You were taught to add. Almost everything that matters — money, populations, infections, cell counts, sound, light — multiplies. These are two different arithmetics, and moving between them without noticing is the single most reliable way to be wrong about growth. Part One builds the multiplicative intuitions from scratch and discovers, from the inside, why fluctuation costs you. Part Two shows that the same tax is being levied all over the natural world.
long-run growth ≈ mean return − ½ σ²
The rate you experience is not the average of the returns. Every point of variance is charged against your growth — and by the end of Part One you will have derived this yourself, twice.
Master fluctuation
40%
One knob sets the amount of fluctuation in every model on the page, both halves. At zero, every gap closes and all thirty averages agree.
Part One
Two arithmetics, and the gap between them
Nothing here is about finance, and none of it needs any. We start with two rulers and a coin, and we do not write down a formula until you have already felt what it says. The nine panels below are meant to be played with in order: each one breaks an additive intuition that the next one repairs.
Field guide: which world am I in?
Most errors about growth are a single mistake — reaching for the left-hand column while standing in the right-hand one. Every row is a place where the two arithmetics quietly disagree.
Additive world
Multiplicative world
combine by
adding: x + y
multiplying: x · y
size of a change
the difference, x − y
the ratio, x / y
the honest average
arithmetic mean
geometric mean — the exponential of the mean log
a symmetric move
+10 then −10 puts you back
×1.1 then ×0.9 does not; only log moves are symmetric
many small shocks
a normal distribution
a lognormal one — normal in the logs, skewed in the levels
the typical case
mean = median; the average is a person
median < mean; the average is nobody, and you live at the median
what to maximise
expected value
expected log value — the growth rate
natural graph paper
linear axes
log axes, where growth rate is a slope
zero
harmless, just a number
absorbing — one multiplication by zero ends the story
fluctuation
averages out, costs nothing
never averages out; costs ½σ² per step, forever
Five quick ones — which world is this?
A fund reports returns of +60%, −40%, +60%, −40%. Its brochure says the average annual return is +10%. What did an investor actually earn?
They lost 7.8%. Each pair multiplies to 1.6 × 0.6 = 0.96, so four years give 0.96² = 0.9216 — about −2.0% a year, not +10%. The brochure averaged in the additive world; the investor lived in the multiplicative one. The 12-point gap is almost exactly ½σ².
A coin flip either doubles your net worth or halves it, 50/50. Expected value per flip: +25%. How many times should you take it?
Never. The growth rate is √(2 × 0.5) − 1 = 0% at best, and any real bet has costs. The +25% is an average across parallel universes; in the one you inhabit, wealth just random-walks in log space forever. This is panel ii-4.
Is the mean income higher than the income of the median earner — and if so, why?
Yes, always, for a lognormal distribution, and incomes are roughly lognormal because they compound multiplicatively (raises are percentages). The mean sits at exp(M + S²/2), the median at exp(M). The gap is ½S² — the same term again, this time as inequality rather than drag.
You drive half the distance at 20 mph and half at 60 mph. Average speed?
30 mph, not 40 — the harmonic mean, because time (not speed) is what adds. This is a third mean for a third kind of adding, and it is exactly why fuel economy must be quoted per distance, not per volume. See Fuel economy and the wrong mean in Part Two.
Two machines each measure 60 dB. Switch both on. What does the meter read?
63 dB — energy adds, decibels do not. Doubling the power is always +3 dB, wherever you start. Your ear is already a log instrument; the scale was built to match it. See Averaging in decibels in Part Two.
Part Two
The curvature tax
You have now seen the tax once, in money. Here it is everywhere else — every time nature averages over a fluctuation through a response that bends.
The −½σ² you derived in Part One was not a fact about wealth. It was a fact about the logarithm — a function whose second derivative at 1 happens to be −1. Swap in any other bent response and the same theorem pays out a different number, with the curvature of that response setting the rate. Below, twenty sciences are quietly running the same calculation. Drag the master knob and watch them come apart at once.
E[ f(X) ] − f( E[X] ) ≈ ½ · f″(μ) · σ²
f″ < 0 (∩ concave) → the gap is a loss — drag · f″ > 0 (∪ convex) → the gap is a gain — anti-drag · f = log, f″(1) = −1 → −½σ²