Skip to content

Instantly share code, notes, and snippets.

@hi-ogawa
Last active August 2, 2020 06:34
Show Gist options
  • Save hi-ogawa/d8f3bc9ac955f18b6fbd245667b275d1 to your computer and use it in GitHub Desktop.
Save hi-ogawa/d8f3bc9ac955f18b6fbd245667b275d1 to your computer and use it in GitHub Desktop.
misc
$$$
% From AsciiMath
\newcommand{\bb}[0]{\mathbf}
\newcommand{\bbb}[0]{\mathbb}
\newcommand{\cc}[0]{\mathcal}
\newcommand{\RR}[0]{\bbb{R}}
\newcommand{\CC}[0]{\bbb{C}}
\newcommand{\ZZ}[0]{\bbb{Z}}
\newcommand{\xx}[0]{\times}
\newcommand{\del}[0]{\partial}
% Quick begin/end
\newcommand{\Ba}[0]{\begin{aligned}}
\newcommand{\Ea}[0]{\end{aligned}}
\newcommand{\Bc}[0]{\begin{cases}}
\newcommand{\Ec}[0]{\end{cases}}
\newcommand{\Bm}[0]{\begin{matrix}}
\newcommand{\Em}[0]{\end{matrix}}
\newcommand{\Bmp}[0]{\begin{pmatrix}} % round bracket
\newcommand{\Emp}[0]{\end{pmatrix}}
% Others
\newcommand{\t}[0]{\text}
\newcommand{\l}[0]{\left}
\renewcommand{\r}[0]{\right}
\newcommand{\iprod}[2]{\left<#1,#2\right>}
$$$
## Mobius function
***Notation***
- $\ZZ_n = {0, 1, .., n-1}$: cyclic group of order n,
- $\ZZ^*_n = \{ x \mid \t{gcd}(x, n) = 1 \} \subset \ZZ_n$:
units of $\ZZ_n$ (multicative group),
- For the special case of $n = 1$, we treat $\ZZ^*_1 = \ZZ_1 = \{ 0\}$.
### Def. 1
$$
\mu(n) \equiv
\Bc
0 & (\t{if}~~ \exists i.~ e_i >= 2)
~~(\t{i.e.}~n: {\t{not square free}}), \\
(-1)^{|I|} & (\t{otherwise}). \\
\Ec
$$
where $n$ has prime factorization $n = \prod_{i \in I} p_i^{e_i}$.
***N.B.***
- It's easy to see $\mu$ is multiplicative.
### Def. 2
$$
\Ba
\t{SR}(n) &\equiv \sum_{k \in \ZZ_n} \exp(i 2 \pi k/n), \\
\t{SP}(n) &\equiv \sum_{k \in \ZZ^*_n} \exp(i 2 \pi k/n).
\Ea
$$
***N.B.***
- For the special case of $n = 1$, we have $\t{SP}(1) = \t{SR}(1) = \exp(0) = 1$.
### Prop. 3
$$
\ZZ_n = \bigcup_{d | n} \frac{n}{d}\cdot\ZZ^*_d \\
$$
and the sets under the union are disjoint each other.
**N.B.**
- $x \cdot S$ means that $x \cdot S \equiv \{ xy \mid y \in S \}$.
- This implies that $\sum_{d | n} \phi(d) = n$ where $\phi$: Euler's totient.
***Proof***
TODO
### Prop. 4
$$
\t{SR}(n) = \sum_{d | n} \t{SP}(d).
$$
***Proof***
$$
\Ba
\sum_{d | n} \t{SP}(d)
&= \sum_{d | n} \sum_{k \in \ZZ^*_d} \exp(i 2 \pi k/d) \\
&= \sum_{d | n} \sum_{k \in \ZZ^*_d} \exp\l(i 2 \pi (\frac{n}{d}k)/n\r) \\
&= \sum_{d | n} \sum_{k' \in \frac{n}{d}\ZZ^*_d} \exp\l(i 2 \pi k'/n\r) \\
&= \sum_{k' \in \ZZ_n} \exp\l(i 2 \pi k'/n\r)
~~(\because \t{Prop. 3}) \\
&= \t{SR}(n).
\Ea
$$
### Prop. 5
$\t{SP}(n)$ is multiplicative, i.e.
$$
\gcd(n, m) = 1 \implies \t{SP}(nm) = \t{SP}(n) \t{SP}(m).
$$
***Proof***
Given $\gcd(n, m) = 1$, we have,
$$
\Ba
\t{SP}(n) \t{SP}(m)
&= \sum_{k \in \ZZ^*_{n}} \exp(i 2 \pi k / n)
\sum_{l \in \ZZ^*_{m}} \exp(i 2 \pi l / m) \\
&= \sum_{(k, l) \in \ZZ^*_{n} \xx \ZZ^*_{m}}
\exp(i 2 \pi (k / n + l / m)) \\
&= \sum_{(k, l) \in \ZZ^*_{n} \xx \ZZ^*_{m}}
\exp(i 2 \pi (km + ln) / (n m)).
\Ea
$$
Thus, it suffices to show $f(l, m) \equiv k m + l n: \ZZ^*_{n} \xx \ZZ^*_{m} \to \ZZ^*_{nm}$
is well-defined and bijective.
For that, we show $\gcd(k m + l n, nm) = 1$ and $f$: injective.
(NOTE: the rest is easy...)
### Prop. 6
$$
\t{SR}(n) =
\Bc
1 & (\t{if}~ n = 1) \\
0 & (\t{otherwise}).
\Ec
$$
***Proof***
Writing $\alpha = 2 \pi / n$, we have,
$$
\Ba
(\exp(i 2 \pi / n) - 1) \t{SR}(n)
&= \sum_{k : 1...n} \exp(i 2 \pi k / n) - \sum_{k : 0...n-1} \exp(i 2 \pi k / n) \\
&= \exp(i 2 \pi n / n) - \exp(i \alpha 0 / n) \\
&= 0.
\Ea
$$
### Prop. 7
For $p$: prime,
- $\t{SP}(p) = -1$.
- $\t{SP}(p^e) = 0$ where $e >= 2$.
***Proof***
From Prop. 4 and Prop. 6, we have:
$$
0 = \t{SR}(p^e) = \t{SP}(1) + \t{SP}(p) + ... + \t{SP}(p^e).
$$
Thus, for $e = 1$, we have $\t{SP}(p) = - \t{SP}(1) = -1$.
For $e >= 2$, by induction, we have $\t{SP}(p^e) = 0$.
### Prop. 8
$$
\mu(n) = \t{SP}(n).
$$
***Proof***
This follows from multiplicativity and the coinciding base case $\mu(p^e) = \t{SP}(p^e)$.
### Prop. 9 (Mobius inversion formula)
$$
g(n) = \sum_{d | n} f(d)
\implies
f(n) = \sum_{d | n} \mu(d) g(\frac{n}{d}).
$$
***Proof***
$$
\Ba
\sum_{d | n} \mu(d) g(\frac{n}{d})
&= \sum_{d | n } \mu(d) \sum_{e | \frac{n}{d}} f(e) \\
&= \sum_{d | n ~\wedge~ d e | n} \mu(d) f(e) \\
&= \sum_{e | n} f(e) \sum_{d | \frac{n}{e}} \mu(d) \\
&= \sum_{e | n} f(e) ~ \delta(n/e, 1)
~~(\because \t{Prop. 4, 6, 8}) \\
&= \sum_{e | n} f(e) ~ \delta(e, n) \\
&= f(n).
\Ea
$$
## Slerp (Spherical linear interpolation)
***Notation***
- $p_0, p_1 \in \RR^{n}$,
- $|p_0| = |p_1| = 1$,
- $\iprod{p_0}{p_1} = \cos(\theta)$,
- $\sin(\theta) \neq 0$ i.e. $p_0$ and $p_1$ are not parallel.
### Def. 1
$$
p(t)
\equiv
\frac{1}{\sin(\theta)}
(
\sin((1 - t) \theta)p_0 +
\sin(t \theta)p_1
).
$$
**N.B.**
- Clearly $p(0) = p_0$ and $p(1) = p_1$.
- By writing $a \equiv (1 - t) \theta$ and $b \equiv t \theta$, we have:
$$
p(t) \equiv
\frac{\sin(a) p_0 + \sin(b) p_1}{\sin(a + b)}.
$$
- $
\del_t p(t)
=
\dfrac{\theta}{\sin(\theta)}
(- \cos((1 - t) \theta) p_0 + \cos(t \theta) p_1)
=
\theta
\dfrac{- \cos(a) p_0 + \cos(b) p_1 }{\sin(a + b)}.
$
- $
\del_t^2 p(t)
=
\dfrac{- \theta^2}{\sin(\theta)}
(\sin((1 - t) \theta) p_0 + \sin(t \theta) p_1)
=
- \theta^2 p(t).
$
### Prop. 1
$$
\iprod{p_0}{p(t)} = \cos(t \theta).
$$
***Proof***
Using $a, b$, we can write:
$$
\Ba
\iprod{p_0}{p(t)}
&= \frac{1}{\sin(a + b)}
(\sin(a) |p_0|^2 + \sin(b) \iprod{p_0}{p_1}) \\
&= \frac{\sin(a) + \sin(b) \cos(a + b)}{\sin(a + b)}.
\\
\cos(t \theta)
&= \cos(b).
\Ea
$$
Thus, we see:
$$
\Ba
\iprod{p_0}{p(t)} = \cos(t \theta)
&\iff
\sin(a) + \sin(b) \cos(a + b) =
\sin(a + b) \cos(b) \\
&\iff
\cos(a + b) \sin(b) - \cos(a + b) \sin(b) = \sin(a).
\Ea
$$
### Prop. 2
$$
|p(t)| = 1.
$$
***Proof***
Noting that $\del_t |p(t)|^2 = 2 \iprod{p(t)}{\del_t p(t)}$ and $|p(0)| = |p_0| = 1$,
it suffices to show that $\iprod{p(t)}{\del_t p(t)} = 0$. Indeed, we see
$$
\Ba
\iprod{p(t)}{\del_t p(t)}
&=
\frac{\theta}{\sin(a + b)^2}
\iprod{\sin(a) p_0 + \sin(b) p_1}{-\cos(a) p_0 + \cos(b) p_1} \\
\Ea
$$
Expanding the last term:
$$
\Ba
&\iprod{\sin(a) p_0 + \sin(b) p_1}{- \cos(a) p_0 + \cos(b) p_1} \\
&=
- \sin(a) \cos(a) + \sin(b) \cos(b)
+ (\sin(a) \cos(b) - \cos(a) \sin(b)) \cos(a + b) \\
&=
\frac{1}{2} \l(- \sin(2a) + \sin(2b) \r)
+ \sin(a - b) \cos(a + b) \\
&=
\frac{1}{2} \l(- \sin((a + b) + (a - b)) + \sin((a + b) - (a - b)) \r)
+ \sin(a - b) \cos(a + b) \\
&= 0.
\Ea
$$
### Prop. 3
$$
|\del_t p(t)| = |\theta|.
$$
***Proof***
We have $|\del_t p(t)|$: constant since:
$$
\del_t |\del_t p(t)|^2
= 2 \iprod{\del_t p(t)}{ \del_t^2 p(t) }
= - 2 \theta^2 \iprod{\del_t p(t)}{p(t)} = 0.
$$
Thus, it suffices to show $|\del_t p(0)| = |\theta|$. Indeed, we have:
$$
|\del_t p(0)|
= |\theta|
\l| \frac{- \cos(\theta) ~ p_0 + p_1}{\sin(\theta)} \r| = 0.
$$
The last equation follows from:
$$
|- \cos(\theta) ~ p_0 + p_1|^2
= \cos^2(\theta) + 1 - 2 \cos(\theta) \iprod{p_0}{p_1}
= 1 - \cos^2(\theta) = \sin^2(\theta).
$$
## Conjugate gradient descent
***Notation***
- $A \in \RR^{n \xx n}$: positive definite
- $x, b \in \RR^n$
### Def. 1
- $x_0 \equiv x$,
- $r_0 \equiv A x_0 - b$,
- $p_0 \equiv r_0$,
- $x_{n + 1} \equiv x_n + \alpha_n p_n$,
- $r_{n + 1} \equiv A x_{n + 1} - b = (A x_n - b) + \alpha_n A p_n = r_n + \alpha_n A p_n$,
- $p_{n + 1} \equiv r_{n + 1} + \beta_n p_n$,
where
- $\alpha_n \equiv - \dfrac{\iprod{p_n}{r_n}}{\iprod{p_n}{A p_n}}$,
- $\beta_n \equiv - \dfrac{\iprod{r_{n+1}}{A p_n}}{\iprod{p_n}{A p_n}}$,
running until we find $r_n = 0$.
**N.B.**
- Writing $f(x) \equiv \dfrac{1}{2} \iprod{x}{A x} - \iprod{x}{b}$,
we have $\del_x f(x) = A x - b$.
- $\del_\alpha f(x + \alpha p)
= \iprod{\del_x f(x + \alpha p)}{p}
= \iprod{A(x + \alpha p) - b}{p}$.
- $\alpha_n$ is taken so that
$\iprod{r_{n + 1}}{p_n} = \del_{\alpha_n} f(x_n + \alpha_n p_n) = 0$
(i.e. line-search minimum).
- $\beta_n$ is taken so that $\iprod{p_{n + 1}}{A p_n} = 0$.
- $|p_{n + 1}|^2 = |r_{n + 1}|^2 + |p_n|^2$ thus
$r_{n + 1} \ne 0 \implies p_{n + 1} \ne 0$.
Therefore, $\alpha_n$ and $\beta_n$ are well-defined as long as $r_n \ne 0$.
- Furthermore
$\iprod{p_n}{r_n} = \iprod{r_n + \beta_{n - 1} p_{n - 1}}{r_n} = \iprod{r_n}{r_n}$,
thus $r_n \ne 0 \implies \alpha_n \ne 0$.
### Prop. 2
- $\iprod{p_i}{A p_j} = \iprod{r_i}{r_j} = 0$ for $i \ne j$.
***Proof***
Prove by induction.
We show $\iprod{r_{n + 1}}{r_k} = 0$:
$$
\Ba
\iprod{r_{n + 1}}{r_n}
&= \iprod{r_{n + 1}}{p_n - \beta p_{n - 1}} \\
&= - \beta \iprod{r_{n + 1}}{p_{n - 1}} \\
&= - \beta \iprod{r_n + \alpha A p_n}{p_{n - 1}} \\
&= - \alpha \beta \iprod{A p_n}{p_{n - 1}} \\
&= 0 ~~(\because \t{IH}), \\
\Ea
$$
and for $k < n$,
$$
\Ba
\iprod{r_{n + 1}}{r_k}
&= \iprod{r_n + \alpha A p_n}{r_k} \\
&= \alpha \iprod{A p_n}{r_k} ~~(\because \t{IH}) \\
&= \alpha \iprod{A p_n}{p_k - \beta p_{k - 1}} \\
&= 0 ~~(\because \t{IH}).
\Ea
$$
Next, we show $\iprod{p_{n + 1}}{A p_k} = 0$.
For $k = n$, it was shown above and for $k \lt n$,
$$
\Ba
\iprod{p_{n + 1}}{A p_k}
&= \iprod{r_{n + 1} + \beta p_n}{A p_k} \\
&= \iprod{r_{n + 1}}{A p_k} ~~(\because \t{IH})\\
&= \iprod{r_{n + 1}}{\frac{r_{k + 1} - r_k}{\alpha}} \\
&= 0 ~~(\because \t{IH}).
\Ea
$$
### Prop. 3
- $\alpha_n
\equiv - \dfrac{\iprod{p_n}{r_n}}{\iprod{p_n}{A p_n}}
= - \dfrac{\iprod{r_n}{r_n}}{\iprod{p_n}{A p_n}}$,
- $\beta_n
\equiv - \dfrac{\iprod{p_n}{A r_{n+1}}}{\iprod{p_n}{A p_n}}
= \dfrac{\iprod{r_{n+1}}{r_{n+1}}}{\iprod{r_n}{r_n}}$.
***Proof***
Noting that
- $r_{n + 1} = r_n + \alpha_{n} p_n$,
- $p_n = r_n + \beta_{n - 1} p_{n - 1}$,
- $\iprod{r_n}{p_{n - 1}} = \iprod{r_{n + 1}}{p_n} = 0$,
- $\iprod{r_{n + 1}}{r_n} = 0$,
- $\iprod{p_{n - 1}}{A p_n} = 0$,
we see:
- $\iprod{r_n}{p_n} = \iprod{r_n}{r_n + \beta_{n - 1} p_{n - 1}} = \iprod{r_n}{r_n}$,
- $\iprod{r_{n + 1}}{A p_n}
= \iprod{r_{n + 1}}{\dfrac{r_{n + 1} - r_n}{\alpha_n}}
= \dfrac{\iprod{r_{n + 1}}{r_{n + 1}}}{\alpha_n}$,
- $\iprod{p_n}{A p_n}
= \iprod{p_n}{\dfrac{r_{n + 1} - r_n}{\alpha_n}}
= - \dfrac{\iprod{p_n}{r_n}}{\alpha_n}
= - \dfrac{\iprod{r_n}{r_n}}{\alpha_n}$.
---
## Singular value decomposition
### Prop. 1 (Polar decomposition)
- $A \in \CC^{n \xx n}$: invertible,
- $A^\dagger A$: positive definite,
- $P \equiv (A^\dagger A)^{1/2}$: positive definite,
- $U \equiv A \l((A^\dagger A)^{1/2}\r)^{-1}$: unitary,
- $A = U P$: polar decomposition.
### Prop. 2 (Existence of SVD)
Given,
- $A \in \CC^{n \xx m}$,
There exist $U \in \CC^{n \xx n}, V \in \CC^{m \xx m}$ : unitary
and $D \in \RR_{\ge 0}^{n \xx m}$ : diagonal s.t.
$$
A = U D V^\dagger.
$$
***Proof***
- Quick proof for the special case of $A \in \CC^{n \xx n}$: invertible.
- By polar decomposition, $A = U P$.
- By spectral decomposition, $P = V D V^\dagger$.
- Thus, $A = (U V) D V^\dagger$.
- Proof for general case (i.e. non-square or non-invertible).
By spectral decomposition, $A^\dagger A = P W P^\dagger \in \CC^{m \xx m}$ where
$P$ is unitary and $W$ is non-negative diagonal.
Next, we find permutation $S \in \CC^{m \xx m}$ s.t. $A P S$ has
column vectors with decreasing norm (i.e. sorting columns of $A P$ by their norm).
Next, by QR decomposition $A P S = Q R$
where $Q \in \CC^{n \xx n}$ and $R \in \CC^{n \xx m}$.
Here notice that $R$ has decreasing norm column vectors since $Q$ is unitary.
Also, we note:
$$
R^\dagger R
= (Q^\dagger A P S)^\dagger Q^\dagger A P S
= S^\dagger P^\dagger A^\dagger Q Q^\dagger A P S
= S^\dagger W S: \t{diagonal}.
$$
Finally, by **Lemma. 3** below, we know $R$ is actually diagonal and
thus $A = Q R (P S)^\dagger$ is singular value decomposition.
### Lemma. 3
Given,
- $R$: upper triangle with column vectors of decreasing norm
- $R^T R$: diagonal
Then, $R$: diagonal.
***Proof***
First, we write:
$$
R =
\Bmp
X & Y \\
0 & Z
\Emp
$$
where $X$: invertible and $Z_{1, 1} = 0$,.
Then,
$$
R^T R =
\Bmp
X^T X & X^T Y \\
Y^T X & Y^T Y + Z^T Z.
\Emp
$$
Since $X$: invertible and $R^T R$: diagonal, we find
- $Y = 0$,
- $X^T X$: diagonal, thus $X^T$: upper triangle, thus $X$: diagonal
- $
R =
\Bmp
X & 0 \\
0 & Z
\Emp
$ and $Z_{1, 1} = 0$ thus $Z = 0$ since decreasing norm.
### Prop. 4 (Orthogonal Procrustes problem)
Define
- $\iprod{A}{B} \equiv \t{tr}(A^T B)$,
- $|A| = \sqrt{\iprod{A}{A}}$.
Given
- $A, B \in \RR^{n \xx m}$,
- $B A^T = U D V^T \in \RR^{n \xx n}$:
singular value decomposition with $D$ decreasing diagonal,
$$
\min\limits_{P : \t{SO}(3)}|A - P B|.
$$
$$
\Ba
|A - P B|^2
&= \iprod{A - P B}{A - P B} \\
&= |A|^2 + |P B|^2 - 2 \iprod{A}{P B} \\
&= |A|^2 + |B|^2 - 2 \iprod{A}{P B} \\
\iprod{A}{P B}
&= \t{tr}(A^T P B)
= \t{tr}(P B A^T)
= \t{tr}(P U D V^T) \\
&= \t{tr}((V^T P U) D).
\Ea
$$
By taking
$$
P =
V
\Bmp
I & 0 \\
0 & \t{det}(V U^T)
\Emp
U^T,
$$
we have $\iprod{A}{PB} = (\sum_{i < n} D_{i, i}) + \t{det}(V U^T) D_{n, n}$.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment