Documente Academic
Documente Profesional
Documente Cultură
Beyond Black-Litterman:
views on non-normal markets
In normally distributed markets, the Black-Litterman technique allows managers to construct
portfolios that account for their views on a set of expected returns. Attilio Meucci extends the
Black-Litterman framework to generic market distributions and shows an application to
portfolio management in a thick-tailed, skewed and highly co-dependent market
B
lack & Litterman’s (1990) (BL) ground-breaking technique allows a is directly inherited from the prior market structure. Finally, a suitable
portfolio manager to smoothly blend his subjective views on the change of co-ordinates allows us to translate the joint distribution of the
market with a prior market distribution. The popularity of the BL views into a joint posterior distribution for the market.
methodology is due to its intuitiveness, flexibility and ease of implemen- The purpose of the proposed copula-opinion pooling (COP) methodolo-
tation. The portfolio manager can express an arbitrary number of views gy is to determine one market distribution (the ‘posterior’) that cogently, in
on generic linear combinations of market expected values and immedi- a sense defined in the article, incorporates two different distributions (the
ately obtain the market distribution that reflects these views in the most ‘views’ and the ‘prior’). In the special case where the market describes the
consistent way. returns on a set of securities, the most natural application is in the context of
Nevertheless, the BL approach suffers from two drawbacks. The first portfolio management, where linear coefficients (exposures) map the mar-
problem is of a practical nature: in the BL approach, both the market prior ket distribution into the portfolio distribution. On the other hand, the market
and the manager’s views are normally distributed. For most markets, the distribution can represent any source of randomness, not necessarily asset
normal assumption is too strong: fat tails, skewness and high dependence returns. Therefore, the proposed methodology can find applications, among
among extreme events characterise the joint distribution of market risk fac- others, on the trading floor. In this case, the prior can be set, for instance, in
tors in many contexts. As far as the views are concerned, the manager terms of quantities implied by the market prices and the views can be pro-
might think, for instance, in terms of ranges within whose boundaries all vided by some technical indicator, thereby yielding a rich-cheap analysis.
market realisations are equally likely. This corresponds to uniformly dis- First, we present the general theory. Then we compare the COP ap-
tributed views, as opposed to the ‘alpha plus normal noise’ specification proach with the traditional BL methodology in a simulated normal market.
in the BL approach. After that, we apply the general theory to a more realistic case. We model
The normality assumption about the market and the views can be over- the thick-tailedness, skewness and high dependence among extreme events
come within the BL framework, if one is willing to accept numerical re- in the market by means of a multivariate skew t distribution and we model
sults instead of closed analytical formulas. the manager’s views as uniformly distributed on arbitrary ranges. We cal-
The second problem in the BL approach is conceptual: due to its culate the highly non-symmetrical posterior market distribution. To pur-
Bayesian nature, in this context the manager is supposed to express views sue the optimal allocation, instead of the classical mean-variance approach,
on the parameters that determine the market distribution. In reality, in gen- which is suitable in elliptical markets, we use the expected return/expected
eral the manager expresses views on the possible realisations of the mar- shortfall (CVAR) approach, which accounts for asymmetries and tail risk.
ket. This is not evident under the normal assumption, because there exists We then conclude.
a clear relationship between the parameters and the realisations of the mar-
ket. Indeed, in this case the parameters on which the manager expresses The general theory
the views (the ‘µµ’) represent the expected values of the market realisations. Consider an N-dimensional market vector M, that is, a set of random vari-
Under different distributional assumptions, such as the classical lognormal ables that determines the randomness in the market. For instance, M could
Black-Scholes hypothesis, the relationship between the parameters (the ‘µ µ’ represent the returns on a set of securities, or a collection of risk factors
of the lognormal distribution) and the realisations of the market is not trans- in an arbitrage pricing theory-like model, as well as the times-to-default of
parent and the practitioner would not be able to express views on the pa- a basket of bonds in a collateralised debt obligation tranche.
rameters in an intuitive way. The statistical properties of the market are described by an ‘official’ prior
As far as this problem is concerned, it is possible to reformulate the BL distribution, which is determined by means of backward-looking estima-
approach in terms of views on the market instead of the market parame- tion techniques, forward-looking market-implied values, equilibrium ar-
ters (see Meucci, 2005). Nevertheless, this reformulation prescribes that the guments, etc. We represent this distribution equivalently in terms of its
manager’s views be input in a rather counter-intuitive way as statements probability density function (PDF), its cumulative distribution function
conditioned on the realisation of the market. Instead, it is more natural to (CDF) or its characteristic function (CF) respectively:
solve the dichotomy between the practitioner’s ‘subjective’ views and the
M ~ ( f M , FM , φM ) (1)
‘official’ prior market distribution in terms of opinion pooling. In this arti-
cle, we draw on this interpretation to improve on the BL approach, pro- As in the BL framework, the views are a set of K ≤ N statements on
viding a posterior market distribution that smoothly blends an arbitrarily generic linear combinations of the market. Unlike in the BL framework,
distributed market prior with arbitrarily distributed manager’s views on the where the portfolio manager expresses views on the market parameters,
realisation of the market. the portfolio manager expresses views directly on the market realisations
The intuition behind our approach is as follows. We use opinion pool- M. The above linear combinations are represented by a K × N dimen-
ing criteria to determine the marginal distribution of each view separate- sional ‘pick’ matrix P. Therefore the views are the following K-variate ran-
ly, whereas the joint co-dependence, that is, the copula, among the views dom vector:
V ≡ PM (2) ⎛ V ≡ PM ⎞
M⇔⎜ ⎟ (8)
⎝ W ≡ P ⊥ M⎠
In the most general setting, the manager expresses the generic kth view in
terms of a subjective univariate distribution, which is represented by its where P⊥ is any (N – K) × N matrix whose rows span the null space of
PDF, CDF or CF: P.1 In the view-adjusted co-ordinates, the posterior distribution of the
fV , FV ,φV first entries was determined in (7), whereas the posterior distribution
k k k (3)
of the remaining entries has been left unaltered. The posterior market
On the other hand, the prior market distribution (1) induces a structure distribution is then defined naturally by reverting to the original market
on the distribution of the views. Indeed, the marginal distribution of the co-ordinates:
generic kth view reads as follows in terms of its CF: −1
d⎛ P ⎞ ⎛V ⎞
=
M
Vk ~ φVk (ω) = φM (p′k ω ) (4) ⎜ ⊥ ⎟ ⎜ W⎟ (9)
⎝P ⎠ ⎝ ⎠
where pk denotes the kth row of the pick matrix P. The COP posterior distribution of the market displays a few intuitive
The K market-induced distributions (4) clash with the manager’s properties. For instance, views that coincide with the market prior leave
subjective views (3). Our aim is to determine a distribution for the the market prior unaltered; this follows immediately from (5). Further-
market, which we call the posterior market distribution, that solves more, as in the BL framework, the whole market distribution is affected
this dichotomy. by the views. We can see this clearly in the limit case where the manag-
To determine the posterior market distribution we adopt a ‘bottom-up’ er’s uncertainty range relative to the generic kth view collapses to a spe-
approach. First, we determine the posterior marginal distribution of each cific value vk and where the confidence on that view is full, for all the
view separately by drawing on the literature on opinion pooling. Then we views. As we show in the technical appendix (see www.symmys.com), in
determine the joint posterior distribution of the views by using the de- this circumstance we obtain the intuitive result that the COP posterior mar-
pendence structure inherited from the market prior. Finally, we determine ket distribution is represented by the prior market distribution conditioned
the joint posterior distribution of the market by embedding the views in a on the realisation of the views:
suitable set of market co-ordinates. d
According to the plan outlined above, we first define the posterior PDF =M
M (10)
PM ≡ v
of the generic kth view as a weighted average of the market-induced prior
PDF and the subjective PDF that expresses the view: This is the case also in the BL approach, where the posterior normal dis-
tribution becomes the conditional normal distribution in the limit of total
Vk ~ fV ≡ (1 − ck ) fVk + ck fV (5) confidence (see Meucci, 2005).
k k
where the weight ck spans the interval [0, 1]. There is a vast literature on A comparison with the BL methodology
the choice of the ‘best’ opinion pooling method (see Genest & Zidek, 1986, Here, we compare the COP approach discussed above with the BL method-
for a review). As discussed in Clemen & Winkler (1997), the simple rule ology. To meet the assumptions of the BL approach, we consider a simu-
(5) is as effective as more complex pooling techniques and it has the ad- lated normal market, where the views are input according to the ‘alpha
vantage of being intuitive. In particular, the weight ck can be interpreted plus normal noise’ prescription.
as the confidence in the manager’s views. If the confidence is null, the pos- We perform our comparison directly on the posterior distributions en-
terior distribution results in the original market-induced distribution; if the suing from the BL approach and the COP approach. Alternatively, we
confidence is full, the posterior distribution coincides with the manager’s could compare the two approaches on the basis of the optimal alloca-
statement. The confidence ck can be input directly by the manager. Alter- tions to which they give rise. However, we do not pursue this route for
natively, it can be imposed exogenously, based on the manager’s track two reasons. First, this article proposes a methodology to calculate a pos-
record. For instance, one could link the confidence to the correlation be- terior distribution that can be used in any context, such as a rich-cheap
tween past views and actual market realisations (see Grinold & Kahn, 1999). analysis for the trading desk, and not necessarily only for portfolio man-
Furthermore, we see that generalisations of (5) to a multi-manager context agement purposes. Second, the outcome of a portfolio optimisation de-
are immediate. pends on the market distribution as much as it depends on the
As a second step in the COP approach, we determine the joint distrib- optimisation technique (mean-variance, full grid-search, genetic algo-
ution of the views. To determine a joint distribution that is consistent with rithms, etc). This would add a spurious degree of subjectivity in the com-
the posterior marginal distributions of the views (5) and reflects the co-de- parison. For an explicit application of the theory to portfolio management,
pendence structure of the original market prior (1), we first calculate the refer to the next section.
copula of the views induced through (2) by the market prior: We consider an artificial ‘prior’ market, which is normally distributed:
M ~ N (m S)
µ, Σ
( )′
d (11)
C = FV1 (V1 ) ,..., FVK (VK ) (6)
This market is composed of N ≡ 25 securities. The parameters of this mar-
Then we impose the marginal structure (5), obtaining the following joint ket are set as follows. The standard deviations read:
posterior distribution for the views:
σ1 ≡ 5%, σ 2 ≡ σ1 + ∆ ,..., σ N ≡ 25% ≡ σ1 + ( N − 1) ∆
(
≡ F −1 (C ) ,..., F −1 (C ) ′ )
d
V V 1 V K
(7)
1 K
The correlations are assumed to be homogeneous. In particular, we com-
where F–1 X denotes the quantile function relative to the CDF FX. pare the two approaches in a highly correlated environment:
The final step in the COP approach is the determination of a posterior ρmn ≡ 0.9, m ≠ n
distribution for the market. To determine a market posterior distribution
that is consistent with the joint posterior distribution of the views (7) we 1As we show in the technical appendix at www.symmys.com, the specific choice of P⊥
notice that the market vector can be expressed equivalently in a set of does not influence the final result. Nevertheless, to improve the efficiency of the
view-adjusted co-ordinates: calculations we impose that the rows of P⊥ be orthogonal
Distance
the mean is bearish:
(
V ≡ M N ~ N − µ N , σ 2N ) (12) 10
{ }
that the applications of the theory go beyond the portfolio manage-
d c (COP, BL) ≡ E ln f M (c) (M ) − ln f M BL (c) (M ) (16) ment context.
COP
We consider an international-equity fund manager, whose investment
where the expectation is calculated according to the prior distribution (11). horizon in one week and whose market M is represented by the returns
This distance is reminiscent of the Kullback-Leibler relative entropy, ex- on N ≡ 4 international stock indexes: the US S&P 500, the UK FTSE 100,
cept that it hinges on the natural benchmark, namely the prior distribu- the French Cac 40 and the German Dax.
tion. We can easily calculate this distance as a function of the overall level In figure 2, we plot the weekly returns of the US index versus the UK
0.05
µ = δΣ
Σweq (18)
0
where weq is the relative capitalisation of the four above indexes and δ
≈ 2.5.
–0.05
Notice that in this context the specification (18) is apparently nonsen-
sical. Indeed, a sufficient condition for a capital asset pricing model (CAPM)-
–0.10
like specification such as (18) to be viable is that the returns be elliptically
distributed (see Ingersoll, 1987). The skew t in general is inconsistent with
a CAPM-like equilibrium specification, because it is not elliptical. Further-
–0.10 –0.05 0 0.05 0.10 0.15
more, in general (18) is not even a CAPM equilibrium specification be-
S&P 500
cause µ does not represent the expected value of the skew t distribution.
However, since α ≈ 0 the skew t distribution reduces to a regular t distri-
bution, which is elliptical and thus CAPM-like equilibrium specifications
index (the plot is qualitatively similar for other combinations). The market are viable. Furthermore, when α ≈ 0, µ represents the expected value of
M is not normally distributed: the marginal distributions of the above in- the returns; therefore (18) is a CAPM-like equilibrium specification. To
dexes display fat tails, potentially some degree of skewness, and the cop- show the calculations in their full generality, we continue the discussion
ula of the above indexes displays larger than normal dependence among under the more general assumption of a general skew t distribution. We
extreme events (see also Mashal & Zeevi, 2002). stress that in the more general case α ≠ 0, the most suitable estima-
We can model these features by means of the skew t distribution (see tion/modelling methodology should be tailored to the needs of the spe-
Azzalini & Capitanio, 2003): cific investor and market.
As far as the manager’s views are concerned, we model them as uni-
M ~ SkT (ψ, µ, Σ, α) (17) formly distributed on given ranges. Any other specification is equally fea-
sible, although parsimonious specifications should be favoured. Indeed,
In this expression ψ is a positive number, µ is an N-dimensional vector, one of the criticisms of the (very parsimonious) ‘alpha plus normal noise’
Σ is an N × N symmetric and positive matrix, and α is an N-dimension- specification of BL is that it already has too many inputs. In formulas, the
al vector. We report in the technical appendix the PDF f SkT
ψ, µ, Σ, α of this CDF of the generic kth view reads as follows:
distribution. The skew t distribution coincides with the Student t distri-
⎧ 0 v ≤ ak
bution when the shape parameter α is null. In this case, µ is the ex- ⎪
pected value and Σ is the (rescaled) covariance matrix. In particular, in FV (v) ≡
k
F[Ua ,b ]
k k
(v) ≡ ⎪⎨ bvk−−aakk v ∈[ ak , bk ] (19)
the limit ψ → ∞ we recover the normal distribution. As the degrees of ⎪
⎪⎩ 1 v ≥ bk
freedom ψ decrease, this distribution displays heavier tails and larger
dependence among extreme events (see Embrechts, Lindskog & McNeil, In particular, we consider the case where the manager expresses one
2003). For non-null values of the shape parameter α, the skew t distri- view, although for the reader’s convenience in future applications, in
bution also displays skewness. what follows we use the multi-views notation. The manager is bearish
We estimate the parameters of the distribution in the example by on the Dax, which he believes will realise between zero and –2% with-
maximum likelihood (see Azzalini, 2005, for details). We report the re- in a week. In other words, the pick matrix and the ranges read respec-
sults below: tively in this case:
where: 0.6
CDF
ξ ≡ Pµ
µ, Φ ≡ PΣ
Σ P′
0.4
and:
0.2
diag (F
Φ) 2 Φ
1
F −1H ′aα
β≡
b
( ))
1
⎛1 + a ′ diag S − 12 S diag S − 12 − HFΦ−1H ′ a 0
(Σ) Σ (Σ)
2
α
⎝ α
− 12 –0.10 –0.08 –0.06 –0.04 –0.02 0 0.02 0.04 0.06 0.08 0.10
H ≡ diag ( S
Σ) Σ
SP ′ Return
The PDFs of each view (20) can then be integrated by means of efficient
numerical techniques similar to those that yield the error function in the
case of the normal distribution: to multiple indexes of satisfaction or dissatisfaction such as value-at-risk,
coherent measures of outperformance with respect to a benchmark, or ex-
FVk = FψSkT
, ξ k , Φ kk ,β k (21) pected utility (see Meucci, 2005, for a thorough discussion of these issues).
In practice, it is very common to optimise the risk-reward profile of a port-
By applying the opinion pooling technique (5), we immediately obtain folio in terms of its mean-variance co-ordinates, because in most cases this
the posterior CDF of the generic kth view: optimisation can be solved by quadratic programming and yields close to
optimal results.
FV ≡ (1 − ck ) FψSkT
, ξ k , Φ kk ,β k + ck F[ak ,bk ]
U
(22) Nevertheless, in the present context the posterior market distribution
k
can potentially be heavily twisted by the manager’s views. Unlike in BL,
where we set a 20% confidence in the views: ck ≡ 0.2. In figure 3, we dis- where the ‘twisted’ posterior distribution is again normal, in the proposed
play the CDF of the prior, the view and the posterior. COP approach the posterior is not even elliptical. Asymmetries and tail risk
The posterior quantile functions of the views are easily obtained by lin- can come to play a major role and therefore the mean-variance approach
ear interpolation of their respective CDFs: can potentially become highly sub-optimal.
Therefore, as in Rockafellar & Uryasev (2000), we evaluate the risk-
FV FV−1 (23) iness of an allocation in terms of its expected shortfall2, with a confi-
k k
dence of, say, γ ≡ 95%. The manager minimises the expected shortfall
To derive the posterior distribution of the market, we first generate a subject to the standard long-only and full-investment constraints, as well
large number J of Monte Carlo scenarios of the prior market distribution as the constraint of a minimum target expected return r. In formulas,
(17) using its equivalent stochastic representation (see Azzalini, 2005): the optimal portfolio reads3:
M( j) ≡ µ
m+
g Y ( j ) + BX( j )
, j ∈ {1,..., J } w ′m ≥r
({
w ( r ) ≡ arg max E R R ≤ FR−1 (1 − γ ) }) (26)
(24)
s( j )
w ′ 1≡1, w ≥ 0
ψ
~ ~
where R ≡ w′M is the portfolio return, F –1 X (α) denotes the α-quantile of
In this expression each (X(j), Y (j))
is a drawing from the standard (N + 1)- the generic random variable X, and m ~ is the expected value of the market
~
variate normal distribution; each s (j)
ψ is an independent random drawing M.. As the target expected return r varies, the optimal portfolios (26) span
from the chi-square distribution with ψ degrees of freedom; g is a suitable an efficient frontier in the risk-reward space defined by expected shortfall
conformable vector; and B is a suitable conformable matrix. and target return. These optimal allocations account for the non-normali-
From the Monte Carlo scenarios of the prior market distribution (24), ty of the market.
the prior marginal CDFs (21) and the posterior quantiles (23), we im- To determine the optimal allocations in our scenario-based posteri-
mediately obtain Monte Carlo scenarios of the posterior market distrib- or market distribution (25), we make use of a result in Rockafellar &
ution (9): Uryasev (2000), according to which (26) can be quickly solved by lin-
ear programming. We provide more details in the technical appendix at
⎛ ⎞
⎜ ⎟ www.symmys.com.
⎛ P⎞
⎜ F −1 FV p1M ( j )
−1 ⎜ Vk 1 ( ( )) ⎟
⎟
In figure 4, we display the efficient prior and posterior frontiers.
The posterior frontier is lower than the prior frontier because the bear-
( j)
M ≡ ⎜ ⊥⎟ ⎜ ⎟ ish view on the Dax shifts the investment opportunities into a less
⎝P ⎠ ⎜ ⎟
⎜ F −1 F p M ( j )
⎜ Vk VK K ( ( )) ⎟
⎟
(25)
favourable region.
In table A, we display the optimal posterior portfolio compositions (26)
⎜ ⎟
⎝ P ⊥M( j) ⎠ 2 The expected shortfall is also known as conditional value-at-risk (CVAR). See Meucci
(2005) for issues regarding the terminology
We can now proceed to determine the optimal asset allocation. In the- 3 This expression is exact for continuous distributions and somewhat different for discrete
~
ory, the manager evaluates a portfolio w in a given market M according distributions. See Rockafellar & Uryasev (2002) for more details
Ingersoll E, 1987
Theory of financial decision making
B. Prior allocation Rowman and Littlefield
P⊥ 7→ AP⊥ , (29)
Therefore:
µ ¶−1 µ ¶µ ¶
d P IK×K 0K×(N−K) e
V
fA
M = (32)
P⊥ 0(N −K)×K A−1 AP⊥ M
µ ¶−1 µ ¶
P e
V d f
= =M
P⊥ P⊥ M
In the case of infinite confidence in one specific value PM ≡ v the above ex-
pression reads:
µ ¶−1 µ ¶ µ ¶−1 µ ¶ ¯¯
f= d P v d P P ¯
M = M¯ . (34)
P⊥ P⊥ M P⊥ P⊥ ¯
PM≡v
Therefore:
f=d
M M|PM≡v , (35)
i.e. the posterior market distribution is the prior distribution conditioned on
the realization of the views.
16
5.3 Comparison between BL posterior and pooling poste-
rior
We need to compute
n¯ ¯o
¯ ¯
ρc (OP, BL) ≡ E ¯ln fM
i OP (c) (M) − ln fMiBL (c) (M) ¯ , (36)
where
M ∼ N (µ, Σ) . (37)
To compute the first function in (36), we first derive the distribution of the
market in the views coordinates. Define
µ ¶
P
R≡ . (38)
P⊥
Then ¡ ¢
V ≡ RM ∼ N ν ≡ Rµ, Ξ ≡ RΣR0 . (39)
From this expression we can compute the pdf of the copula, see e.g. Meucci
(2005) p. 41:
³ ´
N
fν,Ξ QN N
ν 1 ,ξ11 (u1 ) , . . . , Qν N ,ξN N (uN )
fU (u1 , . . . , uN ) = ³ ´ ³ ´ , (40)
fνN1 ,ξ11 QN N N
ν 1 ,ξ11 (u1 ) · · · fν N ,ξN N Qν N ,ξ NN (uN )
where fνNn ,ξnn and QN ν n ,ξnn are the normal quantile and the normal pdf of the
generic n-th marginal entry of V respectively with parameters provided by (39);
N is the joint normal pdf with parameters provided by (39). The posterior
and fν,Ξ
pdf in the views coordinates then reads:
³
fV
h (v 1 , . . . , v N ) = fU FV h1 (v1 ) , . . . , FV
hK (vK ) , (41)
´
FνNK+1 ,ξK+1,K+1 (vK+1 ) , . . . , FνNN ,ξNN (vN )
K h
Y i Y
N h i
fVhn (vn ) fνNn ,ξnn (vn ) ,
n=1 n=K+1
see e.g. Meucci (2005) p. 42. In this expression FνNn ,ξnn is the normal cdf with
parameters provided by (39); fVhn and FVhn follows from (5). In this case we
obtain:
fVhn ≡ (1 − c) fνNn ,ξnn + cfqNn ,Snn . (42)
Finally we represent the posterior in the market coordinates:
¯¡ ¯ 1
¯ −1 ¢ ¡ −1 ¢0 ¯− 2
i OP (c) (m) ≡ ¯ R
fM R ¯ fV h (Rm) . (43)
17
It is very easy to compute the second term in (36). Indeed from (15) we
obtain:
1 ¯¯ ¯¯ N 1 0 −1
i BL (c) (m) = − ln Σc −
ln fM ln (2π) − (m − µc ) Σc (m − µc ) . (44)
2 2 2
To compute the distance in (36) we generate according to (37) a large number
J of Monte Carlo scenarios of the prior market distribution:
M(j) , j = 1, . . . J. (45)
1 X ¯¯ ³ ´ ³ ´¯
J
(j) (j) ¯
ρc (OP, BL) ≈ ¯ln fM
iOP (c) M − ln fM
i BL (c) M ¯. (46)
J j=1
18
In our case the market distribution is discrete, i.e.
J
1 X (M i (j) )
i =
fM δ , (52)
J j=1
where δ (y) is the Dirac delta with point mass in y. Therefore (51) can be
solved by linear programming (LP). Indeed, introducing a J-dimensional vector
of auxiliary variables u, the optimization (51) can be restated equivalently as
follows: ½ ¾
1
w (r) = argmin α+ u0 1 , (53)
{w,u,α}∈C(r) J (1 − γ)
where the set of constraints reads:
n
C (r) : f (j) − α − u(j) ≤ 0
w0 m ≥ r, −w0 M (54)
w ≥ 0, u ≥ 0, w0 1 ≡ 1} ;
and where
J
1 X f (j)
m≡ M . (55)
J j=1
where
⎛ ⎞ ⎛ ⎞
−11×J M/Jf 0 01×J −r
⎜ −M f ⎟
−1J×1 −IJ×J ⎟ , b (r) ≡ ⎜ J×1 ⎜ 0 ⎟
A ≡ ⎜⎝ −I
⎟ (58)
N×N 0N ×1 0N×J ⎠ ⎝ 0N×1 ⎠
0J×N 0J×1 −IJ×J 0J×1
µ ¶
11×J
q0 ≡ 01×N , 1, , c0 ≡ (11×N , 0, 01×J ) ,
J (1 − γ)
19