2

DIMENSIONAL ANALYSIS

1. Physical mathematics. We learned from the ancients—Pythagoras (who died

∼ 497 b.c.), Galileo (1564–1642)—that world structure admits of mathematical
description, “God is a mathematician.” This non-obvious fact seems never to
lose its deeply surprising and mysterious quality in the imaginations of even
the greatest physicists,1 and becomes the more surprising when one appreciates
that
• Most attempts to comprehend natural events—historically, cross-culturally
—have been phrased in qualitative language, the language of (say) myth.
• We are still obliged to use qualitative language when discussing most of
the concepts (beauty, justice, . . . ) and problems (“Why did she do that?”)
encountered in the course of our non-scientific day-to -day affairs. Physics
is, in this respect, “special.”
• Physical calculations, no matter how long/intricate/abstract they may
become, manage somehow to “stay in touch with the world.”
It is interesting to ask: Can one imagine a world so unstructured/chaotic as NOT
to admit of a “mathematical physics?”

Numbers—actually integers, which are promoted to real numbers by acts

of abstraction—enter physics (and the sciences generally) by
• counting, and that special kind of counting called
• measurement.
1
E. P. Wigner’s essay “The unreasonable effectiveness of mathematics in the
natural sciences,” Comm. Pure Appl. Math. 13, 1–14 (1960) is a classic in the
field. See also R.W.Hamming,“The unreasonable effectiveness of mathematics,”
Amer. Math. Monthly 87, No. 2 (February 1981), which can be found on the
web at http://www.lecb.ncifcrf.gov/∼toms/Hamming.unreasonable.html.
2 Dimensional analysis

Interesting subtleties aside, it is, I think, the role assigned to measurement

intruments that distinguishes the physical sciences from all other branches of
knowledge—among them philosophy and mathematics itself. It was historically
the search for patterns among physically generated numbers that inspired the
invention of much fundamental mathematics. But by the beginning of the
19th Century it had become clear that mathematics is an autonomous subject,
related to but unconstrained by contingent world-structure. Thus arose the
distinction between pure and applied mathematics. Physicists—when not
acting as instrument builders, dreamers. . . —are applied mathematicians,
mathematical model makers.
It is useful to attempt to clarify the pure/applied distinction. When a
(pure) mathematician writes

f (x, ẍ; m, k) ≡ mẍ + kx = 0

he has described precisely a problem that lives only in his mind, the problem
that he proposes to discuss. But when a physicist writes that same equation
he has described precisely only how he proposes to model the (approximate!)
behavior of a class of objectively real physical systems. For the physicist (but
not for the mathematician) the variables (x, t) and the parameters (m, k) are
considered to have objective referents; i.e., to have number values derived from
acts of measurement. This circumstance gives the physicist a kind of “heuristic
advantage” which is denied the pure mathematician, and which it will be my
primary objective here to describe.

2. Measurement: units & dimensionality. For the simple purposes at hand it

will be sufficient to consider by way of illustration the measurement of length.
Similar remarks pertain, with interesting variations, to the measurement of
mass/time/temperature. . . indeed, to all the variables/parameters that enter
properly into physics, though it is not clear that they pertain with the same
force to some of the variables (“propensities”) contemplated by economists and
psychologists.
Physics is circular. Before we can undertake the measurement of (say)
length we must have some preliminary sense—however informal/tentative—of
1) what the concept in question (distance between two points) “means”
2) what operations can sensibly (which is to say: consistently with a physics
yet to be invented!) relate to its “measurement.”
Those bridges crossed (and they have from time to time to be reexamined), we
1) pick (arbitrarily/conveniently) a reproducible “standard instance” of the
concept in question (inscribe two marks on a stable bar, these defining our
unit of length) and
2) proceed operationally to quantify natural instances of the concept by stating
“how many times they contain the unit .” I need not describe how one uses
a meter stick, but point out that
Units & dimensionality 3

a) we proceed on the (usually tacit) assumption that meter sticks do not

change their defining properties when transported from one spacetime
region to another;
b) meter sticks differ profoundly from (say) clocks and thermometers, for
one cannot “lay temporal (or thermal) intervals side by side” as one can
and does lay a meter stick beside the object whose length one wants to
measure;
c) to measure very short or very long spatial intervals one can/does not
use a meter stick: the concept of length (which in well-established
practice spans at least forty orders of magnitude) is quantified by a
spliced hierarchy of distinct operational procedures .
Preceding remarks may serve to establish this point: metrology lives at the
frontiers of both technology and philosophy. It may, at first blush, seem dull, but
it is definitely not trivial. A science can be no more secure than its metrological
foundations.
When we say that “the distance from A to B is about x” we mean that
1) we imagine there to exist a “true distance” having “the nature of a
length,” i.e., the dimensionality of length;
2) measurement has or would yield the value

x length units

where x is a real number, known only to within some observational

error ∆x (which may be irreducible-in-principle).
When we wish to indicate the dimensionality (as opposed to the numerical
value) of x we write [x]: thus

[x] = length

It is clear that units and dimensionality are quite distinct concepts , not to
be confused, certainly not interchangable. There are, as will emerge, subtle
senses in which the concept of dimensionality is convention-dependent, but the
conventionality of the unit is obvious/manifest. This is a circumstance
unaffected by the fact that mature physics supplies certain “natural” units
—thus

c = natural unit of velocity

e = natural unit of electric charge

= natural unit of action
..
.

—but these are seldom natural in the sense “most convenient for practical
work.”
4 Dimensional analysis

3. Dimensional interrelationships. Expansion of the phenomenological scope

of our physics tends to increase the number of what we may initially suppose
to be independently dimensioned physical concepts (each measured in its own
units), while the discovery of a (deep) physical law serves often to decrease
the number of independently dimensional concepts, and to establish numerical
interrelationships within our system of units. By way of clarification. . .
Suppose that—experts in length-measurement that we are—we expand our
physics to include the concept of area. We measure lengths in (say) centimeters,
but the new concept (area) requires the introduction of a corresponding new
unit: we adopt (say) the acre. “Research” shows that the area A of any
rectangle (sides of lengths x and y) can be described

A= 1 xy
40468564
This statement is, however, unacceptable as a statement of “natural law”
because its validity is contingent upon a convention (choice of units). To avoid
this formal defect we write (with A in acres, x and y in centimeters)
A =kxy (1)
k = 1 acres/centimeter2
40468564
[k] = area/length2 (2)
The point is that (1) captures—in a convention-independent manner—the
structure of the functional relationship among A, x and y. It is evident that
by unit-adjustment we could in particular arrange to have
k = 1 (adjusted area unit)/(adjusted area unit)2
—this being a mere process - of-convenience that leaves (1) unchanged. Less
evidently, we might insist that
k = 1 is dimensionless
which—by (2)—enforces
[area] = [length]2
This is by nature a dimensional interrrelation, and means that we can discard as
redundant our unit of area. The “composite nature of area” would, of course,
be one of the central features of the “theory of area” which would be the
final product of our “physical” research. If the above seems far-fetched and
belabored, consider the following:
We are (let us suppose) interested in the dynamics of mass points. We
discover the subject to be dominated by four key concepts—mass, length,
time, force—which we quantify by introducing a corresponding quartet of units.
Experimental research leads to the conclusion that
F =k mẍ
(force unit)
k = numeric ·
(mass unit)(length unit)/(time unit)2
Dimensional interrelationships 5

We may interpret the observed universality of F = k mẍ (i.e., the fact that
it is found to work with the same k for all forces F and all masses m) as an
invitation to set

k = 1 (dimensionless)

Then
F = mẍ (3)
gives
[force] = [mass][length][time]−2 (4)
and one of our units has been rendered redundant : this is standardly read as
license to write

force unit = (mass unit)(length unit)/(time unit)2 (5)

Suppose now that we expand our interest to include the dynamics of

electrically charged particles. A new concept— charge —and a corresponding
new unit enter the discussion. Experimentally
q1 q2
F =k
r2
(force unit)(length unit)2
k = numeric · (6)
(charge unit)2
≡ 1 = 8.99 × 109 (Newton)(meter)2 /(Coulomb)2 (7)
4π0
3 −2 −2
= numeric · (mass unit)(length unit) (time unit) (charge unit)

Here again it becomes attractive to set

k = 1 (dimensionless)

Coulomb’s law then assumes the form

q1 q2
F = (8)
r2

giving
1
[charge] = [force] 2 [length]
= [mass] 2 [length] 2 [time]−1
1 3
(9)

and again one of our units (taken most naturally to be the charge unit) has
been rendered redundant. This option gives rise to the so -called “electrostatic
system of units.” Engineers prefer, however, to regard charge as a dimensionally
independent entity; i.e., to live with (6) and (7).
6 Dimensional analysis

Suppose now that we expand our sphere of interest once again, to embrace
the gravitational interaction of material particles. We might expect a new
concept—“gravitational charge” µ—and a corresponding new unit to enter the
picture. The essence of Newton’s Universal Law of Gravitation resides in the
two -part assertion that
µ1 µ2
Fgravitation = K (10.1)
r2
µ = km with k the same for all bodies, all materials (10.2)

—assertions that when taken in combination yield the more commonly

encountered form
m 1 m2
Fgravitation = G (11)
r2
G = Kk 2

The discovered universality2 of (10.2) entitles us to set k = 1 (dimensionless);

i.e., to identify the concepts of gravitational charge and inertial mass.3 Working
from (11) one has

(force unit)(length unit)2

G = numeric ·
(mass unit)2
= 6.672 × 10−11 (Newton)(meter) /(kilogram)
2 2

−1 3 −2
= numeric · (mass unit) (length unit) (time unit)

Were we to yield now to our recently acquired instinct we might attempt

to set
= 1 (dimensionless)
This, however, would require that the dimensions presently assigned to G be
redistributed. Suppose we assume—arbitrarily—that that full responsibility is
2
Had Galileo actually performed the Leaning Tower experiment he would
have been in position to assert, in about 1590, that

ka µ /m
κa/b ≡ = a a = 1 ± 10−2
kb µb /mb

Newton (1686) obtained κa/b = 1 ± 10−3 from the physics of pendula, and by
1832 Bessel had obtained κa/b = 1 ± 10−5 by a refinement of Newton’s method.
Such experiments are usually associated with the name of Baron Lórand von
Eötvös, who in 1922 used a torsion balance technique to achieve κa/b = 1±10−8 .
A lunar laser ranging experiment obtained κa/b = 1 ± 10−12 in 1976, and by
2003 an astronomical technique had given κa/b = 1 ± 10−18 . This continuing
experimental effort is motivated by the fact that the Principle of Equivalence
(see below) is fundamental to general relativity.
3
That mgravitational ≡ minertial is the upshot of the principle of equivalence.
Dimensional interrelationships 7

to be assigned to [mass], and that [length] and [time] are to remain passive
by-standers. If we insist that Newton’s 2nd law is to retain its familiar design
then
m −→ m̃ = αm enforces F −→ F̃ = αF
This said, we observe that if we set α = G then

F = mẍ entails F̃ = m̃ẍ

m m m̃ m̃
Fgravitational = G 1 2 2 entails F̃gravitational = 1 2 2
r r
and the “gravitostatic mass” m̃ becomes dimensionally redundant with [length]
and [time]:
[gravitostatic mass] = [length]3 [time]−2
This, however, is not standardly done. . . for, I suppose, some mix of the
following (mainly practical) reasons:
1) It seems inappropriate to build gravitation into the foundations of our
metrology since
a) gravitational forces are (relatively) so weak
b) gravitational effects are irrelevant to most of our physics.
2) While F and m can be measured by well established laboratory procedures,
F̃ and m̃ cannot be.
3) G is known with insufficient precision to make the program described above
metrologically sound
4) It could be argued—I would argue—that constants of nature

{G, also c, e,
, . . .}

are too important to be metrologically disguised, at least until physics has

reached a point of higher maturity.
5) Mass is much too important/busy a concept to be assigned the awkward
composite dimension stated above.
We conclude from the preceding discussion that in this dusty corner of
physics one confronts—depending upon how much one knows about world-
structure—many options, and that whether one elects to exploit those options
is a question settled on grounds partly utilitarian (the nature of the job at
hand) and partly conventional. Definitions (density ≡ mass/volume, velocity ≡
length/time) lead trivially to concepts of composite dimension, and to
dimensional interrelationships of a trivial nature. On the other hand, physical
laws (F = ma, E = mc 2 , E = hν, . . .)—to the extent that they are “deep/
fundamental—bring to light dimensional interrelationships that contain the
element of surprise. All such interrelationships provide opportunities to contract
the list of “primary dimensions (units).” The constants of Nature which remain
explicit in our physical equations signal by-passed opportunities. The question
How many dimensions are primary? (in the sense that from them can be
assembled all others) has therefore a semi-conventional answer.
8 Dimensional analysis

So much for what might be called the “philosophical” aspects of our topic.
I turn now to discussion of its practical ramifications.

4. The principle of dimensional homogeneity. If, in the interest of concreteness,

we assume mass, length and time to be dimensionally primary then for any
physical variable X we find that measurement supplies a datum of the form
a b c
x (mass units) (length units) (time units) (12)

and to describe the evident dimensionality of X we write

[X] = M a Lb T c (13)

Note that the meaning of (13) is independent of any particular choice of mass/
length/time units, but that the value of the numeric x that assigns measured
value to X depends critically upon such choices. Suppose, for example, we were
to change units as follows:
⎫
mass units = A new mass units ⎪
⎬
length units = B new length units (14)
⎪
⎭
time units = C new time units

where A, B and C are dimensionless numerics. The measured value of X then

becomes
a b c
x̄ (new mass units) (new length units) (new time units)

with
x̄ = x Aa B b C c (15)

Suppose now that x1 , x2 , . . . , xn refer (relative to some prescribed system

of units) to the measured values of a set of variables4 X1 , X2 , . . . , Xn that enter
into the construction of some physical theory, and suppose further that

y = f (x1 , x2 , . . . , xn ) (16)

describes a physical relationship—a “formula”—encountered within that theory.

Write ⎫
[ Y ] = M a Lb T c ⎪
⎪
⎪
⎪
a1 b1 c1 ⎪
⎪
[X1 ] = M L T ⎪
⎪
a2 b2 c2
⎬
[X2 ] = M L T (17)
⎪
⎪
.. ⎪
⎪
. ⎪
⎪
⎪
⎪
an bn cn ⎭
[Xn ] = M L T
4
By “variables” I here mean “parameters, variables, derivatives of variables
with respect to variables, natural constants, etc.”
Simplest elements of dimensional analysis 9

It follows by (15) that a change of units (14) induces

⎫
y −→ ȳ = y Aa B b C c ≡ K y ⎪
⎪
⎪
⎪
x1 −→ x̄1 = x1 Aa1 B b1 C c1 ≡ K1 x1 ⎪
⎪
⎪
⎪
⎬
x2 −→ x̄2 = x2 Aa2 B b2 C c2 ≡ K2 x2 (18)
⎪
⎪
.. ⎪
⎪
. ⎪
⎪
⎪
⎪
⎭
xn −→ x̄n = xn Aan B bn C cn ≡ Kn xn

Equations (18) describe what happens when we exercise our option to modify
our metrological conventions. It is clear that statements—such as (16)—which
refer to world-structure must be stable under (18):

y = f (x1 , x2 , . . . , xn ) ⇐⇒ ȳ = f (x̄1 , x̄2 , . . . , x̄n )

This entails that

Kf (x1 , x2 , . . . , xn ) = f (K1 x1 , K2 x2 , . . . , Kn xn ) (19)

must have the status of an identity in the variables {x1 , x2 , . . . , xn , A, B, C }.

Not all functions f (•) have this property. Those which do are said to be
dimensionally homogeneous. The

principle of dimensional homogeneity asserts that only

functions f (•) that are dimensionally homogeneous can
figure properly in our physics.

The principle derives its power and importance from the facts that it restricts
the class of physically acceptable f (•)-functions, provides a valuable check on
the accuracy/plausibility of conjectured physical functions and often permits
one to guess the form of physical functions even in advance of the
development of a detailed theoretical account of the phenomenon in question.
In this respect the principle of dimensional homogeneity resembles (say) the
principle of Lorentz covariance: both speak formally to the transformational
properties of physically admissible equations. . . and both serve to tell us “where
the physics can’t be,” where it may plausibly be.

PROBLEM 1: Proceeding in the notation of (17), use (19) to show

a) that
y = x1 + x2 + . . . + xn
conforms to the principle of dimensional homogeneity if and only if

[Y ] = [X1 ] = [X2 ] = · · · = [Xn ]

which is to say: if and only if ai = a, bi = b, ci = c (all i).

b) that
y = x1k1 x2k2 · · · xnkn (20.1)
10 Dimensional analysis

conforms to the principle of dimensional homogeneity if and only if

⎫
a1 k1 + a2 k2 + · · · + an kn = a ⎪
⎬
b1 k1 + b2 k2 + · · · + bn kn = b (20.2)
⎪
⎭
c1 k1 + c2 k2 + · · · + cn kn = c

5. Simplest elements of dimensional analysis. Equations (20.2) might be notated

⎛ ⎞
⎛ ⎞ k1 ⎛ ⎞
a1 a2 . . . an ⎜ k2 ⎟ a
⎝ b1 b2 · · · bn ⎠ ⎜ . ⎟=⎝b⎠
⎝ .. ⎠ (21.1)
c1 c2 · · · cn c
kn
3×n matrix, call it M

or again ⎫
a· k = a ⎪
⎬
b· k = b (21.2)
⎪
⎭
c· k = c
where k , a, b, c are n-dimensional column vectors
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
k1 a1 b1 c1
⎜ k2 ⎟ ⎜ a2 ⎟ ⎜ b2 ⎟ ⎜ c2 ⎟
k≡⎜ ⎟
⎝ ... ⎠ , a≡⎜ ⎟
⎝ ... ⎠ , b≡⎜ ⎟
⎝ ... ⎠ , c≡⎜ ⎟
⎝ ... ⎠
kn an bn cn

and where the dot products have the familiar meaning:

a · k ≡ a1 k1 + a2 k2 + · · · + an kn

The linear system (21.1) can be solved

⎛ ⎞
a
k = M –1 ⎝ b ⎠ provided det M = 0
c

in the case n = 3, but

• is overdetermined (more equations than unknown k’s) if n < 3;
• is underdetermined (fewer equations than unknown k’s) if n > 3.

EXAMPLE: Suppose we were to conjecture that

force = (mass)k1 · (velocity)k2 · (acceleration)k3

Dimensionally we would have

M 1 L1 T −2 = (M 1 L0 T 0 )k1 (M 0 L1 T −1 )k2 (M 0 L1 T −2 )k3

Simplest elements of dimensional analysis 11

which in the notation of (21.1) becomes

⎛ ⎞⎛ ⎞ ⎛ ⎞
1 0 0 k1 1
⎝0 1 1 ⎠ ⎝ k2 ⎠ = ⎝ 1 ⎠
0 −1 −2 k3 −2

and by matrix inversion gives

⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
k1 1 0 0 1 1
⎝ k2 ⎠ = ⎝ 0 2 1⎠⎝ 1⎠ = ⎝0⎠
k3 0 −1 −1 −2 1

whence

force = (mass)1 · (velocity)0 · (acceleration)1

= (mass) · (acceleration)

A MORE INTERESTNG EXAMPLE: Suppose—with pendula on our

minds—we were to conjecture that

period = (mass of bob)k1 · (length of rod)k2 · (g)k3

Dimensionally we would have

M 0 L0 T 1 = (M 1 L0 T 0 )k1 (M 0 L1 T 0 )k2 (M 0 L1 T −2 )k3

which in the notation of (21.1) becomes

⎛ ⎞⎛ ⎞ ⎛ ⎞
1 0 0 k1 0
⎝0 1 1 ⎠ ⎝ k2 ⎠ = ⎝ 0 ⎠
0 0 −2 k3 1

and by matrix inversion gives

⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
k1 1 0 0 0 0
⎝ k2 ⎠ = ⎝ 0 1 1 ⎠⎝
2 0 ⎠ = ⎝ + 12 ⎠
k3 0 0 −2
1
1 − 12

whence

period = (mass)0 · (length)1 /(gravitational acceleration)1

Dimensional analysis has by itself informed us that (rather

surprisingly) the period of a pendulum is independent of the mass
of the bob. Detailed dynamical theory serves only to supply a
dimensionless factor of 2π.
12 Dimensional analysis

AN OVERDETERMINED SYSTEM : Suppose—with flapping flags in

mind—we were to conjecture that

period = (mass of flag)k1 · (wind speed)k2

Dimensionally we would have

M 0 L0 T 1 = (M 1 L0 T 0 )k1 (M 0 L1 T −1 )k2

In the notation of (21.1) this becomes

⎛ ⎞ ⎛ ⎞
1 0   0
⎝0 ⎠ k1
1 = ⎝0⎠
k2
0 −1 1

which quite clearly possesses no solution, the implication being that

our conjecture is untenable.
AN UNDERDETERMINED SYSTEM : Suppose our pendulum is subject
not only to gravitational but also to electrical forces. We conjecture
that

period = (mass)k1 · (length)k2 · (g)k3 · (electrical force)k4

Dimensionally we would have

M 0 L0 T 1 = (M 1 L0 T 0 )k1 (M 0 L1 T 0 )k2 (M 0 L1 T −2 )k3 (M 1 L1 T −2 )k4

which in the notation of (21.1) becomes

⎛ ⎞ ⎛ k1 ⎞ ⎛ ⎞
1 0 0 1 0
⎝0 ⎜k ⎟
1 1 1⎠⎝ 2 ⎠ = ⎝0⎠
k3
0 0 −2 −2 1
k4

This linear system can be written

⎛ ⎞⎛ ⎞ ⎛ ⎞
1 0 0 k1 −k4
⎝0 1 1 ⎠ ⎝ k2 ⎠ = ⎝ −k4 ⎠
0 0 −2 k3 1 + 2k4

and gives ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
k1 0 k4
⎝ k2 ⎠ = ⎝ + 12 ⎠ − ⎝ 0 ⎠
k3 − 12 k4
so we are led to a one-parameter family of possibilities:

period = (mass)−u · (length) 2 · (g)−( 2 +u) · (electrical force)u

1 1
Dimensionless products 13

In the absence of electrical force (i.e., at u = 0) we recover the

result developed on the preceding page. In the presence of such a
force we would have to look beyond dimensional analysis—to the
detailed dynamics of the system—to fix the value of u.

6. Dimensionless products. Clearly, one will have

[y ] = [y power ] if and only if y is dimensionless

And from this it follows that equations of (say) the form z = ey , and more
generally of the form
z = power series in y
are physically admissible (conform to the principle of dimensional homogeneity)
if and only if y(x1 , x2 , . . . , xn ) has been assembled dimensionlessly from the
variables/parameters/constants characteristic of the system under study. Thus
do we acquire special interest in dimensionless functions of the system variables.
That interest will soon be reenforced by quite another consideration.
It is evident that if c1 and c2 are dimensionless constants (things like 12
and π) and if y1 and y2 are dimensionless physical variables, then so are y1· y2
and c1 y1 + c2 y2 dimensionless: the set of all dimensionless constants/variables
is closed under both addition and multiplication. Elements of that set are
standardly denoted Π, and have the same numerical values in all systems of
units.
It follows directly from (21) that Π = x1k1 x2k2 · · · xnkn will be dimensionless
if and only if
⎛ ⎞
⎛ ⎞ k1 ⎛ ⎞
a1 a2 . . . an ⎜ k2 ⎟ 0
⎝ b1 b2 · · · bn ⎠ ⎜ . ⎟ = ⎝ 0 ⎠ (22.1)
⎝ .. ⎠
c1 c2 · · · cn 0
kn
3×n matrix, call it M

or more succinctly Mkk = 0 (22.2)

Equivalently ⎫
a· k = 0 ⎪
⎬
b· k = 0 (22.3)
⎪
⎭
c· k = 0

The vectors a, b, c, k live in an n -dimensional vector space where, according

to (22.3), k stands normal to the sub -space spanned by {a a, b, c }:

k ⊥ {a
a, b, c }

How many linearly independent such k -vectors are there? The obvious answer
is 
r is the dimension (1 or 2 or 3) of
p = n − r, where (23)
the sub -space spanned by {aa, b, c }
14 Dimensional analysis

SIMPLE PENDULUM REVISITED: This familiar system presents the

variables

period τ ≡ x1 : [x1 ] = M 0 L0 T 1
bob mass m ≡ x2 : [x2 ] = M 1 L0 T 0
rod length  ≡ x3 : [x3 ] = M 0 L1 T 0
gravitational acceleration g ≡ x4 : [x4 ] = M 0 L1 T −2

Equation (22.1) has become

⎛ ⎞ ⎛ k1 ⎞ ⎛ 0 ⎞
0 1 0 0
⎝0 ⎜k ⎟ ⎜0⎟
0 1 1 ⎠ ⎝ 2 ⎠ = ⎝ ⎠ , abbreviated Mkk = 0
k3 0
1 0 0 −2
k4 0

Mathematica provides a command NullSpace[rectangular matrix]

that provides a list of linearly independent solutions k 1 , k 2 , . . . , k p
of the equation Mkk = 0. In the present instance it supplies a single
solution: ⎛ ⎞
2
⎜ 0 ⎟
k=⎝ ⎠
−1
1
Thus are we led to the dimensionless construct

Π = τ 2 m0 −1 g 1

and from
Π = dimensionless constant
we recover 
τ∼ /g

CHARGED PENDULUM REVISITED: Some train of thought motivates

us to add

electric force E ≡ x5 : [x5 ] = M 1 L1 T −2

to our list of pendular variables. Equation (22.1) has now become

⎛
⎞ ⎛ ⎞
⎛ k1
⎞ 0
0 1 0 0 1 ⎜ k2 ⎟ ⎜ 0 ⎟
⎝0 ⎜ ⎟ ⎜ ⎟
0 1 1 1 ⎠ ⎜ k3 ⎟ = ⎜ 0 ⎟
⎝ ⎠ ⎝ ⎠
1 0 0 −2 −2 k4 0
k5 0
Dimensionless products 15

The NullSpace[rectangular matrix] command now provides two

linearly independent solutions of Mkk = 0 :
⎛ ⎞ ⎛
⎞
2 2
⎜ 0 ⎟ ⎜ −1 ⎟
⎜ ⎟ ⎜ ⎟
k 1 = ⎜ −1 ⎟ and k 2 = ⎜ −1 ⎟
⎝ ⎠ ⎝ ⎠
1 0
0 1

Thus are we led to the dimensionless constructs

Π1 = τ 2 m0 −1 g 1 E0
Π2 = τ 2 m−1 −1 g 0 E1

On page 12 we were led by a different line of argument to a one-

parameter family of dimensionless constructs

Π(u) = τ 2 m2u −1 g 2u+1 E−2u

It is interesting to note in this regard that

( Π1 )p ( Π2 )q = τ 2(p+q) m−q −(p+q) g p Eq

is a two -parameter family of dimensionless constructs that gives

back Π(u) when we set p = 1 + 2u and q = −2u.

GRAVITATIONAL PHYSICS OF A SPHERICAL MASS: We ask ourselves

What velocity, what acceleration are “natural” to the physics of
gravitating spherical mass m of radius R? From the following
material

velocity v ≡ x1 : [x1 ] = M 0 L1 T −1
acceleration a ≡ x2 : [x2 ] = M 0 L1 T −2
mass m ≡ x3 : [x3 ] = M 1 L0 T 0
radius R ≡ x4 : [x4 ] = M 0 L1 T 0
gravitational constant G ≡ x5 : [x5 ] = M −1 L3 T −2

we are led to write

⎛⎞ ⎛ ⎞
⎛ ⎞ k1 0
0 0 1 0 −1 ⎜ k2 ⎟ ⎜ 0 ⎟
⎝ 1 1 ⎜ ⎟ ⎜ ⎟
0 1 3 ⎠ ⎜ k3 ⎟ = ⎜ 0 ⎟
⎝ ⎠ ⎝ ⎠
−1 −2 0 0 −2 k4 0
k5 0

The NullSpace[rectangular matrix] command again provides two

16 Dimensional analysis

linearly independent vectors

⎛ ⎞ ⎛
⎞
−2 −4
⎜ 1 ⎟ ⎜ 1 ⎟
⎜ ⎟ ⎜ ⎟
k1 = ⎜ 0 ⎟ and k 2 = ⎜ 1 ⎟
⎝ ⎠ ⎝ ⎠
1 0
0 1
The null space is clearly closed with under linear combination, and
for the purposes at hand it proves more convenient to work with
⎛ ⎞ ⎛ ⎞
1 0
⎜ 0 ⎟ ⎜ 1 ⎟
⎜ ⎟ ⎜ ⎟
j 1 = 12 (kk 1 − k 2 ) = ⎜ − 12 ⎟ and j 2 = 2kk 1 − k 2 = ⎜ −1 ⎟
⎝ 1⎠ ⎝ ⎠
+2 2
−2 1
−1
which produce
Π1 = v 1 a0 m− 2 R 2 G− 2
1 1 1

Π2 = v 0 a1 m−1 R2 G−1
It follows that for such a system

characteristic velocity v ∼ Gm/R
characteristic acceleration a ∼ Gm/R2 ∼ v 2 /R

Physically the “characteristic velocity” shows up as the escape

velocity, also as the velocity of a satellite in low orbit . If we set
v = c then R ∼ Gm/c2 can be understood to refer to the radius
of a blackhole of mass m. The“characteristic acceleration” is more
familiar as g.
Clearly, it would be misguided to regard

y = xk11 xk22 · · · xknn and (say) z ≡ y 2

as independent constructs, for the value/dimension of z are latent in those of y :

z is redundant with y. Enlarging upon this remark: one says in general that
physical variables of the product structure

y1 = x1k11 x2k12 · · · xnk1n

y2 = x1k21 x2k22 · · · xnk2n
..
.
k k
yp = x1 p1 x2 2p · · · xnkpn

are “dimensionally dependent/independent” according as there do/don’t exist

exponents {h1 , h2 , . . . , hp }—not all of which vanish—such that

(y1 )h1 (y2 )h2 · · · (yp )hp is dimensionless

Buckingham’s theorem 17

—the idea here being that if the product were dimensionless one could write
(say)  
yp = (y1 )h1 (y2 )h2 · · · (yp−1 )hp−1 −hp
and thus render one of the yj —here taken to be yp —redundant. In some
cases it will be possible to continue the process. A quick argument serves
to establish that dimensional independence of {y1 , y2 , . . . , yp } implies and is
implied by the linear independence of the associated {kk 1 , k 2 , . . . k p }-vectors:
i.e., with the statement that there no not exist numbers {h1 , h2 , . . . , hp }—not
all of which vanish—such that

h1 k 1 + h 2 k 2 + · · · + hp k p = 0

Evidently we cannot expect ever to have p > n.

7. Buckingham’s theorem. Understand {x1 , x2 , . . . , xn } to refer as before to

an exhaustive list of the variables, parameters and dimensioned constants that
enter into the formulation of some physical theory, and let {Π1 , Π2 , . . . , Πp } refer
to some/any complete list of independent dimensionless products. Theoretical
statement of the form
y = f (x1 , x2 , . . . , xn ) (24.1)
are admissible if and only if they conform to the principle of dimensional
homogeneity. Statements of the design

Π = ϕ( Π1 , Π2 , . . . , Πp ) (24.2)

automatically conform to that principle (and are therefore invariant with respect
to adjustment of one’s system of units!). In  E. Buckingham established(or
at least conjectured)5 what is in effect the converse of the preceding
observation:

Buckingham’s Π-theorem : Every statement of the form

y = f (x1 , x2 , . . . , xn )

that conforms to the principle of dimensional homogeneity can be

written as a relationship among dimensional products:

Π = ϕ( Π1 , Π2 , . . . , Πp )

This is the upshot of what might be called the fundamental theorm of

dimensional analysis.
5
“On physically similar systems: Illustrations of the use of dimensional
analysis,” Phys. Rev. 4, 345 (1914). See also Lord Rayleigh, “The principle of
similitude,” Nature 95, 66 (1915). Careful discussion of the non-trivial proof
can be found in §24 of H. L. Langhaar, Dimensional Analysis and the Theory
of Models, ().
18 Dimensional analysis

The Π-theorem does tend to make familiar results look often a bit strange.
If we were concerned, for example, with the dynamics of a mass point we would,
instead of writing
F = ma
construct Π ≡ ma/F and write
Π=1
It should, however, be noted that ϕ(•) is always a function of fewer variables
than f (•), and that application of the theorem leads to relations that express
pure world-structure, free from the arbitrary conventions of unit selection.
When we state that β ≡ v/c = 0.62 we do not have to report whether lengths
have been measured in inches, furlongs or light years; time in seconds, hours or
weeks.

PROBLEM 2: Show by direct calculation that the vectors k 1 and k 2

that Mathematica supplied on page 14 do in fact lie within (and in
fact span) the null space of the rectangular matrix in question.

PROBLEM 3: Stars—which are held together by gravitational forces

—are found on close observation to “vibrate.” It seems plausible
that the vibrational frequency ν depends upon the stellar diameter
D, the mean stellar density ρ, and G. Use the method illustrated
on pages 13–15 to construct as many dimensionless products

Π(ν, D, ρ, G)

as you can, then use that information to deduce the necessary form
of ν = f (D, ρ, G). Why do starts of the same density vibrate with
the same frequency?

8. Some examples of physical interest . The examples considered thus far were
designed to illustrate points of computational principle and methodology, but
teach us little or nothing we did not already know about physics. Here I propose
to take the methodology pretty much for granted, and to concentrate on the
illustrative physics. Dimension analysis merits our attention precisely (but not
only) because it does have the power to teach us things we didn’t know, to
provide information that is of value particularly at earliest stages of any effort
to explore experimentally and to account theoretically for the phenomenon of
interest.

planck mass/length/time It was Max Planck who first thought to ask

what mass/length/time would enter most naturally into a (relativistic) quantum
theory of gravity, a theory into which c,
and G enter as characteristic constants.
Recalling that
[c ] = M 0 L1 T −1
[
] = M 1 L2 T −1
[G ] = M −1 L3 T −2
Examples of physical interest 19

we do have
[c k1
k2 Gk3 ] = M k2 −k3 Lk1 +2k2 +3k3 T −k1 −k2 −2k3
and want to have
= M 1 , else L1 , else T 1
In the first instance we write
    
0 1 −1 k1 1
 1 2 3   k2  =  0 
−1 −1 −2 k3 0

and by matrix inversion obtain

   
k1 +1/2
 k2  =  +1/2 
k3 −1/2

whence

Planck mass =
c/G = 2.177 × 10−5 g
= 1.302 × 10+19 proton masses

Similarly

Planck length = G
/c 3 = 1.616 × 10−33 cm

Planck time = G
/c 5 = 5.391 × 10−44 sec

And from these results we infer that (for example)

3
Planck density = c5 /G2
= 5.157 × 10+93 g/cm

Planck energy =
c 5 /G = 1.956 × 1016 erg

Dimensional analysis has by itself supplied the important insight that quantum
gravity, whatever shape such a theory might ultimately take, can be expected to
have things to say about the world only at space/time scales far, far smaller—
and at energies/densities far, far larger—than those encountered in existing
physics.
ball falling in viscous fluid When a spherical ball (liquid or solid, of
mass m and radius r) is dropped into a fluid (gaseous or liquid, of density ρ)
it achieves6 a terminal velocity v determined in part by the viscosity µ of the
fluid. We expect fluid dynamical analysis to supply a formula of the form

v = f (m, r, g, ρ, µ)

6
Unless it floats! That is, unless m < 43 πr 3 ρ. Evidently the Π2 encountered
below provides a dimensionless measure of bouyancy.
20 Dimensional analysis

Working from
[ v ] = M 0 L1 T −1
[ m ] = M 1 L0 T 0
[ r ] = M 0 L1 T 0
[ g ] = M 0 L1 T −2
[ ρ ] = M 1 L−3 T 0
[ η ] = M 1 L−1 T −1
and the requirement that v k1 mk2 rk3 g k4 ρk5 µk6 be dimensionless we are led (by
Mathematica) to the construction of

r3 ρ r2 η
Π1 = √v , Π2 = , Π3 =
gr m mv

The dimensionless triple

r3 ρ m2 g mgρ
Π1 , π1 ≡ Π2 = , π2 ≡ ( Π1 Π3 )−2 = =
m r3 η2 η2

REMARK: The final equality results from our using the

dimensionlessness of Π2 to write [ m/r3 ] = [ ρ ]. The
adjustment {Π2 , Π3 } → {π1 , π2 } can be considered to
be the result of our replacing {kk 2 , k 3 } by certain linear
combinations of those (null) vectors: we have been
led by our physical good sense to over-ride decisions
made by Mathematica.

better serves our immediate purpose, because it permits us to isolate v, writing

√
v = gr · ϕ(π1 , π2 )

A falling body experiences a velocity-dependent drag force Fdrag = Dv, and the
terminal velocity u ≡ vterminal is achieved when

weight = drag : mg = Du

Evidently
u = mg/D
while dimensional analysis has supplied a result that in the simplest instance
reads
√ r3 ρ a mgρ b
u = gr
m η2
= mb−a g b+ 2 r3a+ 2 ρa+b µ−2b
1 1

We are weakly motivated to set b−a = b+ 12 = 1; i.e., to set a = − 12 and b = + 12 ,

in which instance
mg
= (25)
ηr
Examples of physical interest 21

How does this square with the physical facts? In , George Stokes
showed by detailed fluid dynamical analysis that the drag of a sphere of radius
r moving with velocity v through a fluid with density ρ and viscosity η is (in
first approximation) given by

Fdrag = 6πηr · v (26)

which is the upshot of Stokes’ law. The terminal velocity of a falling sphere
would on this basis be given by

mg
uStokes = (27)
6πηr

which was anticipated already at (25).

The surprising ρ -independence of (25) is, as we have seen, not dimensionally
enforced, though it is, in some sense, “dimensionally plausible.”
We expect Stokes’ law to break down when v becomes too great (turbulence
and wave-generation become important, and bring additional variables and
parameters into play) or when r becomes so small that the sphere senses the
“granularity” of the fluid. Robert Millikan found that to make consistent good
sense of his “oil drop” data he had to work with an improved version of (27).7
One would expect

mg
u= 1 + power series in π1 and π2
6πηr

to provide a rational basis for such improvements.

REMARK: To obtain the dimensionality of viscosity one
has only to recall the procedure by which it is defined
(see, for example, §13–10 in D. C. Giancoli, Physics
for Scientists & Engineers (3rd edition, 2000)). One
should, however, become familiar with the exhaustive
compendium of such information that is tabulated near
the end of Section F in the Handbook of Chemistry and
Physics, and with the very clearly presented information
about many physical/mathematical topics that can be
found in Eric Weisstein’s World of Science, on the web
at
http://scienceworld.wolfram.com/

Under Physics see particularly Units & Dimensional

Analysis/Dimensionless Parameters.

7
See R. A. Millikan, Electrons (+ and −), Protons, Neutrons, Mesotrons
and Cosmic Rays (), pages 90–102.
22 Dimensional analysis

PROBLEM 4: Show by dimensional analysis that the centripetal

force F required to constrain a mass point m to a circular orbit is
proportional to m, proportional to the square of the velocity v, and
inversely proportional to the orbital radius r.

PROBLEM 5: The speed u of sound in gas depends upon the pressure

√ √
p and the mass density ρ. Show that u ∼ p and u ∼ 1/ ρ.

PROBLEM 6: Show that the frequency ω of any vibrational mode

of a liquid drop—under the action of surface tension σ (you will
need to know that [σ ] = [ energy/area] = M T −2 )—is proportional
to the square root of σ, inversely proportional to the square root
of the mass density ρ, and inversely proportional to the 32 power of
the diameter d.

PROBLEM 7: We conjecture that the height h of the tide caused by

steady wind blowing across a lake depends upon the mean length L
and depth D of the lake, the mass density ρ of the water, and the
shearing stress τ of the wind on the water (you will need to know
that [stress ] = [ force/area]). Show that

h ∼ D · f (L/D, τ /gρD)

Is h therefore independent of the speed v of the wind?

PROBLEM 8: If a drop of liquid falls into a pool a small column

of liquid splashes up. We expect the “splash height” h to depend
upon the mass m and impact velocity of the drop, the density ρ,
surface tension σ and viscosity η of the liquid, and the gravitational
acceleration g. Show that
1
h = (η 2 /gρ2 ) 3 · f (mgρ/η 2 , ρv 3 /ηg, ρσ 3 /gη 4 )

From the fact that f (•) has so many arguments we infer that this
would be a relatively difficult system to study, either experimentally
or theoretically.

9. What dimensions are primary, and how many are there? There have proceeded
thus far in the unexamined presumption that physical dimension can in every
instance be described

[ physical dimension] = M a Lb T c

and that the dimensionalities M , L and T of mass/length/time are conceptually

independent, irresolvable, “primary,” though it was remarked near the end of
§3 that an element of arbitrariness, of convention enters into this conception.
I propose to explore the matter now in somewhat greater detail, and begin by
expanding upon material sketched already on page 5.
Primary dimensions 23

By early in the 19th Century equations of the form

q1 q2 q q
F =k = 1 1 22
r2 4π. r
had entered the literature of physics as descriptions of the electrostatic
interaction of a pair of bodies bearing electric charges q1 and q2 . Here k
(equivalently .) is a phenomenological constant the value of which depends upon
the specific substance (oil, water, air, . . . ) in which the charges are immersed.
In vacuum (which was at the time a fairly rare and expensive “substance”)
q1 q2 q1 q2
F = k0 = 1
r2 4π.0 r2
The quantification of charge issued from the quantification of electrical current
(Coulomb = Ampere · second, in familiar practical units), and the quantification
of current was accomplished by an electrochemical procedure (measure the rate
at which silver is deposited in a standardized electroplating process). It seemed
natural to assign to electric charge its own independent dimension

[ charge] = Q

One then had

[k0 ] = M 1 L3 T −2 Q−2
= [k all material substances ] (28)

Eventually it became evident that one might advantageously look upon k0 as a

constant of Nature, and write

k material = (correction factor) · k0

With that realization it became natural to make adjustments (see again page 4)
so as to achieve

k0 = 1 (dimensionless)

This done, one or another of the heretofore “fundamental dimensions” M,L,T,Q

has been rendered reduntant with the others. Working from (28) we acquire8

Q = M 2 L 2 T −1
1 3

and Coulomb’s law becomes

q q
F = 1 22
r
↓
q q
= 1 1 2 2 in “rationalized” electrostatic units units
4π r

8
It is always easier—and, since it takes two to interact, often more natural—
to work with Q2 than with Q.
24 Dimensional analysis

That dimensional analysis leads—by different routes—to identical results

whether one considers {M, L, T, Q} or {M, L, T } to be “fundamental” is an
important point that I illustrate by
EXAMPLE: Working from

[electric potential V ] = [energy/charge]

[charge] = [d(current)/dt]
[resistance R ] = [potential/current]
[inductance L ] = [potential]/[d(current)/dt]
= [potential/charge]
[capacitance C ] = [charge/potential]

we obtain

[R ] = M 1 L2 T −1 Q−2 = M 0 L−1 T 0
[L ] = M 1 L2 T 0 Q−2 = M 0 L−1 T 1
[C ] = M −1 L−2 T 2 Q 2 = M 0 L+1 T −1

Were we to ask, within the {M, L, T, Q} system, for dimensionless

products constructable from R, L, C and a frequency ω we would
proceed from ω k1 R k2 Lk3 C k4 to
  
0 1 1 −1 k1
 0 2 2 −2   k2 
   = 0
−1 −1 0 −2 k3
0 −2 −2 2 k4

whence (with the assistance of Mathematica’s NullSpace[etc])

   
1 1
1  −1 
k1 =   , k2 =  
0 1
1 0

Were we to pose the same question within the {M, L, T } system we

would proceed from
   k1 
0 0 0 0
 0 k 
−1 −1 1 2  = 0
k3
−1 0 1 1
k4

and be led to the same set of k -vectors. By either procedure we

find

Π1 = ωRC, Π2 = ωL , Π3 = Π1 Π2 = ω 2 LC
R
Primary dimensions 25

Since it is the resistance R that accounts for the dissipation in an

RLC circuit, we expect on dimensional grounds alone to encounter

I(t) = I0 e−t/(RC) in a RC circuit

−t/(L/R)
I(t) = I0 e in a RL circuit
√
I(t) = I0 cos(t/ LC) in a LC circuit

Fundamental to thermodynamics is the concept of temperature which, prior

to the development of the kinetic theory of gases and of statistical mechanics,
appeared to have no relationship to any mechanical concept. It appeared
therefore to be natural/necessary to assign to temperature its own autonomous
dimension9
[ temperature T ] = θ
But statistical mechanics brought into focus the existence of a dimensioned
constant of Nature; namely Boltzmann’s constant k, with the dimension

[k ] = [ energy/temperature]

Because k is a universal constant (constant of Nature) we are free to use

T ≡ kT : [T ] = [mechanical energy] = M L2 T −2

rather than T itself to quantify temperature. The discovered existence of k has

here been used to contract the system of fudamental dimensions:

{M, L, T, θ} −−−−−→ {M, L, T }

k

Contemporary physics supplies also other universal constants; namely c,
,

the quantum of charge e, and Newton’s G, and those permit us to extend the
contraction process. For example: relativity recommends that we use

x0 ≡ ct : [x0 ] = [ length] = L

rather than t itself to quantify time. And the definition
/mc of the Compton
length recommends that we use

µ ≡ (c/
)m : [µ] = L−1

rather than m itself to quantify mass. At this point we have achieved

{M, L, T } −−−−−→ {M, L} −−−−−→ {L}

c

9
Do not confuse the temperature T with the dimension T of time.
26 Dimensional analysis

Relativistic quantum theory provides no “natural length,” so we have come here

to the end of the line.10

If, however, we were to bring G into play then we could use

length x
7≡  : [7 ] dimensionless
Planck length G
/c 5

rather than x itself to quantify length. Were we to adopt such a procedure

then all physical variables would be rendered dimensionless, and dimensional
considerations would (now in physics as standardly in mathematics) place
no constraint at all on statements of the form11

y = f (x1 , x2 , . . . , xn )

The principle of dimensional homogeneity (the Π-theorem (24.2)) has collapsed

into useless triviality.

If we abandon reference to G (on grounds that gravitational effects are

typically irrelevant to the physics of interest, and the Planck length so absurdly
small) then we would have

[ every physical variable] = L some power

and the argument presented on page 8 would simplify markedly: in place of

(14) we have simply

length units = B new length units

If [X] = Lb then the measured value x of X then becomes x̄ = xB b . In place

10
Notice that no use has been or will be made of e, for the interesting reason
that the
2
fine structure constant α ≡ e is dimensionless

c

so [e2 ] is redundant with {[
], [c ]}.

11
See again equations (16) = (24.1).
Primary dimensions 27

of (18) we have
y −→ ȳ = y B b ≡ K y
x1 −→ x̄1 = x1 B b1 ≡ K1 x1
x2 −→ x̄2 = x2 B b2 ≡ K2 x2
..
.
xn −→ x̄n = xn B bn ≡ Kn xn
The statement (19) of the principle of dimensional homogeneity can now be
notated
B b f (x1 , x2 , . . . xn ) = f (B b1 x1 , B b2 x2 , . . . , B bn xn )
If—following in the footsteps of Euler—we differentiate with respect to B and
then set B = 1 we obtain

b · f (x1 , x2 , . . . xn ) = b1 x1 ∂ + b2 x2 ∂ + · · · + bn xn ∂ f (x1 , x2 , . . . xn )
∂x1 ∂x2 ∂xn

And if, in particular, y = x1k1 x2k2 · · · x2k2 then we must have (compare (20.2))

b1 k1 + b2 k2 + · · · + bn kn = b

which imposes only a single condition on the exponents {k1 , k2 , . . . , kn }.

SIMPLE PENDULUM REVISITED ONCE AGAIN: Proceeding once

again from the conjecture that it is possible to write

period = (mass of bob)k1 · (length of rod)k2 · (g)k3

we have
L1 = L−k1 Lk2 L−k3
giving
1 + k1 − k2 + k3 = 0
whence—for all {k1 , k3 }—

period = (mass of bob)k1 · (length of rod)1+k1 +k3 · (g)k3

We do recover the familiar result if we set k1 = 0 and k3 = − 12 but

now have no particular reason to do so!

Evidently dimensional analysis becomes a weaker and ever weaker tool as the
size of the system of “fundamental dimensions” is progressively contracted .
Which is counterintuitive, for it is the advance of physics—the discovery of
fundamental relations12 and of universal constants—that fuels that contraction.
12
Think of F = ma, which declares measurements performed with spring
scales to be redundant with measurements performed with measuring rods and
clocks.
28 Dimensional analysis

10. Application of dimensional analysis to the design of models. The engineers

and physicists responsible for the design of large/complex/expensive structures
(boat hulls, air frames, bridges, telescope and accelerator components) often
study scaled-down dynamical models13 to gain confidence in and to fine-tune
their designs. It is, however, intuitively evident that cinematic footage of a
toy forest fire, or a toy storm at sea, will not look convincingly like the real
thing, but becomes somewhat more convincing when projected in slow motion.
It is not immediately obvious what real -world information can be inferred
from study of (say) the drag on toy models of ships. And it would clearly
be frivolous to contemplate construction of a functional ten-meter model of
the sun. Or of a galaxy. Or—to compound the absurdity—of a hydrogen
atom! Clearly, the design of informative models is subject to some severe
constraints-in-principle. Historically, it was a desire to clarify the source of
those constraints—to construct of an orderly “theory of models”—that served
as the primary motivation for the invention of dimensional analysis.14 I will
illustrate the points at issue by discussion of a realistic

drag on a yacht Our assignment is to determine how hull-

shape affects the drag on a 20-meter yacht. This we propose to do
by measuring the drag on one-meter models of similar shape and
mass distribution. We expect that detailed dynamical analysis (if
it could be carried out) would culminate in a formula of the form

drag D = f (v, 7, η, ρ, g)

where v refers to the yacht’s speed, 7 to its characteristic length,

η to the viscosity of water, ρ to the density of water, and g enters
because the yacht produces a wake that undulates energetically in
the earth’s gravitational field. Arguing as on page 20, we have

[ D ] = M 1 L1 T −2
[ v ] = M 0 L1 T −1
[ 7 ] = M 0 L1 T 0
[ η ] = M 1 L−1 T −1
[ ρ ] = M 1 L−3 T 0
[ g ] = M 0 L−1 T −2

and from the requirement that Dk1 v k2 7k3 η k4 ρk5 g k6 be dimensionless

13
Architects also construct static models to study æsthetic aspects of their
designs, which is quite a different matter.
14
This effort seems to have originated with Maxwell (“On the mathematical
classification of physical quantities,” Proc. London Math. Soc. 3, 224 (1871)),
but see also the papers by Lord Rayleigh and E. Buckingham cited previously.5
Principle of similitude 29

we are to write
 
k1
   k2   
1 0 0 1 1 0   0
 1 k 
1 1 −1 −3 1  3  = 0
k 
−2 −1 0 −1 0 −2  4  0
k5
k6
and thus (with the assistance again of Mathematica’s NullSpace
command) to the construction of dimensionless expressions
D
Π1 =
ρv 2 72
v7ρ
Π2 ≡ Reynolds’ number R =
η
2
Π3 ≡ Froud number F = v
g7
We expect therefore to have
D = ρv 2 72 · f (R, F )
The physical argument that leads from hull-shape to the specific
design of f (•, •), but we can proceed without that information on
strength of the assumption that the function in question pertains
15
to
 all hulls of the same shape, irrespective of size. Let numbers
D, v, 7, η, ρ, g refer to the yacht, and (in the same units) numbers
D  = KD · D
v  = Kv · v
7  = K · 7 : we have agreed to set K = 1
20
η  = Kη · η
ρ  = Kρ · ρ
g  = Kg · g
refer to the model. If the model is to mimic the behavior of the full
scale yacht then
KD
Π1  = · Π1 must = Π1
Kρ (Kv )2 (K )2
Kv K Kρ
R = ·R must = R
Kη
(Kv )2
F= ·F must = F
Kg K

15
This assumption cannot be maintained under all possible circumstances:
for miniature hulls we expect surface tension—of which we have taken no
account—to contribute importantly to drag.
30 Dimensional analysis

If both yacht and model float in water, and experience the same
gravity, then necessarily Kη = Kρ = Kg = 1, and we must have

KD = (Kv K )2 , Kv K  = 1 and Kv2 = K

which taken together imply Kv = K = KD = 1: the model must be

the same size as the yacht itself! But if the hull-shape is designed
to minimized the effect of viscosity then we can drop the second
condition, and obtain
1
Kv = (K ) 2
KD = (K ) 3
1
If K = 20 then the model should be propelled at 22% of the
intended speed of the yacht, and the measured drag will have to be
multiplied by 8000 to give the predicted drag on the yacht. If, on the
other hand, wake production is negligible (as it is for submarines)
then we can drop the third condition, and obtain

Kv = (K ) –1
KD = 1

The model, if propelled at 20 times the intended speed of the yacht

(submarine), will experience the same drag as the full-scale vessel
(but not really, for we have taken no account of turbulence).

models of hydrogen atoms A good approximation to the

physically correct theory of atoms arises when one looks quantum
mechanically (
) to the electromagnetic interaction (e2 ) of electrons
with one another and with the much more massive nucleus. In
the simplest instance one has the hydrogen atom, with its single
electron. Let R—call it the “Bohr radius”—refer to the natural
“atomic length.”. From R, m, e2 ,
one can assemble but a single
dimensionless expression; namely

Π = Rme2 /
2

One therefore expects to have

Bohr radius R =
2 ≈ 0.529189379 × 10−10 m

2

me

To make an enlarged functional model of the hydrogen atom

one has—since e2 and
are unalterable constants of Nature—no
alternative but to proceed m → m  = Km · m  m. But that is
impossible, since Nature provides no charged particle less massive
than an electron.
Biomechanical consequences of scale 31

One can draw several general lessons from the preceding examples:
• It is sometimes not possible to dilate variables/parameters in such a way
as to preserve the
  values of all the independent dimensionless expressions
Π1 , Π2 , . . . , Πq . One has then to abandon the least important of the Π’s,
and to make do with approximate models.
• The intrusion of natural constants (c, e2 , k,
, G,particle masses, mole
numbers) into the construction of Π1 , Π2 , . . . , Πq tends—since natural
constants are not susceptible to adjustment—to inhibit the construction
of models. This is a difficulty confronted more often by physicists (who for
this reason cannot expect to construct functional scale models of the sun)
than by engineers, whose equations seldom contain natural constants.
• The equations of interest to engineers do, on the other hand, make frequent
reference to the properties of materials (density, surface tension, elastic
moduli, conductivity, etc.) and in Nature’s Stockroom the variability of
those is in most instances severely limited. It would, for instance, be
frivolous to ask the model maker to use a material that is ten times denser
than iron, but one fifth as stiff!

11. Some illustrative biomechanical consequences of scale.16 Let 7 refer to the

characteristic length of an animal—an animal, let us say, with legs. Specifically,
we might set
 1
7 = volume of an equivalent mass of water 3
We are informed that the largest blue whale (the largest animal that has
ever lived, so far as is known) is about 1021 times more massive tahn the
smallest microbe, which suggests that 7max ≈ 107 7min . But the creatures at
those extremes both live (leglessly) in aquatic environments, and the smallest
swimmers inhabit a world that is—for reasons that I will have occasion to
explore17 —profoundly unlike that experienced by the terrestrial creatures to
which we have agreed to restrict our attention. For the latter we expect to have
something like
7terrestrial max ≈ 1 m ≈ 104 7terrestrial max

We expect the rate at which a terrestrial creature can unload metabolically

generated heat to be roughly proportional to its surface area; i.e., to go as 7 2 .
Relatedly, we expect the rate at which such a creature can take in oxygen to be
proportional to the surface area of its lungs: to go again as 7 2 . And we expect
its strength to be proportional to the cross sections of its bones and muscles:
7 2 again. For all these reasons we expect large creatures to be relatively less

16
The following discussion draws heavily upon material presented in
Chapter 1 of J. Maynard Smith, Mathematical Ideas in Biology () and
in the wonderfully detailed and beautifully illustrated monograph On Size and
Life, by Thomas A. McMahon & John Tyler Bonner ().
17
In the meantime, see E. M. Purcell’s classic “Life at small Reynolds
numbers,” AJP 45, 3 (1977).
32 Dimensional analysis

powerful, less strong. We expect smaller creatures to have relatively more

slender legs, lighter musculature. We are not surprised that ants can lift many
times their own weight, while horses can carry only a fraction of their weight.
Look to the maximal running speed of animals with legs. Working from
power = ( force) · (velocity)
and taking force to be determined by the maximal strength of bones and
muscles, we obtain
7 2 ∼ 7 2v
and conclude that in leading approximation all animals have the same top speed .
This surprising conclusion is borne out for animals sized between rabbits and
horses. The refined problem would be to explain why both elephants and bugs
run more slowly than cheetahs (though in terms of body length per unit time
many bugs are much faster than cheetahs!).
When running uphill an animal must do work agtainst the gravitational
field at a rate proportional to
(mass) · g · v ∼ 7 3 v
When set equal to the available power this gives
vmax uphill ∼ 1/v
This conclusion is supported by the facts: hills up which horses are obliged to
walk are hills up which little dogs are able to run.
Because the volume of the metabolic fuel containers (stomach and lungs)
increases as 7 3 while the metabolic rate of energy production (power) increases
less rapidly (as 7 2 ), we expect larger animals to be able to work longer between
meals, and larger aquatic mammals to be able to dive relatively deeper.
Look to Figure 1. The jumper invests energy W = (force) · d ∼ 7 2 · 7 1 in
executing its jump, and rises to a height given by h = W/mg ∼ 7 2 · 7 1 /7 3 = 7 0 .
Thus are we led to expect all animals to jump to roughly the same height, fleas
to be able to jump as high as dogs, dogs as high as kangaroos.
Many animals adjust their gait when they want to go faster, or faster still,
adopting modes of locomotion in which their feet spend progressively less time
on the ground, more time off the ground. Defining the “jumpiness” J by

J≡ air time
ground time
he is able to show that in leading approximation one might expect to have
ground time + air time
= 1 + J ∼ v 2 /7
ground time
which gets larger as the the speed v increases, smaller as the animal gets
larger, in qualitative conformity with the familiar facts (think of rabbits, human
runners, horses, elephants).
Biomechanical consequences of scale 33

Figure 1: The figure, read from left to right, shows a jumper in full
prepatory crouch; the jumper with legs fully extended; the jumper at
the top of its leap. The figure has been adapted from Figure 5 in
J. Maynard Smith’s little monograph.16

PROBLEM 9: Argue why it is that sufficiently small animals are

able to walk up walls and across the ceiling.

PROBLEM 10: Argue why it is that heart rate can be expected to

decrease with increasing size.

12. Life at small Reynolds number. We learned from Newton that—contrary to

what Aristotle imagined to be the case—force is required to accelerate objects
that move in empty space, but no force is required to maintain uniform motion.
But steady velocity-dependent force is required to maintain the steady motion
of an object in a fluid environment, even the absence of viscosity effects. For
this reason: massive fluid elements must be accelerated to “get out of the way,”
to create a co -moving cavity. In this respect, particle dynamics within a fluid
appears on its face to be more Aristotelian than Newtonian.
Thee viscosity-independent force in question is called the “inertial force,”
and by dimensional analysis we expect to have

Finertial = (geometrical factor) · (ρ 3 )(v 2 /) ∼ ρv 2  2

where ρ refers to the density of the fluid,  to a length characteristic of the

object, and the “geometrical factor” refers in a complicated way to its shape.
34 Dimensional analysis

Viscosity is defined by an operational procedure (see again pages 19 & 20)

that entails [η 2 ] = M 2 L−2 T −2 = (M L−3 )(M LT −2 ) = [density] · [force ] so we
expect to have
Fviscous = (geometrical factor) · η 2 /ρ
Therefore
Finertial ρ2 v 2  2
∼ = (Reynolds number R)2
Fviscous η2
which is to say18 
Finertial
R∼
Fviscous

Viscosity η and density ρ both refer to properties of the fluid. In practice

it often proves convenient to work instead with a single conflated parameter
called the “kinematic viscosity”

ν = η/ρ

which has the dimensionality of area/time. In this notation Reynolds’ number

becomes
R = v/ν
For water
νwater ≈ 10−2 cm2 /sec
For a swimming person we might reasonably set  ≈ 50 cm (the diameter of a
144 -pound sphere of water) and v ≈ 100 cm/sec, which would give

Rswimmer ≈ 5 × 10 5

For a guppy we might take  ≈ 5 × 10 −1 cm and v ≈ 10 cm/sec, giving

Rguppy ≈ 5 × 10 2

For a microbe it is, according to Purcell,17 reasonable to set  ≈ 10−4 cm (one

micron) and v ≈ 3 × 10−3 cm/sec, which give

Rmicrobe ≈ 3 × 10−5

If the force that is propelling an object through a viscous fluid is suddenly

turned off, and if also R  1, then we can on dimensional grounds expect the
object to come to rest with a characteristic

stopping time ∼ τ ≡  2 /ν = R/v

and during that time to travel a characteristic

stopping distance ∼ σ ≡ vτ = R

18
I cannot account for
√ the fact that many/most authors—including Purcell—
neglect to include the .
Life at small Reynolds number 35

Our microbe would coast only a few hundred thousandths of a body-length! To

share the swimming experience of a microbe you would have to swim in a pool
filled with a fluid 1010 times more viscous than water. It is intuitively evident
that whatever the “cyclically sequenced contortions” you might devise as a
propulsion strategy, it is unlikely that you would be moved to call that exercise
“swimming.” “Corkscrewing/snaking” would appear—intuitively—to be more
apt, and the microbes themselves seem to agree. Microbes—brainless though
they be—have evolved strategies that enable them to move beyond regious at
risk of becoming over-grazed, and to move beyond the expanding sphere of their
diffusing waste products. For discussion of these and other aspects of “life at
very small Reynolds number” I do urge my readers to have a look at Purcell’s
little article:17 it is easy to read, but highly informative.