Consider the example of trying to estimate exponential growth rate $\gamma$ from some series of measurements with units vs. time $y(t)$ using the model $y = y_0 \exp{(\gamma t)}$. Solving the model function gives $\gamma = \frac{d}{dt}\ln\frac{y}{y_0}$. Now, it's tempting to use the logarithm quotient rule $\ln{\frac{a}{b}} = \ln{a} - \ln{b}$ to get $\gamma = \frac{d}{dt}\left(\ln{y} - \ln{y_0}\right)$ and then drop $\frac{d}{dt}\ln{y_0}$ because $y_0$ doesn't vary with time, but now we're left with $\ln{y}$. If one writes a script to do this calculation, it does work and produces a result which is consistent with fitting the data to the exponential growth model (shown in figure 1, with the additional complication that $y-\mathrm{offset}$ takes the place of $y$).

So, what was wrong with using the log quotient rule, and why did it seem to work, anyway?

Example fit

Related to What is the logarithm of a kilometer? Is it a dimensionless number? , but different because of the quotient rule and example calculation.

  • $\begingroup$ Not exactly clear what your question is, but note that (d/dt)ln(y) =(dy/dt)/y = y'/y or the relative rate of change in y. This quantity is independent of the units that you use to describe y. $\endgroup$ – Samuel Weir Apr 11 '18 at 17:07
  • $\begingroup$ math.stackexchange.com/questions/238390/… Does this help you? $\endgroup$ – Yuzuriha Inori Apr 11 '18 at 17:10
  • $\begingroup$ @SamuelWeir Yes, I know it can be reduced to y'/y and that's why I got away with it. But the log quotient rule is ln(a/b) = ln(a) - ln(b). a and b can both have units, but a/b will be unitless, making it a legal argument to ln. So, the quotient rule is invalid if a and b have units. $\endgroup$ – EL_DON Apr 11 '18 at 18:34
  • $\begingroup$ @YuzurihaInori No, that doesn't help. I understand that ln(1 meter) is illegal. The funny business is ln(2 meters / 1 meter) is legal, but applying the log quotient rule to it creates a problem. $\endgroup$ – EL_DON Apr 11 '18 at 18:35
  • $\begingroup$ $\ln(a/b)=\ln(ka/kb)=\ln(ka)-\ln(kb)$. Now choose $k$ to have units inverse of $a$ and $b$. $\endgroup$ – The Photon Apr 11 '18 at 19:17

The logarithm of any unit is unphysical, so none of the logarithm rules can be used to produce a result that has units in the argument. The specific example works anyway because the problem can be simplified to eliminate the logarithm:

$\gamma = \frac{d}{dt}\ln\frac{y}{y_0} = \frac{1}{y/y_0} \frac{d}{dt}\left(\frac{y}{y_0}\right) = \frac{1}{y}\frac{dy}{dt}$

This gives a perfectly legal equation with no use of the logarithm rules. Also notice that the units of $y$ cancelled out completely, so if the log quotient rule had been used to get rid of $y_0$, the units of $ln(\mathrm{mm})$ would've cancelled out as well; the unphysical log(unit) part only exists in intermediate steps.

If a constant offset is involved as in the example figure, then it's $\gamma = \frac{d}{dt}\ln\frac{y-y_{off}}{y_0} = \frac{1}{(y-y_{off})/y_0} \frac{d}{dt}\left(\frac{y-y_{off}}{y_0}\right) = \frac{1}{y-y_{off}}\frac{dy}{dt}$, so the y-axis label of the bottom plot should be updated to something like y'/(y-off).

For more general problems, there needs to be some way of removing the units, such as by dividing out a standard unit.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.