Académique Documents
Professionnel Documents
Culture Documents
Abstract:
For an idealized model organization of n decision
There is a fairly common perception that large nodes the per capita limit on the sustainable decision
organizations tend to behave much less intelligently rate is proportional to 1/log2(n). Additionally, a
than their size suggests. They often lose the distinct saturation effect with a higher threshold
decisiveness seen in small groups and may seem switches on when individual's decision throughput
"stupid" to people within them who work with ideas limits are reached.
and knowledge. This paper offers the first quantitative
basis for these observations. It adapts some Quantitative tools for managing entropy and
elements of Shannon's information theory to productivity in business firms and in command and
organizations by defining "organizational entropy" control situations are feasible as practical outgrowths
and related notions of decision complexity and of this work. They involve conscious restructuring
productivity. Organizational entropy measures extra designed to limit entropy effects
decision information needed when partitioning
functions onto a structure.
There appears to be a fundamental upper limit on the Keywords:
average per capita decision rate that an organization
can sustain, depending inversely on the Management, information theory, organizational
organization's entropy. If the limit is broached, entropy, decision networks, social networks, business
impaired productivity among knowledge managers reengineering, technology management, decision
may result and large organizations may be dis- 3
complexity, command and control, C I, entropy-
advantaged when performing knowledge -intensive based warfare, knowledge management, complex
tasks that require efficient use of intellectual capital. systems, control theory applications.
Keywords: ......................................................................................................................................... 1
1. Introduction ................................................................................................................................. 3
2. A model for decision processes............................................................................................... 4
2.1. Symbol transmission in information theory 4
2.2. Management decisions and "dits" 4
2.3. The upper limit M(n) on the maximum management decision flow rate: 4
2.4. The maximum "dit" capacity R(n) 5
2.5. Organizational entropy and the decision complexity A(n) 5
3. Approximations ........................................................................................................................... 6
3.1. A form factor for organization structure 6
3.2. An exactly solvable model: the fully connected organization 7
3.3. Sparsifying topology, limiting choice (partitioning into sub-units) 8
4. Applications ................................................................................................................................. 8
References:....................................................................................................................................... 9
Appendix A: The choice of a definition for M(n) ....................................................................... 10
Appendix B: The smallest fully connected organization ......................................................... 10
About the author:........................................................................................................................... 11
A communication channel's capacity C is the The quantity analogous to Shannon's upper limit on
maximum rate for transmitting raw bits/second, symbol transmission is a function M(n) that
determined by physical characteristics of the channel represents an organization's maximum total
and modulation schemes. The Fundamental Theorem management decision flow rate. An organization is
for a Noiseless Channel [2] showed that there is a regarded as a network of n decision nodes
maximum rate S for symbol transmission that is (knowledge workers) that create, consume, and
simply C divided by the entropy H: communicate actionable information related to the
organizations activities. M(n) is simply the quotient:
C
S H
M (n) R (n) (1a)
r i,j
pi(j)
The ratio M0 is the maximum management decision coupling, it will show up as simplification of
flow rate per collaboration channel (below individuals' tasks: i.e., as a reduction of the decision
saturation). To the right of M0 is a dimensionless form complexity coefficients A0i.
factor that depends purely on the organization
Parallel processor arrays obey a superficially similar
structure and entropy.
rule called "Amdahl's Law", which sets an upper limit
With the structure effects separated, equation (4) can on the speedup ratio [7] achievable by adding
be used for comparisons that reflect organization processing nodes. When an algorithm is completely
structure alone, and absolute measurements of the parallelizable the speedup can grow linearly as
coefficients are not essential. processors are added, since the nodes never need to
wait while another is making decisions. Otherwise
3.2. An exactly solvable model: the
the speedup ratio is less than 1. Amdahl's Law,
fully connected organization though, does not apply entropy to a processor
network, so the analogy is limited.
A further approximation is the "fully connected
model", which corresponds to setting r i,j = 1 for all i The per capita maximum decision rate m(n) = M(n)/n
and j in Equation (4). It assumes that paths between is an organization's productivity limit for knowledge-
all pairs of knowledge workers are open and have intensive tasks. It declines as 1/log2(n) for a system
equal weighting, and that all decision-makers below the saturation threshold.
collaborate with equal probability. The maximum dit
rate R(n) reduces to R0.n(n-1), which approaches Table (1) shows some relative values of M and m for
2 a range of organization sizes. For example, a one
R0.n for large n. All of the conditional probabilities
pi(j) are equal to 1/(n-1). The organizational entropy million-person organization can utilize only about
50,000 times as much management decision
becomes H = n.log2(n-1) while the decision complexity
capacity as a single individual, neglecting saturation
becomes A(n) = A0 H. and assuming it is fully connected (the worst case).
This model can be solved exactly as all terms in the
summations are the same, so the sums become Table 1: Form factors for M and m ignoring? saturation
trivial. But it tends to overstate the entropy effects. In N Decision Rate Per Capita
large organizations this model can most realistically N-1/log2(N-1) Factor
be applied to individual functional units or processes
1/log2(N -1)
that are then sparsely linked to each other.
3 2.00 1.00
4 1.89 0.63
The n.log2(n) expression for the entropy can be
5 2.00 0.50
arrived at another way: that expression is also the
combinatorial complexity of the most efficient general 7 2.32 0.39
method for sorting n objects [6]. The decision 10 2.84 0.32
network in this model can alternatively be viewed as 100 14.90 0.15
a sorting machine for management decisions each of 500 56 0.11
which has complexity A0. The entropy is thus 1,000 100 0.10
proportional to the sorting time. 10,000 753.00 0.0753
The maximum sustainable decision rate M(n) for an 100,000 6,020 0.0602
organization below the size nS that triggers saturation 200,000 11,356 0.0568
is: 500,000 24,409 0.0528
1,000,000 50,171 0.0502
M (n ) = M 0 n-1
log 2 (x) = loge(x)/loge(2) = loge(x)/.6931
log 2(n - 1) 5a)
When an organization is "saturated" the total dit rate
M0 n
for large n n S
is limited by nD0, where D0 is an average node's own
log 2(n) maximum internal dit capacity. For fixed R0 the
number of nodes ns that initiates saturation, along
This result grows sub-linearly; that is, more slowly with organizational thrashing and productivity
than n. It makes intuitive sense that people work
implosion, satisfies D0 = R0.(ns - 1). Above saturation
fastest (although not necessarily most effectively) on
a task when they can proceed autonomously on their the total decision rate M(n) actually falls as 1/log2(n)
with further growth, due to the increased structural
own pieces of it. When there is a synergy or scale information that increases entropy, viz:
economy to be gained by increasing collaborative
M (n) = M 0 nS -1 4. Applications
log 2(n - 1) (5b) One way of applying this work to re-engineering is to
nS use the simple modeling results as heuristics in
M0 for large n > ns
conjunction with a set of rules. Some strategies for
log 2(n) improving knowledge managers' productivity that are
The per capita maximum decision rate m (n) declines consistent with these results have been used
intuitively, but in an ad hoc way without benefit of a
as 1/ n.log2(n) in the saturated regime - a factor of n quantitative rationale:
faster than before.
Reduce choice by creating specialized,
Figure (1) plots m (n) for organizations with dedicated organizations whenever they can be
populations from 3 nodes up to 1 million nodes. justified. Define and automate workflows so
Small organizations growing from a few individuals to that they follow customer-centric processes
about 1,000 experience a ten-fold fall-off in their per and cross "silo" boundaries.
capita productivity that should be highly noticeable.
Further growth (without saturation) from 1000 to 1 Match organization size and structure to the
million knowledge managers would reduce complexity of the task.
productivity by only about another factor of 2 - Use information hiding aggressively.
smaller but still with significant economic impact. A Knowledge managers should work at a level of
person from a small startup firm that is acquired by a abstraction where many low level decisions are
large one would feel a sudden culture shock not seen.
Productivity collapses markedly if saturation is
Use knowledge management systems to avoid
reached, as plotted in figure (1) for a range of
"reinventing the wheel", especially in large
choices for ns. In a real organization saturation would firms.
be reached less suddenly than shown.
Alter the rewards system so that managers
3.3. Sparsifying topology, limiting have an incentive to lower entropy. Decouple
choice (partitioning into sub-units) pay from the number of people supervised.
As experience provides knowledge of the model's
The fully connected architecture maximizes entropy.
coefficients, quantitative tools may emerge to make
"Sparsifying" the decision paths reduces it; i.e. it tuning organizations for high performance a much less
subdivides an organization into subunits and lets a chancy and ad-hoc process. For example, the metrics
small fraction of the people do most of the introduced here may be used as components of cost
communication between them. The entropy 4
functions in linear programming methods.
denominator decreases faster than the "dit capacity"
numerator in Equation (1a). In commercial markets efficiency in managing
knowledge increasingly determines competitive
Managers of large organizations often seem to
advantage. Bigness can be a dis-economy of scale if
intuitively understand this principle. They try to
large, functionally diverse, general-purpose
control decision complexity by subdividing into
organizations compete with focused, low entropy
weakly coupled business units or non-hierarchical
firms in markets where intellectual labor costs drive
process teams, sometimes declaring the intent to
product economics. Where traditional scale
emulate "small firm environments". Only a small
economies (like manufacturing or distribution) are
fraction of the knowledge managers interact across
dominant, knowledge managers' productivity has little
unit boundaries.
leverage on bottom line competitiveness and
Sparsification may not gain as much efficiency as is complex, high entropy organizations may
hoped for if the people who handle external decision nonetheless enjoy critical strategic advantages.
interfaces exceed their "saturation" limits Di. This
Defense applications include improving the speed
may happen easily. One estimate [8] of a typical
and productivity of command and control, intelligence
human's conventional communication bandwidth
capacity is 50 bits/sec - about 250 English words per and sensor fusion, and real-time battle management.
3 These functions all depend on rapid execution of
minute . A human's maximum "dit" rate D0 is likely to many complex decisions by networks of skilled
be much smaller. human decision-makers, who are a limited resource.
1.E+00
1.E-01
1.E-02
Decision
1.E-03
No Saturation
Saturation at N = 10K
Saturation at N = 1000
Saturation at N = 100
Per Capita
1.E-04
Saturation at N = 100K
1.E-05
1.E-06
Rich Janow is the principal of The Analytic Solutions Group, LLC, consulting with government, industry, and
academic organizations. He likes working on cross-disciplinary problems at the intersection of R&D,
information systems, business and technology strategy, physical science, and disruptive technology. He is a
faculty member in the Applied Physics Department at New Jersey Institute of Technology.
Rich's association with advanced technology markets and strategy, roadmaps, assessment,
technological forecasting, and futures includes 18 years at Bell Laboratories, 4 years as an
executive in a high technology company, and applied research in computer science, condensed
matter and surface physics. He holds a Ph. D. from CUNY and an A. B. degree from Columbia
College, both in Physics, and resides in South Orange, New Jersey.
He may be reached at janow@att.net, janow@njit.edu or by phone at (973) 762-4987.