ \Bx | A \Bx \le \Bb, C \Bx = d },
\end{aligned}

where the final inequality and equality are component wise.

Proving $$\mathcal{P}$$ is convex:

• Pick $$\Bx_1 \in \mathcal{P}$$, $$\Bx_2 \in \mathcal{P}$$
• Pick any $$\theta \in [0,1]$$
• Test $$\theta \Bx_1 + (1-\theta) \Bx_2$$. Is it in $$\mathcal{P}$$?

\label{eqn:convexOptimizationLecture4:140}
\begin{aligned}
A \lr{ \theta \Bx_1 + (1-\theta) \Bx_2 }
&=
\theta A \Bx_1 + (1-\theta) A \Bx_2 \\
&\le
\theta \Bb + (1-\theta) \Bb \\
&=
\Bb.
\end{aligned}

## Balls

Euclidean ball for $$\Bx_c \in \mathbb{R}^n, r \in \mathbb{R}$$

\label{eqn:convexOptimizationLecture4:160}
\mathcal{B}(\Bx_c, r)
= \setlr{ \Bx | \Norm{\Bx – \Bx_c}_2 \le r },

or
\label{eqn:convexOptimizationLecture4:180}
\mathcal{B}(\Bx_c, r)
= \setlr{ \Bx | \lr{\Bx – \Bx_c}^\T \lr{\Bx – \Bx_c} \le r^2 }.

Let $$\Bx_1, \Bx_2$$, $$\theta \in [0,1]$$

\label{eqn:convexOptimizationLecture4:200}
\begin{aligned}
\Norm{ \theta \Bx_1 + (1-\theta) \Bx_2 – \Bx_c }_2
&=
\Norm{ \theta (\Bx_1 – \Bx_c) + (1-\theta) (\Bx_2 – \Bx_c) }_2 \\
&\le
\Norm{ \theta (\Bx_1 – \Bx_c)}_2 + \Norm{(1-\theta) (\Bx_2 – \Bx_c) }_2 \\
&=
\Abs{\theta} \Norm{ \Bx_1 – \Bx_c}_2 + \Abs{1 -\theta} \Norm{ \Bx_2 – \Bx_c }_2 \\
&=
\theta \Norm{ \Bx_1 – \Bx_c}_2 + \lr{1 -\theta} \Norm{ \Bx_2 – \Bx_c }_2 \\
&\le
\theta r + (1 – \theta) r \\
&= r
\end{aligned}

## Ellipse

\label{eqn:convexOptimizationLecture4:220}
\mathcal{E}(\Bx_c, P)
=
\setlr{ \Bx | (\Bx – \Bx_c)^\T P^{-1} (\Bx – \Bx_c) \le 1 },

where $$P \in S^n_{++}$$.

• Euclidean ball is an ellipse with $$P = I r^2$$
• Ellipse is image of Euclidean ball $$\mathcal{B}(0,1)$$ under affine mapping.

fig. 5. Circle and ellipse.

Given

\label{eqn:convexOptimizationLecture4:240}
F(\Bu) = P^{1/2} \Bu + \Bx_c

\label{eqn:convexOptimizationLecture4:260}
\begin{aligned}
\setlr{ F(\Bu) | \Norm{\Bu}_2 \le r }
&=
\setlr{ P^{1/2} \Bu + \Bx_c | \Bu^\T \Bu \le r^2 } \\
&=
\setlr{ \Bx | \Bx = P^{1/2} \Bu + \Bx_c, \Bu^\T \Bu \le r^2 } \\
&=
\setlr{ \Bx | \Bu = P^{-1/2} (\Bx – \Bx_c), \Bu^\T \Bu \le r^2 } \\
&=
\setlr{ \Bx | (\Bx – \Bx_c)^\T P^{-1} (\Bx – \Bx_c) \le r^2 }
\end{aligned}

## Geometry of an ellipse

Decomposition of positive definite matrix $$P \in S^n_{++} \subset S^n$$ is:

\label{eqn:convexOptimizationLecture4:280}
\begin{aligned}
P &= Q \textrm{diag}(\lambda_i) Q^\T \\
Q^\T Q &= 1
\end{aligned},

where $$\lambda_i \in \mathbb{R}$$, and $$\lambda_i > 0$$. The ellipse is defined by

\label{eqn:convexOptimizationLecture4:300}
(\Bx – \Bx_c)^\T Q \textrm{diag}(1/\lambda_i) (\Bx – \Bx_c) Q \le r^2

The term $$(\Bx – \Bx_c)^\T Q$$ projects $$\Bx – \Bx_c$$ onto the columns of $$Q$$. Those columns are perpendicular since $$Q$$ is an orthogonal matrix. Let

\label{eqn:convexOptimizationLecture4:320}
\tilde{\Bx} = Q^\T (\Bx – \Bx_c),

this shifts the origin around $$\Bx_c$$ and $$Q$$ rotates into a new coordinate system. The ellipse is therefore

\label{eqn:convexOptimizationLecture4:340}
\tilde{\Bx}^\T
\begin{bmatrix}
\inv{\lambda_1} & & & \\
&\inv{\lambda_2} & & \\
& \ddots & \\
& & & \inv{\lambda_n}
\end{bmatrix}
\tilde{\Bx}
=
\sum_{i = 1}^n \frac{\tilde{x}_i^2}{\lambda_i} \le 1.

An example is sketched for $$\lambda_1 > \lambda_2$$ below.

Ellipse with $$\lambda_1 > \lambda_2$$.

• $$\lambda_i$$ tells us length of the semi-major axis.
• Larger $$\lambda_i$$ means $$\tilde{x}_i^2$$ can be bigger and still satisfy constraint $$\le 1$$.
• Volume of ellipse if proportional to $$\sqrt{ \det P } = \sqrt{ \prod_{i = 1}^n \lambda_i }$$.
• When any $$\lambda_i \rightarrow 0$$ a dimension is lost and the volume goes to zero. That removes the invertibility required.

Ellipses will be seen a lot in this course, since we are interested in “bowl” like geometries (and the ellipse is the image of a Euclidean ball).

## Norm ball.

The norm ball

\label{eqn:convexOptimizationLecture4:360}
\mathcal{B} = \setlr{ \Bx | \Norm{\Bx} \le 1 },

is a convex set for all norms. Proof:

Take any $$\Bx, \By \in \mathcal{B}$$

\label{eqn:convexOptimizationLecture4:380}
\Norm{ \theta \Bx + (1 – \theta) \By }
\le
\Abs{\theta} \Norm{ \Bx } + \Abs{1 – \theta} \Norm{ \By }
=
\theta \Norm{ \Bx } + \lr{1 – \theta} \Norm{ \By }
\lr
\theta + \lr{1 – \theta}
=
1.

This is true for any p-norm $$1 \le p$$, $$\Norm{\Bx}_p = \lr{ \sum_{i = 1}^n \Abs{x_i}^p }^{1/p}$$.

Norm ball.

The shape of a $$p < 1$$ norm unit ball is sketched below (lines connecting points in such a region can exit the region).

## Cones

Recall that $$C$$ is a cone if $$\forall \Bx \in C, \theta \ge 0, \theta \Bx \in C$$.

Impt cone of PSD matrices

\label{eqn:convexOptimizationLecture4:400}
\begin{aligned}
S^n &= \setlr{ X \in \mathbb{R}^{n \times n} | X = X^\T } \\
S^n_{+} &= \setlr{ X \in S^n | \Bv^\T X \Bv \ge 0, \quad \forall v \in \mathbb{R}^n } \\
S^n_{++} &= \setlr{ X \in S^n_{+} | \Bv^\T X \Bv > 0, \quad \forall v \in \mathbb{R}^n } \\
\end{aligned}

These have respectively

• $$\lambda_i \in \mathbb{R}$$
• $$\lambda_i \in \mathbb{R}_{+}$$
• $$\lambda_i \in \mathbb{R}_{++}$$

$$S^n_{+}$$ is a cone if:

$$X \in S^n_{+}$$, then $$\theta X \in S^n_{+}, \quad \forall \theta \ge 0$$

\label{eqn:convexOptimizationLecture4:420}
\Bv^\T (\theta X) \Bv
= \theta \Bv^\T \Bv
\ge 0,

since $$\theta \ge 0$$ and because $$X \in S^n_{+}$$.

Shorthand:

\label{eqn:convexOptimizationLecture4:440}
\begin{aligned}
X &\in S^n_{+} \Rightarrow X \succeq 0
X &\in S^n_{++} \Rightarrow X \succ 0.
\end{aligned}

Further $$S^n_{+}$$ is a convex cone.

Let $$A \in S^n_{+}$$, $$B \in S^n_{+}$$, $$\theta_1, \theta_2 \ge 0, \theta_1 + \theta_2 = 1$$, or $$\theta_2 = 1 – \theta_1$$.

Show that $$\theta_1 A + \theta_2 B \in S^n_{+}$$ :

\label{eqn:convexOptimizationLecture4:460}
\Bv^\T \lr{ \theta_1 A + \theta_2 B } \Bv
=
\theta_1 \Bv^\T A \Bv
+\theta_2 \Bv^\T B \Bv
\ge 0,

since $$\theta_1 \ge 0, \theta_2 \ge 0, \Bv^\T A \Bv \ge 0, \Bv^\T B \Bv \ge 0$$.

fig. 8. Cone.

Inequalities:

Start with a proper cone $$K \subseteq \mathbb{R}^n$$

• closed, convex
• non-empty interior (“solid”)
• “pointed” (contains no lines)

The $$K$$ defines a generalized inequality in \R{n} defined as “$$\le_K$$”

Interpreting

\label{eqn:convexOptimizationLecture4:480}
\begin{aligned}
\Bx \le_K \By &\leftrightarrow \By – \Bx \in K
\Bx \end{aligned}

Why pointed? Want if $$\Bx \le_K \By$$ and $$\By \le_K \Bx$$ with this $$K$$ is a half space.

Example:1: $$K = \mathbb{R}^n_{+}, \Bx \in \mathbb{R}^n, \By \in \mathbb{R}^n$$

fig. 12. K is non-negative “orthant”

\label{eqn:convexOptimizationLecture4:500}
\Bx \le_K \By \Rightarrow \By – \Bx \in K

say:

\label{eqn:convexOptimizationLecture4:520}
\begin{bmatrix}
y_1 – x_1
y_2 – x_2
\end{bmatrix}
\in R^2_{+}

Also:

\label{eqn:convexOptimizationLecture4:540}
K = R^1_{+}

(pointed, since it contains no rays)

\label{eqn:convexOptimizationLecture4:560}
\Bx \le_K \By ,

with respect to $$K = \mathbb{R}^n_{+}$$ means that $$x_i \le y_i$$ for all $$i \in [1,n]$$.

Example:2: For $$K = PSD \subseteq S^n$$,

\label{eqn:convexOptimizationLecture4:580}
\Bx \le_K \By ,

means that

\label{eqn:convexOptimizationLecture4:600}
\By – \Bx \in K = S^n_{+}.

• Difference $$\By – \Bx$$ is always in $$S$$
• check if in $$K$$ by checking if all eigenvalues $$\ge 0$$.
• $$S^n_{++}$$ is the interior of $$S^n_{+}$$.

Interpretation:

\label{eqn:convexOptimizationLecture4:620}
\begin{aligned}
\Bx \le_K \By &\leftrightarrow \By – \Bx \in K \\
\Bx \end{aligned}

We’ll use these with vectors and matrices so often the $$K$$ subscript will often be dropped, writing instead (for vectors)

\label{eqn:convexOptimizationLecture4:640}
\begin{aligned}
\Bx \le \By &\leftrightarrow \By – \Bx \in \mathbb{R}^n_{+} \\
\Bx < \By &\leftrightarrow \By – \Bx \in \textrm{int} \mathbb{R}^n_{++}
\end{aligned}

and for matrices

\label{eqn:convexOptimizationLecture4:660}
\begin{aligned}
\Bx \le \By &\leftrightarrow \By – \Bx \in S^n_{+} \\
\Bx < \By &\leftrightarrow \By – \Bx \in \textrm{int} S^n_{++}.
\end{aligned}

## Intersection

Take the intersection of (perhaps infinitely many) sets $$S_\alpha$$:

If $$S_\alpha$$ is (affine,convex, conic) for all $$\alpha \in A$$ then

\label{eqn:convexOptimizationLecture4:680}
\cap_\alpha S_\alpha

is (affine,convex, conic). To prove in homework:

\label{eqn:convexOptimizationLecture4:700}
\mathcal{P} = \setlr{ \Bx | \Ba_i^\T \Bx \le \Bb_i, \Bc_j^\T \Bx = \Bd_j, \quad \forall i \cdots j }

This is convex since the intersection of a bunch of hyperplane and half space constraints.

1. If $$S \subseteq \mathbb{R}^n$$ is convex then\label{eqn:convexOptimizationLecture4:720}
F(S) = \setlr{ F(\Bx) | \Bx \in S }
is convex.
2. If $$S \subseteq \mathbb{R}^m$$ then\label{eqn:convexOptimizationLecture4:740}
F^{-1}(S) = \setlr{ \Bx | F(\Bx) \in S }
is convex. Such a mapping is sketched in fig. 14.

fig. 14. Mapping functions of sets.

# References

[1] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004.