Mean-Variance Optimization for Participating Life Insurance Contracts

Abstract

This paper studies the equity holders’ mean-variance optimal portfolio choice problem for (non-)protected participating life insurance contracts. We derive explicit formulas for the optimal terminal wealth and the optimal strategy in the multi-dimensional Black-Scholes model, showing the existence of all necessary parameters. In incomplete markets, we state Hamilton-Jacobi-Bellman equations for the value function. Moreover, we provide a numerical analysis of the Black-Scholes market. The equity holders on average increase their investment into the risky asset in bad economic states and decrease their investment over time.

Keywords: optimal portfolio, portfolio insurance, mean-variance optimization, participating life insurance, non-concave utility maximization

JEL: C61, G11, G22

1 Introduction

This paper investigates a mean-variance optimization from the perspective of the equity holders of an insurance company for the two standard designs of participating life insurance contracts, i.e., with a protected or non-protected guarantee. In a participating life insurance contract, the policyholder gets a (possibly protected) guarantee and participates proportionally at maturity from the portfolio value exceeding a pre-defined threshold higher than the guarantee. The policyholders receive at least the guarantee value at maturity if the guarantee is protected. If the guarantee is not protected, then the policyholders get, at most, the portfolio value where, initially, the insurance company gives additional equity to the premium for the investment, i.e., the equity has only limited liability.

The portfolio theory research goes back to the 1950s with the pioneering work of Markowitz [17, 18], who used variance to measure the riskiness of stock returns in a one-period setting. Mean-variance was later also analyzed in dynamic settings, see for instance Hakansson [11], Samuelson [27] or Merton [19, 20], and has been extended in various directions. For a more detailed overview of mean-variance portfolio optimizations, see the literature review from Zhang et al. [29]. In these works, the management is typically assumed to maximize the mean-variance of the entire portfolio, and no distinction between equity holders and debt holders or policyholders is made.

This paper aims to provide an explicit formula for the optimal terminal wealth (from the perspective of the equity holders of an insurance company), show that the optimal solution exists (which is non-trivial since the solution includes an additional parameter), and prove an analytical formula for the optimal strategy. Moreover, we derive a Hamilton-Jacobi-Bellman (HJB) equation for incomplete markets and discuss the characteristics of the terminal wealth and the optimal strategy in a numerical analysis.

While the optimal investment problem for participating life insurance contracts was solved for expected utility (EU), an analysis for mean-variance is lacking, which is the contribution of our paper. This analysis is relevant as mean-variance is widely spread in the industry, and most finance papers and textbooks use it as the benchmark to measure risk; see, for instance, Cochrane [5]. The reason is that mean-variance provides a straightforward interpretation of risk vs. reward, admits tractable statistical properties, and also, due to historical contingencies, has become the leading standard taught in business schools, making it easier for a risk manager to justify its use compared to specifying a utility function. Finally, due to the nonlinearity of variance, the mathematical analysis differs from EU. In particular, one needs to specify an equivalent problem which, contrary to EU-maximization, adds an additional parameter $\lambda$ to the model. The existence of such a $\lambda$ is not apparent, and we are only able to derive its existence and compute it in a semi-explicit way in the case of the Black-Scholes model.

As discussed in the beginning, the insurance company offering a participating life insurance contract pays back a surplus to the policyholder in good economic states. Insurance policies with profit participation play an essential role in the life insurance sector. According to the European Insurance Overview 2023 [9], issued by the European Insurance and Occupational Pensions Authority (EIOPA), policyholders spent 2022 around a quarter of their gross premiums on profit participation insurance policies in the life sector (includes life, health, and pension insurance). In Croatia, Italy, and Belgium, these policies have a market share of over 50 % in the life sector. Several publications are concerned with the valuation and hedging of such insurance policies, e.g., Bryis and de Varenne [3], Bacinello and Persson [1], Gatzert and Kling [10], Schmeiser and Wagner [28] or Mirza and Wagner [22]. In recent years, the study of optimal investments for participating life insurance contracts in continuous time started. Lin et al. [16] analyzed 2017 the optimal investment for an EU-problem for a specific S-shaped utility function. Afterwards, Nguyen and Stadje [23] examined a similar setting for more general S-shaped utility functions under a Value at Risk constraint and mortality risk. He et al. [12] made another generalization by considering a weighted utility function between the insurer and the policyholders. Moreover, Dong et al. [8] added to the EU-problem a Value at Risk and a portfolio insurance constraint. Chen et al. [4] considered Value at Risk, expected shortfall, and Average Value at Risk constraints. To the best of our knowledge, we are the first to consider participating insurance contracts in continuous time under mean-variance optimization.

A significant factor of an optimal policy design is the payoff structure: For a participating life insurance contract, the payoff has at least one point where it is non-differentiable, e.g., at the threshold where the proportional surplus participation starts. (The payoff of the non-protected product also has a second point of non-differentiability at the guarantee value.) At these points, the payoff changes its slope, resulting in a payoff that is generally neither convex nor concave (see Figure 2 for an illustration of the insurer’s payoff). In such situations, one often uses concavification techniques; see, for instance, Larsen [14] or Reichlin [26]. Liang et al. [15] used a generalization of this method. For alternative approaches in non-concave portfolio optimization, see, for instance, Kraft and Steffensen [13], Dai et al. [7], or Qian and Yang [25]. In this work, we use the specific structure of a payoff in a complete market to carefully compare different solution candidates and show that the optimal terminal wealth actually has a closed and, up to the implicit parameters whose existence proof and computation in our setting is delicate, simple form. We also use a standard Lagrangian approach to get the optimal terminal wealth and use the price density process to give the analytic formula for the optimal strategy. This method was used to solve several other optimization problems in complete markets; see, for instance, Basak and Shapiro [2], Cuoco et al. [6], Chen et al. [4], Nguyen and Stadje [23], or Mi et al. [21]. To use this approach for variance, we start to give an equivalent problem in the spirit of Zhou and Li [30]. The ansatz described above then leads to having two multipliers instead of only one Lagrangian multiplier. We obtain a system of non-linear equations with several variables yielding implicit functions whose properties, through various intermediate results, then entail the existence of a solution. Although mean-variance does not respect first-order stochastic dominance, we do get similar results compared to EU-optimization regarding the general form of the optimal solution, but with the terminal wealth having a simpler structure. In particular, in a complete market, the optimal terminal wealth is piecewise linear in the price density. On the other hand, in an incomplete market, the optimal solution itself can only be characterized implicitly as the (viscosity) solution of a certain PDE. Finally, our numerical results show a somewhat different investment behavior compared to EU-maximization. Specifically, we observe that the investment gets more conservative with shorter maturities, and the equity holders on average increase their investment into the risky asset in bad economic states. Moreover, the insurer invests more riskily when offering a non-protected participation life insurance product than when offering a protected one due to a reduced downside potential. Since variance is the most widely used risk measure in industry and the cornerstone of most modern portfolio theory and the finance literature, these results might be potentially relevant from a descriptive point of view.

Section 2 describes participating life insurance contracts and briefly introduces the functional to optimize. In Section 3, we explicitly show the optimal terminal wealth, the optimal strategy, and the existence of the necessary parameters in the Black-Scholes model. In Section 4, we give the SDEs for possibly incomplete markets, and Section 5 analyzes some numerical results for the Black-Scholes model. Finally, Section 6 concludes the paper.

2 Model Setup

2.1 Participating life insurance contracts

Participating insurance contracts or, in general, insurance contracts with some profit participation play a crucial role in the life sector. Figure 1 shows that, on average, around a quarter of the gross premiums 2022 in the life sector are spent on policies with profit participation. In some countries, like Croatia, Belgium, or Italy, even more than 50 % of the gross premiums are invested in such policies.

Figure 1: Market share in 2022 of the gross premium separated by the line of business in the life sector. Data source: European Insurance Overview from the EIOPA [9]

This paper focuses on two standard designs of participating life insurance contracts. Both products offer a guarantee value $G$ and a proportional surplus participation rate $\alpha_{2}$ when the portfolio value exceeds a threshold, which we denote by $k_{2}$. Note that $k_{2}$ is always higher or equal to $G$. The difference between the two designs is that the first product offers a non-protected guarantee, whereas the second provides a protected one. Protected means, in this case, that the policyholders get at least their guarantee value, independent of the economic situation at maturity. In contrast, in the non-protected case, the insurance company declares bankruptcy if the portfolio value is below $G$. In this case, the policyholders only get the portfolio value. For the portfolio, it is essential to note that the initial portfolio value $x_{0}$ is the sum of the premiums from the policyholders plus some initial capital from the insurer (i.e., the equity holders). There is no rule on how to set the guarantee value. One possible example of setting $G$ is to take the sum of the premiums in addition to a guaranteed interest rate below the risk-free interest rate. Moreover, product designers often set the threshold $k_{2}$ as the sum of the premiums divided by the share of the policyholders in the portfolio.

Hence, we conclude the following payoffs $V$ for the policyholders (pol) and the insurer (ins) for the contracts with non-protected (non) resp. protected (pro) guarantees and terminal portfolio value $X_{T}\geq 0$:

$$V_{\scriptsize\text{pol}}^{\scriptsize\text{non}}(X_{T}) =\begin{cases}X_{T}&\text{if }X_{T}<G,\\ G&\text{if }G\leq X_{T}<k_{2},\\ G+\alpha_{2}(X_{T}-k_{2})&\text{if }X_{T}\geq k_{2},\end{cases}$$

$$V_{\scriptsize\text{pol}}^{\scriptsize\text{pro}}(X_{T}) =\begin{cases}G&\text{if }X_{T}<k_{2},\\ G+\alpha_{2}(X_{T}-k_{2})&\text{if }X_{T}\geq k_{2},\end{cases}$$

$$V_{\scriptsize\text{ins}}^{\scriptsize\text{non}}(X_{T}) =\begin{cases}0&\text{if }X_{T}<G,\\ X_{T}-G&\text{if }G\leq X_{T}<k_{2},\\ X_{T}-G-\alpha_{2}(X_{T}-k_{2})&\text{if }X_{T}\geq k_{2},\end{cases}$$

$$V_{\scriptsize\text{ins}}^{\scriptsize\text{pro}}(X_{T}) =\begin{cases}X_{T}-G&\text{if }X_{T}<k_{2},\\ X_{T}-G-\alpha_{2}(X_{T}-k_{2})&\text{if }X_{T}\geq k_{2}.\end{cases}$$

Note that solely $V_{\scriptsize\text{ins}}^{\scriptsize\text{pro}}$ can attain negative values. For a more detailed explanation of participating insurance, we refer to Nguyen and Stadje [23].

2.2 Optimization Functional

Before defining the optimization functional, let us introduce the basic financial market, which in Section 3 and 5 will become the Black-Scholes model.
Let $(\Omega,\mathcal{F},(\mathcal{F})_{t\in[0,T]},\mathbb{P})$ be a filtered probability space with time horizon $T>0$. The filtration is generated by the $d$-dimensional Brownian Motion $W$ satisfying the usual conditions. We consider an arbitrage-free market with a risk-free asset $B$ and a deterministic interest rate $r_{t}\geq 0$ following the price process $\mathrm{d}B_{t}=B_{t}r_{t}\mathrm{d}t$, and $d$ risky assets $S^{i}$, $i\in\{1,\ldots,d\}$ with adapted price processes. Let the dynamic strategy $u$ represent the fraction of wealth invested in the corresponding risky asset, while the remaining money is invested in the risk-free asset with the corresponding wealth process $X_{t}$ and initial value $X_{0}=x_{0}$. We denote by $\mathcal{U}$ the set of all admissible strategies given by all $u$, which are progressively measurable and induce an integrable $X_{T}$.

Next, we introduce a general functional to optimize, the previously defined participating life insurance products being special cases. We look for the optimal strategy $\hat{u}\in\mathcal{U}$ such that

$$J(0,T,\hat{u},x_{0})=\sup_{u\in\mathcal{U}}J(0,T,u,x_{0}),$$ (2.1)

where the value functional $J$ is defined as

$$J(0,T,u,x_{0}):=\mathbb{E}[F(0,T,u,x_{0})]-\gamma\text{Var}(F(0,T,u,x_{0}))$$

with the risk aversion parameter $\gamma>0$. Hence, we are looking for the mean-variance optimal strategy for the (continuous) function $F$, which we define for $s<t$ as

$$F(s,t,u,x):= \ \alpha\left((X_{t}-k_{1})_{+}-k_{0}\right)-\alpha_{2}(X_{t}-k_{2})_{+}$$

$$= \ \begin{cases}-\alpha k_{0}&X_{t}<k_{1},\\ \alpha(X_{t}-k_{1}-k_{0})&k_{1}\leq X_{t}<k_{2},\\ \tilde{\alpha}(X_{t}-k_{2})+\alpha(k_{2}-k_{1}-k_{0})&X_{t}\geq k_{2},\end{cases}$$

where $X_{s}=x$, $0\leq k_{0},k_{1}\leq k_{2}<\infty$ with $k_{0}+k_{1}\leq k_{2}$, $0\leq\alpha_{2}<\alpha<\infty$ with $\tilde{\alpha}:=\alpha-\alpha_{2}$. Note that $0<\tilde{\alpha}\leq\alpha$ holds. We also write $F(X_{T})$ instead of $F(0,T,u,x)$ when the trading strategy is clear (for instance, for the optimal terminal wealth $\hat{X}_{T}$), suppressing the initial value.

Now, one can observe that if $\alpha=1$, $k_{0}=0$, and $k_{1}=G$, the function $F$ reduces to the payoff of the insurer for the non-protected participating life insurance contract $V_{\scriptsize\text{ins}}^{\scriptsize\text{non}}$. If $\alpha=1$, $k_{0}=G$, and $k_{1}=0$, $F$ reduces to the payoff of the insurer for the protected product $V_{\scriptsize\text{ins}}^{\scriptsize\text{pro}}$. In the following Figure 2, we show the payoff of the insurer for these two variants of participating life insurance contracts.

Figure 2: Payoffs for the insurer with $k_{1}=2.5$ (left) resp. $k_{0}=2.5$ (right), $k_{2}=7$, $\alpha=1$, $\alpha_{2}=0.5$.

3 Optimization in a Black-Scholes market

This section assumes that the underlying financial market follows the Black-Scholes model, i.e., the market is complete. Then, the price dynamics for the $d$ risky assets $S^{i}$, $i\in\{1,\ldots,d\}$ are given by

$$\mathrm{d}S^{i}_{t} =S^{i}_{t}\mu^{i}_{t}\mathrm{d}t+S^{i}_{t}\sigma^{i}_{t}\mathrm{d}W_{t},$$

where $W$ is the Brownian Motion generating the filtration $\mathcal{F}$, $\mu_{i}$ is the deterministic drift of the $i$’th asset and $\sigma^{ij}$ is the deterministic volatility between the $i$’th and the $j$’th asset. We assume that $\sigma_{t}$ is bounded, bounded away from zero and invertible. Then, there exists a unique price density process $\xi$ with the wealth process $X$ and the price density process $\xi$ admitting the following dynamics

$$\mathrm{d}X_{t} =X_{t}\left[r_{t}+u_{t}^{T}(\mu_{t}-r)\right]\mathrm{d}t+X_{t}u_{t}^{T}\sigma_{t}\mathrm{d}W_{t},$$ (3.1)

$$\mathrm{d}\xi_{t} =-\xi_{t}r_{t}\mathrm{d}t-\xi_{t}\kappa_{t}^{T}\mathrm{d}W_{t},$$ (3.2)

with $X_{0}=x_{0}$ and $\xi_{0}=1$, where $\kappa_{t}=(\sigma_{t})^{-1}(\mu_{t}-r_{t})$ is the Sharpe ratio process and $\cdot^{T}$ denotes the transpose of a vector. The term $\xi_{T}(\omega)$, $\omega\in\Omega$, can be interpreted as the Arrow-Debreu value per probability unit in state $\omega$ at time $T$. Note that $\xi_{T}(\omega)$ can be written as a decreasing function of the stock price, and therefore attains high values in times of a bad economy and low values in times of a good economy. We assume that the processes $r_{t}$ and $\mu_{t}$ are integrable and the processes $\sigma_{t}$ and $\kappa_{t}$ are square-integrable over $[0,T]$ to ensure that the previous SDEs and all of the following integrals are well-defined.

3.1 Derivation of the optimal terminal wealth

The mean-variance optimization (2.1) is challenging to solve directly due to the term $(\mathbb{E}[X])^{2}$ in the decomposition formula $\text{Var}(X)=\mathbb{E}[X^{2}]-(\mathbb{E}[X])^{2}$. Hence, we show in the following Lemma 3.1 that if an optimal strategy exists, we can alternatively look for the optimal strategy considering the value functional $\tilde{J}$ which we define as:

$$\tilde{J}(0,T,u,x_{0}):=\mathbb{E}[\lambda F(0,T,u,x_{0})-\gamma F(0,T,u,x_{0})^{2}]$$ (3.3)

with $\lambda=1+2\gamma\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]$ where $\hat{u}$ is the optimal strategy.

Lemma 3.1.

If $\hat{u}$ is an optimal strategy for $J$, it is also an optimal strategy for $\tilde{J}$.

Consequently, the optimal terminal wealth also coincides for the two problems maximizing $J$ resp. $\tilde{J}$. This result is a slight generalization of Theorem 3.1 in Zhou and Li [30], who showed this lemma in the case of $F$ being the identity. We will include the proof in Appendix A for the reader’s convenience.

Since we have a complete market, we optimize with the following three steps. First, we use a Lagrangian approach to find the optimal terminal wealth $\hat{X}$ (using the alternative problem). Second, we derive the optimal strategy, and third, we determine the $\lambda$ and the Lagrangian multiplier $y$ since the optimal terminal wealth and strategy depend on these values initially. In the following, we suppress the dependence on $\lambda$ and $y$ for the sake of simplicity in the notation unless stated otherwise in some proofs, and we use the convention that $(a,a]=\emptyset$ and $[a,b]=\emptyset$ if $b<a$.

Theorem 3.2.

The optimal terminal wealth $\hat{X}_{T}$ is given by:

$$\hat{X}_{T}:=\begin{cases}k_{2}+\displaystyle\frac{\lambda\tilde{\alpha}-y\xi_{T}}{2\gamma\tilde{\alpha}^{2}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})&\xi_{T}\in(0,\xi_{1}^{*}],\\ k_{2}&\xi_{T}\in(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}],\\ k_{0}+k_{1}+\displaystyle\frac{\lambda\alpha-y\xi_{T}}{2\gamma\alpha^{2}}&\xi_{T}\in(\alpha\hat{\xi},\xi_{3}^{*}],\\ 0&\text{else,}\end{cases}$$ (3.4)

where $y$ is the Lagrangian multiplier which solves $\mathbb{E}[\xi_{T}\hat{X}_{T}(y)]=\xi_{0}x_{0}$, $\lambda=1+2\gamma\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]$, and

$$\hat{\xi} :=\max\left\{0,\displaystyle\frac{\lambda-2\gamma\alpha(k_{2}-k_{1}-k_{0})}{y}\right\},$$

$$\bar{\xi} :=\displaystyle\frac{\lambda\alpha}{y}+\displaystyle\frac{2\gamma\alpha^{2}k_{0}}{y},$$

$$\tilde{\xi}_{1}^{*} :=\tilde{\alpha}\hat{\xi}-\displaystyle\frac{2\gamma\tilde{\alpha}}{y}\left(\sqrt{\max\left\{0,(\alpha(k_{0}+k_{1})-\alpha_{2}k_{2})^{2}-\alpha^{2}k_{0}^{2}+\displaystyle\frac{\lambda}{\gamma}(\alpha k_{1}-\alpha_{2}k_{2})\right\}}-\tilde{\alpha}k_{2}\right),$$

$$\xi_{1}^{*} :=\max\left\{0,\min\left\{\tilde{\alpha}\hat{\xi},\tilde{\xi}_{1}^{*}\right\}\right\},$$

$$\tilde{\xi}_{2}^{*} :=\displaystyle\frac{\alpha\lambda}{y}-\displaystyle\frac{\gamma\alpha^{2}(k_{2}-k_{1})^{2}-2\gamma\alpha^{2}k_{0}(k_{2}-k_{1})+\lambda\alpha k_{1}}{yk_{2}},$$

$$\xi_{2}^{*} :=\max\left\{\tilde{\alpha}\hat{\xi},\min\left\{\alpha\hat{\xi},\tilde{\xi}_{2}^{*}\right\}\right\},$$

$$\xi_{3}^{*} :=\max\left\{\alpha\hat{\xi},\bar{\xi}-\displaystyle\frac{2\gamma\alpha^{2}}{y}\left(\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+\displaystyle\frac{\lambda}{\gamma\alpha}\right)}-k_{1}\right)\right\}.$$

In particular, such $\hat{u},t,y$ exist. Moreover, let $\xi^{*}$ be defined by:

$$\xi^{*}:=\begin{cases}\xi_{3}^{*}&\text{,if }\xi_{3}^{*}>\alpha\hat{\xi},\\ \xi_{2}^{*}&\text{,if }\xi_{3}^{*}=\alpha\hat{\xi},\xi_{2}^{*}>\tilde{\alpha}\hat{\xi},\\ \xi_{1}^{*}&\text{,if }\xi_{3}^{*}=\alpha\hat{\xi},\xi_{2}^{*}=\tilde{\alpha}\hat{\xi}.\end{cases}$$

Then, it holds that $\xi^{*}>0$ and $\hat{X}_{T}>0$ for $\xi\in(0,\xi^{*})$ and $\hat{X}_{T}=0$ for $\xi>\xi^{*}$.

Remark 3.3.

1. The terminal wealth $\hat{X}_{T}$ denotes the wealth before distributing the wealth to the insurer and the policyholders. The terminal wealth of the insurer is given by $F(\hat{X}_{T})=\alpha((\hat{X}_{T}-k_{1})_{+}-k_{0})-\alpha_{2}(\hat{X}_{T}-k_{2})_{+}$. In particular, in the case of a protected participating life insurance contract, i.e., $\alpha=1$, $k_{1}=0$, the insurer makes a loss if $\hat{X}_{T}<k_{0}$. In the case of a non-protected participating life insurance contract, the insurer cannot make a loss by construction.

2. If $\alpha_{2}=\alpha$ resp. $\tilde{\alpha}=0$, the surplus over $k_{2}$ is fully distributed to the policyholders. In this case, it holds that $\tilde{\xi}_{1}^{*}=0$ and $\tilde{\alpha}\hat{\xi}=0$, i.e., the result is similar with the exception that the first case, $\hat{X}_{T}=k_{2}+\frac{\lambda\tilde{\alpha}-y\xi_{T}}{2\gamma\tilde{\alpha}^{2}}-\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})$, does not need to be considered. The proofs are similar, but we exclude this case for the ease of exposition.

3. We can generalize the result for $X_{T}\in\mathbb{R}$ when we extend the function $F$ to $\mathbb{R}_{<0}$ as $F(X_{T})=\alpha(X_{T}-k_{0})$. Then, the optimal terminal wealth has the same structure as before, but with the additional case that $\hat{X}_{T}=k_{0}+\frac{\lambda\alpha-y\xi_{T}}{2\gamma\alpha^{2}}$ if $\xi_{T}>\bar{\xi}$. Hence, the restriction to $X_{T}\geq 0$ is not crucial for the structure of the solution. Note that $F(\hat{X}_{T}(\xi_{T}=\bar{\xi}))=0$, and that for $\xi_{T}>\bar{\xi}$ the function $F$ is linearly decreasing in $\xi_{T}$. Again, the proofs are similar to before, but we nevertheless restrict to $X_{T}\geq 0$ to avoid too technical proofs.

4. Note that the representation of $\hat{X}_{T}$ also holds in a general complete market (without necessarily using a Black-Scholes market model) if we assume the existence of the variables $y$ and $\lambda$. However, that these variables exist is then not clear.

5. The way how to compute the parameters $\lambda$ and $y$ is given in the proof of Proposition A.1.

From the definition formula of $\hat{X}_{T}$, we can derive the following two propositions, which are proven in Appendix A:

Proposition 3.4.

It holds that $(0,\xi_{1}^{*}]\cup(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]\cup(\alpha\hat{\xi},\xi_{3}^{*}]=(0,\xi^{*}]$ with these constants defined in Theorem 3.2, i.e., the three intervals of the terminal wealth are connected.

Proposition 3.5.

It holds that $\hat{X}_{T}$ as a function of $\xi$ is continuous and non-increasing in $(0,\xi^{*})\cup(\xi^{*},\infty)$ with $\xi^{*}$ as in Theorem 3.2. If $k_{1}>0$, then $\hat{X}_{T}$ is always discontinuous at $\xi^{*}$.

Therefore, $\hat{X}_{T}$ as a function of $\xi_{T}$ is non-increasing, taking the value $0$ for $\xi_{T}>\xi^{*}$ and one possible discontinuity point. See the black lines in Figure 3 for a visualization.

• We use the Lagrangian multiplier method to prove this theorem. Therefore, choose $\lambda>0$ and $y>0$ as in Proposition A.1 in the appendix and we define the Lagrangian function $L$ for $X\geq 0$ as

  $$L(X,y):= \ \lambda F(0,T,u,x_{0})-\gamma F(0,T,u,x_{0})^{2}-y\xi X$$ (3.5)

  $$= \ \lambda\left[\alpha\left((X-k_{1})_{+}-k_{0}\right)-\alpha_{2}(X-k_{2})_{+}\right]-\gamma\left[\alpha\left((X-k_{1})_{+}-k_{0}\right)-\alpha_{2}(X-k_{2})_{+}\right]^{2}-y\xi X$$

  $$= \ \lambda\left[\alpha(X-k_{1}-k_{0})\mathbbm{1}_{k_{1}\leq X<k_{2}}+\tilde{\alpha}(X-k_{2})\mathbbm{1}_{X\geq k_{2}}+\alpha(k_{2}-k_{1}-k_{0})\mathbbm{1}_{X\geq k_{2}}-\alpha k_{0}\mathbbm{1}_{X<k_{1}}\right]$$

  $$-\gamma\left[\alpha(X-k_{1}-k_{0})\mathbbm{1}_{k_{1}\leq X<k_{2}}+\tilde{\alpha}(X-k_{2})\mathbbm{1}_{X\geq k_{2}}+\alpha(k_{2}-k_{1}-k_{0})\mathbbm{1}_{X\geq k_{2}}-\alpha k_{0}\mathbbm{1}_{X<k_{1}}\right]^{2}$$

  $$-y\xi X,$$

  $$= \ \lambda\alpha(X-k_{1}-k_{0})\mathbbm{1}_{k_{1}\leq X<k_{2}}+\lambda\left[\tilde{\alpha}(X-k_{2})+\alpha(k_{2}-k_{1}-k_{0})\right]\mathbbm{1}_{X\geq k_{2}}-\lambda\alpha k_{0}\mathbbm{1}_{X<k_{1}}$$

  $$-\gamma\alpha^{2}(X-k_{1}-k_{0})^{2}\mathbbm{1}_{k_{1}\leq X<k_{2}}-\gamma\left[\tilde{\alpha}(X-k_{2})+\alpha(k_{2}-k_{1}-k_{0})\right]^{2}\mathbbm{1}_{X\geq k_{2}}-\gamma\alpha^{2}k_{0}^{2}\mathbbm{1}_{X<k_{1}}$$

  $$-y\xi X$$

  $$= \begin{cases}-\lambda\alpha k_{0}-\gamma\alpha^{2}k_{0}^{2}-y\xi X&X\in[0,k_{1}),\\ \lambda\alpha(X-k_{1}-k_{0})-\gamma\alpha^{2}(X-k_{1}-k_{0})^{2}-y\xi X&X\in[k_{1},k_{2}),\\ \lambda\left[\tilde{\alpha}(X-k_{2})+\alpha(k_{2}-k_{1}-k_{0})\right]-\gamma\left[\tilde{\alpha}(X-k_{2})+\alpha(k_{2}-k_{1}-k_{0})\right]^{2}-y\xi X&X\in[k_{2},\infty),\end{cases}$$

  where we suppress the $\omega$, write $X$ instead of $X_{T}$, $\xi$ instead of $\xi_{T}$, and use $y$ as the multiplier.
  The following paragraph considers $L$ as a function of $X$. Obviously, $L$ is not smooth in $k_{1}$ and $k_{2}$, but continuous in $[0,\infty)$. Hence, we optimize $L$ in the regions $[0,k_{1}]$, $[k_{1},k_{2}]$, and $[k_{2},\infty)$ separately and compare afterwards the minimal points. Now, for the individual optimization, we get:

  $$\displaystyle\frac{\partial L}{\partial X}=\begin{cases}-y\xi&X\in(0,k_{1}),\\ \lambda\alpha-2\gamma\alpha^{2}(X-k_{1}-k_{0})-y\xi&X\in(k_{1},k_{2}),\\ \lambda\tilde{\alpha}-2\gamma\tilde{\alpha}\left[\tilde{\alpha}(X-k_{2})+\alpha(k_{2}-k_{1}-k_{0})\right]-y\xi&X\in(k_{2},\infty).\end{cases}$$

  Hence, it follows that we get the following possible maximal points (if they are in the respective interval): $X_{1}=0$, $X_{2}=k_{0}+k_{1}+\frac{\lambda\alpha-y\xi}{2\gamma\alpha^{2}}$, and $X_{3}=k_{2}+\frac{\lambda\tilde{\alpha}-y\xi}{2\gamma\tilde{\alpha}^{2}}-\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})$. To have $X_{2}$ and $X_{3}$ in the correct interval, we get the following conditions (since $\xi>0$):

  $$X_{2}\in[k_{1},k_{2}] \Leftrightarrow \xi\in\left[\displaystyle\frac{\lambda\alpha-2\gamma\alpha^{2}(k_{2}-k_{1}-k_{0})}{y},\displaystyle\frac{\lambda\alpha}{y}+\displaystyle\frac{2\gamma\alpha^{2}k_{0}}{y}\right]=\left[\alpha\hat{\xi},\bar{\xi}\right],$$ (3.6)

  $$X_{3}\in[k_{2},\infty) \Leftrightarrow \xi\in\left(0,\displaystyle\frac{\lambda\tilde{\alpha}-2\gamma\tilde{\alpha}\alpha(k_{2}-k_{1}-k_{0})}{y}\right]=\left(0,\tilde{\alpha}\hat{\xi}\right].$$ (3.7)

  Moreover, for the maximum of $L$, we must consider the different boundary values of the corresponding intervals. The first boundary value $k_{1}$ is always dominated by $X_{1}$ due to $L$ being non-increasing in $[0,k_{1}]$. The second boundary value $k_{2}$ is dominated by $X_{2}$ or by $X_{3}$ if $\xi$ is in one of the intervals from (3.6) or (3.7). If $\xi>\bar{\xi}$, then $L$ is non-increasing in $[0,\infty)$ since then $\frac{\partial L}{\partial X}\leq 0$ for all $X\geq 0$ and hence $k_{2}$ is dominated by $X_{1}$. If $\xi\in[\tilde{\alpha}\hat{\xi},\alpha\hat{\xi}]$, then $L$ has a local maximum in $k_{2}$ since $\frac{\partial L}{\partial X}\geq 0$ for $X\in(k_{1},k_{2})$ and $\frac{\partial L}{\partial X}\leq 0$ for $X>k_{2}$. Hence, we add

  $$X_{4}=k_{2}\quad\text{for}\quad\xi\in[\tilde{\alpha}\hat{\xi},\alpha\hat{\xi}]$$ (3.8)

  to the list of potential maximum points. We note that the intervals of $\xi$ for $X_{2}$, $X_{3}$, and $X_{4}$ are disjoint except the lower (resp. upper) boundary points of $X_{2}$ (resp. $X_{3}$) with the upper (resp. lower) boundary point of $X_{4}$. However, in this case $X_{2}$ (resp. $X_{3}$) coincides with $X_{4}$. It is remarkable that due to the disjointness of the related intervals for $X_{2}$, $X_{3}$, and $X_{4}$, we do not have to compare these values themselves with each other (for the potential maximum of $L$). Thus, we replace the closed intervals for $\xi$ by left-open, right-closed intervals and denote by $X^{*}$ the combined solution of $X_{2}$, $X_{3}$, and $X_{4}$ with $X^{*}=0$ for $\xi>\bar{\xi}$, i.e.,

  $$X^{*}:=\begin{cases}k_{2}+\displaystyle\frac{\lambda\tilde{\alpha}-y\xi}{2\gamma\tilde{\alpha}^{2}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})&\xi\in(0,\tilde{\alpha}\hat{\xi}],\\ k_{2}&\xi\in(\tilde{\alpha}\hat{\xi},\alpha\hat{\xi}],\\ k_{0}+k_{1}+\displaystyle\frac{\lambda\alpha-y\xi}{2\gamma\alpha^{2}}&\xi\in(\alpha\hat{\xi},\bar{\xi}],\\ 0&\xi\in(\bar{\xi},\infty).\end{cases}$$

  Note that $X^{*}$ takes the same values as $\hat{X}_{T}$ in (3.4), but the intervals are not truncated. Moreover, we observe that $X^{*}$ is continuous in $\xi$ on $(0,\bar{\xi}]$. Therefore, we only have to compare $L(X^{*},y)$, $y$ fixed, with $L(X_{1},y)=-\gamma\alpha^{2}k_{0}^{2}-\lambda\alpha k_{0}$. Thus, we compare these values in the following separately for the three cases, i.e., depending on $\xi$, which value out of $\{X_{2},X_{3},X_{4}\}$ $X^{*}$ attains. Note that we consider, from now on, $L$ as a function of $\xi$. To emphasize this, we write $L(X,y,\xi)$.

  Case 1: $\xi\in(\alpha\hat{\xi},\bar{\xi}]$, i.e., $X^{*}=X_{2}\in[k_{1},k_{2}]$:
  Then, we get:

  $$L(X_{2},y,\xi) =\lambda\alpha\displaystyle\frac{\lambda\alpha-y\xi}{2\gamma\alpha^{2}}-\gamma\alpha^{2}\left(\displaystyle\frac{\lambda\alpha-y\xi}{2\gamma\alpha^{2}}\right)^{2}-y\xi(k_{0}+k_{1})-y\xi\displaystyle\frac{\lambda\alpha-y\xi}{2\gamma\alpha^{2}}$$

  $$=\displaystyle\frac{y^{2}}{4\gamma\alpha^{2}}\xi^{2}-y\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)\xi+\displaystyle\frac{\lambda^{2}}{4\gamma}.$$ (3.9)

  Then $L(X_{2},y,\xi)\geq-\gamma\alpha^{2}k_{0}^{2}-\lambda\alpha k_{0}=L(X_{1},y,\xi)$ if $\xi\leq\xi_{-}$ or $\xi\geq\xi_{+}$ with

  $$\xi_{\pm} =\displaystyle\frac{2\gamma\alpha^{2}}{y^{2}}\left(y\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)\pm\sqrt{y^{2}\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)^{2}-\displaystyle\frac{y^{2}}{\gamma\alpha^{2}}\left(\displaystyle\frac{\lambda^{2}}{4\gamma}+\gamma\alpha^{2}k_{0}^{2}+\lambda\alpha k_{0}\right)}\right)$$

  $$=\displaystyle\frac{\lambda\alpha}{y}+\displaystyle\frac{2\gamma\alpha^{2}(k_{0}+k_{1})}{y}\pm\displaystyle\frac{2\gamma\alpha^{2}}{y^{2}}\sqrt{y^{2}k_{1}^{2}+\displaystyle\frac{y^{2}k_{1}\lambda}{\gamma\alpha}+2y^{2}k_{0}k_{1}}$$

  $$=\displaystyle\frac{\lambda\alpha}{y}+\displaystyle\frac{2\gamma\alpha^{2}(k_{0}+k_{1})}{y}\pm\displaystyle\frac{2\gamma\alpha^{2}}{y}\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+\displaystyle\frac{\lambda}{\gamma\alpha}\right)}$$

  $$=\bar{\xi}+\displaystyle\frac{2\gamma\alpha^{2}}{y}\left(\pm\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+\displaystyle\frac{\lambda}{\gamma\alpha}\right)}+k_{1}\right).$$

  We notice that $\xi_{+}>\bar{\xi}$ and $\max\left\{\alpha\hat{\xi},\xi_{-}\right\}=\xi_{3}^{*}$ where $\xi_{3}^{*}$ is defined as in Theorem 3.2. If $\xi_{-}<\alpha\hat{\xi}$, $X_{2}$ is not in the necessary interval $[k_{1},k_{2})$ which entails that $(\alpha\hat{\xi},\xi_{3}^{*}]=\emptyset$. Otherwise, it holds that $\xi_{-}=\xi_{3}^{*}$ by which we get (3.4) for this case, i.e., the interval for $\xi$ being in the interval $(\alpha\hat{\xi},\xi_{3}^{*}]$ for the optimal wealth $\hat{X}_{T}$.

  Case 2: $\xi\in(0,\tilde{\alpha}\hat{\xi}]$, i.e., $X^{*}=X_{3}\in[k_{2},\infty)$:
  Then it holds:

  $$L(X_{3},y,\xi)= \ \lambda\left(\displaystyle\frac{\lambda\tilde{\alpha}-y\xi}{2\gamma\tilde{\alpha}}-\tilde{\alpha}\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})+\alpha(k_{2}-k_{1}-k_{0})\right)$$

  $$-\gamma\left(\tilde{\alpha}\displaystyle\frac{\lambda\tilde{\alpha}-y\xi}{2\gamma\tilde{\alpha}^{2}}-\tilde{\alpha}\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})+\alpha(k_{2}-k_{1}-k_{0})\right)^{2}$$

  $$-y\xi k_{2}-y\xi\displaystyle\frac{\lambda\tilde{\alpha}-y\xi}{2\gamma\tilde{\alpha}^{2}}+y\xi\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})$$

  $$= \ \displaystyle\frac{y^{2}}{4\gamma\tilde{\alpha}^{2}}\xi^{2}-y\left(k_{2}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}\right)\xi+\displaystyle\frac{\lambda^{2}}{4\gamma}.$$ (3.10)

  Note that $k_{2}-\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})=\frac{\alpha}{\tilde{\alpha}}(k_{0}+k_{1})-\frac{\alpha_{2}}{\tilde{\alpha}}k_{2}$. First we assume that the discriminant, i.e., the term under the square root in the following equation, is non-negative. Then, we have $L(X_{3},y,\xi)\geq-\gamma\alpha^{2}k_{0}^{2}-\lambda\alpha k_{0}=L(X_{1},y,\xi)$ if $\xi\leq\xi_{-}$ or $\xi\geq\xi_{+}$ with

  $$\xi_{\pm}= \displaystyle\frac{2\gamma\tilde{\alpha}^{2}}{y^{2}}\left(y\left(k_{2}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}\right)\right)$$

  $$\pm\displaystyle\frac{2\gamma\tilde{\alpha}^{2}}{y^{2}}\sqrt{y^{2}\left(k_{2}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}\right)^{2}-\displaystyle\frac{y^{2}}{\gamma\tilde{\alpha}^{2}}\left(\displaystyle\frac{\lambda^{2}}{4\gamma}+\gamma\alpha^{2}k_{0}^{2}+\lambda\alpha k_{0}\right)}$$

  $$= \displaystyle\frac{2\gamma\tilde{\alpha}^{2}}{y^{2}}\left(y\left(k_{2}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}\right)\right)$$

  $$\pm\displaystyle\frac{2\gamma\tilde{\alpha}^{2}}{y^{2}}\sqrt{y^{2}\left(\frac{\alpha^{2}}{\tilde{\alpha}^{2}}(2k_{0}k_{1}+k_{1}^{2})+\frac{\alpha_{2}^{2}}{\tilde{\alpha}^{2}}k_{2}^{2}-2\frac{\alpha\alpha_{2}}{\tilde{\alpha}^{2}}(k_{0}+k_{1})k_{2}+\frac{\lambda}{\gamma\tilde{\alpha}^{2}}(\alpha k_{1}-\alpha_{2}k_{2})\right)}$$

  $$= \displaystyle\frac{\lambda\tilde{\alpha}}{y}-\displaystyle\frac{2\gamma\alpha\tilde{\alpha}(k_{2}-k_{1}-k_{0})}{y}+\displaystyle\frac{2\gamma\tilde{\alpha}^{2}k_{2}}{y}$$

  $$\pm\displaystyle\frac{2\gamma\tilde{\alpha}}{y}\sqrt{\alpha^{2}(2k_{0}k_{1}+k_{1}^{2})+\alpha_{2}^{2}k_{2}^{2}-2\alpha\alpha_{2}(k_{0}+k_{1})k_{2}+\frac{\lambda}{\gamma}(\alpha k_{1}-\alpha_{2}k_{2})}$$

  $$= \tilde{\alpha}\hat{\xi}+\displaystyle\frac{2\gamma\tilde{\alpha}}{y}\left(\pm\sqrt{(\alpha(k_{0}+k_{1})-\alpha_{2}k_{2})^{2}-\alpha^{2}k_{0}^{2}+\frac{\lambda}{\gamma}(\alpha k_{1}-\alpha_{2}k_{2})}+\tilde{\alpha}k_{2}\right).$$

  Again, we notice that $\xi_{+}>\tilde{\alpha}\hat{\xi}$ and if the discriminant is non-negative, we get $\xi_{-}=\tilde{\xi}_{1}^{*}$ where we can ignore this case if $\xi_{-}\leq 0$ due to the assumption that $\xi$ has to be in the interval $(0,\tilde{\alpha}\hat{\xi}]$. Now, if the discriminant is negative, then $L(X_{3},y,\xi)\geq-\gamma\alpha^{2}k_{0}^{2}-\lambda\alpha k_{0}$ for all $\xi$ and $\tilde{\xi}_{1}^{*}>\tilde{\alpha}\hat{\xi}$. Hence, $X_{3}$ is optimal for all $\xi\in(0,\tilde{\alpha}\hat{\xi}]$ and it holds that $\xi_{1}^{*}=\tilde{\alpha}\hat{\xi}$ which implies (3.4) for this case.

  Case 3: $\xi\in(\tilde{\alpha}\hat{\xi},\alpha\hat{\xi}]$, i.e., $X^{*}=X_{4}=k_{2}$:
  Now, we get:

  $$L(X_{4},y,\xi) =\lambda\alpha(k_{2}-k_{1}-k_{0})-\gamma\alpha^{2}(k_{2}-k_{1}-k_{0})^{2}-y\xi k_{2}.$$ (3.11)

  Then $L(X_{4},y,\xi)\geq-\gamma\alpha^{2}k_{0}^{2}-\lambda\alpha k_{0}=L(X_{1},y,\xi)$ if

  $$\xi\leq\displaystyle\frac{\lambda\alpha}{y}-\displaystyle\frac{\gamma\alpha^{2}(k_{2}-k_{1})^{2}-2\gamma\alpha^{2}k_{0}(k_{2}-k_{1})+\lambda\alpha k_{1}}{yk_{2}}=\tilde{\xi}_{2}^{*}.$$

  Hence, it follows that $\xi\in(\tilde{\alpha}\hat{\xi},\tilde{\xi}_{2}^{*}]$ which implies (3.4) in this case as well. Hence, (3.4) holds in total.

  Finally, we must have that $\xi^{*}>0$ since otherwise $(0,\xi_{1}^{*}]\cup(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]\cup(\alpha\hat{\xi},\xi_{3}^{*}]=\emptyset$ which implies that $\hat{X}_{T}\equiv 0$. Hence, the budget constraint $\mathbb{E}[\xi_{T}\hat{X}_{T}(y)]=\xi_{0}x_{0}$ cannot be fulfilled, which is a contradiction to Proposition A.1 whose proof is given below. The property of $\hat{X}_{T}$ follows directly from Proposition 3.4 and $y<\infty$. ∎

We deduce from Theorem 3.2:

Corollary 3.6.

The optimal payoff of the insurer is given by:

$$F(\hat{X}_{T})=\begin{cases}\displaystyle\frac{\lambda\tilde{\alpha}-y\xi_{T}}{2\gamma\tilde{\alpha}}&\xi_{T}\in(0,\xi_{1}^{*}],\\ \alpha(k_{2}-k_{1}-k_{0})&\xi_{T}\in(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}],\\ \displaystyle\frac{\lambda\alpha-y\xi_{T}}{2\gamma\alpha}&\xi_{T}\in(\alpha\hat{\xi},\xi_{3}^{*}],\\ -\alpha k_{0}&\text{else,}\end{cases}$$

where $\hat{\xi}$, $\xi_{1}^{*}$, $\xi_{2}^{*}$, and $\xi_{3}^{*}$ are defined as in Theorem 3.2.

Proof. This proof follows immediately by simple calculations when we plug (3.4) into the function $F$.  ∎

In the following Figure 3, we derive the shape of the optimal value $\hat{X}_{T}$ and the corresponding wealth of the insurance company as a function of $\xi_{T}$.

Figure 3: The optimal value $\hat{X}_{T}$ and the optimal terminal wealth of the insurance company as a function of $\xi_{T}$ with $k_{1}=2.5$ (left) resp. $k_{0}=2.5$ (right), $k_{2}=7$, $\alpha=1$, $\alpha_{2}=0.5$, $\lambda=3$, $y=1$, $\gamma=0.25$.

3.2 Derivation of the optimal strategy

Next, we derive the optimal strategy for our problem where the proof is somewhat technical and therefore given in Appendix A.

Theorem 3.7.

The optimal solution $\hat{u}$ is non-negative and given by:

$$\hat{u}_{t}=(\sigma_{t}^{T})^{-1}\kappa_{t}\displaystyle\frac{v_{t}}{\hat{X}_{t}},$$

where

$$v_{t}= \left(k_{2}+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)\displaystyle\frac{e^{-\int_{t}^{T}r_{s}\mathrm{d}s}}{\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}}\varphi\left(d_{1}\left(\xi_{1}^{*},t\right)\right)$$

$$+\displaystyle\frac{y}{2\gamma\tilde{\alpha}^{2}}\xi_{t}e^{\int_{t}^{T}-(2r_{s}-\left\lVert\kappa_{s}\right\rVert^{2})\mathrm{d}s}\left[\Phi\left(d_{2}\left(\xi_{1}^{*},t\right)\right)-\displaystyle\frac{1}{\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}}\varphi\left(d_{2}\left(\xi_{1}^{*},t\right)\right)\right]$$

$$+k_{2}\displaystyle\frac{e^{-\int_{t}^{T}r_{s}\mathrm{d}s}}{\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}}\left(\varphi\left(d_{1}\left(\xi_{2}^{*},t\right)\right)-\varphi\left(d_{1}\left(\tilde{\alpha}\hat{\xi},t\right)\right)\right)$$

$$+\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)\displaystyle\frac{e^{-\int_{t}^{T}r_{s}\mathrm{d}s}}{\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}}\left(\varphi\left(d_{1}\left(\xi_{3}^{*},t\right)\right)-\varphi\left(d_{1}\left(\alpha\hat{\xi},t\right)\right)\right)$$

$$+\displaystyle\frac{y}{2\gamma\alpha^{2}}\xi_{t}e^{\int_{t}^{T}-(2r_{s}-\left\lVert\kappa_{s}\right\rVert^{2})\mathrm{d}s}\left[\left(\Phi\left(d_{2}\left(\xi_{3}^{*},t\right)\right)-\Phi\left(d_{2}\left(\alpha\hat{\xi},t\right)\right)\right)\right.$$

$$\hskip 128.0pt-\left.\displaystyle\frac{1}{\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}}\left(\varphi\left(d_{2}\left(\xi_{3}^{*},t\right)\right)-\varphi\left(d_{2}\left(\alpha\hat{\xi},t\right)\right)\right)\right],$$

$$\hat{X}_{t}= \left(k_{2}+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)e^{-\int_{t}^{T}r_{s}\mathrm{d}s}\Phi\left(d_{1}\left(\xi_{1}^{*},t\right)\right)$$

$$-\displaystyle\frac{y}{2\gamma\tilde{\alpha}^{2}}\xi_{t}e^{\int_{t}^{T}-(2r_{s}-\left\lVert\kappa_{s}\right\rVert^{2})\mathrm{d}s}\Phi\left(d_{2}\left(\xi_{1}^{*},t\right)\right)$$

$$+k_{2}e^{-\int_{t}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{2}^{*},t\right)\right)-\Phi\left(d_{1}\left(\tilde{\alpha}\hat{\xi},t\right)\right)\right)$$

$$+\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)e^{-\int_{t}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{3}^{*},t\right)\right)-\Phi\left(d_{1}\left(\alpha\hat{\xi},t\right)\right)\right)$$

$$-\displaystyle\frac{y}{2\gamma\alpha^{2}}\xi_{t}e^{\int_{t}^{T}-(2r_{s}-\left\lVert\kappa_{s}\right\rVert^{2})\mathrm{d}s}\left(\Phi\left(d_{2}\left(\xi_{3}^{*},t\right)\right)-\Phi\left(d_{2}\left(\alpha\hat{\xi},t\right)\right)\right)$$

with

$$d_{1}(x,t)= \displaystyle\frac{\ln x-\ln\xi_{t}+\int_{t}^{T}(r_{s}-\frac{\left\lVert\kappa_{s}\right\rVert^{2}}{2})\mathrm{d}s}{\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}},$$

$$d_{2}(x,t)= \displaystyle\frac{\ln x-\ln\xi_{t}+\int_{t}^{T}(r_{s}-\frac{3\left\lVert\kappa_{s}\right\rVert^{2}}{2})\mathrm{d}s}{\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}}=d_{1}(x,t)-\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s},$$

where $\Phi$ denotes the cdf and $\phi$ the density of a standard normal distributed random variable. The values $\hat{\xi}$, $\xi_{1}^{*}$, $\xi_{2}^{*}$, and $\xi_{3}^{*}$ are defined as in Theorem 3.2.

4 Optimization in a possibly incomplete market

In this subsection, we consider a financial market without the completeness assumption such that we can write the wealth process as:

$$\mathrm{d}X_{t}=b(X_{t},u_{t})\mathrm{d}t+\sigma(X_{t},u_{t})\mathrm{d}W_{t},$$

where $W$ is again the $d$-dimensional Brownian Motion generating the filtration $\mathcal{F}$, and $b$ and $\sigma$ are measurable functions which satisfy a uniform Lipschitz condition.

First, Lemma 3.1 also holds in incomplete markets since we do not use completeness in its proof. Now, for this optimization, we denote for $(t,x)\in[0,T]\times\mathbb{R}$ the optimal value functional by $V$:

$$V(t,x)=\sup_{u\in\mathcal{U}(t,x)}\tilde{J}(t,T,u,x)=\sup_{u\in\mathcal{U}(t,x)}\mathbb{E}[\lambda F(0,T,u,x_{0})-\gamma F(0,T,u,x_{0})^{2}],$$

where $\mathcal{U}(t,x)$ denotes the subset of $\mathcal{U}$ with processes starting at $t$ and $X_{t}=x$. Remember that $\lambda=1+2\gamma\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]$, i.e., $\lambda$ depends on the optimal strategy $\hat{u}$.

Theorem 4.1.

If $V\in C^{1,2}$, then for every $\lambda\geq 0$ the optimal value functional $V$ is the solution of the following SDE for all $(t,x)\in[0,T)\times\mathbb{R}$:

$$-\displaystyle\frac{\mathrm{d}V}{\mathrm{d}t}(t,x)-\sup_{u\in\mathcal{U}}\mathcal{L}^{u}V(t,x) =0,$$

$$V(T,x) =F(T,T,0,x),$$

where the operator $\mathcal{L}^{u}$ is defined as

$$\mathcal{L}^{u}v(t,x) :=b(x,u)\displaystyle\frac{\mathrm{d}V}{\mathrm{d}x}(t,x)+\frac{1}{2}\sigma(x,u)\sigma^{T}(x,u)\displaystyle\frac{\mathrm{d}^{2}V}{\mathrm{d}x^{2}}(t,x),$$

where tr denotes the trace of a matrix.

Proof. Obviously, $\tilde{J}$ satisfies a quadratic growth condition. Then, the result follows from Chapter 3 in Pham [24].  ∎

The regularity assumption of $V$ is typically challenging to prove. Therefore, we also give the result for viscosity solutions where this assumption is not needed:

Theorem 4.2.

Let the control space $\mathcal{U}$ be compact and the Hamiltonian $H$ be defined as usual, i.e.:

$$H :[0,T)\times\mathbb{R}\times\mathbb{R}\times\mathbb{R}\to\mathbb{R},$$

$$H(t,x,p,M) :=\sup_{u\in\mathcal{U}}\left[b(x,u)\displaystyle\frac{\mathrm{d}V}{\mathrm{d}x}(t,x)+\frac{1}{2}\sigma(x,u)\sigma^{T}(x,u)\displaystyle\frac{\mathrm{d}^{2}V}{\mathrm{d}x^{2}}(t,x)\right].$$

If $V$ is locally bounded on $[0,T)\times\mathbb{R}$, then for every $\lambda\geq 0$, $V$ is a viscosity solution of the following HJB equation for $(t,x)\in[0,T)\times\mathbb{R}$:

$$-\displaystyle\frac{\mathrm{d}V}{\mathrm{d}t}(t,x)-H\left(t,x,\displaystyle\frac{\mathrm{d}V}{\mathrm{d}x}(t,x),\displaystyle\frac{\mathrm{d}^{2}V}{\mathrm{d}x^{2}}(t,x)\right) =0,$$

$$V(T,x) =F(T,T,0,x).$$

Proof. The result follows from Chapter 4 in Pham [24].  ∎

Note that it is possible to relax the assumption of $\mathcal{U}$ being compact, leading to a possible singularity in the Hamiltonian. This leads to a complicated structure of the HJB equation and the terminal value condition.

5 Numerical Results

In the following section, we discuss the numerical results for the Black-Scholes model with a given calibration of the parameters with a setting as in Section 3. The first two figures show the impact of different values for the risk aversion resp. the participation rate and of different optimization strategies (EU instead of mean-variance) on the optimal terminal wealth as a function of $\xi_{T}$. The remaining four figures show simulation outcomes for the optimal terminal wealth and the optimal strategy. We analyze the results for the non-protected resp. the protected participation life insurance product, and compare those (a) to each other, (b) to a product with no surplus participation, and (c) to the results with an EU-optimization.

For the participating life insurance contract, we use the parameters $k_{0}=0$, $k_{1}=2.5$, $k_{2}=7$, $\alpha=1$, and $\alpha_{2}=0.25$ for the non-protected insurance product. We change the values of $k_{0}$ and $k_{1}$, i.e., $k_{0}=2.5$ and $k_{1}=0$, to get the protected product. The initial wealth, i.e., the sum of the premiums and the initial capital from the insurance company, is given by $x_{0}=4$, the time horizon by $T=10$, and the risk aversion by $\gamma=0.25$. We assume a constant risk-free interest rate of $r:=r_{t}\equiv 0.02$ for the financial market and consider one risky asset with constant mean $\mu:=\mu_{t}\equiv 0.08$ and volatility $\sigma_{t}\equiv 0.2$. In particular, the Sharpe ratio is also constant and given by $\kappa_{t}\equiv 0.3$. For the implementation, we used a time step of $\delta=0.01$, i.e., we have 1000 intermediate points. Moreover, we simulate 1000 realizations of the Brownian Motion.

In Figure 4, we present the optimal terminal wealth as a function of $\xi_{T}$ for different values of risk aversion parameter $\gamma$. (Note that the risk aversion parameters are not comparable for the different optimization strategies.) For both the non-protected and the protected case, we compare the optimal terminal wealth stemming from our mean-variance optimization with the EU-optimization for participating life insurance contracts from Lin et al. [16]. To measure the utility, we use, as Lin et al., the following S-shaped utility function: $U(x)=\begin{cases}x^{\tilde{\gamma}}&x\geq 0,\\ -\tilde{\lambda}(-x)^{-\tilde{\gamma}}&x<0\end{cases}$ with $\tilde{\lambda}=2$ and different values for $\tilde{\gamma}$ (Lin et al. [16] used $\tilde{\gamma}=0.5$ in their numerical analysis). Moreover, we added the case that there is no additional surplus participation of the policyholders, i.e., $\alpha_{2}=0$, $k_{0}=0$, and $k_{1}=0$. Both participating optimizations share that the optimal wealth function has three points of significant behavioral change, i.e., at $\xi_{1}^{*}$, $\xi_{2}^{*}$ and $\xi_{3}^{*}$ as in Theorem 3.2. This theorem shows that the optimal wealth decreases before $\xi_{1}^{*}$ and between $\xi_{2}^{*}$ and $\xi_{3}^{*}$, and that the optimal wealth is constant elsewhere. The main difference in the shape of the functions is that we have a piecewise linearity in the mean-variance case but a total non-linearity in the EU case. This effect is already apparent from the respective formulas. Noticeably, the distance between the second and third point of behavioral change is much higher in the EU-optimization than in the mean-variance one. Moreover, the figure shows a moral hazard of the insurance companies offering a non-protected participating life insurance contract since the insurance companies favor a portfolio value of $0$ over a portfolio value at the guarantee value level. We can observe this effect for both the mean-variance and the EU-optimization in the figure due to the drop-down of the optimal terminal wealth to zero. Protected products have no such effect when optimizing mean-variance since the insurance company always profits from higher portfolio values. While this effect also occurs when optimizing EU, the probability is smaller since the drop-down is at a far higher value (compared to the non-protected product).

Figure 4: Comparison of the optimal terminal wealth as a function of $\xi_{T}$ of mean-variance with EU optimized participating life insurance products and no participation in the case of a not protected guarantee (left) and a protected one (right). Note that the risk aversion parameters are not comparable for mean-variance and EU.

In Figure 5, we show the influence of the participation rate $\alpha_{2}$ on the optimal terminal wealth as a function of $\xi_{T}$. We can see that a higher rate $\alpha_{2}$ leads to a more extended plateau at $k_{2}=7$. This effect is not surprising when checking the formula (3.4) of the optimal terminal wealth $\hat{X}_{T}$ since a higher $\alpha_{2}$ leads to a smaller $\tilde{\alpha}$. Thus, the intermediate interval $(\tilde{\alpha}\hat{\xi},\xi_{2}^{*})$ gets bigger since for our parametrization, it always holds that $\xi_{2}^{*}=\alpha\hat{\xi}$. Moreover, we can observe that $\xi^{*}$ (defined in Theorem 3.2 as the point where in the non-protected case the optimal terminal wealth drops down to $0$) is increasing for an increasing participation rate $\alpha_{2}$. This effect stems from the different values of the variables $\lambda$ and $y$ when changing the participation rate. Additionally, the figure shows that for minimal values of $\xi_{T}$, the optimal terminal wealth is higher for large participation rates, which we can derive directly from the formula of $\hat{X}_{T}$ in (3.4). For the economic interpretation, we infer that a higher participation rate of the policyholders leads to a higher portfolio value in bad economic states, because for a higher $\alpha_{2}$, the upside potential of the insurer is reduced. Hence, the optimal strategy should be less risky, which implies a higher portfolio value. For good economic states, the opposite effect happens. A somewhat surprising result is that in very good economic states (happening with an extremely low probability), the optimal terminal wealth is higher for higher participation rates, which occurs due to changes in the Lagrange multiplier of the budget constraint.

Figure 5: Influence of the participation rate $\alpha_{2}$ on the optimal terminal wealth as a function of $\xi_{T}$ of a not protected (left) and a protected (right) participation life insurance.

In Figure 6, we show the optimal wealth process and the optimal share into the risky asset over time for the non-protected participating life insurance product. In both figures, the red line shows the average of the 1000 realizations, and each of the ten black lines shows a single realization. Note that for the optimal strategy, we used a weighted average with the weight given by the absolute investment amount (which also holds for all other averages in this section). Furthermore, we get $\lambda\approx 3.423$ and $y\approx 0.860$. The average optimal wealth process develops from $4$ to approximately $7.8$, where the increase is slightly larger in the beginning than in the end. A terminal wealth of about $7.8$ corresponds to a terminal wealth for the insurer of $5.1$ and a wealth of $2.7$ for the policyholder. Note that this is higher than the guarantee value, which is $2.5$. When exercising the optimal strategy, we start with high investment into the risky asset (approximately $110\%$ of our wealth), which decreases relatively constant over time. The final optimal investment share into the risky asset is around $40\%$. When analyzing the strategies in more detail, one observes that the most risky investments are taken when the economy has poorly evolved until then, i.e., when $\xi_{t}$ is high, while the least risky investments are taken in cases where the economy has developed well. These results are reasonable since when the wealth is below the guarantee, the insurer has nothing to lose anymore, while the wealth is already over the second threshold $k_{2}$ (i.e., the policyholder also participates in the surplus over $k_{2}$), the upside potential of the insurer is reduced, whereas the downside potential is not (or only slightly).

Figure 6: Optimal wealth and optimal strategy for the non-protected participating life insurance product.

In Figure 7, we show the optimal wealth process and the optimal share into the risky asset over time for the protected participating life insurance product. Again, the red line shows the average of the 1000 realizations, and each of the ten black lines shows a single realization. Moreover, the variables $\lambda$ and $y$ are given by $\lambda\approx 2.893$ and $y\approx 1.003$. In this case, the optimal value evolves only to around $6.8$, corresponding to a terminal wealth of the insurance company of $4.3$ since we are below $k_{2}$, i.e., there is no surplus participation of the policyholder. The optimal strategy starts risky but not as risky as in the non-protected case, with around $75\%$ of the wealth invested into the risky asset. It also decreases over time, but the reduction is smaller, and the final optimal investment percentage is at around $30\%$ of the wealth. As in the non-protected case, the realizations in the best economic states correspond with the least risky investments and vice versa. One of the shown strategies is rather inconspicuous over most of the time but has a major change close to the final time point $T=10$, i.e., a drop-down from $u_{9.93}=0.2$ to $u_{9.99}=0.00001$. Such an effect (also in the other way, i.e., an upward move) happens for several realizations when their wealth for time points $t$, which are close to the maturity $T$, is close to the second threshold $k_{2}$ since the optimal terminal wealth as a function of $\xi$ has a small plateau at $k_{2}$, i.e., if $\xi_{t}\in(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]$. Hence, for $\xi_{t}\approx\tilde{\alpha}\hat{\xi}$ (resp. $\xi_{t}\approx\tilde{\alpha}\xi_{2}^{*}$) for $t$ close to $T$, the upside (resp. downside) potential is reduced and the optimal strategy is to invest safely (resp. risky).

Figure 7: Optimal wealth and optimal strategy for the protected participating life insurance product.

In Figure 8, we compare the optimal wealth before splitting it between the policyholders and the insurance company and the optimal strategy of our two participating life insurance products with mean-variance to a non-participating investment, i.e., we set $\alpha_{2}=0$, $k_{0}=0$, and $k_{1}=0$. Moreover, we set the initial wealth such that the mean of $\hat{X}_{T}$ is approximately equal to $7.787$. Therefore, we get the following initial values: $x_{0}={\scriptsize\begin{cases}4&\text{ non-protected product},\\ 4.765&\text{ protected product},\\ 4.738&\text{ no participation}.\end{cases}}$ The lines show again an average of $1000$ realizations each. The figure shows that the investor makes the riskiest investment when offering a non-protected insurance product and approximately identical investments when offering one of the other two products. When offering the protected product compared to the non-participation product, it is remarkable that the investor invests slightly safer in the beginning and somewhat riskier close to maturity. This investment behavior leads to the highest wealth gain for the non-protected product since, on average, the risky asset performs better than the risk-less asset (due to $\mu>r$). For the other two products, the wealth gain is approximately equal. The protected life insurance investment strategy is relatively close to the non-participating strategy since a fixed-guarantee payment only minimally influences the optimal strategy. Then, the only remaining difference in the payoff shape is when the wealth is above $k_{2}$, which happens in our chosen parametrization either relatively late or not at all (see Figure 7). The payoff structure for the non-protected insurance product differs highly in the values due to the reduced downside potential below $k_{1}$.

Figure 8: Comparison of the optimal wealth (left) and strategy (right) between a (non-)protected participating life insurance product and a non-participating investment.

In Figure 9, we compare the optimal wealth and strategy for the non-protected and the protected life insurance contract when considering mean-variance resp. EU-optimization. We take the result for optimizing EU from Lin et al. [16] as in Figure 4 with parameter values $\tilde{\gamma}=0.125$ and $\tilde{\lambda}=2$. As in the other figures, the lines show the average of $1000$ realizations, and we state the total wealth (before splitting it between the policyholders and the insurance company). As in the previous figure, we again choose the initial value such that the terminal wealth is approximately equal for all products, i.e., we take $x_{0}={\scriptsize\begin{cases}4&\text{ non-protected product},\\ 4.765&\text{ protected product}\end{cases}}$ for the mean-variance optimization and $x_{0}={\scriptsize\begin{cases}3.343&\text{ non-protected product},\\ 3.616&\text{ protected product}\end{cases}}$ for the EU-optimization. Both optimizations have in common that the insurance company offering a non-protected product invests riskier than the insurance company offering the protected product due to the reduced downside potential. When comparing the strategies, we observe that the strategies optimizing EU become riskier over time (except for the final time points). In contrast, the strategies optimizing mean-variance get less risky over time, as already discussed. Since we chose the two (not comparable) risk aversion factors such that the initial investments in the risky asset are similar, the wealth gain of the EU-optimization is higher. The decreasing portion invested into the risky asset is typical for a pre-commitment mean-variance optimization strategy (this can be seen by implementing the optimal mean-variance strategy by Zhou and Li [30] when using typical ranges for the parameters), which is structurally different from EU-optimization strategies.

Figure 9: Comparison of the optimal wealth (left) and the optimal strategy (right) between a (non-)protected participating life insurance contract optimizing mean-variance and EU.

6 Conclusion

In this paper, we derived explicit analytic formulas for general contracts, including participating life insurance contracts, when optimizing mean-variance in the multi-dimensional Black-Scholes market. Moreover, by showing the existence of all arising parameters, we showed the existence of the optimal solution. We also gave the HJB equation for the value functional if the market is possibly incomplete. A numerical analysis shows that the mean-variance optimal strategy compared to the EU optimal strategy becomes more conservative the shorter the maturity is and is increased in particular in bad economic states. Future research directions include possibly generalizing this approach to other stock market models.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Appendix

Appendix A Proofs

• We show this lemma by contradiction. In particular, we assume that a $\hat{u}$ exists, which is an optimal solution for $J$, but not for $\tilde{J}$. Consequently, there exists a strategy $u$ which is better than $\hat{u}$ for $\tilde{J}$, i.e.,

  $$\tilde{J}(0,T,u,x_{0})-\tilde{J}(0,T,\hat{u},x_{0}) >0$$

  $$\Leftrightarrow \lambda\left(\mathbb{E}\left[F(0,T,u,x_{0})-F(0,T,\hat{u},x_{0})\right]\right)-\gamma\left(\mathbb{E}\left[F(0,T,u,x_{0})^{2}-F(0,T,\hat{u},x_{0})^{2}\right]\right) >0.$$ (A.1)

  The next step is to define the function $G(x,y):=y-\gamma x+\gamma y^{2}$. We observe that $G$ is a convex function, and it holds that

  $$G\left(\mathbb{E}\left[F(0,T,u,x_{0})^{2}\right],\mathbb{E}\left[F(0,T,u,x_{0})\right]\right)=\mathbb{E}\left[F(0,T,u,x_{0})\right]-\gamma\text{Var}\left(F(0,T,u,x_{0})\right)=J(0,T,u,x_{0}).$$

  We note that $\frac{\partial}{\partial x}G(x,y)=-\gamma$ and $\frac{\partial}{\partial y}G(x,y)=1+2\gamma y$. Then, the convexity of $G$ implies

  $$G\left(\mathbb{E}\left[F(0,T,u,x_{0})^{2}\right],\mathbb{E}\left[F(0,T,u,x_{0})\right]\right)$$

  $$\geq \ G\left(\mathbb{E}\left[F(0,T,\hat{u},x_{0})^{2}\right],\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]\right)$$

  $$+\begin{pmatrix}-\gamma\\ 1+2\gamma\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]\end{pmatrix}\cdot\begin{pmatrix}\mathbb{E}\left[F(0,T,u,x_{0})^{2}\right]-\mathbb{E}\left[F(0,T,\hat{u},x_{0})^{2}\right]\\ \mathbb{E}\left[F(0,T,u,x_{0})\right]-\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]\end{pmatrix}$$

  $$= \ G\left(\mathbb{E}\left[F(0,T,\hat{u},x_{0})^{2}\right],\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]\right)$$

  $$-\gamma\left(\mathbb{E}\left[F(0,T,u,x_{0})^{2}\right]-\mathbb{E}\left[F(0,T,\hat{u},x_{0})^{2}\right]\right)$$

  $$+(1+2\gamma\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right])\left(\mathbb{E}\left[F(0,T,u,x_{0})\right]-\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]\right)$$

  $$> \ G\left(\mathbb{E}\left[F(0,T,\hat{u},x_{0})^{2}\right],\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]\right),$$ (A.2)

  where we used (Proof of Lemma 3.1.) and $\lambda=1+2\gamma\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]$. Hence, (Proof of Lemma 3.1.) implies that $J(0,T,u,x_{0})>J(0,T,\hat{u},x_{0})$, i.e., $\hat{u}$ is not optimal, which is a contradiction and implies the lemma. ∎

• We prove this proposition by showing the following two statements from which the claim follows directly: (i) If $\xi_{3}^{*}>\alpha\hat{\xi}$, then it holds that $\xi_{2}^{*}=\alpha\hat{\xi}$, and (ii) if $\xi_{2}^{*}>\tilde{\alpha}\hat{\xi}$, then it holds that $\xi_{1}^{*}=\tilde{\alpha}\hat{\xi}$.

  We start with the proof of (i). Therefore, we reformulate the condition $\xi_{3}^{*}>\alpha\hat{\xi}$ to get a condition for $\lambda$. When plugging in the formulas as defined in Theorem 3.2, it holds that $\xi_{3}^{*}>\alpha\hat{\xi}$ if and only if

  $$\displaystyle\frac{2\gamma\alpha^{2}(k_{0}+k_{1})-2\gamma\alpha^{2}\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+\frac{\lambda}{\gamma\alpha}\right)}}{y}>-\displaystyle\frac{2\gamma\alpha^{2}(k_{2}-k_{1}-k_{0})}{y}.$$ (A.3)

  Hence, it follows that $y<\infty$ and $k_{2}>\sqrt{k_{1}^{2}+2k_{0}k_{1}+\frac{\lambda k_{1}}{\gamma\alpha}}$ which is again equivalent to $\frac{\lambda k_{1}}{\gamma\alpha}<k_{2}^{2}-k_{1}^{2}-2k_{0}k_{1}$. Thus, we get that $\xi_{3}^{*}>\alpha\hat{\xi}$ holds if and only if $k_{1}=0$ or $\lambda<\gamma\alpha\left(\frac{k_{2}^{2}}{k_{1}}-k_{1}-2k_{0}\right)$ since $k_{2}>0$. Before showing the implication (i), let us also reformulate the equality $\xi_{2}^{*}=\alpha\hat{\xi}$. First, $\xi_{2}^{*}=\alpha\hat{\xi}$ if and only if $\tilde{\xi}_{2}^{*}\geq\alpha\hat{\xi}$ which is equivalent to:

  $$-\displaystyle\frac{\gamma\alpha^{2}(k_{2}-k_{1})^{2}-2\gamma\alpha^{2}k_{0}(k_{2}-k_{1})+\lambda\alpha k_{1}}{yk_{2}}\geq-\displaystyle\frac{2\gamma\alpha^{2}(k_{2}-k_{1}-k_{0})}{y}.$$

  Since $y<\infty$ (otherwise (A.3) would be wrong), this is equivalent to:

  $$2\gamma\alpha^{2}\left(k_{0}(k_{2}-k_{1})+k_{2}(k_{2}-k_{1}-k_{0})\right)-\gamma\alpha^{2}(k_{2}^{2}-2k_{1}k_{2}+k_{1}^{2})\geq\lambda\alpha k_{1}.$$

  This again is equivalent to:

  $$\lambda k_{1}\leq\gamma\alpha\left(2k_{0}k_{2}-2k_{0}k_{1}+2k_{2}^{2}-2k_{1}k_{2}-2k_{0}k_{2}-k_{2}^{2}+2k_{1}k_{2}-k_{1}^{2}\right)=\gamma\alpha\left(k_{2}^{2}-k_{1}^{2}-2k_{0}k_{1}\right).$$

  Now, if $k_{1}=0$ and if $\lambda<\gamma\alpha\left(\frac{k_{2}^{2}}{k_{1}}-k_{1}-2k_{0}\right)$ the inequality is fulfilled since $k_{2}>0$. Thus, statement (i) follows.

  Second, we prove (ii). As in the proof of (i), we start by reformulating the first condition. We note that $\xi_{2}^{*}>\tilde{\alpha}\hat{\xi}$ if and only if $\tilde{\xi}_{2}^{*}>\tilde{\alpha}\hat{\xi}$ which is equivalent to:

  $$\displaystyle\frac{\alpha\lambda}{y}-\displaystyle\frac{\gamma\alpha^{2}(k_{2}-k_{1})^{2}-2\gamma\alpha^{2}k_{0}(k_{2}-k_{1})+\lambda\alpha k_{1}}{yk_{2}}>\displaystyle\frac{\tilde{\alpha}\lambda}{y}-\displaystyle\frac{2\gamma\alpha\tilde{\alpha}(k_{2}-k_{1}-k_{0})}{y}.$$

  From this inequality, it follows that $y<\infty$ and

  $$\lambda(\alpha-\tilde{\alpha})k_{2}-\lambda\alpha k_{1}+\gamma\alpha^{2}\left(-k_{2}^{2}+2k_{1}k_{2}-k_{1}^{2}+2k_{0}k_{2}-2k_{0}k_{1}\right)>2\gamma\alpha\tilde{\alpha}\left(-k_{2}^{2}+k_{1}k_{2}+k_{0}k_{2}\right).$$

  Since $\alpha-\tilde{\alpha}=\alpha_{2}$, this is equivalent to:

  $$\lambda(\alpha_{2}k_{2}-\alpha k_{1})>-\gamma\alpha\left(k_{2}^{2}(\tilde{\alpha}-\alpha_{2})+2k_{1}k_{2}\alpha_{2}+2k_{0}k_{2}\alpha_{2}-\alpha k_{1}^{2}-2\alpha k_{0}k_{1}\right).$$

  As before, we also reformulate the equality $\xi_{1}^{*}=\tilde{\alpha}\hat{\xi}$. This is indeed true if $\tilde{\xi}_{1}^{*}\geq\tilde{\alpha}\hat{\xi}$. If the term under the square root in the definition of $\tilde{\xi}_{1}^{*}$ (see Theorem 3.2) is $0$, then $\tilde{\xi}_{1}^{*}\geq\tilde{\alpha}\hat{\xi}$ is fulfilled since $\frac{2\gamma\tilde{\alpha}^{2}k_{2}}{y}\geq 0$. Now, if this term is positive, $\tilde{\xi}_{1}^{*}\geq\tilde{\alpha}\hat{\xi}$ is equivalent to:

  $$\displaystyle\frac{2\gamma\tilde{\alpha}^{2}k_{2}}{y}\geq\displaystyle\frac{2\gamma\tilde{\alpha}}{y}\sqrt{(\alpha(k_{0}+k_{1})-\alpha_{2}k_{2})^{2}-\alpha^{2}k_{0}^{2}+\frac{\lambda}{\gamma}(\alpha k_{1}-\alpha_{2}k_{2})}.$$

  Then, since $y<\infty$, this is equivalent to:

  $$\tilde{\alpha}^{2}k_{2}^{2}\geq\alpha^{2}k_{0}^{2}+2\alpha^{2}k_{0}k_{1}+\alpha^{2}k_{1}^{2}-2\alpha\alpha_{2}k_{0}k_{2}-2\alpha\alpha_{2}k_{1}k_{2}+\alpha_{2}^{2}k_{2}^{2}-\alpha^{2}k_{0}^{2}+\frac{\lambda}{\gamma}(\alpha k_{1}-\alpha_{2}k_{2}),$$

  which is again equivalent to:

  $$\lambda(\alpha_{2}k_{2}-\alpha k_{1})\geq-\gamma\alpha\left(k_{2}^{2}\left(\frac{\tilde{\alpha}^{2}-\alpha_{2}^{2}}{\alpha}\right)+2k_{1}k_{2}\alpha_{2}+2k_{0}k_{2}\alpha_{2}-\alpha k_{1}^{2}-2\alpha k_{0}k_{1}\right).$$

  Now, we get that $\frac{\tilde{\alpha}^{2}-\alpha_{2}^{2}}{\alpha}=\frac{(\tilde{\alpha}-\alpha_{2})(\tilde{\alpha}+\alpha_{2})}{\alpha}=\tilde{\alpha}-\alpha_{2}$ since $\tilde{\alpha}+\alpha_{2}=\alpha$. Hence, the inequality is fulfilled if $\lambda(\alpha_{2}k_{2}-\alpha k_{1})>-\gamma\alpha\left(k_{2}^{2}(\tilde{\alpha}-\alpha_{2})+2k_{1}k_{2}\alpha_{2}+2k_{0}k_{2}\alpha_{2}-\alpha k_{1}^{2}-2\alpha k_{0}k_{1}\right)$ which implies statement (ii) and finishes the proof. ∎

• Using the two properties proven in the proof of Proposition 3.4, we only have to check the boundary values $\xi_{1}^{*}=\tilde{\alpha}\hat{\xi}$ and $\xi_{2}^{*}=\alpha\hat{\xi}$ for continuity. However, this follows immediately by plugging in the values. Now, if $k_{1}>0$, we see that $\hat{X}_{T}>0$ if $\xi^{*}\neq\xi_{3}^{*}$. If $\xi^{*}=\xi_{3}^{*}$, then we note that

  $$\hat{X}_{T}(\xi_{3}^{*}) =k_{0}+k_{1}+\displaystyle\frac{\lambda\alpha-\lambda\alpha-2\gamma\alpha^{2}k_{0}+2\gamma\alpha^{2}\left(\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+\frac{\lambda}{\gamma\alpha}\right)}-k_{1}\right)}{2\gamma\alpha^{2}}$$

  $$=k_{0}+k_{1}-k_{0}-k_{1}+\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+\frac{\lambda}{\gamma\alpha}\right)}=\sqrt{k_{1}\left(k_{1}+2k_{0}+\frac{\lambda}{\gamma\alpha}\right)}.$$

  Hence, $\hat{X}_{T}(\xi_{3}^{*})>0$ if $k_{1}>0$ and $\lambda>-2\gamma\alpha k_{0}-\gamma\alpha k_{1}$. From the proof of Proposition A.3, we see that $\lambda>C$ with $C\geq-2\gamma\alpha k_{0}$. Thus, $\hat{X}_{T}(\xi_{3}^{*})>0$ if $k_{1}>0$. ∎

• From (3.1) and (3.2), we conclude using Itô’s formula that $\xi_{t}\hat{X}_{t}$ is a martingale. Hence, we get using Theorem 3.2:

  $$\hat{X}_{t}= \mathbb{E}\left[\left.\displaystyle\frac{\xi_{T}}{\xi_{t}}\hat{X}_{T}\right|\mathcal{F}_{t}\right]$$

  $$= \left(k_{2}+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)\mathbb{E}\left[\left.\displaystyle\frac{\xi_{T}}{\xi_{t}}\mathbbm{1}_{\xi_{T}\leq\xi_{1}^{*}}\right|\mathcal{F}_{t}\right]-\displaystyle\frac{y}{2\gamma\tilde{\alpha}^{2}}\mathbb{E}\left[\left.\displaystyle\frac{\xi_{T}^{2}}{\xi_{t}}\mathbbm{1}_{\xi_{T}\leq\xi_{1}^{*}}\right|\mathcal{F}_{t}\right]$$

  $$+k_{2}\mathbb{E}\left[\left.\displaystyle\frac{\xi_{T}}{\xi_{t}}\mathbbm{1}_{\tilde{\alpha}\hat{\xi}<\xi_{T}\leq\xi_{2}^{*}}\right|\mathcal{F}_{t}\right]+\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)\mathbb{E}\left[\left.\displaystyle\frac{\xi_{T}}{\xi_{t}}\mathbbm{1}_{\alpha\hat{\xi}<\xi_{T}\leq\xi_{3}^{*}}\right|\mathcal{F}_{t}\right]$$

  $$-\displaystyle\frac{y}{2\gamma\alpha^{2}}\mathbb{E}\left[\left.\displaystyle\frac{\xi_{T}^{2}}{\xi_{t}}\mathbbm{1}_{\alpha\hat{\xi}<\xi_{T}\leq\xi_{3}^{*}}\right|\mathcal{F}_{t}\right].$$ (A.4)

  Now, the claim for $\hat{X}_{t}$ follows from the formula for the conditional expectation of log-normal distributions (see (B.1)) using that it holds conditionally on $\mathcal{F}_{t}$:

  $$\frac{\xi_{T}}{\xi_{t}} \sim\mathcal{LN}\left(-\int_{t}^{T}r_{s}+\frac{\left\lVert\kappa_{s}\right\rVert^{2}}{2}\mathrm{d}s,\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s\right),$$

  $$\frac{\xi_{T}^{2}}{\xi_{t}} \sim\mathcal{LN}\left(\ln\xi_{t}-2\int_{t}^{T}r_{s}+\frac{\left\lVert\kappa_{s}\right\rVert^{2}}{2}\mathrm{d}s,4\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s\right),$$

  where $\mathcal{LN}$ denotes a log-normal distribution.
  The next step is to calculate the volatility process of $\hat{X}_{t}$, i.e., the term before the $\mathrm{d}W_{t}$ in its SDE, denoted by $\sigma_{\hat{X}_{t}}$ using (3.2). There, we get $\sigma_{\hat{X}_{t}}=(-\kappa_{t}^{T}\xi_{t})\displaystyle\frac{\partial\hat{X}_{t}}{\partial\xi_{t}}=\kappa_{t}^{T}v_{t}$, where we used that $\displaystyle\frac{\partial d_{1}(x)}{\partial\xi_{t}}=\displaystyle\frac{\partial d_{2}(x)}{\partial\xi_{t}}=\displaystyle\frac{1}{-\xi_{t}\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}}$. Now, we get the optimal strategy $\hat{u}_{t}$ by comparing the volatility processes from $\hat{X}_{t}$ which is $\sigma_{\hat{X}_{t}}=\hat{X}_{t}\hat{u}^{T}\sigma_{t}$. Hence, the claim follows. ∎

A.1 Proposition A.1

Since the equation $\mathbb{E}[\xi_{T}\hat{X}_{T}(y)]=\xi_{0}x_{0}$ depends on $\lambda$ and $\lambda=1+2\gamma\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]$ depends on $y$, we have to solve these two equations together. We show that these variables always exist in the following Proposition A.1.

Proposition A.1.

There always exists a solution for $\lambda$ and $y$ and an equation system that can be numerically solved to determine them.

We split this proof into two lemmas. In the first Lemma A.2, we give the equation system to solve for the two parameters $y$ and $\lambda$, where we added $y$ and $\lambda$ as a superscript to help the reader follow the interdependence. In the second Lemma A.3, we show that these parameters exist:

Lemma A.2.

The variables $y$ and $\lambda$ are the solution of the following equation system:

$$x_{0}= \left(k_{2}+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\Phi\left(d_{1}\left(\xi_{1}^{*,y,\lambda},0\right)\right)$$

$$-\displaystyle\frac{y}{2\gamma\tilde{\alpha}^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\Phi\left(d_{2}\left(\xi_{1}^{*,y,\lambda},0\right)\right)$$

$$+k_{2}e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{2}^{*,y,\lambda},0\right)\right)-\Phi\left(d_{1}\left(\tilde{\alpha}\hat{\xi}^{y,\lambda},0\right)\right)\right)$$

$$+\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{3}^{*,y,\lambda},0\right)\right)-\Phi\left(d_{1}\left(\alpha\hat{\xi}^{y,\lambda},0\right)\right)\right)$$

$$-\displaystyle\frac{y}{2\gamma\alpha^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\left(\Phi\left(d_{2}\left(\xi_{3}^{*,y,\lambda},0\right)\right)-\Phi\left(d_{2}\left(\alpha\hat{\xi}^{y,\lambda},0\right)\right)\right),$$

$$\lambda= \ 1-2\gamma\alpha k_{0}+2\gamma\alpha k_{0}\left(\Phi\left(d_{0}\left(\xi_{3}^{*,y,\lambda},0\right)\right)-\Phi\left(d_{0}\left(\alpha\hat{\xi}^{y,\lambda},0\right)\right)+\Phi\left(d_{0}\left(\xi_{2}^{*,y,\lambda},0\right)\right)\right.$$

$$\hskip 102.0pt-\left.\Phi\left(d_{0}\left(\tilde{\alpha}\hat{\xi}^{y,\lambda},0\right)\right)+\Phi\left(d_{0}\left(\xi_{1}^{*,y,\lambda},0\right)\right)\right)$$

$$+\lambda\left(\Phi\left(d_{0}\left(\xi_{3}^{*,y,\lambda},0\right)\right)-\Phi\left(d_{0}\left(\alpha\hat{\xi}^{y,\lambda},0\right)\right)+\Phi\left(d_{0}\left(\xi_{1}^{*,y,\lambda},0\right)\right)\right)$$

$$-y\left(\displaystyle\frac{1}{\alpha}\Phi\left(d_{1}\left(\xi_{3}^{*,y,\lambda},0\right)\right)-\displaystyle\frac{1}{\alpha}\Phi\left(d_{1}\left(\alpha\hat{\xi}^{y,\lambda},0\right)\right)+\displaystyle\frac{1}{\tilde{\alpha}}\Phi\left(d_{1}\left(\xi_{1}^{*,y,\lambda},0\right)\right)\right)$$

$$+2\gamma\alpha(k_{2}-k_{1}-k_{0})\left(\Phi\left(d_{0}\left(\xi_{2}^{*,y,\lambda},0\right)\right)-\Phi\left(d_{0}\left(\tilde{\alpha}\hat{\xi}^{y,\lambda},0\right)\right)\right)$$

with

$$d_{0}(x,t)= \displaystyle\frac{\ln x-\ln\xi_{t}+\int_{t}^{T}r_{s}+\frac{\left\lVert\kappa_{s}\right\rVert^{2}}{2}\mathrm{d}s}{\sqrt{\int_{t}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}}.$$

The values $\hat{\xi}^{y,\lambda}$, $\xi_{1}^{*,y,\lambda}$, $\xi_{2}^{*,y,\lambda}$, and $\xi_{3}^{*,y,\lambda}$ are defined as in Theorem 3.2 and the functions $d_{1}$ and $d_{2}$ are defined as in Theorem 3.7.

Proof. The first formula follows from the definition $\mathbb{E}[\xi_{T}\hat{X}_{T}(y)]=\xi_{0}x_{0}=x_{0}$ and (Proof of Theorem 3.7.) for $t=0$. The second formula follows from the definition $\lambda=1+2\gamma\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]$ and the following properties for arbitrary $0\leq a<b<\infty$:

$$\hat{X}_{T}= \left(k_{2}+\displaystyle\frac{\lambda\tilde{\alpha}-y\xi_{T}}{2\gamma\tilde{\alpha}^{2}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)\mathbbm{1}_{\xi_{T}\in(0,\xi_{1}^{*}]}+k_{2}\mathbbm{1}_{\xi_{T}\in[\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]}$$

$$+\left(k_{0}+k_{1}+\displaystyle\frac{\lambda\alpha-y\xi_{T}}{2\gamma\alpha^{2}}\right)\mathbbm{1}_{\xi_{T}\in[\alpha\hat{\xi},\xi_{3}^{*}]},$$

$$\ln\xi_{T}\sim \ \mathcal{N}\left(-\int_{0}^{T}r_{s}+\frac{\left\lVert\kappa_{s}\right\rVert^{2}}{2}\mathrm{d}s,\int_{0}^{T}\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s\right),$$

$$\mathbb{E}[\xi_{T}\mathbbm{1}_{\xi_{T}\in[a,b]}]= \ \Phi\left(d_{1}(b,0)\right)-\Phi\left(d_{1}(a,0)\right),$$ (A.5)

$$\mathbb{E}[\mathbbm{1}_{\xi_{T}\in[a,b]}]= \ \mathbb{P}(\xi_{T}\in[a,b])=\Phi\left(d_{0}(b,0)\right)-\Phi\left(d_{0}(a,0)\right).$$ (A.6)

Moreover, we conclude from (3.4) (combined with Proposition 3.5) that $\mathbbm{1}_{\hat{X}_{T}\geq k_{2}}=\mathbbm{1}_{\xi_{T}\in(0,\xi_{1}^{*}]}+\mathbbm{1}_{\xi_{T}\in(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]}$, $\mathbbm{1}_{\hat{X}_{T}\in[k_{1},k_{2})}=\mathbbm{1}_{\xi_{T}\in(\alpha\hat{\xi},\xi_{3}^{*}]}$, and $\mathbbm{1}_{\hat{X}_{T}<k_{1}}=\mathbbm{1}_{\hat{X}_{T}=0}=\mathbbm{1}_{\xi_{T}\not\in(0,\xi_{1}^{*}]\cup(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]\cup(\alpha\hat{\xi},\xi_{3}^{*}]}$.
Indeed, we get:

$$\mathbb{E}\left[F(0,T,\hat{u},x_{0})\right]= \mathbb{E}\left[-\alpha k_{0}\mathbbm{1}_{\hat{X}_{T}<k_{1}}+\alpha(\hat{X}_{T}-k_{1}-k_{0})\mathbbm{1}_{\hat{X}_{T}\in[k_{1},k_{2})}+\tilde{\alpha}(\hat{X}_{T}-k_{2})\mathbbm{1}_{\hat{X}_{T}\geq k_{2}}\right]$$

$$+\mathbb{E}\left[\alpha(k_{2}-k_{1}-k_{0})\mathbbm{1}_{\hat{X}_{T}\geq k_{2}}\right]$$

$$= -\alpha k_{0}\mathbb{P}(\hat{X}_{T}<k_{1})+\alpha\mathbb{E}[(\hat{X}_{T}-k_{1}-k_{0})\mathbbm{1}_{\xi_{T}\in(\alpha\hat{\xi},\xi_{3}^{*}]}]$$

$$+\mathbb{E}\left[\left(\tilde{\alpha}(\hat{X}_{T}-k_{2})+\alpha(k_{2}-k_{1}-k_{0})\right)\left(\mathbbm{1}_{\xi_{T}\in(0,\xi_{1}^{*}]}+\mathbbm{1}_{\xi_{T}\in(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]}\right)\right]$$

$$= -\alpha k_{0}+\alpha k_{0}\left(\mathbb{P}(\xi_{T}\in(0,\xi_{1}^{*}])+\mathbb{P}(\xi_{T}\in(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}])+\mathbb{P}(\xi_{T}\in(\alpha\hat{\xi},\xi_{3}^{*}])\right)$$

$$+\displaystyle\frac{\lambda}{2\gamma}\mathbb{P}(\xi_{T}\in(\alpha\hat{\xi},\xi_{3}^{*}])-\displaystyle\frac{y}{2\gamma\alpha}\mathbb{E}\left[\xi_{T}\mathbbm{1}_{\xi_{T}\in(\alpha\hat{\xi},\xi_{3}^{*}]}\right]+\displaystyle\frac{\lambda}{2\gamma}\mathbb{P}(\xi_{T}\in(0,\xi_{1}^{*}])$$

$$-\displaystyle\frac{y}{2\gamma\tilde{\alpha}}\mathbb{E}\left[\xi_{T}\mathbbm{1}_{\xi_{T}\in(0,\xi_{1}^{*}]}\right]+\alpha(k_{2}-k_{1}-k_{0})\mathbb{P}(\xi_{T}\in(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]),$$

where we used the above-discussed results in the second and (3.4) in the third equation. The claim follows with (A.5) and (A.6).  ∎

Lemma A.3.

The equation system from Lemma A.2 admits a solution.

Proof. For this existence proof, we make the following two definitions by writing the two equations as functions depending on $y$ and $\lambda$:

$$f_{1}(y,\lambda):= -x_{0}+\left(k_{2}+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\Phi\left(d_{1}\left(\xi_{1}^{*},0\right)\right)$$

$$-\displaystyle\frac{y}{2\gamma\tilde{\alpha}^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\Phi\left(d_{2}\left(\xi_{1}^{*},0\right)\right)+k_{2}e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{2}^{*},0\right)\right)-\Phi\left(d_{1}\left(\tilde{\alpha}\hat{\xi},0\right)\right)\right)$$

$$+\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)-\Phi\left(d_{1}\left(\alpha\hat{\xi},0\right)\right)\right)$$

$$-\displaystyle\frac{y}{2\gamma\alpha^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\left(\Phi\left(d_{2}\left(\xi_{3}^{*},0\right)\right)-\Phi\left(d_{2}\left(\alpha\hat{\xi},0\right)\right)\right),$$ (A.7)

$$f_{2}(y,\lambda):= \ 1-2\gamma\alpha k_{0}+2\gamma\alpha k_{0}\left(\Phi\left(d_{0}\left(\xi_{3}^{*},0\right)\right)-\Phi\left(d_{0}\left(\alpha\hat{\xi},0\right)\right)+\Phi\left(d_{0}\left(\xi_{2}^{*},0\right)\right)\right.$$

$$\hskip 102.0pt-\left.\Phi\left(d_{0}\left(\tilde{\alpha}\hat{\xi},0\right)\right)+\Phi\left(d_{0}\left(\xi_{1}^{*},0\right)\right)\right)$$

$$+\lambda\left(-1+\Phi\left(d_{0}\left(\xi_{3}^{*},0\right)\right)-\Phi\left(d_{0}\left(\alpha\hat{\xi},0\right)\right)+\Phi\left(d_{0}\left(\xi_{1}^{*},0\right)\right)\right)$$

$$-y\left(\displaystyle\frac{1}{\alpha}\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)-\displaystyle\frac{1}{\alpha}\Phi\left(d_{1}\left(\alpha\hat{\xi},0\right)\right)+\displaystyle\frac{1}{\tilde{\alpha}}\Phi\left(d_{1}\left(\xi_{1}^{*},0\right)\right)\right)$$

$$+2\gamma\alpha(k_{2}-k_{1}-k_{0})\left(\Phi\left(d_{0}\left(\xi_{2}^{*},0\right)\right)-\Phi\left(d_{0}\left(\tilde{\alpha}\hat{\xi},0\right)\right)\right).$$

Note that we suppress again the dependence of $\hat{\xi}$ and $\xi_{i}^{*}$, $i\in\{1,2,3\}$, on $y$ and $\lambda$. By definition, the equation system from Lemma A.2 is solved if there exist $\lambda^{*}$, $y^{*}$ such that $f_{1}(y^{*},\lambda^{*})=0=f_{2}(y^{*},\lambda^{*})$.

To show this, we define

$$C:=\begin{cases}-2\gamma\alpha k_{0}&x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{1},\\ 2\gamma\alpha(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}-k_{0})&x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\in(k_{1},k_{2}],\\ 2\gamma\left(\tilde{\alpha}x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\alpha_{2}k_{2}-\alpha(k_{0}+k_{1})\right)&x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{2},\end{cases}$$

and $h:(C,\infty)\to\mathbb{R}_{\geq 0}$ an arbitrary continuous function with $\lim_{x\to C}h(x)=0$ if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{1}$ resp. $\liminf_{x\to C}h(x)\geq 0$ if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{1}$, and $\lim_{x\to\infty}h(x)=\infty$. Then, we make the following four statements, which are shown at the end of this proof:

$$\lim_{y\to 0}f_{1}(y,\lambda) \geq 0\text{ for all $\lambda>C$},$$ (A.8)

$$\lim_{y\to\infty}f_{1}(y,\lambda) \leq-x_{0}\text{ for all $\lambda>C$},$$ (A.9)

$$\liminf_{\lambda\searrow C}f_{2}(h(\lambda),\lambda) \geq 1,$$ (A.10)

$$\limsup_{\lambda\to\infty}f_{2}(h(\lambda),\lambda) \leq 0.$$ (A.11)

Due to the continuity of $y\mapsto f_{1}(y,\cdot)$ and $\lambda\mapsto f_{2}(h(\lambda),\lambda)$, there exists for each $\lambda>C$ a $y^{*}_{\lambda}\in[0,\infty)$ and for all functions $h$ a $\lambda^{*}_{h}\in(C,\infty)$ such that $f_{1}(y^{*}_{\lambda},\lambda)=0$ and $f_{2}(h(\lambda^{*}_{h}),\lambda^{*}_{h})=0$. If there exists more than one solution, then we take $y_{\lambda}^{*}$ as the smallest one of those (well-defined due to $y$ being bounded from below and the continuity of $f_{1}$ and $f_{2}$). Moreover, we have the following four statements, which are also shown at the end of this proof:

$$\lambda\mapsto y^{*}_{\lambda} \text{ is continuous in $\lambda$ on $(C,\infty)$},$$ (A.12)

$$\lim_{\lambda\searrow C}y^{*}_{\lambda} =0\text{ if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{1}$},$$ (A.13)

$$\liminf_{\lambda\searrow C}y^{*}_{\lambda} \geq 0\text{ if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{1}$},$$ (A.14)

$$\lim_{\lambda\to\infty}y^{*}_{\lambda} =\infty.$$ (A.15)

Now, if we consider $y^{*}_{\lambda}$ as a function of $\lambda$, it fulfills the assumptions of $h$. Hence, there exists a $\lambda^{*}$ such that $f_{2}(y^{*}_{\lambda^{*}},\lambda^{*})=0$. On the other hand, by the definition of $\lambda\mapsto y_{\lambda}$, we have that $f(y_{\lambda^{*}}^{*},\lambda^{*})=0$, which implies the claim.

To finalize the proof, we show the remaining eight statements afterwards. But first, we note that it holds by definition (see Theorem 3.2):

$$0\leq\xi_{1}^{*}\leq\tilde{\alpha}\hat{\xi}\leq\xi_{2}^{*}\leq\alpha\hat{\xi}\leq\xi_{3}^{*}\leq\bar{\xi}.$$ (A.16)

• First of all, we claim that

  $$\xi_{3}^{*}>0\text{ if $\lambda>C$},$$ (A.17)

  $$\xi_{1}^{*}>0\text{ if and only if $\lambda>2\gamma\alpha(k_{2}-k_{1}-k_{0})$},$$ (A.18)

  which we will show directly after the proof of (A.8). So for now assume that (A.17) and (A.18) hold, and let $\lambda>C$.
  By (A.17), $\xi_{3}^{*}\xrightarrow{y\to 0}+\infty$ if $\lambda>C$. Moreover, it follows directly from the definition that

  $$\hat{\xi}\xrightarrow{y\to 0}\begin{cases}0,&\text{if }\lambda\leq 2\gamma\alpha(k_{2}-k_{1}-k_{0}),\\ +\infty,&\text{if }\lambda>2\gamma\alpha(k_{2}-k_{1}-k_{0}).\end{cases}$$ (A.19)

  Thus, if $\lambda\leq 2\gamma\alpha(k_{2}-k_{1}-k_{0})$ then it holds that (i) $\Phi(d_{1}(\xi_{3}^{*},0))-\Phi(d_{1}(\alpha\hat{\xi},0))\xrightarrow{y\to 0}1$. On the other hand, if $\lambda>2\gamma\alpha(k_{2}-k_{1}-k_{0})$, then it holds that (ii) $\Phi\left(d_{1}\left(\xi_{1}^{*},0\right)\right)\xrightarrow{y\to 0}1$ as (A.18) entails that $\xi_{1}^{*}\xrightarrow{y\to 0}\infty$. Now, if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{1}$, it holds that (iii) $\left(k_{0}+k_{1}+\frac{\lambda}{2\gamma\alpha}\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}>x_{0}$ for $\lambda>C=-2\gamma\alpha k_{0}$. Furthermore, if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\in(k_{1},k_{2}]$, we deduce that (iv) $\left(k_{0}+k_{1}+\frac{\lambda}{2\gamma\alpha}\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}>x_{0}$ for $\lambda>C=2\gamma\alpha(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}-k_{0})$ and, if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{2}$, it follows directly that (v) $\left(k_{2}+\frac{\lambda}{2\gamma\tilde{\alpha}}-\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}>x_{0}$ for $\lambda>2\gamma\alpha(k_{2}-k_{1}-k_{0})(\geq C)$. Finally, if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{2}$, then it holds that $C>2\gamma\alpha(k_{2}-k_{1}-k_{0})$ and (vi) $\left(k_{2}+\frac{\lambda}{2\gamma\tilde{\alpha}}-\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}>x_{0}$ for $\lambda>C=2\gamma\left(\tilde{\alpha}x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\alpha_{2}k_{2}-\alpha(k_{0}+k_{1})\right)$. Note that $\tilde{\alpha}\hat{\xi}\leq\xi_{2}^{*}\leq\alpha\hat{\xi}$ (see (A.16)) and hence (vii) $\xi_{2}^{*}$ shows the same limiting behavior as $\hat{\xi}$ for $y\to 0$. Summarizing these results, we conclude for $\lambda>C$:

  1. Properties (ii), (v), and (vi) imply:

     $$\lim_{y\to 0}\left(k_{2}+\displaystyle\frac{\lambda}{2\gamma\tilde{\alpha}}-\displaystyle\frac{\alpha}{\tilde{\alpha}}(k_{2}-k_{1}-k_{0})\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\Phi\left(d_{1}\left(\xi_{1}^{*},0\right)\right)$$

     $$\begin{cases}>x_{0}&\text{ if $\lambda>2\gamma\alpha(k_{2}-k_{1}-k_{0})$ and $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{2}$,}\\ >x_{0}&\text{ if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{2}$,}\\ \geq 0&\text{ else.}\end{cases}$$

  2. One can see directly that $\lim_{y\to 0}\displaystyle\frac{y}{2\gamma\tilde{\alpha}^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\Phi\left(d_{2}\left(\xi_{1}^{*},0\right)\right)=0$.

  3. It holds that $\lim_{y\to 0}k_{2}e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{2}^{*},0\right)\right)-\Phi\left(d_{1}\left(\tilde{\alpha}\hat{\xi},0\right)\right)\right)=0$ due to (vii).

  4. Properties (i), (iii), and (iv) imply:

     $$\lim_{y\to 0}\left(k_{0}+k_{1}+\displaystyle\frac{\lambda}{2\gamma\alpha}\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)-\Phi\left(d_{1}\left(\alpha\hat{\xi},0\right)\right)\right)$$

     $$\begin{cases}>x_{0}&\text{ if $\lambda\in(C,2\gamma\alpha(k_{2}-k_{1}-k_{0})]$ and $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{2}$,}\\ \geq 0&\text{ else.}\end{cases}$$

  5. One can see directly that $\lim_{y\to 0}\displaystyle\frac{y}{2\gamma\alpha^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\left(\Phi\left(d_{2}\left(\xi_{3}^{*},0\right)\right)-\Phi\left(d_{2}\left(\alpha\hat{\xi},0\right)\right)\right)=0$.

  Thus, the claim follows. ∎

• We show this equation by considering the three different cases of $C$:

  Case 1: $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{1}$, i.e., $C=-2\gamma\alpha k_{0}$:
  Let $\lambda>C$, i.e., there exists an $\varepsilon>0$ such that $\lambda=C+2\gamma\alpha\varepsilon$. First, we notice that $\xi_{3}^{*}>0$ if $\tilde{\xi}_{3}^{*}:=\displaystyle\frac{\lambda\alpha}{y}+\displaystyle\frac{2\gamma\alpha^{2}k_{0}}{y}-\displaystyle\frac{2\gamma\alpha^{2}}{y}\left(\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+\displaystyle\frac{\lambda}{\gamma\alpha}\right)}-k_{1}\right)>0$. Now, it holds if we plug in $\lambda$:

  $$\tilde{\xi}_{3}^{*} =\displaystyle\frac{2\gamma\alpha^{2}}{y}\left(-k_{0}+\varepsilon+k_{0}+k_{1}-\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}-2k_{0}+2\varepsilon\right)}\right)=\displaystyle\frac{2\gamma\alpha^{2}}{y}\left(k_{1}+\varepsilon-\sqrt{(k_{1}+\varepsilon)^{2}-\varepsilon^{2}}\right).$$

  Since $y,\gamma,\alpha,\varepsilon>0$ and $k_{1}\geq 0$, we get that $\tilde{\xi}_{3}^{*}>0$ if and only if $(k_{1}+\varepsilon)^{2}>\left(\sqrt{(k_{1}+\varepsilon)^{2}-\varepsilon^{2}}\right)^{2}$, which is equivalent to $\varepsilon^{2}>0$. Thus, the claim, i.e., $\xi_{3}^{*}>0$ for $\lambda>C$.

  Case 2: $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\in(k_{1},k_{2}]$, i.e., $C=2\gamma\alpha(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}-k_{0})$:
  Let $\lambda>C$, i.e., there exists an $\varepsilon>0$ such that $\lambda=C+2\gamma\alpha\varepsilon$. As in the first case, we get that $\xi_{3}^{*}>0$ if $\tilde{\xi}_{3}^{*}>0$ (defined as in Case 1). Then, we get after plugging in $\lambda$:

  $$\tilde{\xi}_{3}^{*} =\displaystyle\frac{2\gamma\alpha^{2}}{y}\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}-k_{0}+\varepsilon+k_{0}+k_{1}-\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+2x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-2k_{1}-2k_{0}+2\varepsilon\right)}\right)$$

  $$=\displaystyle\frac{2\gamma\alpha^{2}}{y}\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\varepsilon-\sqrt{-k_{1}^{2}+2k_{1}x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\right)^{2}+\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\right)^{2}+2\varepsilon k_{1}}\right)$$

  $$=\displaystyle\frac{2\gamma\alpha^{2}}{y}\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\varepsilon-\sqrt{-\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}\right)^{2}+\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\right)^{2}+2\varepsilon k_{1}}\right).$$

  Since $y,\gamma,\alpha,\varepsilon>0$ and $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\geq 0$, we get that $\tilde{\xi}_{3}^{*}>0$ if and only if

  $$\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\varepsilon\right)^{2}>-\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}\right)^{2}+\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\right)^{2}+2\varepsilon k_{1}.$$

  Since $\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\varepsilon\right)^{2}=\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\right)^{2}+2\varepsilon x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\varepsilon^{2}$, this is equivalent to:

  $$0<\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}\right)^{2}-2\varepsilon k_{1}+2\varepsilon x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\varepsilon^{2}=\left(\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}\right)+\varepsilon\right)^{2}.$$

  Now, the claim, i.e., $\xi_{3}^{*}>0$ for $\lambda>C$, follows.

  Case 3: $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{2}$, i.e., $C=2\gamma\left(\tilde{\alpha}x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\alpha_{2}k_{2}-\alpha(k_{0}+k_{1})\right)$:
  Then, we get immediately that $\hat{\xi}>0$ for $\lambda>C(>2\gamma\alpha(k_{2}-k_{1}-k_{0})\text{ since }\alpha_{2}+\tilde{\alpha}=\alpha)$. Hence, it holds that $\xi_{3}^{*}>0$ due to $\xi_{3}^{*}\geq\alpha\hat{\xi}$ (see (A.16)). ∎

• First, let $\lambda>2\gamma\alpha(k_{2}-k_{1}-k_{0})$, i.e., there exists an $\varepsilon>0$ such that $\lambda=2\gamma\alpha(k_{2}-k_{1}-k_{0})+2\gamma\tilde{\alpha}\varepsilon$. We note that $\hat{\xi}>0$ for this $\lambda$ which implies that $\xi_{1}^{*}>0$ if and only if $\tilde{\xi}_{1}^{*}>0$. Now, if the term under the square root in the maximum in the formula of $\tilde{\xi}_{1}^{*}$ (for the definition, see Theorem 3.2) is negative, the claim follows immediately. Hence, we assume that the term is non-negative. If we plug in $\lambda$ into $\tilde{\xi}_{1}^{*}$, we get:

  $$\tilde{\xi}_{1}^{*} =\displaystyle\frac{2\gamma\tilde{\alpha}}{y}\left(\displaystyle\frac{\lambda}{2\gamma}-\alpha(k_{2}-k_{1}-k_{0})+\tilde{\alpha}k_{2}-\sqrt{(\alpha(k_{0}+k_{1})-\alpha_{2}k_{2})^{2}-\alpha^{2}k_{0}^{2}+\displaystyle\frac{\lambda}{\gamma}(\alpha k_{1}-\alpha_{2}k_{2})}\right)$$

  $$=\displaystyle\frac{2\gamma\tilde{\alpha}}{y}\left(\tilde{\alpha}\varepsilon+\tilde{\alpha}k_{2}-\sqrt{\alpha_{2}^{2}k_{2}^{2}+2\alpha^{2}k_{1}k_{2}-2\alpha\alpha_{2}k_{2}^{2}-\alpha^{2}k_{1}^{2}+2\alpha\tilde{\alpha}k_{1}\varepsilon-2\tilde{\alpha}\alpha_{2}k_{2}\varepsilon}\right)$$

  $$=\displaystyle\frac{2\gamma\tilde{\alpha}^{2}}{y}\left(k_{2}+\varepsilon-\frac{1}{\tilde{\alpha}}\sqrt{k_{2}^{2}\alpha_{2}(\alpha_{2}-2\alpha)+\alpha^{2}k_{1}(2k_{2}-k_{1})+2\tilde{\alpha}\varepsilon(\alpha k_{1}-\alpha_{2}k_{2})}\right).$$

  Since $y,\gamma,\tilde{\alpha},\varepsilon>0$ and $k_{2}\geq 0$, it holds that $\tilde{\xi}_{1}^{*}>0$ if and only if

  $$k_{2}^{2}+2\varepsilon k_{2}+\varepsilon^{2}>\frac{1}{\tilde{\alpha}^{2}}\left(k_{2}^{2}\alpha_{2}(-\tilde{\alpha}-\alpha)+\alpha^{2}k_{1}(2k_{2}-k_{1})+2\tilde{\alpha}\varepsilon(\alpha k_{1}-\alpha_{2}k_{2})\right).$$

  This is equivalent to:

  $$L:=k_{2}^{2}\left(1+\displaystyle\frac{\alpha_{2}}{\tilde{\alpha}}+\displaystyle\frac{\alpha\alpha_{2}}{\tilde{\alpha}^{2}}\right)+2\varepsilon\left(k_{2}\left(1+\displaystyle\frac{\alpha_{2}}{\tilde{\alpha}}\right)-k_{1}\frac{\alpha}{\tilde{\alpha}}\right)+\varepsilon^{2}-\displaystyle\frac{\alpha^{2}}{\tilde{\alpha}^{2}}k_{1}(2k_{2}-k_{1})>0.$$

  Now, we notice with $\alpha=\tilde{\alpha}+\alpha_{2}$ that

  $$1+\displaystyle\frac{\alpha_{2}}{\tilde{\alpha}}+\displaystyle\frac{\alpha\alpha_{2}}{\tilde{\alpha}^{2}}=\displaystyle\frac{\tilde{\alpha}^{2}+\alpha_{2}\tilde{\alpha}+\tilde{\alpha}\alpha_{2}+\alpha_{2}^{2}}{\tilde{\alpha}^{2}}=\displaystyle\frac{(\tilde{\alpha}+\alpha_{2})^{2}}{\tilde{\alpha}^{2}}=\displaystyle\frac{\alpha^{2}}{\tilde{\alpha}^{2}}.$$

  Then, it follows with $L$ as above since $1+\frac{\alpha_{2}}{\tilde{\alpha}}=\frac{\alpha}{\tilde{\alpha}}$:

  $$\displaystyle\frac{\tilde{\alpha}^{2}}{\alpha^{2}}L =k_{2}^{2}+2\displaystyle\frac{\tilde{\alpha}}{\alpha}\varepsilon(k_{2}-k_{1})+\displaystyle\frac{\tilde{\alpha}^{2}}{\alpha^{2}}\varepsilon^{2}-2k_{1}k_{2}+k_{1}^{2}$$

  $$=(k_{2}-k_{1})^{2}+2\displaystyle\frac{\tilde{\alpha}}{\alpha}\varepsilon(k_{2}-k_{1})+\left(\displaystyle\frac{\tilde{\alpha}}{\alpha}\varepsilon\right)^{2}=\left((k_{2}-k_{1})+\frac{\tilde{\alpha}}{\alpha}\varepsilon\right)^{2}.$$

  Hence, $L>0$. Second, we note that $\hat{\xi}=0$ (and thus also $\xi_{1}^{*}$) for $\lambda\leq 2\gamma\alpha(k_{2}-k_{1}-k_{0})$ by definition. Hence, the claim $\xi_{1}^{*}>0$ if and only if $\lambda>2\gamma\alpha(k_{2}-k_{1}-k_{0})$ follows. ∎

• By definition, it holds that $\bar{\xi}\xrightarrow{y\to\infty}0$. Then, we get the claim due to (A.16) and $\lim_{x\to 0}\Phi(d_{i}(x,0))=0$ for $i\in\{1,2\}$. ∎

• Recall that $h$ is a continuous function from $(C,\infty)\to\mathbb{R}_{\geq 0}$ with some limiting properties at $C$ and $\infty$, and we plug this function into $f_{2}$ with $y=h(\lambda)$. To start, we give a small overview of the following proof: To show (A.10), we must find the limiting behavior of $\xi_{1}^{*}$, $\tilde{\alpha}\hat{\xi}$, $\xi_{2}^{*}$, $\alpha\hat{\xi}$, and $\xi_{3}^{*}$ when $\lambda\searrow C$. However, this behavior heavily depends on $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}$, i.e., the different values of $C$ and the limiting properties of $h$. Hence, we have to distinguish several cases. First, we differentiate between the different limiting properties for $\lambda\searrow C$ depending if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{1}$ is true. If this is true, we first analyze the limit of $\xi_{3}^{*}$. After that, we have to separate the cases depending if $C$ is bigger, smaller, or equal to $2\gamma\alpha(k_{2}-k_{1}-k_{0})$. For the equality case, we even have to separate along the limiting behavior of $\hat{\xi}$ to derive the limiting behavior for $\xi_{1}^{*}$ and $\xi_{2}^{*}$. If $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{1}$ is not true, we only have to separate along the possible limiting values of $h$.

  Case 1: $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{1}$, i.e., $\lim_{x\to C}h(x)=0$:
  Since $\lim_{\lambda\searrow C}h(\lambda)=0$ and $C>-2\gamma\alpha k_{0}$, it holds that (i) $\xi_{3}^{*}\to+\infty$ for $\lambda\searrow C$, i.e., $y=h(\lambda)\xrightarrow{\lambda\searrow C}0$. Indeed, it holds: If $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{2}$, we get that $C>2\gamma\alpha(k_{2}-k_{1}-k_{0})$ and hence $\hat{\xi}\xrightarrow{\lambda\searrow C}+\infty$ since $y=h(\lambda)\xrightarrow{\lambda\searrow C}0$. Now, (A.16) implies that $\xi_{3}^{*}\xrightarrow{\lambda\searrow C}+\infty$. If $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{2}$, it holds that $C=2\gamma\alpha\left(e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}-k_{0}\right)$ since $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{1}$ by assumption and we get that:

  $$\xi_{3}^{*}\geq\displaystyle\frac{\alpha\lambda+2\gamma\alpha^{2}\left(k_{0}+k_{1}-\sqrt{k_{1}^{2}+k_{1}(2k_{0}+\frac{\lambda}{\alpha\gamma})}\right)}{h(\lambda)}=:\displaystyle\frac{D}{h(\lambda)}.$$

  When calculating the limit for $\lambda\searrow C$ for $D$, it holds that

  $$\lim_{\lambda\to C}D =2\gamma\alpha^{2}\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}-k_{0}+k_{0}+k_{1}-\sqrt{k_{1}^{2}+k_{1}(2k_{0}+2x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-2k_{1}-2k_{0})}\right)$$

  $$=2\gamma\alpha^{2}\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-\sqrt{-k_{1}^{2}+2k_{1}x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}}\right).$$

  Now, we see that $\lim_{\lambda\searrow C}D>0$ if and only if $\left(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}\right)^{2}>0$ which is true by assumption. Thus, $\xi_{3}^{*}\xrightarrow{\lambda\searrow C}+\infty$ since $h(\lambda)\xrightarrow{\lambda\searrow C}0$, i.e., (i) is proven.
  Next, if $C>2\gamma\alpha(k_{2}-k_{1}-k_{0})$, we get that ((ii).1) $\xi_{1}^{*}\xrightarrow{\lambda\searrow C}+\infty$ due to (A.18) and $h(\lambda)\xrightarrow{\lambda\searrow C}0$ and thus also ((ii).2) $\alpha\hat{\xi},\xi_{2}^{*},\tilde{\alpha}\hat{\xi}\xrightarrow{\lambda\searrow C}+\infty$ due to (A.16). If $C<2\gamma\alpha(k_{2}-k_{1}-k_{0})$, we get that ((iii).1) $\alpha\hat{\xi}\xrightarrow{\lambda\searrow C}0$ due to (A.19) and $y=h(\lambda)\xrightarrow{\lambda\searrow C}0$ which can be factored out. Hence, also ((iii).2) $\xi_{1}^{*},\tilde{\alpha}\hat{\xi},\xi_{2}^{*}\xrightarrow{\lambda\searrow C}0$ due to (A.16). The remaining case that $C=2\gamma\alpha(k_{2}-k_{1}-k_{0})$ needs a closer look. First, we note that then $k_{2}=x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}$ and hence $k_{2}>k_{1}$ by the assumption of Case 1. There, we have to distinguish three more cases for the limiting behavior of $\xi_{1}^{*}$, $\tilde{\alpha}\hat{\xi}$, $\xi_{2}^{*}$, and $\alpha\hat{\xi}$ depending on the behavior of $\hat{\xi}$. Note that $\hat{\xi}$ can have multiple accumulation points for $\lambda\searrow C$ (all in [0,$\infty$]) despite $\hat{\xi}$ being continuous in $\lambda$. Now, let $\lambda_{n}$ be a sequence converging to the $\liminf$, i.e., $\lim_{n\to\infty}f_{2}(h(\lambda_{n}),\lambda_{n})=\liminf_{\lambda\searrow C}f_{2}(h(\lambda),\lambda)$. Note that $\lambda_{n}\xrightarrow{n\to\infty}C$ from above. We show (A.10) then for each accumulation point separately by possibly switching to subsequences, i.e., we can assume without loss of generality that $\lim_{n\to\infty}\hat{\xi}^{\lambda_{n}}$ exists (in [0,$\infty$]) and have to consider the following cases:

  Case 1.1: $\hat{\xi}^{\lambda_{n}}\xrightarrow{n\to\infty}0$:
  Then, it follows immediately that (iv) $\xi_{1}^{*,\lambda_{n}},\xi_{2}^{*,\lambda_{n}}\xrightarrow{n\to\infty}0$ due to (A.16).

  Case 1.2: $\hat{\xi}^{\lambda_{n}}\xrightarrow{n\to\infty}+\infty$:
  Here, we first note that when applying the same argument as in the proof of (A.18) (with $\varepsilon=0$) to the second term in the definition of $\tilde{\xi}_{1}^{*,\lambda_{n}}$, this second term is non-negative for $\lambda=C$, i.e., $\tilde{\xi}_{1}^{*,\lambda_{n}}\geq\tilde{\alpha}\hat{\xi}^{\lambda_{n}}$ when taking the limit $n\to\infty$. Then, we obtain ((v).1) $\xi_{1}^{*,\lambda_{n}}\xrightarrow{n\to\infty}+\infty$. In particular, we get ((v).2) $\xi_{1}^{*,\lambda_{n}},\tilde{\alpha}\hat{\xi}^{\lambda_{n}},\xi_{2}^{*,\lambda_{n}},\alpha\hat{\xi}^{\lambda_{n}}\xrightarrow{n\to\infty}+\infty$ due to (A.16).

  Case 1.3: $\hat{\xi}^{\lambda_{n}}\xrightarrow{n\to\infty}c\in(0,\infty)$:
  First, since $k_{2}>k_{1}$, we get by plugging in $\lambda=2\gamma\alpha(k_{2}-k_{1}-k_{0})$ into $\tilde{\xi}_{2}^{*}$ that $\tilde{\xi}_{2}^{*,\lambda_{n}}\xrightarrow{n\to\infty}+\infty$. The reason is that $y_{n}=h(\lambda_{n})\xrightarrow{n\to\infty}0$ and the nominator in $\tilde{\xi}_{2}^{*,\lambda_{n}}$ converges to a positive number, i.e.:

  $$\alpha\lambda k_{2}-\gamma\alpha^{2} (k_{2}-k_{1})^{2}+2\gamma\alpha^{2}k_{0}(k_{2}-k_{1})-\lambda_{n}\alpha k_{1}$$

  $$\xrightarrow{n\to\infty} \ 2\gamma\alpha^{2}k_{2}(k_{2}-k_{1}-k_{0})-\gamma\alpha^{2}(k_{2}-k_{1})^{2}+2\gamma\alpha^{2}k_{0}(k_{2}-k_{1})-2\gamma\alpha^{2}k_{1}(k_{2}-k_{1}-k_{0})$$

  $$= \ 2\gamma\alpha^{2}(k_{2}-k_{1}-k_{0})(k_{2}-k_{1})-\gamma\alpha^{2}(k_{2}-k_{1})(k_{2}-k_{1}-2k_{0})$$

  $$= \ \gamma\alpha^{2}(k_{2}-k_{1})(2k_{2}-2k_{1}-2k_{0}-k_{2}+k_{1}+2k_{0})=\gamma\alpha^{2}(k_{2}-k_{1})^{2}>0.$$

  Hence, ((vi).1) $\xi_{2}^{*,\lambda_{n}}\xrightarrow{n\to\infty}\alpha c$. Second, we note that $\tilde{\xi}_{1}^{*,\lambda_{n}}\geq\tilde{\alpha}\hat{\xi}^{\lambda_{n}}$ for $n\to\infty$ by the same argument as in Case 1.2. Thus, ((vi).2) $\xi_{1}^{*,\lambda_{n}}\xrightarrow{n\to\infty}\tilde{\alpha}c$.

  Summarizing, we get in all subcases of Case 1, that $\Phi(d_{0}(\xi_{3}^{*},0))\xrightarrow{\lambda\searrow C}1$ since $\lim_{x\to\infty}\Phi\left(d_{0}\left(x,0\right)\right)=1$ due to (i) and $-\Phi(d_{0}(\alpha\hat{\xi}^{\lambda_{n}},0))+\Phi(d_{0}(\xi_{2}^{*,\lambda_{n}},0))-\Phi(d_{0}(\tilde{\alpha}\hat{\xi}^{\lambda_{n}},0))+\Phi(d_{0}(\xi_{1}^{*,\lambda_{n}},0))\xrightarrow{n\to\infty}0$ due to either (ii), (iii), (iv), (v), or (vi) depending on $C$. Hence, it holds that $\Phi(d_{0}(\xi_{3}^{*,\lambda_{n}},0))-\Phi(d_{0}(\alpha\hat{\xi}^{\lambda_{n}},0))+\Phi(d_{0}(\xi_{2}^{*,\lambda_{n}},0))-\Phi(d_{0}(\tilde{\alpha}\hat{\xi}^{\lambda_{n}},0))+\Phi(d_{0}(\xi_{1}^{*,\lambda_{n}},0))\xrightarrow{n\to\infty}1$ since $\lim_{x\to\infty}\Phi\left(d_{0}\left(x,0\right)\right)=1$ and $\lambda_{n}\xrightarrow{n\to\infty}C$ from above. In particular, we get that $\lambda(-1+\Phi(d_{0}(\xi_{3}^{*,\lambda_{n}},0))-\Phi(d_{0}(\alpha\hat{\xi}^{\lambda_{n}},0))+\Phi(d_{0}(\xi_{1}^{*,\lambda_{n}},0)))+2\gamma\alpha(k_{2}-k_{1}-k_{0})(\Phi(d_{0}(\xi_{2}^{*,\lambda_{n}},0))-\Phi(d_{0}(\tilde{\alpha}\hat{\xi}^{\lambda_{n}},0)))\xrightarrow{n\to\infty}0$ since $C=2\gamma\alpha(k_{2}-k_{1}-k_{0})$ and $\lambda_{n}\xrightarrow{n\to\infty}C$ from above. Thus, the claim follows.

  Case 2: $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\leq k_{1}$, i.e., $\liminf_{x\searrow C}h(x)\geq 0$ and $C=-2\gamma\alpha k_{0}$:
  For this case, we have to distinguish two more subcases depending on the possible limit of $h$. However, $h$ can have multiple accumulation points. As before, we consider each accumulation point separately, denote by $\lambda_{n}$ a sequence converging to the $\liminf$, and assume, without loss of generality by possibly switching to a subsequence, that $\lambda_{n}\xrightarrow{n\to\infty}C$ from above and $\lim_{n\to\infty}h(\lambda_{n})$ exists (in $[0,\infty]$). This gives us the following cases:

  Case 2.1: $\lim_{n\to\infty}h(\lambda_{n})>0$:
  Then, we get that $\bar{\xi}^{\lambda_{n}}\xrightarrow{n\to\infty}0$ since $\bar{\xi}^{\lambda_{n}}=\frac{\alpha(\lambda_{n}-C)}{h(\lambda_{n})}$ for $C=-2\gamma\alpha k_{0}$ which implies that $\hat{\xi}^{\lambda_{n}},\xi_{1}^{*,\lambda_{n}},\xi_{2}^{*,\lambda_{n}},\xi_{3}^{*,\lambda_{n}}\xrightarrow{n\to\infty}0$ using (A.16). Thus, it holds that $\lim_{n\to\infty}f_{2}(h(\lambda_{n}),\lambda_{n})=1-2\gamma\alpha k_{0}-C=1$ using that $\lim_{x\to 0}\Phi(d_{i}(x,0))=0$ for $i\in\{0,1\}$.

  Case 2.2: $\lim_{n\to\infty}h(\lambda_{n})=0$:
  Here, we first rewrite $f_{2}$ into:

  $$f_{2}(h(\lambda_{n}),\lambda_{n}):= \ 1+2\gamma\alpha k_{0}\left(-1+\Phi\left(d_{0}\left(\xi_{3}^{*,\lambda_{n}},0\right)\right)-\Phi\left(d_{0}\left(\alpha\hat{\xi}^{\lambda_{n}},0\right)\right)+\Phi\left(d_{0}\left(\xi_{1}^{*,\lambda_{n}},0\right)\right)\right)$$

  $$+\lambda_{n}\left(-1+\Phi\left(d_{0}\left(\xi_{3}^{*,\lambda_{n}},0\right)\right)-\Phi\left(d_{0}\left(\alpha\hat{\xi}^{\lambda_{n}},0\right)\right)+\Phi\left(d_{0}\left(\xi_{1}^{*,\lambda_{n}},0\right)\right)\right)$$

  $$-h(\lambda_{n})\left(\displaystyle\frac{1}{\alpha}\Phi\left(d_{1}\left(\xi_{3}^{*,\lambda_{n}},0\right)\right)-\displaystyle\frac{1}{\alpha}\Phi\left(d_{1}\left(\alpha\hat{\xi}^{\lambda_{n}},0\right)\right)+\displaystyle\frac{1}{\tilde{\alpha}}\Phi\left(d_{1}\left(\xi_{1}^{*,\lambda_{n}},0\right)\right)\right)$$

  $$+2\gamma\alpha(k_{2}-k_{1})\left(\Phi\left(d_{0}\left(\xi_{2}^{*,\lambda_{n}},0\right)\right)-\Phi\left(d_{0}\left(\tilde{\alpha}\hat{\xi}^{\lambda_{n}},0\right)\right)\right)$$

  $$= \ 1+(\lambda_{n}-C)\left(-1+\Phi\left(d_{0}\left(\xi_{3}^{*,\lambda_{n}},0\right)\right)-\Phi\left(d_{0}\left(\alpha\hat{\xi}^{\lambda_{n}},0\right)\right)+\Phi\left(d_{0}\left(\xi_{1}^{*,\lambda_{n}},0\right)\right)\right)$$

  $$-h(\lambda_{n})\left(\displaystyle\frac{1}{\alpha}\Phi\left(d_{1}\left(\xi_{3}^{*,\lambda_{n}},0\right)\right)-\displaystyle\frac{1}{\alpha}\Phi\left(d_{1}\left(\alpha\hat{\xi}^{\lambda_{n}},0\right)\right)+\displaystyle\frac{1}{\tilde{\alpha}}\Phi\left(d_{1}\left(\xi_{1}^{*,\lambda_{n}},0\right)\right)\right)$$

  $$+2\gamma\alpha(k_{2}-k_{1})\left(\Phi\left(d_{0}\left(\xi_{2}^{*,\lambda_{n}},0\right)\right)-\Phi\left(d_{0}\left(\tilde{\alpha}\hat{\xi}^{\lambda_{n}},0\right)\right)\right),$$

  where we used that $C=-2\gamma\alpha k_{0}$. Now, the claim follows, i.e., $\liminf_{\lambda\searrow C}f_{2}(h(\lambda),\lambda)\geq 1$ since $\lim_{n\to\infty}h(\lambda_{n})=0$, $\lambda_{n}\xrightarrow{n\to\infty}C$ from above, $\xi_{2}^{*,\lambda_{n}}\geq\tilde{\alpha}\hat{\xi}^{\lambda_{n}}$, and $\Phi(d_{0}(\cdot,0))$ being non-decreasing. ∎

• For this proof, we have to consider three cases depending on the limiting behavior of $\frac{\lambda}{h(\lambda)}$ for $\lambda\to\infty$. As in the proof of (A.10), we have here possibly multiple accumulation points that we treat separately and denote by $\lambda_{n}$ the sequence converging to the $\limsup$. Note that $\lambda_{n}\xrightarrow{n\to\infty}\infty$ in this case. Hence, we can assume, without loss of generality by possibly switching to a subsequence, that $\lim_{n\to\infty}\frac{\lambda_{n}}{h(\lambda_{n})}$ exists (in $[0,\infty]$) and consider the three cases that $\lim_{n\to\infty}\frac{\lambda_{n}}{h(\lambda_{n})}=0$, $\lim_{n\to\infty}\frac{\lambda_{n}}{h(\lambda_{n})}=c\in(0,\infty)$, and $\lim_{n\to\infty}\frac{\lambda_{n}}{h(\lambda_{n})}=\infty$:

  Case 1: $\lim_{n\to\infty}\frac{\lambda_{n}}{h(\lambda_{n})}=0$:
  It holds that $\bar{\xi}^{\lambda_{n}}\xrightarrow{n\to\infty}0$ using $y_{n}=h(\lambda_{n})$ and the assumption. Therefore, we also get that $\hat{\xi}^{\lambda_{n}},\xi_{1}^{*,\lambda_{n}},\xi_{2}^{*,\lambda_{n}},\xi_{3}^{*,\lambda_{n}}\xrightarrow{n\to\infty}0$ due to (A.16). Thus, the claim follows using $\lim_{x\to 0}\Phi(d_{i}(x,0))=0$ for $i\in\{0,1\}$.

  Case 2: $\lim_{n\to\infty}\frac{\lambda_{n}}{h(\lambda_{n})}=c\in(0,\infty)$:
  In this case, it holds by assumption that $\hat{\xi}^{\lambda_{n}}\xrightarrow{n\to\infty}c$, and $\bar{\xi}^{\lambda_{n}},\xi_{3}^{*,\lambda_{n}}\xrightarrow{n\to\infty}\alpha c$. Moreover, it holds that $\lim_{n\to\infty}\xi_{2}^{*,\lambda_{n}}\in[\tilde{\alpha}c,\alpha c]$ by (A.16) and $\xi_{1}^{*,\lambda_{n}}\xrightarrow{n\to\infty}\tilde{\alpha}c$ since $\frac{\lambda_{n}-\sqrt{l(\lambda_{n})}}{h(\lambda_{n})}\xrightarrow{n\to\infty}c$ for any affine function $l$. Next, we note that $\Phi(d_{i}(ac,0))\in(0,1)$ for all $a>0$ and $i\in\{0,1\}$. In total, we get that $\lim_{n\to\infty}-1+\Phi(d_{0}(\xi_{3}^{*,\lambda_{n}},0))-\Phi(d_{0}(\alpha\hat{\xi}^{\lambda_{n}},0))+\Phi(d_{0}(\xi_{1}^{*,\lambda_{n}},0))=-1+\Phi(d_{0}(\tilde{\alpha}c,0))<0$ and $\lim_{n\to\infty}\frac{1}{\alpha}\Phi(d_{1}(\xi_{3}^{*,\lambda_{n}},0))-\frac{1}{\alpha}\Phi(d_{1}(\alpha\hat{\xi}^{\lambda_{n}},0))+\frac{1}{\tilde{\alpha}}\Phi(d_{1}(\xi_{1}^{*,\lambda_{n}},0))=\frac{1}{\tilde{\alpha}}\Phi(d_{1}(\tilde{\alpha}c,0))>0$. Thus, the claim follows since $\lambda_{n},h(\lambda_{n})\xrightarrow{n\to\infty}+\infty$.

  Case 3: $\lim_{n\to\infty}\frac{\lambda_{n}}{h(\lambda_{n})}=\infty$:
  In this case, we get that $\xi_{1}^{*,\lambda_{n}}\xrightarrow{n\to\infty}\infty$ as $\lim_{n\to\infty}\frac{\lambda_{n}-\sqrt{l(\lambda_{n})}}{h(\lambda_{n})}=\infty$ for any affine function $l$. Hence, it holds that $\hat{\xi}^{\lambda_{n}},\bar{\xi}^{\lambda_{n}},\xi_{2}^{*,\lambda_{n}},\xi_{3}^{*,\lambda_{n}}\xrightarrow{n\to\infty}\infty$ due to (A.16). Thus, we get the claim using $\lim_{n\to\infty}h(\lambda_{n})=\infty$, $\lim_{n\to\infty}\lambda_{n}=\infty$, and $\lim_{x\to\infty}\Phi(d_{i}(x,0))=1$ for $i\in\{0,1\}$. ∎

• To show this, it is sufficient to show that (i) $\lambda\mapsto f_{1}(\cdot,\lambda)$ is strictly increasing, (ii) $y\mapsto f_{1}(y,\cdot)$ is non-increasing and (iii) $f_{1}$ is jointly continuous on $\mathbb{R}_{>0}\times(C,\infty)$. Indeed, let $\lambda,\lambda_{n}>C$ with $\lambda_{n}\xrightarrow{n\to\infty}\lambda$. Due to the existence of the zero root (see the main part of the proof) and the uniqueness of the zero root (since $\lambda\mapsto f_{1}(\cdot,\lambda)$ is strictly increasing), there exist then unique $y_{\lambda}^{*},y_{\lambda_{n}}^{*}\in(0,\infty)$ such that $f_{1}(y_{\lambda}^{*},\lambda)=0=f_{1}(y_{\lambda_{n}}^{*},\lambda_{n})$. Then, we have to show that $y_{\lambda_{n}}^{*}\xrightarrow{n\to\infty}y_{\lambda}^{*}$. Due to $y\mapsto f_{1}(y,\cdot)$ being non-increasing and $\lambda_{n}\xrightarrow{n\to\infty}\lambda$, it holds that $y_{\lambda_{n}}^{*}\in[y_{\max_{k\in\mathbb{N}}\{\lambda,\lambda_{k}\}}^{*},y_{\min_{k\in\mathbb{N}}\{\lambda,\lambda_{k}\}}^{*}]$ (maximum and minimum exist due to the convergence of $\lambda_{n}$ to $\lambda$). Then, there exist a subsequence $y_{\lambda_{n_{l}}}^{*}$ and a $\tilde{y}$ such that $y_{\lambda_{n_{l}}}^{*}\xrightarrow{l\to\infty}\tilde{y}$. Moreover, we know that $f_{1}(y_{\lambda}^{*},\lambda)=0=f_{1}(y_{\lambda_{n_{l}}}^{*},\lambda_{n_{l}})\xrightarrow{l\to\infty}f_{1}(\tilde{y},\lambda)$ due to the joint continuity of $f_{1}$. Now, the uniqueness of the zero root implies that $\tilde{y}=y_{\lambda}^{*}$. Hence, all subsequences converge to $y_{\lambda}^{*}$ and thus also the sequence itself. Therefore, (A.12) would follow provided we can show (i), (ii), and (iii) from the beginning:

  First, the joint continuity (i.e., property (iii)) follows directly from the definition of $f_{1}$.

  Second, we derive property (i), i.e., the strict monotonicity in $\lambda$, from the formula of $\hat{X}_{T}$ (see Theorem 3.2), which gives us the claim. Note that we add in the following paragraph a superscript to $\hat{X}_{T}$, $\xi_{1}^{*}$, $\xi_{2}^{*}$, and $\xi_{3}^{*}$ when we take these values for a certain fixed $\lambda$.
  Indeed, if for all $\omega\in\Omega$, we can show that $\hat{X}_{T}^{\lambda_{1}}(\xi_{T}(\omega))\geq\hat{X}_{T}^{\lambda_{2}}(\xi_{T}(\omega))$ with a strict inequality for a set with positive probability for all $\lambda_{1}>\lambda_{2}\,(>C)$, then also $f_{1}(\cdot,\lambda_{1})>f_{1}(\cdot,\lambda_{2})$ since $f_{1}=\mathbb{E}[\xi_{T}\hat{X}_{T}]-x_{0}$ and $\xi_{T}>0$. Therefore, let $\lambda_{1}>\lambda_{2}\,(>C)$: Due to (A.17) and Proposition 3.4, we conclude that for all $\lambda>C$ (and hence for $\lambda_{1}$ and $\lambda_{2}$) $\hat{X}_{T}\not\equiv 0$, i.e., $\hat{X}_{T}(\xi_{T})>0$ for $\xi_{T}$ small enough. The formula of $\hat{X}_{T}$ (see Theorem 3.2) implies that for fixed $y$ the slope remains unchanged in each interval $(0,\xi_{1}^{*}]$, $(\tilde{\alpha}\hat{\xi},\xi_{2}^{*}]$, resp. $(\alpha\hat{\xi},\xi_{3}^{*}]$ when changing $\lambda$, but the interval boundaries change. Note that the slope (as a function of $\xi_{T}$) is strictly negative in $(0,\xi_{1}^{*}]$ and $(\alpha\hat{\xi},\xi_{3}^{*}]$, and constant otherwise. Therefore, due to $\hat{X}_{T}$ being non-increasing and a non-increasing function getting larger when being shifted to the right, it is sufficient to show that all interval boundaries do not decrease when $\lambda$ increases and at least one boundary value strictly increases when $\lambda$ increases:
  It follows directly from the definition of $\hat{X}_{T}$ that $\lim_{\xi_{T}\to 0}\hat{X}_{T}^{\lambda_{1}}(\xi_{T})>\lim_{\xi_{T}\to 0}\hat{X}_{T}^{\lambda_{2}}(\xi_{T})\,(>0)$. Hence, due to having the same slope and the continuity of $\hat{X}_{T}^{\lambda_{1}}$ (resp. $\hat{X}_{T}^{\lambda_{2}}$) in $\xi_{T}$ on $(0,\xi_{1}^{*,\lambda_{1}}]$ (resp. $(0,\xi_{1}^{*,\lambda_{2}}]$), we conclude that $\xi_{1}^{*,\lambda_{1}}>\xi_{1}^{*,\lambda_{2}}$ since $\hat{X}_{T}^{\lambda_{1}}(\xi_{T}=\xi_{1}^{*,\lambda_{1}})=k_{2}=\hat{X}_{T}^{\lambda_{2}}(\xi_{T}=\xi_{1}^{*,\lambda_{2}})$, i.e., the interval is strictly increasing in $\lambda$. Next, we observe immediately from its definition that $\hat{\xi}^{\lambda_{1}}\geq\hat{\xi}^{\lambda_{2}}$. For $\xi_{2}^{*}$ and $\xi_{3}^{*}$, we show this property by proving that their derivatives with respect to $\lambda$ are non-negative. Let $\tilde{\xi}_{2}^{*}$ be defined as in Theorem 3.2 and $\tilde{\xi}_{3}^{*}:=\bar{\xi}-\frac{2\gamma\alpha^{2}}{y}\left(\sqrt{k_{1}^{2}+k_{1}\left(2k_{0}+\frac{\lambda}{\gamma\alpha}\right)}-k_{1}\right)$ with $\bar{\xi}$ as in Theorem 3.2. Then, we get:

  $$\frac{\partial}{\partial\lambda}\tilde{\xi}_{2}^{*} =\frac{\alpha}{y}-\frac{\alpha k_{1}}{yk_{2}}=\frac{\alpha}{y}\left(1-\frac{k_{1}}{k_{2}}\right)\geq 0$$

  $$\Rightarrow \frac{\partial^{-}}{\partial\lambda}\xi_{2}^{*} \geq 0,$$

  $$\frac{\partial}{\partial\lambda}\tilde{\xi}_{3}^{*} =\begin{cases}\displaystyle\frac{\alpha}{y}&\textit{ if $k_{1}=0$},\\ \displaystyle\frac{\alpha}{y}-\displaystyle\frac{2\gamma\alpha^{2}}{y}\displaystyle\frac{\frac{1}{2}\cdot\frac{k_{1}}{\gamma\alpha}}{\sqrt{k_{1}^{2}+k_{1}(2k_{0}+\frac{\lambda}{\gamma\alpha})}}=\frac{\alpha}{y}\left(1-\displaystyle\frac{k_{1}}{\sqrt{k_{1}^{2}+k_{1}(2k_{0}+\frac{\lambda}{\gamma\alpha})}}\right)&\textit{ if $k_{1}>0$},\end{cases}$$

  $$\geq 0,$$

  $$\Rightarrow \frac{\partial^{-}}{\partial\lambda}\xi_{3}^{*} \geq 0,$$

  where $\frac{\partial^{-}}{\partial\lambda}$ denotes the left side derivative with respect to $\lambda$. Summarizing, $\xi_{1}^{*}$ strictly increases and $\xi_{2}^{*}$ and $\xi_{3}^{*}$ non-decrease in $\lambda$. Thus, property (i) follows.

  Third, we prove property (ii), i.e., that $f_{1}$ is non-increasing in $y$. Let $\hat{X}_{T}(y)$ be a function of $y$ as in Theorem 3.2. It follows directly from the definitions of $\hat{\xi}$, $\xi_{1}^{*}$, $\xi_{2}^{*}$, and $\xi_{3}^{*}$ that they are non-increasing in $y$ for fixed $\lambda$. Now, let $y_{1}<y_{2}$ and $\xi_{T}>0$ arbitrary. Then, the claim follows if $\hat{X}_{T}(y_{1},\xi_{T})\geq\hat{X}_{T}(y_{2},\xi_{T})$. If $\xi_{T}$ is in the same interval for both values $y_{1}$ and $y_{2}$, the claim is obvious due to (A.7). Since $\hat{\xi}$, $\xi_{1}^{*}$, $\xi_{2}^{*}$, and $\xi_{3}^{*}$ are non-increasing in $y$ and $\hat{X}_{T}$ is non-increasing in $\xi_{T}$ (check Remark 3.3(c)), the claim also follows if $\xi_{T}$ is in different intervals for $y_{1}$ and $y_{2}$. ∎

• For the proof of this equation, we have to consider the two cases $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}<k_{2}$ and $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\geq k_{2}$. Note that if $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}=k_{2}$, we get that $C=2\gamma\alpha(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}-k_{0})=2\gamma\left(\tilde{\alpha}x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\alpha_{2}k_{2}-\alpha(k_{0}+k_{1})\right)$. Moreover, notice that $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}>k_{1}$ by assumption.

  Case 1: $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}<k_{2}$, i.e., $C=2\gamma\alpha(x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}-k_{1}-k_{0})$:
  For $\lambda_{\varepsilon}=C+2\gamma\alpha e^{\int_{0}^{T}r_{s}\mathrm{d}s}\varepsilon$, it holds that $\hat{\xi}=0$ for $\varepsilon$ small enough and hence also $\xi_{1}^{*}=\xi_{2}^{*}=0$ due to (A.16). However, it holds with (A.17) that $\xi_{3}^{*}>0$. Note that $\lim_{x\to 0}\Phi(d_{j}(x,0))=0$ for $j\in\{0,1,2\}$. Thus, $f_{1}$ reduces for all $\varepsilon>0$ small enough to

  $$f_{1}(y,C+2\gamma\alpha e^{\int_{0}^{T}r_{s}\mathrm{d}s}\varepsilon)= -x_{0}+\left(k_{0}+k_{1}+\displaystyle\frac{C+2\gamma\alpha e^{\int_{0}^{T}r_{s}\mathrm{d}s}\varepsilon}{2\gamma\alpha}\right)e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)$$

  $$-\displaystyle\frac{y}{2\gamma\alpha^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\Phi\left(d_{2}\left(\xi_{3}^{*},0\right)\right)$$ (A.20)

  $$= -x_{0}+(x_{0}+\varepsilon)\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)-\displaystyle\frac{y}{2\gamma\alpha^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\Phi\left(d_{2}\left(\xi_{3}^{*},0\right)\right).$$

  Now, if $\limsup_{\varepsilon\to 0}y_{\lambda_{\varepsilon}}^{*}>0$, it follows that $\liminf_{\varepsilon\to 0}f_{1}(y_{\lambda_{\varepsilon}}^{*},\lambda_{\varepsilon})<0$ (since $\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)<1$ for all $y>0$) which is a contradiction to (A.8). Thus, the claim follows, i.e., $\lim_{\lambda\to C}y_{\lambda}^{*}=0$, when taking the limit of $\varepsilon\to 0$ (and hence $\lambda\to C$) on both sides, since $y_{\lambda}^{*}\geq 0$. Note that for every $y$, the existence of a solution $\lambda_{y}$ of (Proof of (A.13).) was already ensured in the main part of the proof.

  Case 2: $x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\geq k_{2}$, i.e., $C=2\gamma\left(\tilde{\alpha}x_{0}e^{\int_{0}^{T}r_{s}\mathrm{d}s}+\alpha_{2}k_{2}-\alpha(k_{0}+k_{1})\right)$:
  For $\lambda_{\varepsilon}=C+2\gamma\tilde{\alpha}e^{\int_{0}^{T}r_{s}\mathrm{d}s}\varepsilon$, it holds that:

  $$f_{1}(y,\lambda_{\varepsilon})= -x_{0}+(x_{0}+\varepsilon)\Phi\left(d_{1}\left(\xi_{1}^{*},0\right)\right)-\displaystyle\frac{y}{2\gamma\tilde{\alpha}^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\Phi\left(d_{2}\left(\xi_{1}^{*},0\right)\right)$$

  $$+k_{2}e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\left(\Phi\left(d_{1}\left(\xi_{2}^{*},0\right)\right)-\Phi(d_{1}(\tilde{\alpha}\hat{\xi},0))\right)$$

  $$+\left(\displaystyle\frac{\tilde{\alpha}}{\alpha}(x_{0}+\varepsilon)+\displaystyle\frac{\alpha_{2}}{\alpha}k_{2}e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\right)\left(\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)-\Phi(d_{1}(\alpha\hat{\xi},0))\right)$$

  $$-\displaystyle\frac{y}{2\gamma\alpha^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\left(\Phi\left(d_{2}\left(\xi_{3}^{*},0\right)\right)-\Phi(d_{2}(\alpha\hat{\xi},0))\right)$$

  $$\leq -x_{0}+x_{0}\left(\Phi\left(d_{1}\left(\xi_{1}^{*},0\right)\right)+\Phi\left(d_{1}\left(\xi_{2}^{*},0\right)\right)-\Phi(d_{1}(\tilde{\alpha}\hat{\xi},0))+\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)-\Phi(d_{1}(\alpha\hat{\xi},0))\right)$$

  $$+\varepsilon\left(\Phi\left(d_{1}\left(\xi_{1}^{*},0\right)\right)+\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)-\Phi(d_{1}(\alpha\hat{\xi},0))\right)$$

  $$-\displaystyle\frac{y}{2\gamma}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s}\left(\frac{1}{\alpha^{2}}\Phi\left(d_{2}\left(\xi_{3}^{*},0\right)\right)-\frac{1}{\alpha^{2}}\Phi(d_{2}(\alpha\hat{\xi},0))+\frac{1}{\tilde{\alpha}^{2}}\Phi\left(d_{2}\left(\xi_{1}^{*},0\right)\right)\right)$$

  $$\leq -x_{0}+x_{0}+\varepsilon-\displaystyle\frac{y}{2\gamma\tilde{\alpha}^{2}}e^{\int_{0}^{T}-2r_{s}+\left\lVert\kappa_{s}\right\rVert^{2}\mathrm{d}s},$$

  where we used that $k_{2}e^{-\int_{0}^{T}r_{s}\mathrm{d}s}\leq x_{0}$, $\tilde{\alpha}+\alpha_{2}=\alpha$, and $\frac{\tilde{\alpha}}{\alpha}\leq 1$ in the first inequality. Moreover, we used (A.16) and $\Phi\left(d_{1}\left(\xi_{3}^{*},0\right)\right)\leq 1$ for all $y>0$ in the second inequality. Hence, the claim follows, i.e., $\lim_{\varepsilon\to 0}y_{\lambda_{\varepsilon}}^{*}=0$ by the same argument as in Case 1. Note that the existence of a solution was already ensured in the main part of the proof. ∎

• There is nothing to show since we already have that $y_{\lambda}^{*}\in[0,\infty)$. ∎

• We prove this by contradiction. Therefore, we assume that there exists an $L>0$ such that $\limsup_{\lambda\to\infty}y^{*}_{\lambda}\leq L$. Under this assumption, it holds that $\xi_{1}^{*}\xrightarrow{\lambda\to\infty}\infty$ since $\lambda-\sqrt{l(\lambda)}\xrightarrow{\lambda\to\infty}\infty$ for all affine functions $l$ and $y_{\lambda}^{*}$ being bounded. Then (A.16) implies that $\hat{\xi},\bar{\xi},\xi_{2}^{*},\xi_{3}^{*}\xrightarrow{\lambda\to\infty}\infty$ and we know that $\Phi(d_{1}(\xi_{3}^{*},0))\geq\Phi(d_{1}(\alpha\hat{\xi},0))$. However, then $\lim_{\lambda\to\infty}f_{1}(y,\lambda)=\infty$ for all $y\leq L$ which is a contradiction to $y^{*}_{\lambda}$ being the zero root of $f_{1}$. ∎

Since all statements are proved, we have shown the lemma.  ∎

Appendix B Additional lemma

The following lemma restates a well-known result for log-normal distributions, which is used, e.g., in deriving the pricing formula of a put or call in a Black-Scholes model. The proof is just a straightforward calculation. However, we will give it for the sake of completeness.

Lemma B.1.

Let $0\leq a\leq b\leq+\infty$ and $X\sim\mathcal{LN}(\mu,\sigma^{2})$, where $\mathcal{LN}$ denotes a log-normal distribution. Then, it holds that:

$$\mathbb{E}\left[X\mathbbm{1}_{X\in[a,b]}\right]=e^{\mu+\frac{\sigma^{2}}{2}}\left(\Phi\left(\frac{\ln b-\mu-\sigma^{2}}{\sigma}\right)-\Phi\left(\frac{\ln a-\mu-\sigma^{2}}{\sigma}\right)\right),$$ (B.1)

where $\Phi$ denotes the cdf of a standard normal distribution with $\Phi(+\infty)=1$ and $\Phi(-\infty)=0$. The formula remains unchanged when we replace the interval $[a,b]$ by $(a,b]$, $[a,b)$, or $(a,b)$.

Proof. It holds with $y:=\ln x$ and $z:=\frac{y-\mu-\sigma^{2}}{\sigma}$:

$$\mathbb{E}\left[X\mathbbm{1}_{X\in[a,b]}\right] =\int_{\ln a}^{\ln b}\displaystyle\frac{1}{\sqrt{2\pi\sigma^{2}}}\exp\left(y-\frac{(y-\mu)^{2}}{2\sigma^{2}}\right)\mathrm{d}y$$

$$=\exp\left(-\frac{1}{2\sigma^{2}}\left(\mu^{2}-(\mu+\sigma^{2})^{2}\right)\right)\int_{\frac{\ln a-\mu-\sigma^{2}}{\sigma}}^{\frac{\ln b-\mu-\sigma^{2}}{\sigma}}\displaystyle\frac{1}{\sqrt{2\pi}}\exp\left(\frac{1}{2}z^{2}\right)\mathrm{d}z$$

$$=\exp\left(\mu+\frac{\sigma^{2}}{2}\right)\left(\Phi\left(\frac{\ln b-\mu-\sigma^{2}}{\sigma}\right)-\Phi\left(\frac{\ln a-\mu-\sigma^{2}}{\sigma}\right)\right).$$

 ∎