python3
single: Robustness
- depth
2
In addition to what's in Anaconda, this lecture will need the following libraries:
!pip install --upgrade quantecon
single: Bellman Equation
This lecture modifies a Bellman equation to express a decision-maker's doubts about transition dynamics.
His specification doubts make the decision-maker want a robust decision rule.
Robust means insensitive to misspecification of transition dynamics.
The decision-maker has a single approximating model.
He calls it approximating to acknowledge that he doesn't completely trust it.
He fears that outcomes will actually be determined by another model that he cannot describe explicitly.
All that he knows is that the actual data-generating model is in some (uncountable) set of models that surrounds his approximating model.
He quantifies the discrepancy between his approximating model and the genuine data-generating model by using a quantity called entropy.
(We'll explain what entropy means below)
He wants a decision rule that will work well enough no matter which of those other models actually governs outcomes.
This is what it means for his decision rule to be "robust to misspecification of an approximating model".
This may sound like too much to ask for, but ….
… a secret weapon is available to design robust decision rules.
The secret weapon is max-min control theory.
A value-maximizing decision-maker enlists the aid of an (imaginary) value-minimizing model chooser to construct bounds on the value attained by a given decision rule under different models of the transition dynamics.
The original decision-maker uses those bounds to construct a decision rule with an assured performance level, no matter which model actually governs outcomes.
Note
In reading this lecture, please don't think that our decision-maker is paranoid when he conducts a worst-case analysis. By designing a rule that works well against a worst-case, his intention is to construct a rule that will work well across a set of models.
Let's start with some imports:
import pandas as pd
import numpy as np
from scipy.linalg import eig
import matplotlib.pyplot as plt
%matplotlib inline
import quantecon as qe
Our "robust" decision-maker wants to know how well a given rule will work when he does not know a single transition law ….
… he wants to know sets of values that will be attained by a given decision rule F under a set of transition laws.
Ultimately, he wants to design a decision rule F that shapes these sets of values in ways that he prefers.
With this in mind, consider the following graph, which relates to a particular decision problem to be explained below
The figure shows a value-entropy correspondence for a particular decision rule F.
The shaded set is the graph of the correspondence, which maps entropy to a set of values associated with a set of models that surround the decision-maker's approximating model.
Here
- Value refers to a sum of discounted rewards obtained by applying the decision rule F when the state starts at some fixed initial state x0.
- Entropy is a non-negative number that measures the size of a set of models surrounding the decision-maker's approximating model.
- Entropy is zero when the set includes only the approximating model, indicating that the decision-maker completely trusts the approximating model.
- Entropy is bigger, and the set of surrounding models is bigger, the less the decision-maker trusts the approximating model.
The shaded region indicates that for all models having entropy less than or equal to the number on the horizontal axis, the value obtained will be somewhere within the indicated set of values.
Now let's compare sets of values associated with two different decision rules, Fr and Fb.
In the next figure,
- The red set shows the value-entropy correspondence for decision rule Fr.
- The blue set shows the value-entropy correspondence for decision rule Fb.
The blue correspondence is skinnier than the red correspondence.
This conveys the sense in which the decision rule Fb is more robust than the decision rule Fr
- more robust means that the set of values is less sensitive to increasing misspecification as measured by entropy
Notice that the less robust rule Fr promises higher values for small misspecifications (small entropy).
(But it is more fragile in the sense that it is more sensitive to perturbations of the approximating model)
Below we'll explain in detail how to construct these sets of values for a given F, but for now ….
Here is a hint about the secret weapons we'll use to construct these sets
- We'll use some min problems to construct the lower bounds
- We'll use some max problems to construct the upper bounds
We will also describe how to choose F to shape the sets of values.
This will involve crafting a skinnier set at the cost of a lower level (at least for low values of entropy).
If you want to understand more about why one serious quantitative researcher is interested in this approach, we recommend Lars Peter Hansen's Nobel lecture.
Our discussion in this lecture is based on
HansenSargent2000
HansenSargent2008
For simplicity, we present ideas in the context of a class of problems with linear transition laws and quadratic objective functions.
To fit in with our earlier lecture on LQ control, we will treat loss minimization rather than value maximization.
To begin, recall the infinite horizon LQ problem, where an agent chooses a sequence of controls {ut} to minimize
subject to the linear law of motion
xt + 1 = Axt + But + Cwt + 1, t = 0, 1, 2, …
As before,
- xt is n × 1, A is n × n
- ut is k × 1, B is n × k
- wt is j × 1, C is n × j
- R is n × n and Q is k × k
Here xt is the state, ut is the control, and wt is a shock vector.
For now, we take {wt} := {wt}t = 1∞ to be deterministic --- a single fixed sequence.
We also allow for model uncertainty on the part of the agent solving this optimization problem.
In particular, the agent takes wt = 0 for all t ≥ 0 as a benchmark model but admits the possibility that this model might be wrong.
As a consequence, she also considers a set of alternative models expressed in terms of sequences {wt} that are "close" to the zero sequence.
She seeks a policy that will do well enough for a set of alternative models whose members are pinned down by sequences {wt}.
Soon we'll quantify the quality of a model specification in terms of the maximal size of the expression
If our agent takes {wt} as a given deterministic sequence, then, drawing on intuition from earlier lectures on dynamic programming, we can anticipate Bellman equations such as
Jt − 1(x) = minu{x′Rx + u′Qu + β Jt(Ax + Bu + Cwt)}
(Here J depends on t because the sequence {wt} is not recursive)
Our tool for studying robustness is to construct a rule that works well even if an adverse sequence {wt} occurs.
In our framework, "adverse" means "loss increasing".
As we'll see, this will eventually lead us to construct the Bellman equation
J(x) = minumaxw{x′Rx + u′Qu + β [J(Ax + Bu + Cw) − θw′w]}
Notice that we've added the penalty term − θw′w.
Since w′w = ∥w∥2, this term becomes influential when w moves away from the origin.
The penalty parameter θ controls how much we penalize the maximizing agent for "harming" the minimizing agent.
By raising θ more and more, we more and more limit the ability of maximizing agent to distort outcomes relative to the approximating model.
So bigger θ is implicitly associated with smaller distortion sequences {wt}.
So what does J in rb_wcb0
look like?
As with the ordinary LQ control model, J takes the form J(x) = x′Px for some symmetric positive definite matrix P.
One of our main tasks will be to analyze and compute the matrix P.
Related tasks will be to study associated feedback rules for ut and wt + 1.
First, using matrix calculus, you will be able to verify that
where
𝒟(P) := P + PC(θI − C′PC) − 1C′P
and I is a j × j identity matrix. Substituting this expression for the maximum into rb_wcb0
yields
x′Px = minu{x′Rx + u′Qu + β (Ax + Bu)′𝒟(P)(Ax + Bu)}
Using similar mathematics, the solution to this minimization problem is u = − Fx where F := (Q + βB′𝒟(P)B) − 1βB′𝒟(P)A.
Substituting this minimizer back into rb_owb
and working through the algebra gives x′Px = x′ℬ(𝒟(P))x for all x, or, equivalently,
P = ℬ(𝒟(P))
where 𝒟 is the operator defined in rb_d
and
ℬ(P) := R − β2A′PB(Q + βB′PB) − 1B′PA + βA′PA
The operator ℬ is the standard (i.e., non-robust) LQ Bellman operator, and P = ℬ(P) is the standard matrix Riccati equation coming from the Bellman equation --- see this discussion.
Under some regularity conditions (see HansenSargent2008
), the operator ℬ ∘ 𝒟 has a unique positive definite fixed point, which we denote below by P̂.
A robust policy, indexed by θ, is u = − F̂x where
F̂ := (Q + βB′𝒟(P̂)B) − 1βB′𝒟(P̂)A
We also define
K̂ := (θI − C′P̂C) − 1C′P̂(A − BF̂)
The interpretation of K̂ is that wt + 1 = K̂xt on the worst-case path of {xt}, in the sense that this vector is the maximizer of rb_mp0
evaluated at the fixed rule u = − F̂x.
Note that P̂, F̂, K̂ are all determined by the primitives and θ.
Note also that if θ is very large, then 𝒟 is approximately equal to the identity mapping.
Hence, when θ is large, P̂ and F̂ are approximately equal to their standard LQ values.
Furthermore, when θ is large, K̂ is approximately equal to zero.
Conversely, smaller θ is associated with greater fear of model misspecification and greater concern for robustness.
What we have done above can be interpreted in terms of a two-person zero-sum game in which F̂, K̂ are Nash equilibrium objects.
Agent 1 is our original agent, who seeks to minimize loss in the LQ program while admitting the possibility of misspecification.
Agent 2 is an imaginary malevolent player.
Agent 2's malevolence helps the original agent to compute bounds on his value function across a set of models.
We begin with agent 2's problem.
Agent 2
- knows a fixed policy F specifying the behavior of agent 1, in the sense that ut = − Fxt for all t
- responds by choosing a shock sequence {wt} from a set of paths sufficiently close to the benchmark sequence {0, 0, 0, …}
A natural way to say "sufficiently close to the zero sequence" is to restrict the summed inner product
However, to obtain a time-invariant recursive formulation, it turns out to be convenient to restrict a discounted inner product
Now let F be a fixed policy, and let JF(x0, w) be the present-value cost of that policy given sequence w := {wt} and initial condition x0 ∈ ℝn.
Substituting − Fxt for ut in rob_sih
, this value can be written as
where
xt + 1 = (A − BF)xt + Cwt + 1
and the initial condition x0 is as specified in the left side of rob_fpv
.
Agent 2 chooses w to maximize agent 1's loss JF(x0, w) subject to rb_dec
.
Using a Lagrangian formulation, we can express this problem as
where {xt} satisfied rob_lomf
and θ is a Lagrange multiplier on constraint rb_dec
.
For the moment, let's take θ as fixed, allowing us to drop the constant βθη term in the objective function, and hence write the problem as
or, equivalently,
subject to rob_lomf
.
What's striking about this optimization problem is that it is once again an LQ discounted dynamic programming problem, with w = {wt} as the sequence of controls.
The expression for the optimal policy can be found by applying the usual LQ formula (see here).
We denote it by K(F, θ), with the interpretation wt + 1 = K(F, θ)xt.
The remaining step for agent 2's problem is to set θ to enforce the constraint rb_dec
, which can be done by choosing θ = θη such that
Here xt is given by rob_lomf
--- which in this case becomes xt + 1 = (A − BF + CK(F, θ))xt.
Define the minimized object on the right side of problem rb_a2o
as Rθ(x0, F).
Because "minimizers minimize" we have
where xt + 1 = (A − BF + CK(F, θ))xt and x0 is a given initial condition.
This inequality in turn implies the inequality
where
The left side of inequality rob_bound
is a straight line with slope − θ.
Technically, it is a "separating hyperplane".
At a particular value of entropy, the line is tangent to the lower bound of values as a function of entropy.
In particular, the lower bound on the left side of rob_bound
is attained when
To construct the lower bound on the set of values associated with all perturbations w satisfying the entropy constraint rb_dec
at a given entropy level, we proceed as follows:
- For a given θ, solve the minimization problem
rb_a2o
. - Compute the minimizer Rθ(x0, F) and the associated entropy using
rb_pdt22
. - Compute the lower bound on the value function
$R_\theta(x_0, F) - \theta \ {\rm ent}$ and plot it against${\rm ent}$ . - Repeat the preceding three steps for a range of values of θ to trace out the lower bound.
Note
This procedure sweeps out a set of separating hyperplanes indexed by different values for the Lagrange multiplier θ.
To construct an upper bound we use a very similar procedure.
We simply replace the minimization problem rb_a2o
with the maximization problem
where now θ̃ > 0 penalizes the choice of w with larger entropy.
(Notice that θ̃ = − θ in problem rb_a2o
)
Because "maximizers maximize" we have
which in turn implies the inequality
where
The left side of inequality robboundmax
is a straight line with slope θ̃.
The upper bound on the left side of robboundmax
is attained when
To construct the upper bound on the set of values associated all perturbations w with a given entropy we proceed much as we did for the lower bound
- For a given θ̃, solve the maximization problem
rba2omax
. - Compute the maximizer Vθ̃(x0, F) and the associated entropy using
rbpdt223
. - Compute the upper bound on the value function
$V_{\tilde \theta}(x_0, F) + \tilde \theta \ {\rm ent}$ and plot it against${\rm ent}$ . - Repeat the preceding three steps for a range of values of θ̃ to trace out the upper bound.
Now in the interest of reshaping these sets of values by choosing F, we turn to agent 1's problem.
Now we turn to agent 1, who solves
where {wt + 1} satisfies wt + 1 = Kxt.
In other words, agent 1 minimizes
subject to
xt + 1 = (A + CK)xt + But
Once again, the expression for the optimal policy can be found here --- we denote it by F̃.
Clearly, the F̃ we have obtained depends on K, which, in agent 2's problem, depended on an initial policy F.
Holding all other parameters fixed, we can represent this relationship as a mapping Φ, where
F̃ = Φ(K(F, θ))
The map F ↦ Φ(K(F, θ)) corresponds to a situation in which
- agent 1 uses an arbitrary initial policy F
- agent 2 best responds to agent 1 by choosing K(F, θ)
- agent 1 best responds to agent 2 by choosing F̃ = Φ(K(F, θ))
As you may have already guessed, the robust policy F̂ defined in rb_oc_ih
is a fixed point of the mapping Φ.
In particular, for any given θ,
- K(F̂, θ) = K̂, where K̂ is as given in
rb_kd
- Φ(K̂) = F̂
A sketch of the proof is given in the appendix <rb_appendix>
.
Now we turn to the stochastic case, where the sequence {wt} is treated as an IID sequence of random vectors.
In this setting, we suppose that our agent is uncertain about the conditional probability distribution of wt + 1.
The agent takes the standard normal distribution N(0, I) as the baseline conditional distribution, while admitting the possibility that other "nearby" distributions prevail.
These alternative conditional distributions of wt + 1 might depend nonlinearly on the history xs, s ≤ t.
To implement this idea, we need a notion of what it means for one distribution to be near another one.
Here we adopt a very useful measure of closeness for distributions known as the relative entropy, or Kullback-Leibler divergence.
For densities p, q, the Kullback-Leibler divergence of q from p is defined as
Using this notation, we replace rb_wcb0
with the stochastic analog
J(x) = minumaxψ ∈ 𝒫{x′Rx+u′Qu+β [∫J(Ax+Bu+Cw) ψ(dw)−θDKL(ψ,ϕ)]}
Here 𝒫 represents the set of all densities on ℝn and ϕ is the benchmark distribution N(0, I).
The distribution ϕ is chosen as the least desirable conditional distribution in terms of next period outcomes, while taking into account the penalty term θDKL(ψ, ϕ).
This penalty term plays a role analogous to the one played by the deterministic penalty θw′w in rb_wcb0
, since it discourages large deviations from the benchmark.
The maximization problem in rb_wcb1
appears highly nontrivial --- after all, we are maximizing over an infinite dimensional space consisting of the entire set of densities.
However, it turns out that the solution is tractable, and in fact also falls within the class of normal distributions.
First, we note that J has the form J(x) = x′Px + d for some positive definite matrix P and constant real number d.
Moreover, it turns out that if (I − θ − 1C′PC) − 1 is nonsingular, then
where
κ(θ, P) := θln [det (I − θ − 1C′PC) − 1]
and the maximizer is the Gaussian distribution
ψ = N((θI−C′PC) − 1C′P(Ax+Bu),(I−θ − 1C′PC) − 1)
Substituting the expression for the maximum into Bellman equation rb_wcb1
and using J(x) = x′Px + d gives
x′Px + d = minu{x′Rx+u′Qu+β (Ax+Bu)′𝒟(P)(Ax+Bu)+β [d+κ(θ,P)]}
Since constant terms do not affect minimizers, the solution is the same as rb_owb
, leading to
x′Px + d = x′ℬ(𝒟(P))x + β [d + κ(θ, P)]
To solve this Bellman equation, we take P̂ to be the positive definite fixed point of ℬ ∘ 𝒟.
In addition, we take d̂ as the real number solving d = β [d + κ(θ, P)], which is
The robust policy in this stochastic case is the minimizer in rb_wcb2
, which is once again u = − F̂x for F̂ given by rb_oc_ih
.
Substituting the robust policy into rb_md
we obtain the worst-case shock distribution:
wt + 1 ∼ N(K̂xt, (I − θ − 1C′P̂C) − 1)
where K̂ is given by rb_kd
.
Note that the mean of the worst-case shock distribution is equal to the same worst-case wt + 1 as in the earlier deterministic setting.
Before turning to implementation, we briefly outline how to compute several other quantities of interest.
One thing we will be interested in doing is holding a policy fixed and computing the discounted loss associated with that policy.
So let F be a given policy and let JF(x) be the associated loss, which, by analogy with rb_wcb1
, satisfies
JF(x) = maxψ ∈ 𝒫{x′(R+F′QF)x+β [∫JF((A−BF)x+Cw) ψ(dw)−θDKL(ψ,ϕ)]}
Writing JF(x) = x′PFx + dF and applying the same argument used to derive rb_mls
we get
x′PFx + dF = x′(R + F′QF)x + β [x′(A−BF)′𝒟(PF)(A−BF)x+dF+κ(θ,PF)]
To solve this we take PF to be the fixed point
PF = R + F′QF + β(A − BF)′𝒟(PF)(A − BF)
and
If you skip ahead to the appendix <rb_appendix>
, you will be able to verify that − PF is the solution to the Bellman equation in agent 2's problem discussed above <rb_a2>
--- we use this in our computations.
The QuantEcon.py package provides a class called RBLQ
for implementation of robust LQ optimal control.
The code can be found on GitHub.
Here is a brief description of the methods of the class
d_operator()
andb_operator()
implement 𝒟 and ℬ respectivelyrobust_rule()
androbust_rule_simple()
both solve for the triple F̂, K̂, P̂, as described in equationsrb_oc_ih
--rb_kd
and the surrounding discussionrobust_rule()
is more efficientrobust_rule_simple()
is more transparent and easier to follow
K_to_F()
andF_to_K()
solve the decision problems ofagent 1 <rb_a1>
andagent 2 <rb_a2>
respectivelycompute_deterministic_entropy()
computes the left-hand side ofrb_pdt
evaluate_F()
computes the loss and entropy associated with a given policy --- seethis discussion <rb_coq>
Let us consider a monopolist similar to this one, but now facing model uncertainty.
The inverse demand function is pt = a0 − a1yt + dt.
where
and all parameters are strictly positive.
The period return function for the monopolist is
Its objective is to maximize expected discounted profits, or, equivalently, to minimize
To form a linear regulator problem, we take the state and control to be
Setting b := (a0 − c)/2 we define
For the transition matrices, we set
Our aim is to compute the value-entropy correspondences shown above <rb_vec>
.
The parameters are
a0 = 100, a1 = 0.5, ρ = 0.9, σd = 0.05, β = 0.95, c = 2, γ = 50.0
The standard normal distribution for wt is understood as the agent's baseline, with uncertainty parameterized by θ.
We compute value-entropy correspondences for two policies
- The no concern for robustness policy F0, which is the ordinary LQ loss minimizer.
- A "moderate" concern for robustness policy Fb, with θ = 0.02.
The code for producing the graph shown above, with blue being for the robust policy, is as follows
# Model parameters
a_0 = 100
a_1 = 0.5
ρ = 0.9
σ_d = 0.05
β = 0.95
c = 2
γ = 50.0
θ = 0.002
ac = (a_0 - c) / 2.0
# Define LQ matrices
R = np.array([[0., ac, 0.],
[ac, -a_1, 0.5],
[0., 0.5, 0.]])
R = -R # For minimization
Q = γ / 2
A = np.array([[1., 0., 0.],
[0., 1., 0.],
[0., 0., ρ]])
B = np.array([[0.],
[1.],
[0.]])
C = np.array([[0.],
[0.],
[σ_d]])
# ----------------------------------------------------------------------- #
# Functions
# ----------------------------------------------------------------------- #
def evaluate_policy(θ, F):
"""
Given θ (scalar, dtype=float) and policy F (array_like), returns the
value associated with that policy under the worst case path for {w_t},
as well as the entropy level.
"""
rlq = qe.robustlq.RBLQ(Q, R, A, B, C, β, θ)
K_F, P_F, d_F, O_F, o_F = rlq.evaluate_F(F)
x0 = np.array([[1.], [0.], [0.]])
value = - x0.T @ P_F @ x0 - d_F
entropy = x0.T @ O_F @ x0 + o_F
return list(map(float, (value, entropy)))
def value_and_entropy(emax, F, bw, grid_size=1000):
"""
Compute the value function and entropy levels for a θ path
increasing until it reaches the specified target entropy value.
Parameters
==========
emax: scalar
The target entropy value
F: array_like
The policy function to be evaluated
bw: str
A string specifying whether the implied shock path follows best
or worst assumptions. The only acceptable values are 'best' and
'worst'.
Returns
=======
df: pd.DataFrame
A pandas DataFrame containing the value function and entropy
values up to the emax parameter. The columns are 'value' and
'entropy'.
"""
if bw == 'worst':
θs = 1 / np.linspace(1e-8, 1000, grid_size)
else:
θs = -1 / np.linspace(1e-8, 1000, grid_size)
df = pd.DataFrame(index=θs, columns=('value', 'entropy'))
for θ in θs:
df.loc[θ] = evaluate_policy(θ, F)
if df.loc[θ, 'entropy'] >= emax:
break
df = df.dropna(how='any')
return df
# ------------------------------------------------------------------------ #
# Main
# ------------------------------------------------------------------------ #
# Compute the optimal rule
optimal_lq = qe.lqcontrol.LQ(Q, R, A, B, C, beta=β)
Po, Fo, do = optimal_lq.stationary_values()
# Compute a robust rule given θ
baseline_robust = qe.robustlq.RBLQ(Q, R, A, B, C, β, θ)
Fb, Kb, Pb = baseline_robust.robust_rule()
# Check the positive definiteness of worst-case covariance matrix to
# ensure that θ exceeds the breakdown point
test_matrix = np.identity(Pb.shape[0]) - (C.T @ Pb @ C) / θ
eigenvals, eigenvecs = eig(test_matrix)
assert (eigenvals >= 0).all(), 'θ below breakdown point.'
emax = 1.6e6
optimal_best_case = value_and_entropy(emax, Fo, 'best')
robust_best_case = value_and_entropy(emax, Fb, 'best')
optimal_worst_case = value_and_entropy(emax, Fo, 'worst')
robust_worst_case = value_and_entropy(emax, Fb, 'worst')
fig, ax = plt.subplots()
ax.set_xlim(0, emax)
ax.set_ylabel("Value")
ax.set_xlabel("Entropy")
ax.grid()
for axis in 'x', 'y':
plt.ticklabel_format(style='sci', axis=axis, scilimits=(0, 0))
plot_args = {'lw': 2, 'alpha': 0.7}
colors = 'r', 'b'
df_pairs = ((optimal_best_case, optimal_worst_case),
(robust_best_case, robust_worst_case))
class Curve:
def __init__(self, x, y):
self.x, self.y = x, y
def __call__(self, z):
return np.interp(z, self.x, self.y)
for c, df_pair in zip(colors, df_pairs):
curves = []
for df in df_pair:
# Plot curves
x, y = df['entropy'], df['value']
x, y = (np.asarray(a, dtype='float') for a in (x, y))
egrid = np.linspace(0, emax, 100)
curve = Curve(x, y)
print(ax.plot(egrid, curve(egrid), color=c, **plot_args))
curves.append(curve)
# Color fill between curves
ax.fill_between(egrid,
curves[0](egrid),
curves[1](egrid),
color=c, alpha=0.1)
plt.show()
Here's another such figure, with θ = 0.002 instead of 0.02
Can you explain the different shape of the value-entropy correspondence for the robust policy?
We sketch the proof only of the first claim in this section <rb_eq>
, which is that, for any given θ, K(F̂, θ) = K̂, where K̂ is as given in rb_kd
.
This is the content of the next lemma.
Lemma. If P̂ is the fixed point of the map ℬ ∘ 𝒟 and F̂ is the robust policy as given in rb_oc_ih
, then
K(F̂, θ) = (θI − C′P̂C) − 1C′P̂(A − BF̂)
Proof: As a first step, observe that when F = F̂, the Bellman equation associated with the LQ problem rob_lomf
-- rb_a2o
is
P̃ = − R − F̂′QF̂ − β2(A − BF̂)′P̃C(βθI + βC′P̃C) − 1C′P̃(A − BF̂) + β(A − BF̂)′P̃(A − BF̂)
(revisit this discussion if you don't know where rb_a2be
comes from) and the optimal policy is
wt + 1 = − β(βθI + βC′P̃C) − 1C′P̃(A − BF̂)xt
Suppose for a moment that − P̂ solves the Bellman equation rb_a2be
.
In this case, the policy becomes
wt + 1 = (θI − C′P̂C) − 1C′P̂(A − BF̂)xt
which is exactly the claim in rb_kft
.
Hence it remains only to show that − P̂ solves rb_a2be
, or, in other words,
P̂ = R + F̂′QF̂ + β(A − BF̂)′P̂C(θI − C′P̂C) − 1C′P̂(A − BF̂) + β(A − BF̂)′P̂(A − BF̂)
Using the definition of 𝒟, we can rewrite the right-hand side more simply as
R + F̂′QF̂ + β(A − BF̂)′𝒟(P̂)(A − BF̂)
Although it involves a substantial amount of algebra, it can be shown that the latter is just P̂.
(Hint: Use the fact that P̂ = ℬ(𝒟(P̂)))