Sponsors



« This just in: BBC editors didn't pass calculus | Main | The Army's contribution to New Urbanism »

Speaking of the calculus...

In a previous post, I derided Ray Kurzweil for showing a straight line on a log-log graph and calling that an "exponential trend." As I said in that post:

Click to see larger image
...the math savvy should recognize that a straight line on a log-log graph is not indicative of exponential growth but instead illustrates a power law relationship. So does that chart contradict the exponential developments that are the cornerstone of his arguments?

Well, as it turns out, the chart does not contradict the exponential trend. You might ask, how can you get an exponential trend to look like a straight line on a log-log graph? The answer is very subtle, actually. Let's take a simple model of exponential growth:

n = n0 exp(kt)

Where n is something like number of inventions, and t is time (the other two parameters, n0 and k, are there just for the sake of generalization). The key to making this a straight line in log-log space is to plot the number of inventions vs. the time until next invention. With this subtle change, the plot becomes linear on a log-log plot. The time until next invention, τ(t), is fairly simple to obtain:

τ(t) = dt/dn = (n0 k)-1 exp(-kt)

When you plot τ(t) vs. n(t), you get a straight line. So I was wrong, and Kurzweil was right. I still maintain that the way Kurzweil plots this is very deceptive. The reason that he didn't plot it on a normal semilog plot (like all the others he normally shows), is that the scatter would look huge. By compressing both axes into log scale, it makes it look as if the correlation is better than it really is; as far as I'm concerned, this is a rather big no-no for serious academics. I think many peer-reviewers would recoil upon seeing data plotted in such a way.

Just my two cents.



Comments

By compressing both axes into log scale, it makes it look as if the correlation is better than it really is; as far as I'm concerned, this is a rather big no-no for serious academics. I think many peer-reviewers would recoil upon seeing data plotted in such a way.
Excuse my belated boldness but this is completely untrue. Many natural and technical processes follow an exponential course - take bacteria growth or nuclear chain reactions just as examples. Many technical diagrams are scaled logarithmically, and a correlation which can be seen in such a form is still very much a correlation ;) And if you browse those obscure and boring some-hundred pages zines with scientific publications you'll quickly see there's an abundance of these lg-scaled diagrams. I agree though it creates a distorted visual impression for folks who skipped exponential functions in school ;-P

Response:

I'm going to ignore your condescending tone and address your point directly. I have no issues with logarithms. My issue is the data itself. For instance, look at the right-hand side of the curve, and notice the spread on the y-axis. Now look on the left-hand side of the curve, and notice its spread. On the left, the spread is +/- 10^9 on both axes! Perhaps you'd be willing to tell me what the correlation coefficient is for such data?

P.S. Please Pubmed me to confirm that I've published in one or two of those "obscure and boring some-hundred pages zines with scientific publications."

The forum is a brighter place thanks to your posts. Tnhaks!

Really impressed! Everything is very open and very clear explanation of issues. It contains truly information. Your website is very useful. Thanks for sharing. Looking forward to more! Great visualization and unique informativ blog. I had a great time reading this.

lly shows), is that the scatter would look huge. By compressing both axes into log scale, it makes it lo

The forum is a brighter place thanks to your posts. Tnhaks!

Post a comment


Type the characters you see in the picture above.