This article was downloaded by: [130.226.44.

8] On: 10 December 2023, At: 05:24

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Management Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

Imitation of Complex Strategies

Jan W. Rivkin,

To cite this article:

Jan W. Rivkin, (2000) Imitation of Complex Strategies. Management Science 46(6):824-844. https://doi.org/10.1287/
mnsc.46.6.824.11940

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

© 2000 INFORMS

Please scroll down for article—it is on subsequent pages

With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 Imitation of Complex Strategies

Jan W. Rivkin
Harvard Graduate School of Business Administration, Morgan Hall 239, Boston, Massachusetts 02163
[email protected]
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

R esearchers examining loosely coupled systems, knowledge management, and comple-

mentary practices in organizations have proposed, informally, that the complexity of a
successful business strategy can deter imitation of the strategy. This paper explores this
proposition rigorously. A simple model is developed that parametrizes the two aspects of
strategic complexity: the number of elements in a strategy and the interactions among those
elements. The model excludes conventional resource-based and game-theoretic barriers to
imitation altogether. The model is used to show that complexity makes the search for an
optimal strategy intractable in the technical sense of the word provided by the theory of
NP-completeness. Consequently, would-be copycats must rely on search heuristics or on
learning, not on algorithmic “solutions,” to match the performance of superior firms.
However, complexity also undermines heuristics and learning. In the face of complexity, firms
that follow simple hill-climbing heuristics are quickly snared on low “local peaks,” and firms
that try to learn and mimic a high performer’s entire strategy suffer large penalties from small
errors. The model helps to explain why some winning strategies remain unmatched even
though they are open to public scrutiny; why certain bundles of organizational practices
diffuse slowly even though they lead to superior performance; and why some strategies yield
superior returns even after many of their critical ingredients are adopted by competitors. The
analysis also suggests roles for management science and managerial choice in a world of
complex strategies.
(Imitation; Complexity; Interactions; Competitive Advantage)

1. Introduction least some firms meet this goal. Although information

The profound influence of imitation on industrial about new products and processes leaks out to rivals
dynamics is, by now, well established. The actions of within about a year (Mansfield 1985) and superior
imitators promote the rapid diffusion of new prod- financial performance at the firm level typically fades
ucts, processes, and organizational arrangements. The within five years (Ghemawat 1991), a handful of firms
prospect of imitation may dull the incentive to inno- maintain superior profitability—and presumably defy
vate, but actual imitation may spur past innovators on imitation— over long periods (Mueller 1986). Particu-
to fresh invention (Schumpeter 1942). Rapid imitation, larly striking is the ability of some firms to resist
it has been argued, reduces the profitability of an imitation despite extensive public scrutiny of their
industry, the degree to which production is concen- strategies. Firms such as Dell Computer, Southwest
trated in a few hands, and the ability of successful Airlines, and Toyota enjoy higher rates of return and
firms to maintain productivity advantages (Nelson faster growth than rivals even though journal articles,
and Winter 1982, Porter 1980). case studies, analyst reports, and books by founding
Accordingly, the strategists of successful firms seek executives reveal the ingredients of their successful
to deter imitation. Empirical evidence suggests that at recipes.

Management Science © 2000 INFORMS 0025-1909/00/4606/0824$05.00

Vol. 46, No. 6, June 2000 pp. 824 – 844 1526-5501 electronic ISSN
 RIVKIN
Imitation of Complex Strategies

What prevents the imitation of well-known, win- acterize the item of knowledge . . . ” Zander and
ning business strategies? Past research focuses on two Kogut (1995) measure the complexity of an innovation
classes of barriers to imitation. Proponents of the by the number of distinct competencies it combines.
resource-based view of the firm highlight factors that Tyre (1991) gauges complexity by “the number, nov-
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

make imitation difficult: impediments to factor accu- elty, and technological sophistication of new features
mulation, social complexity, causal ambiguity, tacit and improved concepts introduced” in an innovation,
knowledge, economies of scale and scope, adjustment while MacMillan et al. (1985) consider a new product
costs, and first-mover advantages (e.g., Barney 1991, complex if it “requires extensive reorganization of
Dierickx and Cool 1989, Lippman and Rumelt 1982). interdependent procedures and/or the coordination
Industrial economists, in contrast, emphasize moves of many skills and multiple departments.” The defini-
by incumbents that make imitation unrewarding, even if tion of complexity which I adopt from Simon (1962)
it is fully feasible. In many game-theoretic models, for encompasses the two major concepts from the knowl-
instance, incumbent firms undertake costly commit- edge management literature: the sheer number of
ments that alter their own future incentives, make retal- elements in an item of knowledge and the degree of
iation threats credible, and thereby fend off copycats. interaction among those elements.
This paper offers a sharply different explanation for The proposal that complexity can deter imitation
inimitability. I propose, and argue formally, that the also draws theoretical support from the extensive,
sheer complexity of a strategy can raise a barrier to though diffuse, research on loosely and tightly cou-
imitation. In my formulation, two factors make a pled systems. This line of inquiry has focused not on
strategy complex: the number of decisions that com- imitation, but on the more general topic of change.
prise the strategy and the degree of interaction among Students of both biological systems (e.g., Glassman
those decisions (Simon 1962). As the elements of a 1973) and organizations (e.g., Weick 1976) have noted
firm’s decision problem grow numerous and interde- that systems whose elements are tightly coupled to
pendent, imitation of a successful strategy becomes one another find it difficult to learn and adapt to
very difficult. Indeed it can become “intractable” in a environmental change. Extensive interaction across
technical sense of the word. Hence, a firm can deter elements can immobilize systems in general, just as
imitation by doing a large number of related things they do the imitators in the analyses presented in this
well even if none of its individual actions is inimitable. paper. Among the research efforts on change in cou-
Because of interactions among the components of a pled systems, the recent work of Levinthal (1997) is
strategy, a would-be imitator could understand most particularly pertinent because it employs an analytical
of the ingredients that make up a successful business framework very similar to the one I use below. The
system yet still fail to grasp the recipe. concerns motivating this paper and Levinthal’s are
Although this paper provides the first formal anal- distinct: Levinthal emphasizes the interplay between
ysis of complexity as a barrier to imitation, there are organizational adaptation and population-level selec-
ample precedents—theoretical, empirical, and clini- tion forces while I examine the process of imitation.
cal—for the notion that complexity deters imitation. While I show how tight coupling among elements of a
Theoretical precedents come, in part, from the litera- strategy protects a successful firm from imitation,
ture on knowledge management. (In this line of re- Levinthal identifies a dark side of coupling: successful
search, the object of imitation is typically a product or firms with tightly coupled systems find adaptation
a process, not a business strategy, but the insights are difficult in the face of environmental change. To-
still relevant.) Conceptual discussions of knowledge gether, the papers illustrate a tension between the
management argue that complex knowledge is hard to protective power and immobilizing influence of tight
transfer. Authors differ, however, in how they define coupling. I return to this in the concluding section.
“complex knowledge.” To Winter (1987), complexity Beyond theoretical precedents, a handful of empir-
reflects “the amount of information required to char- ical studies support the notion that complexity raises a

Management Science/Vol. 46, No. 6, June 2000 825

 RIVKIN
Imitation of Complex Strategies

barrier to imitation. In their analysis of the steel- imitation begins in § 2. There I construct a model in
finishing industry, Ichniowski et al. (1997) document a which (a) a firm must make decisions concerning
bundle of interwoven human resource practices that multiple elements of its strategy, (b) the elements
resists imitation. In a study of pharmaceutical re- interact, and (c) all other barriers to imitation are
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

search, Cockburn and Henderson (1996) show that absent. The model is an adaptation of a tool developed
many firms are slow to match the organizational by Kauffman (1993) in the context of evolutionary
practices of rivals that make science-driven R&D ef- biology and introduced into management science by
fective. They attribute the inertia, in part, to “complex Levinthal (1997). Critically, it parametrizes the two
complementarities” among the practices. Examining aspects of complexity: the number of elements in the
commercial banking products, MacMillan et al. (1985) strategy and the degree of interaction among them.
find that imitation proceeds slowly when a new prod- This allows one to tune the degree of complexity and
uct entails interdependent procedures that span de- analyze how an increase in complexity affects the ease
partment borders. Case studies of individual firms of imitation.
also suggest that highly complex strategies can with- I next show (§ 3) that interactions among decisions
stand imitation attempts. In their discussion of “fit” render the firm’s decision problem NP-complete, i.e.,
and complementarity, Milgrom and Roberts (1995) intractable to algorithmic solution. When interactions
suggest informally that rich interactions among Lin- are pervasive, no general, step-by-step procedure ex-
coln Electric’s many choices may explain why rivals ists that can locate the globally optimal strategy
have not replicated that firm’s well-documented suc- quickly. Hence interactions among decisions play at
cess. Porter (1996) argues that tight fit among numer- least a crude role in determining the ease of imitation.
ous activities helps companies such as Southwest When interactions are common and decision problems
Airlines, Vanguard, and Ikea stave off copycats. More intractable, a firm that finds a good resolution of its
casually, discussions in popular business circles com- problem can hope that it will be hard for others to
monly claim that some inability to “get the whole discover its formula for success.
system right” stymies imitators. If complexity renders global optimization intracta-
ble, however, one must presume that thoughtful man-
Though the proposition that complexity deters im-
agers realize the nature of the challenge they face and
itation has numerous precedents, a formal analysis of
opt not to attempt global optimization. Rather, they
the proposal is worthwhile. The analysis highlights,
likely employ judgment and heuristics to find good,
for instance, how interaction among the elements of a
albeit not necessarily optimal, sets of decisions (March
strategy—not the mere proliferation of unrelated ele-
and Simon 1958). Thus it is crucial that I show how
ments—makes imitation difficult. It pinpoints the type
complexity hampers not just global optimization, but
of interaction that stymies copycats: interactions that
also the heuristics that imitating firms plausibly em-
apply in all business situations (e.g., a complex prod-
ploy. Sections 4 and 5 demonstrate how complexity
uct line always necessitates a highly trained sales force)
undermines reasonable forms of heuristic search. Sec-
pose little challenge. Interactions that vary from one
tion 6 discusses the robustness of the results, and a
situation to another (e.g., a complex product line
concluding section summarizes empirical and practi-
sometimes calls for a multidivisional structure) are far
cal implications.
more problematic. The formal analysis identifies gen-
eral structural features that obstruct strategy making,
and it allows us to talk with precision about the 2. A Model with Tunable
computational burden associated with a business Complexity
strategy. It also allows us to see how great the time This section develops a model with “tunable complex-
required to evaluate a strategy must be before com- ity.” The model allows one to set the complexity of the
plexity truly trips up copycats. strategic problem that firms in an industry face, gen-
The rigorous case for complexity as a barrier to erate a sample of problems with that level of complex-

826 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

ity, examine the ease of imitating a good resolution to mit production of a complex product line are often
each problem, and adjust to a different level of com- made more valuable on the margin if a firm also
plexity. By repeating this process, one can isolate how employs a highly trained sales force that can explain
complexity affects imitation. the product line to customers. The marginal value of
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

product inspection activities may be reduced by man-

2.1. Formal Model
ufacturing activities that eliminate defects in the pro-
The Parameters. Two parameters, N and K, gov- duction process. A large number of recent studies
ern the complexity of a modeled firm’s decision prob- document how choices often complement one another
lem. N is the first aspect of complexity, the number of or substitute for each other. 3
decisions that a firm faces. In the formal model, each Conceptually, N implies that managers are not
of N decisions can be resolved in two ways. 1 Hence a searching along a single dimension for an optimal
particular strategy s, a configuration of decisions, is a decision, as is sometimes posited in economic models.
vector {s 1 , s 2 , . . . , s N }, with each decision s i set to Rather, they are searching a very-high-dimensional
either 0 or 1. Note that there are 2 N possible configu- “decision space” for an optimal combination of choices.
rations of decisions. K controls the second aspect of This suggests the following image of a landscape: each
complexity, the degree to which the decisions interact. of N decisions constitutes a “horizontal” axis in a
The efficacy of each decision is affected not only by the high-dimensional space, and each decision offers dif-
choice (0 or 1) made concerning that decision, but also ferent options. Resulting from each combination of
by the choices regarding K other decisions. In the choices is a value for the firm, which is plotted on the
model, each decision i makes a contribution C i to vertical axis. The goal of the strategy formulation
overall firm value. C i depends not only on s i but also process is to occupy a high spot on this landscape, i.e.,
on how K other randomly assigned decisions are to select a combination of choices that, together, pro-
resolved: C i ⫽ C i (s i ; s i1 , . . . , s iK ). K ranges from 0 to duce a great deal of value for the owners of the firm.
N ⫺ 1. When K ⫽ 0, the contribution of each decision As I discuss below, the interactions introduced by K
depends only on the choice made concerning that cause the landscape to become rugged and multi-
decision: C i ⫽ C i (s i ). When K ⫽ N ⫺ 1, the contri- peaked, and this makes the search for a high peak
bution of each decision is affected by all N ⫺ 1 others: profoundly more difficult.
C i ⫽ C i (s). Note that a “strategy” in this formulation is pre-
N reflects the reality that managers must make cisely a complete configuration of N decisions. Re-
numerous decisions. A firm’s strategy is embodied in
searchers adopt widely varying definitions of the term
this nexus of choices. K captures the fact that the
“strategy.” Whatever definition one adopts, however,
choice made concerning one decision may affect the
a strategy is realized in the marketplace as a set of
marginal benefit or cost associated with another deci-
choices that influence organizational performance. Ac-
sion. 2 For instance, investments in machines that per-
cordingly, this paper focuses directly on sets of
choices. The choices encompassed by the N decisions
1
This formulation assumes that the decisions concern discrete, not include elements that many researchers would con-
continuous, variables, i.e., that each can be resolved in a finite or
sider organizational structure, not strategy per se. For
countably infinite number of ways. In reality, many strategic
decisions present discrete options. The restriction to binary choice is that reason, some readers may prefer to interpret this
made purely for convenience and does not affect any of the paper as addressing the imitation of entire organiza-
qualitative results of this paper. tional configurations (Miller et al. 1984) or organiza-
2
This is an adaptation of Milgrom and Roberts’ (1990) definition of tional forms (Levinthal 1997) rather than strategies.
“complementarity.” Under their definition, complementarity arises
when performing more of activity A increases the marginal benefit of
3
doing more of activity B. I use the term “interaction” rather than See, for instance, Cockburn and Henderson (1996), Hwang and
“complementarity” to encompass both complementarity and sub- Weil (1997), Ichniowski et al. (1997), MacDuffie (1995), Milgrom and
stitution between decisions. Roberts (1990, 1995), Porter (1996), and Siggelkow (1998).

Management Science/Vol. 46, No. 6, June 2000 827

 RIVKIN
Imitation of Complex Strategies

Landscape Generation. The landscape, or equiva- firm’s decision problem. I discuss this further in the
lently the decision problem, which firms face in the final section.
model is generated by a stochastic process. Although
Imitation. Once generated, a landscape is em-
the precise landscape is a product of chance, the
ployed in one of two ways. In §3, I examine the
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

parameters that define the complexity of the land-

computational burden associated with global optimi-
scape, N and K, are controlled by the modeler. A
zation on a typical landscape and show how that
payoff is assigned to each of the 2 N combinations of
burden grows with strategic complexity. In §§4 and 5,
decisions as follows. Recall that the contribution C i of
I consider imitators that use simple heuristics to try to
each decision to overall firm value is affected by K
match a benchmark firm. The effectiveness of each
other randomly assigned decisions. For each possible
heuristic is assessed by a series of Monte Carlo simu-
realization of (s i ; s i1 , . . . , s iK ), a value contribution is
lations. I begin the simulations by fixing N and K. For
drawn at random from a uniform U[0, 1] distribution.
a given level of N and K, a number of landscapes are
The overall value associated with a configuration is
generated. On each landscape, a high point, s 夹, is
the average over the N value contributions:
identified as the best outcome of 100 exploratory

冋冘 册
N
searches, and a benchmark firm is assigned this strat-
P共s兲 ⫽ C i 共s i ; s i1 , . . . , s iK 兲 /N. egy. Five hundred would-be imitators are then re-
i⫽1 leased at random initial locations on the landscape.
That is, each is assigned an initial decision configura-
Note that when K ⫽ 0, P(s) is the average over N tion, or strategy, by chance. Each imitator uses a
contributions, each of which depends only on a single heuristic to search for higher ground— better sets of
choice. At the other extreme, when K ⫽ N ⫺ 1, P(s) decisions—and I record how successful the typical
is the average over N contributions, each of which is imitator is for that level of complexity. I then adjust N
affected by all N choices. and K and repeat the process. This isolates how the
Decision problems are generated by a process that is complexity of the underlying decision problem makes
random in two ways: the K decisions that affect a focal the imitation heuristic more or less effective.
decision are assigned by chance, as is the contribution Three forms of search heuristic are considered.
C i corresponding to a particular configuration of Under the incremental improvement heuristic, an imitat-
(s i ;s i1 , . . . , s iK ). At first it may seem odd that decision ing firm considers all alternative strategies that in-
problems are generated randomly. Recall, however, volve changing M or fewer of its current decisions. It
that the objective is to determine how complexity accepts an alternative if and only if the alternative
affects the imitation process typically, not for a partic- generates a higher payoff. Once it has adopted the
ular decision problem. The best way to accomplish new strategy, the imitator considers another incre-
this is to generate at random a number of decision mental change, and so forth until the firm reaches a
problems of a particular degree of complexity, deter- (possibly local) peak. The parameter M governs how
mine average behavior on this set, adjust the complex- local the search is. M ⫽ 1 implies a polar version of
ity, and repeat the process. local search, employed by researchers such as
The role of chance in the modeling procedure re- Levinthal (1997). M ⫽ N constitutes global search.
flects an important implicit assumption. The N deci- Under follow-the-leader imitation, a firm at s attempts
sions are choices whose relationships are not known a a major reconfiguration toward s 夹. J of its N decisions
priori. If it is universally true and known that, say, two are adjusted to match the benchmark firm’s choices.
choices always “go together,” then those two choices The probability that each decision is matched correctly
are treated as a single decision in this paper’s model. is ␪ ⱕ 1. ␪ is ordinarily less than 1, reflecting the fact
A key purpose of strategy scholarship is to identify that an imitator may not understand the benchmark
sets of choices that normally accompany one an- firm’s operations perfectly, does not have full access to
other—and thereby reduce the dimensionality of a the template (Nelson and Winter 1982) provided by

828 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

the original success, and may not be able to control its K, I isolate the effect of complexity on the evolution of
own operations precisely. the benchmark. These additional simulations and the
In contrast to a follow-the-leader imitator, an incre- primary simulations envision quite different industry
mental improver is “imitating” the benchmark firm settings. The primary simulations reflect a mature
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

only in a weak sense of the word. An incremental business in which an imitator attempts to mimic a
improver does not use any information about the well-established leader. The additional simulations
strategy of the benchmark firm actively. Rather, it apply better to an emerging industry in which the
operates under the hope that incremental improve- standards for performance are still in flux.
ment will lead it to rediscover the benchmark’s suc- It is important to note what the model excludes.
cess. Hence the cognitive models underlying the two Imitators face no costs of adjustment, no threats of
heuristics are very different. retaliation from the incumbent high performer, no fear
I also consider a third search heuristic, a hybrid in that colocation with the high performer will lead to
which a firm jumps toward the benchmark in follow- intense competition, no second-mover disadvantages,
the-leader fashion, tweaks uphill to a local peak, no barriers to modifying individual decisions, no
jumps again if necessary, and so forth. This hybrid economies of scale or scope, and no impediments to
captures three aspects of reality. It recognizes that resource accumulation. Search costs are set to zero;
firms are capable of both major reorientations and indeed, the incremental improver knows the value of
incremental change (Miller et al. 1984). It models all neighboring strategies within a radius of M per-
imitation as an on-going effort, not a one-shot attempt fectly without cost or commitment. The real and
to mimic success. And it permits rivals to start their important fact that some decisions are harder to
imitation attempts (after the initial leap) clustered change than others is purposely suppressed. In short,
around the benchmark rather than spread uniformly none of the conventional deterrents to imitation are
over the landscape. present. Any barriers that imitators face derive purely
As noted above, the benchmark in each simulation from the complexity of the underlying decision prob-
is pinpointed by a series of exploratory searches. One lem, coupled with modest limits on what the firm
knows and can do.
hundred firms, released at random locations on the
landscape, search for better strategies, and the top 2.2. Antecedents in Evolutionary Biology
performer serves as a standard. In all cases, the search Some readers will recognize the modeling approach as
heuristic employed in the exploratory phase is the a modification of the NK model developed by evolu-
same as that used by the subsequent imitators. The tionary biologist Stuart Kauffman and his colleagues
results are robust with respect to the method used to (Kauffman 1993 and references therein). Kauffman
determine the benchmark (§6). devised the NK model in order to examine how
The primary simulation results of the paper exam- biological entities such as proteins and organisms
ine how well imitators match a benchmark firm. evolve. Just as managerial decision making can be
Additional results show how the benchmark itself conceived of as a hill-climbing search in a high-
evolves over time. Specifically, I release thousands of dimensional decision space for an internally consistent
firms sequentially on unexplored landscapes, allow set of choices, so evolution has often been seen as a
each firm to seek a successful strategy using one of the hunt in a high-dimensional genetic space for a peak
heuristics described above, and record when “leader- combination of genes (Wright 1931).
ship” in the market changes hands. The first firm Of course, the biological analogy is not perfect.
released on the landscape sets an initial benchmark for Managers can scan the strategies adopted by their
performance. The second firm may match that bench- rivals, learn from them, and purposely reconfigure
mark, achieve a worse performance, or exceed the their choices to match the competition. Mice, on the
benchmark—in which case it sets a new benchmark. other hand, cannot modify their genetic makeup to be
By repeating this analysis for various levels of N and more elephantine, no matter how much they admire

Management Science/Vol. 46, No. 6, June 2000 829

 RIVKIN
Imitation of Complex Strategies

and study elephants. For this reason, I cannot import First, consider the extreme cases of K ⫽ 0 and K
Kauffman’s biological techniques, unmodified, into ⫽ N ⫺ 1. When K ⫽ 0, choices make independent
the realm of management science. contributions to firm value. In this situation, alteration
When transporting intellectual goods across disci- of a single decision changes the contribution of that
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

pline borders, as I am doing here, one must declare decision alone. Adjacent locations on the terrain—
clearly what is old, what is borrowed, and what is configurations that differ by how a single decision is
new. The particular process for generating landscapes resolved— cannot differ in elevation by more than
and the use of simulation are taken, unmodified, from 1/N, so the landscape is relatively smooth. Moreover,
Kauffman (1993). As subsequent references make from any initial location on the landscape, firms can
clear, the structural features of landscapes I discuss rise to the global peak via a series of value-improving,
below are identified by Kauffman and other users of single-decision “tweaks.” Management simply adjusts
the NK model. Prior efforts, however, do not explore each choice, one by one, to the option that allows that
the collective effect of those features on imitation decision to make the greater contribution in isolation.
attempts. Of the search heuristics I explore, the incre- Hence when K ⫽ 0, one has a smooth landscape with
mental improvement mode is identical to the search a single peak whose basin of attraction is the entire
process employed by most of Kauffman’s biological decision space.
entities. The informed follow-the-leader heuristic and In contrast, K ⫽ N ⫺ 1 implies that every decision
the hybrid heuristic are new. Indeed, such purposeful influences the contribution of every other choice.
search processes make little sense in the context of Then, a small step on the landscape—a change in a
biological evolution. Similarly, the use of a benchmark single decision—alters the value contributions of all N
toward which followers aim is an innovation. Kauff- elements. Adjacent decision configurations have levels
man (1993) speculates on the computational burden of value (elevations) that are altogether uncorrelated
imposed by the NK model but, since the agents in his with one another, and the impact of a single decision
models cannot devise algorithms, he need not prove is bounded above by 1 and below by ⫺1. The land-
the intractability of optimization as I do below. Al- scape is fully random with many local peaks.
though developed in the context of evolutionary biol-
The landscapes considered here exist in such high-
ogy, NK models are beginning to find applications in
dimensional spaces that they cannot be depicted di-
management science (McKelvey 1999 and references
rectly, but random walks on typical landscapes can
therein). None of these efforts examine imitation.
convey a sense of ruggedness. Figure 1 does this for N
2.3. Interactions and Landscape Ruggedness ⫽ 24 and various values of K. As K rises, the landscape
A crucial feature of the modeling approach is that, by grows rugged, and the elevation at any step becomes a
means of K, it parametrizes the richness of interaction poor predictor of the elevation at the next step.
among the elements of the firm’s strategy. Much of the The intuition of why interactions may deter imita-
intuition concerning the effects of K depends on the tion now begins to become clear. In a richly interactive
following insight: as the pieces of a strategy become system, global optimization is difficult. In graphic
more interdependent, the landscape corresponding to terms, it is hard to pinpoint and scale the highest peak
a firm’s decision problem becomes increasingly rug- on rugged terrain. Incremental improvement is inef-
ged. That is, the correlation between the altitudes of fective. In enhancing the performance of one element
adjacent points on the landscape—points that differ by of the system, managers may inadvertently under-
how one decision is made—falls, and local peaks mine other parts. The firm quickly becomes ensnared
proliferate. In the biological context, Kauffman (1993, in a web of conflicting constraints. Graphically, the
pp. 54 – 63) goes to great lengths to show that this is firm is trapped on a low local peak. A wholesale
the case. Here I give two examples that provide some reconfiguration might free the firm from the local peak
intuition concerning the link between K and landscape and take it to a much higher location. Unfortunately,
ruggedness. interactions also make such large-scale changes risky.

830 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

Figure 1 Random Walks on N ⴝ 24 Landscapes

Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

Description. For N ⫽ 24 and various values of K, landscapes were generated as described in the text. On each landscape, a firm was released at a random initial
location and allowed to wander, altering one decision at random each period for 200 periods. The charts show the elevation of each firm over time. The correlation
between the height each period and the height in the previous period is also reported.
Interpretation. As K rises, the landscape in decision space becomes rugged.

In a long jump on rugged terrain, a miscalculation on imitation intercedes). Complexity alone would pose
any one of many dimensions can cause a firm to land no obstacle to imitation.
in a trough instead of atop the intended peak. This section proves that complexity indeed makes
The landscape metaphor alone gives us this intu- global optimization hard. When decisions are numer-
ition. The mathematical model of the landscape allows ous and cross-coupled, the strategy formulation prob-
us to deepen our understanding of how complexity lem defies algorithmic solution. It becomes intractable,
deters imitation. in the technical sense of algorithm theory.

3.1. The Theory of NP-Completeness 4

3. Intractability of the Strategy Algorithm designers in computer science, mathemat-
Formulation Problem ics, and operations research have developed a lan-
Complexity cannot deter imitation unless it first guage system and a set of ideas concerning the
makes global optimization difficult. Suppose that the
decision problems posed by the model are easily 4
This subsection draws substantially from Garey and Johnson’s
solved by some algorithm, regardless of N or K. Then
(1979) seminal text, especially Chapters 1–3. Algorithm designers
a benchmark firm that finds even the best possible set draw a distinction between decision problems and search problems.
of decisions will see its strategy quickly discovered by This section does not emphasize the distinction, which is not
others and replicated (unless some other barrier to relevant in the present context.

Management Science/Vol. 46, No. 6, June 2000 831

 RIVKIN
Imitation of Complex Strategies

tractability of decision problems. The tractability of a solve an NP-complete problem of large size in reasonable
particular problem hinges on its time complexity func- time.
tion, which relates the size of a specific instance of a
problem (n) to the time that a given algorithm re- 3.2. Strategic Complexity and NP-Completeness
Now consider the implications for the strategy formu-
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

quires to solve the problem (t(n)). An algorithm is any

lation problem as formalized in this paper. Suppose a
general, step-by-step procedure for solving a problem.
group of managers faces a decision problem, an NK
A critical distinction exists between polynomial time
landscape, whose topography is unknown. Given
algorithms, for which t is some polynomial function of
some time, the managers can correctly evaluate the
n, and exponential time algorithms, for which t is an
payoff of any proposed strategy. Can they write down
exponential function of n. This distinction is crucial
an algorithm that will quickly locate the best set of
because exponential functions grow far faster than
choices? It turns out that the answer hinges on K. As
polynomial functions. Algorithms with exponential
the degree of interaction among a firm’s decisions
time complexity become impractical for large in-
rises, the NK problem goes from being a member of
stances while algorithms with polynomial time com-
the P class of problems to being NP-complete. The
plexity remain feasible. Problems for which polyno-
time required for any algorithm to find the best
mial time algorithms exist are said to fall into the P
strategy switches from being a polynomial function of
class and are considered by algorithm designers to be
N to being an exponential function of N (under the
well solved. Problems for which no polynomial time
standard assumption that P ⫽ NP).
algorithms exist are labeled “intractable.”
To see this, first consider the NK model with no
Definitively proving intractability, however, turns
interactions, K ⫽ 0. Trivially, this can be solved in
out to be difficult, and this difficulty has led to two polynomial time by an algorithm that optimizes each
additional concepts: the NP class of problems and of the N decisions one-by-one, by choosing the better
NP-completeness. Many problems for which no poly- of the two alternatives. The time complexity function
nomial time algorithm has been found could, in the- associated with this problem is simply a linear func-
ory, be solved in polynomial time if one had a “non- tion of N.
deterministic computer,” a machine with unlimited An appendix available from the author proves,
parallel processing capabilities. Such problems fall however, that the NK model with K ⬎ 2 produces an
into the nondeterministic polynomial, or NP, class. P NP-complete problem. The proof proceeds roughly as
債 NP. Moreover, it is widely believed (but not yet follows: The NK problem with K ⬎ 2 is first shown to
proven) that P 傺 NP, i.e., there are NP problems with belong to the NP class. A particular problem that is
no polynomial time solutions. The NP-complete class of widely known to be NP-complete, the so-called “K-
problems is a subset of NP problems which have been satisfiability problem,” is then chosen. Since the
shown to be as hard as any in the NP class. Transfor- K-satisfiability problem is NP-complete, it is as hard
mations have been devised that permit one to convert as any in the NP class. It is next shown that any
each NP-complete problem into others in polynomial K-satisfiability problem can be transformed in polyno-
time. Therefore, if one could find a polynomial time mial time into an NK model with K ⬎ 2. Such an NK
algorithm for any NP-complete problem, one could problem is therefore at least as hard as the K-satisfi-
devise such an algorithm for all NP problems. Con- ability problem; if the NK problem could be solved in
versely, if any problem in the NP class is intractable, polynomial time, then one could also solve the K-
all NP-complete problems are intractable. Hundreds satisfiability problem in polynomial time by trans-
of problems have now been shown to fall into the forming it into an NK problem and solving the trans-
NP-complete class (Garey and Johnson 1979). It is formed problem. In sum, the NK problem with K ⬎ 2
widely believed, though not yet proven, that the is in the NP class and is at least as hard as the
NP-complete problems are all intractable. In all like- K-satisfiability problem, which itself is at least as hard
lihood, no general, step-by-step procedure can be devised to as any problem in NP. One concludes that the NK

832 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

problem is at least as hard as any problem in NP; it is tions of N and evaluation time per strategy. (Evalua-
NP-complete. Under the standard assumption that P tion time reflects either the time required to conduct a
⫽ NP, the NK problem is intractable for K ⬎ 2. study of an alternative strategy or the time necessary
to implement a test and receive feedback.) Only when
3.3. Implications
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

N is very low and the evaluation time very brief does

The first consequence of NP-completeness, the one
most germane to this paper, is that K delineates two exhaustive search appear practical.
categories of decision problems. In the low-K cate- The NP-completeness result alone does not imply
gory, decision problems are easy, and great strategies that complexity deters imitation. The result ensures
are quickly deciphered and matched. When K is high, only that complexity undermines global optimization.
problems are complex, and a successful resolution of a When facing a complex decision problem, thoughtful
decision problem may defy discovery for an exponen- managers presumably understand the futility of try-
tially long period. The generality of this result is worth ing to “solve” the problem and match the performance
emphasizing. It applies to any step-by-step search of a benchmark firm by any algorithmic means.
procedure that one could write down. It encompasses Rather, they likely use simple heuristics and attempt
procedures ranging from simple exhaustive search to to learn directly from the successful firm. The remain-
sophisticated approaches that send “probes” into the ing challenge, then, is to show that complexity also
landscape and react to the results. undermines the heuristics and attempts at learning.
For K ⬎ 2, the strategy decision problem is intrac- Sections 4 and 5 tackle this challenge.
table in the technical sense of the word. The problem The NP-completeness result suggests interpreta-
becomes intractable in the vernacular sense, however, tions for two phenomena that are unrelated to imita-
only when two further conditions hold: the number of tion. First, it is commonly posited that managers
decisions N is substantial, and the time involved in satisfice rather than optimize (e.g., March and Simon
each algorithmic step is not trivial. Some examples 1958, Nelson and Winter 1982). Satisficing seems a
make this clear. When N ⫽ 5 and a proposed strategy sensible course of action if the strategy formulation
can be evaluated in an hour, for example, a very problem is NP-complete; managers have little choice
simple algorithm— exhaustive search—will pinpoint but to use heuristics and judgment to seek good, albeit
the global optimum in a mere 2 5 ⫽ 32 hours. When N not necessarily optimal, configurations of decisions.
grows to 20, however, the solution time associated Second, the NP-completeness finding suggests that in
with exhaustive search rises to 120 years. Table 1 complex settings, some very good strategies may go
shows the time required for exhaustive search to undiscovered for long periods. Thus a successful new
discover the global optimum under various combina- way of doing business may emerge “out of the blue”

Table 1 Time Required for Exhaustive Search

Time required to evaluate one possible strategy

1 second 1 hour 1 day 1 year

N ⫽ 5 32 seconds 32 hours 32 days 32 years

N ⫽ 10 17 minutes 43 days 2.8 years ⬃1,000 years
N ⫽ 20 12.1 days 120 years ⬃3,000 years ⬃1 million years
N ⫽ 30 34 years ⬃120,000 years ⬃3 million years ⬃1 billion years
N ⫽ 40 ⬃35,000 years ⬃125 million years ⬃3 billion years ⬃1 trillion years

Description. The table reports the time required to evaluate all 2 N possible strategies, for various
combinations of N and evaluation time per strategy.
Interpretation. Exhaustive search is practical only when N is low and evaluation time is brief.

Management Science/Vol. 46, No. 6, June 2000 833

 RIVKIN
Imitation of Complex Strategies

occasionally. Insightful or lucky managers may dis- raises a dilemma concerning the comparative static
cover superior strategies—virgin peaks— even though result: is it appropriate to focus on the global optimum
tastes and technologies have not changed. when that optimum might be very hard to discover?
Finally, the NP-completeness of the strategy formu- How relevant are comparative statics when optimiza-
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

lation problem poses a question for the recent, impor- tion might be a poor model of behavior?
tant stream of research on complementarities (e.g., The dilemma is eased, to a certain extent, by the fact
Milgrom and Roberts 1990, 1995; Topkis 1998 and that Milgrom and Roberts et al. examine comparative
references therein). Like this paper, the complementa- statics only for supermodular functions, in which an
rity research drops the assumption, often made by increase in each decision variable always raises the
economists, that managers must make only a handful incremental returns to all others. (NK landscapes are
of decisions. It also emphasizes the role played by much more general. With a payoff function generated
interactions among choice variables. In contrast to this by the NK model, altering a decision may increase or
paper, however, the complementarities research is decrease the incremental returns to another decision,
interested primarily in comparative statics: how do and the sign of the effect may depend on the choices
optimal sets of decisions concerning complementary made concerning other decisions.) Finding a good set
variables shift as underlying parameters change? Con- of decisions on a supermodular landscape is, in certain
sider, for instance, the array of decisions which an senses, easier than finding one on a general NK
auto manufacturer must make concerning price, ma- landscape. On a supermodular landscape, for in-
chine flexibility, product variety, the length of produc- stance, one can test whether a candidate point is a
tion runs, inventory, vertical integration, and human global optimum by surveying just the two orthants
resource practices. How does the best set of decisions above and below the candidate (Milgrom and Roberts
shift as exogenous factors change—say, as the costs of 1995, p. 185); on a general landscape, one would have
data communication, computation, and flexible ma- to consider all 2 N orthants. Starting at any point on a
chines fall? supermodular landscape, a firm that scours only those
The power of the complementarity approach is that two orthants is sure to get at least half of the gains that
it identifies the weakest conditions under which one an exhaustive search would yield (Milgrom and Rob-
can say anything about such questions. Roughly, erts 1995, p. 185); on a general landscape, such a
Topkis, Milgrom and Roberts, and others show that limited search might produce none of the gains. For
one can make clean comparative static predictions if particular restricted classes of problems with super-
and only if the auto maker’s profit ␲(x, p) is super- modular payoff functions, polynomial-time algo-
modular in decision variables x and outside parame- rithms are known to exist. (Specifically, the ellipsoid
ters p. That is, the marginal return to increasing any method can solve problems involving supermodular
variable must increase with any other variable, for all set functions in polynomial time (Grötschel et al. 1988,
values of all other variables. Under those conditions, Chapter 10).)
the optimal choice x 夹(p) is monotone nondecreasing Although the structure imposed by supermodularity
in p. Milgrom and Roberts (1990) use these results to makes global optimization easier, it by no means makes
show how declining costs of communication, comput- it easy. The time required to survey two orthants may
ers, and flexible machines could prompt an optimizing easily rise exponentially with the size of a problem.
firm to lower prices, invest in flexible production, Polynomial-time algorithms are not known to exist for
broaden product lines, shorten production runs, re- most types of problems with supermodular payoff func-
duce inventory, buy inputs rather than make them, tions. Even when they do exist, they can be very time-
and adopt team-based human resource practices. consuming: “theoretically efficient” but not “practically
The NP-completeness result, however, illustrates a efficient” (Grötschel et al. 1988, Chapter 3).
situation in which interactions among numerous de- In sum, before applying comparative static results
cisions make the global optimum difficult to find. This in situations involving sets of interactive variables,

834 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

one should be confident that the global optima to get snared on a local peak rather than make its way to
which the results apply can indeed be located. This is the benchmark strategy.
more likely to be true when the number of decisions The multiplicity of peaks is not quite so daunting
involved is modest; when little time is required to when K is less than N ⫺ 1, but it remains very large.
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

evaluate a candidate solution; when interactions are For various values of N and K, I release 100 firms at
sparse; when search is permitted for a long period; random locations on a landscape and allow them to
and when a large number of firms are tackling the edge uphill by single-decision improvements. Table 2
decision problem. Supermodularity imposes some reports the number of distinct local peaks that the
structure that aids in finding good solutions, but it firms discover. When N and K are low, only a handful
does not guarantee that the global optimum will be of peaks exist, and all 100 firms gravitate toward them.
found readily. When N and K are high, nearly every firm reaches its
own local peak. Note that both aspects of complexity
are necessary for local peaks to proliferate.
4. Imitation by Incremental Low Peaks. The multiplicity of peaks on rough
Improvement terrain would not be harmful to imitators if those
This section and the next turn to heuristic-based peaks were highly profitable. It turns out, however,
imitation. I begin with the plight of the incremental that not only do local peaks proliferate as K rises, but
improver. I use simulation to show how strategic the “average” peak sinks. For the extreme cases of K
complexity undercuts such a firm’s ability to approach ⫽ 0 and K ⫽ N ⫺ 1, average peak height is easily
the benchmark firm. Prior to a full-scale simulation, calculated using order statistics and the Central Limit
however, I discuss two structural features of land- Theorem. As shown in Table 3, N has no effect on peak
scapes that bode ill for the incremental improvement elevation when decisions are independent of one
effort. another (K ⫽ 0). In contrast, when interactions are
pervasive (K ⫽ N ⫺ 1), increasing N causes the
4.1. Structural Deterrents to Incremental expected peak height to fall. In the large-N limit, firm
Improvement value on a local peak declines to 0.5, which is no better
Consider a rival that hopes to replicate the benchmark than the value expected from a randomly selected set
firm’s success by making a series of incremental, of choices. Calculating the heights of local optima is
single-decision improvements (M ⫽ 1). As the under-
lying decision problem becomes complex—as N and K Table 2 Number of Distinct Peaks Discovered by 100 Firms
rise—the rival’s prospects worsen for two structural
reasons. K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11

Many Peaks. First, the number of local peaks on N ⫽ 4 1.0 1.9 3.3
which the imitator may get trapped rises rapidly with (0.0) (0.2) (0.4)
N and K. When K ⫽ 0, there exists only one peak N ⫽ 8 1.0 2.4 8.6 23.5
regardless of N. Facing such a decision problem, the (0.0) (0.5) (0.7) (0.7)
N ⫽ 12 1.0 5.1 20.2 55.6 80.4
incremental imitator will surely replicate the high
(0.0) (1.1) (1.2) (1.2) (1.1)
performer’s strategy eventually. At the other extreme, N ⫽ 24 1.0 15.1 68.8 97.8 99.6
K ⫽ N ⫺ 1, the elevations of adjacent strategies are (0.0) (1.8) (1.5) (0.4) (0.2)
fully uncorrelated. Of the 2 N possible strategies,
2 N /(N ⫹ 1) are expected to be local peaks (Kauffman Description. For various values of N and K, 100 firms were released on a
landscape and allowed to move uphill, one decision at a time, until reaching a
1993, p. 47). This number is large for even modest N.
local peak. This table records the number of distinct local peaks discovered in
For N ⫽ 20, for instance, the typical K ⫽ N ⫺ 1 this fashion. Each data point is the average over ten landscapes, and the standard
landscape has nearly 50,000 local optima. On such deviation of the average is shown in parentheses below each data point.
terrain, the incremental improver will almost surely Interpretation. As N and K rise, local peaks proliferate.

Management Science/Vol. 46, No. 6, June 2000 835

 RIVKIN
Imitation of Complex Strategies

Table 3 Expected Height of a Local Peak Table 4 Results of Incremental Improvement Efforts

N ⫹ 1 K ⫽ 0 K ⫽ N ⫺ 1 Panel A: What portion of imitators match or exceed the benchmark firm’s

performance eventually?
10 0.667 0.648
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

20 0.667 0.624 K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
30 0.667 0.610
40 0.667 0.600 N ⫽ 4 100.0% 70.2% 50.1%
50 0.667 0.593 (0.0%) (4.6%) (4.3%)
100 0.667 0.573 N ⫽ 8 100.0% 63.5% 28.0% 7.6%
(0.0%) (5.9%) (3.8%) (0.9%)
Description. This table shows the expected height of a local peak for various N ⫽ 12 100.0% 49.1% 17.9% 3.2% 1.3%
values of N and for extreme values of K. (0.0%) (7.3%) (3.0%) (0.4%) (0.3%)
Interpretation. As the underlying decision problem grows complex, the typical N ⫽ 24 100.0% 22.6% 5.7% 1.4% 0.7%
local peak falls. (0.0%) (4.1%) (1.8%) (0.4%) (0.2%)

Panel B: How much worse off is the typical imitator than the benchmark firm,
more difficult for intermediate levels of K. Weinberger as a percent of value?
(1991), however, succeeds in doing so via the Central
K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
Limit Theorem and a series of approximations. He
finds that for values of K that are large but possibly N ⫽ 4 0.0% 2.2% 8.9%
much less than N, the heights of local peaks are (0.0%) (0.5%) (1.6%)
asymptotically normally distributed with approxi- N ⫽ 8 0.0% 2.6% 5.3% 11.7%
mate mean ␮ ⫽ 21 ⫹ 公ln(K ⫹ 1)/6(K ⫹ 1). Note that (0.0%) (0.5%) (0.6%) (0.9%)
N ⫽ 12 0.0% 2.3% 6.4% 11.9% 15.5%
d ␮ /dK ⬍ 0 for K ⱖ 2 and lim K3⬁ ␮ ⫽ 21. For large K,
(0.0%) (0.5%) (0.4%) (0.8%) (0.8%)
the typical local optimum on which an incremental N ⫽ 24 0.0% 2.5% 7.4% 9.9% 10.8%
imitator gets trapped corresponds to mediocre perfor- (0.0%) (0.2%) (0.5%) (0.5%) (0.6%)
mance. As the web of conflicting constraints thickens,
opportunities for incremental improvement can van- Panel C: Once it has reached a local peak, how many decisions does the
typical imitator make differently from the benchmark firm?
ish even at low levels of performance.

4.2. Simulation Results K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11

The structural features suggest that incremental im-
N ⫽ 4 0.0 0.7 1.3
provement will fail in the face of complexity. Simula- (0.0) (0.1) (0.1)
tions confirm that this is the case. To provide a polar N ⫽ 8 0.0 1.2 3.4 3.8
case, I set M, the number of decisions modified in each (0.0) (0.2) (0.3) (0.1)
incremental step, to 1. For various values of N and K, N ⫽ 12 0.0 2.8 4.6 5.9 6.0
500 imitators are released on each of ten landscapes (0.0) (0.4) (0.3) (0.1) (0.0)
N ⫽ 24 0.0 3.8 9.3 11.9 12.0
and allowed to “tweak” uphill until encountering a
(0.0) (0.5) (0.4) (0.1) (0.0)
(possibly local) peak.
Table 4 reports the results of this process. Panel A Description. For each level of N and K shown here, ten landscapes were
records the portion of incremental imitators that meet generated. On each landscape, a benchmark strategy was pinpointed as the best
or exceed the performance of the benchmark strategy. of 100 exploratory searches. Five hundred imitators were then released on each
The table also gives the gap between the benchmark landscape with randomly assigned initial strategies and were allowed to improve
their strategies incrementally (M ⫽ 1) until reaching a local peak. The panels
strategy and the highest point found by the typical
address various questions regarding the imitators’ performance. Each data point
imitator. This gap is measured in terms of both firm is the average over the ten landscapes, and the standard deviation of the average
value (panel B) and distance in decision space (panel is shown in parentheses below each data point.
C). As expected, when K ⫽ 0 incremental improve- Interpretation. As the underlying decision problem becomes complex, incremen-
ment always delivers the imitator to the single (global) tal improvement grows less successful.

836 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

peak, regardless of N. As interactions proliferate, Table 5 Evolution of the Benchmark Under the Incremental
however, the portion of copycats that match the Improvement Heuristic

benchmark strategy declines. By the N ⫽ 24, K ⫽ 11 Panel A: How many times is a new benchmark established (maximum ⫽
case, only 37 of the 5,000 imitators match the bench- 1,000)?
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

mark. Local peaks become numerous and the typical

imitator gets stranded well below and far away from K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
the best strategy. Panel B highlights the crucial role
N ⫽ 4 1.0 1.2 1.8
played by K, relative to N. It is the interaction among
N ⫽ 8 1.0 1.6 2.1 4.0
numerous decisions, not the mere proliferation of N ⫽ 12 1.0 1.6 2.6 4.2 5.5
decisions, that harms the incremental improver. N ⫽ 24 1.0 2.6 4.3 6.3 7.2
Table 5 examines the evolution of the benchmark
under the incremental improvement heuristic. For Panel B: How many times is an existing benchmark matched (but not
exceeded) by a subsequent firm (maximum ⫽ 999)?
each level of N and K, one thousand firms are released
sequentially on each of 20 landscapes, and each firm K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
exhausts its opportunities for local improvement.
Panel A records the number of times that a new firm N ⫽ 4 999 710 466
establishes a new, higher standard for performance. N ⫽ 8 999 600 251 62
N ⫽ 12 999 507 184 28 6
Panel B reports the number of times that a new firm
N ⫽ 24 999 205 57 7 0
matches (but does not exceed) the existing benchmark.
Panel C analyzes the number of firms released before Panel C: How many firms are released on average before a newly established
each new benchmark is matched or exceeded; i.e., it benchmark is matched or exceeded?
measures how long each new benchmark remains
K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
unbeaten. New, superior strategies are discovered
more frequently on complex landscapes (panel A), N ⫽ 4 1.0 1.7 1.7
and such strategies are far more rarely matched in N ⫽ 8 1.0 1.3 2.4 13.5
complex settings (panel B). Consequently, benchmark N ⫽ 12 1.0 2.4 4.5 13.7 55.6
firms have much longer runs of success when deci- N ⫽ 24 1.0 3.0 24.7 66.9 79.8
sions are numerous and intertwined (panel C).
Description. For each level of N and K shown here, 20 landscapes were
generated. On each landscape, 1,000 firms were released sequentially, each at
5. Imitation by Following the a random location. Each firm improved its strategy incrementally (M ⫽ 1) until
reaching a local peak. Each time a firm exceeded the highest value attained so
Leader far, that firm became the new benchmark. The panels address various questions
Surely, one might argue, incremental improvement is concerning the evolution of the benchmark. Each data point is an average over
a poor caricature of the imitation process. An imitator the 20 landscapes.
does not try to match a successful rival solely by Interpretation. As the underlying decision problem becomes complex, new
standards for performance are set more often, but those standards are much less
making uninformed, local “tweaks” to its own choices.
readily matched by others.
Rather, it observes the high performer closely: reverse-
engineers its products, interviews its customers and
former employees, speaks with its suppliers, dissects tive, just as it undermines incremental improvement. I
its public literature, benchmarks its every activity, and discuss two structural reasons for this, then turn to an
so forth. Once well informed, the imitator changes a integrated simulation.
large number of its choices—leaps in decision space
5.1. Structural Deterrents to Follow-the-Leader
toward the configuration of the successful firm—in
Imitation
hope of landing somewhere near the profitable strat-
egy. From there, it edges uphill. Vanishing Basins of Attraction. Suppose an imi-
Complexity makes this style of imitation less effec- tating firm can alter numerous choices at once.

Management Science/Vol. 46, No. 6, June 2000 837

 RIVKIN
Imitation of Complex Strategies

Management cannot guarantee that it will recreate the its distance in decision space from the highest peak
strategy of a high-performing rival exactly, however, found on the landscape. For N ⫽ 24 and K ⫽ 1, note
because it does not understand the rival’s operations the strong negative correlation. The higher peaks tend
precisely, because management’s control is imperfect, to be near “Everest” and the lower peaks far away. As
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

or because perfect replication is prohibitively expen- K rises, the correlation fades and high peaks spread
sive. Nonetheless, the would-be imitator hopes that it apart. The location of a high peak carries less and less
will come close enough to the successful set of deci- information about where other high summits can be
sions that subsequent “tweaks” will take it to the peak. found.
Such a wholesale reconfiguration is successful only if The intuition behind this is straightforward. When
the imitator’s leap lands in the basin of attraction choices are loosely connected, the strategy of a suc-
below the high performer’s peak. cessful firm can typically be decomposed into separa-
Unfortunately for the imitator, complexity under- ble subsystems of well-made choices (Simon 1962).
mines its leap. When decisions are numerous and Reconfiguring one of the subsystems has limited effect
richly cross-coupled, the attraction basins of high on the others and generates a new local peak that is
peaks become small, and the odds of jumping into quite close in decision space to the original one. The
them grow correspondingly poor. To see this, consider mountain range around Everest consists of Everest-
the extreme cases of K ⫽ 0 and K ⫽ N ⫺ 1, with M like decision configurations with individual sub-
⫽ 1. For K ⫽ 0, of course, the basin of attraction of the systems modified. In contrast, when choices are
global peak is the entire landscape regardless of N, tightly interlocked, the strategy of a successful firm
and incremental improvements will take a firm from cannot be broken into subsystems. High peaks are
any set of choices to the best configuration. The then likely to be altogether different ways of doing
accuracy of a long jump is irrelevant. For K ⫽ N ⫺ 1, business rather than slight variations on a theme.
however, the situation is very different. Consider the When interactivity is strong, slight variations destroy
global peak, which typically has among the broadest the theme.
basins of attraction. The portion of points from which Note the special role played by K in determining
a firm can tweak its way to the global peak dwindles these structural features. Never does an increase in N
rapidly with N. The basin of the global peak covers alone, the mere number of decisions a firm faces,
only 7% of the landscape when N ⫽ 8, K ⫽ 7, and less impede imitators. Both structural obstacles discussed
than 0.5% when N ⫽ 16, K ⫽ 15. here (as well as the two considered in §4.1) hinge
crucially on K. Interactions, then, are at the core of the
Spreading Peaks. The problem of the dwindling
imitation barrier proposed in this paper. High values
basins would not be so vexing for would-be copycats of N are important in deterring imitation largely
if high peaks on the landscape clustered together. If because they create rich opportunities for interactions.
peaks lay in “mountain ranges,” then an imitator
would, on average, find a long jump toward a high 5.2. Simulation Results
performer beneficial even if it failed to land in pre- Simulations confirm what the structural features pro-
cisely the right basin of attraction. The imitator might pose: complexity makes follow-the-leader imitation
not wind up atop Mount Everest, but at least it would less successful. Recall the key parameters governing
be in the Himalayas. the search processes: J, the number of decisions recon-
Once again, however, interactions work against the figured in a leap toward the benchmark, and ␪, the
imitator. As K increases, peaks on the landscape spread probability that each of the J decisions will be imitated
apart. This effect is illustrated in Figure 2. For N ⫽ 24 correctly in the reorientation.
and various values of K, I release 100 firms and allow First consider reorientations in isolation. As in §4.2,
them to move uphill incrementally until they discover I release 500 imitators on each of ten landscapes, but
local peaks. For each peak, Figure 2 plots two items now allow them to “leap” in decision space toward the
against one another: (1) the height of the peak, and (2) benchmark strategy. Of course, when J ⫽ N and ␪ ⫽ 1,

838 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

Figure 2 Spreading of Peaks for N ⴝ 24

Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

Description. For N ⫽ 24 and various values of K, 100 firms were released on a landscape and allowed to tweak uphill, one decision at a time, until reaching a local
peak. Each gray point represents a local peak discovered in this fashion. The point shows the height of the peak and the “strategic distance” of the peak from the
highest discovered on the landscape. The strategic distance between two peaks is defined as the number of decisions made differently in the two strategies that
constitute the peaks. This figure was inspired by Kauffman (1993), fig. 2.7.
Interpretation. For low K, the higher peaks tend to be nearer the highest peak. For high K, this clustering is not evident.

the imitator jumps to precisely the benchmark peak in high likelihood that imitators will fare poorly. Once
a single bound. However, even a small departure of ␪ again, panel B emphasizes the role of K, not N alone,
from 1 to 0.9 creates a substantial problem for the in making the imitator worse off.
imitator when the underlying decision problem is Next, I couple follow-the-leader and incremental
complex (Table 6). For high N and K, the typical search. The imitator’s hope here is that the two modes
imitator is much worse off than the benchmark firm of search will work together to overcome complexity:
(panel B). Note the interplay between N and K. As N a reconfiguration will position the firm in the valley
grows, the likelihood that the imitator will accurately beneath a good strategy, and subsequent incremental
replicate all of the benchmark firm’s decisions without improvement will achieve that strategy. Given enough
a single mistake ( ␪ N ) declines (panel A). As K in- cycles of following the leader and improving incre-
creases, the impact of each mistake rises (panel B). A mentally, the imitator will (almost) surely match the
high N makes errors likely, while a high K makes benchmark firm. It turns out, however, that the time
errors matter. In a strategy whose pieces are numerous required to do so may be lengthy. As results avail-
and tightly knit, small probabilities that each element able from the author show, the median number of
will be replicated incorrectly cumulate to produce a periods required for imitation to succeed is an order of

Management Science/Vol. 46, No. 6, June 2000 839

 RIVKIN
Imitation of Complex Strategies

Table 6 Results of Follow-the-Leader Imitation Efforts Table 7 Evolution of the Benchmark Under the Follow-the-Leader
Heuristic
Panel A: What portion of imitators match or exceed the benchmark firm’s
performance?
Panel A: How many times is a new benchmark established (maximum ⫽
1,000)?
K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
N ⫽ 4 66.1% 66.6% 66.4%
(0.8%) (0.6%) (0.8%)
N ⫽ 4 1.0 1.3 1.7
N ⫽ 8 43.8% 42.7% 42.7% 43.0%
N ⫽ 8 1.0 1.2 2.2 3.2
(0.4%) (0.8%) (0.7%) (0.6%)
N ⫽ 12 1.0 1.5 2.3 4.0 5.2
N ⫽ 12 28.1% 29.0% 30.9% 28.4% 28.4%
N ⫽ 24 1.0 2.6 3.8 5.9 7.3
(0.5%) (0.9%) (1.5%) (0.5%) (0.4%)
N ⫽ 24 11.6% 11.2% 10.3% 9.8% 8.9%
Panel B: How many times is an existing benchmark matched (but not
(0.5%) (0.9%) (0.8%) (0.7%) (0.6%)
exceeded) by a subsequent firm (maximum ⫽ 999)?
Panel B: How much worse off is the typical imitator than the benchmark firm,
as a percent of value? K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11

K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11 N ⫽ 4 999 927 762

N ⫽ 8 999 876 607 326
N ⫽ 4 5.0% 6.7% 9.9% N ⫽ 12 999 771 527 218 99
(0.3%) (0.9%) (1.4%) N ⫽ 24 999 631 366 98 35
N ⫽ 8 4.9% 7.2% 11.3% 15.9%
(0.4%) (0.5%) (0.4%) (1.2%) Panel C: How many firms are released on average before a newly established
N ⫽ 12 4.6% 7.9% 11.4% 17.8% 18.4% benchmark is matched or exceeded?
(0.3%) (0.5%) (0.8%) (0.7%) (0.8%)
N ⫽ 24 4.7% 7.2% 10.6% 15.4% 18.9% K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
(0.2%) (0.3%) (0.3%) (0.5%) (0.6%)
N ⫽ 4 1.0 1.0 1.2
Panel C: How many decisions does the typical imitator make differently from N ⫽ 8 1.0 1.3 1.5 2.3
the benchmark firm? N ⫽ 12 1.0 1.7 2.1 4.0 6.7
N ⫽ 24 1.0 1.7 2.5 6.3 17.9
K ⫽ 0 K ⫽ 1 K ⫽ 3 K ⫽ 7 K ⫽ 11
Description. For each level of N and K shown here, 20 landscapes were
N ⫽ 4 0.4 0.4 0.4 generated. On each landscape, 1,000 firms were released sequentially, each at
(0.0) (0.0) (0.0) a random location. Each firm reoriented itself toward the current benchmark ( ␪
N ⫽ 8 0.8 0.8 0.8 0.8 ⫽ 0.9, J ⫽ N/ 2), then moved incrementally (M ⫽ 1) until reaching a local
(0.0) (0.0) (0.0) (0.0) peak. Each time a firm exceeded the highest value attained so far, that firm
N ⫽ 12 1.2 1.2 1.2 1.2 1.2 became the new benchmark. The panels address various questions concerning
(0.0) (0.0) (0.0) (0.0) (0.0) the evolution of the benchmark. Each data point is the average over the 20
N ⫽ 24 2.4 2.4 2.4 2.4 2.4 landscapes.
(0.0) (0.0) (0.0) (0.0) (0.0) Interpretation. As the underlying decision problem becomes complex, new
Description. For each level of N and K shown here, ten landscapes were standards for performance are set more often, but those standards are much less
generated. One each landscape, a benchmark strategy was pinpointed as the readily matched by others.
best of 100 exploratory searches. Five hundred imitators were then released on
each landscape with randomly assigned initial strategies and were allowed to
attempt a “long jump.” That is, each imitator tried to replicate the N decisions of magnitude greater when N ⫽ 24 and K ⫽ 11 than it
the benchmark firm, with a ␪ ⫽ 90% chance of success on each decision. The is when N ⫽ 4 and K ⫽ 0.
panels address various questions regarding the imitators’ performance. Each A similar pattern emerges when one examines the
data point is the average over the ten landscapes, and the standard deviation of
evolution of the benchmark strategy under the follow-
the average is shown in parentheses below each data point.
Interpretation. Follow-the-leader efforts become ineffective as the underlying the-leader heuristic (Table 7). Good strategies are
decision problem grows complex. High N makes mistakes likely, and high K matched far less often when decisions are numerous
makes mistakes matter. and intertwined. As a result, firms that attain the best

840 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

performance to date have a much longer run of replicate it precisely (␪ ⫽ 1). Second, the barrier posed
success when strategic complexity is high. by complexity declines as incremental imitators are
allowed to search more broadly (M ⬎ 1). Though
diminished, the barrier remains so long as M ⬍ N.
6. Robustness of the Results
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

When imitators are permitted global search (M ⫽ N),

It is notoriously difficult to demonstrate the robust- the barrier disappears altogether. Similarly, and as
ness of simulation results conclusively. This is true noted in §3.3, the NP-completeness result loses its
because the space of possible model specifications and force if imitators can evaluate alternative strategies
parameter values is vast, far too large to explore by very rapidly and N is modest. Exhaustive search then
exhaustive search. Nonetheless, several features of becomes practical. Of these three conditions, the one
§§3–5 should reassure the reader that the main result, that seems most likely to arise in reality is that N, the
that complexity deters imitation, is robust. First, the number of decisions over which managers have true
analytical techniques employed from algorithm the- latitude, is limited. This might occur, for example, if
ory do not rely on simulation. Second, the reported managerial choice is rigidly constrained by national
results are averages over large numbers of simulation institutions.
trials. Third, the simulations span the full range of
parameter values permitted by a state-of-the-art per-
sonal computer. Fourth, the simulations explore a 7. Conclusion
variety of search heuristics. Fifth and finally, the The complexity of a strategy, coupled with limits on
discussion tries to elucidate the structural features that what managers know about rivals and can implement,
give rise to the simulation results, and the intuition raises a barrier to imitation. When the decisions that
behind those features, before it turns to simulation. embody a strategy are numerous and tightly linked to
An appendix available on request subjects the one another, a firm that discovers an effective combi-
model to a further battery of robustness tests. The nation of choices is protected against imitation in three
results are not sensitive to the arbitrary assumption ways. First, interactions among the decisions make the
that contributions C i are drawn from a U[0, 1] distri- strategy formulation problem intractable. When the
bution. Nor are they sensitive to the assumptions that number of decisions is large and strategy evaluation is
the K decisions affecting each C i are assigned at time consuming, this undermines any logical, algo-
random and that all decisions are equally influential rithmic attempt to rediscover the high performer’s
ex ante. I also consider three alternative ways to strategy. Second, a would-be imitator that tries to
identify the benchmark strategy: as the globally opti- match the successful strategy by incremental improve-
mal strategy; as the highest point discovered in an ment will soon find itself snared in a web of conflicting
exhaustive search over a fixed fraction of the land- constraints. Finally, a would-be imitator that attempts
scape; and as the highest point discovered in an to match the high performer’s entire system, via a
exhaustive search over a fixed number of strategies. In wholesale reconfiguration, is also likely to fail. Sup-
all cases, complexity makes imitators worse off, and pose it succeeds in imitating most of the high perform-
results are largely comparable to those reported here. er’s decisions but misses on a handful. Then because
The result of virtually any model can be shaken, the decisions’ effects are tightly knit together, the
however, if one alters the model dramatically enough. overall imitation attempt will probably be a failure.
An important question is, what are the least extreme Due to the complexity barrier to imitation, success-
modifications that undo the main result? In exploring ful strategies may remain unmatched even though
the model under a wide range of conditions, I noted they are open to public scrutiny and even after many
three modifications that neutralize complexity as an of their critical ingredients are adopted by competi-
obstacle to imitation. First, the barrier to follow-the- tors. This may help to explain the persistent superior
leader imitation disappears if imitators understand performance of certain firms, as mentioned in §1. It
the high performer’s strategy thoroughly and can may also shed light on the slow diffusion of complex

Management Science/Vol. 46, No. 6, June 2000 841

 RIVKIN
Imitation of Complex Strategies

bundles of practices whose superior performance is nated recent analyses of imitation. My claim is not that
well documented: e.g., modern manufacturing tech- other barriers to imitation are unimportant. Indeed,
niques (Milgrom and Roberts 1990), flexible produc- strategic complexity amplifies many of those barriers
tion systems in the auto industry (MacDuffie 1995), (and vice versa). Suppose, for instance, that lack of
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

bundles of high-performance work practices in the some piece of tacit knowledge prevents a would-be
steel finishing industry (Ichniowski et al. 1997), mod- imitator from matching one of the benchmark firm’s
ern logistical practices in the textile business (Hwang decisions. The damage this does to the copycat is
and Weil 1997), and science-driven research in phar- greater if the affected decision interacts widely with
maceuticals (Cockburn and Henderson 1996). There many other choices than if it is isolated. A small
may be some truth in the popular business wisdom misstep is more dangerous on rugged terrain than on
that the whole of a strategy can resist imitation more a smooth surface. 5
effectively than the sum of its parts. Though complementary in some ways to game-
The analysis is also consistent with Sorenson’s theoretic and resource-based views of strategy, a focus
(1997) finding that tight coupling confers a contempo- on complexity does lead one to an unconventional
raneous advantage on computer workstation makers. view of what choices are “strategic.” Traditionally,
Sorenson takes vertical integration as a proxy for the strategy scholars have paid closest attention to inher-
extent of interaction across decisions (K) and finds ently irreversible choices with large performance im-
that more integrated organizations have a lower mor- plications (e.g., Ghemawat 1991). In the present paper,
tality rate at a point in time. His interpretation is that however, it is difficult to label any isolated decision as
integration allows firms to realize operating efficien- strategic. No single choice is innately sticky. What
cies, exploit market power, tap synergies across verti- does appear strategic by traditional standards is the
cal stages, or reduce uncertainty. It is also possible, choice of a particular logic of configuration, a local
however, that tight integration is advantageous be- peak. The stickiness of individual choices is then
cause it protects a successful workstation maker from endogenous; once a firm adopts a specific configura-
copycats. tion or peak, changes in individual choices become
In addition to illuminating well-established pat- costly. Likewise, the consequences of individual deci-
terns, the analysis suggests new empirical fingerprints sions become large because they undermine the effec-
that complexity might leave behind. Where decisions
are numerous and interconnected, one expects imita- 5
Among the imitation barriers associated with the resource-based
tion attempts to be less successful, and perhaps less view of the firm, the nearest relative to complexity is “causal
common, than they are where strategies are relatively ambiguity.” Lippman and Rumelt (1982) argue on logical grounds
that a basic ambiguity concerning causal connections between
simple. This suggests that the distribution of firm
actions and results can make imitation an uncertain venture. They
performance will be broader and convergence to mean then show how uncertainty in imitation can lead to persistent
performance will be slower in settings with more performance differences and above-normal rates of return despite
complex strategies. Indeed, industries differ substan- free entry, profit-maximizing behavior, no scale economies, and no
tially in the range of performance levels they exhibit market power. They never explain, however, the roots of causal
ambiguity, nor do they specify how it deters imitation. The complex-
and in their speed of convergence (Waring 1996), and
ity explored here can certainly underlie causal ambiguity. When a
conventional industrial organization variables explain winning strategy consists of numerous, interlocking decisions, it is
these cross-industry differences incompletely. A nat- nearly impossible to assign credit for any part of the success to any
ural hypothesis is that proxies for industry complexity isolated choice. Success arises from the whole, not the parts, so
would help to explain the residual differences. cause and effect are blurred. On the other hand, complexity can
deter imitation even when causal ambiguity is altogether absent.
The impediment to imitation I propose arises purely
This arises, for instance, if the managers of the would-be imitator
from the character of the decision problem and the understand causal connections perfectly but have imperfect control
nature of search. It owes nothing to the game-theoretic over internal operations. Hence, complexity may underpin causal
considerations or resource barriers that have domi- ambiguity, but it may also operate independently of ambiguity.

842 Management Science/Vol. 46, No. 6, June 2000

 RIVKIN
Imitation of Complex Strategies

tiveness of accompanying choices. In such a setting, immutable. In reality, managers make investments to
decisions that seem minor or easily altered can take on build connections across decisions or to sever links,
surprising salience. Recent research on complemen- i.e., to modularize (Baldwin and Clark 2000). Hence a
tary bundles of organizational practices gives at least successful firm might build new and elaborate connec-
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

some evidence that this occurs. tions between, say, its production and marketing
The world depicted in this paper could lead one to operations in order to fend off imitators. Alternatively,
a fatalistic view of management: seemingly minor an imitator may try to break a successful strategy into
decisions can be salient, relationships among decisions discrete pieces and attack each piece in turn. A firm
are thoroughly unknown a priori, algorithmic analysis might expand its scope of operations, increasing N, or
of decision problems is doomed, and reasonable heu- focus on a portion of the total value chain, reducing N.
ristics leave firms trapped in webs of conflicting In altering the complexity of its strategy, a firm will
constraints. What role is there, then, for management face the tension identified in §1: a more modular, less
science or managerial choice in a complex world? My coupled system adjusts more readily to environmental
belief is that major roles remain. I highlight two. change and is less subject to inertia (Levinthal 1997,
First, the analysis in this paper should make clear Sorenson 1997), but it is more easily attacked by
the power and importance of anticipating the relation- imitators. Baldwin and Clark (2000) observe precisely
ships among decisions. An imitator who understands this tension in the evolution of the computer industry.
that a particular pair of choices should always be By adopting modular architectures in the mainframe
adjusted in unison effectively reduces the number of industry and later in the personal computer market,
decisions and the number of interactions that she must IBM made itself and its products far more responsive
actively consider. As the numerical results make clear, to market needs. This approach, however, allowed
modest reductions in the dimensionality of the deci- widespread entry by imitators. Loosely coupled orga-
sion problem can make imitation significantly easier. nizations, it appears, are responsive to change, but
Even an understanding that certain variables usually, vulnerable to imitation. A natural implication, worthy
though not always, move together can be powerful. of further research, is that designers of organizations
must decide which they fear more: inertia or imita-
This is one reason that recent work on complementa-
tion. 6
rity and supermodularity have generated such inter-
est. As discussed in §3, the structure imposed by 6
I am grateful to George Baker, Pankaj Ghemawat, Tarun Khanna,
supermodularity can greatly simplify managerial Daniel Levinthal, and Michael Porter for especially incisive com-
search. This also helps to explain recurring efforts to ments and to Howard Brenner for computer-programming wiz-
classify strategies into a manageable number of ge- ardry. Stuart Kauffman influenced the paper profoundly through
neric types (e.g., Miles and Snow 1978, Porter 1980). his writings. Thanks to anonymous referees and editors, Susan
Athey, Carliss Baldwin, Stephen Bradley, Adam Brandenburger,
Such taxonomies hold out the hope that only a hand-
Richard Caves, Ken Corts, Rebecca Henderson, Robert Kennedy,
ful of coherent sets of choices exist. If true, this Elon Kohlberg, David Laibson, Roger Martin, Anita McGahan,
simplifies the managerial search problem dramati- Cynthia Montgomery, John Pratt, Mark Seasholes, Nicolaj
cally. Empirical support for taxonomies is mixed (e.g., Siggelkow, Olav Sorenson, Donald Topkis, Arthur Veinott, and
Miller and Dess 1993), suggesting that they may participants in seminars at the University of Chicago, Columbia,
Harvard, and UCLA for formative discussions and comments.
simplify decision problems too much. An important
Errors remain my own.
role of management science is to isolate relationships
that truly, reliably exist among decisions and thereby References
to simplify decision problems (without simplifying Baldwin, C. Y., K. B. Clark. 2000. Design Rules: The Power of
Modularity. MIT Press, Cambridge, MA.
them falsely).
Barney, J. 1991. Firm resources and sustained competitive advan-
Analyses in the paper also hint at the potency of tage. J. Management 17 99 –120.
altering the relationships among decisions. In the for- Cockburn, I., R. M. Henderson. 1996. Exploring inertia: Com-
mal models described here, managers take N and K as plementarities and firm effects in pharmaceutical research.

Management Science/Vol. 46, No. 6, June 2000 843

 RIVKIN
Imitation of Complex Strategies

Working Paper, Sloan School of Management, Massachusetts , . 1995. Complementarities and fit: Strategy, structure, and
Institute of Technology, Cambridge, MA. organizational change in manufacturing. J. Accounting Econom.
Dierickx, I., K. Cool. 1989. Asset stock accumulation and sustain- 19 179 –208.
ability of competitive advantage. Management Sci. 35(12) 1504 – Miller, A., G. G. Dess. 1993. Assessing Porter’s (1980) model in terms
1511. of its generalizability, accuracy, and simplicity. J. Management
Downloaded from informs.org by [130.226.44.8] on 10 December 2023, at 05:24 . For personal use only, all rights reserved.

Garey, M. R., D. S. Johnson. 1979. Computers and Intractability: A Stud. 30 553–585.

Guide to the Theory of NP-Completeness. Freeman, San Francisco. Miller, D., P. H. Friesen, H. Mintzberg. 1984. Organizations: A
Ghemawat, P. 1991. Commitment: The Dynamic of Strategy. Free Press, Quantum View. Prentice-Hall, Englewood Cliffs, NJ.
New York. Mueller, D. 1986. Profits in the Long Run. Cambridge University
Glassman, R. B. 1973. Persistence and loose coupling in living Press, Cambridge, UK.
systems. Behavioral Sci. 18 83–98. Nelson, R., S. Winter. 1982. An Evolutionary Theory of Economic
Grötschel, M., L. Lovász, A. Schrijver. 1988. Geometric Algorithms and Change. Belknap, Cambridge, MA.
Combinatorial Optimization. Springer-Verlag, Berlin, Germany. Porter, M. E. 1980. Competitive Strategy. Free Press, New York.
Hwang, M., D. Weil. 1997. Production complementarities and the . 1996. What is strategy? Harvard Bus. Rev. 74(6) 61–78.
diffusion of modern manufacturing practices: Evidence from Schumpeter, J. A. 1942. Capitalism, Socialism, and Democracy. Harper,
the U.S. apparel industry. Working paper, Boston University, New York.
Boston, MA. Siggelkow, N. 1998. Benefits of focus, evolution of fit, and agency
Ichniowski, C., K. Shaw, G. Prennushi. 1997. The effects of human issues in the mutual fund industry. Unpublished Ph.D. Disser-
resource management practices on productivity: A study of tation, Harvard University, Cambridge, MA.
steel finishing lines. Amer. Econom. Rev. 87 291–313. Simon, H. A. 1962. The architecture of complexity. Proc. Amer.
Kauffman, S. A. 1993. The Origins of Order: Self-Organization and Philos. Soc. 106 467– 482.
Selection in Evolution. Oxford University Press, Oxford, U.K. Sorenson, O. J. 1997. The complexity catastrophe in the computer
Levinthal, D. 1997. Adaptation on rugged landscapes. Management industry: Interdependence and adaptability in organizational
Sci. 43 934 –950. evolution. Unpublished Ph.D. Dissertation, Stanford Univer-
Lippman, S., R. Rumelt. 1982. Uncertain imitability: An analysis of sity, Stanford, CA.
interfirm differences in efficiency under competition. Bell J. Topkis, D. M. 1998. Supermodularity and Complementarity. Princeton
Econom. 12 418 – 438. University Press, Princeton, NJ.
MacDuffie, J. P. 1995. Human resource bundles and manufacturing Tyre, M. 1991. Managing the introduction of new process technol-
performance: Organizational logic and flexible production sys- ogy: International differences in a multi-plant network. Res.
tems in the world auto industry. Indust. Labor Relations Rev. 48 Policy 20 57–76.
197–221. Waring, G. F. 1996. Industry differences in the persistence of
MacMillan, I., M. L. McCaffery, G. van Wijk. 1985. Competitors’ firm-specific returns. Amer. Econom. Rev. 86 1253–1265.
responses to easily imitated new products: Exploring commercial Weick, K. E. 1976. Educational organizations as loosely coupled
banking product introductions. Strategic Management J. 6 75– 86. systems. Admin. Sci. Quart. 21 1–19.
Mansfield, E. 1985. How rapidly does new industrial technology Weinberger, E. D. 1991. Local properties of Kauffman’s N-k
leak out? J. Indust. Econom. 34 217–223. model: A tunably rugged energy landscape. Phys. Rev. A 44
March, J. G., H. A. Simon. 1958. Organizations. Wiley, New York. 6399 – 6413.
McKelvey, B. 1999. Avoiding complexity catastrophe in coevolu- Winter, S. G. 1987. Knowledge and competence as strategic assets.
tionary pockets: Strategies for rugged landscapes. Organ. Sci. 10 D. J. Teece, ed. The Competitive Challenge: Strategies for Industrial
294 –321. Innovation and Renewal. Ballinger, Cambridge, MA.
Miles, R., C. Snow. 1978. Organizational Strategy, Structure, and Wright, S. 1931. Evolution in Mendelian populations. Genetics 16
Process. McGraw-Hill, New York. 97–159.
Milgrom, P., J. Roberts. 1990. The economics of modern manufac- Zander, U., B. Kogut. 1995. Knowledge and the speed of the transfer
turing: Technology, strategy, and organization. Amer. Econom. and imitation of organizational capabilities: An empirical test.
Rev. 80 511–528. Organ. Sci. 6 76 –92.

Accepted by Rebecca Henderson; received December 19, 1997. This paper was with the author 12 months for 2 revisions.

844 Management Science/Vol. 46, No. 6, June 2000