Dimensional Analysis and Its

Applications in Statistics
WEIJIE SHEN
The Pennsylvania State University, University Park, PA 16802, USA

TIM DAVIS
We Predict Ltd., Technium 1, Kings Road, Swansea, SA1 8PH, UK

DENNIS K. J. LIN
The Pennsylvania State University, University Park, PA 16802, USA

CHRISTOPHER J. NACHTSHEIM
University of Minnesota, Minneapolis, MN 55455, USA

Dimensional analysis (DA) is a well-developed, widely-employed methodology in the physical and en-
gineering sciences. The application of dimensional analysis in statistics leads to three advantages: (1) the
reduction of the number of potential causal factors that we need to consider, (2) the analytical insights
into the relations among variables that it generates, and (3) the scalability of results. The formalization of
the dimensional-analysis method in statistical design and analysis gives a clear view of its generality and
overlooked significance. In this paper, we first provide general procedures for dimensional analysis prior to
statistical design and analysis. We illustrate the use of dimensional analysis with three practical examples.
In the first example, we demonstrate the basic dimensional-analysis process in connection with a study
of factors that a↵ect vehicle stopping distance. The second example integrates dimensional analysis into
the regression analysis of the pine tree data. In our third example, we show how dimensional analysis
can be used to develop a superior experimental design for the well-known paper helicopter experiment.
In the regression example and in the paper helicopter experiment, we compare results obtained via the
dimensional-analysis approach to those obtained via conventional approaches. From those, we demonstrate
the general properties of dimensional analysis from a statistical perspective and recommend its usage based
on its favorable performance.

Key Words: Buckingham’s ⇧ Theorem; Design of Experiment; Dimensions; Statistical Analysis.

1. Introduction Dr. Davis is the Director of Timdavis Consulting Ltd.,

Chief Technical Officer at We Predict Ltd., and Professor at

D IMENSIONAL ANALYSIS (DA) is a well-established

method in physics (see Sonin (2001), Szirtes
(2007)). Bridgman (1931) stated that “The principal
the University of Warwick. His email address is tim@timdavis
.co.uk.

Dr. Lin is a Distinguished Professor of Statistics and Sup-

use of dimensional analysis is to deduce from a study
ply Chain in the Department of Statistics, Pennsylvania State
of the dimensions of the variables in any physical sys- University. His email address is [email protected].
tem certain limitations on the form of any possible
Dr. Nachtsheim is a Professor, the Frank A. Donaldson
Chair of Operations Management, and Chair of the Supply
Mr. Shen is a Doctoral Student in the Department of Statis- Chain and Operations Department in the Carlson School of
tics, Pennsylvania State University. His email address is weijie Management, University of Minnesota. His email address is
[email protected]. [email protected].

Vol. 46, No. 3, July 2014 185 www.asq.org

186 WEIJIE SHEN ET AL.

relationship between those variables. The method is abstract spaces, whereas in statistics, dimension typ-
of great generality and mathematical simplicity”. It ically refers to the number of variables in a design
is mainly used to find the relations among physical problem or a data set. However, physical dimen-
quantities in complicated physical systems by their sions refer to the measurement systems to charac-
dimensions. A variety of literature has applied di- terize certain objects. Each physical dimension has
mensional analysis in various fields. See Asmussen several empirical scales of the measurements and
and Heebooll-Nielsen (1955), Islam and Lye (2009), they are called “units”. Ignoring nuclear e↵ects such
and Stahl (1962), for examples. Through these anal- as isospin, charm, and strangeness, there are seven
yses, some simple rules among those quantities can fundamental physical dimensions: namely, mass M,
be extracted. As a dimension-reduction and feature- length L, time T, temperature ⇥, electric current I
extraction methodology, dimensional analysis could (or charge Q), amount of substance mol, and lumi-
be of great use to the field of statistics. This a pri- nous intensity Iv . The corresponding units, defined
ori analysis gives us a conceptual and analytical view by SI (International System of Units), are kilogram,
of the problem we are dealing with, thereby provid- meter, second, kelvin, ampere, mole, and candela, re-
ing guidance in both the design and analysis steps. spectively. All other physical quantities are combina-
Furthermore, the physical origin of dimensional anal- tions of these fundamental quantities and their units
ysis improves the interpretability of the final results, are combinations of the units of the corresponding
which is particularly desirable to the fields of physics fundamental quantities, combined in the same way.
and engineering. For example, speed has the dimension of length per
time, for which the SI unit is meters per second.
Unfortunately, statisticians seem to have over-
looked the advantages of dimensional analysis. 2.2. Background of Dimensional Analysis
Finney (1977) commented that “I am surprised by
the lack of attention given to dimensions as a check Physical quantities cannot be constructed unre-
on the theory and practice of statistics. The basic strictedly. For example, it makes no sense to add
ideas, readily appreciated, should form part of the “length” to “mass” due to the natural constraints in
stock-in-trade of every statistician”. In this paper, the physical quantities. The main constraint is that
we focus on building functional relationships between “a physical law must be independent of the units
inputs and outputs. We will first introduce the basic used to measure the physical quantities”. This was
concept and general procedure of dimensional anal- first proposed by Joseph Fourier in the 19th cen-
ysis. Illustrated by two examples, the pine tree and tury (see Mason (1962)). This principle has been
paper helicopter, we show how to apply dimensional formalized in two important theorems, Bucking-
analysis in real problems and also compare the results ham’s ⇧-theorem (Buckingham (1914, 1915a,b)) and
with classic approaches. We summarize the proper- Bridgman’s principle of absolute significance of rela-
ties and discuss the advantages of dimensional anal- tive magnitude (Bridgman, 1931). Buckingham’s ⇧-
ysis in the statistical context. theorem shows that physical equations must be di-
mensionally homogeneous. In other words, any mean-
The rest of the paper is organized as follows. In ingful equations (and inequalities) must have the
the Section 2, we introduce the definitions and gen- same dimensions in both the left and right sides.
eral procedures of dimensional analysis with an il- Bridgman’s principle of absolute significance of rela-
lustrative example. In Section 3, we use dimensional tive magnitude shows that such formulae should be in
analysis for data analysis and show its generality and the power-law form. Basically, Bridgman’s principle
importance. In Section 4, dimensional analysis is ap- allows us to transform physical quantities properly,
plied to design of experiments. The last two sections especially into dimensionless forms. The method of
summarize general properties, followed by some con- using dimensionless quantities and Buckingham’s ⇧-
cluding remarks and prospective research. theorem to remove such constraints is called dimen-
sional analysis. Next, we introduce Buckingham’s ⇧-
2. Definitions of Dimensional Analysis theorem and how to use it in practice.
and General Procedure with
Illustrative Example 2.3. General Procedure
2.1. Physical Dimensions
We recommend applying dimensional analysis be-
In mathematics, dimension typically refers to the fore statistical analysis to give a general view of the
number of coordinates required to define points in problems and the variables involved. From the physi-

Journal of Quality Technology Vol. 46, No. 3, July 2014

 DIMENSIONAL ANALYSIS AND ITS APPLICATIONS IN STATISTICS 187

cal perspective, procedures can be found in Taylor et i = 1, . . . , p. If not, dimensional homogeneity is vio-
al. (2007) and others. From the statistical perspec- lated. This assumption leads to the existence of the
tive, the general procedure of dimensional analysis basis quantities but they are not unique. However,
can be specified as follows: the number of basis quantities is a fixed constant.
The concept of basis in linear algebra is a very good
Step 1. Determine the input and output variables
analogy to the concept of basis quantities.
and their dimensions, respectively.
Step 2. Determine the basis quantities. Step 3

Step 3. Transform input and output variables into Transform input and output variables into di-
dimensionless quantities by using basis mensionless quantities by using basis quantities. We
quantities in step 2. mainly transform variables that are not basis quan-
Step 4. Re-express the model functions via trans- tities, i.e., Q0 , Qt+1 , . . . , Qp , based on Buckingham’s
formed variables in step 3. ⇧-theorem. Due to the two properties of basis quan-
tities in step 2, we can have [Qi ] = [Q1 ]di1 · · ·
Step 1 [Qt ]dit , i = 0, t + 1, t + 2, . . . , p. Consequently, the
transformed dimensionless quantities are ⇧i = Qi ·
Determine the input and output variables of
Q1 di1 · · · Qt dit , i = 0, t + 1, t + 2, . . . , p, because
the system we consider. Denote input variables as
Q1 , . . . , Qp and the output variable (response) as Q0 . [⇧i ] = [Qi Q1 di1 · · · Qt dit ]
The conventional model will be Q0 = f (Q1 , . . . , Qp ), = [Q1 ]di1 · · · [Qt ]dit [Q1 ] di1
· · · [Qt ] dit
where f is the model function to be estimated. Note
= 1.
that, in dimensional analysis, Qi may include rel-
evant physical constants with dimensions, such as
gravitational constant = 6.67300 ⇥ 10 11 m3 kg 1 s 2 . Step 4
The units are often standardized to avoid dimension- Re-express the response functions. Before using di-
less multiplicative constants. But standardization is mensional analysis, we have Q0 = f (Q1 , . . . , Qt , Qt+1 ,
not always necessary because these constants are . . . , Qp ). Using ⇧i instead of Qi , we have the follow-
combined into unknown functional relationships. Af- ing expression:
ter checking the physical meaning of all the variables
we consider, we determine the relevant fundamen- ⇧0 Qd101 · · · Qdt 0t
tal physical quantities in the system of the seven SI d d
= f (Q1 , . . . , Qt , ⇧t+1 Q1t+1,1 · · · Qt t+1,t , . . . ,
units as shown in the previous section: denote them d d
⇧p Q1p1 · · · Qt pt )
as q1 , . . . , q7 . Further denote the dimensions of Qi as
[Qi ] and qj as [qj ] for i = 0, 1, . . . , p, j = 1, . . . , 7. We and
express the dimensions of Q0 , Q1 , . . . , Qp in terms
of [q1 ], . . . , [q7 ] as [Qi ] = [q1 ]ri1 · · · [q7 ]ri7 , for some ⇧0 = Q1 d01 · · · Qt d0t

d d
proper choices of {rij } with i = 0, 1, . . . , p, j = ⇥ f (Q1 , . . . , Qt , ⇧t+1 Q1t+1,1 · · · Qt t+1,t , . . . ,
1, . . . , 7. d d
⇧p Q1p1 · · · Qt pt ),
Step 2 where f is the function we hope to estimate. So we
Determine the basis quantities. The basis quanti- can rewrite it as
ties constitute a subset of the inputs. We reorder and ⇧0 = g(Q1 , . . . , Qt , ⇧t+1 , . . . , ⇧p ),
denote them as Q1 , . . . , Qt , where t  7 as discussed
above and t  p. The basis quantities should satisfy where ⇧i , i = 0, t + 1, . . . , p are dimensionless and
two conditions: (1) “Representativity”: the dimen- Q1 , . . . , Qt are “independent”. Buckingham’s theo-
sions of any other quantities, [Q0 ], [Qt+1 ], . . . , [Qp ] rem indicates that Q1 , . . . , Qt should not be in the
can be expressed by the combinations of the dimen- formula. This implies ⇧0 = g(⇧t+1 , . . . , ⇧p ) to be the
sions of the basis quantities, [Q1 ], . . . , [Qt ]. The com- final model.
binations take the form of power law. (2) “Indepen-
2.4. Example: Vehicle Stopping Distance
dence”: the dimension of any basis quantities cannot
be expressed by the combinations of the dimensions Here we use the vehicle stopping distance exam-
of other basis quantities. Furthermore, assume that ple to illustrate the general procedure. In the exper-
[Q0 ] can be expressed by the combinations of [Qi ], iment, we estimate the stopping distance for cars, a

Vol. 46, No. 3, July 2014 www.asq.org

188 WEIJIE SHEN ET AL.

key indication of their safety. Assume that the driver TABLE 1. Dimensions of Variables in
requires a certain amount of time for reaction to an the Vehicle Stopping Distance
emergency and that the wheels are not locked when
braking. We show the dimensional analysis below us- Dimension D v ⌧ m F µ
ing the procedure described in the previous section.
Step 1. Identify input and output variables and their L (length) 1 1 0 0 1 0
dimensions as follows: T (time) 0 1 1 0 2 0
M (mass) 0 0 0 1 1 0
Q0 = D : vehicle stopping distance [D] = L.
Q1 = v : velocity of the vehicle [v] = LT 1 .
Q2 = ⌧ : thinking time [⌧ ] = T. From the procedure and the example, we can see
that, in physical phenomena, we have certain restric-
Q3 = m : mass of the car [m] = M.
tions in the forms of f , satisfying certain dimensional
Q4 = F : braking force on the brake discs requirements. After dimensional analysis, the poten-
[F ] = MLT 2 . tial e↵ects on responses are attributable to the com-
Q5 = µ : friction coefficient of brakes [µ] = 1. binations of quantities considered. These quantities
act like groups. If we base our estimated function
The response is {D} and the predictors are g on the group values, we do not have dimensional
{v, ⌧, m, F, µ}. The model function is D = restrictions.
f (v, ⌧, m, F, µ). The dimensions of respective
variables are listed above and we summa- 3. Dimensional Analysis
rize corresponding {rij } in Table 1. There for Data Analysis
are three fundamental quantities in this sys-
tem. Their dimensions are length L, time T, Statistics extracts information from the data of
and mass M. The entries in the table are the experiments to find or justify properties, laws, and
power of the fundamental dimensions (de- performance. Based on the fact that those experi-
noted by rows) for dimensions of each vari- ments are results of physical phenomena and the data
able (denoted by columns). are physical quantities measured in experiments, it
is often justifiable and beneficial to perform dimen-
Step 2. Determine the basis quantities. We choose sional analysis in the first place. For data analysis,
{Q1 = v, Q2 = ⌧, Q3 = m} in this case. the advantages of using dimensional analysis stand
Step 3. Determine the dimensionless transformation out clearly: it is rather straightforward. Incorporat-
for the remaining three quantities {Q0 = D, ing dimensional analysis only transforms the data
Q4 = F , Q5 = µ}, and formulate {⇧0 , ⇧4 , in a predetermined fashion; the basic development
⇧5 } as follows: strategies remain the same. Furthermore, it is sup-
ported by physics to make dimensionless transfor-
[v⌧ ] = L, [⌧ ] = T, [m] = M, [µ] = 1
mations. We conjecture that making variables di-
and mensionally independent is helpful in making them
D F⌧ statistically independent, although this important is-
⇧0 = , ⇧4 = , ⇧5 = µ. sue needs further investigation. After dimensional
v⌧ mv
analysis, the information from inputs and outputs
Step 4. Re-express the model. Here our objective is more concentrated, leading to statistical models
will be to estimate the function g, with ⇧0 with fewer variables and simpler analysis. Below, we
as the response variable and only two input compare data-analysis procedures with and without
variables, ⇧4 and ⇧5 : dimensional analysis based on the pine tree data from
Bruce and Schumacher (1935, p. 226). From there, we
⇧0 = g(Q1 , Q2 , Q3 , ⇧4 , ⇧5 )
show how to perform dimensional analysis for data
= g(⇧4 , ⇧5 ), analysis and its potential benefits.
or equivalently, 3.1. Pine Tree Example
✓ ◆
D F⌧ The pine tree data has been used by various au-
=g ,µ .
v⌧ mv thors to illustrate the use of diagnostics and transfor-

Journal of Quality Technology Vol. 46, No. 3, July 2014

 DIMENSIONAL ANALYSIS AND ITS APPLICATIONS IN STATISTICS 189

mation methods in linear regression (see, e.g., Atkin-

son (1994)). The data arise from the measurements of
70 shortleaf pine trees. Three measurements of inter-
est here are d, the diameter of the tree in feet taken
at “breast height” above the ground; h, the height of
the tree in feet; and v, the volume in cubic feet. The
objective of the analysis is to establish a relationship
between the volume v and the variables d and h. In
other words, we hope to predict the volume of a tree
from its known diameter and height. The complete
data set is given in the Appendix. FIGURE 1. Studentized Residuals Plot of Model (1).

3.2. Regression Method Without Dimensional

Analysis for Pine Tree Data Set coefficients of diameter and height e↵ects (1.943 and
1.05) by the integers 2 and 1, respectively. After fix-
The conventional linear regression assumes that ing the coefficients, we obtain a regression model on
vi = ↵ + 1 di + 2 hi + ✏i , (1) the original scale without intercept:

with ✏i ⇠ i.i.d. N(0, ). It gives us the following es-

2
vi = d2i hi · i , (3)
timated function with standard errors of estimations
in subscripts: with ✏i = ln i ⇠ i.i.d. N(0, 2
). The estimated re-
gression function is
v̂ = 45.3(5.0) + 77.2(5.9) d + 0.12(0.11) h, (2)
with ˆ = 9.87. The coefficient of h is not signifi- v̂ = 0.4411(0.0081) d2 h. (4)
cant. Although a univariate analysis makes the resid- On the other hand, the Box–Cox transformation on
uals appear to be reasonably normally distributed the response variable recommends a transformation
with constant variance, we notice that (i) because parameter of ˆ = 0.384. Thus, a cubic-root trans-
the data (see Appendix) are ordered by diameter d, formation ( = 1/3 = 0.333) seems appropriate.
Figure 1 shows a distinct trend of residuals relative The use of the cubic-root transformation of the re-
to the diameters; and (ii) there is a potential outlier sponse (without intercept) suggests the following lin-
of tree #70 (the largest tree). Both of these diagnos- ear model:
tics strongly suggest that model (2) is inadequate.
We could proceed by applying the log transforma- 1/3
vi = 1 di + 2 hi + ✏i , (5)
tion to all variables before the linear model fitting,
which leads to the following result: with ✏i ⇠ i.i.d. N(0, 2 ). The estimated regression
function turns out to be
d =
ln(v) 1.06(0.24) + 1.943(0.038) ln(d)
+ 1.054(0.055) ln(h), vd
1/3 = 2.084
(0.044) d + 0.01471(0.00059) h. (6)

with ˆ = 0.0673. It is appealing to approximate the The residual plots in Figures 2a and 2b show that

FIGURE 2. Studentized Residuals Plot of Models (4) in (a) and (6) in (b).

Vol. 46, No. 3, July 2014 www.asq.org

190 WEIJIE SHEN ET AL.

both log transformation and Box–Cox transforma- TABLE 2. Dimensions of Quantities in Pine Tree Data Set
tion have fixed the problems highlighted in Figure 1.
Note that tree #53 deserves special attention, but Name Quantity Dimension Unit
will not be further studied in this paper.
The preceding analyses were, of course, conducted Volume v L3 feet3
without using dimensional analysis. However, a fur- Diameter d L feet
ther look at equations (4) and (6) reveals that both Height h L feet
methods provide dimensionally homogeneous solu-
tions to the prediction problem. The physical dimen-
sions are coherent in the equations. Model (4) has the
dimension of cubic length on both sides while model ˆ v = 0.4363(0.0036) ⇧2d ,
⇧
(6) has the dimension of length on both sides.
i.e.,
3.3. Regression Method with Dimensional v̂ = 0.4363(0.0036) d2 h. (8)
Analysis for Pine Tree Data Set
The result is similar to the log-transformed model (4)
Following the general procedure that we proposed and it also accommodates the problematic 70th case.
in Section 2.3, the dimensional analysis can be im- The di↵erences in parameter estimates can be at-
plemented as below. A similar approach was outlined tributed to di↵erences in associated error structures.
in Vignaux and Scott (1999).
In summary, certain assumptions regarding the
1. Our objective is to predict the output volume form of the function g will give corresponding pa-
v as a function of diameter d and height h: v = rameterizations and results. If ⇧v = k⇧d , then v =
f (d, h). We determine the physical dimensions kd h3 . If ⇧v = (A⇧d + B)3 , then v 1/3 = Ad + Bh.
of these quantities in Table 2. The dimension These are exactly the two models previously obtained
of v is cubic length L3 with unit feet3 . Both in equations (4) and (6). The dimensionally homoge-
dimensions of d and h are length L with units neous results we derived from regression analyses are
in feet. merely special cases of choosing di↵erent functions
2. Because the only dimension involved is length, g after dimensional analyses. The procedure ensures
let h be the basis quantity. that the results are dimensionally homogeneous and
3. Transform other quantities into dimensionless intuitively interpretable, while leaving choice of the
forms, function g to the investigator, as informed by the
data. It gives us a guide of how to model in an ef-
v d
⇧v = and ⇧d = . ficient and parsimonious way based on the physical
h3 h laws.
4. By Buckingham’s ⇧-theorem, the predicted
function should be ⇧v = g(⇧d ), or equivalently,
✓ ◆
v d
=g .
h3 h

Suppose ⇧v = g(⇧d ), and we choose g(⇧d ) = k⇧d .

After taking the logarithm of both sides, we obtain
the linear model,
ln(⇧v,i ) = ln(k) + ln(⇧d,i ) + ✏i ,
with ✏i ⇠ N (0, 2 ). The estimated regression func-
tion is
ln(⇧ˆ v ) = 1.07(0.16) + 1.942(0.036) ln(⇧d ) (7)
Alternatively, we might prefer that ⇧v = g(⇧d ) =
k⇧2d . Figure 3 shows the data and linear fits in terms FIGURE 3. Plot of Pine Tree Data Set in Terms of Di-
of ⇧v and ⇧2d . In fact, fitting a linear model yields mensionless Variables.

Journal of Quality Technology Vol. 46, No. 3, July 2014

 DIMENSIONAL ANALYSIS AND ITS APPLICATIONS IN STATISTICS 191

In contrast with the analysis and discussion of From a practical perspective, dimensional anal-
the same dataset by Atkinson (1994), dimensional ysis is applicable when modeling physical relation-
analysis points to an “automatic” transformation ships. It is a straightforward method before collect-
of the data, without prior assumption of the cone ing data and modeling. It fits all kinds of data struc-
shape, and posterior diagnosis of “transform both tures and modeling requirements. Furthermore, di-
sides model” as Atkinson (1994) did. Moreover, be- mensional analysis does not lose generality when
cause the dimensional analysis of each tree is the transforming the data. As shown above, it does not
same, there is less need to worry about individual constrain the forms of estimating functions for di-
trees influencing the choice of transformation, and mensionally homogeneous solutions. Due to its an-
hence regression methods such as constructed vari- alytical nature, the proposed procedure is easy to
able plots, which Atkinson (1994) used, are no longer implement.
necessary.
4. Dimensional Analysis for
3.4. Remarks the Design of Experiments
From an analytic perspective, dimensional analy-
Dimensional analysis can also serve as a guid-
sis o↵ers several advantages relative to the conven-
ance in the design of experiments. For the design
tional procedures. First, it decreases the number of
of experiments, incorporating dimensional analysis
variables, which may lead to a simpler model. Sec-
can significantly improve efficiency by reducing the
ond, physical independence may establish a simpler
number of experimental variables. In this section, we
statistical relationship. Third, it gives more sensi-
demonstrate the use of dimensional analysis in the
ble interpretations by having dimensionless variables
design and analysis of the popular paper helicopter
and coefficients. Physicists and engineers often fa-
experiment. We also compare the results to those ob-
vor dimensionless coefficients as indices for describ-
tained using the conventional design of experiments
ing systems. Fourth, dimensional analysis produces
approaches.
scalable results, which is necessary for extrapolation,
although the scalability depends on a good choice 4.1. The Paper Helicopter Experiment
of the model. This is attributed to the ratio form
of dimensionless variables that is invariant to scale The paper helicopter experiment is a widely used
changes. Often, extrapolating in the original scale re- teaching device for the design of experiments. The
sults in values interpolating in the transformed scale. objective is to predict the “flight time” performance
Fifth, it captures the inherent nonlinear relationship for a particular configuration of the helicopter dimen-
between physical quantities. sions in Figure 4. Upon its launch, twin blades spin

FIGURE 4. Paper Helicopter Illustration.

Vol. 46, No. 3, July 2014 www.asq.org

192 WEIJIE SHEN ET AL.

around the central ballast shaft to provide lift as it setting, they designed a full-factorial experiment in-
descends. It can be easily constructed from a sheet volving four important factors. Two key lessons from
of paper using only scissors and tape. As displayed their work, among others, are (1) with the help of
in Figure 4, the upper part of the model consists of response surface and steepest ascent, sequential de-
two wings (or rotors). The central part is the body signs search for an optimum point e↵ectively and ef-
and the lower part is the tail folded into the ballast. ficiently and (2) minimum variance or dispersion of
The length and width of each part can be varied to flight time is included in addition to longest flight
achieve di↵ering levels of performance. We wish to time to enrich the meaning of optimum. Annis (2005)
predict the flight time for a given configuration. The derived the aerodynamics of this flying object in a
usual design factors include the rotor radius (r), the rigorous physical sense before designing the experi-
rotor width (w), the tail length (l), the tail width ment. He presented a physical model of flight time in
(d), paper clip (p), and the tape (t). terms of the length and width of wing and body. He
employed two three-level factors in a single replicate
4.2. Conventional Design of the Paper of a 3 ⇥ 3 factorial design for wing length and width
Helicopter Experiment Without Using and employed response surface methods to identify
Dimensional Analysis the optimal operating condition. The use of physics
to identify promising factors in advance turned out
Johnson et al. (2006a, b) studied experimental de-
to be extremely advantageous.
sign on the paper helicopter as part of a Six Sigma
Black Belt project. They considered a Resolution VII Table 3 summarizes the results of the above three
design with seven two-level factors in a half fraction, papers, including Johnson et al., (2006a,b), Box and
with two replicates. This led to 27 1 ⇥2 (= 128) runs Liu (1999), and Annis (2005). The directions of
in total. They provided a step-by-step routine to de- respective e↵ects are provided in the parentheses.
sign experiments and maximize the flight time. Box Three variables are included in all of the three ex-
and Liu (1999) and Box (1999) discussed the use of periments; namely, body length, body width, and
sequential design in the paper helicopter experiment. wing length. Additionally, Johnson et al. (2006a, b)
First, they conducted a two-level Resolution IV frac- considered (i) paper type, (ii) whether or not tap-
tional factorial design with eight factors in four repli- ing body and wing, and (iii) whether or not clipping
cates, i.e., 28IV-4 ⇥4, for a total of 64 runs. In the second in the bottom. They also examined the two-way in-

TABLE 3. Summary of Literature on Paper Helicopter

Johnson et al. (2006a,b) Box and Liu (1999) Annis (2005)

Input variables Body length ( ) Body length ( ) Body length ( )*

Body width ( ) Body width ( ) Body width ( )*
Wing length (+) Wing length (+) Wing length (+)
Paper type ( ) Paper type ( ) Wing width (dip)
Taped body ( ) Taped body ( )
Taped wing ( ) Taped wing ( )
Clip ( ) Clip ( )
Interactions Fold ( )

Design -1 ⇥ 2 = 128
27VII 28IV-4 ⇥ 4 = 64 32 = 9

Feature Interactions Sequential learning Physical insight

Significant variables Body length ( ) Body length ( ) Body length ( )

Body width ( ) Body width ( ) Body width ( )
Wing length (+) Wing length (+) Wing length (+)

*Not a design factor.

Journal of Quality Technology Vol. 46, No. 3, July 2014

 DIMENSIONAL ANALYSIS AND ITS APPLICATIONS IN STATISTICS 193

teractions between the variables of interest. Box and two. Following Gearhart (2004), we define
Liu (1999) considered whether to fold the paper and the two dimensionless variables as v =
use an additional 50 runs to search for the optimum. vra ⇢b g c and m = mrd ⇢e g f . Dimensions of
Their second design included a full-factorial exper- v and m are, thus, as follows:
iment on both the length and width of both body
[ v]= (LT 1 )(L)a (ML 3 )b (LT )
2 c
and wing. Annis (2005) took wing width as an ad-
ditional variable and chose wing length and width = L1+a 3b+c Mb T 1 2c
to be experimental variables and considered e↵ects [ m ] = (ML)d (ML 3 )e (LT 2 )f
of body length and width as given by the physical = Ld 3e+f M1+e T 2f .
formula. The “dip” e↵ect of wing width means the
relationship is not monotone. In general, the signs in We enforce nondimensionality and solve the
the parentheses indicate the e↵ects of each variable two sets of linear equations,
on the flight time. For example, wing length has a
1+a 3b + c = 0
positive e↵ect on the flight time, meaning that the
flight time will increase if wing length is increased. b=0
All the above studies concluded that the wing length 1 2c = 0
is the most important factor for determining flight
and
time, with factors a↵ecting helicopter mass also be-
ing important. d 3e + f = 0
4.3. Design of the Paper Helicopter 1+e=0
Experiment Using Dimensional Analysis 2f = 0.
The physics of falling objects in a gravitational From the first set, we obtain a = c = 1/2
field follows the following assumptions: (1) the flight and b = 0 and, from the second, d = 3, e =
time (T ) is determined by the launch height (H) and 1, and f = 0. The transformed variables
average velocity (v), i.e., T = H/v; (2) the falling ob- are thus
ject reaches terminal velocity quickly after dropped, v h m
when the drag force of the air becomes equal to the v = p = p ; m = . (10)
gr T gr ⇢r3
force of gravity; (3) the drag force depends on the
density of dry air (⇢ = 1.20412 kgm 1 at sea level at Step 4. The final equation is obtained as the follow-
20 C), drag coefficient (cd , dimensionless) and the ing form:
shape of the helicopter (rotor radius r and rotor v = g( m ). (11)
width w, or their combinations, such as ratio r/w and
Because there is only one input variable
area rw); (4) the weight of the helicopter depends on
( m ), we conduct the paper helicopter fly-
the mass (m) and acceleration due to gravity (g = 9.8
ing experiment with four runs: m = 0.937,
ms 2 ).
2.087, 3.088, and 4.642. The resulting flight
Suppose the flight time follows the model below, times (T) are 5.18, 3.87, 3.48, and 2.98, re-
T = f1 (m, g, r, cd , ⇢, H). spectively. This implies the values of v to
be 0.873, 1.264, 1.537, and 1.795. The results
Because cd is dimensionless and T = H/v, the ex-
pression can be represented as
TABLE 4. Dimensions of Quantities of Paper Helicopter
v = f2 (m, g, ⇢, r). (9)
Quantity
Next, we apply dimensional analysis step-by-step.
Quantity name symbol Dimension Unit
Step 1. Table 4 displays the dimensions and units
of the variables. We have three fundamental Velocity v = h/T LT 1
ms 1
dimensions, length L, time T, and mass M. Mass m M kg
Step 2. Variables r, ⇢, and g are chosen as the base Gravity acceleration g LT 2
ms 2
quantities. Air density ⇢ ML 3
kgm 3
Step 3. From Buckingham’s ⇧ theorem, we can re- Wing length r L m
duce the number of variables from five to

Vol. 46, No. 3, July 2014 www.asq.org

194 WEIJIE SHEN ET AL.

TABLE 5. Table of Design and Results of Paper Helicopter Experiment

m Paper Helicopter Rotor Flight v

p
No. (m/⇢r3 ) type mass(m) radius(r) time(T ) (h/(T gr))

1 0.937 80 g/m2 3.09 g 140 mm 5.18 s 0.873

2 2.087 120 g/m2 4.34 g 120 mm 3.87 s 1.264
3 3.088 100 g/m2 3.72 g 100 mm 3.48 s 1.537
4 4.642 160 g/m2 5.59 g 100 mm 2.98 s 1.795

Note: The results are the averages of three flights recorded independently twice.

are displayed in Table 5. Figure 5 is the scat- 5. For validation of the final model, we conducted
ter plot of v and m . eight confirmation runs with various combinations
of 80/100/120/160 g/m2 A4 paper and 100/120/140
The data are modeled using a simple linear regres- mm rotor radius. The actual flight times versus pre-
sion, giving ˆ v = 0.700.09 + 0.250.03 m . Converting dicted flight times are displayed in Figure 6. The
to the original variables, this becomes points align closely to the line y = x, indicating a
h good model indeed.
T̂ = p 0.25 . (12)
gr(0.70 + ⇢r3 ) Table 6 provides a comparison of results between
designs using dimensional analysis and those without
Alternatively, we regress data on a log scale. This it. It can be seen that the key factors are exactly the
gives us log( v ) = 0.1020.01 +0.460.01 log( m ). The same, while the settings for maximal time are close.
power of 0.46 suggests a square-root transformation However, the design we used had only one dimen-
of m . We thus take a square-root transformation on sionless variable; therefore, fewer runs were needed (4
m and fit a linear pmodel without intercept, obtain- runs vs. 128 runs in Johnson et al. (2006b), 64 runs in
ing ˆ v = 0.8590.014 m . Converting to the original Box and Liu (1999), and 9 runs in Annis (2005)). In
variables, this can be expressed as addition, the prediction model is elegant and easy
r to interpret: the response is proportional to some
hr ⇢
T̂ = . (113)
0.859 mg

In both models (12) and (13), the coefficients are

dimensionless and the equations are dimensionally
homogeneous. We prefer model (13) because it is able
to capture the potential curvature shown in Figure

FIGURE 5. Plot of Simple Linear Regression on Paper FIGURE 6. The Plot of Predicted Flight Time and Actual
Helicopter. Flight Time of the Confirmation Runs.

Journal of Quality Technology Vol. 46, No. 3, July 2014

 DIMENSIONAL ANALYSIS AND ITS APPLICATIONS IN STATISTICS 195

TABLE 6. Comparison Between Designs with and Without DA

Without DA With DA

Variables 2 or 3 Levels Continuous (interpolate and extrapolate)

Design On variables On dimensionless quantities

Number of experimental runs 128, 64 and 9 4

Key variables Wing length (+) Wing length (+)

Body length ( ) Body length ( )

Settings for maximal flight time L = 15.2c m, w = 4.60 cm L = 14 cm, w = 5.25 cm

m = A4 (gsm unstated) m = 3.09 g (80 gsm)

Validation model Full factorial Confirmation runs

power of each factor. The ratio form also produces 1. Dimensional analysis starts from basic and nat-
scalable results. Furthermore, after transformation, ural physical assumptions. The resulting fac-
the dimensionless quantities seem to be linearly re- tors and their coefficients are dimensionless and
lated to the response, which avoids the complicated easy to interpret for practitioners.
relationships such as “dip” e↵ect found in Box and
2. Dimensional analysis combines and eliminates
Liu (1999) and Annis (2005). Last, robustness analy-
unnecessary variables using Buckingham’s ⇧
sis through transmitted variation (Morrison (1957))
Theorem. This leads to dimension reduction,
can be directly applied to the fitted model, avoiding
which is especially helpful for the design of ex-
costly replications to determine this variation empir-
periments.
ically. Further results on a dimensional-analysis ap-
proach to the paper helicopter can be found in Davis 3. The power law is used in the combination, re-
(2013). vealing that the nature of relationships between
variables with di↵erent dimensions is often not
4.4. Remarks linear. It is believed that, after transformation,
From the design of experiments perspective, the the dimensionless quantities may become more
number of experimental runs required tends to in- independent (fewer interactions) and their re-
crease with the number of experimental factors. Di- lationships with response are simpler, as shown
mensional analysis combines variables according to in our examples.
physical laws and creates new design variables that
can be incorporated into the design. It reduces the 4. It is compatible with all kinds of methods, as it
number of factors and consequently reduces the re- is a “data-free” method. It transforms the vari-
quired number of runs. It potentially allows the in- ables according to their dimensions, not values.
clusion of variables not included in the original ex- It can be done even before we get the real data
periment, as long as they can be expressed in the di- or any pilot experiment. The subsequent statis-
mensionless variables. We also benefit from the scal- tical procedures are valid without changes.
ability and the interpretability of the solutions. See, 5. The resulting models are often scalable. For ex-
for example, Albrecht et al. (2013) for a detailed ample, extrapolation in linear regression is of-
treatment of designing experiments to dimensional- ten misleading, because the assumed form of
analysis models. the model may not be valid beyond the range
of the data. The dimensionless models devel-
5. General Properties oped using dimensional analysis do not depend
From the above examples, we summarize the fol- on absolute quantities, but are rather defined
lowing advantages for dimensional analysis on gen- in terms of relative amounts; thus, scale is not
eral cases. relevant in most cases.

Vol. 46, No. 3, July 2014 www.asq.org

196 WEIJIE SHEN ET AL.

Drawbacks of dimensional analysis may include the interweave the knowledge of engineers and that of
requirement of physical knowledge about the exper- statisticians. Special designs or analyses may be pre-
imental environment and the possibility of severe ferred after conducting dimensional analysis. Fourth,
problems if any related variable is excluded (Albrecht we believe dimensional analysis could be generalized
et al. (2013)). Piepel (2013) also raises the issue of into fields outside of physics and engineering. Certain
spurious correlation. A comprehensive discussion of common measure units in economics, biology, or soci-
statistical issues of dimensional analysis is presented ology could be candidates to enlarge the application
in Lin and Shen (2013). of dimensional analysis. Fifth, it seems promising to
generalize the idea of combining variables. PCA is
Conclusion one kind of combining under linear schemes in feature
extraction. Dimensional analysis implies that combi-
Dimensional analysis has been well developed in
nations may be done nonlinearly by power law.
physics, engineering, and other fields. However, its
significance was overlooked for years by statisticians. References
Little e↵ort was made to incorporate it into statisti-
cal practice. In this paper, we describe the use of the Albrecht, M. C.; Nachtsheim, C. J.; Albrecht, T. A.; and
Cook, R. D. (2013). “Experimental Design for Engineering
dimensional-analysis method for both data analysis Dimensional Analysis”. Technometrics 55(3), pp. 257–270.
and design of experiments. Additional examples and Annis, D. H. (2005). “Rethinking the Paper Helicopter: Com-
comments can be found in Davis (2011) and Lin and bining Statistical and Engineering Knowledge”. The Amer-
Shen (2013). Our purpose is to promote greater inte- ican Statistician 59(4), pp. 320–326.
gration of dimensional analysis into statistical design Asmussen, E. and Heebooll-Nielsen, K. (1955). “A Di-
mensional Analysis of Physical Performance and Growth in
and analysis by statisticians, engineers, and scien-
Boys”. Journal of Applied Physiology 7(6), pp. 593–603.
tists. Atkinson, A. C. (1994). “Transforming Both Sides of a Tree”.
The fundamental insight of dimensional analysis The American Statistician 48(4), pp. 307–313.
Box, G. E. P. (1999). “Statistics as a Catalyst to Learning by
is to identify key dimensionless variables from physi- Scientific Method Part II—A Discussion”. Journal of Quality
cal considerations, and then use data and statistical Technology 31(1), pp. 16–29.
analysis to understand them. Engineers provide the- Box, G. E. P. and Liu, P. Y. T. (1999). “Statistics as a Cata-
ory for guidance in the statistical analysis using phys- lyst to Learning by Scientific Method Part I—An Example”.
ical prior knowledge. Statisticians design and ana- Journal of Quality Technology 31(1), pp. 1–15.
Bridgman, P. (1931). Dimensional Analysis, 2nd edition.
lyze experiments for the unknown physical structure,
Yale University Press.
check the validity of physical assumptions and rec- Bruce, D. and Schumacher, F. X. (1935). Forest Mensura-
ommend further experiments. Complementing each tion, p. 226. New York, NY: McGraw Hill.
other in this fashion often leads to solutions that nei- Buckingham, E. (1914). “On Physically Similar Systems; Il-
ther could achieve alone. lustrations of the Use of Dimensional Equations”. Physical
Review 4(4), pp. 345–376.
This paper introduces the basic idea of dimen- Buckingham, E. (1915a). “Model Experiments and the Forms
sional analysis and its potential applications in statis- of Empirical Equations”. Transactions of the American So-
tics, notably in regression analysis and the design ciety of Mechanical Engineers 37, pp. 263–296.
Buckingham, E. (1915b). “The Principle of Similitude”. Na-
of experiments. There are many more issues to be
ture 96(2406), pp. 396–397.
studied. First, the error structure should be fur- Davis, T. P. (2011). “Dimensional Analysis in Experimental
ther investigated for dimensional analysis. Latent Design”. Talk at Isaac Newton Institute. Available online at
errors in covariates and the robustness need to be http://sms.cam.ac.uk/media/1191742/.
taken into account as well. For example, an orthog- Davis, T. P. (2013). “Comment: Dimensional Analysis in Sta-
onal distance regression can be applied when both tistical Engineering”. Technometrics 55(3), pp. 271–274.
Finney, D. J. (1977). “Dimensions of Statistics”. Applied
sides of the modeling equation have errors. More- Statistics 26(3), pp. 285–289.
over, errors could propagate very di↵erently for dif- Gearhart, C. (2004). “Using Dimensional Analysis to Build
ferent nonunique DA representations. Second, once a Better Transfer Function”. SAE Technical Paper, 2004–
designing the transformed dimensionless quantities, 01–1129.
the corresponding design on the original quantities is Islam, M. and Lye, L. M. (2009). “Combined Use of Di-
mensional Analysis and Statistical Design of Experiment
not unique. The various design options on those op-
Methodologies in Hydrodynamics Experiments”. Ocean En-
erating quantities o↵er a way to test the validity of gineering 36, pp. 237–247.
Buckingham’s ⇧-theorem statistically. Third, a for- Johnson, J. A.; Widener, S.; Gitlow, H.; and Popvich, E.
mal sequential or recursive scheme is suggested to (2006a). “A ‘Six Sigma’ Black Belt Case Study: G.E.P. Box’s

Journal of Quality Technology Vol. 46, No. 3, July 2014

 DIMENSIONAL ANALYSIS AND ITS APPLICATIONS IN STATISTICS 197

Appendix
The Shortleaf Pine Tree Data Set

ID Diameter Height Volume ID Diameter Height Volume

(feet) (feet) (feet3 ) (feet) (feet) (feet3 )

1 0.383 33 2.2 36 0.917 71 25.8

2 0.367 38 2.0 37 0.925 81 32.8
3 0.417 40 3.0 38 0.933 91 35.4
4 0.425 49 4.3 39 0.958 66 26.0
5 0.425 37 3.0 40 0.975 65 29.0
6 0.433 41 2.9 41 1.000 72 30.2
7 0.433 41 3.5 42 1.017 66 28.2
8 0.458 39 3.4 43 1.017 72 32.4
9 0.458 50 5.0 44 1.042 90 41.3
10 0.467 69 7.2 45 1.075 88 45.2
11 0.492 58 6.4 46 1.083 63 31.5
12 0.492 50 5.6 47 1.092 69 37.8
13 0.625 45 7.7 48 1.092 65 31.6
14 0.633 51 10.3 49 1.117 73 43.1
15 0.633 49 8.0 50 1.150 69 36.5
16 0.650 59 12.1 51 1.150 77 43.3
17 0.667 56 11.1 52 1.192 64 41.3
18 0.675 86 16.8 53 1.192 77 58.9
19 0.700 59 13.6 54 1.217 91 65.6
20 0.717 78 16.6 55 1.233 90 59.3
21 0.742 93 20.2 56 1.242 68 41.4
22 0.758 65 17.0 57 1.258 96 61.5
23 0.767 67 17.7 58 1.267 91 66.7
24 0.775 76 19.4 59 1.267 97 68.2
25 0.775 64 17.1 60 1.275 95 73.2
26 0.817 71 23.9 61 1.283 89 65.9
27 0.825 72 22.0 62 1.308 73 55.5
28 0.825 79 23.1 63 1.325 99 73.6
29 0.825 69 22.6 64 1.333 90 65.9
30 0.842 71 22.0 65 1.400 90 71.4
31 0.850 80 27.0 66 1.483 91 80.2
32 0.850 82 27.0 67 1.525 96 93.8
33 0.858 81 27.4 68 1.525 100 97.9
34 0.867 75 25.2 69 1.617 94 107.0
35 0.883 75 25.5 70 1.950 104 163.5

Vol. 46, No. 3, July 2014 www.asq.org

198 WEIJIE SHEN ET AL.

Paper Helicopter Experiment. Part A”. Quality Engineering Sonin, A. A. (2001). The Physical Basis of Dimensional
18, pp. 413–430. Analysis, 2nd edition. Cambridge, MA: Department of Me-
Johnson, J. A.; Widener, S.; Gitlow, H.; and Popvich, E. chanical Engineering, Massachusetts Institute of Technology.
(2006b). “A ‘Six Sigma’ Black Belt Case Study: G.E.P. Box’s web.mit.edu/2.25/www/pdf/DA unified.pdf.
Paper Helicopter Experiment. Part B”. Quality Engineering Stahl, W. R. (1962). “Similarity and Dimensional Methods
18, pp. 431–442. in Biology”. Science 137(3525), pp. 205–212.
Lin, D. K. J. and Shen, W. (2013). “Comments: Some Sta- Szirtes, T. (2007). Applied Dimensional Analysis and Mod-
tistical Concerns on Dimensional Analysis”. Technometrics eling, 2nd edition. Oxford, UK: Elsevier Butterworth-
55(3), pp. 281–285. Heinemann.
Mason, S. F. (1962). A History of the Sciences. New York, Taylor, M.; Diaz, A. I.; Jodar-Sanchez, L. A.; and
NY: Collier Books. Villanueva-Mico, R. J. (2007). “100 Years of Dimensional
Morrison, S. J. (1957). “The Study of Variability in Engi- Analysis: New Steps toward Empirical Law Deduction”.
neering Design”. Applied Statistics 6(2), pp. 133–138. arXiv: Physics 0709.3584v3.
Piepel, G. F. (2013). “Comment: Spurious Correlation and Vignaux, G. A. and Scott, J. L. (1999). “Simplifying Re-
Other Observations on Experimental Design for Engineering gression Models Using Dimensional Analysis”. The Austral-
Dimensional Analysis”. Technometrics 55(3), pp. 286–289. ian and New Zealand Journal of Statistics 41(1), pp. 31–42.

Journal of Quality Technology Vol. 46, No. 3, July 2014