Abstract
We consider a consumption-investment problem on infinite time horizon maximizing discounted expected HARA utility for a general incomplete market model. Based on dynamic programming approach we derive the relevant H–J–B equation and study the existence and uniqueness of the solution to the nonlinear partial differential equation. By using the smooth solution we construct the optimal consumption rate and portfolio strategy and then prove the verification theorems under certain general settings.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
In this paper we consider an optimal consumption-investment problem on infinite time horizon for a general incomplete market model. The market model considered here consists of \(m+1\) securities, one of which is a risk-less asset and the other \(m\) assets are risky ones. The price of risk-less asset is governed by an ordinary differential equation, while the prices of risky ones are defined by the stochastic differential equations. We suppose that all coefficients appearing in those dynamics of the asset prices are affected by the economic factors. The dynamics of the economic factors are also formulated by the stochastic differential equations (cf. (2.1)–(2.3)). In such a general market model, the investor divides his (her) wealth among those \(m+1\) securities and decides the rate for consumption. The goal is to select consumption and investment strategies which maximize the total expected (discounted) power utility of consumption on a long run (cf. (2.5)).
Our approach is based on the dynamic programming principle, in which the H–J–B (Hamilton–Jacobi–Bellman) equation (cf. (2.10) and (2.11)) are derived relevant to the consumption-investment problem. One can construct an optimal consumption-investment strategy from the smooth solution to the equation. Indeed, the optimal consumption rate and optimal portfolio strategy can be explicitly expressed in terms of the smooth solution. In this approach, a pioneering work by Merton for the market with the risky asset price having constant volatility and return has been done in [14] and recent progress of the further studies in this direction is seen in [4–8] etc.. In [4–6] one dimensional H–J–B equations are considered, where the ordinary differential equations concern. On the other hand, in [7, 8], they prove existence of the smooth solution to the H–J–B equation in general dimension by employing a modification of the Leray-Schauder fixed point theorem. Then, constructing the optimal consumption-investment strategy by using the solution, the verification theorems are shown under certain incomplete market settings. In their proofs they have also the best possible discount factor. The current paper is motivated by these works [7, 8]. Following their methods we prove existence of the solution to the H–J–B equation (cf. (2.11)) by giving sub- and super- solutions under the general settings, while we obtain newly the uniqueness theorems on the solution, which is one of our main concern. In the current paper the definition of the set of admissible strategies is given by using the unique solution of the H–J–B equation and thus our uniqueness theorems have a crucial meaning. As a result our set of admissible strategies generalizes the one defined in [8] and the proofs of the verification theorems have become quite natural and simple in the current paper. Indeed, by introducing the new measure \({\overline{P}}^{\hat{h}}\) defined by (5.3) from the unique solution \(z(x)\) of H–J–B equation (2.11), the relevant criterion function to the optimal strategy \(({\hat{c}}_s,{\hat{h}}_s)\) turns out to be described as
by using a multiplicative functional \(\varphi _t=e^{-\int _0^te^{-\frac{z(X_s)}{1-\gamma }}ds}=e^{-\int _0^t{\hat{c}}_s^{\gamma }e^{-z(X_s)}ds}\) as you see in (5.2) and (5.5). Thus, identification of the value function with the solution to the H–J–B equation can be done by showing \(\varphi _{\infty }=0, \) a.s. Comparing the value function with the criterion for strategy \((c_s,h_s)\) is also seen by looking at \(e^{z(x)}{\overline{E}}^{ h}[-\int _0^{\infty }d{\tilde{\varphi }}_t]\) with \({\tilde{\varphi }}_t=e^{-\int _0^tc_s^{\gamma }e^{-z(X_s)}ds}\) (cf. Proofs of Theorem 5.1 and 5.2).
There are slight difference between the market models discussed in [7, 8] and ours although the both are factor models. They treat a linear Gaussian model and another factor model with the boundedness assumptions, where all coefficients appearing in the dynamics for the asset prices and the factor process are bounded. In the current paper, the general factor model including linear Gaussian models is discussed without boundedness assumptions on the returns of the price process and the drift coefficient of the factor process. Such difference may cause certain technical difference to treat.
We mention some other works with different approach from ours. Approach using the martingale methods for a complete market model appears in [2, 11, 19]. Analytical solutions are given in [3, 12, 20]. More attentive introduction to the historical works can be seen in [7].
The paper is organized as follows. Derivation of the H–J–B equation and our assumptions are described in Sect. 2 under the setting of the factor model. We devote Sect. 3 to construction of sub- and super- solutions to the H–J–B equation. In Sect. 4, we present the existence and uniqueness theorems for the H–J–B equation and the proofs of uniqueness are given. The proofs of existence of the solution following the methods due to Hata and Sheu [7] is completed in Appendix 1. We give the proofs of the verification theorems in Sect. 5. Some notes on the useful gradient estimates for the proofs are given in Appendix 2.
2 Derivation of H–J–B Equations and Assumptions
Consider a market model with \(m+1\) securities and \(n\) factors, where the bond price is governed by ordinary differential equation
The other security prices and factors are assumed to satisfy stochastic differential equations
and
where \(W_t=(W_t^k)_{k=1,\ldots ,(n+m)}\) is an \(m+n\)-dimensional standard Brownian motion process on a probability space \((\Omega ,{\mathcal F},P)\). Let \(N_t^i\) be the number of the shares of \(i\)th security. Then the total wealth that the investor possesses is defined as
the portfolio proportion invested to \(i\)th security as
We assume that self-financing condition holds for a consumption investment strategy \((h_t,C_t)\):
where \(C_t\) is the instantaneous nominal consumption. Setting \(C_t=c_tV_t\), we have
Thus the equation describing the dynamics of wealth \(V_t=V(c,h)_t\) is given by
Here \(h^*\) stands for the transposed vector of \(h\) and \({\mathbf 1}=(1,1,\ldots ,1)^*\). As for the filtration to be satisfied by admissible investment strategies,
is relevant in the present problem and we introduce the following definition.
Definition 2.1
\(h(t)_{0\le t\le T}\) is said an investment strategy if \(h(t)\) is an \(R^m\) valued \({\mathcal G}_t\)- progressively measurable stochastic process such that
The set of all investment strategies is denoted by \({\mathcal H}(T)\). For a given \(h\in {\mathcal H}(T)\), the process \(V_t=V_t(h)\) representing the total wealth of the investor at time \(t\) is determined by the stochastic differential equation as was seen above. For \(\rho \ge 0\), let us consider the following problem
The following equations are equivalent to (2.5)
in each case of \(0<\gamma <1\) and \(\gamma <0\), respectively. It is easy to see that
where, \(v_0\) is the initial wealth and
Therefore
Now let us assume that \((\Omega ,{\mathcal F})\) be a standard measurable space (cf. [18]). Then, if \(M^h_t\) defined by \(M^h_t=\gamma \int _0^th_s^*\sigma (X_s)dW_s\) satisfies
then there is a probability measure \(P^h\) satisfying
Then we have
and thus
and
Note that, under the probability measure \(P^h\),
is a Brownian motion process and the stochastic differential equation for the economic factor \(X_t\) is written as
When setting
the H–J–B equation for \(v(x)\) turns out to be
By taking a transformation \(z(x)=\log v(x)\), we obtain
which can be written as
where
On the other hand, for \(1>\gamma >0\), we set
Then, the H–J–B equation of \({\tilde{v}}(x)\) is seen to be
and, by taking a transformation \({\tilde{z}}(x)=\log {\tilde{v}}(x)\), we obtain
which turns out to be the same equation as (2.11).
When \(\gamma <0\), we assume that
and that, in the case of \(\rho =0\),
On the other hand, when \(0<\gamma <1\), we assume (2.13)–(2.15),
and the following condition
Note that
hold under these assumptions.
In considering (2.6) and (2.7) we formulate the set of strategies defined by
Then, we confine the sets of admissible strategies in each case of \(0<\gamma <1\) and \(\gamma <0\) as follows. Set
where \(z(x)\) is the unique solution to H–J–B equation (2.11) (see Theorem 4.1 and Theorem 4.2 ), and consider strategies satisfying
For such \(h_s\) we have a probability measure \({\overline{P}}^h\) on \((\Omega ,{\mathcal F})\) such that
The set \({\mathcal A}_1\) of admissible strategies is defined by
3 Sub- and Super-Solution
3.1 Risk-Averse Case (\(\gamma <0\))
We first note that there exists a positive constants \(c_0\) and \(c\) such that
under assumption (2.16). We consider the following stochastic differential equation
and set
and
Then, we have the following lemma.
Lemma 3.1
Assume assumptions (2.13)–(2.16). Then, \({\underline{z}}(x)\) (respectively \({\overline{z}}(x)\)) is a sub-solution (resp. super-solution) to (2.11) for \(\rho =0\). Further
Proof
Set
Then, it satisfies
and thus \({\underline{z}}\) does
Since \(\frac{1}{1-\gamma }I\le N_{\gamma }^{-1}\), \({\underline{z}}\) turns out to be a sub-solution to (2.11).
On the other hand, set
Then, it satisfies
Therefore, \({\overline{z}}\) satisfies
The left hand side is obtained when taking in the right hand of (2.12) \(h=\frac{1}{1-\gamma }(\sigma \sigma ^*)^{-1}{\hat{\alpha }}(x)\) and the constant \(c>0\) satisfying (3.1), and thus we see that \({\overline{z}}\) is a super-solution to (2.11) with \(\rho =0\).
It is a direct consequence from the following lemma that \({\underline{z}}(x)\le {\overline{z}}(x)\). \(\square \)
Lemma 3.2
The following inequality holds
Proof
Set
Then, we have
Hence, we obtain the present lemma. \(\square \)
Remark
When \(\rho >0\), we do not need to assume (2.16) because there exist positive constants \(c\) and \(c_0\) such that
and thus, considering \({\tilde{U}}_{\gamma }:=\frac{\gamma }{2(1-\gamma )}{\hat{\alpha }}^*(\sigma \sigma ^*)^{-1}{\hat{\alpha }}+\gamma r-\rho \) in place of \(U_{\gamma }\) the parallel arguments to the above apply.
3.2 Risk Seeking Case (\(0<\gamma <1\))
It is not easy to construct a super-solution to (2.11) in risk seeking case, \(0<\gamma <1\). Let us start with considering the H–J–B equation of risk-sensitive portfolio optimization without consumption:
To study the existence and uniqueness of the solution to (3.4), we introduce the discounted type equation:
We first note that (3.5) can be written as
where
and
We have the following lemma.
Lemma 3.3
Under assumptions (2.13)–(2.15) and (2.17), (3.5) has a solution \(z_{\epsilon }\in C^2(R^n)\) such that \(z_{\epsilon }-z_0\) is bounded above.
Proof
In light of assumption (2.17), we can assume that \(z_0\ge 0\) and also take \(R_0\) such that, for \(R\ge R_0\),
Set
Then, \(\Phi _{\epsilon }(x)\) turns out to be a super-solution to (3.5). In proving the existence of the solution to (3.5), we first consider the Dirichlet problem for \(R>R_0\):
Owing to Theorem 8.3 [13, Chapter 4], Dirichlet problem (3.7) has a solution \(z_{\epsilon ,R}\in C^{2,\mu }(R^n)\). We extend \(z_{\epsilon ,R}\) to the whole Euclidean space as
Then, we can see that \({\tilde{z}}_{\epsilon ,R}(x)\) is non-increasing with respect to \(R\). Indeed, for \(R<R'\),
Therefore, from the maximum principle, we see that
since \({\tilde{z}}_{\epsilon ,R}(x)={\tilde{z}}_{\epsilon ,R'}(x),\; x\in \partial B_{R'}\). We further note that \({\tilde{z}}_{\epsilon ,R}(x)\ge 0\) for each \(R\) because \(z_1(x)\equiv 0\) is a sub-solution to (3.5) for \(\gamma >0\) and the maximum principle again applies.
Similarly to the proof of Proposition 3.2 in [17], we have the following gradient estimate: for each \(R\), \(r<\frac{R}{2}\), and \(x\in B_r\),
where \(C\) is a positive constant independent of \(R\) and \(\epsilon \), \(|f|_r=|f|_{L^{\infty }(B_r(x))}\), and \(\Psi =\lambda N_{\gamma }^{-1}\lambda ^*\). Thus, by similar arguments to the proof of Lemma 2.8 in [9], we can see that \({\tilde{z}}_{\epsilon ,R}(x)\) converges \(H^1_{loc}\) strongly and uniformly on each compact set to the solution \(z_{\epsilon }\in C^2(R^n)\). Since \({\tilde{z}}_{\epsilon ,R}(x)\le \Phi _{\epsilon }(x)\) for each \(R>R_0\) we see that \(z_{\epsilon }\le \Phi _{\epsilon }\), and hence, \( z_{\epsilon }-z_0\) is bounded above. \(\square \)
Lemma 3.4
Assume assumptions (2.13)–(2.15) and (2.17). Then, the solution \(z_{\epsilon }\) to (3.5) such that \( z_{\epsilon }-z_0\) is bounded above satisfies
Proof
Let \(z_{\epsilon }\) be a solution to (3.5) such that \(z_{\epsilon }-z_0\) is bounded above and set
Then, we have
where
Then, similarly to the proof of Lemma 2.1 in [16], we can see that \(z_{\epsilon }-z_0\rightarrow -\infty \) as \(|x|\rightarrow \infty \) since \(V(x)+\epsilon z_{\epsilon } \rightarrow \infty ,\;\;|x|\rightarrow \infty \). \(\square \)
Lemma 3.5
Under assumptions (2.13)–(2.15) and(2.17), (3.4) has a solution \(({\hat{\chi }}, {\hat{z}})\) such that \({\hat{z}}-z_0\) is bounded above and \({\hat{z}}\in C^2(R^n)\). Moreover, the solution \((\chi ,z)\) such that \( z-z_0\) is bounded above is unique, when admitting ambiguity of additive constants with respect to \(z\).
Proof
Let us first note that the same estimate as (3.8) holds for \(z_{\epsilon }\) for each \(\epsilon >0\). Owing to assumptions (2.13) - (2.15), estimate (3.8) implies that
where \(C\) is a positive constant independent of \(\epsilon \). According to Lemma 3.4, \(z_{\epsilon }(x)-z_0(x) \rightarrow -\infty \) as \(|x|\rightarrow \infty \). Therefore, we can take \(x_{\epsilon }\) such that
Then, at \(x_{\epsilon }\)
Therefore, from (3.9), we have
and thus, \(V(x_{\epsilon })\le 0\). Since \(V(x)\rightarrow \infty \) as \(|x|\rightarrow \infty \), there exists a compact set \(K\) such that \(x_{\epsilon }\in K\) for each \(\epsilon >0\). Therefore, we can take a subsequence \(\{ x_{\epsilon _n}\} \subset \{x_{\epsilon }\}\) and \({\hat{x}}\) such that \(x_{\epsilon _n}\rightarrow {\hat{x}}\), \(n\rightarrow \infty \). Once more again from (3.9) we see that
Thus, by taking a subsequence if necessary, \(\epsilon _nz_{\epsilon _n}(x_{\epsilon _n})\rightarrow {\hat{\chi }}\), \(n\rightarrow \infty \). On the other hand, by using (3.10) we can see that \(\{z_{\epsilon _n}(x)-z_{\epsilon _n}({\hat{x}})\}\) forms a sequence of uniformly bounded and equicontinuous functions on each compact set \(K' \) including \(K\). Thus, similarly to the proof of Theorem 3.1 in [9], we can see that it converges to \({\hat{z}}(x)\in C^2(R^n)\) in \(H^1_{loc}\) strongly and uniformly on each compact set, by using estimate (3.8) for \(z_{\epsilon }\), and that \(({\hat{\chi }}, {\hat{z}}(x))\) satisfies (3.4). Further, we can see that \(\epsilon _nz_{\epsilon _n}({\hat{x}})\rightarrow {\hat{\chi }}\). Note that
and the left hand side converges to \({\hat{z}}(x)-z_0(x)\). Therefore we see that \({\hat{z}}(x)-z_0(x)\) is bounded above by the constant \(-z_0({\hat{x}})\) which is the limit of the right hand side.
Further, we can see that \({\hat{z}}(x)-z_0(x)\rightarrow -\infty \) as \(|x|\rightarrow \infty \) similarly to the proof of Lemma 3.4. Therefore, similarly to the proof of Lemma 3.2 in [16], we see that the solution to (3.4) such that \({\hat{z}}(x)-z_0(x)\) is bounded above is unique when admitting additive constant with respect to \(z\). \(\square \)
Let us define the operator by
Then,we have the following lemma.
Lemma 3.6
Under the assumptions of Lemma 3.3, the diffusion process with the generator \({\hat{L}}\) is ergodic.
Proof
We have shown that \(z_0(x)-{\hat{z}}(x)\rightarrow \infty \) as \(|x|\rightarrow \infty \) in the proof of Lemma 3.5. We moreover see that
Since \(-V-{\hat{\chi }}\rightarrow -\infty \) as \(|x|\rightarrow \infty \), we see that the Has’minskii’s conditions are satisfied and that \({\hat{L}}\) is ergodic. \(\square \)
Lemma 3.7
Besides the assumptions of Lemma 3.3, we assume (2.17’). Then, \({\hat{z}}(x)\) is bounded below.
Proof
Note that
Then, from (3.6), it follows that
Set
Then, \(|\epsilon {\bar{z}}_{\epsilon }|<M\) for some positive constant \(M\) and we can take \(R>0\) such that
under assumption (2.17’). Therefore, by using Itô’s formula, we have
since \(z_{\epsilon }(x)\ge 0\), where \((Y_t,P_x)\) is the diffusion process with the generator \(L\) and
Therefore, as \(T\rightarrow \infty \) we see that
Since \({\tilde{z}}_{\epsilon }\) converges to \({\hat{z}}\) uniformly on each compact set we have \(\inf _{|x|=R}{\tilde{z}}_{\epsilon }(x)\ge -K,\;\; K>0\) and hence obtain \({\hat{z}}(x)\ge -K\). \(\square \)
Now, we consider H–J–B equation (2.11) for \(0<\gamma <1\).
Proposition 3.1
Assume assumptions (2.13)–(2.15), 2.17’ and (2.17). Then, when taking \(C\) to be sufficiently large, \({\overline{z}}(x)={\hat{z}}(x)+ C\) is a super-solution to (2.11) with \(\rho >{\hat{\chi }}\). Moreover, \({\underline{z}}(x)\equiv -C'\) is a sub-solution to (2.11) if \(C'>0\) is sufficiently large.
Proof
Take \(\epsilon >0\) such that \(\rho -{\hat{\chi }}>\epsilon \). Then, we can see that
by taking \(C\) to be sufficiently large because of Lemma 3.7. Since \(({\hat{\chi }},{\hat{z}})\) is a solution to (3.4), \({\overline{z}}={\hat{z}}(x)+ C\) turns out to be a super-solution to (2.11).
It is easy to see that \({\underline{z}}(x)\) is a sub-solution to (2.11). \(\square \)
4 Existence and Uniqueness
We first prepare the following lemma.
Lemma 4.1
Assume assumptions (2.13)–(2.16) and \(\gamma <0\). Then, the bounded above solution to H–J–B equation (2.11) is unique.
Proof
Note that each smooth solution \(z\) to (2.11) satisfies the estimate
for a positive constant \(C>0\) under our assumptions (cf. Appendix 2). Let \(z\) and \(z_1\) be bounded above solutions to (2.11) and set
Then, we have
since \(\frac{1}{1-\gamma }I\le N_{\gamma }^{-1}\) for \(\gamma <0\).
Let \(Y_t\) be a solution of the stochastic differential equation
and set
Then, from (4.1) it follows that
Setting \(t=\tau _G\wedge T\), where
we have
Note that
where \(K\) is a positive constant such that \(e^{\frac{1}{1-\gamma }z_1(x)}\le K\). Thus, we see that
Sending \(T\) to \(\infty \) we have \(\phi (x)-1\ge 0\) and so \(z(x)\ge z_1(x)\).
Exchanging a role of \(z_1\) by \(z\), we obtain converse inequality \(z_1(x)\ge z(x)\). \(\square \)
Let \( {\underline{z}}(x)\) and \({\overline{z}}(x)\) be, respectively sub- and super- solution to (2.11) with \(\rho =0\) obtained in Lemma 3.1. Then, we have the following theorem.
Theorem 4.1
For \(\gamma <0\), we assume assumptions (2.13)–(2.16). Then, for \(\rho = 0\), (2.11) has a solution \(z\) such that \({\underline{z}}(x)\le z(x)\le {\overline{z}}(x)\). Moreover, the bounded above solution to (2.11) is unique.
Proof
Since we have a sub- and a super- solution as was seen in Lemma 3.1, the existence of a solution can be shown in a similar manner to the proof of Theorem 3.5 in [7]. We complete the proof in Appendix 1.
Let us prove uniqueness. Let \(z(x)\) be the solutions to (2.11) with \(\rho =0\) such that \({\underline{z}}(x)\le z(x)\le {\overline{z}}(x)\). Under our assumptions we can see that
holds and so \(z(x)\) is bounded above. Further, owing to Lemma 4.1 we have the uniqueness of the bounded above solution to (2.11). \(\square \)
Remark
It is to be noted that even in the case of \(\rho =0\), H–J–B equation (2.11) has the unique solution without ambiguity of additive constants with respect to \(z(x)\). Moreover, considering (2.11) with \(\rho >0\) and without assumption (2.16) can be reduced to the case of \(\rho =0\) with assumption (2.16) (cf. Remark in Sect. 3).
Theorem 4.2
For \(0<\gamma <1\), we assume assumptions (2.13)–(2.15), 2.17’ and (2.17), and let \({\underline{z}}(x)\) and \({\overline{z}}(x)\) be, respectively sub- and super- solutions to (2.11) appeared in Proposition 3.1. Then, for each \(\rho > {\hat{\chi }}\), (2.11) has a solution \(z\) such that \({\underline{z}}(x)\le z(x)\le {\overline{z}}(x)\) and that it satisfies
Moreover, the solution satisfying (4.2) is unique.
Proof
As in the proof of the previous theorem, the existence of a solution is given in a similar manner to the proof of Theorem 3.5 in [7] (cf. Appendix 1). We give the proof of unique existence of the solution satisfying (4.2). First note that the solution \(z(x)\) to (2.11) necessarily satisfies (4.2) under our assumptions. Indeed, we have seen that \({\hat{z}}(x)-z_0(x)\rightarrow -\infty \) as \(|x|\rightarrow \infty \) in the proof of Lemma 3.5, and thus, (4.2) follows from \(z(x)\le {\hat{z}}(x)+C\). Let us prove uniqueness. Set
for a solution \(z\) to (2.11). Then,
Therefore,
where
and
Let \(z_1\) and \(z_2\) be solutions to (2.11) satisfying (4.2) and set \(\psi _i= z_i-z_0\), \(i=1,2\). Assume that there exists \(x_0\) such that \(z_2(x_0)>z_1(x_0)\). Then, \(\psi _2(x_0)>\psi _1(x_0)\) and \(\psi _i(x)\rightarrow -\infty \), as \(|x|\rightarrow \infty \). Therefore, for each \(\epsilon >0\) there exists \(x_\epsilon \) such that
At \(x_{\epsilon }\) we have
Therefore,
On the other hand, \(D\psi _2=D\psi _1e^{\epsilon \psi _1-\epsilon \psi _2}\) because \(D\psi _{\epsilon }=0\) at \(x_\epsilon \), and thus,
Then, from (4.6), it follows that
and so,
by taking \(\epsilon \) such that \(\epsilon <\frac{1}{1-\gamma }\). Thus, we obtain
from (4.7). Since \(V(x)\rightarrow \infty \) as \(|x|\rightarrow \infty \), there exists \(R>0\) independent of \(\epsilon \) such that \(x_{\epsilon }\in B_R\) and \(\epsilon _n\) such that \(x_{\epsilon _n}\rightarrow {\hat{x}}\in {\overline{B}_R}\). From (4.5), we have
and, by letting \(\epsilon _n\rightarrow 0\), we obtain
From (4.7), at \(x_{\epsilon }\), we have
and letting \(\epsilon _n\rightarrow 0\), the right-hand side tends to \(-\infty \), which is a contradiction. Therefore \(\psi _2(x)\le \psi _1(x)\) for each \(x\). In the same way, we have the converse inequality, and hence, we proved uniqueness of the solution to (2.11). \(\square \)
Let us set the operator \({\overline{L}}\) by
Then, inspired by Lemma 3.6 we have the following proposition useful in the proof of the verification theorem.
Proposition 4.1
Under the assumptions of Theorem 4.2, the diffusion process with the generator \({\overline{L}}\) is ergodic.
Proof
Let us set
Then, as was seen in the proof of Theorem 4.2, \({\bar{\psi }}(x)\rightarrow \infty \), as \(|x|\rightarrow \infty \). Further, from (4.3) it follows that
Since \(z_0=z+{\bar{\psi }}\) we have
Thus, we can see that \({\overline{L}}{\bar{\psi }}(x)\rightarrow -\infty \) as \(|x|\rightarrow \infty \) and hence the proof is complete. \(\square \)
5 Verification Theorems
Let us set the value functions
for \(\gamma <0\), and
for \(0<\gamma <1\), where \({\mathcal A}_1\) is the sets of admissible strategies defined in the end of Sect. 2. For a solution \(z(x)\) to (2.11), we shall prove that \(e^{z(x)}={\hat{v}}(x)\), for \(\gamma <0\), under assumptions (2.13)-(2.15) (resp. (2.13)-(2.16)) for \(\rho >0\) (resp. \(\rho =0\)), and that \(e^{z(x)}={\check{v}}(x)\), for \(0<\gamma <1\), under the assumptions of Theorem 4.2. For the solution \(z(x)\) to (2.11), define a function \({\hat{h}}(x)\) by
and
We define also
Let us prepare the following lemma for the proof of the verification theorem.
Lemma 5.1
Let \(Y_t\) be a solution to the stochastic differential equation
where \(\sigma \) and \(\mu \) are locally Lipschitz continuous with respect to \(x\) and continuous in \(t\). We moreover assume that \( |\mu (t,x)|\le C(1+|x|) \) and \(\sigma \) is bounded. For a given continuous function \(H(t,x)\) satisfying \( |H(t,x)|\le C(1+|x|), \) define \(\rho _t\) by
Then, we have
The proof of this lemma is similar to that of Lemma 4.1.1 in [1] and we omit the proof.
We have the following theorem.
Theorem 5.1
For \(\rho =0\) (resp. \(\rho >0\)), assume assumptions (2.13)–(2.16) (resp. (2.13)–(2.15)). Then, for a solution \(z(x)\) to (2.11), we have
with \(\gamma <0\).
Proof
Let us first note that \({\hat{h}}_s\) satisfies (2.8). Indeed, because of the gradient estimates for the solution \(z\) given in (7.1) in Appendix 2, we can see it by using the above lemma. Thus, we have a probability measure \(P^{\hat{h}}\) under which \(X_t\) satisfies the stochastic differential equation
where
Set
Then, by Itô’s formula, we have
Note that
and hence,
Therefore,
where
Once again, owing to the above lemma, we see that \(M_t\) satisfies (2.19) with respect to \(P^{{\hat{h}}}\) and therefore \(({\hat{c}},{\hat{h}})\in {\mathcal A}_1\). Thus, we obtain
since
When setting \(\varphi _t=e^{-\int _0^t{\hat{c}}_s^{\gamma }e^{-z(X_s)}ds}=e^{-\int _0^te^{-\frac{z(X_s)}{1-\gamma }}ds}\), we have
From Theorem 4.1, we have
with \(K_{\gamma }=c^{-\frac{\gamma }{1-\gamma }}(\frac{-1}{\gamma c_0})^{-\frac{1}{1-\gamma }}\) and thus, \( \lim _{T\rightarrow \infty }\varphi _T=\varphi _{\infty }=0\). Therefore we see that
Hence,
Now, we shall prove that
For the controlled process defined by
we have
from H–J–B equation (2.11). Therefore,
where
Thus, we obtain
where \({\overline{P}}^{h}\) is a probability measure defined by
Let us first assume that \(c_t\le K\), \(\forall t\), for some positive constant \(K>0\). In this case,
holds since \(z(x)\) is bounded above by a constant \(C\). Then we have
and thus obtain
Hence (5.4) holds.
For general \((c,h)\in {\mathcal A}_1\) we set \(c_s^{(n)}:=\min \{c_s,n\}\). Then, we have
and therefore
Hence, by monotone convergence theorem we have (5.4). \(\square \)
Theorem 5.2
Under the assumptions of Theorem 4.2, for a solution \(z(x)\) to (2.11), we have
Proof
Similarly to the proof of Theorem 5.1, we see that \(({\hat{c}}_s,{\hat{h}}_s)\in {\mathcal A}_1\) and have
where
We first note that \((X_t,{\overline{P}}^{{\hat{h}}})\) is the ergodic diffusion process with the generator \({\overline{L}}\) defined by (4.8) according to Proposition 4.1. Moreover,
and, for each \(R>0\), there exists a positive constant \(C_R\) such that
Therefore,
as \(T\rightarrow \infty \), where \(m(dx)\) is the invariant measure of \((X_t,{\overline{P}}^{{\hat{h}}})\). Therefore, we see that
as \(T\rightarrow \infty \), and thus we have \(\varphi _T\rightarrow 0,\;\;{\overline{P}}^{\hat{h}} \; \text{ a.s. } \) since
Thus, we have the first equality.
To prove
we use H–J–B equation (2.12) similarly to the above, and then we arrive at
\(\square \)
References
Bensoussan, A.: Stochastic Control of Partially Observable Systems. Cambridge University Press, Cambridge (1992)
Cox, J.C., Huang, C.-F.: Optimal consumption and portfolio policies when asset prices follow a diffusion process. J. Econom. Theory 49, 33–83 (1989)
Chacko, G., Viceira, L.M.: Dynamic consumption and portfolio choice with stochastic volatility in incomplete markets. Rev. Financ. Stud. 18, 1369–1402 (2005)
Fleming, W.H., Hernandez-Hernandez, D.: An optimal consumption model with stochastic volatility. Financ. Stoch. 7, 245–262 (2003)
Fleming, W.H., Hernandez-Hernandez, D.: The tradeoff between consumption and investment in incomplete markets. Appl. Math. Optim. 52, 219–235 (2005)
Fleming, W.H., Pang, T.: An application of stochastic control theory to financial economics. SIAM J. Cont. Optim. 52, 502–531 (2004)
Hata, H., Sheu, S.J.: On the Hamilton–Jacobi–Bellman equation for an optimal consumption problem: I. Existence of Solution. SIAM J. Cont. Optim. 50, 2373–2400 (2012)
Hata, H., Sheu, S.J.: On the Hamilton–Jacobi–Bellman equation for an optimal consumption problem: II. Verification theorem. SIAM J. Cont. Optim. 50, 2401–2430 (2012)
Kaise, H., Nagai, H.: Ergodic type Bellman equations of risk-sensitive control with large parameters and their singular limits. Asymptot. Anal. 20, 279–299 (1999)
Kaise, H., Sheu, S.J.: On the structure of solutions of ergodic type bellman equation related to risk-sensitive control. Ann. Prob. 34, 284–320 (2006)
Karatzas, I., Lehoczky, J.P., Shreve, S.E.: Explicit solution of a general consumption/investment problems. Math. Oper. Res. 11, 261–294 (1986)
Liu, J.: Portfolio selection in stochastic environments. Rev. Financ. Stud. 20, 1–39 (2007)
Ladyzhenskaya, O.A., Ural’tseva, N.N.: Linear and Quasilinear Elliptic Equations. Academic Press, New York (1968)
Merton, R.C.: Optimal consumption and portfolio rules in a continuous time model. J. Econom. Theory 3, 373–413 (1971)
Nagai, H.: Ergodic control problems on the whole Euclidean space and convergence of symmetric diffusions. Forim Mathematicum 4, 159–173 (1992)
Nagai, H.: Bellman equations of risk-sensitive control. SIAM J. Control Optim. 34, 74–101 (1996)
Nagai, H.: Downside risk minimization via a large deviations approach. Ann. Appl. Prob. 22, 608–669 (2012)
Parthasarathy, K.R.: Probability Measures on Metric Spaces. AMS Chelsea Publishing, Providence (1967)
Pliska, S.R.: A stochastic calculus model of continuous trading: optimal portfolio. Math. Oper. Res. 11, 239–246 (1986)
Wachter, J.A.: Portfolio and consumption decisions under mean-reverting returns: an exact solution for complete markets. J. Financ. Quant. Anal. 37, 69–91 (2002)
Acknowledgments
The author would like to express his thanks to professor S. J. Sheu who attracted his interest to the consumption-investment problems with useful discussions about them, and also encouraged him to extend his private notes on the problems to an article to be published somewhere. Thanks are also due to the referee for his (her) careful reading of the manuscript and correcting a number of typos.
Author information
Authors and Affiliations
Corresponding author
Additional information
Supported in part by a grant-in-aid for Scientific Research 25400150, 23244015, 23340028 JSPS.
Appendices
Appendix 1
Here, we give the proof of the existence of a solution to (2.11) along the line of Hata and Sheu [7], which complete the proofs of Theorems 4.1 and 4.2. We rewrite the equation (2.11) as follows.
where
To prove the existence of a solution to (6.1), we first consider the Dirichlet problem on \(B_R\):
where \({\underline{z}}\) is a smooth sub-solution to (2.11). For \(\gamma <0\), \({\underline{z}}\) is the sub-solution appeared in Lemma 3.1 and for \(0<\gamma <1\), \({\underline{z}}=-C\) with sufficiently large \(C\). In proving the existence of a solution to (6.2) we appeal to the following theorem, a modification of the Leray-Schauder fixed point theorem, according to the scheme due to Hata and Sheu ([7]).
Theorem 6.1
(Hata and Sheu [7]) Let \({\mathcal B}\) be a Banach space with the norm \(\parallel \cdot \parallel _{{\mathcal B}}\) and \(T\) a continuous, compact operator from \({\mathcal B}\times [0,1]\) to \({\mathcal B}\). Assume that there exists a constant \(M>0\) such that \(\parallel \xi \parallel _{{\mathcal B}}<M\) for all \((\xi ,\tau )\) satisfying \(\xi =T(\xi ,\tau )\), or \(\xi =\tau T(\xi ,0)\). Then, there exists a fixed point \(\xi \in {\mathcal B}\): \(\xi = T(\xi ,1)\).
In applying this theorem, we consider the following problem for each \(\tau \in [0,1]\):
We note that, under assumptions (2.13)–(2.15), for \(x\in B_R,\; |z|\le M\), we have
where \(c_1,c_2\) and \(c_3\) are positive constants. To study (A.3), we consider the linear partial differential equation
for a given function \(w\in C^{1,\mu }({\overline{B}_R})\), \(1>\mu >0\). Under assumptions (2.13)–(2.15), we have a unique solution \(z\in C^{2,\mu \mu '}({\overline{B}_R})\) since \({\underline{z}}\in C^{2,\mu }({\overline{B}_R})\) (cf. [13]). Thus, we can define a continuous, compact mapping \(T(w,\tau )\) of \({\mathcal B}\times [0,1]\) into \({\mathcal B}\) as \(T(w,\tau )=z\), where \(z\) is the solution to (6.4) for a given function \(w\in {\mathcal B}:=C^{1,\mu }({\overline{B}_R})\) and \(\tau \in [0,1]\). Indeed, since \(b(x,w,Dw;\tau \gamma )\in C^{0,\mu }({\overline{B}_R}) \) for \(w\in C^{1,\mu }({\overline{B}_R})\), and we assume that \(\tau {\underline{z}} \in C^{2,\mu }({\overline{B}_R})\), \(T(w,\tau )\), for every \(\tau \), transforms the function \(w\in C^{1,\mu }({\overline{B}_R})\) into \(z(x;\tau )\) in \(C^{2,\mu '\mu }({\overline{B}_R})\). Further, we have
where \(f\) is a continuous monotonically increasing function of \(t\in [0,\infty )\). Since an arbitrary bounded set in \(C^{2,\mu '\mu }({\overline{B}_R})\) is compact in the space \(C^{1,\mu }({\overline{B}_R})\), \(T(w,\tau )\) maps each bounded set of the pairs \((w,\tau )\) in \(C^{1,\mu }({\overline{B}_R})\times [0,1]\) into a compact set in \(C^{1,\mu }({\overline{B}_R})\). On the other hand, when \(\parallel w_1-w_2\parallel _{C^{1,\mu }({\overline{B}_R})} \) goes to \(0\), \(\parallel b(x,w_1,Dw_1;\tau \gamma )- b(x,w_2,Dw_2;\tau \gamma )\parallel _{C^{0,\mu }({\overline{B}_R})}\) tends to \(0\) uniformly with respect to \(\tau \). Moreover, \(T(w,\tau )\) is continuous in \(\tau \) uniformly with respect to \((x, w, Dw) \in {\overline{B}_R}\times \{u\in R^n;|u|\le c\}\times \{p\in R^n; |p|\le c\}\). Therefore \(T(w,\tau )\) is a continuous map of \((w,\tau )\in C^{1,\mu }({\overline{B}_R})\times [0,1]\) into \(C^{1,\mu }({\overline{B}_R})\).
Note that a fixed point \(z^{(\tau )}\) of \(T\), \(z^{(\tau )}=T(z^{(\tau )},\tau )\), is a solution to (6.3) and \(z^{(1)}\) is a solution to (6.2). On the other hand, if we set \(z_0=T(w,0)\), then \(z_0\) satisfies
with
Therefore, \(z_0^{(\tau )}=\tau T(w;0)\) turns out to be a solution to the equation
Hence, a fixed point \({\hat{z}}_0^{\tau }=\tau T({\hat{z}}_0^{\tau };0)\) is a solution to
Now, let us give the proof of the existence of a solution to (6.1).
Step 1. Proof of the existence of a solution to (6.2).
Owing to Theorem 3.8 in Hata and Sheu [7], we have the estimate for \(z^{(\tau )}\):
where \(f_0(x)\) is the solution to
Moreover, we see that
for sufficiently large \(C>0\). Indeed, when \(\gamma <0\), for \(x\in \{x; {\underline{z}}(x)\le 0\}\) \(\tau {\underline{z}}(x)\ge {\underline{z}}(x)>-C\) for each \(\tau \in (0,1]\) since \({\underline{z}}(x)\) is bounded in \({\overline{B}_R}\). \(\tau {\underline{z}}(x)>-C\) holds as well in \(x\in \{x; {\underline{z}}(x)\le 0\}^c\). Moreover, \(-C\) becomes a sub-solution of (6.3) by taking \(C\) to be sufficiently large. Therefore we see (6.10) owing to Lemma 3.6 in [7]. Similarly, (6.10) holds also in the case of \(0<\gamma <1\).
Further, owing to Theorem 3.9 in Hata and Sheu [7], we have
where \(\sigma _R=\inf \{ t; |X_t|=R\}\), and \(X_t\) is the solution to the stochastic differential equation:
Therefore, from Theorem 4.1 and 6.1, Chapter 4 in [13], we obtain the estimates
for a positive constant \(M\) independent of \(\tau \), \(z^{(\tau )}\) and \(z_0^{(\tau )}\). Then, from Theorem 4.1 and 6.1, Chapter 4 in [13], we obtain
where the constants \(M_1\), \(M_2\) and \(\mu '\) are determined by \(n\), \(M\), \(c_1\), \(c_2\), and \(c_3\). Thus, we can consider the operator \(T(w;\tau )\) only on the space
for \(\epsilon >0\). Then, we see that there exists \({\tilde{M}}\) such that
for any fixed points \(z^{(\tau )}\) and \(z_0^{(\tau )}\). Hence, Theorem 6.1 applies and we see that \(z\in C^{2,\mu \mu '}({\overline{B}_R})\) such that \(z=T(z,1)\).
Step 2. Proof of the existence of the solution to (6.1).
Let us take a sequence \(\{R_n\}\) such that \(R_n\rightarrow \infty \), as \(n\rightarrow \infty \), and a sequence of solutions \(z_{R_n}\) to (6.2). Since \({\underline{z}}\) is a subsolution to (6.1), we can see that \(z_{R_n}\) is nondecreasing because of the maximum principle. Further, we can see that \(z_{R_n}\) is dominated by the super-solution \({\overline{z}}(x)\) again by the maximum principle. Therefore, there exists \(z(x)\) to which \(z_{R_n}\) converges as \(n\rightarrow \infty \). Note that there exists a constant \(M\) independent of \(n\) such that
for \(r<R_n\), which can be seen in a similar manner to the proof of Proposition 3.2 in [17] (cf. Appendix 2 and also (3.8)). Therefore, we can see that \(z_{R_n}\rightarrow z \), \(W^{1,p}_{loc}\) weakly \(\forall p>1\) by taking a subsequence if necessary. The convergence can be strengthen as \(\nabla z_{R_n}\) converges in \(L^2_{loc}\) strongly to \(\nabla z\). As a result we can see that \(z\in W^{1,p}_{loc}\) is a weak solution to (6.1). Then, from the regularity theorem we see that \(z\in C^{2,\mu ''}\) and that it is a classical solution to (6.1). \(\square \)
Appendix 2
Let us give the gradient estimates for the solution to H–J–B equation (2.11).
Lemma 7.1
Under assumptions(2.13)–(2.15), the solution \(z\) to (2.11) satisfies the following estimate.
for some positive constant \(C>0\).
The proof of this estimate is almost the same as the one of Proposition 3.2 in [17]. Here we only give some remarks that the proof could proceed in almost parallel to it. One could see [17] to be more precise.
Proof
Set \(Q^{ij}:=(\lambda N_{\gamma }^{-1}\lambda ^*)^{ij}\) and differentiate (2.11). Then, we have
Set \(F=|\nabla z|^2=\sum _{k=1}^n|D_kz|^2\) and
Then, we have
Here we have used (7.2) and the matrix inequality \( (\text{ tr }[AB])^2\le nC\text{ tr }[AB^2]\), for symmetric matrix \(B\) and nonnegative definite symmetric matrix A having the maximum eigenvalue \(C\). Thus, in a similar manner to the proof of Proposition 3.2 in [17] (cf. also [9, 10, 15]) we can obtain estimate (7.1) by using H–J–B equation (2.11) again in the last line in the above. Although we have the term \((1-\gamma )e^{-\frac{z}{1-\gamma }}\) in the equation it does not affect the proof since it is nonnegative. \(\square \)
Rights and permissions
About this article
Cite this article
Nagai, H. H–J–B Equations of Optimal Consumption-Investment and Verification Theorems. Appl Math Optim 71, 279–311 (2015). https://doi.org/10.1007/s00245-014-9258-0
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00245-014-9258-0
Keywords
- Utility maximization
- Risk-sensitive stochastic control
- Factor models
- H–J–B equation
- Infinite time horizon