Seven is the Magic Number in Nature

THOMAS SAATY
Distinguished University Professor
University of Pittsburgh

Joshua and Israel marched around Jericho seven times while seven
priests blew seven trumpets before the walls came crashing down.
(Joshua 6:3-4)

Introduction

Where there is structure, the parts of the structure must function together
with a degree of consistency and purpose. Specifically, I am thinking of
dynamic systems in which there are action and reaction among the parts
and their functions and also friction and resistance. Natural systems,
such as the cells in our bodies, and man-made systems, such as a watch,
are constructed in a hierarchic way so that the different parts in each
level work together consistently—that is, each group performs a function
to fulfill some purpose. Thus, the number of functions working together
determines the structure through which materials or energy pass. The
number of functions that can work together is determined by the consis-
tency of the interactions of these functions. Conversely, consistency
among the functions depends on the number of interacting components;
if there is a large number, the possibility of inconsistency is greater. How
large should the number of functions be to fulfill a purpose? The answer
given here has important implications for constructing both physical and
social systems. The current paper shows with mathematics supported by
examples that 7 to 8 seem to be the maximum number for any compo-
nent of a complex system.
A system consists of a structure, flows in the structure, functions or
actions that the flows perform, and a purpose for the system to fulfill.
There can be multiple flows, functions, and purposes served. For
example, to survive the human body must perform a few interacting
functions through its flows, such as circulating blood, breathing,
digesting, reproducing, sending hormones, firing nerves, moving
muscles, obtaining support from bones, and relying on integumentary
parts (e.g., hairs, nails). The last two or three serve to support the func-
tions of the other organs and are fairly independent of them.

PROCEEDINGS OF THE AMERICAN PHILOSOPHICAL SOCIETY VOL. 160, NO. 4, DECEMBER 2016

[       335     ]
336 thomas saaty

The functions themselves are a synthesis of lesser functions; diges-

tion involves chewing, tasting, swallowing, secreting chemicals,
breaking down complex sugars into simple sugars and proteins into
amino acids, and emulsifying fats, absorbing the nutrients of the food
we eat, and excreting the waste. These sub-function themselves can
each be broken down to lesser sub-functions. Thus, the structure of any
system needs to be broken down hierarchically into modules to facili-
tate the flows in that system and their functions. Modularity is a general
principle for managing complexity. By breaking down a complex
system into discrete pieces—which can then communicate with one
another only through standardized interfaces within a standardized
architecture—one can eliminate what would otherwise be an unman-
ageable tangle of system-wide interconnections.
The functions interact and depend on each other—each one of
them is important for the maintenance and survival of the other
functions. However, for a system or subsystem to survive, there cannot
be an excessive number of functions. Such an idea is not new in the
literature of technological design (Simon, 1962).1 The aforementioned
theory is thought to have been operating as a law of nature from the
beginning even if, as some claim (Baldwin and Clark, 1997), modu-
larity is becoming more important today because of the increased
complexity of modern technology. We can apply the idea of modularity
not only to technological design but also to social organizations.
The structure of a system is designed to accommodate certain flows
that pass through it. Subsystems of the system have different functions
that interact, which lead to the fulfillment of the overall purpose. The
functions must therefore work together (i.e., be interdependent and
conjoint and give feedback) to achieve the purpose. When one or more
functions are faulty, the purpose the system is designed to serve fails in
different degrees. The functions can take place sequentially, conjointly, or
in combination. When they are sequential, as in a relay race, there is no
problem in being consistent (except perhaps in handing the baton). The
important question to be examined in the current paper is how consis-
tently interdependent functions combine to achieve the desired purpose.
For the Nobel Laureate Herbert Simon,1 a complex system is:
. . . one made up of a large number of parts that interact in a
non-simple way. In such systems, the whole is more than the sum
of the parts, at least in the important pragmatic sense that, given
the properties of the parts and the laws of their interaction, it is not
a trivial matter to infer the properties of the whole.
seven is the magic number in nature 337

Simon also says “complexity is both a matter of the sheer number

of distinct parts the system comprises and of the nature of the
interconnectedness among those parts.”
Looking at it differently, however, modularity has an even longer
pedigree in the social sciences. We can think of the “obvious and simple
system of natural liberty” in Adam Smith’s Wealth of Nations (1776),
where he showed that a complex modern society with its social and
economic institutions needs modular design to be more productive.2, 3
A hierarchy is one of two ways to structure a system that is
composed of interrelated subsystems that are each hierarchic in turn.
The other way to structure a system is as a network. In formal organi-
zations, the number of subordinates who report directly to a single
boss is called his or her “span of control.” Analogously, the span of a
system is the number of subsystems into which it is partitioned. Simon1
says that a hierarchic system is flat at a given level if it has a wide span
at that level. A diamond has a wide span at the crystal level but not at
the next level down (i.e., the molecular level).
One important difference exists between physical and biological
hierarchies, on the one hand, and social hierarchies, on the other. Most
physical and biological hierarchies are described in spatial terms. We
detect the organelles in a cell in the way we detect the raisins in a
cake—they are “visibly” differentiated substructures localized spatially
within the larger structure. In social hierarchies, one considers who
interacts with whom, not who lives next to whom. The width of span
in a hierarchic system is of concern in this paper.
We are not thinking of “dead” parts, such as the wires in circuits that
conduct electricity to destinations. There can be millions of them. Also
we are not thinking about collections of objects arranged in orderly ways
to form a structure. We are thinking about objects that are dynamic and
function together according to natural or manmade forces that act to
fulfill a purpose. None of the parts can function well or at all without the
presence of the others, as, for example, in the case of a car’s cylinders or
a clock’s wheels, and within natural organisms (i.e., the parts or organ-
elles of a living cell that need one another to survive). In the case of
organelles, interactions are not mechanically direct but rather act through
chemistry and a medium, the cytoplasm. The organs of our body use the
circulatory system and the blood supported by the materials they produce
to help nurture each other and the rest of the body. They all work
together and influence each other—they are interdependent in performing
their function. The effect on the organism may take a longer time to
manifest these influences, good or bad; their influence may take a shorter
338 thomas saaty

time to be felt and noticed. If we stop breathing, it can be the end of us

because of the lack of oxygen as the heart stops pumping blood to the
brain and other organs.
Underlying this interdependence and feedback is the degree of
consistency or harmony in the interaction of the functions. Consistency
in the workings of the parts of the system determines the degree of
stability of the system. Inconsistency can lead to instability and to the
system faltering and ceasing to function. Inconsistency varies in inten-
sity from extremely inconsistent, to randomly inconsistent, to moder-
ately inconsistent, and, finally, to perfectly consistent. There can be
measurements associated with the degree of inconsistency with which
any system of multiple parts and functions is operating. It is possible
that there could also be underlying simple laws of form, which a
rational mind might apprehend to explain complexity.
The philosopher, Arthur Schopenhauer said, “Every truth is the refer-
ence of a judgment to something outside it, and intrinsic truth is a contra-
diction.” Since there are no absolutes, comparisons must be used, which
inevitably lead to judgments and the possibility of inconsistency because
of the subjectivity and variability of judgments. When we deal with
intangible factors, which by definition have no scales of measurement,
we can compare them in pairs according to the dominance of one over
another with respect to a common property. We can not only determine
the preferred object but also discriminate among intensities of prefer-
ence. When we compare functions, each working in its own domain, we
can compare how well a function is doing with how well it was doing
before. The common property in such a comparison is how well it fulfills
its purpose. But when we compare two functions, what is the common
property? They may have very different purposes. The common property
needs to be, as also affirmed by Simon, some emergent entity that comes
from their interaction, “contributing to maintaining synchronous
timing?” for example. Another possibility to compare functions pairwise
within a given hierarchical level could be: “With respect to the higher
purpose (a node/function in the above higher level) which function better
defines, and to what extent, this higher purpose?”

Why Consistency is Essential for the Workings of any System

The Oxford English dictionary defines inconsistency as a “want of

agreement or harmony between two things or different parts of the
same thing.” Webster’s dictionary defines consistency as “agreement or
harmony in parts or of different things.” This definition is the common-
sense view of consistency, but there is also a mathematical version of
consistency derived by considering the elements of the system in pairs.
seven is the magic number in nature 339

Consistency forms the basis of causal thinking, but it also applies to the
workings of things, as the dictionary says. The insightful Julian Huxley4
wrote that “something like the human mind might exist even in lifeless
matter.” Herms Romijn5 has published a substantial paper in which he
argues persuasively that photons have consciousness. The article
suggests that photons carry subjectivity or consciousness as a given
property, which is possible in principle because irreducible properties
(nothing is smaller than a photon) are present at this level. He argues
that it is more reasonable than the current approach, which suggests
that the new property of consciousness can be produced by banging
together previously unconscious bits of matter.
It is with the consistent interaction of functions that the purpose is
fulfilled. If the functions are inconsistent, the purpose is less perfectly
satisfied. The question is: What should the number of functions be, and
what level of inconsistency can the purpose tolerate before it begins to
show signs of deterioration?
To be consistent is not to lead to contradictions. This definition is
independent of time. When a system is dynamic and depends on time,
the foregoing definition of consistency involves time in a different
manner. Is there consistency or harmony among the parts of the system
so they continue to work together? How bad can inconsistency be? If
we are close to consistency, we expect that the system will continue to
function well. That closeness to consistency is sufficient because no
system is perfectly consistent.
To say that A is twice as heavy as B and B is 3 times as heavy as C and
conclude that A is 6 times heavier than C is a consistent way of thinking.
If one were to conclude that A is 5 times as heavy as C, one would think it
is not as wrong as saying A is 100 times as heavy as C. Consistency in
language means that reasoning does not lead to contradictory outcomes,
and this example is a mathematical way to express that idea.

More about Consistency and Inconsistency in Science,

Mathematics, and Engineering

The idea of consistency, with some exceptions, is not used much in

philosophy or mathematics. One speaks of the consistency of a set of
axioms in that they do not produce contradictory results. When a set of
equations are all satisfied by at least one set of values for the variables,
they are said to be consistent. If they are not all satisfied by any one set
of values for the variables, they are said to be inconsistent. We also
assume that the real world is consistent, and it is our job to describe it
in a consistent way. But even in physics, it does not always happen; the
340 thomas saaty

theory of relativity and quantum theory have not been reconciled in a

consistent way.
When we say that the cylinder of an engine is inconsistent in its
function with its intended design, we mean that one cylinder is not
functioning as closely to its design as other cylinders may be. If we
compare the relative inconsistency of these cylinders with the intended
design, we would say that this cylinder behaves “equally as well,” “a
little better,” or “strongly better” than the other. In the end, we can
obtain a measure of the priorities of the cylinders according to their
consistency with their design.
This example can be generalized to any system, and we can use the
same measure of consistency. We note that it is easier to compare the
cylinders among themselves for the degree of consistency because we
can observe them. This approach is clearer than comparing each with
the design ideal that it is assumed to fulfill, since one may have little
information about the design and its implementation. By comparing
the cylinders with one another, despite their inconsistent functioning,
we can seek better performance of the system of cylinders. It is clear
that the more consistent cylinders tend to compensate for the inconsis-
tency of the less consistent ones. This kind of interdependence is what
we are referring to. In addition, many inconsistent cylinders can cause
the good cylinders to become less consistent in their attempt to improve
performance. It should now be clear that if the number of cylinders is
large and many of them are inconsistent, the compensation of the other
cylinders becomes less sensitive, and the system now gradually slows
down in attaining its purpose, which is the reasoning behind the
measurement of inconsistency. Of course, the number of cylinders can
also be small but inconsistent. However in this case, one can identify
the most inconsistent cylinder and attempt to repair it. Many inconsis-
tent cylinders cause “overcompensation” by the consistent cylinders,
which can wear out/decrease performance.
The same ideas apply in social situations. The members of a jury
can be considered to reason with the facts. However, some of them are
better at drawing conclusions of “guilty” or “not guilty” from these
facts. When we compare them, we find that some jurors seem to be
more inconsistent than others. A large number of jurors prevents us
from determining which of the jurors are inconsistent in their treat-
ment of the facts. In biology, the internal organs of the body depend on
each other’s functioning for their survival; the more organs there are,
the more difficult it becomes for them to compensate for inconsistency
in other organs. If they compensate strongly, the system will become
seven is the magic number in nature 341

dysfunctional because they respond only with their own chemistry and
material to balance the system.
Let us now examine in greater depth the idea of consistency and
inconsistency in our minds. Because inconsistency also occurs in nature
according to the definition given earlier, the ideas developed below can
also be generalized to inconsistency in nature. A good example of
inconsistency, in the world of sports, is that team A beats team B, team
B beats team C, but team C beats team A. Here, inconsistency is a
natural occurrence and not a mental aberration. It is surprising that we
create axioms for economics that preclude such intransitivity, which
abounds in people’s expression of many different kinds of preference. It
is easy to identify inconsistency in simple situations such as the one
noted above: if A is 3 times more important than B and B is 2 times
more important than C, then A should be 6 times more important than
C, not 5 times. Here, the logical approach works well, but if there are
many things to compare, it becomes difficult to be perfectly consistent
in making judgments. Thus how many elements there are influences
tracking the consistency of the system, whether in our mind’s thinking
or in physical systems of the natural world.
To emphasize the point, interacting with consistency means to
work together in harmony, agreement, or concord to fulfill a purpose.
There is a cogent, logical, and mathematical reason to decompose any
complex system with interactive parts. The clearest way to deconstruct
the system is in a hierarchical fashion, breaking it down into small
groups or levels of homogeneous parts.6 These parts must be “similar,”
or more precisely be of the same “order of magnitude,” to work
together consistently.

Hierarchic Decomposition of a System: The Role of

Modularity to Allow Different Flows to Serve Different
Functions

Simon1 introduces the topic of evolution of a complex system with a

parable. There once were two watchmakers named Hora and Tempus
who manufactured very fine watches. Both of them were highly
regarded, and the phones in their workshops rang frequently; new
customers were constantly calling. However, Hora prospered whereas
Tempus became poorer and poorer and finally lost his shop. What was
the reason? The watches the men made consisted of about 1,000 parts
each. Tempus had constructed his so that if he had one partly assembled
and had to put it down (to answer the phone, for example) it immedi-
ately fell to pieces and had to be rebuilt. The more the customers liked
342 thomas saaty

his watches, the more they phoned him. Therefore, it became difficult
for him to find enough uninterrupted time to finish a watch.
The watches that Hora made were no less complex than those of
Tempus. But he had designed them so that he could put together
subassemblies of about 10 elements each. Ten of these subassemblies,
again, could be put together into a larger subassembly; and a system of
10 of the latter subassemblies constituted the whole watch. Hence,
when Hora had to put down a partly assembled watch to answer the
phone, he lost only a small part of his work and he assembled his
watches in only a fraction of the man-hours it took Tempus.
It is rather easy to make a quantitative analysis of the relative diffi-
culty of the tasks of Tempus and Hora. In the end, however, what
makes Tempus’s unfinished watches so unstable is not the sheer number
of distinct parts involved; Rather, it is the interdependency among the
parts in his design that cause the watches to fall apart. In a
non-decomposable system, the successful operation of any given part is
likely to depend on the characteristics of many other parts throughout
the system. So when such a system is missing parts (because it is not
finished, for example, or because some of the parts are damaged), the
whole ceases to function.
By contrast, in a decomposable system, the proper working of a
given part will depend highly on the characteristics and consistency of
other parts within its subassembly—but it will also depend on the
characteristics and consistency of parts outside of that subassembly. As
a result, a decomposable system may be able to limp along even if some
subsystems are damaged or incomplete. In organizational and social
systems—and even in mechanical ones as well—it is possible to think
of interdependence and interaction among the parts as a matter of
information transmission or communication. The flows in the system
are the means of communication, whether mental or physical, like
opening a car door with a remote control. Communication of informa-
tion requires consistency with existing knowledge or, more abstractly,
with the function of an existing structure. Consistency is the most
important criterion in building information. The test for consistency is
comparative. In fact, no absolute method for testing the consistency of
a set of assumptions has ever been found. For a mathematical defini-
tion of consistency in a hierarchy see Saaty (2010).7
A good example of a decomposable system is the central nervous
system made of nerve cells and their interactions. It is made up of the
brain, the spinal cord, and the peripheral nervous system. The brain is
made of three main parts: the forebrain, midbrain, and hindbrain (Figure
1). The forebrain consists of the cerebrum (little brain), associated with
higher brain function such as thought and action; the thalamus; and the
seven is the magic number in nature 343

Figure 1. A simplified hierarchic model of the brain.

hypothalamus (part of the limbic system or “emotional brain”). The

midbrain consists of the tectum and the tegmentum. The hindbrain is
made of the cerebellum, pons, and medulla. Often, the midbrain, pons,
and medulla are referred to together as the brainstem, which is respon-
sible for basic vital life functions such as breathing, heartbeat, and blood
pressure. The spinal cord is made of a bundle of nerves running up and
down the spine. The peripheral nervous system consists of the nerves and
ganglia outside of the brain and spinal cord. Its function is to connect the
central nervous system to the limbs and organs. It is divided into the
somatic nervous system and the autonomic nervous system, including
sensory systems. Most of the 12 cranial nerves are part of it. Figure 1
shows that we have a complex hierarchic structure that again needs
overall measurement of the consistency of its functions, which is
developed later in the current paper.

How to Measure Inconsistency

What is known in the theory of measurement6 as the Fundamental

Scale is used to make paired comparison judgments. This scale can be
derived from the logarithmic stimulus-response function of Weber-
Fechner in psychophysics.7
In a less mathematical vein, we are able to distinguish between
high, medium, and low at one level, and for each of them in a second
level below, we also are able to distinguish between high, medium, and
344 thomas saaty

low for each of the three giving us nine different categories. We assign
the value 1 to (low, low), which is the smallest, and the value 9 to (high,
high). Doing so allows us to cover the spectrum of possibilities between
two elements and to give the value nine for the top of the paired
comparisons scale compared with the lowest value on the scale (valida-
tion examples of this scale appear later). The mathematician and cogni-
tive neuropsychologist Stanislas Dehaene8 writes: “Introspection
suggests that we can mentally represent the meaning of numbers 1
through 9 with actual acuity. Indeed, these symbols seem equivalent to
us. They all seem equally easy to work with, and we feel that we can
add or compare any two digits in a small and fixed amount of time like
a computer.” The fundamental scale in measurement theory comprises
the numbers 1 to 9.
In making paired comparisons, numbers are assigned to pairs of
elements using judgment about dominance. The elements being compared
must be homogeneous, requiring no greater number than 9. An element
compared with itself with respect to a certain criterion is always equal to
1. Therefore, the main diagonal entries of the pairwise comparison
matrix are all 1. The numbers 3, 5, 7, and 9 correspond to the verbal
judgments “moderately more dominant,” “strongly more dominant,”
“very strongly more dominant,” and “extremely more dominant,” with
2, 4, 6, and 8 between the previous values. Reciprocal values are auto-
matically entered in the transpose position. We are permitted to interpo-
late values between the integers, if desired, or use numbers from an
actual ratio scale of measurement. The decision-making theory known as
the Analytic Hierarchy Process (AHP) uses the integers 1 to 9 as its
Fundamental Scale of Absolute Numbers corresponding to the afore-
mentioned verbal statements for the comparisons.
The ordinary way of measuring tangibles uses scales that have a
unit to measure and assigns a number to each object one at a time. In
the paired comparisons process, a number is assigned not to the objects
but to the relation of dominance between two objects or functions at a
time. A priority scale is then derived from all the dominance measure-
ments for the objects. In this manner, we are able to both derive priori-
ties for tangibles, and for intangibles for which there is no measurement.
Although numbers obtained by using a scale are permanent and are
always the same, priorities are only useful for the problem at hand (i.e.,
the group of objects being compared) and need not be the same for
another problem with the members of the group changed.
Let A = (aij ) be an n-by-n positive reciprocal matrix, so all aii = 1
and aij = 1/aji , for all i, j 1,…,n. Let w = ( w1 ,..., wn ) be the principal
 seven is the magic number in nature 345

right eigenvector of A, which captures the priority of “domi-

nating” for each element in the group, and let v = (v1 ,..., vn ) be the

principal left eigenvector of A, , which captures the priority of

“being dominated.” Now if a dominance matrix is consistent, then one

can write its entries as ratios of its priority vector entries so that
aij = wi / w j .
Because practice AA isis usually
Because inin practice Because
usually inin practice
inconsistent,
Because inconsistent, wewe Awrite
practice isis usually
write
A usually inconsistent,
inconsistent, . .we
MM
w
Because in practice A is usually inconsistent, we write
Because in usually A is usually
practiceinconsistent, inconsistent, we write
e A isA usually
actice Because
is usually in practice
inconsistent,
inconsistent, we Awe is write
write we .write . Moreover,
Moreover, have, and
. Moreover,
s usually
stent, inconsistent, we write .. Moreover,
we write , ,
and
and
Moreover, substituting
substituting
. Moreover, , ,we
for
for wehave ,, andsubstituting
substituting
and . . forfor , w ,
isA, have
and substituting for , and substituting
, we have for , we have .
, and,Because
Because
andBecause in
in for
substituting
substituting
substituting
practice
practice
in for
for weAA,,have
,a
practice is
we
we
usually
usually
is
have
inconsistent,
inconsistent,
usually . inconsistent,
.
.
we
we we write
writewrite . .. Moreov
Moreo
. Mo
, andpractice
orin substituting
, we have A isforusually , we. inconsistent,
have ij we . write . Moreover,
The
Thecomputation:
computation: The
Thecomputation:
computation:
ractice AA is
practice is usually
usually Theinconsistent,
inconsistent,
The computation:
computation: we ,, write
we
The computation: and
and, substituting
write substituting
and substitutingfor
for for,, we
we,.have
.have
Moreover,
weMoreover,
have ..
.
n: , and substituting for , we have .

,, and
and substituting
substituting for for ,, we
we havehave ..
The
TheThe computation:
computation:
computation:
putation: reveals
revealsthat
reveals that
that . .Moreover,
Moreover,since
Moreover, reveals
revealsthat
since that for
for
for .all
Moreover,
all
. all xx>x>0,>0,with
Moreover, 0, since
with equality
equalityififand
since
with andonlyfor fora
only
reveals that . Moreover, since for all x > 0, with equality if and only if x = 1,
on:
on: reveals that . Moreover, since for all x > 0, with equality
. Moreover,
. Moreover, since since
wewe
equality
we see see
see
that that
thatand
if forifonly forifxif
all
and all
and
>and
if
only 0, >with
xifonly
=all 0,1,with
only we we
ifequality
if=allallequality
we
1, see
see
which see= that
=1,
ifthat 1,if
that
and
is and
onlyonly
which
which
equivalent ifisifxto
is equivalent
ifequivalent
if= and
and x=
ifhaving
and 1, only1,
only
only
all iftoifall
to all =obtained
having
ifhaving
all all 1,1,which
=all whichisiseqo
over, for since all x > 0, with equality for all xif> and 0, with if x = 1,if and only if x = 1,
onlyequality
we see that if and only if all = 1, which is equivalent to having all
if and if andonlyonly allifwhen
if reveals
reveals all
when
when
=that
that ,=
1,thethe 1,judgments
which
which which
judgments
is .are
. isconsistent.
equivalent
are
equivalent
equivalent
Moreover,areconsistent.
Moreover, consistent.
toto to when
having
since
sincehaving having
whenallallthe all
aijjudgments
the w
=judgments
for
for xare
wobtained
i / all
all jx
obtainedobtained
are
>> consistent.
0, consistent.
when the ifif and ifif xx =i
nly
hichifisallequivalent = 1, which reveals the
to having
judgments
isthat
equivalent to having
. Moreover, all since obtainedfor all x0,>with
with
0, with equality
equality equality and if onlyonlyonly
and
hat . Moreover, sinceall when the
obtained
for
judgments all xare > 0, with
consistent. equality if and only if x = 1,
ents are consistent.
are consistent. judgments TheThe The priorities
priorities
are priorities are
are obtained
consistent. areobtained
obtained
by raising bybytheraising The
raising
matrixThe priorities
thethe matrix
matrix
topriorities
arbitrarily areto to
are obtained
large powersby
arbitrarily
arbitrarily
obtained to raising
large
large
by capturepowers
raising powers the
the thematrix
toobtai
to cap
matr ca
sistent... Moreover,
Moreover, we
we see see
wesince
sincethat
that
see that ifif and
and if
for
for only
only
and
all
all xx if
if>>all
only all
0,
0, if all
with
with == 1,
1,
equality=
equalitywhich
which
1, which
if
if is
is
and
andequivalent
equivalent
isonlyequivalent
only if
if xx to
to
== having
having
1,
1,to havingall
all all obtai o
at if andtransitivity
only if allof dominance
transitivity
transitivity of of=dominance
1, The
which
dominance along priorities
isalong
equivalent
chains
alonglarge ofare obtained
toof
arbitrarily
transitivity
chains
chains
transitivity having
ofpowers large
of by
arbitrarilyofall
arbitrarily raising
lengths
dominance large
dominance large thelengths
[13]. matrix
Let
alongobtained
lengths
along uschainsto
now
[13]. arbitrarily
[13].
chains introduce
of
Let
Let usanow
ofarbitrarily large
usarbitrarily
nowintr po
la
int
srities are obtained
are obtained whenby
when
byThe
raising
the
raising
thematrix the
judgments
judgments
prioritiesthe
matrix matrix
are
areare to to
consistent.
consistent.
obtainedarbitrarily
arbitrarily by raising large
powers the matrixto to
capture capture the
to arbitrarily large the
he tained ifif by
matrix andto
and raising
arbitrarily
only
only the
ififwhen
all
measureall large
or==index
the 1,to
1, powersarbitrarily
transitivity
which
judgments which
for to
is
is large
capture
of
areequivalentdominance
equivalent
consistent.
inconsistency, orpowers
the to
tomeasure
the having
havingto capture
along
deviation all
all
ofchains the
fromof arbitrarily obtained
obtained
n.inconsistency,
We from represent large the lengths
consistency [13]. Let
judgments
ominance
nance are
alongalong measure
measure
consistent.
chains chains of or or index
of arbitrarily
arbitrarily index largefor
for inconsistency,
large inconsistency,
lengths lengths [13]. or or
measure
[13].
Let Let the the
usthe or or index
deviation
deviation
us now
now index for
introduceof of
for
introduce a from
inconsistency, a n.n.We oror
We thethedeviation
represent
represent deviationthe
theconofcoo
ong chains
trarily large of arbitrarily
lengths powers The
The
[13].The priorities
priorities
tolarge
Letcaptureuslengths
priorities or
measure are
are
now the obtained
obtained
[13].
introduce Let
transitivity
areindex
obtained by
by us
a raising
raising
now
of the
introduce
dominance
by raising theormatrix
for inconsistency, matrix
matrix a to
to
along arbitrarily
the deviationarbitrarily
chains
to arbitrarily oflarge
large
of large arbi-powers
powers
from powers to
to capture
capture
n. Wetorepre cap
ments
ments are
are consistent.
consistent.
rhe priorities
x inconsistency,
for are or
inconsistency, obtained
index
transitivity
transitivity
index
orby:
the
index
the
of
of
by:
by
deviation
by:
raising
deviation
dominance
dominance of the of alongmatrix
from
along from n. toWe
chains
chains
arbitrarily
n. We
represent
of
of
index
represent
arbitrarily
arbitrarily
by:
large powers
the
the consistency
large
large consistency to capture the
lengths
lengths [13].
[13]. Let
Let us
us now
now introduc
introduc
istency,
ation of or thefrom deviationn. We
transitivity represent
ofthe from
dominance the6 n. We
consistencyrepresent index by:
theofconsistency
iorities
orities are
y of dominanceare obtained
obtained trarily
along by large
chainslengths.
by raising
raising the
of matrix
matrix Let
arbitrarily usalong
to
to now chains
arbitrarily
arbitrarily
large introduce
lengths large
large [13].
arbitrarily
apowers
measure
powers Let us tolarge
to or
capture index
capture
now
lengths
thefor incon-
the
introduce
[13]. Let us now intro
a
measure
measure or
or
Thusindex
index index
for
for by:
. inconsistency,
inconsistency,
if and onlyor or ifthe
theA deviation
deviation
is consistent. of
of What from about n.
from n.
theWeWe term represent
represent
“n-1” inthe the theconsiste
consiste
dominance
ordominance
index foralong along measure
chains
chains of
inconsistency,
sistency,
or
of
Thus
Thus
or
index for
orarbitrarily
arbitrarily
the deviation
the . . inconsistency,
deviation large
largeof oflengths
lengths
if if and
and from
from
or
[13].
[13].
only
only theif
n.Let
n. ifdeviation
Let
Thus
WeWe AAus
Thus us isnow
now
isrepresent of. introduce
consistent.
represent introduce
consistent. . the from
if
What
What
theconsistency a n.
ifaand
consistencyand Weonly
about
about represent
only ifthe
the ifAterm the
isis
Aterm consiscon
“n-1”
“n-1
cons
denominator? Because the elements on the main diagonal are each equal to 1, their sum is equal to
dexex .for
for inconsistency,
inconsistency, or
orAthethe Thus . n. n. ifrepresent
and onlythe A is consistent. What about the
ifconsistency
if and if index
index
and
only
n.
by:
by:
denominator?
denominator?
only
index
indexIf if
we by: ifisAdeviation
denote
deviation
isBecause
Because
consistent.
the
of
consistent.
eigenvalues
of theWhat from
from
theofelements
elements
What
A about
that are
We
We
denominator?
onon the
denominator?
about the
different
represent
the
the
term main
mainterm
from “n-1” the
Because
diagonal
diagonal
Because
“n-1” consistency
in by the
are areelements
inthe each
each , itequal
elements
the equal
is onon
knowntothe
tothe 1,main
1,that theirthe diagona
their
main sum
sum
diago isis
isf and only ifWhat
consistent. A is about consistent. the What“n-1”
term aboutinthe the term “n-1” in the
denominator? Because the elements on the main diagonal are each equal to 1, t
ecause the elements n.n. If
on Ifall
wewe
the denote
denote
main the
theeigenvalues
eigenvalues of n.to
ofequal AIf
An. that wewe
1,that
Ifto aredenote
are different
different
denote the the eigenvalues
from from
eigenvalues by byofA Athat
of that are ,itdifferent
,are itis knowfr
isknown
different
se the elements
nlements
diagonal on are the each
on
sum
main
theof
Thus
Thus main
diagonal
equal
the diagonal
to 1,n.
Thus are ..diagonal
eigenvalues
their each sum
are
ofififare
Aeach
isand
is and
equal
. denote
each
equal
and
tothe
equal
if
equal
only
only
1,
and
to
only
totheirthe
onlyifsum
if if
A sum AAtheir
is 1,
of
ifis isA their
the
consistent.
is
sum
equal
sum
elements
consistent.
consistent.
isA that
is
toWhat
consistent.
is
equaldown equal
What
Whatto the to
about
What
main thediagonal,
about
about aboutthe
termthe the which
term
term term
is
“n-1”
“n-1” “n-1”in
in
hus . if and only If
if we
A is consistent. eigenvalues
What about of the are
term different
“n-1” in fromthe by ,
the eigenvalues
eigenvalues A of
ofdenominator?
that A
sum
sum of
that
are of all
all
are the
differentthe eigenvalues
differenteigenvalues
from from of of
by A A is is
by sum
equal
equal
sum of
to, to
of
itall
thethe
is,
all the
sum
it sum
is
the
known eigenvalues
of of
known the
the
eigenvalues
that elements
elements
that
the ofthe
ofA Aisdown equal
down
is the
equal toto
the the
main
main the sum ofofthe
diagonal,
diagonal
sum th
ues of Afrom
fferent that denominator?
are different
“n-1” onlyin ififfrom
byonly
denominator? Because
Because
the ,denominator?
itis is
Because the
the
byelements
known elements
that Because
theeigenvalues
elements theon
,on it the
the
is
on main
main
known
the elements diagonal
diagonal
that the are
are
on“n-1” theeach eachmain equal
equal diagonalto
to 1,1,totheir
their sum
sumsum is
is equa
equa
ator?..Becauseififthe and
and
 elements onAAsumthe is consistent.
consistent.
of
main all diagonal
the What
What
are each about
aboutofthe A the
equal
main
the
is equalterm
to
diagonal
term
1, to “n-1”
theirthesum are
sum in
iniseach
ofthe
thetheequal
equal elements
to
1, down their the mai is e
genvalues
alues of A of is Aequal
n.
n. is equal
to the to the
sum sum
of the of the
elements elements down A down
Aof the the
main maindiagonal, diagonal, which which is is
m A of
Because
Becauseis equal
the elements
the
the theIf
toelements
elements
Ifare
n.we
sumwe
down
denote
on
denote
of
Ifeach
wethe
on theequal
the
the
the eigenvalues
elements
denote
main
mainto
main
eigenvalues
the
diagonal,
diagonal
diagonaldown
1,eigenvalues
their are theof
which
sum
are
ofmain
is
each
each
that
that
isequal
equal
equal
are
are
diagonal,
A that toto different
different
toare
n.1,1,which 
Ifdifferent
theirwe
their
from
from
is from
denote
sum
sum is the
isknown
equal
equal
by
by
toby the
eigenvalues
to
,, itit is is, itknown
known
is that
known that
denote the eigenvalues  of A that are different from   by , it is that
sum
sumsum of
of all
allofthe
the eigenvalues
eigenvalues of AAof
of is
is A equal
equal to
to thethe 
el the AA allare 
the different
eigenvalues is equal tosum sum
the sum ofof thetheof elements
elements
the elements down
down down the
the main
sumisthe
mainmain diagonal,
diagonal, diagonal, whic
whicw
the eigenvalues of Ais equal to the sum of the elements down the main diagonal,the
the eigenvalues
eigenvalues of that are from by
by , it is
is known
known that
that the
of A that are different from , it is known that the of
 which
eigenvalues of AA is
igenvalues of is equal
equal to
to the
the sum
sum of
of the
the elements
elements down
down the
the main
main
 diagonal,
 diagonal, which
which is
is

 



 346 thomas saaty

Figure 2. Area example.

all the eigenvalues of A is equal to the sum of the elements down the

main diagonal, which is equal to n in this case. We have

sscase.
case.
case.We
We
Wehave
have
have ,,,so
so
so and
and
and isis the
is the
the negative

average of the non-principal eigenvalues of A.

of
ofofthe
the
thenon-principal
non-principal
non-principaleigenvalues
eigenvalues
eigenvaluesofof A.
A.
ofA.
We said before that we need a large order matrix to improve the
before
before that
thatwe
efore that we
weneed
need
needaaalarge
large
largeorder
order
ordermatrix
matrix
matrixto
to
toimprove
improve
improvethe
the
thevalidity
validity
validityof
of
ofour
our
ourresults
results
resultsby
by
by
validity of our results by reproducing answers that correspond to the
wers
wersthat
wers that
thatcorrespond
correspond
correspondto to
tothe
the
thereal
real
realunderlying
underlying
underlyinganswer.
answer.
answer.The
The
Thefollowing
following
followingsection
section
sectionwill
will
will
real underlying answer. The following section will attempt to determine
mine
mine how
howlarge
ine how large
largethe
the
thematrix
matrix
matrixshould
should
shouldbe.
be.
be.
how large the matrix should be.

H1]
H1]
H1] Validation

this
thisapproach
this approach
approachis isisnot
not
notaillustrate
To aanumber-crunching
number-crunching
number-crunching scheme
that this approach schemebut
scheme but
but rather
is notrather
rather relates
relatesclosely
closely
closelyto
a number-crunching
relates to
tothe
the
the
scheme but
rather relates closely to the reality of actual measurement, consider a
measurement,
measurement,consider
measurement, consider
consider
personaaawhoperson
person
person would who
who
who would
would
would
like like
liketo
like
to estimate to
toestimate
estimate
estimate
the the
thearea
the
relative relative
relative
relative area
area
area
of the of
ofgeometric
of
five
cic
c shapes
shapes
shapesgiven
given
givenin shapes
in
in Figure
Figure
Figure given
2.2.ItItItin
2. isisFigure
is an
an 2. It isof
anexample
example
example ofan
of example of with
measurement
measurement
measurement measurement
with
withrespect
respect
respectto towith
toaaa respect
to a tangible criterion. For the purpose of this illustration, the relative
n.
.. For
For
Forthe
the
thepurpose
purpose
purposearea of
ofthis
of this
this illustration,
illustration,
illustration,
inside each shape the
therelative
the relative
relativearea
obtained area
area
from inside
inside
inside
actualeach
each
each shape
shapeobtained
shape
measurement obtained
obtained
is also given.
surement
urement is
urement isisalso
also Of
alsogiven.
given.course,
given.Of Of in
Ofcourse,
course,
course,in real inlife
inreal situations,
real
reallife
life the
lifesituations,
situations, relative
situations,the the areas
therelative
relative would
relativeareas
areas wouldbe known
areaswouldnot
would
to the person. He or she needs to estimate the relative sizes of the
othethe
theperson.
person.
person.He He
Hefigures
or
orshe
or she
shebyneeds
needs
needscomparingto
totoestimate
estimate
estimate
themthethe
the relative
relative
in relative
pairs. sizes
sizesof
A sizes
pairwise of
ofthe
the
the figures
figuresby
figures
comparison by
by
consists of
m inin
in pairs.
pairs.
pairs. AAA pairwise
pairwise
pairwise comparison
comparison
comparison consists consists
consists of of
of identifying
identifying
identifying the the
the figure
figure
figure with
with
with thethe
the
he
he two
twoand
e two and
andestimating
estimating
estimatingnumerically
numerically
numericallyhow
how
howmany
many
manytimes
times
timeslarger
larger
largerthe
the
thearea
area
areaof
of
ofthe
the
thelarger
larger
larger
earea
area
areaof
of
ofthe
the
thesmaller
smaller
smallerone
one
oneusing
using
usingthe
the
thefundamental
fundamental
fundamentalscale.
scale.
scale.The
The
Thesmaller
smaller
smallerfigure
figure
figureis
isisthen
then
then
ciprocal
iprocal value
iprocal value
value when
when
when compared
compared
compared with
with
with the
the
the larger
larger
larger one.
one.
one. These
These
These comparisons
comparisons
comparisons are
are
are
seven is the magic number in nature 347

Table 1. Matrix of judgments, outcomes, and actual relative sizes of the five geo-
metric shapes.

Table 2. Relative consumption of drinks.

identifying the figure with the smaller area of the two and estimating
numerically how many times larger the area of the larger figure is than
the area of the smaller one using the fundamental scale. The smaller
figure is then assigned the reciprocal value when compared with the
larger one. These comparisons are arranged in a five-by-five matrix as
illustrated in Table 1. Conventionally, the item on the left side of the
matrix is compared with that on top. If it is larger, the whole number
corresponding to the judgment is put in that cell. If it is smaller, the
reciprocal value is put in the cell.
A second example is about estimating relative drink consumption
in the United States. To make the comparisons, the types of drinks are
listed on the left and at the top of Table 2, and judgment is made as to
how strongly the consumption of a drink on the left dominates that
348 thomas saaty

Table 3. Random index.

of a drink at the top. At the right of Table 2, it is apparent that the

derived values and the actual values (obtained from various pages of
Statistical Abstract of the United States) are close by nearly any
measure of closeness.
The theory itself also provides us with a compatibility index
between the derived and actual results without the need for statistical
theory. We denote by x = ( xi ) and y = ( y j ) , respectively, the derived
and actual scale priorities, and by c = (cij ) the matrix of
cij = ( xi / x j )( y j / yi ) of one matrix of ratios of the two scales and the
transpose of the other matrix of ratios. We then sum all the elements of
C and divide by n2 to obtain the Compatibility Index, a number that
represents the deviation from perfect consistency of the two vectors.
The index for the drinks example is 1.036. If the two vectors were
identical, the index would be 1. The less compatible they are, the higher
the value will be above 1.

Statistical Demonstration of the Leveling Off of

Inconsistency as the Number of Elements Increases

As the number of elements becomes too large, we will show that the
inconsistency levels off, and it becomes literally impossible to use it to
diagnose the faulty elements. Computers can do the tedious work of
double-checking logical proofs, even for very large order matrices. To
get some feel for what the consistency index might be telling us about a
positive n-by-n reciprocal matrix A , consider the following simulation:
choose the entries of A above the main diagonal at random from the
17 values of the Fundamental Scale {1/9, 1/8,…,1, 2,…,8, 9}. Then fill
in the entries of A below the diagonal by taking reciprocals. Put 1’s
down the main diagonal and compute the consistency index. Do this
50,000 times and take the average, which is the random index. Table 3
shows the values obtained from one set of such simulations and also
their first and second order differences, for matrices of size 1, 2,…,15.
Since it would be pointless to try to discern any priority ranking
from a set of random comparison judgments, we should be uncomfort-
able proceeding unless the consistency index of a pairwise comparison
 and divide by n2 to obtain the Compatibility Index, a number that represents the deviation from
ct consistency ofseven
the twoisvectors.
the magic number
The index for the in nature
drinks 349vectors
example is 1.036. If the two
identical, the index would be 1. The less compatible they are, the higher the value will be
e 1.

Statistical Demonstration of the Leveling Off of Inconsistency as the Number of Elements

ases[H1]
he number of elements becomes too large, we will show that the inconsistency levels off, and it
mes literally impossible to use it to diagnose the faulty elements. Computers can do the tedious
of double-checking logical proofs, even for very large order matrices. To get some feel for
Figure 3. Plot of random index.
the consistency index might be telling us about a positive n-by-n reciprocal matrix ,
der the following simulation: choose the entries of above the main diagonal at random from
7 values of the Fundamental Scale {1/9, 1/8,…,1, 2,…,8, 9}. Then fill in the entries of
w the diagonal by taking reciprocals. Put 1’s down the main diagonal and compute the
stency index. Do this 50,000 times and take the average, which is the random index. Table 3
s the values obtained from one set of such simulations and also their first and second order
rences, for matrices of size 1, 2,…,15.
Since it would be pointless to try to discern any priority ranking from a set of random
parison judgments, we should
Figure 4. Plot ofbefirst
uncomfortable proceeding unless the consistency index of a
order differences.

wise comparison matrix

matrix is much
is much smaller
smaller than
thanthe
thecorresponding
correspondingrandom randomindex
index value
valuein Table
in Table 3. The consistency ratio (CR) of a pairwise comparison matrix
e consistency ratio (CR)
is the of aofpairwise
ratio comparison
its consistency indexmatrix
to theiscorresponding
the ratio of itsrandom
consistency
indexindex to
valueindex
orresponding random (RI) invalue
Table 3. in Table 3.
(RI)
Figure 4 is the plot of RI and shows the importance of the number
Figure 4 is the7plot
when of taking
RI anddifferences
shows the of
importance
the randomof the
valuesnumber 7 when
of Figure 3 astaking differences
the graph
e random valueslevels off when the number of elements is around 7. The number 8 is a
of Figure 3 as the graph levels off when the number of elements is around 7.
cutoff point beyond which the differences are less than 0.10. It follows
number 8 is a cutoff
that point
the CR beyond
is nowhich
longerthe differenceswhen
meaningful are less
many than 0.10.elements
more It followsarethat the
being compared in the matrix because the differences are small (0.05 or
s no longer meaningful when many
less), meaning more
that the RIelements
values inare
thebeing
second compared in the3matrix
row of Table get toobecause
close(0.05
ifferences are small pastorn less),
= 7 to providethat
meaning useful information
the RI values in the for second
the CRrow to be
of used
Table 3 get
determine the most inconsistent judgment. a
close past n = 7 to provide useful information for the CR to be used determine the most
a In addition, although the RI levels off as n increases beyond 8 (and the first order a
nsistent differences become negligible beyond judgment.
n = 7), the matrix consistency index will most likely
keep increasing beyond n = 7 (the more elements, the more difficult to be consistent), leading


In addition, although the RI levels off as n increases beyond 8 (and the first order differences
become negligible beyond n = 7), the matrix consistency index will most likely keep increasing


der of magnitude” is essential in any mathematical consideration of thomas saaty
350
or example, when one has a numerical value between 1 and 10 for some
o determine whetherThe notion
change in of
this“order
valueofismagnitude”
significant isoressential
not, thein any mathematical
consideration of changes in measurement. For example, when one has
can be used. A achange
numericalof a value
whole between
integer value
1 andis10 critical because
for some it
measurement and wants
to determine whether change in this
identity of the original number significantly. If the change in value is a value is significant or not, the
following line of reasoning can be used. A change of a whole integer
e small (by two value
ordersisofcritical
magnitude)
because andit thus wouldthebemagnitude
changes consideredand identity of the
original(one
perturbation is a decimal number
ordersignificantly.
of magnitudeIfsmaller),
the change we arein value
likely is a percent or less,
it would be small (by two orders of magnitude) and thus would be
ue by this decimal without losing
considered the significance
negligible. However ifand thisidentity of the is a decimal (one
perturbation
t understood it to be. Thus in synthesizing near consistent judgment the original value
order of magnitude smaller), we are likely to modify
by this decimal without losing the significance and identity of the orig-
large can cause inal
dramatic
number change
as wein ourfirstunderstanding,
understood itand values
to be. Thusthatin synthesizing near
consistentWejudgment
nge in our understanding. are left with values, changesof that
only values are too
one order of large can cause
dramatic change in our understanding, and values that are too small
can deal with incrementally
cause no changeto change our understanding.
in our understanding.ItWe follows thatwith only values of
are left
one order
be no more than about 0.10.ofThemagnitude
requirement smaller
of 10%that cannot
we canbedeal madewith incrementally to
change our understanding. It follows that our allowable CR should be
%, without trivializing
no more the than
impact of inconsistency.
about Assuming that
0.10. The requirement all cannot be made
of 10%
sistent contradictssmaller, such as 1%
the experience that or 0.1%, continued
requires without trivializing
updating of the impact of incon-
sistency. Assuming that all knowledge should be consistent contradicts
the experience that requires continued updating of understanding.
If the CR is larger than desired, we do three things: (1) find the most
han desired, we do three things: (1) find the most inconsistent judgment
inconsistent judgment in the matrix (for example, the judgment for
the judgment for which
which is largest);
largest); (2)
(2) determine
determinethe therange
range of values to which
the judgment can be changed so that the inconsistency would be
gment can be changed
improved; so that
andthe
(3) inconsistency
ask the judgewould be improved;
to consider changing his or her judg-
ment changing
k the judge to consider to a plausible
his orvalue in that range.
her judgment If he orvalue
to a plausible she is unwilling, we try
with the second most inconsistent judgment and so on.
unwilling, we try withIfthe nosecond
judgmentmostisinconsistent
changed, judgment
the decision and is
so postponed
on. until better
understanding of the stimulus is obtained.
changed, the decision is postponed until better understanding of the In my experience in decision
making in numerous applications, judges who understand the theory
y experience in decision
are alwaysmaking in numerous
willing to revise applications,
their judgments judges who and subsequently
partially
always willing to examine the second
revise their judgments mostpartially
inconsistent judgment and so on. It can
and subsequently
happen that a judge’s knowledge does not permit one to improve his or
nconsistent judgmenther and so on. It can
consistency, andhappen
morethat a judge’s knowledge
information is required to improve the
consistency of judgments.
ove his or her consistency, and more information is required to improve
To see how large random inconsistency can get in using the Funda-
s. mental Scale, experiments were made with 3 x 3 up to 9 x 9 matrices
with 9 and 1/9 alternating in each row and column, for matrices of
random inconsistency can get in using the Fundamental Scale,
to a more inconsistent system. In conclusion, beyond n = 7, it is not possible (at least from a
ith 3 x 3 up to 9practical
x 9 matrices
point of with
view) 9
to and
adjust1/9
anyalternating
inconsistency,inwhile
each at row
the same time it is also much
easier to be inconsistent.
f order 3, 5, 7, and 9 yielded the following respective results for :


 powers of a matrix, one can see that λmax increased by the same amount
powers of a matrix, one can see that λmax increased by the same amount 9.1111 from one mat
22, 28.3333, and 37.4444 roundedtooff thetonext.four decimal places. For successive odd
9.2222, 28.3333,
to 10.1111,
theandnext.37.4444 19.2222, rounded 28.3333,off toand four37.4444
decimalrounded places. For off to four decimal
successive odd places. For successiv
atrix, one can see10.1111,
that λmax19.2222,increased 28.3333,
by the same and 37.4444
amount rounded 9.1111 from off to onefour decimal
matrix places. For successiv
a matrix, one canpowerssee thatof
powers
a matrix,
ofλamax increased
matrix,
one by
one can
canThe thesee
see
that λamount
quality
that λof
same increased
max response 9.1111toby the one
stimuli
from same
max increased by the same amount 9.1111 from one
amount 9.1111
ismatrix
determined from one
by three facto
seven is the magic number in nature
The quality of response to stimuli is determined by three factors: accuracy or validi 351
to the next. consistency, and efficiency. Our judgment is much more sensitive
to the next.
consistency, and efficiency. Our judgment is much more sensitive and responsive to la
ality of responseorderto stimuli3, 5, is 7, determined
and perturbations.
9 yielded by three theWhen factors:
followingwe speak accuracy or results
of perturbation,
respective validity,we for areλmaxreferring
: to numeri
quality of response
perturbations. The
to Thestimuli quality
When of
iswedetermined
speakresponse to
bytothree
of perturbation, stimuli factors:
weis determined
areaccuracy
referring by
ortofourthree
validity,
numerically factors: accuracy
changing consist or va
nd efficiency. Our 10.1111,
judgment quality
19.2222,is much of
ratios. response
28.3333,
more
Conversely, and
sensitive stimuli
37.4444the and isrounded
smaller determined
responsive off to by
to
the inconsistency, three
large factors:
decimal
the more accuracy
difficult itorisva f
, and efficiency. consistency,
places.
ratios. Our For
judgment
Conversely, and
successive efficiency.
is much
theefficiency. odd
smaller theOur more Our
powers judgment
of
sensitive
inconsistency, a matrix,
andis
the much one
responsive
more difficultmore
can see
tosensitive
that
large
it is for usλ and responsive
to know where b to
When we speak of consistency,
perturbation, and wesame are changes
referring to judgment
numerically ischanging
much more sensitive max
and is responsive
attained, it to
increased by the the amountshould 9.1111 befrom made. one Once matrixconsistent
nearto the consistency
next. b
ns. When we the perturbations.
speak of
changes perturbation,
should When bewe we are
made. speakreferring
Once of perturbation,
to
near numerically
consistency we are referring
changing
is attained, to
consistent itnumerically
becomes changing
uncertain con
wh
perturbations.
ely, the smaller the The quality
inconsistency, When of
the we
more speak
response
coefficients of
to
difficult perturbation,
shouldstimuli
it isbefor is us
perturbed we
determined
to knowarebyreferring
by three
where
small to numerically
best
amounts factors: changing
to transform a con
nea
ratios.
versely, the smalleraccuracy
the
coefficients Conversely,
or validity,
inconsistency,
should be the thesmaller
consistency,
perturbed more the inconsistency,
difficult
by and
small efficiency.
it is
amounts for ustheOur
to
to more
know
transform difficult
judgment
where a isit
best
near is
much for us
consistent to know
matrixwheto
hould be made. ratios.
Once Conversely,
more near consistency
sensitive and
the smaller
consistent is attained,
responsive one. the inconsistency,
to large it becomes perturbations.
the more difficult
uncertain When which it is for us to know whe
we speak
the changes
s should be consistent
made. Once one.near should be made.
consistency Once near
is attained, it consistency
becomes changing is attained,
uncertain whichit becomes uncertain
ould be perturbed the
of by changes
perturbation,
small amounts should webeto are made.referring
transform Onceatonear near consistency
numerically
consistent is attained,
matrix a it becomes uncertain
toconsistent
coefficients
ratios.byConversely,
should be perturbed should
small be
amountstheperturbed
smaller
to transform by inconsistency,
the small
a near amounts the
consistent to more
transform
matrix difficulttoaa anear
it is consistent matri
coefficients should be perturbed by small amounts to transform near consistent matri
for us to know where best the changes should
[H1] Inconsistency and Perturbation of the Principal Eigenvector of Prio be made. Once near
ne. consistent one.
consistent
consistency one. is attained,
[H1] Inconsistency and Perturbation it becomes of the uncertain
Principalwhich Eigenvectorcoefficients should
of Priorities[H1]
be perturbed by small By using amounts the inconsistency
to transform index a nearfor a pairwise
consistent matrix comparison matrix,
By using the
to a consistent inconsistency
one. index for a pairwise comparison matrix, we have seen that
ency and Perturbation of the Principal maximumEigenvector numbers of Priorities[H1]
are 7 or 8. Now we use perturbations of the principa
[H1] Inconsistency
sistency and Perturbation
maximum numbers are and
of the Principal 7 orPerturbation
Eigenvector
8. Now we use ofofthe Principal Eigenvector
Priorities[H1]
perturbations of the principal of Priorities[H1]
eigenvector w to sh
[H1] Inconsistency
inconsistency index for a pairwise and
that aPerturbation
comparison
similar result ofisthe
matrix, wePrincipal
obtained. have seen Eigenvectorthat theof Priorities[H1] 1
Inconsistency and Perturbation aof the Principal
thatBy
he inconsistency By
usingfor
a index
similar
using
thea inconsistency
result pairwise
is obtained.
theperturbations
inconsistency comparison index for
index formatrix, pairwise
a pairwise we have comparison
seen thatmatrix, the we have seen th
bers are 7 or 8. Now we use
Eigenvector of Priorities Weof the have principal
two eigenvector
formulas forcomparison
w1 to showmatrix,
representing we have seen
the perturbations; th
the fir
maximum numbers are 7 or 8. Now we use perturbations of the principal eigenvector w to
numbers are 7 ormaximum
8. Now
We have we use twoperturbations
formulas 7 orafor of the we
representing principal
use the eigenvector
perturbations; ofwfor totheshowfirst is thanks to Wilkins1
esult is obtained. By using numbers are for
[20] 8. Now
general perturbations 1the principal eigenvector w1 to
the inconsistency index matrix, for a pairwise and the second comparison positive
matrix, reciprocal matrice
[20]that
ar result is obtained.for aahave
asimilar
general result
matrix, is obtained.
andmaximum
the secondnumbers for positive arereciprocal matrices, which is easier to u
e two formulas that
we
for similar
representing seen resulttheisisperturbations;
that obtained.
the
attributed the first
to Vargas [19].is thanks
[Query 7toor 8.
OKNow
#:Wilkinson we use
as edited?]
have two formulas perturbations We
for representing
is attributed haveoftwo
to Vargas the formulas
the
[19]. [Queryfor
principal
perturbations; representing
#:eigenvector
OKthe as first
edited?]wisthe toperturbations;
1 thanks show that a the
to Wilkinson first is thanks to Wil
similar
ral matrix, and the second
result
We have
for
is obtained.
two
positive formulas
reciprocal for representing
matrices, which the isperturbations;
easier to use,the first is thanks to Wil
eneral matrix, and[20]the forsecond
a general matrix, and
for positive the second
reciprocal matrices, for positivewhich reciprocal
is (1) tomatrices,
easier use, which is easier
[20] We
Vargas [19]. [Query for
#: OK ahave astwo
general formulas
matrix,
edited?] and for therepresenting
second the
for positive perturbations;
reciprocal the first which is easier
matrices,
(1)
d to Vargas [19].is isattributed
thanks#:to toWilkinson
Vargas [19]. [Query #: OK as edited?]
9 for a general matrix, and the second for posi-
is[Query
attributed OK
to as edited?]
Vargas
tive reciprocal matrices, which [19]. [Query is (2) #: OKtoasuse.
easier edited?]
10
(1)
(2) (1)
(1) In the first formula (1)
(1) , represents the principal eigenvector (
(2) In the first formula , represents the principal eigenvector (priority vector), and
(2) (2) other w’s represent the remaining right eigenvectors of matrix A, wher
w’s (2)
first formula ,other represents the principal eigenvector (priority vector), andA,the
represent the remaining right eigenvectors of matrix whereas the v’s represent
In theIn firstfirst
the left
formula, formula w1 represents
eigenvectors , of A
represents theand theprincipal
principal
the ’seigenvector
represent
eigenvector its eigenvalues.
(priority (priority vector),We usea
he first formula , represents
In the firsttheformula
principal , eigenvector
represents (priority
the principalvector), and the (priority
eigenvector vector),
left vector),
esent the remaining eigenvectors
right and of Aother
the
eigenvectors andof the
w’s ’s represent
represent
matrix A, itsthe
the remaining
whereas eigenvalues.
v’s right
represent Wethe
eigenvectors use the of first formula onlya
other
matrix w’s A,represent
whereas showthe that
the remaining n
v’sofrepresentmust be
right small.
the eigenvectors We
left eigenvectorsthen use the
of represent
matrix second
of AA,and the the v’s represen
whereasformula in a large
epresent the show
remaining
otherthat w’s
n right
must eigenvectors
represent
be the remaining
small. We thenmatrix userightA,the whereas
eigenvectors
second the
formulav’sof matrix
in a A,the
large whereas
number the v’s represe
rs of A and the ’s’srepresent
representits itseigenvalues.
eigenvalues.
show that the We We useuse
maximum thethe firstfirst
value formula
formula of n only is only
7 to or showto that ofonsimulations
8 depending how sens
ectors of A andshowleft
thethat
n eigenvectors
must ’stherepresent
be small.
maximum of
We A and
itsA value
then the
eigenvalues.
use
of n ’sthe ’s
is We represent
7 oruse
second its
theitsfirst
formula
8 depending eigenvalues.
informula
aon howonly
large We
number use
to
sensitive the
of first formula o
ust be small. Weleft then eigenvectors
use the second of
system andis the
formula to in a largerepresent
inconsistency. number of eigenvalues.
simulations Weto use theand firstresponsive
formula o
simulations
show that to show
n must that the
be small. Wemaximum then use thevalue second of nformula
is 7 or 8independing a large
n must be small. We
system isthen use the second
to inconsistency. formula intheasensitive
large thenumber of simulations to number of simulati
maximum valueshow on n that
of how 7normust
is sensitive be
andsmall.
8 depending InWe
responsive on thehowthen
first usesystem
formula, second
isandto formula
note thatinthe
weinconsistency.
responsive a large
the number of(priority
eigenvector simulati
show Inthatthe the
first maximum
orformula, value
we
we note of n is 7 or 8 depending on how sensitive and responsi
he maximum value
consistency. show In ofthatn isthe
the 7 maximum
first 8 depending
formula, value on
of how
note nthat is 7the
that oreigenvector
sensitive
the and responsive
8eigenvector
depending (priority
on howthe
(priority vector)
vector) andwill
sensitive be ve
responsi
w
system1 will is be
to very sensitive
sensitive
inconsistency. to
to perturbations
perturbations in A
A if is
is close
close toto any of
any of the other eigen
o inconsistency. system
sensitive
thenote toisperturbations
other toeigenvalues.
inconsistency. inWhenA if (priority isiswellclosevector) to any offrom
separated the otherthe eigenvalues.
other eigen- When λ1 is w
first formula, we that the eigenvector will be very
In the first offormula, we note T that
he first formula,values we note and none that the the
In thewfirst formula, we note products
eigenvector v
(priority w
i that i
the eigenvectors
of vector)
left eigenvector will
the eigenvector (priority v
be(priority
i and
very right vector)
vector)
will be
will be
rturbations in A eigenvectors
if is close to i isanysmall, of the the othereigenvector eigenvalues. When λ1 is to
w1 corresponding well theeigen-
perturbations insensitive
A if
value
sensitive
toisperturbations
will
to close  to any in
be comparatively
perturbations of A
in
if otheriseigenvalues.
Atheifinsensitive
close
is close to to
 anyWhen
perturbations
to
of the other
any of theλother 1in is A. eigenvalues. When λ1 i
well
eigenvalues. When λ1 i


 



 352 thomas saaty

separated from the other eigenvalues and none

right eigenvectors is small, the eigenvector

separated
separated
separated
separated
from
fromfrom
from
thethe the
other
the other
other
other eigenvalues
eigenvalues
eigenvalues
eigenvalues
and
and and
none
and
nonenone
none
of of
theof
of
the the products
products
the
products
products of of ofeigenvectors
left
of
leftleft
left eigenvectors
eigenvectors
eigenvectors
and aa
and
comparatively insensitive to perturbations in A.
right right
right
right eigenvectorsis issmall,
eigenvectors
eigenvectors
eigenvectors is small,
is small,
small, the eigenvector
thetheeigenvector
the
eigenvector
eigenvector w11 corresponding
corresponding
w1wcorresponding
1wcorresponding to toeigenvalue
the eigenvalue will
λwill
The aretothe
to
the
the
eigenvalue
eigenvalue
interdependent λin1λawill
1λ11 will
way be be
tha
comparatively
comparatively
comparatively
comparatively insensitive toperturbations
perturbations A.
inA.
A.in
A.
Table insensitive
4. insensitive
insensitive
to to
Perturbation perturbations
to
perturbations
and first order in in i = 1,…,n is large. Thus if one of them is arbitraril
differences.

TheThe The areare

The are interdependent
interdependent
areinterdependent
interdependent in in ainway
a way
in aaway
way
that
thatthat
want precludes
precludesthem
precludes
that theto
precludes the the
be possibility
small,
possibility
possibility
the i.e.,
that
possibility thatthat
near
just 1.
just
that just
one The
one
just one
eigenvecto
one ,
The viT wi are interdependent in a way that precludes the possi-
=ii1,…,n
==1,…,n
i =i 1,…,n 1,…,n
is is
bility isis
large.
large.
that large.
Thus
Thus
large.
just Thus
ifone
Thus one ifif
if one1of/one
oneofT of them is arbitrarily large, they are all arbitrarily large. However,
vthem
them
of is is
them arbitrarily
arbitrarily
large,
i wi , i = 1,…,n is large.
large,
is arbitrarily they
they
large,
(1) are
Thus are
they
The all all
ifarearbitrarily
one arbitrarily
all of
perturbationthem large.
arbitrarilylarge.
isHowever,
However,
is large.
small as we
thewe
However, c
want arbitrarily
them to large,
besmall,
small, they
i.e., are
near all arbitrarily
1.eigenvector
The large. However,
eigenvectoris stable we want them
want
want them
want them
totobeto
them besmall,
be
tosmall,
small,
be i.e.,
i.e., i.e.,
near
near near
i.e., 1.
near
1. 1.
The
The The
1. Theeigenvector
eigenvector
eigenvector isisstable
w1 isisstable
stable
stable
(2) when:
when: when:
iswhen:
when: well separated from ; when A

(1)(1)
The(1)
(1)
(1)TheThe
The perturbation
perturbationis is
perturbation
perturbation
The perturbation issmall
is
small small
small
small
is as as
the as
as
the
as the
theconsistency
consistency
consistency
the (3)
consistency The
index
consistency index
index
might
index might
index
product
mightmightsuggest;
ofsuggest;
left and right eigenve
suggest;
suggest;
might
suggest;
(2)is is
(2)(2)(2)
(2) well iswell
well
isis well separated
separated
separated
well from
separated
separated fromfrom
from
from ; when;;when
when
;; when
whenAAis AAconsistent,
Aisis isconsistent,
consistent,
for a consistent (and near-consistent) ma
consistent,
consistent,
is
(3) The product of left and right eigenvectors is not excessively large,
(3) The product ofand
left and right eigenvectors here onexcessively
the relativelarge,
isexcessively
not dominance
which isscale
the co
(3)(3)
TheThe
(3) product
Theproduct
which of of
product
is the left
left
of
case and
left
for aright
andright
eigenvectors
right
consistenteigenvectorsis is
eigenvectors notis
not not
(and near-consistent) excessively
matrix large,
if large,
excessively which
which
thelarge, is is
whichtheis
thecase
case
the ca
forelements are homogeneous (compared here on criterion of comparison;
the elements
relative domi- and
forfor
afor aa consistent
consistent
consistent
a consistent (and
(and (and near-consistent)
near-consistent)
near-consistent)
(and near-consistent) matrix
matrix
matrix
if if
matrixtheififelements
the the
theelements
nance scale of the Fundamental Scale) with respect to the criterion areare
elements are homogenous
homogenous
homogenous
are (compa
(compared
homogenous (compared
(compar
hereoftheon the
comparison; relative
and dominance scale of the (4)Fundamental
If the numberScale)of theirwithentries is small
respect to
here
here onon
here on therelative
therelative dominance
relative dominance
dominance scale
scaleof ofthe
scale oftheFundamental
the Fundamental
Fundamental Scale)
Scale) with
Scale) with respect
with respect to tothe
respect thet
to
(4) If the number of their entries is small. 1
We note that n, the order of the matrix, s
criterion
criterion
criterion of of
criterion ofcomparison;
comparison;
comparison;
comparison;
of and and and
and
We note that n, the order of the matrix, should
not[16].
get not be
enough overly small
information from the few comp
(4)(4) (4)
If(4)
becauseIf Ifnumber
theIf
the the
the
then
number
number of of
number
one
oftheir
their their
their
of
does entries entries
entries
not is is
entries
getsmallissmall
small
small
is
enough[16].
[16].
[16].
information from the few
We
comparison note that n,order
judgments the order
toof ofmatrix,
obtainthe matrix,
valid world
should
results measurement.
for not be
real overly
world To determine
small
measure- because thethen
magnitud
one d
We We note
We notethat
note n,
that n,
thatthen,
the order
the of
order thethe
of matrix,
the should
matrix, should notnot
should be be
notoverly
overly
be small
overlysmall because
smallbecause then
because then one
thenonedoes
onedoes
do
ment. To determine the magnitude of n, inconsistency we need to examine on the to the
order effect
n ofvalid
a positive recipro
notnotnot
getget
not get
enough
of enough
enough
get random information
information
enough information
information
inconsistencyfrom
from from
the
fromonthe the
fewfew
the
the few comparison
comparison
few comparison
order n of a judgments
comparison judgments
judgments
positive to to
judgments obtain
reciprocal obtain
obtain
to valid
obtainvalid
matrix results
valid results
results forfor
results for
real
real
for r
re
world
world that
measurement.
world would
A measurement.
world measurement.
measurement. ToToinfluence
To
determine
Todetermine the
determine
determine number
thethe the magnitude
magnitude
magnitude
the magnitude terms
of terms
of of inin
of
n,of
n,
we n,
we the
n,the
we
need
we sum
sum
needneed
to to
need defining
defining
toexamine
examineexamine
examine
to thethe theWith
. effect
effect
the effectlarge
of of
effect random inco
ofrando
rand
random
of
inconsistency
inconsistency
inconsistency
inconsistencyonontheon
the
on the
order order
n of
order
the order nanof
n of aofpositive
aapositive
positivepositive reciprocal
reciprocal
reciprocalmatrix
reciprocal matrix
components
matrix
matrix AAwould
of
A that
A thatthat
that would
wwould
is influence
arbitrarily
influence
1 would influencethethe
influence the
small.
numbernumber
Thus,
number
the of of
number
terms terms
termsin in
terms theinsum
the
in the
sum
the sum defining. With
defining
sum defining
defining .. With
. With With
large large
large inconsistency
stability.one
inconsistency
inconsistency
large inconsistency Note
oneone
cannot cannot
also
cannot
one that guarantee
we need
guarantee
cannot guarantee that
guaranteethatthat
to
none
none
that none
keep the
of of
none theofnt
the
of
components of1wwis values of all the components are ofcondition
the same
components
components
componentsof ofwof
1wis isis arbitrarily
11arbitrarily
arbitrarilysmall.small.
small.
arbitrarily Thus,Thus,
Thus,
small. near-consistency
near-consistency
near-consistency
Thus, is isa isis
near-consistency a sufficient
sufficient
asufficient
asufficientcondition
conditionforforf
condition
matrices are the archetypical
stability. Note also that we need to keep the number of elements relatively small, so that [Query #: OK
seven is the magic number in nature 353

Figure 5. Plot of perturbations in second column of Table 4.

Figure 6. Plot of first order difference perturbations in the third column of Table 4.

With large inconsistency one cannot guarantee that none of the compo-
nents of w1 is arbitrarily small. Thus, near-consistency is a sufficient
condition for stability. Note also that we need to keep the number of
elements relatively small, so that the values of all the components are
of the same order. The foregoing suggests that reciprocal matrices are
the archetypical matrices, which produce stable eigenvectors on small
perturbations of the consistent case. The conclusion is that n must be
small but not too small, and one must compare homogeneous elements
(homogeneity is one of the axioms of the AHP).
Let us now turn to the second perturbation formula. It consists of
the difference between the principal eigenvectors of the perturbed
matrix and the original matrix. In the current paper, we use the
 354 thomas saaty
positive reciprocal matrix is and is the perturbation matr
“Matlab” are software positive
produced tobyreciprocal
simulate
using random matrix
the is
second
values theand
perturbation
from Fundamentalformula.Scale. is the Thenperturbatio
the c
Suppose
positive
positive the
reciprocal
reciprocal original
matrix matrix
is is positivepositive and andreciprocal
reciprocal matrix
matrix is
is theisperturbation
the perturbation and
and
matrix, matrix,
both both
of whic isofth
and follows
is the [13]:
are produced
the perturbation
perturbationmatrix, by using
matrix, random
both
both of of whichvalues from
which are produced by the Fundamental Scale. The
are produced
are produced
using by using
random by values
using
random
(1) random
fromvalues
Calculate are
thevaluesproduced
from from
the
Fundamental
the by using
the Fundamental
Fundamental
principal Scale. random
Then
eigenvectorvalues
Scale.Scale.
theThen from
Then
the
calcula- the
the
w1 calculationsFundamenta
of calculations
A using are a
follows [13]:
om values from thetionsFundamental
are as follows Scale.
6 : Then follows the calculations are as
follows
follows [13]: [13]: [13]:
(1) Calculate the principal eigenvector w1 ofth A
, where is the (i, j) entry of the k powe
(1)(1) (1) Calculate the principal (1) Calculate w1 thewof1 principal A usingeigenvector w
Calculate
Calculate thethe principalprincipaleigenvector weigenvector
eigenvector
1 of A using the formula:
of
A using the the formula for
e principal eigenvector w1 of A using the formula: , where is the (i, j) entry of the kt
(2) Construct the, perturbed
, where where
where isaij(the k)
matrix
is(i,the (i,
j) entry of theofk, ththe
byj) entry where
kth power
power of is of
, where
the the (i, j)op
themat
the
matrix Ae
th
, where is the (i, j) entry of the k power of the matrix A;
kth power ofelementwise
the matrix(2) Construct
A;(Hadamard) the perturbed
product of matrix by
the matrices A and P; , where
(2) Construct
(2)(2)Construct
Construct the the perturbed
theperturbed
perturbed matrix
matrix (2)
matrix
by by
by Construct the perturbed
,, where
where matrix
, where by
the operation
the operation
the “ ” is“ th ”
rturbed matrix by operation “  ”, is
(3) Calculate
elementwise
where
the the the principal
operation
elementwise (Hadamard)“ ”eigenvector
(Hadamard) is the product productof of
theofthe perturbed
matrices
the A andmatrix P; u
elementwise
elementwise
matrices(Hadamard)
A and (Hadamard)
P; (3)product product elementwise
of the ofmatrices
the matrices (Hadamard)
andAP;
A eigenvector product of
and P; of the perturbed matrithe matrices A an
mard) product of the matrices A and P; Calculate the principal
'
(3)(3)Calculate
Calculate
(3) Calculate the principal
the principal
the principal eigenvector (3) w
eigenvector
eigenvector Calculate
1 of ofthe theof thetheprincipal
,perturbed
where
perturbed matrix
perturbedmatrix the A ' (i, using
is eigenvector
matrix j) entryusing
theof formula
ofthe
thethe per
for
ncipal eigenvector of the perturbed matrix using the formula: ( k )'
using the formula: whereaij isis the (i, j) entry o
, ,where
matrix ;
the (i, j) entry of the k th power , where , whereis theis (i,
th the matrix A ' ;
of
thej)(i,entryj) entry of ,the where
of ktheth
kth is
power powerthe th(o
of
, where is the (i, (4) j) entry
Calculateof the
matrix vector thek power
; perturbation of thevector of the principal eigenvector u
(4) Calculate the perturbation of the principal eigenvector using
matrix matrix ; ; ; matrix ;
the formula (4) Calculate the perturbation vector of the principal eigenve
(5) (4)
Run Calculate
(4) Calculate thethe program the perturbation
50,000 times (4)
for Calculate
vector
eachof size thetheprincipal
ofmatrix perturbation
from 15 to vector
2 to eigenvector ofthe
using theformul
princ
the fo
erturbation vector of the do principal
(5)perturbation
Run
(4)—thus eigenvector
the program
calculate the
vector ; the
50,000
usingperturbation
the
formula
principal
times for eacheigenvector
of the principal
size matrix usingfrom 2 to 15
;
; calculate ;
eigenvector; (5)the Run perturbation
the program of the 50,000principal timeseigenvector;
for each size matrix from 2
(6)
(5) Run Compute
(5) Run the
the(6)
the program norm
program of
50,000
Compute
each
50,000of
times
thethe (5)
these
times
norm for Run
50,000for
each theeach
size
of each ofofthese program
perturbation
size
matrix 50,000
vectors
matrix
50,000from
from times
byto 215to
2eigenvector;
perturbation for
to15each
do size(4)—
to(4)—thu
vectorsdo bym
calculate perturbation theitsprincipal
m 50,000 times for each size matrix from 2 to 15 to do (4)—thus
taking the square root of the sums of the squares of entries;
calculate the perturbation of the principal eigenvector
(7) calculate
calculate
Add the all of the perturbation
perturbation
root
these of(6)
50,000 theofprincipal
theofCompute
sums
norms the
ofoftheprincipal
the
the squares
norm
vectors eigenvector;
eigenvector;
ofof
and itsform
each entries;
of the
these 50,000 perturbation vect
average
ation of the principal eigenvector;
value,
(6) Compute
(6) Compute which
the(7) leads
the
norm Add to
norm Table
ofalleachofthe 4 and
each
of sums
these (6)
these Compute
Figures
of these
50,000 50,000 5 and
50,000
norms the
6. norm
ofperturbation
perturbation the of each
vectors
vectors ofbythese
vectors
and form 50,000
bythe
taking taking pertu
the squar
average the vs
root of of the squares of its entries;
m of each of these 50,000 perturbation vectors by taking the square
rootNoteroot
of thethe
ofsumspeak
the of shown
sums
to theofsquares
Table the in
4 Add Figure
squares
and of its
Figuresroot
of 6its ofand
5in
entries; the
the6.sums
entries; valueofofthethe squares
first order of its entries;
(7) all of these 50,000 norms of the vectors and form the ave
he squares of its entries;
differences for seven elements.
(7) Add
all ofall of these (7) Add all of these 50,000 norms of thevalue,
vectors and
(7) Add these to50,000
50,000 norms
Table 4norms
of the
and ofvectors
Figures the vectors
5 and and6.form and form the average
the average value, whichwhich lead
50,000 norms of the vectors and form the average value, which leads
to Tableto Table
Consistency 4 andof 4Figures
and
Note
a Figures
5 and
the
Hierarchical 56.andshown
peak 6. to Table
System in and 4 and
Figure of6Figures
a in the5value
Network andSystem 6.of the first order diffe
es 5 and 6.
The current elements.
section moves Notefrom the calculating
peak shownthe in consistency
Figure 6 in ofthea single value of the first orde
pairwise
Note Note comparison
the peak the peak shown matrix
shown
in Figure to
in Figure Note
calculating the
6 in 6thein value peak
the
the value shown
consistency
of theof first in Figure
for
the first a
orderorder 6 indifferences
hier-
differences the valuefor seve of t
for
elements.
wn in Figure 6 inarchy and for
the value of athenetwork.
first order differences for seven
elements. [H1] Consistency ofelements.
elements. a Hierarchical System and of a Network System[/H1]
The Consistency of a Hierarchy
[H1] Consistency of a Hierarchical System and of a Network System[/H1
[H1]
The Consistency
consistency
[H1] Consistency of a Hierarchical of
of aa hierarchically
The Hierarchical [H1]
currentSystem Consistency
System
constructed
section andmoves
ofand offrom
a Network aaNetwork
ofsystem Hierarchical
of many
System[/H1]
calculating System
System[/H1]
parts and of a Network
theandconsistency of a
rchical System and of a Network
subparts is obtained System[/H1]
11 by first taking sums of products of each consistency

index withcomparison
the composite matrix
Thepriorityto calculating
current of itssection the moves
criterion. consistency
The ratio fromisfor a hierarchy
calculating
then formed the andconsistency
for a netwo
The The current current
section section
moves moves fromto The
from current
calculating
calculating section
the the moves
consistency
consistency froma of calculating
a single the
paa
comparison matrix calculating the consistency forofa hierarchy
single pairwis
and for
n moves from calculating the consistency of a single pairwise
comparison
comparison matrix matrix toConsistency
to calculating
[H2]The calculatingthecomparison
the
of aconsistency
consistency matrix
for a for
Hierarchy[/H2] to calculating
a hierarchy
hierarchy and for the
andaconsistency
for a network.
network. for a hierar
ating the consistency for a hierarchy and for a network.
[H2]The Consistency of a Hierarchy[/H2]
[H2]The
[H2]The Consistency
Consistency  of a Hierarchy[/H2] [H2]The Consistency of a Hierarchy[/H2]
of a Hierarchy[/H2]
 The consistency of a hierarchically constructed system of many parts and su
obtained [15] by first taking sums of products of each consistency index with the c
seven is the magic number in nature 355
priority of its criterion. The ratio is then formed from this number with the sums of the
of the
from thisrandom
numberconsistency
with the sums indexofforthethat order matrix
products of the with
randomthe composite
consis- priority of its
tency index for
In general, thethat ratioorder
shouldmatrix
be inwith the compositeofpriority
the neighborhood 0.10 toofminimize
its concern fo
criterion. In general, the ratio should be in the neighborhood of 0.10 to
improvements
minimize concerninforthe judgments.
needed improvements in the judgments.
Let nj,Let
j = n1,j, 2, ..., h be the
j = 1, 2, ..., h be number of elements
the number in theinjththe
of elements level
jth of
leveltheof the hierarchy.
hierarchy. Let w be the compositeth weight of thethi th criterion of the jth
the composite ijweight of the i criterion of the j level, and let μi,j+1 be the consistency
level, and let μi,j+1 be the consistency
st index of all elements in the (j+1) th
st
all elements in the (j+1) level compared with respect
level compared with respect to the ith criterion of the jth level. The
to the i criterion of the jth l

consistency
consistency index
index ofofa ahierarchy
hierarchyisisgiven
given by
by where wij = 1 for j = 1,
,, where

wijis= the
1 for j = 1,ofand
number ni,j+1 is of
elements thethenumber of elements
(j+1)st level of theto(j+1)
with respect the ith level
st
criterion of the jth lev
with respect to the i criterion of the j level.
th th

The Inconsistency
[H2]The Consistencyofofa aSystem
System[/H2] [Query #: This section is a replica of the
Wesection.
want toOKrepresent bothor
to delete, the
is inconsistency
another sectionalong paths beginning with a
missing?]
goal and the inconsistency in cycles. For paths, we want the initial priori-
The consistency
ties of the elements. of we
For cycles, a hierarchically
want the limitconstructed
priorities of system of many parts and su
the elements.
Weobtained
need to [15]
evaluate inconsistency
by first taking sums by of
theproducts
weight ofof the
eachcorresponding
consistency index with the c
elements. Also, we need the influence priority of an element of a compo-
priority
nent of its criterion.
to compare elements The ratio iscomponent.
in another then formedInfrom
the this
end,number
we needwith
to the sums of the
weight
of thebyrandom
the priorities KC of the
consistency index supercriteria in the
for that order control
matrix hierarchy:
with the global priority of its cr
general, the ratio௛ should
௡೔ೕశభ be in the neighborhood
௦ ௡ೖ of ȁ௖0.10
೓ȁ to minimize concern fo
‫ܥ‬௦ ൌ ෍ ‫ܭ‬௖  ෍ ሺ෍ ෍ ‫ݓ‬௜௝ ߤ௜௝ାଵ ൅  ෍ ‫ܭ‬௖ ෍ ෍ ‫ݓ‬௝௞ ෍ ‫ݓ‬ሺ௞ሻሺ௛ሻఓೖ ሺ௝ǡ௛ሻ ሻ
௝ୀଵ ௜ୀଵ ௞ୀଵ ௝ୀଵ ௛ୀଵ
improvements
௖௢௡௧௥௢௟
௖௥௜௧௘௥௜௔
௔௟௟
௖௛௔௜௡௦

in the judgments. ௖௢௡௧௥௢௟
௖௥௜௧௘௥௜௔

th
where nj =Let j = n1,2,…,h
j, j = 1, 2, theh,number
is ..., be the number of elements
of elements in the jthinlevelthe jand level
μij+1of the hierarchy.
th th st
the consistency
is the composite weight index of all ielements
of the criterion
in of j level,
thethe(j+1) levelandwith μi,j+1 be the consistency
let respect
st th th
all elements in the (j+1) level compared with respect
to the (k)(h) to the i criterion of the j l
ith criterion of the jth level. In the second term, w is the priority
of the influence of the hth component on the kth component, and wjk is
theconsistency
limit priority of the
index
th element in the kth component. In the case of
of aj hierarchy is given by, , where wij = 1 for j
a hierarchy, there are no cycles and the second term is equal to zero. As
in the measurement of consistency of a hierarchy, this index must be
ni,j+1 is the number of elements of the (j+1)st level with respect to the ith criterion of the j
divided by the corresponding index with random inconsistencies.
In both hierarchies and networks, it can be shown that the incon-
sistency cannot be worse than that of the inconsistency of the most
[H1]The Inconsistency
inconsistent subset whoseoffunctionsa Systemare [/H1]
also pairwise compared.11
We want to represent both the inconsistency along paths beginning with a goa
More Examples
inconsistency in cycles. For paths, we want the initial priorities of the elements. For c
want
Here arethe limitexamples
a few prioritiestoofillustrate
the elements. We need idea
the foregoing to evaluate inconsistency
that 7 or 8 is a by the wei
natural bound on the number of interacting elements imposed by the


356 thomas saaty

need for consistency. Although examples are not a proof, still we present
them here because it is interesting how often that bound comes up.

Number of Jurors

We begin with a social system example about people’s interdependent

judgments. Useful observations about how many jurors is the best
number has been extensively studied over the years by many people.
Condorcet’s Jury Theorem12 says that a larger jury, on average, reaches a
more accurate decision. However, Bag, Levine, and Spencer13 of the
University of Surrey have shown that with a larger jury size, the proba-
bility of reaching a correct verdict may, in fact, decrease, contrary to the
Condorcet Jury Theorem. They showed that if the jurors coordinate on
any one of a number of (equally plausible) asymmetric equilibriums other
than the symmetric equilibrium,14 the probability of accuracy reaches a
maximum for a particular jury size and remains unchanged with larger
juries. In referring to a part of their research and statistical findings,
Nagel and Neef15 write:
. . . the most important aspect is the point at which the weighted
sum of errors is least. This point is reached at a jury size some-
where between six and eight; the nearest whole number is seven.
The model therefore predicts that a jury of seven members will
minimize errors in the fashion we assume would be optimum.
Subject to the limitations on the coin-flipping model . . . , we can
refer to a seven-member jury as the optimum jury size for unani-
mous juries.

Stimulus-response Theory

This example is from psychology. G. A. Miller [9] wrote about the

magic number 7 ± 2 . He observed that in responding to successive
stimuli, performance is nearly perfect up to 7 different stimuli but
declines as the number of different stimuli is increased. He says that the
memory span of young adults is approximately seven items. He also
concluded that memory span is not limited in terms of bits but rather
in terms of chunks. A chunk is the largest meaningful unit in the
presented material that the person recognizes.

Maslow’s Human Needs

This example is also from psychology. The following 7 human needs

were identified by Maslow17: the “Basic needs or Physiological needs”
of a human being, “Safety and security needs,” “Love and Belonging”
needs, “Esteem” level needs, “Cognitive” level needs, “Aesthetic” level
seven is the magic number in nature 357

needs, and finally, at the top of the pyramid, the “Need for
Self-actualization.”

Music
The numbers 7 or 8 again play an important role in music in which
octaves (Latin: octavus, or “eighth”) notes are grouped together. The
octave relationship is a natural phenomenon that has been referred to as
the “basic miracle of music,” the use of which is common in most
musical systems.

Computer Science

In computer science, we do not have more than 7 to 8 interacting

component functions in spite of the complexity of today’s computers
and network systems. For example in the Von Neumann architecture,
according to functions we have: (1) input devices; (2) output devices;
(3) the control unit; (4) the arithmetic unit, composing the central
processing unit; and (5) the memory unit. Even when going into greater
detail, the hardware of a personal computer is composed of a case, a
power supply unit, the motherboard, expansion cards, peripheral
devices, storage devices, and input/output devices. Similarly, microcom-
puter design includes the microprocessor, read and write memory
(RAM), read only memory (ROM), Input/Output unit, address bus,
data bus, and control bus (totaling 7 elements) in modern micro-com-
puter architecture.18 This does not mean that there are not millions of
parts, but rather that they are  always hierarchically structured and
grouped in no more than 7 or 8 interacting components. We have
millions of bytes of data or segments in a hard disk; however, they are
not interacting parts but rather parts of a whole, as are the billions
of neurons in the brain. The same thing applies to networks where we
can have thousands of interconnected devices, but they are not actually
communicating altogether—they are used as intermediates in a hierar-
chical structure, even in the case of a fully connected network, as are
mesh networks. The byte itself was designed as a unit of information,
consisting of 8 binary digits (bits).

Biology

The next two examples are from biology. Animal cell functions and
organelles are linked to each other for the overall behavior of the cell.
The cell has the following organelles: Golgi Apparatus, Lysosomes,
Mitochondria, Ribosomes, Endoplasmic Reticulum, Vacuole, and
protein receptors to bring in needed material to and take waste out of
358 thomas saaty

the cell. In Eukaryotic cells, all organelles are controlled by the centrally
located Nucleus. As we said before, they communicate through the
Cytoplasm. Plant cells also have Chloroplasts for making Chlorophyll.
Our hierarchic body subsystems are also a good example of this prin-
ciple. We have the (1) Cardiovascular or Circulatory System; (2) Respira-
tory System; (3) Digestive System; (4) Endocrine System; (5) Reproductive
System; (6) Nervous System; (7) Muscular System; (8) Skeletal System;
and (9) Integumentary System (i.e., skin, hair, nails, sweat glands). The
last three systems provide the framework for support and movement, and
the brain acts as the controller.

Mechanical Engineering

The next example is from mechanical engineering. John Newman of

Vintage Emperor Clock Consultant, THE VILLAGE CLOCKSMITH,
Old Prattville, Prattville, Alabama, answered the question: What is the
maximum number of wheels in a manufactured clock or watch that
keeps time only without chimes or strike? He answered that usually,
there are 5 or 6 wheels included.

Biblical Example

Finally, 7 is a number of great significance in the Bible, as one reviewer

of this paper learned in a recent trip to Israel. God created the world in
6 days and rested the seventh day. God, being God, neither required 6
days to create anything nor to rest afterward, but He may have wanted
to give us a pattern; work days and a rest day are independent of each
other. This 7-day pattern is so important that He explicitly commands
in the Fourth Commandment: “Six Days you shall work, but on the
seventh day you shall rest” (Exodus 34:21). In general, the number 7 in
the Bible represents “divine perfection, totality or completion and is
mentioned at least 490 times.”19

Conclusions

In this paper, systems of elements are considered that are interdepen-

dent, mutually interact, and compensate for the movement of other
elements. The forces that cause such interaction can be gravitational,
electromagnetic, mechanical, or mental.
It was shown above that there is a limit to the number of elements
that can work together interdependently without breakdown in their
cooperative effort. Every system, including the human body, consists of
a hierarchy of parts, subparts, and still smaller parts. It should always
seven is the magic number in nature 359

be possible to identify a part when it becomes inconsistent with the

workings of the other parts. As the system ages, some of its parts
weaken more than the other parts, and if it is very large, the system
would have difficulty identifying the defective parts. It has been demon-
strated that 7 or 8 is a limit on the number of interdependent elements
working together in a module of a system.

References
1. Simon, H. A. “The Architecture of Complexity.” Proceedings of the American Philo-
sophical Society 106, no. 6 (1962): 467–82.

2. Langlois, R. N. “Modularity in Technology and Organization.” Journal of Eco-

nomic Behavior & Organization, 49, no.1 (2002): 19–37.

3. Smith, A. (1776). The Wealth of Nations.

4. Huxley, J. Man in the Modern World, vol. 31 (London: Chatto & Windus, 1947).

5. Romijn, H. “About the Origin of Consciousness: A New, Multidisciplinary Perspec-

tive on the Relationship between Brain and Mind.” Proceedings of the Royal Nether-
lands Academy of Arts and Sciences 1, no. 100 (1997): 181–267.

6. Roberts, S. F. Measurement Theory with Applications to Decision Making, Utility,

and the Social Sciences (Reading, Mass.: Addison-Wesley Publishing Company,
1979).

7. Saaty, T. L. Principia Mathematica Decernendi: Mathematical Principles of Decision

Making (Pittsburgh, P.A.: RWS Publications, 2010).

8. Dehaene, S. The Number Sense: How the Mind Creates Mathematics (Oxford:
Oxford University Press, 2011).

9. Wilkinson, J. H. The Algebraic Eigenvalue Problem, vol. 87 (Oxford: Clarendon

Press, 1965).

10. Vargas, L. G. “The Consistency Index in Reciprocal Matrices: Comparison of

Deterministic and Statistical Approaches.” European Journal of Operational
Research 191, no. 2 (2008): 454–63.

11. Saaty, T. L. and M. S. Ozdemir. “Why the Magic Number Seven Plus or Minus
Two.” Mathematical and Computer Modelling 38 (2003): 233–44.

12. Baldwin, C. Y., & Clark, K. B. (2003). Managing in an age of modularity. Manag-
ing in the Modular Age: Architectures, Networks, and Organizations, 149.

13. Bag, P., Levine, P., & Spencer, C. (2006). A note on: jury size and the free rider prob-
lem. Economics Bulletin, 4(3), 1-12.

14. Mukhopadhaya, K. (2003). Jury size and the free rider problem. Journal of Law,
Economics, and Organization, 19(1), 24-44.
360 thomas saaty

15. Nagel, S. S., & Neef, M. (1975). Deductive modeling to determine an optimum
jury size and fraction required to convict. Wash. ULQ, 933.

16. Miller, G. A. (1956). The magical number seven, plus or minus two: some limits on
our capacity for processing information. Psychological review, 63(2), 81.

17. Maslow, A. H. (1970). Motivation and Personality. New York: Harper & Row.

18. Rafiquzzaman, M. (2005). Fundamentals of Digital Logic and Computer Design.

New York: John Wiley & Sons.

19. Wellman, J. (2014). What Does Number Seven (7) Mean or Represent in the Bible.
Cristian Crier, September. Retrieved from: http://www.patheos.com/blogs/
christiancrier/2014/09/26/what-does-the-number-seven-7-mean-or-represent-in-
the-bible/