doi: 10.1111/j.1467-6419.2012.00730.

THE NORMALIZED CES PRODUCTION

FUNCTION: THEORY AND EMPIRICS
Rainer Klump
Goethe University, Frankfurt am Main, and Center for Financial Studies

Peter McAdam
European Central Bank & University of Surrey

Alpo Willman
European Central Bank

Abstract. The elasticity of substitution between capital and labor and, in turn, the direction of
technical change are critical parameters in many fields of economics. Until recently, though, the
application of production functions with specifically non-unitary substitution elasticities (i.e., non-
Cobb–Douglas) was hampered by empirical and theoretical uncertainties. As recently revealed,
‘normalization’ of production-technology systems holds out the promise of resolving many of
those uncertainties. We survey and assess the intrinsic links between production (as conceptualized
in a production function), factor substitution (as made most explicit in Constant Elasticity of
Substitution functions) and normalization (defined by the fixing of baseline values for relevant
variables). First, we recall how the normalized Constant Elasticity of Substitution function came
into existence and what normalization implies for its formal properties. Then we deal with the key
role of normalization in recent advances in the theory of business cycles and of economic growth.
Next, we discuss the benefits normalization brings for empirical estimation and empirical growth
research. Finally, we identify promising areas of future research.

Keywords. Constant Elasticity of Substitution Production Function; Estimation; Factor-

Augmenting Technical Change; Growth Theory; Identification; Normalization.

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31


C 2012 Blackwell Publishing Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden,

MA 02148, USA.
2 KLUMP ET AL.

Until the laws of thermodynamics are repealed, I shall continue to relate outputs to inputs – i.e. to
believe in production functions.
Samuelson (1972) (p. 174)
All these results, negative and depressing as they are, should not surprise us. Bias in technical progress
is notoriously difficult to identify.
Kennedy and Thirlwall (1973) (p. 784)
... an increase in the elasticity of substitution will be shown to have far more importance for society’s
future welfare than a similar increase in the rate of technical progress.
La Grandville (2009) (p. xi)

1. Introduction
Substituting scarce factors of production by relatively more abundant ones is a key element of economic
efficiency and a driving force of economic growth. A measure of that force is the elasticity of
substitution between capital and labor which is the central parameter in production functions, and in
particular Constant Elasticity of Substitution (CES) ones. CES production functions allow the elasticity
of substitution to be any positive number; in contrast, the more well-known Cobb–Douglas variant
imposes that elasticity to be unity.
Until recently, the application of production functions with non-unitary substitution elasticities was
hampered by empirical and theoretical uncertainties. As has recently been revealed, ‘normalization’ of
production functions and production-technology systems holds out the promise of resolving many of
those uncertainties and allowing considerations such as the role of the substitution elasticity and biased
technical change to play a deeper role in growth and business-cycle analysis.
Normalization essentially implies representing production relations in consistent indexed number
form. Without normalization, it can be shown that the production function parameters have no economic
interpretation since they are dependent on the normalization point and the elasticity of substitution itself.
This feature significantly undermines estimation and comparative-static exercises, among other things.
Notwithstanding, this leaves open the issue of how we set and interpret normalization points. All
CES production functions, unless explicitly normalized, are at least implicitly normalized in the point
where input values equal one. However, that implicit normalization is empirically counter-factual and,
from a theory standpoint, unattractive. Accordingly, normalization points tend to be pinned down
by some prevailing economic theory or empirical counterpart – typically some steady-state or initial
condition, or in some neighborhood of particular interest to the researcher. Beyond that, normalization
(explicit normalization) also implies internal consistency between other aspects of the data or model.
As we shall see, if this internal consistency condition is violated – as typically has been the case
in theoretic analysis based on the non-normalized (or ‘trivially’ normalized) CES function – then
analysis concerning the effects of alternative elasticity of substitution values on economic development
is flawed.
Let us first, though, place the importance of the topic in perspective. Due to the central role of
the substitution elasticity in many areas of dynamic macroeconomics, the concept of CES production
functions has recently experienced a major revival. The link between economic growth and the size
of the substitution elasticity has long been known. As already demonstrated by Solow (1956) in
the neoclassical growth model, assuming an aggregate CES production function with an elasticity
above unity is the easiest way to generate perpetual growth. Since scarce labor can be completely
substituted by capital, the marginal product of capital remains bounded away from zero in the long
run. Nonetheless, as we argue below, the case for an above-unity elasticity appears empirically weak
and theoretically anomalous.1
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 3

It has been shown that integration into world markets is also a feasible way for a country to increase
the effective substitution between factors of production and thus pave the way for sustained growth
(Ventura 1997; Klump 2001; Saam 2008). On the other hand, it can be shown in several variants of the
standard neoclassical (exogenous) growth model that introducing an aggregate CES production function
with an elasticity of substitution below unity can generate multiple growth equilibria, development traps
and indeterminacy (Azariadis 1996; Klump 2002; Kaas and von Thadden 2003; Guo and Lansing 2009).
Public finance and labor economics are other fields where the elasticity of substitution has been
rediscovered as a crucial parameter for understanding the impact of policy changes. This relates to
the importance of factor substitution possibilities for the demand for each input factor. As pointed
out by Chirinko (2002), the lower the elasticity of substitution, the smaller the response of business
investment to variations in interest rates caused by monetary or fiscal policy.2 In addition, the welfare
effects of tax policy changes, specifically, appear highly sensitive to the values of the substitution
elasticity. Rowthorn (1999) also stresses its importance in macroeconomic analysis of the labor market
and, in particular, how incentives for investment exercise a significant effect on unemployment when
the elasticity of substitution departs from unity.
Indeed, there is now mounting empirical evidence that aggregate production is better characterized
by a non-unitary (and in particular below unitary) elasticity of substitution (see e.g., Chirinko et al.
1999; Klump et al. 2007; León-Ledesma et al. 2010a). Chirinko (2008)’s recent survey suggests that
most evidence favors elasticities ranges of 0.4–0.6 for the United States. Moreover, Jones (2003);
,2005)3 argued that capital shares exhibit such protracted swings and trends in many countries as to
be inconsistent with Cobb–Douglas or CES with Harrod-neutral technical progress (see also Blanchard
1997; McAdam and Willman 2013). Such variability would also suggest the presence of biases in
technical change.
The coexistence of capital and labor-augmenting technical change has different implications for the
possibility of balanced or unbalanced growth. A balanced growth path (BGP) – the dominant assumption
in the theoretical growth literature – suggests that variables such as output, consumption, etc. tend to
a common growth rate, whilst key underlying ratios (e.g., factor income shares, capital–output ratio)
are constant (Kaldor 1961). Neoclassical growth theory suggests that, for an economy to possess a
steady state with positive growth and constant factor income shares, the elasticity of substitution must
be unitary (i.e., Cobb–Douglas) or technical change must be Harrod-neutral.
As Acemoglu (2009) (ch. 15) comments, however, there is little reason to assume technical change is
necessarily labor-augmenting.4 In models of ‘biased’ technical change (e.g., Kennedy 1964; Samuelson
1965; Acemoglu 2003; Sato 2006), scarcity, reflected by relative factor prices, generates incentives
to invest in factor-saving innovations. In other words, firms reduce the need for scarce factors and
increase the use of abundant ones. Acemoglu (2003) further suggested that while technical progress is
necessarily labor-augmenting along the BGP, it may become capital-biased in transition. Interestingly,
given a below-unitary substitution elasticity this pattern promotes the stability of income shares while
allowing them to fluctuate in the medium run.
However, when analytically investigating the significance of non-unitary factor substitution and non-
neutral technical change in dynamic macroeconomic models, one faces the issue of ‘normalization’,
even though the issue is still not widely known. The (re)discovery of the CES production function in
normalized form in fact paved the way for the new and fruitful, theoretical and empirical research on
the aggregate elasticity of substitution which has been witnessed over recent years.
In La Grandville (1989b) and Klump and de La Grandville (2000) the concept of normalization was
introduced in order to prove that the aggregate elasticity of substitution between labor and capital can
be regarded as an important and meaningful determinant of growth in the neoclassical growth model.
In the meantime this approach has been successfully applied in a series of theoretical papers (Klump
2001; Papageorgiou and Saam 2008; Klump and Irmen 2009; Xue and Yip 2013; Guo and Lansing
2009; Wong and Yip 2010) to a wide variety of topics.
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
4 KLUMP ET AL.

A particular striking example of how neglecting normalization can significantly bias results and how
explicit normalization can help to overcome those biases is presented in Klump and Saam (2008). The
effect of a higher elasticity of substitution on the speed of convergence in a standard Ramsey type
growth model is shown to double if an implicitly normalized CES function is replaced by a reasonably
normalized one.
Further, as Klump et al. (2007, 2008) demonstrated, normalization also has been a significant
development for empirical research on the parameters of aggregate CES production functions,5 in
particular when coupled with the system estimation approach. Empirical research has long been
hampered by the difficulties in identifying at the same time an aggregate elasticity of substitution
as well as growth rates of factor-augmenting technical change from the data. Following Diamond et
al. (1978), the received wisdom suggests that their joint identification was infeasible. Accordingly,
for more than a quarter of a century following Berndt (1976), common opinion held that the US
economy was broadly characterized by aggregate Cobb–Douglas technology, leading, in turn, to its
default incorporation in economic models (and, accordingly, the neglect of possible biases in technical
progress in empirical work).6
Translating normalization into empirical production-technology estimations allows the presetting of
the value of the distribution parameter (or, if estimated, facilitates the setting of reasonable initial
conditions); it provides a clear correspondence between theoretical and empirical production function
parameters and allows us ex post validation of estimated parameters. In a series of papers, León-Ledesma
et al. (2010a); ,2010b) showed the empirical advantages in estimating and identifying production-
technology systems when normalized. Further, McAdam and Willman (2011b) showed that normalized
factor-augmenting CES estimation, in the context of estimating ‘New Keynesian’ Phillips curves, helped
better identify the volatility in the driving variable (real marginal costs) that most previous researchers
had not detected.
Here, we analyze the intrinsic links between production (as conceptualized in a production function),
factor substitution (as made most explicit in CES production functions) and normalization. The paper
is organized as follows. In Section 2 we recall how the CES function came into existence and what this
implies for its formal properties. Sections 3 and 4 will deal with the role of normalization in recent
advances in the theory of business cycles and economic growth. Section 5 will discuss the merits
normalization brings for empirical growth research. The last section concludes and identifies promising
area of future research.

2. The Normalized CES Production Function and Variants

It is common knowledge that the first rigid derivation of the CES production function appeared in the
famous Arrow et al. (1961) paper (hereafter ACMS).7 However, there were important forerunners, in
particular the explicit mentioning of a CES-type production technology (with an elasticity of substitution
equal to 2) in the Solow (1956) paper (done, Solow wrote, to add a ‘bit of variety’) on the neoclassical
growth model. There was also the hint to a possible CES function in its Swan (1956) counterpart (on
the Swan story see Dimond and Spencer 2008).8 Shortly before, though, Dickinson (1954) (p. 169,
fn 1) had already made use of a CES production technology in order to model ‘a more general kind
of national-income function, in which the factor shares are variable’ compared to the Cobb–Douglas
form. It has even been conjectured that the famous and mysterious tombstone formula of von Thünen
dealing with ‘just wages’ can be given a meaningful economic interpretation if it is regarded as derived
from an implicit CES production function with an elasticity of substitution equal to 2 (see Jensen
2011).
In this section we want to demonstrate that the formal construction of a CES production
function is intrinsically linked to normalization, and how it is linked. The function may be defined
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 5

as follows:
 σ −1 σ −1  σ −1
σ

Yt = F (K t , Nt ) = C π K t σ + (1 − π) Nt σ (1)

where distribution parameter π ∈ (0, 1) reflects capital intensity in production; C is an efficiency

parameter, and σ is the elasticity of substitution between capital, K , and labor, N . Like all standard
CES functions, equation (1) nests a Cobb–Douglas function when σ = 1; a Leontief function with
fixed factor proportions when σ = 0 and a linear production function with perfect factor substitution
when σ → ∞.
The construction of such an aggregate production technology with a CES property starts from the
formal definition of the elasticity of substitution which had been introduced independently by Hicks
(1932) and Robinson (1933) (on the differences between the two approaches to the concept, see Hicks
1970). It is there defined (in the case of two factors of production, capital and labor) as the elasticity
of K /N with respect to the marginal rate of substitution between K and N (the percentage change in
factor proportions due to a change in the marginal rate of technical substitution) along an isoquant:9

d (K /N ) / (K /N ) d log (K /N )
σ ∈ [0, ∞) = = (2)
d (FN /FK ) / (FN /FK ) d log (FN /FK )

As Hicks notes this concept of elasticity can be equally expressed in terms of the second derivative
of the production function, but only under the assumption of constant returns to scale (due to Euler’s
theorem).
Since under this assumption the marginal factor productivities would also equal factor prices and
the marginal rate of substitution would be identical with the wage/capital rental ratio, the elasticity of
substitution can also be expressed as the elasticity of income per person y with respect to the marginal
product of labor in efficiency terms (or the real wage rate, w), that is, Allen’s theorem (Allen 1938).
Given that income per person is a linear homogeneous function y = f (k) of the capital intensity
k = K /N , the elasticity of substitution can also be defined as:

dy w f´(k) [ f (k) − k f´(k)]

σ = · =− (3)
dw y k f´´´(k) f (k)

Although it is rarely stated explicitly, the elasticity of substitution is implicitly always defined as a
point elasticity. This means that it is related to one particular baseline point on one particular isoquant
(see Figures 1 and 2). From there a whole system of non-intersecting isoquants is defined which all
together create the CES production function. Even if it is true that a given and constant elasticity of
substitution would not change along a given isoquant or within a given system of isoquants, it is also
evident that changes in the elasticity of substitution would of course alter the system of isoquants.
Following such a change in the elasticity of substitution, the old and the new isoquant are not
intersecting at the baseline point but are tangents, if the production function is normalized (by the values
of the baseline points). And they should not intersect because given the definition of the elasticity of
substitution (i.e., the percentage change in factor proportions due to a change in the marginal rate of
technical substitution) at this particular point (as in all other points which are characterized by the
same factor proportion) the old and the new CES function should still be characterized by the same
factor proportion and the same marginal rate of technical substitution.
Just as there are two possible definitions of σ following (3) – from dw dy
· wy and then from
f´(k)[ f (k)−k f´(k)]
− k f´´´(k) f (k) – thus there are two ways of uncovering the normalized production function. These, we
cover in the following two sub-sections.
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
6 KLUMP ET AL.

Figure 1. Isoquants of Normalized CES Production Functions.

Figure 2. Normalized per-capita CES Production Functions.

2.1 Derivation via the Power Function

d log(y) w
Let us start from the definition σ = d log(w)
= dy
dw
· y
, integration of which gives the power function,

y = cw σ (4)

where c is some integration constant.10 Under the assumption of constant returns to scale (or perfectly
competitive factor and product markets), and applying the profit-maximizing condition that the real
wage equals the marginal product of labor, and with the application of Allen’s theorem, we can
transform this equation into the form y = c(y − k dy
dk
)σ .
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 7

Accordingly, after integration and simplification, this leads us to a production function with the
constant elasticity of substitution function (see La Grandville 2009, p. 83ff for further details):
 σ −1  σ σ−1
y = βk σ + α (5)

and,
 σ −1 σ −1
 σ σ−1
Y = β K σ + αL σ (6)

in the extensive form.

It should be noted that (5) and (6) contain the two constants of integration β and α = c− σ , where
1

the latter directly depends on σ . Identification of these two constants makes use of baseline values
for the power function (4) and for the functional form (5) at the given baseline point in the system
of isoquants. In a dynamic setting this baseline point must (as we will see later) also be regarded as
holding at a particular point in time, t = t0 :
y0 = cw0σ (7)

 σ −1  σ σ−1
y0 = βk0 σ + α (8)

Together with (5) this leads to the normalized CES production function,
   σ −1  σ σ−1
k σ
y = y0 π0 + (1 − π0 ) (9)
k0

and
   σ σ−1   σ σ−1  σ σ−1
K N
Y = Y0 π0 + (1 − π0 ) (10)
K0 N0

in the extensive form. Parameter π0 = y0 −w

y0
0
= r0YK0 0 denotes the capital share in total income at the
point of normalization. As a test of consistent normalization, we see from (10) that for t = t0 we
11

retrieve Y = Y0 .

2.2 Derivation via the Homogeneous Production Function

It was shown by Paroush (1964), Yasui (1965) and McElroy (1967) that the rather narrow assumption
of Allen’s theorem is not essential for the derivation of the CES production function which can start
directly from the original Hicks definition (2). This definition can be transformed into a second-order
differential equation whose solution also implies two constants of integration.
Following Klump and Preissler (2000) we start with the definition of the elasticity of substitution in
the case of linear homogeneous production function Yt = F(K t , Nt ) = Nt f (kt ) where kt = K t /Nt is
the capital–labor ratio in efficiency units. Likewise yt = Yt /Nt represents per-capita production.
The definition of the substitution elasticity, σ = − f´(k)[k f´f´(k)
(k)−k f´(k)]
f (k)
,can then be viewed as a second-
order differential equation in k having the following general CES production function as its solution
(intensive and extensive forms),
 σ −1  σ σ−1
yt = a kt σ + b (11)

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31

C 2012 Blackwell Publishing Ltd
8 KLUMP ET AL.

 σ −1 σ −1  σ −1
σ

Yt = a K t σ + bNt σ (12)

where parameters a and b are two arbitrary constants σ

of integration with the following correspondence
with the parameters in equation (1): C = a(1 + b) σ −1 and π = 1/(1 + b).
A meaningful identification of these two constants is given by the fact that the substitution elasticity
is a point elasticity relying on three baseline values: a given capital intensity k0 = K 0 /N0 , a given
marginal rate of substitution [FK /FN ]0 = w0 /r0 and a given level of per-capita production y0 = Y0 /N0 .
Accordingly, (1) becomes

   σ σ−1   σ σ−1  σ σ−1

Kt Nt
Yt = Y0 π0 + (1 − π0 ) (13)
K0 N0

where π0 = r0 K 0 /(r0 K 0 + w0 N0 ) is the capital income share evaluated at the point of normalization.
Rutherford (2003) calls (13) [or (10)] the ‘calibrated form’.

2.3 A Graphical Representation

Normalization as understood by La Grandville (1989b), Klump and de La Grandville (2000) and
Klump and Preissler (2000) is again nothing else but identifying these two arbitrary constants in an
economically meaningful way. Normalizing means the fixing (in the K − N plane as in Figure 1) of a
baseline point (which can be thought of as a point in time at t = t0 ), characterized by specific values
of N , K , Y and the marginal rate of technical substitution μ0 – in which isoquants of CES functions
with different elasticities of substitution but with all other parameters equal – are tangents.
Normalization is helpful to clarify the conceptual relationship between the elasticity of substitution
and the curvature of the isoquants of a CES production function (see La Grandville 1989a for a
discussion of various misunderstandings on this point). Klump and Irmen (2009) point out that in the
point of normalization (and only there), there exists an inverse relationship between the elasticity of
substitution and the curvature of isoquant of the normalized CES production function. This relationship
also has an interpretation in terms of the degree of complementarity of both input factors. At the
normalization point, a higher elasticity of substitution implies a lower degree of complementarity
between the input factors. The link between complementarity between input factors and the elasticity
of substitution is also discussed in Acemoglu (2002a) and in Nakamura and Nakamura (2008).
Equivalently (in the k − y plane as in Figure 2) the baseline point can be characterized by specific
values of k, y and the marginal productivity of capital (or the real wage rate). If base values for
these three variables are selected, this means of course that also a baseline value for the elasticity
of production with respect to capital input is fixed which (under perfect competition) is equal to the
capital share in total income.

2.4 Normalization as a Means to Uncover Valid CES Representations

Normalization thus creates specific ‘families’ of CES functions whose members all share the same
baseline point but are distinguished by the elasticity of substitution (and only the elasticity of
substitution).
As shown in Klump and Preissler (2000), normalization also helps to distinguish those variants of
CES production functions which are functionally identical with the general form (1) from those which
are inconsistent with (5) in one way or another.
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 9

Consider, first, the ‘standard form’ of the CES production function, as it was introduced by ACMS,
restated below:
 σ −1 σ −1
 σ σ−1
Y = C π K σ + (1 − π) N σ (14)

This variant is clearly identical with (10), albeit (and this is a crucial aspect) with the ‘efficiency
parameter’ C and the ‘distribution parameter’ π being defined in the following way (solving for
completeness in terms of both σ and the substitution parameter ρ = 1−σ σ
):
 1+ρ 1+ρ
− ρ1  1/σ 1/σ
 σ σ−1
r 0 K 0 + w0 N 0 r 0 K 0 + w0 N 0
C (σ, ·) = Y0 = Y0 (15)
r 0 K 0 + w0 N 0 r 0 K 0 + w0 N 0

1+ρ 1/σ
r0 K 0 r0 K 0
π (σ, ·) = 1+ρ 1+ρ
= 1/σ 1/σ
(16)
r0 K 0 + w0 N 0 r0 K 0 + w0 N 0
Expressions (15) and (16) reveal that, in the implicitly normalized case, both ‘parameters’ (apart from
being dependent on the scale of the normalized variables) change with variations in the elasticity of
substitution, unless the particular case of K 0 and N0 are exactly equal arises, implying k0 = 1.
This makes the implicitly normalized form in general inappropriate for comparative-static exercises
in the substitution elasticity. It is the interaction between the normalized efficiency and distribution
terms and the elasticity of substitution which guarantees that within one family of CES functions the
members are only distinguished by the elasticity of substitution. Given the accounting identity (and
abstracting from the absence of an aggregate mark-up),
Y 0 = r 0 K 0 + w0 N 0 (17)
it also follows from this analysis that treating C and π in (14) as deep parameters is equivalent to
assuming k0 = 1. In the case σ = 0, we have a perfectly symmetrical Leontief function.
As explained in Klump and Saam (2008) the Leontief case can serve as a benchmark for the choice of
the normalization values for k0 in calibrated growth models. The baseline capital intensity corresponds
to the capital intensity that would be efficient if the economy’s elasticity of substitution were zero.
For k < k0 the economy’s relative bottleneck resides in this case in its capacity to make productive
use of additional labor, as capital is the relatively scarce factor. For k > k0 the same is true for capital
and labor is relatively scarce. Since the latter case is most characteristic for growth model of capitalist
economies, calibrations of these model can be based on the assumption k > k0 .
In the following sub-sections, we will illustrate how normalization can reveal whether certain
production functions used in the literature are legitimate.

2.4.1 David and van de Klundert (1965) Version

Consider the CES variant proposed by David and van de Klundert (1965):
 σ −1 σ −1
 σ σ−1
Y = (B K ) σ + (AN ) σ (18)

This variant is identical with (10) as long as the two ‘efficiency levels’ are defined in the following
way:
Y0 σ σ−1
B= π (19)
K0 0
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
10 KLUMP ET AL.

Y0 σ
A= (1 − π0 ) σ −1 (20)
N0

Again, it is obvious that the efficiency levels change directly with the elasticity of substitution.

2.4.2 Ventura (1997) Version

Consider now a CES variant used by Ventura (1997):
 σ −1 σ −1
 σ σ−1
Y = K σ + (AN ) σ (21)

At first glance (21) could be regarded as a special case of (14) with B being equal to one. With a
view on the normalized efficiency level it becomes clear, however, that B = 1 is not possible for given
baseline values and a changing elasticity of substitution. Given that Ventura (1997) makes use of (21)
in order to study the impact of changes in the elasticity of substitution on the speed of convergence,
in the light of this inconsistency his results should be regarded with caution. Indeed, as shown in
Klump (2001), Ventura’s results are unnecessarily restrictive; working with a correctly normalized
CES technology leads to much more general results.

2.4.3 Barro and Sala-i-Martin (2004) Version

Next consider the CES production function proposed by Barro and Sala-i-Martin (2004):
 σ −1 σ −1
 σ σ−1
Y = C π (B K ) σ + (1 − π) ((1 − B) N ) σ (22)

Normalization is helpful in this case in order to show that (22) can be transformed without any problems
into (10) and/or (14) so that the terms B and 1 − B simply disappear. If for any reason these two
terms are considered necessary elements of a standard CES production function, they cannot be chosen
independently from the normalized values for C and π, but they remain independent from changes
in σ .

2.5 The Normalized CES Function with Technical Progress

So far, we have treated efficiency levels as constant over time. If we now consider factor-augmenting
technical progress one has to keep in mind the intrinsic links between rising factor efficiency and
the distribution of income. This brings us to one further justification for normalizing CES production
function which is closely related to the concept of neutral technical progress and which was first
articulated by Kamien and Schwartz (1968).
Normalization implies that there may be considered a reference (or representative) value for the
capital income share (and thus for income distribution) at some given point. Technical progress that
does not change income distribution over time is called Harrod-neutral (or labor-augmenting) technical
progress. There are other types of neutral technical change, however, that would not have this effect.12
So the whole concept of whether technical progress is neutral with respect to the income distribution
relies on the idea that one has to check whether or not a given income distribution at one point in
time remains constant. This given income distribution, which is used to evaluate possible distribution
effect of technical progress, is exactly the income distribution in the baseline point of normalization at
a fixed point in time, t = t0 .
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 11

2.5.1 Constant growth rates of normalized factor efficiency levels

A CES production function with factor-augmenting technical progress can be written as
 σ −1  σ −1  σ −1
σ

Y = E tK K σ + E tN N σ (23)

where E tK and E tN represent the levels of efficiency of both input factors.13

Thus, whereas the ACMS specification seems to imply that technological change is always Hicks-
neutral, the above specification allows for different growth rates of factor efficiency. To circumvent
problems related to Diamond-McFadden’s Impossibility theorem (Diamond et al. 1978; Diamond and
McFadden 1965), we assume a certain functional form for the growth rates of both efficiency levels
and define
E ti = E 0i eγi (t−t0 ) (24)
where γi denotes growth in technical progress associated with factor i and t represents time. The
combination γ K = γ N > 0 denotes Hicks-neutral technical progress; γ K > 0,γ N = 0 yields Solow-
neutrality; γ K = 0,γ N > 0 represents Harrod-neutrality; and γ K > 0 = γ N > 0 indicates general factor-
augmenting technical progress.14
E 0i are the fixed points of the two efficiency levels, taken at the common baseline time, t = t0 .
Again, normalization of the CES function implies that members of the same CES family should all
share the same fixed point but differ in their elasticity of substitution values (Figures 1 and 2 showed
this graphically). In order to ensure that this property also holds in the presence of growing factor
efficiencies, it follows that
  1−σ
σ   1−σ σ
Y0 1 Y0 1
E 0N = ; E 0K = (25)
N 0 1 − π0 K 0 π0
Note that at t = t0 , eγi (t−t0 ) = 1. This ensures that at the common fixed point the factor shares are
not biased by the growth of factor efficiencies but are just equal to the distribution parameters π0 and
1 − π0 .
Inserting equation (24) and the normalized values (25) into (23) leads to a normalized CES function
that can be rewritten in the following form that again resembles the ACMS variant:
   σ σ−1   σ σ−1  σ σ−1
Y0 γ K (t−t0 ) Y0 γ N (t−t0 )
Y = π0 ·e · Kt + (1 − π0 ) ·e · Nt (26)
K0 N0
or equivalently,
 σ  σ −1 σ  σ −1  σ σ−1
Y = Y0 π0 K 0σ −1 K t · eγ K (t−t0 ) σ
+ (1 − π0 ) N0σ −1 Nt · eγ N (t−t0 ) σ
(27)
In this specification of the normalized CES function, with factor-augmenting technical progress, the
growth of efficiency levels for capital and labor is now measured by growth in the expressions
K 0 eγ K (t−t0 ) and N0 eγ N (t−t0 ) , respectively, and t0 is the baseline year. Again, we see from (27) that for
t = t0 we retrieve Y = Y0 .
Special cases of (27) are the specifications used by Rowthorn (1999), Acemoglu (2002a) or Bentolila
and Saint-Paul (2003), where N0 = K 0 = Y0 = 1 is implicitly assumed, or by Antràs (2004) who sets
N0 = K 0 = 1. Blanchard (1997), Caballero and Hammour (1998) and Berthold et al. (2002) work with
a version of (27) where in addition to N0 = K 0 = 1, γ K = 0 is also assumed so that technological
change is purely labor-augmenting.
It is also worth noting that for constant efficiency levels γ N = γ K = 0 our normalized function
(27) is formally identical with the CES function that Jones (2003) (p. 12) has proposed for the
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
12 KLUMP ET AL.

characterization of the ‘short term’. In his terminology, the normalization values k0 , y0 and π0 are
‘appropriate’ values of the fundamental production technology that determines long-run dynamics.
This long-run production function is then considered to be Cobb–Douglas with constant factor shares
equal to π0 and 1 − π0 and with a constant exogenous growth rate. Actual behavior of output and
factor inputs is thus modeled as permanent fluctuations around ‘appropriate’ long-term values. For a
similar approach in which Cobb–Douglas parameter values are used to normalize a CES production
function, see Guo and Lansing (2009).

2.5.2 Growth Rates in Normalized Technical Progress Functions: Time-Varying Frameworks

Following recent theoretical discussion about possible biases in technical progress (e.g., Acemoglu
2002a), it is not clear that growth rates of technical progress components should always be constant.
An innovation of Klump et al. (2007) was to allow deterministic but time-varying technological progress
terms where curvature or decay terms could be uncovered from the data in economically meaningful
ways. For this they used a Box and Cox (1964) transformation in a normalized context:
λi
γi t
gi (γi , λi , t, t0 ) = t0 − 1 ,i = K, N (28)
λi t0
The curvature parameter λi determines the shape of the technical progress function. For λi = 1,
technical progress functions, gi , are the (textbook) linear specification; if 0 < λi < 1 they are
exponential; if λi = 0 they are log-linear and λi < 0 if they are hyperbolic functions in time. Note, the
re-scaling of γi and t by the fixed point value t0 in (28) allows us to interpret γ N and γ K directly as the
rates of labor- and capital-augmenting technical change at the fixed-point period.
Asymptotically, function (28) would behave as follows:
⎧
lim gi = ∞ for λi ≥ 0
⎨ t→∞
gi (γi , λi , t, t0 ) = γ
⎩ lim gi = − i t0 for λi < 0
t→∞ λi
⎧
 λi −1 ⎪ ∂ gi
∂ gi t ⎨ = γi for λi = 1
= γi ⇒ ∂t
∂t t0 ⎩ lim ∂ gi = 0 for λi < 1
⎪
t→∞ ∂t

This framework allows the data to decide on the presence and dynamics of factor-augmenting technical
change rather than it being imposed a priori by the researcher. If, for example, the data supported an
asymptotic steady state, this would arise from the estimated dynamics of these curvature functions [i.e.,
labor-augmenting technical progress becomes dominant (linear), that of capital absent or decaying].
In addition, as McAdam and Willman (2013) pointed out, the framework also allows one to nest
various strands of economic convergence paths towards the steady state. For instance, the combination,
γ N > 0, λ N = 1; γ K = 0, λ K = 0 (29)
coupled with the assumption, σ >> 1 corresponds to that drawn upon by Caballero and Hammour
(1998) and Blanchard (1997), in explaining the decline in the labor income share in continental Europe.
Another combination speculatively termed ‘Acemoglu-Augmented’ Technical Progress by McAdam
and Willman (2013) can be nested as
γ N , γ K > 0; λ N = 1, λ K < 1 (30)
where σ < 1 is more natural.
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 13

Consider two cases within (30). A ‘weak’ variant, λ K < 0, implies that the contribution of capital
augmentation to TFP is bounded with its growth component returning rapidly to zero; a ‘strong’
case, where 0 < λ K < 1, capital imparts a highly persistent contribution with (asymptotic convergence
to) a zero growth rate. Both cases are asymptotically consistent with a BGP, where output growth
converges to that of labor-augmenting technical progress, γ N , plus the rate of growth of the labor
force. Accordingly, the interplay between |γ N − γ K | and λ K , λ N can be considered sufficient statistics
of BGP divergence. Normalization, moreover, makes this kind of classification quite natural since we
are looking at biases in technical progress relative to some representative point.

2.6 An Aside: The Substitution Elasticity and Factor Income Shares

Normalization allows us to have a valid idea of the reference points for factor income shares and the
bias of non-neutral technical change – in both cases, biases relative to a benchmark or reference point.
Regarding the former, we know (e.g., McAdam and Willman 2013) that factor shares vary over time.
Indeed such variation – as La Grandville (2009) notes – has been a major motivation for moving to
more flexible functional forms for production. We now explain the link between them.
Non-neutral technical change will matter in so far it influences developments in output, relative
prices, factor intensities, income shares and cost pressures. Movements in these variables affect the
inter-temporal decisions of consumers and firms. Some indications of the key role played by factor
substitution can be gauged from the following. Assuming competitive markets and profit maximization,
relative factor income shares are:
 K  σ σ−1
rt K t π0 t K t /K 0
t = = (31)
wt N t 1 − π0 t Nt /N0
N

where tK = eγ K (t−t0 ) etc.

It is straightforward to show that the effect of technical bias and capital deepening on factor income
shares is related to whether factors are gross complements or gross substitutes:
   
∂ ∂
sgn , sgn  = sgn {σ − 1} (32)
∂ (K /N ) ∂ tK / tN
Hence, an increase in factor J -augmenting (J = K , N ) technical change ‘favors’ factor J (i.e.,
∂ F /∂ J
implying ∂ FJI /∂ I > 1, J = I , and raising J ’s income share for given factor proportions) if factors are
gross substitutes (σ > 1). The effects reverse if factors are gross complements.
Thus, it is only in the gross-substitutes case that a factor J-augmenting change in technology is
J-biased. Naturally, the relations between the substitution elasticity, technical bias and factor shares
evaporate under Cobb–Douglas: factor income shares are constant and relative factor prices are purely
determined by capital deepening.
Equation (31) illustrates the impact of technology shocks on factor payments depends on the
substitution elasticity and the factor bias of the shock. This influences the dynamic response of interest
and wages (and hence hours) to technology shocks. Note, though, statement (32) defines factor demand
reactions to technology changes. They therefore abstract from labor-supply reactions (although we will
return to this in Section 5).

3. The Elasticity of Substitution as an Engine of Growth

Although one of the first references to a CES structure of aggregate production appears in the Solow
(1956) paper, it had been for a long time impossible to answer the question of what effect the size of the
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
14 KLUMP ET AL.

substitution elasticity has on the steady-state values in the standard neoclassical growth model. Common
sense would certainly suggest that easier factor substitution – via helping to overcome diminishing
returns – should lead to a higher level of development. But a formal proof of this conjecture seemed for
a long time out of reach. In fact, when Harbrecht (1975) tried to answer this question with the help of
a (implicitly normalized) David and van de Klundert (1965) CES variant, he found the contrary result!
His analysis was, of course, biased by the dependency of the distribution and efficiency parameters on
the elasticity of substitution, when the CES function is not correctly normalized.
Already some years earlier, as mentioned in Section 2.4, Kamien and Schwartz (1968) had presented
a proof of the central relationship between the substitution elasticity and output but only for the special
case in which the baseline values for K and N were equal. Their proof is based on the General Mean
property of the CES function, which had already been recognized by ACMS.
A General Mean of order p is defined as
 n 1
 p p
M( p) = f i xi (33)
ti=1

where xi , . . . , xn are positive numbers (of the same dimension) and where the weights f i , . . . , f n sum
to unity. Special cases of the General Mean are the arithmetic, the geometric and the harmonic means
where the order p would be 1, 0 and −1, respectively. If p tends to −∞, the mean becomes the
minimum of the numbers (xi , . . . , xn ).
One of the most important theorems about a General Mean is that it is an increasing function of its
order (Hardy et al. 1934, p. 26f; Beckenbach and Bellman 1961, p. 16–18; see also the proof in La
Grandville 2009, p. 111–113). More exactly, it says that the mean of order p of the positive values xi
with weights f i is a strictly increasing function in p unless all the xi are equal. With the two factors
K and N (and implicit normalization K 0 = N0 ) this leads to the following statement:
Enlargement of the elasticity of substitution results in an increase in output from every combination
of factors except that for which the capital labor ratio is equal to one. (Kamien and Schwartz 1968,
p. 12)
Of course, this result can be generalized provided that all numbers have the same dimension which
is precisely achieved by normalizing numbers of different dimensions.
La Grandville (1989b) developed a graphical representation of normalized CES structures. He
demonstrated that the general relationship between the elasticity of substitution and the level of
development is usually positive. Moreover, when there are two factors of production, numerical results
suggest that the function has a single inflection point (La Grandville and Solow, 2006): in other words,
between its limiting values,lim p∈(−∞,∞) M( p), the function M( p) is first convex then concave. For
typical production-function weights (i.e., f 1 = 0.4; f 2 = 1 − f 1 ) that inflection point occurs around
p ≈ 0 (i.e., the Cobb–Douglas neighborhood).
This means that within some relevant interval around that even small perturbations of the substitution
elasticity (however such a change may be implemented) might have extremely large implications for
an economy. In short, raising your elasticity of substitution can raise your growth rate and its effect
may be potentially even larger than that traditionally studied in the case of (equivalent percentage)
improvements in the savings rate and/or technical progress (such reasoning is reflected in the third
quote that started our paper).
The formal proof for the conjecture was then presented by Klump and de La Grandville (2000),
based on a very general normalized CES production function. An alternative proof is presented in
Klump and Irmen (2009) who also deal with normalized CES functions in a Diamond-type version
of the neoclassical growth model. It distinguishes efficiency and distribution effects of changes in the

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31

C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 15

elasticity of substitution which can work in different directions if not all individuals have the same
savings pattern so that redistribution matters. The interaction of both effects creates an acceleration
effect for capital accumulation which can have a positive or a negative effect on the steady state. It
can be shown, however, that even in this setting a higher elasticity of substitution leads to a higher
steady-state level as long as the efficiency effect dominates the distribution effect, which is the most
likely case.
Klump and Preissler (2000) extend the analysis of the standard neoclassical growth model with a
normalized CES production function by calculating the effect of the size of the elasticity of substitution
on the speed of convergence towards the steady state. Earlier studies of this problem, for example,
Ramanathan (1975), which were not considering normalization had not derived convincing results.
With an explicitly normalized CES production function, it is possible to show that an increase in the
elasticity of substitution reduces the speed of convergence if the steady-state value of capital intensity
is higher than its baseline value (which seems the most likely case).
Klump (2001) presents the analysis of a Ramsey type (intertemporal optimizing) growth model with
a normalized CES production function. He is able to prove that as long as the steady-state value of
the capital intensity is higher than its baseline value the comparative-static effect of a change in the
elasticity of substitution on the steady state is strictly positive. The result were only recently reproduced
by Xue and Yip (2013) using a different approach. For the effect of the elasticity of substitution on
the speed of adjustment the same results as in the Solow model can be derived in the Ramsey model
(see Klump and Saam 2008). This result holds irrespective of the value of the elasticity of substitution,
whereas Ventura (1997) making use of an implicitly normalized CES production function could only
generate meaningful results for σ < 1.
Summing up, an increase in σ increases the steady-state level of production and capital intensity
while lengthening the convergence time to the new steady state. From the standpoint of short run
growth, this leaves open the question of whether growth in the short run will increase or decrease
relative to an initially lower σ comparative value.
Temple (2008) has criticized the use of normalized CES functions for calculating convergence effects
of a higher factor substitution because of an unclear economic meaning of the chosen baseline value
for the capital intensity. However, as has been clarified by Klump and Saam (2008) the essence of
normalization does not consist in the arbitrary choice of baseline values but in forcing the researcher
to give an explicit statement about the relationship between baseline and steady-state (ss) values. As
growth models are generally motivated by the idea that labor is relatively scarce in the steady state
it seems reasonable to normalize such that k ss > k0 . In addition, in a growing economy, it is always
feasible to assume capital intensity would be below the steady state, whereas values above the steady
state raise the question of how the economy has found itself in such a starting position with surplus
capital stock.
The setting may be different in the business-cycle literature, where fluctuations around the
(typically zero growth) steady state are studied. In this case it makes sense to use steady-state
values as normalization parameters (Guo and Lansing 2009; Cantore et al. 2010, hereafter CLMW
(2010)).
Finally, Irmen (2011) is able to show in an endogenous growth framework with a normalized
CES production function that the steady-state growth rate of output per worker increases with the
elasticity of substitution. The efficiency effect induced by a higher degree of factor substitution makes
innovation investments more profitable that raise permanently the productivity of labor. All analysis
using normalized CES production functions confirm that the elasticity of substitution is among the most
powerful determinants of growth. La Grandville (2009), 2012) suggests that changes in the elasticity of
substitution have a much higher effect on social welfare than changes in the rate of technical progress
– see chapter 13, pp. 316-319, where the author compares the ratio of sensitivities of a given value

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31

C 2012 Blackwell Publishing Ltd
16 KLUMP ET AL.

function, V (e.g., savings or consumption flows over time), to a change in the elasticity of substitution
and to a change in the rate of (Harrod-neutral) technical progress: eV ,σ /eV ,γ N .

4. Estimated Normalized Production Function

Previous sections of our paper introduced the concept of normalization and largely its importance in
theoretical analysis. Here we discuss how the idea of normalization should be applied in empirical
analysis and, more importantly, whether it makes the estimation of the parameters of the CES
production function easier and more robust? We show that its merits are strong especially if system
approach (containing cross-equation restrictions) is used. In this context the scepticism on the proper
identification of the elasticity of substitution and technical progress from each other aroused by the
famous Diamond–McFadden impossibility theory largely loses its practical importance. In fact, we
argue that a general factor-augmenting specification results in markedly less biased estimates of the
elasticity of substitution parameter than imposing a priori neutrality constraints. In the context of single
equation approach, though, normalization is of lesser use.
An added problem,15 however, is that often the predictions of different elasticity and technical
change combinations can have similar implications for variables of interest, such as factor income
shares and factor ratios. Notwithstanding, whether factor income movements are driven by high or low
substitution elasticities and with different combinations of technical change is profoundly important
in terms of their different implications for, for example, growth accounting, inequality, calibration in
business-cycle models, public policy issues etc.
By way of illustration, Tables 1 and 2 present an overview of empirical results obtained for the
elasticity of substitution. We concentrate on the results from time-series or panel studies on aggregate
data. In the case of the United States, which has been widely studied, it is possible to find values of
the elasticity of substitution above unity (with Harrod-neutral technical progress), at unity (with Hicks-
neutral progress) and below unity (with Hicks-neutral progress and with technical progress augmenting
both factors). The situation for other countries is no better; for Germany, values of above, below and at
unity have been estimated. Using information about the degree of factor substitution from other sources
does not resolve this puzzle, either. It has been recognized, for example, by Lucas (1969) that older
time-series studies for the United States have generally provided lower estimates than cross-section
studies that were supportive of the Cobb–Douglas function. More recent cross-section analysis based
on micro data that were used to estimate the relationship between business capital formation and user
costs (e.g., Chirinko et al. 1999) estimates very low elasticities of substitution ranging from 0.25 to
0.40. A drawback of these kinds of studies, however, is their inability to quantify any growth rate(s)
of technical progress.
That there should be diversity in production function estimates – even for countries whose data
properties are relatively stable and well-understood – is not surprising. It doubtlessly reflects the
familiar trapdoor of empirical pitfalls: data quality; a priori modeling choices (such as whether to
test for certain types of factor neutrality or impose them); the performance of various estimators
(e.g., single equation, systems) and algorithms; as well as more prosaic data problems (e.g., outliers,
uncertain auto-correlation, structural breaks, quality improvements, measurement errors etc.).
At a simple level, normalization removes the problem that arises from the fact that labor and capital
are measured in different units – although as we have seen its importance goes well beyond that.
Under Cobb–Douglas, normalization plays no role since, due to its multiplicative form, differences in
units are absorbed by the scaling constant. The CES function, by contrast, is highly non-linear, and
so, unless correctly normalized, excluding technical progress, out of its three key parameters – the
efficiency parameter, the distribution parameter, the substitution elasticity – only the latter is ‘deep’.
The other two parameters turn out to be affected by the size of the substitution elasticity and factor
income shares.
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd

Table 1. Empirical Studies of Aggregate Elasticity of Substitution and Technological Change in the US.

Estimated Annual Rate Of Efficiency Change

Assumption Estimated
on Technological Elasticity of Neutral: Laboraugmenting: Capital-

C 2012 Blackwell Publishing Ltd

Study Samplea Change Substitution:σ γN = γK γN Augmenting: γ K

Arrow et al. (1961) 1909-1949 Hicks-Neutral 0.57 1.8 – –

Kendrick and Sato (1963) 1919-1960 Hicks-Neutral 0.58 2.1 – –
Brown and De Cani (1963) 1890-1918 Factor Augmenting 0.35 Labor saving (γ N – γ K = 0.48)
1919-1937 0.08 Labor saving (γ N – γ K = 0.62)
1938-1958 0.11 Labor saving (γ N – γ K = 0.36)

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31

1890-1958 0.44 ?
David and van de Klundert (1965) 1899-1960 Factor Augmenting 0.32 – 2.2 1.5
Bodkin and Klein (1967) 1909-1949 Hicks-neutral 0.50-0.70 1.4-1.5
Wilkinson (1968) 1899-1953 Factor Augmenting 0.50 Labor saving (γ N – γ K = 0.51)
Sato (1970) 1909-1960 Factor Augmenting 0.50 – 0.70 – 2.0 1.0
Panik (1976) 1929-1966 Factor Augmenting 0.76 Labor saving (γ N – γ K = 0.27)
Berndt (1976) 1929-1968 Hicks-neutral 0.96-1.25 ? – –
Kalt (1978) 1929-1967 Factor Augmenting 0.76 – 2.2 0.01
Antràs (2003) 1948-1998 Hicks-neutral 0.94-1.02 1.14 – –
Factor-augmenting 0.80 Labor saving (γ N – γ K = 3.15)
Klump et al. (2007b)b 1953-2002 Factor-augmenting 0.7 – – –
León-Ledesma et al. (2010) 1960- 2004 Factor-augmenting 0.60-0.70 – 1.60 0.70

Notes:
a
All studies are estimated on annual frequency data except León-Ledesma et al. (2010b) which is quarterly.
Although, to aid comparability, we annualized their estimates for technical change in the table.
b
We do not report technical change estimates for Klump et al. (2007b) since they estimate with structural
breaks in their technical progress terms. For reasons of space, the reader is referred to León-Ledesma et al.
THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS

(2010a) for the original references.

17

18
Table 2. Empirical Studies of Aggregate Elasticity of Substitution in Selected Other Countries.

Assumption For Estimated

Sample Technological Elasticity Of
Study Countries (Frequency) Change Substitution: σ

Lewis and Kirby (1988) Australia 1967-1987 (Weekly) Hicks-Neutral 0.78

Easterly and Fischer (1995) Soviet Union 1950-1987 (Annual) Hicks-Neutral 0.37

C 2012 Blackwell Publishing Ltd

Andersen et al. (1999) Panel of 17 OECD countries 1966-1996 (Annual) Hicks-Neutral 1.12
Bolt and van Els (2000) Austria 1971-1996 (Quarterly) Hicks-Neutral 0.24
Belgium 0.78
Germany 0.53
Denmark 0.61
Spain 1

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31

Finland 0.34
France 0.73
Italy 0.52
Netherlands 0.27
Sweden 0.68
UK 0.6
KLUMP ET AL.

US 0.82
Japan 0.3
Duffy and apageorgiou (2000) Panel of 82 developed and 1960-1987 Annual) Hicks-Neutral 1.4
developing countries
Ripatti and Vilmunen (2001) Finland 1975-1999 (Quarterly) Factor-Augmenting 0.6
Willman (2002) Euro area 1970-1997 (Quarterly) Solow-Neutral 0.95-1.05
McAdam and Willman (2004) Germany 1983-1999 (Quarterly) Hicks-Neutral 0.7, 1, 1.2
Berthold et al. (2002) US, Germany, France 1970-1995 (Semi-Annual) Harrod-Neutral 1.15, 1.45, 2.01
Bertolila and Saint-Paul (2003) 13 industries in 12 OECD 1972-1993 Harrod-Neutral 1.06
countries
McAdam and Willman (2004a) Germany 1983-1999 (Quarterly) Hicks-Neutral 0.7, 1, 1.2
Klump et al. (2007b) Euro Area 1970-2003 (quarterly) Factor-Augmenting 0.7
Luoma and Luoto (2010) Finland 1902-2004 (annual) Factor-Augmenting 0.5

Notes: For reasons of space, the reader is referred to Klump et al. (2007b) for most of these original references.
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 19

If one compares the explicitly normalized with the implicitly normalized function, as before, that is,
  σ −1  σ −1  σ −1
σ

Yt = C π tK K t σ + (1 − π) tN Nt σ

   σ −1   σ −1  σ σ−1
K σ N σ
t Kt t Nt
Yt = Y0 π0 K
+ (1 − π0 ) N
0 K0 0 N0

we may be unsure as to where the estimation benefits of normalization derive. After all, both equations
contain the same number of parameters. In fact, the latter equation seemingly adds complexity by
incorporating normalized reference points into the estimation (the empirical choice of the normalization
point is a particular aspect discussed in Section 4.2).
The answer as to why normalization should improve matters empirically reflects the following. The
distribution and efficiency parameters (respectively, Y0 ; π0 and C; π) can now either be imposed prior
to estimation or at least have a deep interpretation in terms of the data (i.e., the representative capital
income share). Effectively normalization allows us to reduce the number of freely estimated parameters
by two.
This follows straightforwardly from our earlier analysis. In the implicitly normalized formulation the
parameters C and π above have no clear theoretic or empirical meaning. Instead, they are composite
parameters conditional on, besides the selected fixed points, the elasticity of substitution (re-stating
equations (15) and (16)):
 1/σ
 σ
1/σ σ −1
r 0 K 0 + w0 N 0
C (σ, ·) = Y0
r 0 K 0 + w0 N 0

1/σ
r0 K 0
π (σ, ·) = 1/σ 1/σ
r 0 K 0 + w0 N 0
The additional merit in using the normalized instead of the implicitly normalized form is that all
parameters have a clear empirical correspondence. In particular, the distribution parameter is identified
as the capital income share of total factor income at the fixed point. Hence, a suitable choice for the
fixed point may alleviate the estimation of the deep parameters and, to repeat, makes the estimated
production function suitable, for example, for subsequent comparative-static analysis.16
Table 3 presents some consistent sets of (deterministic) initial values for generating data and
the implied ranges of the true values of C and π for σ ∈ [0.2, 1.3]. In the first row we assume
K 0 = N0 = 1. This allows us to solve Y0 from the first row – with initial values of 0K = 0N = 1.
In fact this represents a special case because indexing by the point of normalization equaling one is
neutral implying that the true value of C = 1 and π = π0 = r0 = 0.3 ∀σ (this, in turn, implies solving
the normalized real wage rate as (1−π 0 )Y0
N0
). In this special case it does not matter if the same initial
values of parameters are used, whether the system is estimated in normalized or implicitly normalized
form.
In all other cases, however, this is not so. To illustrate, in these other cases we have adjusted the
initial conditions for output to make them consistent with an initial (and arguably more reasonable)
value for r (the real user cost of capital) equal to 5%. The sample average normalization insulates
the normalized system from the effects of changes in initial values in generating the data but the true
values of composite parameters C and π vary widely: C ∈ [0.23, 0.79], π ∈ [0.23, 0.99]. Thus, we
confirm that the actual income distribution of the data is completely unrelated to the true value of π.
This illustrates the difficulty that a practitioner faces when trying to estimate implicitly normalized
forms since the actual data scarcely give any guidelines for appropriate choices for the initial parameter
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
20 KLUMP ET AL.

Table 3. Consistent Normalization Values.

C π

when when when when

(1 − π0 ) Y0
N0 π0 r0 K0 Y0∗ = Y0 = r0
π0
K0 w0 = N0
σ = 1.3 σ = 0.2 σ = 1.3 σ = 0.2

1 0.3 0.3 1 1 0.7 1 0.4

1 0.3 0.05 5 0.833 0.583 0.546 0.225 0.228 0.996
1 0.3 0.05 8 1.333 0.933 0.788 0.225 0.210 0.999

Notes: C and π in the final two columns are calculated according to equations (15) and (16) for σ ∈ [0.2,1.3].
Outside of the “special case” note the following partial derivatives showing how ceteris paribus changes in initial
values
change these last two parameters, C and π :
Cσ , CY0 , C K 0 , Cw0 > 0, C N0 < 0
πσ , π N0 < 0, π K 0 , πw0 > 0

values of C and π. As León-Ledesma et al. (2010a) have documented, that results in serious estimation
problems. They estimated normalized and implicitly normalized forms where in the latter case the
initial parameter values for C and π are selected randomly from their given range such as in
Table 3. When C  and π substantially depart from their true, theoretical values, there are significant
and quantitatively important biases in the estimated substitution elasticity and technical change.

4.1 Estimation Forms

The recognition of normalization says nothing specifically about the way production and production-
technology should be estimated and how normalization impacts those estimation choices. Typical
estimation forms found in the literature include: the non-linear CES production function; the linear
first-order conditions of profit maximization; linear approximation of the CES function and ‘system’
estimation incorporating the production function and the first-order conditions.17
To proceed let us express the CES function, equation (27), in log form:
     σ σ−1   σ σ−1 
Y σ γ K (t−t0 ) K t γ N (t−t0 ) Nt
log = log π0 e + (1 − π0 ) e (34)
Y0 σ −1 K0 N0
From this we can derive the marginal profit-maximization conditions,18
     
Y0 1 K0 1 Y σ −1
log (r ) = log π0 + log + log + (γ K (t − t0 ))
K0 σ Y0 σ K σ (35)
  
αr

     
Y0 1 N0 1 Y σ −1
log (w) = log (1 − π0 ) + log + log + (γ N (t − t0 ))
N0 σ Y0 σ N σ (36)
  
αw

Where, as before, γ N and γ K are the respective growth rates of labor and capital augmenting technical
progress. Equations (35) and (36) represent the first-order conditions with respect to capital and labor,
respectively.
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 21

Estimation of production and technology parameters based on the first-order conditions and other
single-equation approaches is hampered by the fact that they only admit estimates of technical progress
terms contained by their presumed FOC choice (in that sense any bias in technical progress is,
by definition, not separately identifiable). This apparent drawback is presumably (in the minds of
the researcher at least) compensated by their tractable form and linearity. Accordingly, these forms
are common (more common, for instance, than direct non-linear CES estimation): For example,
equation (35) has been widely used in the investment literature (e.g., Caballero 1994) and (36) was the
form used by ACMS amongst many others.
A notable feature of the above three equations is that if estimated in single-equation mode, the
normalization points (denoted by the curly lower brackets) are absorbed by the respective constants, αr
and αw . Thus, from an estimation stand point, it is only when the non-linear CES function is estimated
directly or where the system approach is used, does formal normalization play an explicit empirical
role.
Another possible vehicle of estimation is the Kmenta (1967) approximation (which became an
important, if apparently unacknowledged, pre-cursor to the translog form). This is a Taylor-series
expansion of the log CES production function around σ = 1.19
yt = π0 kt + λkt2
2λ 2λ
+ π0 1 + kt γ K
t + (1 − π0 ) 1 − kt γ N
t + λ [γ K − γ N ]2 · 
t2 (37)
π0 1 − π0
  
tfp

where  t = t − t0 , yt = log[(Yt /Y0 ) /(Nt /N0 )], kt = log[(K t /K 0 )/(Nt /N0 )], t f p = Log(T F P) and λ =
(σ −1)π0 (1−π0 )
2σ
. Equation (37) shows that the output–labor ratio can be decomposed into capital deepening
and technical progress, weighted by factor shares and the substitution elasticity (where sgn(λ) =
sgn(σ − 1) and lim λ ∈ [−∞, 12 π0 (1 − π0 )]). In addition, (37) shows that, when σ = 1and γ K =
σ ∈[0,∞)
γ N > 0, additional (quadratic20 ) curvature is introduced into the estimated production function.
With the predetermined normalization point, the advantage of (37) over the Kmenta approximation
of the implicitly normalized CES is – as usual – that, since all variables appear in indexed form,
the estimates are invariant to a change in units of measurement. Another advantage is that in the
neighborhood of the normalization point (i.e., K t = K 0 , Nt = N0 , π = π0 ) and without σ deviating
‘too much’ from unity, as the approximation also assumes, the terms including the normalized capital
intensity and multiplying linear trend have only second-order importance and, without any significant
loss of precision, can be dropped, yielding,
y = π0 kt + λkt2
+ [π0 γ K + (1 − π0 ) γ N ]
t + λ [γ K − γ N ]2 
t2
      (38)
θ
  
tfp

Estimation of equation (38) yields four parameters, π0 ,  λ, 

θ , , for four primitives, π0 , σ , γ K ,
2λ
γ N . Using π0 allows us to identify σ from composite parameter λ, that is, σ = (1 − π0 (1−π 0)
)−1 .
However, without a priori information on which one of two technical progress components dominates
and, in addition, that the signs of estimates λ and are (or are constrained to be) the same, one
cannot identify γ K and γ N . This leads to the following
 weak identification result: for γ N >γ K we
obtain γ N = θ + π0  and γ = 
θ − (1 − π ) 
and for γ N < γ K we obtain, γ N = θ − π0  and
 λ  
N 0 λ λ

γN =  
θ + (1 − π0 ) λ .
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
22 KLUMP ET AL.

Given this, even under the helpful environment of normalization, we can say that although the
Kmenta approximation can be used to estimate σ , it cannot effectively identify the direction of the
biased technical change.21 However, the approximation is a useful vehicle to, ex post, calculate TFP.

4.2 The Point of Normalization – Literally!

To be empirically operational, the point of normalization must be defined (i.e., what these Y0 , K 0
are in practice). If the estimation data were deterministic, this would be unproblematic: every sample
point would be equally suitable for the point of normalization. For instance, in theoretic settings, the
normalization point is often calibrated around the non-stochastic steady state (e.g., Klump and Saam
2008; CLMW 2010).
However, since actual data are inevitably stochastic (and the intensity with which factors are utilized
is also unobserved), this convenience does not carry over because the production function does not
hold exactly in any sample point. Therefore, to diminish the size of cyclical and stochastic components
in the point of normalization, an appealing procedure is to calibrate the normalization point in terms
of sample averages for the underlying variables – typically geometric averages for growing variables
(such as factors of production) and arithmetic ones for approximately stationary variables (e.g., factor
income shares, the real interest rate and user cost).
The non-linearity of the CES function, however, in turn, implies that the sample average of production
need not exactly coincide with the level of production implied by the production function with sample
averages of the right hand variables even with a deterministic DGP. To circumvent this problem, Klump
et al. (2007) introduced an additional parameter ξ whose expected value is around unity (which we
might call the normalization constant22 ). Hence, we can define Y0 = ξ Y , K 0 = K , N0 = N ,π0 = π
and t0 = t where the bar refers to the respective sample average (geometric or, as in the last two,
arithmetic).
An advantage of the explicitly normalized system over the implicitly normalized system, in turn,
is that the distribution parameter π has a clear data-based interpretation. Therefore, it can either be
fixed prior to estimation or, at least, the sample average can be used as a very precise initial value of
the distribution parameter. Likewise, a natural choice for the initial value of normalization constant,
ξ , is one. Estimated values of these two parameters should not deviate much from their initial values
without casting serious doubts on the reasonableness of estimation results. In the implicitly normalized
case, by contrast, no clear guidelines exist in choosing the initial values of distribution parameter π and
efficiency parameter C. This is especially the case if the function that the estimation method optimizes
contains several local optima, as demonstrated to be a concrete problem in analyzing both real data
(Klump et al. 2007) and Monte Carlo generated data (León-Ledesma et al. 2010b). Hence, in the
context of non-linear estimation this may imply a well-defined advantage of the explicitly normalized
over the implicitly normalized system.

5. Normalization in Growth and Business-Cycle Models

Production functions are a key part of business-cycle and growth macro-models. Over the last few
decades, the two main competing models in the macro profession have been the RBC (real business-
cycle) model and the NK (New Keynesian) model. Both imply relatively tight, theory-led dynamics
and are typically furnished with a rich number of stochastic shocks which displace the agent from his
optimal plans. The standard RBC model is a variant of the representative agent neoclassical model,
where business cycles are due to non-monetary sources (primarily, changes in technology). The NK
one supplements that with various real and nominal rigidities to better capture the data.
However, what both models tend to share in practice is a focus on Cobb–Douglas aggre-
gate production. This is especially puzzling given that such models tend to be motivated as
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 23

business-cycle frameworks. Yet over business-cycle frequencies one might precisely expect relatively
little (and presumably below unitary) factor substitutability as well as the presence of non-neutral
technical change to capture factor income share developments. By introducing and assessing non-
unitary production forms, the potential for a better understanding of technology and policy transmission
and for a richer decomposition of historical time-series is likely to be considerable.
The introduction of normalized technologies in simple business-cycle models is relatively
straightforward and can be illustrated using the canonical RBC model. The model is relatively well
known and can therefore be introduced compactly. The standard model with CES production technology
in the supply side would be given by:
t = β Et {t+1 [1 + rt+1 − δ]} (39)

ς
Nt
wt = υ (40)
t
   σ σ−1   σ σ−1  σ σ−1
K t γt K N t γt N
Y t = Y 0 π0 e + (1 − π0 ) e (41)
K0 N0

  σ σ−1   σ1
Y0 γ t N Yt
wt = (1 − π0 ) e (42)
N0 Nt

  σ σ−1   σ1
Y0 ztK Yt
r t = π0 e (43)
K0 Kt

Ct + K t − (1 − δ)K t−1 ≤ Yt (44)

j j
γt = ρ j γt−1 + ηtj (45)
where t , wt and rt are, respectively, the marginal utility of consumption (Ct ),wages and the interest
rate (all expressed in real terms). Parameters β, δ and υ represent, respectively, the discount factor, the
j
capital depreciation rate and a scaling constant. Processes γt are technology shocks – as equation (45)
shows usually modeled as a stationary AR(1) process – for j = K , N (i.e., capital-augmenting and
labor-augmenting shocks, respectively). Equations (39) and (40) represent the household’s optimal
consumption and labor supply choices given, for example, the separable utility function,
1+ς
Ct1−σc N
U (C, N ) = −υ t (46)
1 − σc 1+ς
where σc is the coefficient of relative risk aversion and ςis the inverse of the Frisch elasticity. This
particular utility function implies t = Ct−σc . If the researcher wanted to simulate this model conditional
on different values of the substitution elasticity, s/he would do the following:
(i) Imposed key normalization parameters: r0 = β1 − 1 + δ, and, for some given N0 , K 0 , π0 , solve
out Y0 = πr00 K 0 and w0 = (1 − π0 ) NY00 following Table 3 (for temporary shocks, these normalization
points will be chosen to be the same as the presumed steady state);
(ii) Reset the leisure scaling parameter v to equate the real wage expressions in (40) and (42), implying
(1−α )r σc
v = (r0 −δα0 0 )0σc .
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
24 KLUMP ET AL.

In this simple way, conventional dynamic exercises can be performed on the model (e.g., examining
the effect of technology shocks) which are robust to changes in the substitution elasticity.
In this vein, CLMW (2010) looked at the relationship between technology shocks and hours worked –
a key controversy between NK and RBC explanation of the business cycle – by expressing both models
in consistent normalized form. They showed that, depending on the value of the substitution elasticity
and the source of the shock (capital- or labor-augmenting), both models could generate positive or
negative hours responses (thus, largely overturning conventional wisdom on the mechanisms in the
models).

6. Conclusions and Future Directions

The elasticity of substitution between capital and labor represents a key parameter in many fields
of economics: for example, business-cycle and growth outcomes, income distribution, stabilization
policy, labor market dynamics etc. The empirical evidence – for the United States and other developed
economies – is clear in rejecting the unitary-elasticity Cobb–Douglas specification in favor of (generally
below unity) CES aggregate production functions.23 When investigating the ramifications of a non-
unitary substitution elasticity, one necessarily faces the issue of normalization.
The importance of explicitly normalizing CES functions was discovered in the seminal paper by La
Grandville (1989b), then further explored by Klump and de La Grandville (2000), Klump and Preissler
(2000), La Grandville and Solow (2009), and first empirically implemented and investigated by Klump
et al. (2007). Normalization starts from the observation that a family of CES functions whose members
are distinguished only by different elasticities of substitution needs a common benchmark point. Since
the elasticity of substitution is defined as a point elasticity, one needs to fix benchmark values for
the level of production, the inputs of capital and labor and for the marginal rate of substitution, or
equivalently for per-capita production, capital intensity and factor income shares.
Overall, we can say the following points.

(a) Normalization is necessary for identifying in an economically meaningful way the constants of
integration which appear in the solution to the differential equation from which the CES function
is derived (and thus makes it suitable for comparative-static analysis).
(b) Normalization helps to distinguish among the various functional forms, which have been developed
in the CES literature and thus to choose which CES production functions are legitimate.
(c) Normalization is necessary for securing the basic property of CES production in the context of
growth theory, namely the strictly positive relationship between the substitution elasticity and the
output level given the CES function’s representation as a ‘General Mean’.
(d) In situations where the researchers wish to gauge the sensitivity of results (steady-state or dynamic)
to variations in the substitution elasticity, normalization is imperative.
(e) Normalization alleviates the estimation of the deep parameters of the aggregate production
function, in particular the elasticity of substitution and the growth rates of factor-augmenting
technical progress.
(f) Normalization is convenient when biases in the direction of technical progress are to be empirically
determined, since it fixes a benchmark value for factor income shares. This is important when it
comes to an empirical evaluation of changes in income distribution arising from technical progress.
If technical progress is biased in the sense that factor income shares change over time the nature
of this bias can be best classified with regard to a given baseline value.

That said, in our view there are at least five promising and related areas for future research on
normalization:
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 25

(1) Following Jones (2005) one can regard a macroeconomic production function as a reduced form
which should be derived from micro-foundations. This conjecture has been taken up by Growiec
(2008) who shows that a CES production function can be linked to Weibull distributions of unit
factor productivities, whereas a Cobb–Douglas function responds to Pareto distributions. It is still
an open question, however, how the normalization values of the CES function can be linked in a
meaningful way to the parameters of the underlying unit factor productivity distributions.24
(2) A better understanding of the micro-foundations of the normalized CES function would also help
to better understand reasons for possible international differences and intertemporal changes in
the elasticity of substitution. Both seem to be linked to some deeper economic, social and cultural
parameters as well as to the level of development measured by capital per units of labor or income
per capita (Klump and de La Grandville 2000; Duffy and Papageorgiou 2000; Masanjala and
Papageorgiou 2004; Mallick 2010, Wong and Yip 2010).
(3) La Grandville (2009); ,2012) has suggested that an increase in the elasticity of substitution has a
much stronger effect on aggregate wealth than an increase in the rate of technical progress. This
suggestion can be viewed in two ways. First, in terms of comparative statics. For example, if
two countries are otherwise the same but have different (though constant) substitution elasticities,
one can trace the effects of that difference on their growth prospects. Second, even though the
substitution elasticity empirically is (or appears to be) quite stable, such structural changes (e.g.,
larger internal or external markets) are possible, that make factor substitution easier and may
launch some kind of ‘sigma-augmenting’ technical progress at work whose exact mechanisms
are not yet understood. Kamien and Schwartz (1968) had already pointed out that changes in
the elasticity of substitution have similar effects on relative factor prices and on the distribution
of income as the augmentation in factor efficiency. This is obvious from expression (31) given
above. As in other areas of induced technical change, these changes in relative factor prices and
income distribution might trigger biases in the direction of technical change which are worth being
analyzed in more detail.25
(4) In business-cycle and growth models, the Cobb–Douglas aggregate production function is the
default choice. However, the convenience/centrality of Cobb–Douglas production functions in
macro is likely to be obscuring important issues. We would therefore expect that normalization –
which leads to a clarified and deeper understanding of CES properties – will end up being more
widely used in such models. In so doing, its use should help shed light on the propagation and
decomposition of business-cycle shocks and policy transmissions.
(5) Production functions are often single-level variants, given that, in macroeconomics certainly, only
two factors of production are considered. However, potentially that obscures the interactions
between and within different factor categories. For example, there are high- and low-skill types of
labor and different strata of capital such as equipment, software, and buildings and infrastructure.
In this respect, an important departure from the aggregative framework was made in the seminal
contributions of Kazuo Sato and Zvi Griliches. Sato (1967) generalized the CES production
function by nesting the CES at two levels and augmenting the list of possible inputs.

A popular focus for work on the two-level CES function is on explanations of the increase in the
skill premium observed in western economies during the last three decades. Does the premium reflects
capital–skill complementarity (as in Griliches 1969 and Krussel et al. 2000)? Or can the premium can
be attributed to technical change that was biased in favor of skilled workers (e.g., Katz and Murphy
1992, Acemoglu 2002b; Autor et al. 2008)?26 Both approaches rely on particular nestings and estimated
values for the elasticities of substitution between different categories of production factors and would
be highly amenable to the simplification that normalization offers [see León-Ledesma et al. (2012) for
some early work in that direction]. Knowledge of this might also deepen our understanding of skill
differences between and within developing countries.
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
26 KLUMP ET AL.

To conclude our survey, we re-stress that production functions are ubiquitous in theoretical and
empirical models, and ubiquitously Cobb–Douglas! This appears to us not only an unjustifiable
simplification but also an impediment to understanding various economic phenomena. The paper by
CLMW (2010) is hopefully a useful contribution in fashioning otherwise standard Real Business Cycle
and New Keynesian models around a normalized CES supply side and showing the impact in terms
of overturning the prediction that these separate models made for the technology–hours correlation.
But this is merely one example. We would hope therefore that normalized CES functions would be an
integral part of the make-up of macroeconomic models, in the same way as nominal and real rigidities
have become.

Acknowledgments
We thank two anonymous referees, the editor Colin Roberts and an associate editor, as well as Cristiano
Cantore, Olivier de La Grandville, Jakub Growiec, Miguel León-Ledesma, Marianne Saam, and Ryuzo Sato
for comments and support. The views expressed are not necessarily those of the ECB.

Notes
1. The critical threshold level for the substitution elasticity (to generate such perpetual growth) can be
shown to be increasing in the growth of labor force and decreasing in the saving rate (see La Grandville
1989b).
2. This may be one reason why estimated investment equations struggle to identify interest-rate channels.
3. Jones’ work essentially builds on Houthakker (1955)’s idea that production combinations reflect the
(Pareto) distribution of innovation activities, Jones proposes a ‘nested’ production function. Given such
parametric innovation activities, this will exhibit a (far) less than unitary substitution elasticity over
business-cycle frequencies but asymptote to Cobb–Douglas.
4. Moreover, the point that a BGP cannot coexist with capital augmentation is becoming increas-
ingly questioned in the literature (see Growiec 2008; La Grandville 2012; León-Ledesma and
Satchi 2010).
5. It should be noted that the advantages of re-scaling input data to ease the computational burden of highly
nonlinear regressions has been the subject of some study (e.g., ten Cate 1992). And some of this work
was in fact framed in terms of production-function analysis (De Jong 1967; De Jong and Kumar (1972)).
See also Cantore and Levine (2011).
6. It should be borne in mind, however, that Berndt’s result concerned only the US manufacturing sector.
7. It is still not widely known that the famous ACMS paper was in fact the merging of two separate
submissions to the Review of Economics and Statistics following a paper from Arrow and Solow, and
another from Chenery and Minhas.
8. In the inaugural ANU Trevor Swan Distinguished Lecture, Peter L. Swan (Swan 2006) writes, ‘While
Trevor was at MIT he pointed out that a production function Solow was utilizing had the constant
elasticity of substitution, CES, property. In this way, the CES function was officially born. Solow and
his coauthors publicly thanked Trevor for this insight (see Arrow et al, 1961)’.
9. Alternatively, the substitution elasticity is sometimes expressed in terms of the parameter of factor
substitution, ρ ∈ [−1, ∞), where ρ = 1−σ σ
.
10. ACMS started from the empirical observation that the relationship between per-capita income and the
wage rate might best be described with the help of the power function. Note, σ = 1 implies a linear
relationship between y and w which would, in turn, imply that labor’s share of income was constant.
However, instead of a linear y − w scatter plot, ACMS found a concave relationship in the US data. The
authors then tested a logarithmic and power relationship and concluded in favor of σ < 1. Integration of
power function (4) then leads to a production function with constant elasticity of substitution, consistent
with definitions (2) and (3).
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 27

11. Under perfect competition, this distribution parameter is equal to the capital income share but, under
imperfect competition with non-zero aggregate mark-up, it equals the share of capital income in total
factor income.
12. See the seminal contribution of Sato and Beckmann (1970) for such a classification.
13. In the case where there is such technical progress, the question of whether σ is greater than or below
unity takes on added importance. Recall, when σ < 1, factors are ‘gross complements’ in production and
‘gross substitutes’ otherwise. Thus, it can be shown that with gross substitutes, substitutability between
factors allows both the augmentation and bias of technological change to favor the same factor. For
gross complements, however, a capital-augmenting technological change, for instance, increases demand
for labor (the complementary input) more than it does capital, and vice versa. By contrast, when σ = 1
an increase in technology does not produce a bias towards either factor (factor shares will always be
constant since any change in factor proportions will be offset by a change in factor prices).
14. Neutrality concepts associate innovations to related movements in marginal products and factor ratios.
An innovation is Harrod-neutral if relative input shares remain unchanged for a given capital-output
ratio. This is also called labor-augmenting since technical progress raises production equivalent to an
increase in the labor supply. More generally, for F(X i , X j , . . . , A), technical progress is X i -augmenting
if FA A = FX i X i .
15. See the discussion in León-Ledesma et al. (2010b) and possible observational equivalence in examining
income share developments and inferring the associated bias in technical progress.
16. We used the word ‘deep’ when we discuss the C and π parameters. By deep, we meant that it is
not dependent on any parameter other than itself. In the implicitly normalized case, the parameters of
the production function (the efficiency and distribution parameter) are functions of σ (except for the
counterfactual case of K = N = 1), and are therefore not deep.
17. We confine ourselves to constant-returns production functions. This is largely done to be consistent with
much of the aggregate evidence (e.g., Basu and Fernald 1997).
18. Given that the real user cost and real interest rate can be sometimes negative in historical samples
(particularly in the 1970s), the user cost conditions is usually expressed in levels rather than logarithms.
Note, the last two conditions in some estimation cases are merged in many papers:
   
Kt wt
log = αi + σ log + (γ N − γ K ) (1 − σ ) t
Nt rt

 
Kt σ rt K t
log = αj − log + (γ N − γ K ) t
Nt 1−σ wt Nt

19. Linearization around a unitary substitution is algebraically the most convenient form, as can be easily
verified.
20. This is quadratic or higher depending on the order of the approximation.
21. The Kmenta approximation, both empirically and in terms of general identification, has enjoyed limited
success (see Kumar and Gapinski 1974; Thursby 1980; León-Ledesma et al. 2010a).
22. Only in the log-linear case of Cobb–Douglas would one expect ξ to exactly equal unity. Hence, in
choosing the sample average as the point of normalization we lose precision because of the CES’s non-
linearity. If, alternatively, we choose the sample mid-point as the normalization point, we should also
lose because of stochastic (and in actual data, cyclical) components that would also imply non-unitary ξ.
23. Duffy and Papageorgiou (2000) suggest developing countries may be better empirically represented by
an above-unity aggregate substitution elasticity.
24. Prompted by this remark in a first draft of our survey, Jakub Growiec explored the link in Growiec
(2011).
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
28 KLUMP ET AL.

25. Since Hicks (1932), the value of the substitution elasticity has often been seen as reflecting economic
flexibility and thus deep institutional factors such as labor bargaining power, the taxation burden, degree
of economic openness, the characteristics of national education system, etc. Accordingly, some view
changes in the substitution elasticity as drivers of endogenous growth and as such potentially even
more important than traditionally studied growth factors such as savings and technical progress (La
Grandville 2009; Yuhn 1991). This is termed the ‘de La Grandville Hypothesis’ following Yuhn (1991)
(i.e., the conjecture that the high growth rate of east Asian countries was due not to a higher rate of
technical progress, but to a higher elasticity of substitution). Also, earlier, Solow (1956) and Pitchford
(1960) showed in the neoclassical growth model that a CES function with an elasticity of substitution
greater than one can generate sustained growth (even without technical progress). See also Masanjala and
Papageorgiou (2004) for a theoretical model where the aggregate elasticity of substitution is endogenized
and depends on the level of economic development.
26. The capital–skill hypothesis gained particular currency given the sharp decline in the constant-quality
relative price of equipment, for example, Gordon (1990), particularly for information and communication
technologies. This decline naturally led to an uptake in usage of such capital. Given complementarity
between capital and skilled labor, the faster usage of such capital increased the relative demand for
skilled labor and – despite the apparent increase in the supply of such labor – the skill or wage premium
relative to unskilled labor increased in a dramatic and persistent manner. On the other hand, authors such
as Katz and Murphy (1992), Acemoglu (2002b) and Autor et al.(2008) claimed that the skill premium
can be attributed to technical change that was biased in favor of skilled workers. Given that skilled and
unskilled workers are gross substitutes, an increase in skilled labor efficiency led to an increase in the
relative wages (and factor shares) of skilled workers.

References
Acemoglu, D. (2002a) Directed technical change. Review of Economic Studies 69:781–809.
Acemoglu, D. (2002b) Technical change, inequality, and the labor market. Journal of Economic Literature
40(1):7–72.
Acemoglu, D. (2003) Labor- and capital-augmenting technical change. Journal of the European Economic
Association 1:1–37.
Acemoglu, D. (2009) Introduction to Modern Economic Growth. Boston, MA: MIT Press.
Allen, R. G. D. (1938) Mathematical Analysis for Economists. London: MacMillan and Co Ltd.
Antràs, P. (2004) Is the US aggregate production function Cobb-Douglas? New estimates of the elasticity of
substitution. Contributions to Macroeconomics 4: Article 4.
Arrow, K. J., Chenery, H. B., Minhas, B. S. and Solow, R. M. (1961) Capital-labor substitution and economic
efficiency. Review of Economics and Statistics 43(3):225–250.
Autor, D. H., Katz, L. F. and Kearney, M. S. (2008) Trends in U.S. Wage Inequality: Revising the Revisionists.
Review of Economics and Statistics 90(2):300–323.
Azariadis, C. (1996) The economics of poverty traps - Part One: Complete markets. Journal of Economic
Growth 1:449–486.
Barro, R. J. and Sala i-Martin, X. (2004) Economic Growth. 2nd edn, Cambridge, MA: MIT Press.
Basu, S. and Fernald, J. (1997) Returns to Scale in U.S. Manufacturing: Estimates and Implications. Journal
of Political Economy 105(2):249–283.
Beckenbach, E. F. and Bellman, R. (1961) Inequalities. Berlin: Springer.
Bentolila, S. and Saint-Paul, G. (2003) Explaining movements in the labor share. Contributions to
Macroeconomics 3:Article 9.
Berndt, E. R. (1976) Reconciling alternative estimates of the elasticity of substitution. Review of Economics
and Statistics 58(1):59–68.
Berthold, N., Fehn, R. and Thode, E. (2002) Falling labour share and rising unemployment: Long-run
consequences of institutional shocks? German Economic Review 3:431–459.
Blanchard, O. J. (1997) The medium run. Brookings Papers on Economic Activity 2:89–158.
Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31
C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 29

Box, G. and Cox, D. (1964) An analysis of transformations. Journal of the Royal Statistical Society 26, Series
B:211–243.
Caballero, R. J. (1994) Small sample bias and adjustment costs. Review of Economics and Statistics 85:153–65.
Caballero, R. J. and Hammour, M. L. (1998) Jobless growth: Appropriablity, factor substitution and
unemployment. Carnegie-Rochester Conference Proceedings 48:51–94.
Cantore, C., León-Ledesma, M. A., McAdam, P. and Willman, A. (2010) Shocking Stuff: Technology, Hours
and Factor Substitution. ECB Working Paper 1278.
Cantore, C. and Levine, P. (2011) Getting Normalization Right. Mimeo, University of Surrey.
Chirinko, R. S. (2002) Corporate taxation, capital formation, and the substitution elasticity between labor
and capital. National Tax Journal 60:339–355.
Chirinko, R. S. (2008) σ : The long and short of it. Journal of Macroeconomics 30(2):671–686.
Chirinko, R. S., Fazzari, S. M. and Meyer, A. P. (1999) How responsive is business capital formation to its
user cost? Journal of Public Economics 74(1):53–80.
David, P. A. and van de Klundert, T. (1965) Biased efficiency growth and capital-labor substitution in the
US, 1899–1960. American Economic Review 55(3):357–394.
De Jong, F. J. (1967) Dimensional Analysis for Economists. North Holland.
De Jong, F. J. and Kumar, T. K. (1972) Some considerations on a class of macro-economic production
functions. De Economist 120(2):134–152.
Diamond, P. A. and McFadden, D. (1965) Identification of the Elasticity of Substitution and the bias of
Technical Change: An Impossibility Theorem. Working Paper No. 62, University of California Berkeley.
Diamond, P. A., McFadden, D. and Rodriguez, M. (1978) Measurement of the elasticity of substitution
and bias of technical change. In M. Fuss and D. McFadden (eds), Production Economics (Vol. 2, pp.
125–147) Amsterdam: North Holland.
Dickinson, H. D. (1954) A note on dynamic economics. Review of Economic Studies 22(3):169–179.
Dimond, R. W. and Spencer, B. J. (2008) Trevor Swan and the neoclassical growth model. NBER Working
Paper 13950 99(2):256–259.
Duffy, J. and Papageorgiou, C. (2000) A cross-country empirical investigation of the aggregate production
function specification. Journal of Economic Growth 5(1):86–120.
Gordon, R. (1990) The Measurement of Durable Goods Prices. University of Chicago Press.
Griliches, Z. (1969) Capital-skill complementarity. Review of Economics and Statistics 51(4):465–468.
Growiec, J. (2008) Production functions and distributions of unit factor productivities: Uncovering the link.
Economics Letters 101(1):87–90.
Growiec, J. (2011) A Microfoundation for Normalized CES Production Functions with Factor-Augmenting
Technical Change. Working Paper No. 98, National Bank of Poland.
Guo, J. T. and Lansing, K. J. (2009) Capital-labour substitution and equilibrium indeterminacy. Journal of
Economic Dynamics and Control 33(12):1991–2000.
Harbrecht, W. (1975) Substitutionselastizitat und Gleichgewichtswachstum. Jahrbucher fur Nationalokonomie
und Statistik 189:190–201.
Hardy, G. H., Littlewood, J. E., and Pólya, G. (1934) Inequalities. Cambridge University Press, 2nd edn.
1952.
Hicks, J. R. (1932) The Theory of Wages. London: Macmillan.
Hicks, J. R. (1970) Elasticity of substitution again: Substitutes and complements. Oxford Economic Papers
22:289–296.
Houthakker, H. S. (1955) The Pareto distribution and the Cobb-Douglas production function in activity
analysis. Review of Economic Studies 23:27–31.
Irmen, A. (2011) Steady-state growth and the elasticity of substitution. Journal of Economic Dynamics and
Control 35(8):1215–1228.
Jensen, B. S. (2011) Capital, production functions and growth theory - von Thünen, Douglas and Solow.
Mimeo, University of Southern Denmark.
Jones, C. I. (2003) Growth, Capital Shares, and a New Perspective on Production Functions. Mimeo,
University of California Berkeley.
Jones, C. I. (2005) The shape of production functions and the direction of technical change. Quarterly
Journal of Economics 120(2):517–549.

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31

C 2012 Blackwell Publishing Ltd
30 KLUMP ET AL.

Kaas, L. and von Thadden, L. (2003) Unemployment, factor substitution, and capital formation. German
Economic Review 4:475–495.
Kaldor, N. (1961) Capital accumulation and economic growth. In F. A. Lutz and D. C. Hague (eds), The
Theory of Capital. St. Martin’s Press.
Kamien, M. I. and Schwartz, N. L. (1968) Optimal Induced technical change. Econometrica 36(1):1–17.
Katz, L. and Murphy, K. (1992) Changes in relative wages, 1963-1987: Supply and demand factors. Quarterly
Journal of Economics 107:35–78.
Kennedy, C. (1964) Induced bias in innovation and the theory of distribution. Economic Journal 74:541–
547.
Kennedy, C. and Thirlwall, A. P. (1973) Technological change and the distribution of income: A belated
comment. International Economic Review 14(3):780–784.
Klump, R. (2001) Trade, money and employment in intertemporal optimizing models of growth. Journal of
International Trade and Economic Development 10(4):411–428.
Klump, R. (2002) Verteilung und Wirtschaftswachstum: Eine neoklassische Verallgemeinerung des
Wachstumsmodell von Kaldor. In L. Menkhoff and F. Sell (eds), Zur Theorie, Empirie und Politik
der Einkommensverteilung (pp. 11–26). Berlin-Heidelberg-New York.
Klump, R. and de La Grandville, O. (2000) Economic growth and the elasticity of substitution: Two theorems
and some suggestions. American Economic Review 90(1):282–291.
Klump, R. and Irmen, A. (2009) Factor substitution, income distribution and growth in a generalized
neoclassical model. German Economic Review 10(4):464–479.
Klump, R., McAdam, P. and Willman, A. (2007) Factor substitution and factor-augmenting technical progress
in the US. Review of Economics and Statistics 89(1):183–92.
Klump, R., McAdam, P. and Willman, A. (2008) Unwrapping some Euro area growth puzzles: Factor
substitution, productivity and unemployment. Journal of Macroeconomics 30(2):645–666.
Klump, R. and Preissler, H. (2000) CES production functions and economic growth. Scandinavian Journal
of Economics 102(1):41–56.
Klump, R. and Saam, M. (2008) Calibration of normalized CES production functions in dynamic models.
Economics Letters 99(2):256–259.
Kmenta, J. (1967) On estimation of the CES production function. International Economic Review
8(2):180–189.
Krussel, P., Ohanian, L., Rios-Rull, J. V. and Violante, G. (2000) Capital-skill complementarity and inequality.
Econometrica 68(5):1029–1053.
Kumar, T. K. and Gapinski, J. H. (1974) Nonlinear estimation of the CES production function parameters:
A Monte Carlo. Review of Economics and Statistics 56(4):563–567.
La Grandville, Olivier de. (1989a) Curvature and the elasticity of substitution: Straightening it out. Journal
of Economics 66(1):23–34.
La Grandville, Olivier de. (1989b) In quest of the Slutzky diamond. American Economic Review 79:468–481.
La Grandville, Olivier de. (2009) Economic Growth: A Unified Approach. (pp. 389–416) Cambridge University
Press.
La Grandville, Olivier de. (2012) How Much Should a Nation Save? A New Answer. In La Grandville,
Olivier de. (ed), Frontiers of Economics and Globalization (Vol. 11). Emerald. forthcoming.
La Grandville, Olivier de. and Solow, R. M. (2006) A conjecture on general means. Journal of Inequalities
in Pure and Applied Mathematics 7(3).
La Grandville, Olivier de. and Solow, R. M. (2009) Capital-labour substitution and economic growth. In
Economic Growth: A Unified Approach. (pp. 389–416) Cambridge University Press (chapter 5).
León-Ledesma, M. A., McAdam, P. and Willman, A. (2010a) Identifying the elasticity of substitution with
biased technical change. American Economic Review 100(4):1330–1357.
León-Ledesma, M. A., McAdam, P. and Willman, A. (2010b) In Dubio pro CES: Supply Estimation with
Mis-Specified Technical Change. European Central Bank, Working Paper No. 1175.
León-Ledesma, M. A., McAdam, P. and Willman, A. (2012) Aggregation, the skill premium, and the two-
level production function. In de La Grandville, O. (ed), Economic Growth and Development (Frontiers
of Economics and Globalization) (Vol. 11, chapter 15, pp. 417–436). Emerald Group Publishing
Limited.

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31

C 2012 Blackwell Publishing Ltd
 THE NORMALIZED CES PRODUCTION FUNCTION: THEORY AND EMPIRICS 31

León-Ledesma, M. A. and Satchi, M. (2010) A note on balanced growth with a less than unitary elasticity
of substitution. Studies in Economics 1007, Department of Economics, University of Kent.
Lucas, R. E. (1969) Labor-capital substitution in US manufacturing. In A. C. Harberger and M. J. Bailey
(eds ), The Taxation of Income from Capital (pp. 223–274). Washington, D. C.: Brookings Institution.
Mallick, D. (2010) Capital-labor substitution and balanced growth. Journal of Macroeconomics
32(4):1131–1142.
Masanjala, W. H. and Papageorgiou, C. (2004) The Solow model with CES technology: Nonlinearities and
parameter heterogeneity. Journal of Applied Econometrics 19:171–201.
McAdam, P. and Willman, A. (2013) Medium run redux. Macroeconomic Dynamics (forthcoming).
McAdam, P. and Willman, A. (2011b) Technology, Utilization and Inflation: What Drives the New Keynesian
Phillips Curve? ECB Working Paper 1368.
McElroy, F. W. (1967). Notes on the CES production function. Econometrica 35:154–156.
Nakamura, H. and Nakamura, M. (2008) Constant-elasticity-of-substitution production function.
Macroeconomic Dynamics 12(5):694–701.
Papageorgiou, C. and Saam, M. (2008) Two-level CES production technology in the Solow and Diamond
growth models. Scandinavian Journal of Economics 110(1):119–143.
Paroush, J. (1964) A note on the CES production function. Econometrica 32:213–156.
Pitchford, J. D. (1960) Growth and the elasticity of substitution. Economic Record 36:491–504.
Ramanathan, R. (1975) The elasticity of substitution and the speed of convergence in growth models.
Economic Journal 85:612–613.
Robinson, J. (1933) The Economics of Imperfect Competition. London: MacMillan and Co Ltd.
Rowthorn, R. (1999) Unemployment, wage bargaining and capital-labour substitution. Cambridge Journal of
Economics 23(4):413–425.
Rutherford, T. (2003) Lecture notes on constant elasticity functions. Mimeo, University of Colorado.
Saam, M. (2008) Openness to trade as a determinant of the macroeconomic elasticity of substitution. Journal
of Macroeconomics 30:691–702.
Samuelson, P. A. (1965) A theory of induced innovations along Kennedy-Weisacker Lines. Review of
Economics and Statistics 47(4):344–356.
Samuelson, P. A. (1972) Collected Scientific Papers (Vol. 2). Cambridge, MA: MIT Press.
Sato, K. (1967) A two-level constant-elasticity-of-substitution production function. Review of Economic
Studies 34(2):201–218.
Sato, R. (2006). Biased Technical Change and Economic Conservation Laws. Springer: New York.
Sato, R. and Beckmann, M. J. (1970). The estimation of biased technical change and the production function.
Review of Economic Studies 35(1):57–266.
Solow, R. M. (1956) A contribution to the theory of economic growth. Quarterly Journal of Economics
70(1):65–94.
Swan, P. L. (2006) ANU Inaugural Trevor Swan Distinguished Lecture. Available at
[Link]
Swan, T. (1956) Economic growth and capital accumulation. Economic Record 32:334–361.
Temple, J. R. W. (2008) The calibration of CES production functions. Discussion Paper 606, Department of
Economics, University of Bristol.
ten Cate, A. (1992) Data scaling with highly non-linear regression. Computational Statistics 7:59–65.
Thursby, J. (1980) Alternative CES estimation techniques. Review of Economics and Statistics 62(2):259–299.
Ventura, J. (1997) Growth and interdependence. Quarterly Journal of Economics 62:57–84.
Wong, T. N. and Yip, C. K. (2010) Indeterminacy and the elasticity of substitution in one-sector models.
Journal of Economic Dynamics and Control 34(4):623–635.
Xue, J. and Yip, C. K. (2013) Factor substitution and economic growth: A unified approach. Macroeconomic
Dynamics. forthcoming.
Yasui, T. (1965) The CES production function: A note. Econometrica 33:646–648.
Yuhn, K.-H. (1991) Economic growth, technical change biases, and the elasticity of substitution: A test of
the De La Grandville hypothesis. Review of Economics and Statistics 73(2):340–346.

Journal of Economic Surveys (2012) Vol. 0, No. 0, pp. 1–31

C 2012 Blackwell Publishing Ltd