Abstract
We discuss the turnpike property for optimal investment and consumption problems. We find there exists a threshold value that determines the turnpike property for investment policy. The threshold value only depends on the Sharpe ratio, the riskless interest rate and the discount rate. We show that if utilities behave asymptotically like power utilities and satisfy some simple relations with the threshold value, then the turnpike property for investment holds. There is in general no turnpike property for consumption policy. We also provide the rate of convergence and illustrate the main results with examples of power and non-HARA utilities and numerical tests.
Similar content being viewed by others
Notes
A utility function U and its dual function V are equivalent and can be recovered by each other from the relations \(V(y)=\sup _{x>0}(U(x)-xy)\) and \(U(x)=\inf _{y>0}(V(y)+xy)\). For example, if U is a power utility \(U(x)=(1/p)x^p\) for \(x>0\) and \(p<1\), then its equivalent dual function is \(V(y)=-(1/q)y^q\) for \(y>0\) and \(q=p/(p-1)\). Furthermore, \(U_i(x)=V_i(y)+xy\) if and only if \(U'_i(x)-y=0\). Therefore, from
$$\begin{aligned} {U_i'(x)\over x^{p_i-1}}= \left( {y^{q_i-1}\over -V_i'(y)}\right) ^{p_i-1} \end{aligned}$$and \(U_i'(\infty )=0\), we get the equivalence of conditions (1.3) and (1.4).
The result of Theorem 2.4 can be stated equivalently in terms of \(p_i\) defined in (1.3), that is, if \(p_1>p^*\) or \(p_2\ge p^*\), where \(p^*=q^*/(q^*-1)\ge 0\), then the turnpike property holds and the optimal amount of investment can be approximated by \( \lim _{T\rightarrow \infty }A(x,t)=(\theta /\sigma )(1-\max \{p_1,p_2\})^{-1}x\).
Note that \(q^*\) only depends on the market price of risk \(\theta \), the riskless interest rate r and the utility discount rate \(\delta \).
Condition (1.9) is equivalent to the limits of the Arrow–Pratt coefficients of relative risk aversion of utilities \(U_i\) being \(1-p_i\), that is,
$$\begin{aligned} \lim _{x\rightarrow \infty }\left( -{xU_i''(x)\over U_i'(x)}\right) = 1-p_i,\quad i=1,2, \end{aligned}$$(1.10)and \(p_1,p_2>p^*\). This is due to the dual relation of \(U_i\) and \(V_i\) defined in (1.5), which implies that if \(y=U_i'(x)\) then \(x=-V_i'(y)\) and \(U_1''(x)=-1/V_1''(y)\), and therefore the equivalence of conditions (1.9) and (1.10).
The limits in conditions (1.4) and (2.6) are \(-1\) which can be replaced by some constants. Specifically, if there exist \(k_i>0\), such that \(\lim _{y\rightarrow 0}\frac{V_i'(y)}{y^{q_i-1}}=-\,k_i\), then for \(x\in R_+\), we still have the turnpike property (1.7). If, for some \({\hat{q}} <1\), \(k\ge 0\), \(\lim _{y\rightarrow \infty } \frac{V'_2(y)}{y^{{\hat{q}}-1}}=-\,k\), then, we have \(\lim _{t\rightarrow \infty }\hat{R}(t)^{\frac{{\hat{q}}-1}{q_2-1}}C(x,t) =k x^{\frac{ {\hat{q}}-1}{\min \{q_1,q_2\}-1}}\), where \(\hat{R}(t)=k_1e^{\lambda _1 t}+k_2\frac{e^{\lambda _2 t}-1}{\lambda _2}\) if \(q_1=q_2\), \(k_1^{\frac{q_2-1}{\hat{q}-1}}e^{\frac{q_2-1}{q_1-1}\lambda _1 t}\) if \(q_1<q_2\), and \(k_2\frac{e^{\lambda _2 t}-1}{\lambda _2}\) if \(q_1>q_2\). The proof is the same as that of Theorem 2.4 with some obvious changes to include \(k_i\) and k.
It is easy to verify that both (1.4) and (1.9) imply (2.8), so (2.8) is the weakest condition among three conditions. The function \(V'(y)=y^{q-1}\ln y (y\le Y<1)\) is an example that satisfies (1.9), but not (1.4) (see [8], page 1346). The function \(V'(y)=-y^{q-1}e^{hy\sin \frac{1}{y}}\), where \(h<\frac{1-q}{2}\), is an example that satisfies (1.4), but not (1.9) (see Back et al. [1], page 178). However, if \(\lim _{y\rightarrow 0} V_i''(y)/y^{q_i-2}\) exists, then both (1.4) and (1.9) are satisfied with L’Hospital’s Rule.
Assume f and g are well defined functions on \(R_+\times R_+\) and \(R_+\), respectively. We say f(x, t) converges to g(x) exponentially (or polynomially) as \(t\rightarrow \infty \) if there exist constants \(c>0\) and \({\bar{T}}>0\) and a well defined function D on \(R_+\), such that \(|f(x,t)-g(x)|\le D(x)e^{-ct}\) (or \(|f(x,t)-g(x)|\le D(x)t^{-c}\)) for all \(x\in R_+\) and \(t\ge {\bar{T}}\). Exponential convergence is much faster than polynomial convergence.
If \(V_2\equiv 0\), we need only \(\bar{q}\in (\max \{q_1, 0\}, 1]\).
We have \({\bar{q}}={1\over 2} (q+1)>q\) and \({\bar{q}}-1={1\over 2}(q-1)\). It is easy to verify that U is strictly increasing and strictly concave, \(U(0)=0\), \(U(\infty )=\infty \), \(U'(0)=\infty \) and \(U'(\infty )=0\). Therefore U is a utility function in the classical sense. This utility (for \(p=3/4\)) is used in Bian and Zheng [3] to illustrate the turnpike property and the convergence rate for a terminal wealth utility maximization problem.
If \(\{e_n\}\) is a sequence of errors with exponential convergence, then there are positive constants M and c such that \(|e_n|\le Me^{-cn}\). To find c, we may assume \(|e_n|\approx Me^{-cn}\), which gives \( c\approx -\ln \left( |e_{n+1}|/|e_n|\right) \). In our numerical tests, we have chosen time horizon \(t=1,2,5,10,25,50,100\), which does not have equally spaced intervals. An adjustment is needed to reflect this. Specifically, we estimate c by \( c_n:=- (1/m)\ln \left( |e_{n+m}|/ |e_n|\right) \), where m is an integer indicating the distance of indices n and \(n+m\) for errors \(e_n\) and \(e_{n+m}\). For \(n:=1,2,5,10,25,50,100\), the distance between adjacent points are \(m=1,3,5,15,25,50\). We form a sequence \(\{c_n\}\) to see if there is a limit which would indicate the approximate exponent of the exponential convergence rate.
References
Back, K., Dybvig, P.H., Rogers, L.C.G.: Portfolio turnpikes. Rev. Financial Stud. 12, 165–195 (1999)
Bian, B., Miao, S., Zheng, H.: Smooth value functions for a class of nonsmooth utility maximization problems. SIAM J. Financial Math. 2, 727–747 (2011)
Bian, B., Zheng, H.: Turnpike property and convergence rate for an investment model with general utility functions. J. Econ. Dyn. Control 51, 28–49 (2015)
Bingham, N.H., Goldie, C.M., Teugels, J.L.: Regular Variation. Cambridge University Press, Cambridge (1987)
Cox, J., Huang, C.: A continuous time portfolio turnpike theorem. J. Econ. Dyn. Control 2, 491–507 (1992)
Guasoni, P., Kardaras, C., Robertson, S., Xing, H.: Abstract, classic, and explicit turnpikes. Finance Stoch. 18, 75–114 (2014)
Huang, C., Zariphopoulou, T.: Turnpike behaviour of long-term investments. Finance Stoch. 3, 15–34 (1999)
Huberman, G., Ross, S.: Portfolio turnpike theorems, risk aversion, and regularly varying functions. Econometrica 51, 1345–1361 (1983)
Jin, X.: Consumption and portfolio turnpike theorems in a continuous-time finance model. J. Econ. Dyn. Control 22, 1001–1026 (1998)
Robertson, S., Xing, H.: Long term optimal investment in matrix valued factor models. SIAM J. Financial Math. 8, 400–434 (2017)
Acknowledgements
The authors are grateful to the anonymous reviewer whose comments and suggestions have helped to improve the previous two versions.
Author information
Authors and Affiliations
Corresponding author
Additional information
Baojun Bian: Research of this author was supported by NSFC 11371280.
Appendix
Appendix
Derivation of optimal trading strategy (2.1) and budget equation (2.2) We may use a standard stochastic control method to solve problem (1.1). Define the value function u by
for \((x,t)\in R_+\times [0,T]\). Then u satisfies the HJB equation
for \((x,t)\in R_+\times [0,T)\) with the terminal condition \(u(x,T)=U_1(x)\) for \(x\in R_+\), where \(u_x(x,t)\) is the partial derivative of u with respect to x and evaluated at (x, t) (we have omitted (x, t) in equation (5.1) to simplify notations), \({\partial \over \partial t}u\) and \(u_{xx}\) are defined similarly.
Let \(V_i\) be the dual functions of \(U_i\), \(i=1,2\), defined in (1.5). Then \(V_i\) are nonnegative (since \(U_i(0)=0\)), continuously differentiable, strictly decreasing, strictly convex functions.
The optimal investment and consumption strategies in the HJB equation (5.1) are given by
The HJB equation (5.1) can be written as
Equation (5.3) is a fully nonlinear PDE and is in general difficult to solve. In the literature equation (5.3) is solved with some trial-and-error method, which works when \(U_1\) and \(U_2\) are the same power utility functions. Bian et al. [2] and Bian and Zheng [3] apply the dual stochastic control method to solve problem (1.1) and show that the HJB equation (5.1) has a classical solution that can be represented in terms of the dual function of the corresponding dual control problem. The value function v(y, t) of the dual control problem satisfies a linear PDE
for \((y,t)\in R_+\times [0,T)\), with the terminal condition \(v(y,T)=V_1(y)\). Its solution v is continuous on \(R_+\times [0,T]\) and \(C^{2,1}\) on \(R_+\times [0,T)\), and v is strictly decreasing and strictly convex in y for fixed \(t<T\). A classical solution to the HJB equation (5.1) is given by
where y is the unique solution to the equation
This leads to \(u_x(x,t)=y\) and \(u_{xx}=-\,1/v_{yy}\). From (5.2) we conclude that the optimal amount of investment A(x, t) and the optimal consumption rate C(x, t) are given by
To simplify the notation, we define \(\tau =T-t\), the time horizon. Then \(T\rightarrow \infty \) is equivalent to \(\tau \rightarrow \infty \). We still use t to represent a time horizon variable, instead of \(\tau \).
It is easy to verify that \(v_y(y,t)\) is the solution of the initial value problem
for \((y,t)\in R_+\times R_+\) with initial condition \(v_y(y,0)=V'_1(y),\ y\in R_+\). By Poisson’s formula, the solution of equation (5.5) is given by
Preliminaries of mathematical proofs We need some technical results in proofs. Simple calculus gives that for any constant A,
Hence for any \(q<1\),
and
For any \(y>0\), the convexity of \(V_i\) implies that
which, together with the decreasing property and the nonnegativity of V, gives
for all \(y\ge 1\).
We need the following algebraic inequality in the proof:
for all \(x,y\ge 0\) and \(0<c<\infty \). (5.11) can be proved as follows: assume \(x>y>0\) and let \(z=x/y\). Then \(z>1\) and (5.11) is equivalent to \(z^c -1 \le (1+c)(z^c-z^{c-1})\). Let \(g(z)=z^c -1-(1+c)(z^c-z^{c-1})\). Then \(g'(z)=cz^{c-1}-(1+c)(cz^{c-1}-(c-1)z^{c-2})=-c^2(1-z)z^{c-2}-z^{c-2}\le 0\) for \(z\ge 1\). So g is decreasing and \(g(z)\le g(1)=0\) for \(z\ge 1\).
We have, for any \(q<\bar{q}<1\), \(\lambda =\lambda (q), \bar{\lambda }=\lambda (\bar{q})\)
and
Proof of Theorem 2.3
Since \(U'_i(x)= x^{p_i-1}\) for \(i=1,2\), the dual functions are given by \(V'_i(y)=-y^{q_i-1}\) where \(q_i=p_i/(p_i-1)\). Substituting \(V'_i(y)\) into (5.6), also noting (5.7), we have
Hence we derive (2.4) from (2.1), where y is the solution to the equation (2.5).
We next complete the proof by directly discussing three cases.
Case 1 When \(q_1<q^*\) or \(q_2\le q^*\), we have \(\lambda _1>0\) or \(\lambda _2\ge 0\). We know that the solution \(y=u_x\) to equation (2.5) must tend to \(\infty \) as t tends to \(\infty \) (otherwise the right side of (2.5) tends to \(\infty \)). We can find the dominating term in (2.5) as t tends \(\infty \) for different \(q_1\) and \(q_2\).
Assume \(q=q_1=q_2 \le q^*\). It is easy to see from (2.1) and (2.5) that \(A(x,t)=\frac{\theta }{\sigma } (1-q)x\) and \(C(x,t)=-V'_2 (u_x(x,t))=(u_x(x,t))^{q_2-1}=E(t)x\), where \(E(t)=(e^{\lambda _1 t} +\frac{e^{\lambda _2 t }-1}{\lambda _2})^{-1}\). This implies \(R(t)C(x,t)=x\).
Assume \(q_1<\min \{q_2, q^*\}\), then \(\lambda _1>\max \{\lambda _2, 0\}\). Noting that
we obtain
From (2.5) and (5.13), we deduce that \(\lim _{t\rightarrow \infty }\frac{e^{\lambda _2 t }-1}{\lambda _2}y^{q_2-1}=0\) and \(\lim _{t\rightarrow \infty }e^{\lambda _1 t}y^{q_1-1}=x\). Hence from (2.4)
We have by (2.4)
which implies \(\lim _{t\rightarrow \infty } R(t) C(x,t)=x^{\frac{q_2-1}{q_1-1}}\).
Assume \(q_2< q_1\) and \(q_2\le q^*\), then \(\lambda _2> \lambda _1\) and \(\lambda _2\ge 0\). We have
for \(t\ge 1\). Then by (2.5), for \(t\ge 1\)
We deduce from (2.5) and (5.13) \(\lim _{t\rightarrow \infty }e^{\lambda _1 t}y^{q_1-1}=0\) and \(\lim _{t\rightarrow \infty }\frac{e^{\lambda _2 t }-1}{\lambda _2}y^{q_2-1}=x\). Furthermore, we get
and \(\frac{e^{\lambda _2 t }-1}{\lambda _2}C(x,t)= \frac{e^{\lambda _2 t }-1}{\lambda _2}y^{q_2-1}\), which gives \(\lim _{t\rightarrow \infty } R(t)c(x,t)=x\).
Case 2 Since \(q_1> q^*\) and \(q_2>q^*\), we have \(\lambda _1, \lambda _2< 0\). From (2.5), we deduce that \(y=u_x(x,t)\) must have upper bound in t (otherwise, we would have \(x=0\), a contradiction). Therefore \(x=\lim _{t\rightarrow \infty }(-1/\lambda _2)u_x(x,t)^{q_2-1}\). Taking the limit in (2.4) leads to the desired results.
Case 3 Since \(q_1= q^*\) and \(q_2>q^*\), we have \(\lambda _1=0\) and \(\lambda _2<0\). Again from (2.5), we deduce that \(y=u_x(x,t)\) must have upper bound. Let Y be the unique solution to the nonlinear equation
then we have \(\lim _{t\rightarrow \infty }u_x=Y\). Taking the limit in (2.4) leads to the desired results. \(\square \)
Proof of Theorem 2.4
From (1.4) and (5.10), we conclude that there is constant \(\bar{q}\in (\max \{q_1,q_2,0\}, 1)\), for any fixed \(\epsilon >0\), there is \(K_\epsilon >0\), such that
for \(i=1, 2\). By (5.17)(i=1) and (5.7), we get
for all \((y,t)\in R_+\times (0,\infty )\). Taking \(\epsilon =1\) and denoting \(\bar{K}=K_1\), we get, by (5.12) and (5.13)
Since \(0<\frac{\bar{q}-1}{q_1-1}<1\), we see that \((y^{q_1-1}e^{\lambda _1 t})^\frac{q_1-\bar{q}}{q_1-1} \le 2 I_1(y,t)+ \bar{K}\) if \(y^{q_1-1}e^{\lambda _1 t}\ge 1\). Hence
Similarly, we obtain for \(t\ge 1\)
Let \(H_1(t)=(u_x(x,t))^{q_1-1}e^{\lambda _1 t}\), \(H_2(t)=(u_x(x,t))^{q_2-1}\frac{e^{\lambda _2 t}-1}{\lambda _2}\), \(M_1(t)=(u_x(x,t))^{\bar{q}-1}e^{\bar{\lambda } t}\), \(M_2(t)=(u_x(x,t))^{\bar{q}-1}\frac{e^{\bar{\lambda } t}-1}{\bar{\lambda }}\), \(H(t)=H_1(t)+H_2(t)\) and \(M(t)=M_1(t)+M_2(t)\). Then by (2.2)
Recalling \(E(t)=(e^{\lambda _1 t}+\frac{e^{\lambda _2 t}-1}{\lambda _2})^{-1}\). For \(q_1<q^*\) or \(q_2 \le q^*\), we see \(\lambda _1>0\) or \(\lambda _2 \ge 0\). Hence \(E(t)\le 1\) for \(t\ge 1\) and \(\lim _{t\rightarrow \infty }E(t)=0\). We have for all \(y>0\),
Hence
This implies, for \(t\ge 1\)
and \(\lim _{t\rightarrow \infty }u_x(x,t)=\infty \). Using (5.12) and (5.13), we derive the following estimates:
By (5.17), (2.2) and (5.7), we get
Noting that (5.13), (5.21), and letting \(t\rightarrow \infty \) and then \(\epsilon \rightarrow 0\) in (5.22), we deduce the limiting form of (2.5)
By (5.17),(2.1), (5.8) and (5.9), we conclude that
for all \((y,t)\in R_+\times [1,\infty )\). Let
Noting \(q_1,q_2<\bar{q}\), for fixed \(x>0\), we obtain
From estimates (5.18) and (5.21), letting \(t\rightarrow \infty \) and then \(\epsilon \rightarrow 0\) in (5.24), we derive
Next, dividing the cases: \(q_1=q_2 \le q^*\), \(q_1<\min \{q_2, q^*\}\) and \(q_2< q_1, q_2\le q^*\) as in Theorem 2.3, we derive the turnpike property (1.7) from (5.23) and (5.25). To derive (2.7), noting \(C(x,t)=-V'_2(u_x(x,t))\), (2.6), (5.23) and \(\lim _{t\rightarrow \infty } u_x(x,t)=\infty \), we have
The last equality is derived by discussing three cases for \(q_1\) and \(q_2\), the same as that in deriving (1.7). \(\square \)
Proof of Theorem 2.5
The proof is similar to that of Theorem 2.4. With strengthened assumptions, we can give better estimations to some inequalities in the proof of Theorem 2.4. From (5.10) we conclude that assumption (2.9) holds for all \(y>0\) with a changed constant L. This implies that (5.22) becomes
and (5.18) becomes
We have also the estimate (5.21). (5.24) becomes
where \(D= 1-\bar{q}+\frac{2Ee^{\lambda ^+_2}+1}{a\sqrt{\pi }}\). Note that \(\tilde{q}< \min \{\hat{q}, q_2\}\), (2.10) holds for all \(y>0\). Setting \(c:=\frac{{\hat{q}} -1}{\min \{q_1,q_2\}-1}\), we have, from (5.11)
Noting that \(R(t)^{\frac{{\hat{q}}-1}{q_2-1}}\le (E(t))^{\frac{{\hat{q}}-1}{q_2-1}}\), we have, from (2.10) and (5.20)
This gives the estimate for the first term in (5.29).
For \(q=q_1=q_2\le q^*\), we deduce from (5.28) and (5.26),
Let \(G=G(x,c)=(1+c)H^{c-1}\) if \(c>1\) and \(G=G(x,c)=(1+c)x^{c-1}\) if \(c\le 1\). It is clear that \(x<H\) since \(\frac{q_1-1}{q_1-\bar{q}}>1\) from (5.27). Noting \(R(t)u_x^{q-1}=H(t)\) and \(x \le \max \{H(t),x\} \le H\), we deduce the estimate for the second term in (5.29) by (2.10), (5.26) and (5.11),
We deduce convergence rate by (5.21).
If \(q_1<\min \{q_2, q^*\}\), we have, by (5.14),
Note that \(R(t)=e^{\frac{q_2-1}{q_1-1}\lambda _1 t}\) in this case, we have
If \(q_2< q_1, q_2\le q^*\), we have, by (5.16), for \(t\ge 1\)
where \(E:=\frac{\lambda _2}{1- e^{-\lambda _2}}\). We deduce from (5.26) and (5.28)
Noting that \(R(t)=\frac{e^{\lambda _2 t}-1}{\lambda _2}\), from (5.26),
Note that E(t) converges to 0 exponentially as \(t\rightarrow \infty \) if \(q_1<q^*\) or \(q_2 < q^*\), and polynomially if \(q_1\ge q^*\) and \(q_2=q^*\). All other terms converge to 0 exponentially. Combining all discussions above, we have proved the results. \(\square \)
Proof of Theorem 2.6
In this case, \(\lambda _1\le 0\), \(\lambda _2< 0\). From (1.4) and (5.10), we conclude that there is constant \(\bar{q}\in (\max \{q_1,q_2,0\}, 1)\), and for any fixed \(\epsilon >0\), there is \(K_\epsilon >0\), such that
Let \(h(y,t)=\int _{0}^t I_2(y,\tau )d\tau \); then from (5.30) (\(i=2\))
Hence h(y, t) is increasing and bounded above in t. We deduce that \(\lim _{t\rightarrow \infty }h(y,t)=h(y)\), which implies that h(y) is well defined and
Since \(|V'_2(y)|\) is decreasing, we conclude that h(y) is strictly decreasing, \(h(\infty )=0\) and \(h(0)=\infty \) by (5.32). Hence equation (2.13) admits a unique solution for any fixed \(x>0\). We denote this unique solution by \(Y=Y(x)\).
Since \(h(y)-h(y,t)=\int _{t}^\infty I_2(y,\tau )d\tau \) and \(yh'(y)-yh_y(y,t)=\int _{t}^\infty J_2(y,\tau )d\tau \), we have from (5.30) (\(i=2\))
and
By (5.30)(i=1) and (5.7), we get
Hence for fixed \(x>0\)
Note that \(\lambda _1=0\) if \(q_1=q^*\), \(\lambda _1<0\) if \(q_1>q^*\) and \(\bar{\lambda }<0\). Taking \(\epsilon =\frac{1}{2}\), we see from (5.35) and (5.31) that there exists a constant \(\bar{Y}=\bar{Y}(x)\) such that \(u_x(x,t)\le \bar{Y}\). Using (5.31) and taking \(\epsilon =\frac{1}{2}\), we deduce \(\lim _{y\rightarrow 0} h(y,t)=0\) uniformly in t when \(t\ge 1\). Hence there exists a lower bound \(\underline{Y}=\underline{Y}(x)\), \(u_x(x,t)\ge \underline{Y}\). Now assume, for any sequence \({t_k}\), \(\lim _{k\rightarrow \infty }t_k=\infty \) and \(\lim _{k\rightarrow \infty }u_x(x,t_k)=\hat{Y}\). Then (5.33) yields \(\lim _{k\rightarrow \infty }h(u_x(x,t_k),t_k)=h(\hat{Y})\). Letting \(k\rightarrow \infty \) and then \(\epsilon \rightarrow 0\) in (5.35) with \(t=t_k\), we conclude \(\hat{Y}^{q_1-1}1_{\{q_1=q^*\}}+h(\hat{Y})=x\) and \(\hat{Y}=Y\), by the uniqueness of solution for equation (2.13). Therefore, \(\lim _{t\rightarrow \infty }u_x(x,t)=Y\). This yields (2.12) directly.
By (5.30), (5.6) and (5.7), we get
hence
Noting \(\lambda _2, \bar{\lambda } <0\), it yields (2.11) by (5.34).
If \(V_2\) satisfies \(V_2'(y)=-y^{q_2-1}\) with \(q_2<0\), then a simple calculus shows that when \(q_2>q^*\), we have \( h(y)=-(1/\lambda _2)y^{q_2-1}\) and \(yh'(y)=(q_2-1)h(y)\). If \(q_1>q^*\), then relations (2.11)-(2.13) give (2.14). \(\square \)
Proof of Theorem 2.7
From (1.9), for any fixed \(0<\gamma <1-\max \{q_1,q_2\}\), there is \(0<Y=Y_\gamma <1\) such that
Since \(R_i(y)=-yV_i''(y)/V_i'(y)\), we have
for \(y>0\). Hence, there exist positive constants \(l=l_\gamma \) and \(L=L_\gamma \), such that
We now give some estimates for \(I_1(u_x,t)\) and \( \int _{0}^t I_2(y,\tau )d\tau |_{y=u_x}\). Suppose \(q_1<\min \{q_2, q^*\}\), we choose \(\gamma <\min \{\frac{q_2-q_1}{2}, q^*-q_1\}\). Then
Suppose that \(\ln \frac{Y}{y}+2(\alpha -q_1-\gamma )a^2 t\ge 0\); then we get \(x\ge I_1 \ge \frac{l}{2}u_x^{q_1+\gamma -1}e^{\lambda (q_1+\gamma ) t}\). Otherwise, \(\ln \frac{Y}{y}+2(\alpha -q_1-\gamma )a^2 t\le 0\), we have \(u_x \ge Y e^{2(\alpha -q_1-\gamma )a^2 t}\). Noting that \(\alpha -q\ge \alpha -q^*>0\) for \(q\le q^*\) and \(\lambda (q)+2(q-1)(\alpha -q)a^2=-[a^2(q-1)^2+r]<0\), we conclude that
Next assume \(q_2<\min \{q_1, q^*\}\), we chose \(\gamma <\min \{\frac{q_1-q_2}{2}, q^*-q_2\}\). Similarly, we have
Let \(k_0=\frac{\lambda (q_2+\gamma )}{2a^2(1 -q_2-\gamma )(\alpha -q_2-\gamma )}\in (0,1)\) and \(E_0=\frac{\lambda (q_2+\gamma )}{1-e^{(k_0-1)\lambda (q_2+\gamma )}}>0\). Assume \(\ln \frac{Y}{y}+2(\alpha -q_2-\gamma )a^2 k_0 t\ge 0\), then \(x\ge \frac{l}{2E_0}u_x^{q_2+\gamma -1}e^{\lambda (q_2+\gamma ) t}\) for \(t\ge 1\). Otherwise, we conclude that \(u_x \ge Y e^{2(\alpha -q_2-\gamma )a^2 k_0 t}\). We get, for \(t\ge 1\),
Since \(q_2+\gamma <q^*\), we have \(\lambda (q_2+\gamma )>0\) and \(e^{\lambda (q_2+\gamma )t}\rightarrow \infty \) as \(t\rightarrow \infty \). Noting \(q_2+\gamma -1<0\) and (5.38), we conclude that \(\lim _{t\rightarrow \infty }u_x(x,t)=\infty \). On the other hand, from (5.10), we obtain with a changed L
Hence we deduce upper bounds
and
Note that
From (2.15) and (5.10), we conclude that for any fixed \(\epsilon >0\), there is \(K=K_\epsilon >0\), such that
Substituting (5.41) into (5.6), we get by (5.42)
Letting \(y=u_x\), \(t\rightarrow \infty \) and then \(\epsilon \rightarrow 0\) in (5.43), also noting \(\bar{q}>q^*\), \(\bar{\lambda }<0\) and \(\lim _{t\rightarrow \infty }u_x(x,t)=\infty \), we obtain
If \(q_1=q_2 < q^*\), we deduce (1.7) directly. Assume \(q_1<\min \{q_2, q^*\}\). By (5.37) and (5.40), we get \(\lim _{t\rightarrow \infty }\int _{0}^t I_2(u_x,\tau )d\tau =0\), then (1.7). Assume \(q_2<\min \{q_1, q^*\}\). By (5.38) and (5.39), we get \(\lim _{t\rightarrow \infty }I_1=0\) and (1.7). \(\square \)
Proof of Theorem 2.8
From representation theorem for functions of regular variation ([4], Theorem 1.3.1 and equation (1.5.1)), we conclude that (5.36). Assume \(q_1\ne q_2\), we chose \(\gamma <\min \{\frac{|q_1-q_2|}{2}, 1-q_1, 1-q_2\}\) in (5.36). Hence, we get \(\lim _{y\rightarrow 0} {V_2'(y)\over V'_1(y)} = 0\) if \(q_1 < q_2\) and \(\lim _{y\rightarrow 0} {V_1'(y)\over V'_2(y)} = 0\) if \(q_2 < q_1\).
We claim that, for any fixed \(t_0>0\), \(v_y(y,t_0)\) satisfies (1.9) with index \(1-\min \{q_1,q_2\}\) and (2.15). Suppose \(q_1 <q_2 \). From Potter’s bound ([4], Theorem 1.5.6.iii) (see also Back et al. [1], equation (73)), (5.6) and (5.10), for any fixed \(t_0>0\), we get, by the dominated convergence theorem, \(\lim _{y\rightarrow 0} {v_y(y,t_0)\over V'_1(y)} = e^{\lambda _1 t_0}\) and \(\lim _{y\rightarrow 0} {yv_{yy}(y,t_0)\over V'_1(y)} = (q_1-1)e^{\lambda _1 t_0}\). We have
Hence \(v_y(y,t_0)\) satisfies (1.9) with index \(1-q_1\). By (5.6), (5.10) and Potter’s bound, we deduce (2.15). Other cases can be prove similarly. We have shown that \(u_x(x,t_0)\) satisfies (1.10) with index \(1-\max \{p_1, p_2\}\). Consider the following problem: find a solution \(u=\bar{u}(x,t)\) satisfying the equation
with the initial condition \(\bar{u}(x,0)=u(x,t_0)\) for \(x\in R_+\), where t represent a time horizon variable. Note that \(u(x,t_0)\) is a utility function. Then we get \(\bar{u}(x,t)=u(x,t+t_0)\) and \(\bar{A}(x,t)=A(x,t+t_0)\) for all \(x\in R_+\) by the uniqueness of solution to the Cauchy problem of the Bellman equation. From Theorem 2.7 and the dual representation, we deduce \(\lim _{t\rightarrow \infty }\bar{A}(x,t)=\frac{\theta }{\sigma }(1-\min \{q_1,q_2\})x\), and the turnpike property (1.7) holds. \(\square \)
Rights and permissions
About this article
Cite this article
Bian, B., Zheng, H. Turnpike property and convergence rate for an investment and consumption model. Math Finan Econ 13, 227–251 (2019). https://doi.org/10.1007/s11579-018-0226-3
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11579-018-0226-3