Linearizing a Gaussian

EIA Field production of crude in the US, logistic (Hubbert) fit based only on 1958-2005 data, and Gaussian fit (quadratic fit to log of all the data). Source: EIA for the data.

I didn't think to make this picture the other day when writing about Predicting US Production with Gaussians. It seems to explain a lot.

There seem to be two salient points:
  • A Gaussian turn-on explains why the data lie above the linear fit early on in the life of the production history.
  • The Gaussian extrapolates forward very close to the straight line. If a Gaussian is a better fit, we don't have to throw out the linearization technique for extrapolating. (Well, if the result holds true over a range of K, anyway).
(The second point was the one that I suddenly started wondering about at 5am while in bed. I had to run down to check. I'm going back to bed now.)
The gaussian seems to fit well at the beginning. Did you use the same equations computed in the log(P) vs time domain:
log(P)= -5.27e-4 x t^2 + 2.09 x t - 2065

What is the utility of PQ vs Q domain if it<s not used for regression?<br> I was wondering what is the link between the regression coeff and the Gaussian parameters:
y= ax^2 + bx + c
y= a(x + b/(2a))^2 + c - sign(a).b^2/(4|a|)

Consequently:

var= -1/(2a)
mean= -b/(2a)
URR= exp(c-sign(a).b^2/(4|a|)) x sqrt(2xPIxvar)

So, in this case, var= 949.0 and mean= 1982.9.

Yes, the fit was done in the logP vs t domain. I have a bit more of an exploration of the difference coming in a long post I didn't manage to finish last night, but hope to tonight.
Also BTW (since it's taking a little while creating an account at peakoil.com) I wanted to make a comment on your post over there adding random Gaussian noise and then doing kind of a bootstrap type error estimation. I think that approach will underestimate the size of the error bars because you're not taking account of the structure of the errors. I don't know if the residuals of the various models are just auto-correlated or actually systematic, but they are very clearly not iid random. You might want to plot the autocorrelation r^2 as a function of lag in the residual time series from your model fit. If the residual autocorrelations fall off exponentially with lag, IIRC, very roughly you can consider that you have one independent observation per lifetime of the falloff.
Thanks for the comment. You'right, the noise is clearly not white and should be characterized.

Since, I've performed a more rigourous bootstrapping analysis using the R software (I also put the code):

Bootstrapping Technique Applied to the Hubbert Linearization

I looked at the world production (BP data) but I will post results on the US production probably tonight.

Since there is no good mechanistic explanation why the logistics curve is appropriate, modelling it with the gaussian is equally valid in my books.  Both have the serious short coming that it is primarily curve fitting.  
If someone could come up with a theory for why one or the other of these curves is followed by oil production I would be very interested.  

This theory would then also be useful in explaining why the curve is symetrical and could be used to counter the arguments of the folks in the EIA who believe we will be able to increase production rates well past the midpoint.  

Hi nero.

I don't have a full theory, but:

. Population growth follows a logistic curve;

. World oil production per capita is flat since 1982 (the year when the linearization method starts working).

I think there must be a strong link between the two.

Are you saying that global baby production is proportional to global oil production.  Or is the production of funerals inversly proportional to oil production? Or both?
That kinf of ridiculous comments doesn't help. Please read previous post on population growth. If you have the time check on dieoff.org.
Well, the early days fit really seems to be important. The logistic simply can't do it.

Still I must remember that we are looking at data from at least two discovery cycles.

It would be interesting to see a gaussian fit on a single discovery cycle (for instance lower-48, without Alaska).

I was curious, what software are you using to do the analysis? R, S-plus?
Excel. However, I am kind of reaching the limit of that and started dragging out my rusty Mathematica skills the other day.

This is more or less the same type of plot that Deffeyes had in his 2005 book.

I too am uncomfortable with the fact that there isn't a solid theoretical underpinning for what type of curve to fit - there are too many variables, including population growth, changes in extraction technology, market forces and all of the rest.  Then again, it is hard to deny that this does fit the data quite well.  I suppose the next step would be to try the same sort of thing for the North Sea oil and see if we get similar results.  You could do the same thing for Prudhoe bay by itself I imagine, but both of these cases are essentially just one oilfield, so we aren't quite modelling the same thing.

Something else that would be interesting I suppose would be to plot the predicted depletion rates as a function of time.  The first couple of years after a peak, production won't go down much at all.  It could be 5-10 years after peak before we start to get into some of the steeper depletion, so it isn't just a matter of predicting the steepest depletion, but in having a guess as to how long it will be before we get there.

Fitting the North Sea Oil with just one Gaussian curve is quite impossible, you have a twin peak.

Using two logistic curves, one for each discovery cycle, the fit is nearly perfect. I don't know what you'd get fitting two gaussians in this case...


I just played with these numbers a little bit.  The main question I had was how do the depletion rates vary with time.

There is one oddity in the graph though - the 'peak' happens in 1983 or thereabouts...

For US production, 10 years from the peak (year 1993), the production was dropping at about 1.1%/year.

20 years from the peak (year 2003), it was dropping at about 2.1%/year.

30 years from the peak (year 2013), it ought to be dropping at about 3.2%/year.

40 years from the peak (year 2023), it ought to be dropping at about 4.2%/year.

50 years from the peak (year 2033), it ought to be dropping at about 5.2%/year.

My point here is that at least in this model, the steeper depletion (which is relatively modest compared to some of the worst case numbers) is something that you slowly ease into.  If we assume that the world peaked this year for example, production for the first 10 years or so is likely to be fairly flat with a fairly small decrease from year to year.  Once you get 20 or 30 years out, then you are in the thick of it - that is where you are forced to make larger changes on an annual basis.

No guarantee of course that world production will follow a nice gaussian though...

    ... there is no good mechanistic explanation
    why the logistics curve is appropriate ...

But there is: essentially, a logistics curve
says that the growth of [take your choice:]

  * number of trans-Atlantic voyages of
    discovery before 1700,
  * number of Hitchock films,
  * quantity of oil produced,

tends to increase based on previous actions, but
is held back by the increasing amount learned or
the increasing difficulty of finding and doing.

Mostly, people look at it as an `S' shaped
growth curve.  In biological models, the maximum
comes from the carrying capacity of the niche.
It is somewhat odd to see linearization.
Usually, graphs are for an `S' or a bell.

I don't know of an equal mechanical argument for
an oil-based Gaussian.  The mechanical argument
for a Gaussian is that a central value has
errors that are equal on both sides and that
occur less frequently the further away from the
central value.  Gauss invented the curve for
analysing astronomical observations.  He figured
that a celestial object was in a defined orbit,
but that astronomers made mistakes or lacked
good equipment, but did not try to sway their
observations one way or another.

A 15 year old book,

    The Rise and Fall of Infrastructures
    Dynamics of Evolution and Technological
    Change in Transport by Arnulf Grübler
    reprinted 1999, ISBN 3-7045-0135-2

gives examples of numbers of cars registered,
kilometers of roads blacktopped, and the like.
It uses logistic curves extensively.

Theodore Modis wrote a book on logistic curves
called "Predictions" (that is where I got the
choices listed above). That book was copyright in
1992, ISBN 0-617-75917-5

Modis quantified the uncertainties in
determining logistic curve fits, given just the
beginning of a curve.  This could be useful.

I can make a fair argument for some sort of "S" type curve but it is only so much hand waving.  My criticism was about the argument that the logistics equation ought to be used over some other equally apropriate curve.  

The logistics curve has a logical mechanical reason for applying to bacterial growth where in an unconstrained situation the rate of increase is proportional to the current bacteria population and the availability of the constraining resource.  

However for oil production it isn't the past cumulative historical production (Q=sum(P)) that determines the rate of production growth but the current size of the exploration industry (I). Here is an alternative simple model that has to my mind some more relevant significance to the terms.

dI/dt = (h(Qt-Q)-j)I,  
dP/dt= r(Qt-Q)I - nP,

where:
r is a parameter related to the exploratory success rate
n is the average infield decline rate
j is a depreciation factor
h is a parameter associated with the insentives to increase exploratory effort.

If we add some apropriate parameters this also forms a nice bell curve.  Is this a better model than the logistic curve?  I couldn't say, but it does have the advantage that it makes some mechanistic sense.

 

Nero is darn near close to what I advocate, as far as I can tell:

I(t) = Discoveries,  
dQ/dt = I(t) - n Q(t),
P(t) = a Q  

I go through a few more 1st-order transforms because you have to consider latencies corresponding to fallow periods, construction periods, and maturation periods. I have the math all worked out, have the source code to do the numerical integration, and it basically looks like this if you assume that the Discoveries curve follows a quadratic growth curve for the USA (peaking after 1930)

That blue curve comes out to a quadratic (i.e. peaked parabola) convolved with a 4-th order gamma curve (i.e. 4 exponentials of the same rate convolved sequentially). It accurately maps over a range 5 orders of magnitude.

http://mobjectivist.blogspot.com/2006/01/would-you-believe.html

This is completely off topic I realize, however, all the charts and graphs in the world will be worth less if this thing goes through.  I hope the link works.

http://www.globalresearch.ca/index.php?context=viewArticle&code=%20CH20060103&articleId=1714

This may be just so much saber rattleing, but if Iran calles their bluff, do you really think George Bush is going to back down.

If we knew where the crucial scientists, engineers and technicians slept at night we could bomb their homes with GPS guided weapons carried by F-117s. Afterwords we just deny any involvement and blame the explosions on terrorists.
Below a normal-quantiles plot useful in order to compare distributions and in particular tail deviations:


original image

The production is mainly gaussian but deviates a little bit at the beginning.

Nice!
I forgot to give the different parameters I used:

Gaussian:
log(P)= -6.6824e-004 x t^2 + 2.6406 x t - 2.6075e+003
mean= 1.9758e+003
variance= 748.23
URR= 220.5 Gb

Logistic:
K= 6.11%
URR= 221.6 Gb

The fit was performed using a robust fit technique (function robustfit in Matlab) with the data from 1936 to 2005.