Where this method comes from

Four centuries of physicists finding that one constraint, correctly identified, is worth more than a complete model. A companion to lecture 1. CCBS · EST 5053 · Fall 2026 · lecture 1 · lineage · written 2026-09-01 Lecture 1 teaches a method in ninety-five minutes and does not have time to say where it came from. This page says where it came from. It is not required reading and nothing on it will be examined. It exists because the examples in the lecture are not arbitrary teaching devices. Each is a real episode in which someone got an important number out of almost nothing, and knowing which episode you are repeating changes how much you trust yourself when you repeat it.

The claim this page argues for Order-of-magnitude reasoning is not a technique for getting approximate answers quickly. It is a method for finding the one constraint that decides a problem, and the history of the method is a chain of instances of exactly that. Every canonical episode below is remembered for the constraint it exposed, not for the number it produced.

Part 1

The claim

Why a course about cells opens with a lecture about bombs and potatoes.

1Two ways to know a number

You can build a model of the thing, or you can find what the thing cannot do. The second route is cheaper and it survives being wrong. Suppose you want to know how long it takes a signalling protein to cross a cell. There are two routes. The model route writes down what the protein does. It needs a diffusion coefficient, a description of the crowded cytoplasm, perhaps binding partners, perhaps active transport. Each ingredient is a place to be wrong, and the errors multiply. When the model disagrees with an experiment you cannot tell which ingredient failed. The constraint route asks instead what the protein cannot do, whatever else is true. It cannot cross faster than diffusion allows unless something is spending energy to carry it. That single fact, with one number, gives an answer in one line and the answer is robust, because it did not depend on the ingredients.

Figure 1. The trade the method makes. A model buys detail with fragility. A constraint buys robustness by refusing to describe the mechanism at all. When both are available they check each other, and when only one is available it is nearly always the constraint.

2What is being claimed here

A history can be told two ways, and the two disagree about what the method is for. There is a competing account of everything below, and it is worth stating properly because parts of it are true. On that account, order-of-magnitude estimation is a collection of useful tricks for landing within a factor of ten, valuable mainly as a sanity check on real calculations, and its famous episodes are anecdotes about unusually clever individuals rather than instances of a method that can be taught. The test between the two accounts is what the people involved said they were doing. So this page quotes them rather than summarising them, and where a quotation cuts against the argument being made, it is quoted anyway. Section 10 is the clearest case: Haldane states the constraint exactly and then gets the mechanism wrong, which the tidy version of the story would leave out.

Figure 2. The lineage. Positions are not to scale. The three centuries between Galileo's scaling argument and Buckingham's theorem are the interesting part: the method was in continuous use long before anyone could say why it worked.

Part 2

The theorem, 1638 to 1914

From a drawing of a thickened bone to a counting argument that explains why the drawing was right.

3Galileo's bone

The first scaling argument in physics is about why a giant would collapse, and it is made with a picture. In the Second Day of the Two New Sciences of 1638, Galileo has Salviati explain why animals cannot simply be scaled up.¹ Weight goes as volume, so as the cube of length. The strength of a bone goes as its cross-sectional area, so as the square. Triple the animal and you multiply the load by twenty-seven while multiplying the support by nine.

Galileo, Second Day "if one wishes to maintain in a great giant the same proportion of limb as that found in an ordinary man he must either find a harder and stronger material for making the bones, or he must admit a diminution of strength in comparison with men of medium stature; for if his height be increased inordinately he will fall and be crushed under his own weight." And, immediately, the empirical check: "a small dog could probably carry on his back two or three dogs of his own size; but I believe that a horse could not carry even one of his own size."

Three things in that passage are the whole method in embryo. First, no model of bone is offered, only a statement about how two quantities scale differently. Second, the conclusion is a prohibition: a giant of ordinary proportions cannot exist. Third, Galileo immediately tests it against something anyone can check, dogs and horses, rather than against a measurement nobody has made. He also does the thing that separates an estimate from a slogan. Simplicio objects that whales are enormous and manage perfectly well, and Salviati concedes that this "suggests another principle" that he had missed. The objection is right: buoyancy removes the load the argument was about. A scaling law is a claim about which effects dominate, and it is falsified by naming an effect it left out.

4The method before it had a theorem

For nearly three centuries the argument was used constantly and justified never. Between Galileo and 1914 the reasoning became standard practice without becoming a subject. Fourier noticed that equations of physics must be homogeneous in their units and made it a check on his own work. Ship designers tested scale models and needed a rule for translating the results to full size. Rayleigh used the argument so routinely, and to such effect, that he eventually wrote a short and slightly exasperated note observing how much could be settled by attending to units alone. What was missing was not the technique but the licence. Everyone could see that the argument worked. Nobody could say how many independent conclusions it was entitled to produce, or when it had been exhausted. That is the gap Buckingham closed.

5Buckingham counts

The 1914 paper's achievement is not a new technique. It is a theorem that says exactly how much the old technique is worth. Buckingham was at the National Bureau of Standards, an institution whose business is units, which is the right place from which to notice that units are a constraint rather than a bookkeeping convention. His paper is titled On physically similar systems; illustrations of the use of dimensional equations, and the illustrations are half of it.² The argument is short. A physical relation among

n

quantities must hold whatever units you measure them in. Changing the size of the fundamental units multiplies each quantity by a known factor, and demanding that the relation survive this imposes one equation per fundamental unit. With

k

independent fundamental units, that is

k

equations on the exponents, leaving

n - k

free combinations. Buckingham's own phrasing of the count is that the number of independent arguments "is the maximum number of independent dimensionless products ... which can be made by combining the

n

quantities", and that there are "

(n - k)

equations" determining them.

The theorem has been restated and generalised many times since, and a centenary survey collects the versions.¹⁴ None of them changes what a working estimate does with it, which is the count and nothing else.

Why the theorem changed the status of the method

Before 1914, a dimensional argument was a piece of luck: you noticed a combination and it happened to work. After 1914 it is an algorithm with a termination condition. You list the quantities, count the units, subtract, and you know in advance how many independent facts you are going to get. The special case $n - k = 1$ is the one that makes careers, because then there is nothing to choose and the answer is forced.

6What the theorem withholds

It gives the shape and never the size. Knowing which half you have is most of the skill.

The theorem hands you the functional form and refuses to hand you the number in front of it. The blast wave really does grow as $t^{2/5}$ , and the coefficient depends on the ratio of specific heats of air, which no amount of counting will produce. Taylor spent a paper computing it.

This division is more useful than it sounds, for a reason worth stating plainly. The prefactor is usually of order one, and the exponent is never. Getting an exponent wrong compounds without limit as you extrapolate, and getting a prefactor wrong costs you a fixed factor of two or three. So the theorem gives away the cheap half and keeps the half you can afford to lose.

There is a second thing it withholds, and it is the one that catches people. The theorem takes your list of quantities as given. If you leave one out, the count is wrong and the answer is confidently wrong, and nothing in the formalism will warn you. Deciding what belongs on the list is physics, not arithmetic, and it is the only part of move 1 that can fail.

Part 3

The demonstration, 1941 to 1950

The most famous instance of the method, and a comparison that is usually reported unfairly.

7A paper about a bomb that did not exist

Taylor's blast-wave analysis was written four years before the explosion it is remembered for.

The opening of Taylor's 1950 paper is worth reading as history rather than as physics.³ It states that the work "was written early in 1941 and circulated to the Civil Defence Research Committee of the Ministry of Home Security in June of that year", that the author "had been told that it might be possible to produce a bomb in which a very large amount of energy would be released by nuclear fission", and, in an aside that dates the document precisely, that "the name atomic bomb had not then been used".

The question he was asked was practical and strange: in an ordinary bomb the damage is done by hot gas suddenly generated in a confined space, so would similar effects be produced if energy were released in a highly concentrated form unaccompanied by the generation of gas? Nobody could answer by experiment. The device did not exist.

What Taylor did is the constraint route in its purest form. He replaced the bomb with an idealisation, a finite amount of energy released at a point, and asked what the surrounding air must then do. Four quantities, three dimensions, one group, and the radius must go as the two-fifths power of time. Then he did the hard part and computed the coefficient.

The prediction nobody quotes

The same 1941 paper contains a result that matters for section 9 and is almost always dropped: at 20 atmospheres "45 % of the energy has been degraded into heat which is not available for doing work", leading "to the prediction that an atomic bomb would be only half as efficient, as a blast-producer, as a high explosive releasing the same amount of energy." Taylor knew, before there was a bomb, that a blast measurement of a nuclear explosion would report roughly half the energy released.

8Three estimates of one explosion

Trinity was measured by a stopwatch and scraps of paper, by twenty-five photographs, and by chemistry. The three answers are the argument for the method.

Fermi was present at the Trinity test on 16 July 1945. As the blast front passed he dropped small pieces of paper and watched how far they were displaced, got about 2.5 m, and announced roughly ten kilotons. The estimate was made in the seconds after the shock arrived, before any instrument had been read.⁵

Taylor, in England, waited for a strip of high-speed photographs to be declassified. The fireball radius at each frame was public and the yield was not, so he read twenty-five radii and times off the published images and applied his own formula.⁴ His table is reproduced and re-analysed below, and both halves of the prediction survive: the exponent comes out at 0.406 against the theoretical 0.400, and every frame after the first implies the same energy.

Figure 3. The 1941 prediction against the 1945 event. Left, Taylor's own twenty-five measurements with the two-fifths slope that follows from counting alone. Right, the yield each frame implies. The point is the flatness: a scaling law derived before the object existed describes it across three decades in time.

9What the comparison actually shows

The popular version scores Fermi against Taylor against chemistry and declares a winner. All three are measuring different things.

The three numbers are: Fermi about 10 kilotons, Taylor 16,800 tons with an alternative of 23,700, and the accepted radiochemical value about 21 kilotons. Told carelessly, this says Fermi was a factor of two low, Taylor was twenty per cent low, and the chemists were right.

Figure 4. Not the same quantity three times. Fermi and Taylor both measured the mechanical energy delivered to the air. Radiochemistry measures the total energy released, including what left as radiation. Taylor's figure is explicitly for "that part of the energy which was not radiated outside the ball of fire", so a shortfall is expected.

The careless telling is unfair to Fermi, who was measuring blast with a method that cannot see radiation, and unfair to Taylor, who said which quantity he was estimating and predicted the discrepancy four years in advance. It is also unfair to the method, because it invites the conclusion that back-of-envelope work is roughly half right, when the actual record is that two independent blast estimates made by wildly different means agreed with each other.

The habit this should install

Before comparing your estimate to a published number, check that both refer to the same quantity. Most apparent factor-of-two disagreements in the literature are definitional, and the ones that are not are the interesting ones. This is cheap to check and almost never done.

Part 4

The transfer to life, 1926 to 1977

What changes when the object of the estimate is alive, and what does not.

10Haldane gets it right and wrong

The jump-height argument taught in lecture 1 is in an essay from 1926, complete with the caveat, and with the wrong mechanism attached.

J. B. S. Haldane's On Being the Right Size is four pages that founded a genre.⁸ It contains the mouse down the mine shaft, the surface-to-volume argument about insect respiration, and, in one sentence, the entire content of §6 of the lecture 1 core page.

Haldane, 1926

"although Galileo demonstrated the contrary more than three hundred years ago, people still believe that if a flea were as large as a man it could jump a thousand feet into the air. As a matter of fact the height to which an animal can jump is more nearly independent of its size than proportional to it. A flea can jump about two feet, a man about five. To jump a given height, if we neglect the resistance of the air, requires an expenditure of energy proportional to the jumper's weight. But if the jumping muscles form a constant fraction of the animal's body, the energy developed per ounce of muscle is independent of the size, provided it can be developed quickly enough in the small animal."

Read that carefully, because there are four separate things in it.

First, he credits Galileo, which is where the lineage in Figure 2 comes from. Second, the claim is hedged: "more nearly independent of its size than proportional to it", not "independent". His own two numbers differ by a factor of 2.5, and he is telling you what an order-of-magnitude claim actually claims. Third, the derivation is exactly the one in the lecture, in two clauses.

Fourth, and best, the clause I have italicised is the release-time constraint. "Provided it can be developed quickly enough in the small animal" is $\tau = L/v$ in words, in 1926, before anyone had filmed a flea.

And then the sentence that follows it

Haldane continues: "As a matter of fact an insect's muscles, although they can contract more quickly than our own, appear to be less efficient; as otherwise a flea or grasshopper could rise six feet into the air."

That is the wrong answer. The modern understanding is that small jumpers do not power the jump with muscle at all. They contract slowly against an elastic element, hold it with a latch, and release it in under a millisecond, so the spring sets the power and the muscle only sets the energy. Haldane identified the constraint correctly and misdiagnosed how the animal escapes it.

This is the most instructive episode on the page, and it is instructive in a way that flatters nobody. A first-rate scientist did the estimate right, noticed the anomaly, and reached for the nearest available explanation, which happened to be a slur on insect muscle. The estimate told him where to look. It could not tell him what he would find, and he did not go and look.

The lesson is not that Haldane was careless. It is that an estimate's output is a question, and the question still has to be asked of the world. If lecture 1 sends you away with one habit, it should be that one.

The number was settled in 1967, and the textbooks went on repeating 1926

Haldane's "about two feet" is $61\,\mathrm{cm}$ . The first careful measurement of a flea's jump put it at $\pf{3.5\,\mathrm{cm}}$ , and it was published in 1967.¹⁹

H. C. Bennet-Clark and E. C. A. Lucey filmed the rabbit flea Spilopsyllus cuniculus at 1000 frames per second, on stock shot for the BBC. They measured a take-off speed of $0.8$ to $1.2\,\mathrm{m}/\mathrm{s}$ , took $1\,\mathrm{m}/\mathrm{s}$ as the working figure, added air resistance at about half of gravity, and got $3.5\,\mathrm{cm}$ . Their summary opens with the phrase "a jump of 3.5 cm. height by the rabbit flea".

Forty-four years later Sutton and Burrows filmed a different species, Archaeopsyllus erinacei, at 5000 frames per second, and published a full table.¹⁷ Take-off speed $1.3 \pm 0.21\,\mathrm{m}/\mathrm{s}$ at $39 \pm 5.7^\circ$ , from a body $1.8\,\mathrm{mm}$ long massing $0.7\,\mathrm{mg}$ . The vertical component is $0.82\,\mathrm{m}/\mathrm{s}$ , which lifts the centre of mass about $3\,\mathrm{cm}$ . Even their fastest jump, $1.9\,\mathrm{m}/\mathrm{s}$ , gives under $9\,\mathrm{cm}$ with every bit of the speed pointed straight up and the air deleted.

Two species, two laboratories, two film speeds, forty-four years apart, and the same answer to within the width of the animal. Haldane is high by a factor of roughly twenty. His follow-up conjecture, that a flea would "rise six feet into the air" if only its muscle were as efficient as ours, is high by a factor of sixty, and it was the conjecture he offered instead of going to look.

The part worth sitting with is the chronology. The correction has been in the primary literature since 1967. The essay has been quoted in textbooks and lectures continuously since 1926, this course's included, and the two have not been in the same room. A number does not become right by being repeated, and it does not become wrong by being contradicted somewhere nobody checked.

Notice which half of Haldane's sentence survived. The scaling relation is exact enough to be worth teaching a century later. The datum attached to it was wrong by more than an order of magnitude, and the wrong datum is what made the anomaly he then misexplained. His hedge, "more nearly independent than proportional", was sound method spent on a bad number.

And the same estimate, run properly, found the mechanism

The 1967 paper is worth reading for a second reason. Bennet-Clark and Lucey did not only measure the jump. They ran the lecture's own method on it, and the method delivered the spring.

The chain is four steps and no step needs anything a reader of lecture 1 does not have.

One. The energy. A $0.45\,\mathrm{mg}$ flea leaving at $1\,\mathrm{m}/\mathrm{s}$ carries $\tfrac{1}{2}mv^2 = 2.25$ ergs, or $0.225\,\mathrm{\mu J}$ .

Two. The time. Blur lengths on the film put the acceleration at $0.75$ to $1.0\,\mathrm{ms}$ . The latent period of insect muscle was known to be around $3\,\mathrm{ms}$ . A muscle cannot pull appreciably inside an impulse three times shorter than the delay before it starts, so the muscle is not driving the jump.

Three. The rate. Bumblebee flight muscle was known to produce about $6\times 10^{5}$ ergs per gram per second. Charging the flea 20% of its body mass as jumping muscle gives $56$ ergs per second for the whole animal. Delivering $2.25$ ergs at that rate takes $\pf{about\ 0.05\,\mathrm{s}}$ . So the energy must be stored, and stored for something like fifty milliseconds.

Four. The test. They went back to the film. Every flea preparing to jump sits with its hind femora raised, motionless, for $\pf{0.10}$ to $\pf{0.25\,\mathrm{s}}$ before launching. The pause had been photographed before and its significance missed. Cool a flea to $6^\circ\mathrm{C}$ and the pause stretches to seconds.

Then they found the store: a pad of resilin between notum and pleuron, which takes the methylene-blue stain, has an elastic efficiency near 96%, and holds about $1.4\times 10^{-4}\,\mathrm{mm}^{3}$ . At $1.5\times 10^{4}$ ergs per cubic millimetre that is $2.1$ ergs a leg and $4.2$ ergs a flea, against the $2.25$ the jump needs. Adequate, with margin for friction, which is what an honest estimate looks like when it lands.

This is the whole lecture in one paper: an order-of-magnitude argument that could not be settled by more thinking, converted into a number, converted into a prediction about a pause on a film, and then checked. Haldane had the same constraint in 1926 and stopped at the anomaly. Bennet-Clark carried it one step further and got the answer.

The correction has a correction

It would be tidy to say Haldane overstated the flea and stop. The fuller picture is stranger. Ilton and co-authors show that take-off speed across jumping insects is not monotone in size.¹⁸ It peaks in the tens of milligrams and falls away on both sides, because below some threshold the spring can no longer be loaded without failing. Froghoppers reach $5-6\,\mathrm{m}/\mathrm{s}$ at $20-50\,\mathrm{mg}$ and pygmy mole crickets the same at $10\,\mathrm{mg}$ , against the flea's $1.3\,\mathrm{m}/\mathrm{s}$ .

Bennet-Clark saw the shape of this in 1967 without the data to pin it. His discussion notes that the argument "accounts for the far shorter jump of the flea, which has a similar proportion of muscle involved in the jumping mechanism" as a locust or a frog. He is comparing spring users to spring users and finding the flea wanting, which is Ilton's conclusion in words, half a century early and without the material threshold that explains it.

So the flea is not merely below Haldane's figure. It is below its own relatives. The canonical illustration of "small animals need a spring" turns out to be a mediocre example of what a spring can do, and the reason is a second scaling law that Haldane had no way to reach. Two ways of being wrong sit on top of each other here, and only the outer one is his.

11Purcell deletes inertia

A festschrift talk in 1976 that is still the best single piece of writing on what changes when a thing is small.

Purcell's Life at Low Reynolds Number was a talk in honour of Victor Weisskopf, transcribed from tape, and the printed version deliberately keeps the spoken tone.⁹ He opens by saying he would not have the nerve to give it under any other circumstances, and then delivers a piece of physics that reorganised how biologists think about the cell.

The argument is one number. The Reynolds number for a micron-scale swimmer is "about $10^{-4}$ or $10^{-5}$ . For these animals inertia is totally irrelevant. We know that F = ma, but they could scarcely care less."

Then he converts the number into something you can feel. A bacterium swimming at 30 micrometres per second, if it stops pushing, coasts "about 0.1 angstrom", and takes 0.6 microseconds to stop. Not a fraction of its body length. A fraction of an atom.

From that he draws the scallop theorem: with no inertia the flow depends only on the sequence of shapes a swimmer passes through, not on how fast it passes through them, so any stroke that retraces itself produces no net motion however violently it is performed. A scallop, having one hinge, cannot swim at all. Escaping the theorem requires a stroke that traverses a loop in shape space, which is why bacteria turn helical flagella rather than paddling.

The moment worth stealing

Comparing transport by stirring, $l/v$ , against transport by diffusion, $l^{2}/D$ , Purcell forms the ratio $lv/D$ and writes: "I'm sure this ratio has someone's name but I don't know the literature and I don't know whose number that's called. Call it S for stirring number."

It is the Péclet number, and it had been named for decades. A Nobel laureate constructing a standard dimensionless group from scratch, in public, and saying plainly that he does not know what it is called, is the method's own argument for itself: the reasoning does not require the literature, and pretending to a familiarity you lack costs more than admitting the gap.

12When the estimate is the answer

The honest limit of the null-model framing: sometimes the organism is already at the bound, and there is no gap to explain.

The standard way to use an estimate in biology is as a null model. Compute what physics alone would do, compare with the observation, and the discrepancy is where the biology is. Section 9 of the core page sets this out.

Berg and Purcell's Physics of chemoreception, from the same year as the Reynolds paper, is the case that complicates it.¹⁰ They asked how accurately a bacterium could possibly determine a chemical concentration, given that molecules arrive at its surface by diffusion and arrival is a counting process with its own noise. The answer is a bound that depends on the cell's size, the diffusion coefficient, and the averaging time, and no receptor design can beat it.

Then they compared with what bacteria actually achieve in chemotaxis, and found them close to it.

This is the third outcome, and it is worth being clear that it undercuts the framing rather than confirming it. There is no gap. There is nothing left to attribute to design, because physics already forbids doing better. Which is a stronger result than any gap would have been: it converts "why is the cell built this way" into "the cell is built the only way that works", and it tells you to stop looking for a cleverer mechanism.

Part 5

The teaching, 1938 to 2010

How a habit of individual physicists became a course with a syllabus, and what got lost on the way.

13The Fermi problem

The name attached itself to a genre, and the genre is now general-education material.

Fermi's habit of demanding estimates from students, the piano tuners of Chicago being the canonical example, gave the genre its name. What Trinity added was proof that the habit was not a parlour trick: the same man who asked about piano tuners produced a bomb yield with scraps of paper that agreed to within a factor of two with a season of fluid dynamics.

The pedagogical literature that grew from this is now substantial and is mostly about general education rather than about training physicists.¹³ The argument there is that estimation teaches a citizen to notice when a number in the news is impossible, which is a different and more modest goal than the one this course has.

14Caltech, Ph 101

The direct ancestor of this lecture is a Caltech course and its unfinished book.

The course this lecture descends from is Caltech's Order-of-Magnitude Physics, taught by Peter Goldreich and Sterl Phinney, with Sanjoy Mahajan writing it up. The draft book carries a title that is itself an argument: Order-of-Magnitude Physics: Understanding the World with Dimensional Analysis, Educated Guesswork, and White Lies, dated 1 August 1999.⁶

Two decisions in it shaped everything downstream. First, the material is organised by method rather than by subject: dimensional analysis, then scaling, then special cases, with examples pulled from astrophysics, materials and biology as needed. Second, the title admits to the lies. An order-of-magnitude treatment deliberately asserts things that are false in detail, and the course says so on the cover rather than hoping students will not notice.

15Lying on purpose

Mahajan's contribution is to make the approximations themselves the subject, and to name the moment when you are allowed to lie.

Mahajan carried the Caltech material to MIT as 6.055, The Art of Approximation in Science and Engineering, which became Street-Fighting Mathematics.⁷ The change of emphasis is visible in the titles. Caltech's book is about the physical world. Mahajan's is about the reasoning, and the physics is the exercise material.

The clearest short statement of the position is a 2005 note on estimating the fuel consumption of a car, based on a talk he had titled Lying and estimating for general education.¹⁵ The whole calculation is carried through in a few pages, and the interest is not the answer but the running commentary on which lies are safe.

Which lies are safe

A lie is safe when the quantity it distorts enters the answer weakly, and dangerous when it enters strongly. Rounding a drag coefficient to one is safe because the answer is proportional to it and you needed a factor of ten. Rounding an exponent is never safe, because its error compounds with the range you extrapolate over. This is the same asymmetry as §6: exponents are expensive, prefactors are cheap.

16Biology gets its numbers

The last ingredient is the least glamorous and the most necessary: someone had to collect the numbers.

An estimate needs an anchor, and until recently a biologist wanting the concentration of a typical metabolite or the copy number of a typical protein had no place to look. Rob Phillips and Ron Milo made this an explicit programme, arguing in 2009 that biology needed the numerical fluency physics takes for granted, and then building the infrastructure to make it possible: the BioNumbers database, and the book Cell Biology by the Numbers.¹¹¹²

The unglamorous half matters more than it looks. §7 of the core page is a case where the whole chain of reasoning was correct and the conclusion was wrong by two orders of magnitude, because one anchor, the potato, had never been checked. A curated number with a provenance is what stands between an estimate and a fabrication.

Part 6

Now

What is different in 2026, and what is still open.

17What the arithmetic being free changes

An agent will do every step of every estimate on this page instantly. It will not tell you which estimate to make.

Every calculation in lecture 1 can now be produced on demand, correctly, in seconds. It is worth asking what that does to the case for teaching the method, and the honest answer is that it strengthens it, because of how the method decomposes.

Table 1. The three moves against what an agent supplies. The pattern is that automation ate the middle column and left the outer ones.
Move	What it requires	Automatable?
1. Count the dimensions	Listing the relevant quantities, then arithmetic	The arithmetic, yes. Choosing the list, no.
2. Name the bottleneck	Judging which quantity is actually fixed	No. This is the physics.
3. Calibrate on an anchor	An experience you trust, and a check on it	Looking up a number, yes. Trusting it, no.

The parts that survive are the parts that were always the skill. Deciding that air density belongs in the blast-wave list and viscosity does not is a physical judgement made before any counting begins, and get it wrong and the theorem returns a confident falsehood. Noticing that the spacing between cars is not free but equal to $v\tau$ is the same kind of act. Neither is arithmetic.

And move 3 is where the exposure has actually increased. An agent will supply a plausible number for anything you ask, instantly, with no signal about whether it was measured or inferred or invented. The potato is the object lesson: a wrong anchor propagated cleanly through correct reasoning to a wrong conclusion about the moon, and the only thing that catches it is the habit of asking where the number came from. That habit used to be enforced by the difficulty of getting numbers at all. It is not enforced by anything now.

18What is still an argument

Where this page is on firm ground, and where it is telling a story.

The historical claims here rest on full texts that were read, and the quotations are given rather than paraphrased so that you can check the reading. Galileo's scaling argument, Buckingham's count, Taylor's 1941 date and his half-efficiency prediction, Haldane's four moves and his wrong mechanism, Purcell's numbers and his admission about the Péclet number: all are quoted from the sources listed below. The reading itself is logged in the lecture's research folder, including the corrections it forced.¹⁶

Three things are weaker, and it is better to say so.

The unifying claim is an interpretation. That these episodes are instances of one method, rather than six clever people solving six problems, is an argument this page makes and not something the participants said. They did not think of themselves as members of a tradition. The tradition was assembled afterwards, partly by the Caltech course, and this page is a further step in the same assembly.

The flea number is unresolved. §10 says so at the point of use.

The claim in §17 is a prediction. That the judgement half of the method resists automation while the arithmetic half does not is a bet about the next few years, not an observation. It is falsifiable and worth watching, and if it turns out to be wrong then this course's design is wrong with it. Given the argument the course makes for itself, that is the right thing to have staked.

†References

Full texts of everything marked PDF are in the course's literature/ archive, with an annotated index and a corrections log.

Galileo Galilei, Dialogues Concerning Two New Sciences (1638), trans. H. Crew and A. de Salvio (1914), Second Day. PDF The square-cube argument, the thickened-bone figure, the small dog and the horse, and Simplicio's objection about whales.
E. Buckingham, "On physically similar systems; illustrations of the use of dimensional equations", Phys. Rev. 4, 345–376 (1914). PDF
G. I. Taylor, "The formation of a blast wave by a very intense explosion. I. Theoretical discussion", Proc. R. Soc. A 201, 159–174 (1950). PDF Written 1941. The 1941 provenance, the framing question, and the half-efficiency prediction are all in the first two paragraphs.
G. I. Taylor, "The formation of a blast wave by a very intense explosion. II. The atomic explosion of 1945", Proc. R. Soc. A 201, 175–186 (1950). PDF Table 1, page 176, is the data in Figure 3.
J. I. Katz, "Fermi at Trinity", arXiv:2103.05784 (2021). PDF Reconstructs how the paper-scrap displacement could have been converted to a yield, and gives the radiochemical comparison.
P. Goldreich, S. Mahajan and S. Phinney, Order-of-Magnitude Physics, draft of 1 August 1999. PDF
S. Mahajan, The Art of Approximation in Science and Engineering, MIT 6.055 draft (2009); published as Street-Fighting Mathematics, MIT Press (2010). PDF
J. B. S. Haldane, "On Being the Right Size", in Possible Worlds and Other Papers (1926). PDF
E. M. Purcell, "Life at low Reynolds number", Am. J. Phys. 45, 3–11 (1977). PDF
H. C. Berg and E. M. Purcell, "Physics of chemoreception", Biophys. J. 20, 193–219 (1977). PDF
R. Phillips and R. Milo, "A feeling for the numbers in biology", PNAS 106, 21465 (2009). PDF
R. Milo et al., "BioNumbers", Nucleic Acids Res. 38, D750 (2010). PDF · book.bionumbers.org
C. Efthimiou and R. Llewellyn, "Cinema, Fermi problems and general education", Phys. Educ. 42, 253 (2007). PDF
D. Jonsson, "Dimensional analysis: a centenary update", arXiv:1411.2798 (2014). PDF
S. Mahajan, "Estimating gas mileage: an example of order-of-magnitude physics", arXiv:physics/0512209 (2005). PDF
lectures/research/lecture01/: design.md for the thesis and chain of argument, analysis/oom_numbers.py for every number, analysis/make_figures.py for Figure 3. The literature archive's README.md carries the corrections this pass produced, and NEEDED.md the sources it could not obtain.
G. P. Sutton and M. Burrows, "Biomechanics of jumping in the flea", J. Exp. Biol. 214, 836–847 (2011). PDF Table 2, page 838. The measurement that closes the Haldane figure. Arrived after this page was first written, and it changed §10.
M. Ilton et al., "The principles of cascading power limits in small, fast biological and engineered systems", Science 360, eaao1082 (2018). PDF Why take-off speed peaks in the tens of milligrams rather than rising all the way down. The froghopper and mole cricket figures are theirs.
H. C. Bennet-Clark and E. C. A. Lucey, "The jump of the flea: a study of the energetics and a model of the mechanism", J. Exp. Biol. 47, 59–76 (1967). PDF The measurement that settled the flea's jump height at 3.5 cm, forty-four years before Sutton and Burrows, on a different species. Also the origin of the resilin spring, and the source of the energy-storage-time argument quoted in §10: the numbers are on pages 62 to 64, the observed 0.10–0.25 s pre-jump pause on page 63, and the summary statement of the 3.5 cm jump on page 75. Note that its 20% assumption for jumping-muscle fraction is nearly double the 11% Sutton and Burrows carry over from locusts.