livetrax/doc/tempo.tex

% vim: tw=80 spl=en
% Copyright (C) 2018  Julien "_FrnchFrgg_" RIVAUD
%
% This program is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation, either version 3 of the License, or
% (at your option) any later version.
%
% This program is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
% GNU General Public License for more details.
%
% You should have received a copy of the GNU General Public License
% along with this program.  If not, see <http://www.gnu.org/licenses/>.
\documentclass[10pt]{article}


%\usepackage{fontspec}
%\usepackage{polyglossia}
%\setmainlanguage{english}
\usepackage[english]{babel}

%\usepackage{lualatex-math}
\usepackage{amsmath}
%\usepackage[math-style=ISO,bold-style=ISO]{unicode-math}

\usepackage{geometry}
\geometry{reset,margin=1.5cm,left=6cm,includefoot,marginparwidth=4.5cm}

\usepackage{cellprops}

\begin{document}



\def\define#1{\textbf{#1}}
\def\d{\mathrm{d}}

\section{A natural definition for tempo ramping}

\subsection{Some definitions}

The \define{time}, often denoted $t$, is the duration in some multiple unit of
seconds (samples, superclock, etc).

We define $b$ to be the \define{beats function}: $b$~is the number of beats that
occurred since $t = 0$ until the current time. It is assumed to be continuous and
even derivable rather than being a staircase function: $b$~will thus value~$4.5$
when exactly in the middle of beat~$4$, halfway\footnote{Not necessarily halfway
in time units, though.} between pulsations $4$~and~$5$.

The \define{tempo function}, or simply tempo, is the number~$T$ of pulsations
per unit of time. Thus $T = \dfrac{\d b}{\d t}$.

\def\DT{\Delta T}
\def\Dt{\Delta t}
\def\Db{\Delta b}

A \define{tempo ramp} is a definition of $T$ for $0 \le t \le \Dt$ such that
$T_{t=0} = T_0$ and $T_{t=\Dt} = T_{\text{end}} = T_0 + \DT$ where $\Dt$, $T_0$
and $T_{\text{end}}$~or~$\DT$ are given. Most of the times they are set by the
user, but $\Dt$ can be sometimes defined in beats, that is be the unique~$\Dt$
such that $b_{|t=\Dt} = \Db$.

\subsection{Linear tempo ramping}

The simplest definition that comes to mind is a tempo ramp where the tempo
increases from $T_0$~to~$T_0 + \DT$ \emph{linearly} with time.

Here $T = T_0 + \dfrac{\DT}{\Dt}\times t$, and thus
$b = \dfrac{\DT}{2\Dt}\times t^2 + T_0 \times t$, assuming $b_{|t=0} = 0$.

If $\Dt$~is defined implicitly from a duration~$\Db$ in beats, then
\[ \Db = \dfrac{\DT}{2\Dt}\times \Dt^2 + T_0 \Dt
  = \left( \frac{1}{2}\DT + T_0 \right) \times\Dt \]
thus
\[
    \Dt = \dfrac{2\Db}{\DT + 2T_0}
\]

\subsection{Exponential tempo ramping}

\def\e{\mathrm{e}}

Humans are not very good at keeping track precisely of absolute time for long
durations, and are better at comparing short durations, that is maintain a
reasonably stable time span between successive pulsations, and have a precise
sense of elapsed time \emph{in terms of these pulsations}.

In particular, when changing the tempo, it is less far-fetched to imagine a
human feeling the tempo increase as a function of elapsed \emph{beats} rather
than absolute time in seconds which is a pulsation that is very hard to maintain
constant and even sense when competing with the (increasing) musical pulsation.

In other words, a definition more suitable than that of linear ramping would be
that the tempo increases \emph{linearly} with the value of~$b$. We thus want
\[ T = T_0 + \dfrac{\DT}{\Db}\times b
    \text{, that is }
\dfrac{\d b}{\d t} - \dfrac{\DT}{\Db}\times b = T_0 \]

The solutions of this order 1 linear differential equation are\footnote{The
exponential is the solution of the associated homogeneous equation and the
second term is the constant solution of the equation.}
$b = A \e^{\frac{\DT}{\Db}t} - \dfrac{\Db T_0}{\DT}$, and
since we want $b_{|t=0} = 0$, we get $A = \dfrac{\Db T_0}{\DT}$ that is
\[ b = \dfrac{\Db T_0}{\DT} (\e^{\frac{\DT}{\Db}t} - 1) \]

Denoting $\omega = \dfrac{\DT}{\Db}$ we find
$b = \frac{T_0}{\omega} (\e^{\omega t} - 1)$.
(Note: $\omega$ is called $c$ in Nick's text, but I like $\omega$ better since
it is a frequency).

The \emph{tempo} is $T = \dfrac{\d b}{\d t}
= T_0 \e^{\frac{\DT}{\Db}t} = T_0 \e^{\omega t}$, and this is where the name
exponential tempo ramp comes from.

Recall that $\Db$, $\DT$ and $T_0$ are parameters of the ramp, while $t$~is the
arbitrary time within the ramp at which we want to know $b$~and~$T$.

Let us compute the reciprocal of $b$:
\[
    b = \dfrac{\Db T_0}{\DT} (\e^{\frac{\DT}{\Db}t} - 1)
    \iff
    \dfrac{\DT}{\Db T_0} b = \e^{\frac{\DT}{\Db}t} - 1
    \iff
    1 + \dfrac{\DT}{\Db T_0} b = \e^{\frac{\DT}{\Db}t}
\]
\[
    \iff
    \log \left( 1 + \dfrac{\DT}{\Db T_0} b \right) = \frac{\DT}{\Db}t
    \iff
    t = \frac{\Db}{\DT} \log \left( 1 + \dfrac{\DT}{\Db T_0} b \right)
\]

If $\Dt$~is defined directly instead of implicitly from a duration~$\Db$ in
beats, then we have to find out the value of~$\Db$. At $t=\Dt$, we have
$b = \Db$ thus
\[
    \Dt = \frac{\Db}{\DT} \log \left( 1 + \dfrac{\DT}{\Db T_0} \Db \right)
    \iff
    \Dt = \frac{\Db}{\DT} \log \left( 1 + \dfrac{\DT}{T_0} \right)
    \iff
    \Db = \frac{\DT\Dt}{\log \left( 1 + \dfrac{\DT}{T_0} \right) }
\]

A more useful expression is
\[
    \omega = \frac{\DT}{\Db}
    = \frac{1}{\Dt}\log\left(1+ \frac{\DT}{T_0}\right)
\]
which gives
\[
    b = \frac{\Dt T_0}{\log\left(1+ \frac{\DT}{T_0}\right)}
    \e^{\frac{t}{\Dt}\log\left(1+ \frac{\DT}{T_0}\right)}
    = \frac{\Dt T_0}{\log\left(1+ \frac{\DT}{T_0}\right)}
    \left(1+ \frac{\DT}{T_0}\right)^{\frac{t}{\Dt}}
\]
and
\[
    T
    = T_0 \e^{\frac{t}{\Dt}\log\left(1+ \frac{\DT}{T_0}\right)}
    = T_0 \left(1+ \frac{\DT}{T_0}\right)^{\frac{t}{\Dt}}
\]

We at last compute the reciprocals:
\[
    b = \frac{\Dt T_0}{\log\left(1+ \frac{\DT}{T_0}\right)}
        \e^{\frac{t}{\Dt}\log\left(1+ \frac{\DT}{T_0}\right)}
    \iff
    \frac{\log\left(1+ \frac{\DT}{T_0}\right)}{\Dt T_0} b
    = \e^{\frac{t}{\Dt}\log\left(1+ \frac{\DT}{T_0}\right)}
\]
\[
    \iff
    \log\left( \frac{\log\left(1+ \frac{\DT}{T_0}\right)}{\Dt T_0} b \right)
    = \frac{t}{\Dt}\log\left(1+ \frac{\DT}{T_0}\right)
    \iff
    t = \frac{
    \Dt
    \log\left( \frac{\log\left(1+ \frac{\DT}{T_0}\right)}{\Dt T_0} b \right)
    }{
    \log\left(1+ \frac{\DT}{T_0}\right)
    }
\]

\subsection{Summary with $T$ as main tempo representation}

The following table gathers all results. For code factorization we might want
to only consider the red formulas, that is compute (and maybe cache) $\omega$
beforehand. Note that in the formulas, $\omega$ is \emph{always} the inverse of
a time, never in per-beats.

\[\everymath{\displaystyle}
    \cellprops{ td:nth-child(1) { math-mode: text; } }
    \begin{array}{lccc}
    \hline
    When you know & \Dt & \Db & \omega \\
    \hline
    Compute $\omega$ &
        \color{red}
        \frac{1}{\Dt}\log\left(1+ \frac{\DT}{T_0}\right)
        &
        \color{red}
        \frac{\DT}{\Db}
        &
        \\
    Compute $b$ from $t$ &
        \frac{\Dt T_0}{\log\left(1+ \frac{\DT}{T_0}\right)}
        \left(1+ \frac{\DT}{T_0}\right)^{\frac{t}{\Dt}}
        &
        \frac{\Db T_0}{\DT} (\e^{\frac{\DT}{\Db}t} - 1)
        &
        \color{red}
        \frac{T_0}{\omega} (\e^{\omega t} - 1)
        \\
    Compute $T$ from $t$ &
        T_0 \left(1+ \frac{\DT}{T_0}\right)^{\frac{t}{\Dt}}
        &
        T_0 \e^{\frac{\DT}{\Db}t}
        &
        \color{red}
        T_0 \e^{\omega t}
        \\
    Compute $t$ from $b$ &
        \frac{
            \Dt
            \log\left( \frac{\log\left(1+ \frac{\DT}{T_0}\right)}
                        {\Dt T_0} b \right)
        }{
            \log\left(1+ \frac{\DT}{T_0}\right)
        }
        &
        \frac{\Db}{\DT} \log \left( 1 + \frac{\DT}{\Db T_0} b \right)
        &
        \color{red}
        \frac{1}{\omega} \log \left( 1 + \frac{\omega b}{T_0} \right)
        \\
    Compute $T$ from $b$ &
        T_0 + \frac{1}{\Dt}\log\left(1+ \frac{\DT}{T_0}\right) \times b
        &
        T_0 + \frac{\DT}{\Db}\times b
        &
        \color{red}
        T_0 + \omega b
        \\
    \hline
\end{array}\]

Note that there are often library functions to compute $\log(1+x)$ directly
from~$x$ more precise than computing $1+x$ then its logarithm. The main reason
is that when adding 1to some very small number you loose a lot of precision
because the exponent is now tailored to the representation of~$1$, and that
the Taylor-McLaurin series for $\log(1+x)$ is a very efficient mean to compute
an approximate value for it.

\subsection{Using $S = \frac{1}{T}$ to represent tempo}

\def\DS{\Delta S}

\everymath{\displaystyle}

Since the base time unit in ardour is not seconds but far smaller than
that\footnote{It was samples in previous versions of Ardour, and is even more
precise in later versions.}, any value of $T$~or~$\DT$ will be very small. A
better value to store, that can easily be rounded to integer without a huge
loss of precision, is $S = 1/T$.

Instead of having $T_0$~and~$T_1 = T_0 + \DT$, we thus assume that when
computing the formulas we have access to $S_0$~and~$S_1$.

Then from $\Dt$,
$\omega = \frac{1}{\Dt}\log\left(1 + \frac{\DT}{T_0}\right)
 = \frac{1}{\Dt}\log\left(\frac{T_1}{T_0}\right)
 = \frac{1}{\Dt}\log\left(\frac{S_0}{S_1}\right)$.

From beats, $\omega = \frac{\DT}{\Db} = \frac{S_0 - S_1}{S_0S_1\Db}$ (or just
use $\omega = \left( \frac{1}{S_1} - \frac{1}{S_0} \right) / \Db$).

To compute $b$~from~$t$ we can use
$b = \frac{\e^{\omega t}-1}{S_0 \omega}$.

And for $S$~from~$t$ we have
$S = \frac{S_0}{\e^{\omega t}} = S_0\e^{-\omega t}$.

As for $t$~from~$b$ the formula becomes
$t = \frac{1}{\omega} \log \left( 1 + S_0 \omega b \right)$.

Lastly, for $S$~from~$b$ we can see:
\[ S = 1/T = \frac{1}{1/S_0 + \omega b} = \frac{S_0}{1+S_0\omega b}\]

Maybe we can also to compute and cache
$\omega' = S_0 \omega = \frac{S_0}{\Dt}\log\left(\frac{S_0}{S_1}\right)
    = \frac{S_0 - S_1}{S_1\Db}$
Which is in « per-beats » and occurs often in the formulas (but not
exclusively). Note that the fact that $\omega$~or~$\omega'$ is relevant is
\emph{not} depending on whether we compute from time or from beats, or even if
the ramp length is defined in time or in beats.

\end{document}