Acknowledgements: The authors thank Paul Invernizzi for insightful discussions. To sharpen an earlier sufficient condition for nonnegative volatility, the authors proposed using the resolvent of the first kind and a semimartingale decomposition. They used OpenAI’s GPT-6 Astra model to help adapt the proof to this approach; the sharp bounds for exponential kernels had already been derived by the authors.
Introduction
Empirical evidence suggests that the volatility of financial markets depends on the past evolution of the asset price [25 , 26 , 10 , 8 , 6 ] . Guyon and Lekeufack [17 ] introduced a path-dependent volatility (PDV) model based on past returns and past squared returns weighted by memory kernels. We study when its volatility1 stays nonnegative. From a modeling perspective, verifying the nonnegativity of volatilities is important because it ensures that the correlations specified between the driving Brownian motions retain their interpretation as instantaneous correlations between asset returns (in multi-asset models) or between asset returns and changes in some extra stochastic processes, such as stochastic interest rates or stochastic variance (in single- or multi-asset models). For a Brownian motion W W W and deterministic continuous functions g 1 : R + → R g_1:\mathbb{R}_+\to\mathbb{R} g 1 : R + → R and g 2 : R + → [ 0 , ∞ ) g_2:\mathbb{R}_+\to[0,\infty) g 2 : R + → [ 0 , ∞ ) , we define the PDV model through the coupled system, for t ≥ 0 t\geq0 t ≥ 0 ,
d S t S t = σ t d W t (PDV) \frac{dS_t}{S_t}=\sigma_t\,dW_t \quad\text{(PDV)} S t d S t = σ t d W t (PDV)
σ t : = β 0 + β 1 R ‾ 1 , t + β 2 R ‾ 2 , t , R ‾ 1 , t : = g 1 ( t ) + ∫ 0 t K 1 ( t − s ) σ s d W s , R ‾ 2 , t : = g 2 ( t ) + ∫ 0 t K 2 ( s , t ) σ s 2 d s , \begin{aligned}
\sigma_t &:= \beta_0 + \beta_1\overline{R}_{1,t} + \beta_2\sqrt{\overline{R}_{2,t}},\\
\overline{R}_{1,t} &:= g_1(t) + \int_0^t K_1(t-s)\sigma_s\,\mathrm{d}W_s,\\
\overline{R}_{2,t} &:= g_2(t) + \int_0^t K_2(s,t)\sigma_s^2\,\mathrm{d}s,
\end{aligned} σ t R 1 , t R 2 , t := β 0 + β 1 R 1 , t + β 2 R 2 , t , := g 1 ( t ) + ∫ 0 t K 1 ( t − s ) σ s d W s , := g 2 ( t ) + ∫ 0 t K 2 ( s , t ) σ s 2 d s ,
where β 0 , β 2 > 0 \beta_0,\beta_2 > 0 β 0 , β 2 > 0 and β 1 < 0 \beta_1 < 0 β 1 < 0 . The nonnegative convolution kernel K 1 : ( 0 , ∞ ) → [ 0 , ∞ ) K_1 : (0,\infty) \to[0,\infty) K 1 : ( 0 , ∞ ) → [ 0 , ∞ ) and the nonnegative kernel K 2 : Γ ‾ → R + K_2 : \overline{\Gamma} \to\mathbb{R}_+ K 2 : Γ → R + weight past returns and past squared returns, respectively, with Γ ‾ : = { ( s , t ) ∈ R + 2 : s ≤ t } \overline{\Gamma} := \{(s,t) \in\mathbb{R}_+^2 : s \le t\} Γ := {( s , t ) ∈ R + 2 : s ≤ t } .
When K 1 K_1 K 1 and K 2 K_2 K 2 are respectively M M M - and N N N -exponential kernels (defined below), with integers M , N ≥ 1 M,N \ge1 M , N ≥ 1 , the PDV model admits a finite-dimensional Markovian representation. The kernels K 1 K_1 K 1 and K 2 K_2 K 2 of the four-factor PDV model are both 2-exponential kernels. Existence and pathwise uniqueness of strong solutions were established for the four-factor model in [23 Theorem 1.2] and for the PDV model with more general Volterra kernels in [5 Theorem 1] . For K 1 ( t ) = λ 1 , 1 e − λ 1 , 1 t K_1(t) = \lambda_{1,1}e^{-\lambda_{1,1}t} K 1 ( t ) = λ 1 , 1 e − λ 1 , 1 t and K 2 ( s , t ) = λ 2 , 1 e − λ 2 , 1 ( t − s ) K_2(s,t) = \lambda_{2,1}e^{-\lambda_{2,1}(t-s)} K 2 ( s , t ) = λ 2 , 1 e − λ 2 , 1 ( t − s ) , Guyon and Lekeufack [17 Section 4.1.5] observed that λ 2 , 1 < 2 λ 1 , 1 \lambda_{2,1} < 2\lambda_{1,1} λ 2 , 1 < 2 λ 1 , 1 ensures nonnegativity when σ 0 ≥ 0 \sigma_0 \ge0 σ 0 ≥ 0 . Nutz and Riveros Valdevenito [23 Theorem 2.6] obtained a strictly positive pathwise lower bound when σ 0 > 0 \sigma_0 > 0 σ 0 > 0 . Andrès and Jourdain [5 Theorem 2] extended the positivity result to possibly nonconvolution return kernels K 1 ( s , t ) = g ( s ) e h ( t ) K_1(s,t) = g(s)e^{h(t)} K 1 ( s , t ) = g ( s ) e h ( t ) and general squared-return kernels K 2 ( s , t ) K_2(s,t) K 2 ( s , t ) satisfying ∂ t K 2 ( s , t ) − 2 h ′ ( t ) K 2 ( s , t ) ≥ 0 \partial_tK_2(s,t) - 2h'(t)K_2(s,t) \ge0 ∂ t K 2 ( s , t ) − 2 h ′ ( t ) K 2 ( s , t ) ≥ 0 .
Nonnegativity and invariance for stochastic Volterra equations are studied in [3 , 4 , 2 ] . Bondi and Pulido [7 ] establish Feller-type tests for scalar convolution equations with nonsingular kernels. Their tests do not directly apply to (PDV) because σ \sigma σ depends on both R ‾ 1 \overline{R}_1 R 1 and R ‾ 2 \overline{R}_2 R 2 , and the initial inputs depend on time. As in [7 Lemma 1] , we derive a semimartingale decomposition for σ \sigma σ and obtain sufficient conditions for positivity by controlling the drift term arising from memory.
Strikingly, for every T > 0 T > 0 T > 0 , when σ ≥ 0 \sigma\ge0 σ ≥ 0 and condition (11) below holds, Theorem 1 below gives a deterministic upper bound on sup t ∈ [ 0 , T ] S t \sup_{t\in[0,T]} S_t sup t ∈ [ 0 , T ] S t , despite the Brownian driving noise. This bound depends only on the kernels, coefficients and initial inputs. Hence S S S is a bounded martingale on [ 0 , T ] [0,T] [ 0 , T ] . Beyond martingality, this bound restricts the support of the price distribution, ensures finiteness of all positive moments, and implies zero call prices for strikes above the bound. We also extend this price-bound principle to several stochastic volatility models in the limiting case ρ = − 1 \rho= -1 ρ = − 1 of perfect negative correlation between the Brownian motions driving the asset price and volatility. In the PDV model, this result and the strict local martingality criterion in [19 Corollary 2.6 (ii)] yield a dichotomy according to the sign of volatility: for smooth return kernels K 1 K_1 K 1 admitting a resolvent of the first kind, S S S is a strict local martingale on [ 0 , T ] [0,T] [ 0 , T ] whenever σ t 0 < 0 \sigma_{t_0} < 0 σ t 0 < 0 has positive probability for some t 0 ∈ [ 0 , T ) t_0 \in[0,T) t 0 ∈ [ 0 , T ) .
For a completely monotone return kernel K 1 K_1 K 1 (defined below), Theorem 2 below gives explicit conditions on the kernels and initial state that ensure nonnegative volatility, together with an explicit process that bounds σ \sigma σ from below. In particular, its specialization to the four-factor PDV model provides sufficient conditions on both the kernel parameters and the initial factors, see the second row of Table 2 below. It thus addresses the question of finding sufficient conditions for nonnegative volatility in that model, left open by Nutz and Riveros Valdevenito [23 Section 3.2] .
Table 2. Conditions from Proposition 3 and Theorem 3 when K 2 K_2 K 2 is an N N N -exponential kernel. The deterministic initial factors and inputs satisfy (5) and R ‾ 2 , 0 = g 2 ( 0 ) > 0 \overline{R}_{2,0}=g_2(0)>0 R 2 , 0 = g 2 ( 0 ) > 0 as in (1) . The sufficient conditions assume σ 0 ≥ 0 \sigma_0\ge0 σ 0 ≥ 0 , and the necessary conditions assume σ 0 > 0 \sigma_0>0 σ 0 > 0 . The sufficient and necessary conditions coincide for M = 1 M=1 M = 1 when σ 0 > 0 \sigma_0>0 σ 0 > 0 . For M ≥ 2 M\ge2 M ≥ 2 , the sufficient and necessary rate conditions involve λ 1 , M \lambda_{1,M} λ 1 , M and λ 1 , e f f \lambda_{1,\mathrm{eff}} λ 1 , eff , respectively; the sufficient conditions on the initial factors involve both λ 2 , e f f \lambda_{2,\mathrm{eff}} λ 2 , eff and λ 2 ( R 2 , 0 ) \lambda_2(\mathbf{R}_{2,0}) λ 2 ( R 2 , 0 ) , while the necessary condition involves λ 2 ( R 2 , 0 ) \lambda_2(\mathbf{R}_{2,0}) λ 2 ( R 2 , 0 ) . For M ≥ 3 M\ge3 M ≥ 3 , the sufficient conditions bound the maximum of zero and all tail sums ∑ i = p M − 1 λ ^ i η 0 ( λ 1 , i ) \sum_{i=p}^{M-1}\widehat{\lambda}_i\eta_0(\lambda_{1,i}) ∑ i = p M − 1 λ i η 0 ( λ 1 , i ) , whereas the necessary condition bounds only the full sum corresponding to p = 1 p=1 p = 1 .
When K 1 K_1 K 1 and K 2 K_2 K 2 are respectively M M M - and N N N -exponential kernels, Theorem 3 below gives sufficient conditions under which volatility is negative with positive probability at every positive time. By contraposition, these yield necessary conditions for nonnegative volatility. For M = 1 M = 1 M = 1 , Corollary 1 below shows that when the model is initialized from past market returns with σ 0 > 0 \sigma_0 > 0 σ 0 > 0 , nonnegative volatility and martingality of S S S on every finite time interval are equivalent to the same explicit condition on the kernel rates. For M ≥ 2 M \ge2 M ≥ 2 , Table 2 below displays the remaining gap between the sufficient and necessary conditions.
Setting and well-posedness
The filtered probability space ( Ω , F , ( F t ) t ≥ 0 , P ) (\Omega,\mathcal{F},(\mathcal{F}_t)_{t\geq0},\mathbb{P}) ( Ω , F , ( F t ) t ≥ 0 , P ) satisfies the usual conditions, and W W W is an ( F t ) (\mathcal{F}_t) ( F t ) -Brownian motion. Throughout the paper, the initial asset price S 0 > 0 S_0>0 S 0 > 0 is deterministic. For the PDV model, the initial values satisfy
R ‾ 1 , 0 = g 1 ( 0 ) , R ‾ 2 , 0 = g 2 ( 0 ) > 0 , σ 0 = β 0 + β 1 g 1 ( 0 ) + β 2 g 2 ( 0 ) . (1) \overline{R}_{1,0}=g_1(0), \qquad\overline{R}_{2,0}=g_2(0)>0, \qquad\sigma_0=\beta_0+\beta_1g_1(0)+\beta_2\sqrt{g_2(0)}. \tag*{(1)} R 1 , 0 = g 1 ( 0 ) , R 2 , 0 = g 2 ( 0 ) > 0 , σ 0 = β 0 + β 1 g 1 ( 0 ) + β 2 g 2 ( 0 ) . (1)
Notation 1. If X X X has almost surely locally absolutely continuous paths, X ˙ \dot{X} X ˙ denotes their ordinary time derivative, defined for almost every time on each such path. In particular, almost surely, X t = X 0 + ∫ 0 t X ˙ s d s X_t=X_0+\int_0^t\dot{X}_s\,\mathrm{d}s X t = X 0 + ∫ 0 t X ˙ s d s for every t ≥ 0 t\geq0 t ≥ 0 . For a locally integrable function f f f and a locally finite measure μ \mu μ on R + \mathbb{R}_+ R + , we define the convolution by ( f ∗ μ ) ( t ) : = ∫ [ 0 , t ] f ( t − s ) μ ( d s ) (f\ast\mu)(t):=\int_{[0,t]}f(t-s)\,\mu(\mathrm{d}s) ( f ∗ μ ) ( t ) := ∫ [ 0 , t ] f ( t − s ) μ ( d s ) . The same notation applies to the convolution of two functions.
Convolution Kernels and initial inputs
Definition 2.1. A convolution kernel K K K on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) with values in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) is completely monotone if K ∈ C ∞ ( ( 0 , ∞ ) ) K\in C^\infty((0,\infty)) K ∈ C ∞ (( 0 , ∞ )) and
∀ n ∈ N ∪ { 0 } , ∀ t > 0 , ( − 1 ) n K ( n ) ( t ) ≥ 0. \forall n\in\mathbb{N}\cup\{0\},\quad\forall t>0,\quad(-1)^nK^{(n)}(t)\geq0. ∀ n ∈ N ∪ { 0 } , ∀ t > 0 , ( − 1 ) n K ( n ) ( t ) ≥ 0. By Bernstein’s theorem, complete monotonicity together with ∫ 0 ∞ K ( t ) d t = 1 \int_0^\infty K(t)\,\mathrm{d}t=1 ∫ 0 ∞ K ( t ) d t = 1 is equivalent to the representation
∀ t > 0 , K ( t ) = ∫ λ ∈ ( 0 , ∞ ) λ e − λ t ν ( d λ ) , \forall t>0,\quad K(t)=\int_{\lambda\in(0,\infty)}\lambda e^{-\lambda t}\,\nu(\mathrm{d}\lambda), ∀ t > 0 , K ( t ) = ∫ λ ∈ ( 0 , ∞ ) λ e − λ t ν ( d λ ) , where ν \nu ν is a unique probability measure on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
Assumption 1. The return kernel K 1 K_1 K 1 is completely monotone and admits the normalized representation
∀ u > 0 , K 1 ( u ) = ∫ λ ∈ ( 0 , ∞ ) λ e − λ u ν ( d λ ) , \forall u>0,\quad K_1(u)=\int_{\lambda\in(0,\infty)}\lambda e^{-\lambda u}\,\nu(\mathrm{d}\lambda), ∀ u > 0 , K 1 ( u ) = ∫ λ ∈ ( 0 , ∞ ) λ e − λ u ν ( d λ ) , where ν \nu ν is a probability measure on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Moreover, the second moment of ν \nu ν is finite:
∫ λ ∈ ( 0 , ∞ ) λ 2 ν ( d λ ) < ∞ . (2) \int_{\lambda\in(0,\infty)}\lambda^2\,\nu(\mathrm{d}\lambda)<\infty. \tag*{(2)} ∫ λ ∈ ( 0 , ∞ ) λ 2 ν ( d λ ) < ∞. (2) The initial input g 1 g_1 g 1 has the representation
∀ t ≥ 0 , g 1 ( t ) = ∫ λ ∈ ( 0 , ∞ ) e − λ t R 1 , 0 ( λ ) ν ( d λ ) , (3) \forall t\geq0,\quad g_1(t)=\int_{\lambda\in(0,\infty)}e^{-\lambda t}R_{1,0}(\lambda)\,\nu(\mathrm{d}\lambda), \tag*{(3)} ∀ t ≥ 0 , g 1 ( t ) = ∫ λ ∈ ( 0 , ∞ ) e − λ t R 1 , 0 ( λ ) ν ( d λ ) , (3) where the deterministic measurable initial factor profile R 1 , 0 : ( 0 , ∞ ) → R R_{1,0}:(0,\infty)\to\mathbb{R} R 1 , 0 : ( 0 , ∞ ) → R satisfies
∫ λ ∈ ( 0 , ∞ ) ( 1 + λ ) ∣ R 1 , 0 ( λ ) ∣ ν ( d λ ) < ∞ . (4) \int_{\lambda\in(0,\infty)}(1+\lambda)\lvert R_{1,0}(\lambda)\rvert\,\nu(\mathrm{d}\lambda)<\infty. \tag*{(4)} ∫ λ ∈ ( 0 , ∞ ) ( 1 + λ ) ∣ R 1 , 0 ( λ )∣ ν ( d λ ) < ∞. (4) The return component R ‾ 1 , t \overline{R}_{1,t} R 1 , t then admits the factor representation
∀ t ≥ 0 , R ‾ 1 , t = ∫ λ ∈ ( 0 , ∞ ) R 1 , t ( λ ) ν ( d λ ) , \forall t\geq0,\quad\overline{R}_{1,t}=\int_{\lambda\in(0,\infty)}R_{1,t}(\lambda)\,\nu(\mathrm{d}\lambda), ∀ t ≥ 0 , R 1 , t = ∫ λ ∈ ( 0 , ∞ ) R 1 , t ( λ ) ν ( d λ ) , ∀ t ≥ 0 , ∀ λ > 0 , d R 1 , t ( λ ) = − λ R 1 , t ( λ ) d t + λ σ t d W t . \forall t\geq0,\quad\forall\lambda>0,\quad\mathrm{d}R_{1,t}(\lambda)=-\lambda R_{1,t}(\lambda)\,\mathrm{d}t+\lambda\sigma_t\,\mathrm{d}W_t. ∀ t ≥ 0 , ∀ λ > 0 , d R 1 , t ( λ ) = − λ R 1 , t ( λ ) d t + λ σ t d W t . Thus, once the kernels and coefficients are fixed, the model is specified by S 0 S_0 S 0 , g 2 g_2 g 2 and the initial factor profile R 1 , 0 ( ⋅ ) R_{1,0}(\cdot) R 1 , 0 ( ⋅ ) .
Notation 2. Under Assumption 1, we set
Λ 1 : = ∫ λ ∈ ( 0 , ∞ ) λ ν ( d λ ) > 0 , λ 1 , e f f : = ∫ λ ∈ ( 0 , ∞ ) λ 2 ν ( d λ ) Λ 1 . \Lambda_1 := \int_{\lambda\in(0,\infty)} \lambda\,\nu(\mathrm{d}\lambda) > 0, \qquad\lambda_{1,\mathrm{eff}} := \frac{\int_{\lambda\in(0,\infty)} \lambda^2 \,\nu(\mathrm{d}\lambda)}{\Lambda_1}. Λ 1 := ∫ λ ∈ ( 0 , ∞ ) λ ν ( d λ ) > 0 , λ 1 , eff := Λ 1 ∫ λ ∈ ( 0 , ∞ ) λ 2 ν ( d λ ) .
Condition (2) extends K 1 K_1 K 1 canonically to a function in C 1 ( R + ) C^1(\mathbb{R}_+) C 1 ( R + ) ; we use the same symbol for this extension. It also gives Λ 1 , λ 1 , e f f < ∞ \Lambda_1,\lambda_{1,\mathrm{eff}} < \infty Λ 1 , λ 1 , eff < ∞ and
K 1 ( 0 ) = Λ 1 , K 1 ′ ( 0 ) = − Λ 1 λ 1 , e f f . K_1(0) = \Lambda_1, \qquad K_1'(0) = -\Lambda_1\lambda_{1,\mathrm{eff}}. K 1 ( 0 ) = Λ 1 , K 1 ′ ( 0 ) = − Λ 1 λ 1 , eff .
Definition 2.3. For a locally integrable convolution kernel K K K taking values in R \mathbb{R} R , a nonnegative locally finite measure R R R is a resolvent of the first kind if
for almost every t > 0 , ∫ [ 0 , t ] K ( t − s ) R ( d s ) = 1. \text{for almost every } t > 0, \qquad\int_{[0,t]} K(t-s) R(\mathrm{d}s) = 1. for almost every t > 0 , ∫ [ 0 , t ] K ( t − s ) R ( d s ) = 1.
For a nonzero completely monotone kernel K ∈ C 1 ( R + ) K \in C^1(\mathbb{R}_+) K ∈ C 1 ( R + ) , there exists a unique resolvent of the first kind. It is the sum of 1 K ( 0 ) δ 0 \frac{1}{K(0)}\delta_0 K ( 0 ) 1 δ 0 and an absolutely continuous measure with a completely monotone density [16 Chapter 5, Theorems 5.2 and 5.4] . Under Assumption 1, the kernel K 1 K_1 K 1 is completely monotone. We denote by R R R the resolvent of the first kind of K 1 K_1 K 1 .
Definition 2.4. An N N N -exponential kernel has the form
∀ u ≥ 0 , K ( u ) = ∑ j = 1 N θ j λ j e − λ j u , θ j > 0 , ∑ j = 1 N θ j = 1 , λ 1 > ⋯ > λ N > 0. \forall u \ge0,\qquad K(u) = \sum_{j=1}^{N} \theta_j\lambda_j e^{-\lambda_j u},\qquad\theta_j > 0,\qquad\sum_{j=1}^{N}\theta_j = 1,\qquad\lambda_1 > \cdots> \lambda_N > 0. ∀ u ≥ 0 , K ( u ) = j = 1 ∑ N θ j λ j e − λ j u , θ j > 0 , j = 1 ∑ N θ j = 1 , λ 1 > ⋯ > λ N > 0. It is therefore completely monotone, with
ν = ∑ j = 1 N θ j δ λ j . \nu= \sum_{j=1}^{N}\theta_j\delta_{\lambda_j}. ν = j = 1 ∑ N θ j δ λ j .
When K 1 K_1 K 1 is an M M M -exponential kernel and K 2 K_2 K 2 is an N N N -exponential kernel, their weights and rates are denoted by ( θ 1 , i , λ 1 , i ) 1 ≤ i ≤ M (\theta_{1,i},\lambda_{1,i})_{1\le i\le M} ( θ 1 , i , λ 1 , i ) 1 ≤ i ≤ M and ( θ 2 , j , λ 2 , j ) 1 ≤ j ≤ N (\theta_{2,j},\lambda_{2,j})_{1\le j\le N} ( θ 2 , j , λ 2 , j ) 1 ≤ j ≤ N . The deterministic initial factors satisfy R 1 , 0 ( λ 1 , i ) ∈ R R_{1,0}(\lambda_{1,i}) \in\mathbb{R} R 1 , 0 ( λ 1 , i ) ∈ R for every i ∈ { 1 , … , M } i \in\{1,\ldots,M\} i ∈ { 1 , … , M } and R 2 , 0 ( λ 2 , j ) ≥ 0 R_{2,0}(\lambda_{2,j}) \ge0 R 2 , 0 ( λ 2 , j ) ≥ 0 for every j ∈ { 1 , … , N } j \in\{1,\ldots,N\} j ∈ { 1 , … , N } . The initial inputs are specified by
∀ t ≥ 0 , g 1 ( t ) = ∑ i = 1 M θ 1 , i e − λ 1 , i t R 1 , 0 ( λ 1 , i ) , g 2 ( t ) = ∑ j = 1 N θ 2 , j e − λ 2 , j t R 2 , 0 ( λ 2 , j ) . (5) \forall t \ge0,\qquad g_1(t) = \sum_{i=1}^{M}\theta_{1,i}e^{-\lambda_{1,i}t}R_{1,0}(\lambda_{1,i}),\qquad g_2(t) = \sum_{j=1}^{N}\theta_{2,j}e^{-\lambda_{2,j}t}R_{2,0}(\lambda_{2,j}). \tag*{(5)} ∀ t ≥ 0 , g 1 ( t ) = i = 1 ∑ M θ 1 , i e − λ 1 , i t R 1 , 0 ( λ 1 , i ) , g 2 ( t ) = j = 1 ∑ N θ 2 , j e − λ 2 , j t R 2 , 0 ( λ 2 , j ) . (5)
In this exponential-kernel setting, K 2 ( s , t ) = K 2 ( 0 , t − s ) K_2(s,t) = K_2(0,t-s) K 2 ( s , t ) = K 2 ( 0 , t − s ) for 0 ≤ s ≤ t 0 \le s \le t 0 ≤ s ≤ t . Since R 2 , 0 ( λ 2 , j ) ≥ 0 R_{2,0}(\lambda_{2,j}) \ge0 R 2 , 0 ( λ 2 , j ) ≥ 0 and g 2 ( 0 ) = R ‾ 2 , 0 > 0 g_2(0) = \overline{R}_{2,0} > 0 g 2 ( 0 ) = R 2 , 0 > 0 by (1) ,
∀ t ≥ 0 , g 2 ( t ) ≥ e − λ 2 , 1 t g 2 ( 0 ) = e − λ 2 , 1 t R ‾ 2 , 0 > 0. (6) \forall t \ge0,\qquad g_2(t) \ge e^{-\lambda_{2,1}t}g_2(0) = e^{-\lambda_{2,1}t}\overline{R}_{2,0} > 0. \tag*{(6)} ∀ t ≥ 0 , g 2 ( t ) ≥ e − λ 2 , 1 t g 2 ( 0 ) = e − λ 2 , 1 t R 2 , 0 > 0. (6)
Indeed, applying [11 Section 4] to the Laplace transform of K K K , K ^ ( z ) = ∑ j = 1 N θ j λ j z + λ j \widehat{K}(z)=\sum_{j=1}^{N}\frac{\theta_j\lambda_j}{z+\lambda_j} K ( z ) = ∑ j = 1 N z + λ j θ j λ j , and using K ∗ R = 1 K\ast R=1 K ∗ R = 1 gives, for z > 0 z>0 z > 0 (so that z K ^ ( z ) > 0 z\widehat{K}(z)>0 z K ( z ) > 0 ),
R ^ ( z ) = 1 z K ^ ( z ) = 1 ∑ j = 1 N θ j λ j + 1 z + ∑ p = 1 N − 1 1 α p ∑ j = 1 N θ j λ j ( λ j − α p ) 2 1 z + α p . \widehat{R}(z)=\frac{1}{z\widehat{K}(z)}=\frac{1}{\sum_{j=1}^{N}\theta_j\lambda_j}+\frac{1}{z}+\sum_{p=1}^{N-1}\frac{1}{\alpha_p\sum_{j=1}^{N}\frac{\theta_j\lambda_j}{(\lambda_j-\alpha_p)^2}}\frac{1}{z+\alpha_p}. R ( z ) = z K ( z ) 1 = ∑ j = 1 N θ j λ j 1 + z 1 + p = 1 ∑ N − 1 α p ∑ j = 1 N ( λ j − α p ) 2 θ j λ j 1 z + α p 1 .
Taking inverse Laplace transforms yields (7) .
Well-posedness
Assumption 2. The initial input g 2 g_2 g 2 is strictly positive and locally absolutely continuous. The squared-return kernel K 2 K_2 K 2 is nonnegative and measurable. For every s ≥ 0 s\geq0 s ≥ 0 , the map t ↦ K 2 ( s , t ) t\mapsto K_2(s,t) t ↦ K 2 ( s , t ) is locally absolutely continuous on [ s , ∞ ) [s,\infty) [ s , ∞ ) . For every T > 0 T>0 T > 0 , the following bounds hold:
sup t ∈ [ 0 , T ] sup s ∈ [ 0 , t ] K 2 ( s , t ) < ∞ , ess sup t ∈ [ 0 , T ] ∫ 0 t ∣ ∂ t K 2 ( s , t ) ∣ d s < ∞ . (8) \sup_{t\in[0,T]}\sup_{s\in[0,t]}K_2(s,t)<\infty,\qquad\operatorname*{ess\,sup}_{t\in[0,T]}\int_0^t\left|\partial_tK_2(s,t)\right|\,ds<\infty. \tag*{(8)} t ∈ [ 0 , T ] sup s ∈ [ 0 , t ] sup K 2 ( s , t ) < ∞ , t ∈ [ 0 , T ] ess sup ∫ 0 t ∣ ∂ t K 2 ( s , t ) ∣ d s < ∞. (8)
Proposition 1. Under Assumptions 1 1 1 and 2 2 2 , g 1 g_1 g 1 belongs to C 1 ( R + ) C^1(\mathbb{R}_+) C 1 ( R + ) . The model (PDV) has a pathwise unique continuous ( F t ) (\mathcal{F}_t) ( F t ) -adapted solution on R + \mathbb{R}_+ R + . Moreover, P \mathbb{P} P -almost surely, R ‾ 2 \overline{R}_2 R 2 is locally absolutely continuous and satisfies, for almost every t ≥ 0 t\geq0 t ≥ 0 ,
R ‾ ˙ 2 , t = g ˙ 2 ( t ) + K 2 ( t , t ) σ t 2 + ∫ 0 t ∂ t K 2 ( s , t ) σ s 2 d s , (9) \dot{\overline{R}}_{2,t}=\dot{g}_2(t)+K_2(t,t)\sigma_t^2+\int_0^t\partial_tK_2(s,t)\sigma_s^2\,ds, \tag*{(9)} R ˙ 2 , t = g ˙ 2 ( t ) + K 2 ( t , t ) σ t 2 + ∫ 0 t ∂ t K 2 ( s , t ) σ s 2 d s , (9)
and
∀ t ≥ 0 , R ‾ 2 , t ≥ g 2 ( t ) > 0. \forall t\geq0,\qquad\overline{R}_{2,t}\geq g_2(t)>0. ∀ t ≥ 0 , R 2 , t ≥ g 2 ( t ) > 0.
Proof. Differentiation under the integral in (3) , justified by (4) , gives g 1 ∈ C 1 ( R + ) g_1\in C^1(\mathbb{R}_+) g 1 ∈ C 1 ( R + ) . The kernels satisfy the estimates used in the proof of [5 Theorem 1] for 0 ≤ s ≤ t 0\leq s\leq t 0 ≤ s ≤ t . In that article, the initial inputs are defined differently, by integrals over the prescribed history, including a Brownian integral over negative times for g 1 g_1 g 1 . However, the same existence and uniqueness argument applies to deterministic continuous inputs bounded on compact intervals, with g 2 g_2 g 2 bounded away from zero on each compact interval. For every T > 0 T>0 T > 0 , continuity and strict positivity of g 2 g_2 g 2 in Assumption 2 give inf t ∈ [ 0 , T ] g 2 ( t ) > 0 \inf_{t\in[0,T]}g_2(t)>0 inf t ∈ [ 0 , T ] g 2 ( t ) > 0 . Fubini’s theorem gives (9) . In (PDV), K 2 ≥ 0 K_2\geq0 K 2 ≥ 0 and g 2 > 0 g_2>0 g 2 > 0 by Assumption 2, hence R ‾ 2 , t ≥ g 2 ( t ) > 0 \overline{R}_{2,t}\geq g_2(t)>0 R 2 , t ≥ g 2 ( t ) > 0 .
Remark 2.6. The integral defining R ‾ 2 \overline{R}_2 R 2 may exist even when K 2 K_2 K 2 is singular on the diagonal. Our proof uses K 2 ( t , t ) < ∞ K_2(t,t)<\infty K 2 ( t , t ) < ∞ because (9) contains the term K 2 ( t , t ) σ t 2 K_2(t,t)\sigma_t^2 K 2 ( t , t ) σ t 2 .
Positivity
Positivity and martingality
The asset-price bound holds under weaker hypotheses than Assumptions 1 1 1 and 2 2 2 : complete monotonicity of K 1 K_1 K 1 and differentiability of the kernels and initial inputs are not required. We first derive the integrated resolvent identity used in its proof.
Lemma 3.1. Suppose that K 1 ∈ L l o c 2 ( R + ) K_1\in L^2_{\mathrm{loc}}(\mathbb{R}_+) K 1 ∈ L loc 2 ( R + ) has a resolvent of the first kind R R R . For T > 0 T>0 T > 0 , suppose that (PDV) has a continuous adapted solution on [ 0 , T ] [0,T] [ 0 , T ] , with g 1 ∈ C ( [ 0 , T ] ) g_1\in C([0,T]) g 1 ∈ C ([ 0 , T ]) and R ‾ 2 ≥ 0 \overline{R}_2\geq0 R 2 ≥ 0 . Almost surely, the following identity holds:
∀ t ∈ [ 0 , T ] , ( R ∗ ( σ − β 0 − β 1 g 1 − β 2 R ‾ 2 ) ) ( t ) = β 1 ∫ 0 t σ s d W s . (10) \forall t\in[0,T],\qquad\left(R\ast\left(\sigma-\beta_0-\beta_1g_1-\beta_2\sqrt{\overline{R}_2}\right)\right)(t)=\beta_1\int_0^t\sigma_s\,dW_s. \tag*{(10)} ∀ t ∈ [ 0 , T ] , ( R ∗ ( σ − β 0 − β 1 g 1 − β 2 R 2 ) ) ( t ) = β 1 ∫ 0 t σ s d W s . (10)
Proof. Identity (2.14) in [1 Lemma 2.6] applies with
X t : = R ‾ 1 , t − g 1 ( t ) , X 0 = 0 , M t : = ∫ 0 t σ s d W s , K = K 1 , L = R . X_t:=\overline{R}_{1,t}-g_1(t),\qquad X_0=0,\qquad M_t:=\int_0^t \sigma_s\,\mathrm{d}W_s,\qquad K=K_1,\qquad L=\mathcal{R}. X t := R 1 , t − g 1 ( t ) , X 0 = 0 , M t := ∫ 0 t σ s d W s , K = K 1 , L = R .
Indeed, (PDV) gives X = K 1 ∗ d M X=K_1\ast\mathrm{d}M X = K 1 ∗ d M , and both X X X and σ \sigma σ are continuous and adapted. Hence, almost surely, we have
∀ t ∈ [ 0 , T ] , ( R ∗ ( R ‾ 1 − g 1 ) ) ( t ) = ∫ 0 t σ s d W s . \forall t\in[0,T],\qquad\left(\mathcal{R}\ast(\overline{R}_1-g_1)\right)(t)=\int_0^t \sigma_s\,\mathrm{d}W_s. ∀ t ∈ [ 0 , T ] , ( R ∗ ( R 1 − g 1 ) ) ( t ) = ∫ 0 t σ s d W s .
Multiplying by β 1 \beta_1 β 1 and using σ − β 0 − β 1 g 1 − β 2 R ‾ 2 = β 1 ( R ‾ 1 − g 1 ) \sigma-\beta_0-\beta_1g_1-\beta_2\sqrt{\overline{R}_2}=\beta_1(\overline{R}_1-g_1) σ − β 0 − β 1 g 1 − β 2 R 2 = β 1 ( R 1 − g 1 ) proves (10) . □
Theorem 1. Let T > 0 T>0 T > 0 . Assume that g 1 , g 2 ∈ C ( [ 0 , T ] ) g_1,g_2\in C([0,T]) g 1 , g 2 ∈ C ([ 0 , T ]) , with g 2 ≥ 0 g_2\geq0 g 2 ≥ 0 , and that (PDV) has a unique continuous ( F t ) (\mathcal{F}_t) ( F t ) -adapted solution on [ 0 , T ] [0,T] [ 0 , T ] under P \mathbb{P} P . Assume that σ t ≥ 0 \sigma_t\geq0 σ t ≥ 0 for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , P \mathbb{P} P -almost surely. Suppose that K 2 K_2 K 2 is nonnegative and measurable and that K 1 K_1 K 1 is a nonnegative convolution kernel in L l o c 2 ( R + ) L^2_{\mathrm{loc}}(\mathbb{R}_+) L loc 2 ( R + ) , with a resolvent of the first kind R \mathcal{R} R , and that
C 2 , T : = sup r ∈ [ 0 , T ] sup u ∈ [ 0 , r ] ∫ [ 0 , r − u ] K 2 ( u , r − s ) R ( d s ) < ∞ . (11) C_{2,T}:=\sup_{r\in[0,T]}\sup_{u\in[0,r]}\int_{[0,r-u]}K_2(u,r-s)\mathcal{R}(\mathrm{d}s)<\infty. \tag*{(11)} C 2 , T := r ∈ [ 0 , T ] sup u ∈ [ 0 , r ] sup ∫ [ 0 , r − u ] K 2 ( u , r − s ) R ( d s ) < ∞. (11) Then, P \mathbb{P} P -almost surely, the following bound holds:
sup t ∈ [ 0 , T ] S t ≤ S 0 exp ( R ( [ 0 , T ] ) ( β 0 − β 1 + sup t ∈ [ 0 , T ] ( − g 1 ( t ) ) + + β 2 − β 1 sup t ∈ [ 0 , T ] g 2 ( t ) ) + β 2 2 2 β 1 2 R ( [ 0 , T ] ) C 2 , T ) . (12) \sup_{t\in[0,T]}S_t\leq S_0\exp\left(\mathcal{R}([0,T])\left(\frac{\beta_0}{-\beta_1}+\sup_{t\in[0,T]}(-g_1(t))_++\frac{\beta_2}{-\beta_1}\sqrt{\sup_{t\in[0,T]}g_2(t)}\right)+\frac{\beta_2^2}{2\beta_1^2}\mathcal{R}([0,T])C_{2,T}\right). \tag*{(12)} t ∈ [ 0 , T ] sup S t ≤ S 0 exp ( R ([ 0 , T ]) ( − β 1 β 0 + t ∈ [ 0 , T ] sup ( − g 1 ( t ) ) + + − β 1 β 2 t ∈ [ 0 , T ] sup g 2 ( t ) ) + 2 β 1 2 β 2 2 R ([ 0 , T ]) C 2 , T ) . (12) Hence, under P \mathbb{P} P , S S S is a bounded martingale on [ 0 , T ] [0,T] [ 0 , T ] .
Remark 3.2. Under Assumptions [1 ] and [2 ] , the identity (10) explains why nonnegative volatility limits how large ∫ 0 t σ s d W s \int_0^t\sigma_s\,\mathrm{d}W_s ∫ 0 t σ s d W s can be. Indeed, since β 1 < 0 \beta_1<0 β 1 < 0 , an increase in this integral pushes σ \sigma σ downward. The Brownian integral is therefore controlled and, since it drives the logarithm of the asset price, this yields an upper bound on the price.
Remark 3.3. Under the assumptions of Theorem 1, for fixed deterministic initial inputs, let B T B_T B T denote the right-hand side of (12) . Then, for every p > 0 p>0 p > 0 ,
E [ sup t ∈ [ 0 , T ] S t p ] ≤ B T p < ∞ . \mathbb{E}\left[\sup_{t\in[0,T]}S_t^p\right]\leq B_T^p<\infty. E [ t ∈ [ 0 , T ] sup S t p ] ≤ B T p < ∞.
European call options with maturity T T T and strike k ≥ B T k\geq B_T k ≥ B T have zero payoff almost surely, and hence zero price: E [ ( S T − k ) + ] = 0 \mathbb{E}[(S_T-k)_+]=0 E [( S T − k ) + ] = 0 . The Black–Scholes implied volatility smile is therefore identically zero for strikes k ≥ B T k\geq B_T k ≥ B T . Theorem 1 provides a theoretical basis for the vanishing right wing of the implied volatility smile observed numerically by Guyon and Lekeufack [17 ] and by Gazzani and Guyon [15 ] . This vanishing behaviour motivated the quadratic correction introduced by Guyon and Lekeufack [17 Section 4.5] to reproduce the rising right wing of the SPX smile. Alternatively, Guyon and Parent [18 ] have shown that a discrete-time version of the model with fat-tailed innovations captures market smiles very well, even for large strikes.
Remark 3.4. If K 1 K_1 K 1 is continuous at zero and K 1 ( 0 ) > 0 K_1(0)>0 K 1 ( 0 ) > 0 , then R ( { 0 } ) = 1 K 1 ( 0 ) \mathcal{R}(\{0\})=\frac{1}{K_1(0)} R ({ 0 }) = K 1 ( 0 ) 1 by K 1 ∗ R = 1 K_1\ast\mathcal{R}=1 K 1 ∗ R = 1 . Definition (11) gives
1 K 1 ( 0 ) sup r ∈ [ 0 , T ] sup u ∈ [ 0 , r ] K 2 ( u , r ) ≤ C 2 , T ≤ R ( [ 0 , T ] ) sup r ∈ [ 0 , T ] sup u ∈ [ 0 , r ] K 2 ( u , r ) . \frac{1}{K_1(0)}\sup_{r\in[0,T]}\sup_{u\in[0,r]}K_2(u,r)\leq C_{2,T}\leq\mathcal{R}([0,T])\sup_{r\in[0,T]}\sup_{u\in[0,r]}K_2(u,r). K 1 ( 0 ) 1 r ∈ [ 0 , T ] sup u ∈ [ 0 , r ] sup K 2 ( u , r ) ≤ C 2 , T ≤ R ([ 0 , T ]) r ∈ [ 0 , T ] sup u ∈ [ 0 , r ] sup K 2 ( u , r ) .
Then condition (11) holds when K 2 K_2 K 2 is bounded on { ( u , r ) : 0 ≤ u ≤ r ≤ T } \{(u,r):0\leq u\leq r\leq T\} {( u , r ) : 0 ≤ u ≤ r ≤ T } , for example when K 2 ( s , t ) = K 2 ( 0 , t − s ) K_2(s,t)=K_2(0,t-s) K 2 ( s , t ) = K 2 ( 0 , t − s ) and u ↦ K 2 ( 0 , u ) u\mapsto K_2(0,u) u ↦ K 2 ( 0 , u ) belongs to C ( R + ) C(\mathbb{R}_+) C ( R + ) .
Remark 3.5. The Hölder regularity conditions
( ∫ 0 T − h ∣ K 1 ( u + h ) − K 1 ( u ) ∣ 2 d u + ∫ 0 h K 1 ( u ) 2 d u ) 1 2 ≤ C T h δ 1 , (13) \left(\int_{0}^{T-h}\lvert K_1(u+h)-K_1(u)\rvert^2\,\mathrm{d}u+\int_{0}^{h}K_1(u)^2\,\mathrm{d}u\right)^{\frac{1}{2}}\le C_T h^{\delta_1}, \tag*{(13)} ( ∫ 0 T − h ∣ K 1 ( u + h ) − K 1 ( u ) ∣ 2 d u + ∫ 0 h K 1 ( u ) 2 d u ) 2 1 ≤ C T h δ 1 , (13)
sup t ∈ [ 0 , T − h ] ( ∫ 0 t ∣ K 2 ( s , t + h ) − K 2 ( s , t ) ∣ d s + ∫ t t + h K 2 ( s , t + h ) d s ) ≤ C T h δ 2 , (14) \sup_{t\in[0,T-h]}\left(\int_{0}^{t}\lvert K_2(s,t+h)-K_2(s,t)\rvert\,\mathrm{d}s+\int_{t}^{t+h}K_2(s,t+h)\,\mathrm{d}s\right)\le C_T h^{\delta_2}, \tag*{(14)} t ∈ [ 0 , T − h ] sup ( ∫ 0 t ∣ K 2 ( s , t + h ) − K 2 ( s , t )∣ d s + ∫ t t + h K 2 ( s , t + h ) d s ) ≤ C T h δ 2 , (14)
for every T > 0 T>0 T > 0 and h ∈ [ 0 , T ] h\in[0,T] h ∈ [ 0 , T ] , with a constant C T > 0 C_T>0 C T > 0 , follow from K 1 ∈ C 1 ( R + ) K_1\in C^1(\mathbb{R}_+) K 1 ∈ C 1 ( R + ) in Notation 2 and (8) , with δ 1 = 1 2 \delta_1=\frac{1}{2} δ 1 = 2 1 and δ 2 = 1 \delta_2=1 δ 2 = 1 . Under the increment conditions (13) and (14) , [19 Theorem 1 (B.i)] shows that nonnegative volatility implies martingality. (For simplicity, the result is stated there for a convolution kernel K 2 K_2 K 2 , but the proofs remain unchanged.) Theorem 1 of the present paper gives moreover a deterministic pathwise bound on S S S over [ 0 , T ] [0,T] [ 0 , T ] , under an additional condition on the return kernel K 1 K_1 K 1 : it needs to admit a resolvent of the first kind. Assumption 1 ensures this, while Assumption 2 implies (11) .
Proof of Theorem 1 . By Lemma 3.1 and σ − β 0 − β 1 g 1 − β 2 R ‾ 2 = β 1 ( R ‾ 1 − g 1 ) \sigma-\beta_0-\beta_1g_1-\beta_2\sqrt{\overline{R}_2}=\beta_1(\overline{R}_1-g_1) σ − β 0 − β 1 g 1 − β 2 R 2 = β 1 ( R 1 − g 1 ) , almost surely, we have
∀ t ∈ [ 0 , T ] , ∫ [ 0 , t ] ( R ‾ 1 , t − s − g 1 ( t − s ) ) R ( d s ) = ∫ 0 t σ s d W s . \forall t\in[0,T],\qquad\int_{[0,t]}(\overline{R}_{1,t-s}-g_1(t-s))\,\mathcal{R}(\mathrm{d}s)=\int_{0}^{t}\sigma_s\,\mathrm{d}W_s. ∀ t ∈ [ 0 , T ] , ∫ [ 0 , t ] ( R 1 , t − s − g 1 ( t − s )) R ( d s ) = ∫ 0 t σ s d W s .
Moreover, σ ≥ 0 \sigma\ge0 σ ≥ 0 and β 1 < 0 \beta_1<0 β 1 < 0 imply
∀ v ∈ [ 0 , T ] , R ‾ 1 , v − g 1 ( v ) = β 0 + β 2 R ‾ 2 , v − σ v − β 1 − g 1 ( v ) ≤ β 0 − β 1 + β 2 − β 1 R ‾ 2 , v + sup u ∈ [ 0 , T ] ( − g 1 ( u ) ) + . \forall v\in[0,T],\qquad\overline{R}_{1,v}-g_1(v)=\frac{\beta_0+\beta_2\sqrt{\overline{R}_{2,v}}-\sigma_v}{-\beta_1}-g_1(v)\le\frac{\beta_0}{-\beta_1}+\frac{\beta_2}{-\beta_1}\sqrt{\overline{R}_{2,v}}+\sup_{u\in[0,T]}(-g_1(u))_+. ∀ v ∈ [ 0 , T ] , R 1 , v − g 1 ( v ) = − β 1 β 0 + β 2 R 2 , v − σ v − g 1 ( v ) ≤ − β 1 β 0 + − β 1 β 2 R 2 , v + u ∈ [ 0 , T ] sup ( − g 1 ( u ) ) + .
Since S 0 > 0 S_0>0 S 0 > 0 , the stochastic exponential solution of (PDV) satisfies S t > 0 S_t>0 S t > 0 , so log ( S t S 0 ) \log\left(\frac{S_t}{S_0}\right) log ( S 0 S t ) is well defined. Itô’s formula, the nonnegativity of R \mathcal{R} R , and R ( [ 0 , t ] ) ≤ R ( [ 0 , T ] ) \mathcal{R}([0,t])\le\mathcal{R}([0,T]) R ([ 0 , t ]) ≤ R ([ 0 , T ]) therefore give
log S t S 0 = ∫ 0 t σ u d W u − 1 2 ∫ 0 t σ u 2 d u = ∫ [ 0 , t ] ( R ‾ 1 , t − s − g 1 ( t − s ) ) R ( d s ) − 1 2 ∫ 0 t σ u 2 d u (15) \log\frac{S_t}{S_0}=\int_{0}^{t}\sigma_u\,\mathrm{d}W_u-\frac{1}{2}\int_{0}^{t}\sigma_u^2\,\mathrm{d}u=\int_{[0,t]}(\overline{R}_{1,t-s}-g_1(t-s))\,\mathcal{R}(\mathrm{d}s)-\frac{1}{2}\int_{0}^{t}\sigma_u^2\,\mathrm{d}u \tag*{(15)} log S 0 S t = ∫ 0 t σ u d W u − 2 1 ∫ 0 t σ u 2 d u = ∫ [ 0 , t ] ( R 1 , t − s − g 1 ( t − s )) R ( d s ) − 2 1 ∫ 0 t σ u 2 d u (15)
≤ R ( [ 0 , T ] ) ( β 0 − β 1 + sup u ∈ [ 0 , T ] ( − g 1 ( u ) ) + ) + β 2 − β 1 ∫ [ 0 , t ] R ‾ 2 , t − s R ( d s ) − 1 2 ∫ 0 t σ u 2 d u . (16) \le\mathcal{R}([0,T])\left(\frac{\beta_0}{-\beta_1}+\sup_{u\in[0,T]}(-g_1(u))_+\right)+\frac{\beta_2}{-\beta_1}\int_{[0,t]}\sqrt{\overline{R}_{2,t-s}}\,\mathcal{R}(\mathrm{d}s)-\frac{1}{2}\int_{0}^{t}\sigma_u^2\,\mathrm{d}u. \tag*{(16)} ≤ R ([ 0 , T ]) ( − β 1 β 0 + u ∈ [ 0 , T ] sup ( − g 1 ( u ) ) + ) + − β 1 β 2 ∫ [ 0 , t ] R 2 , t − s R ( d s ) − 2 1 ∫ 0 t σ u 2 d u . (16)
By definition of (PDV),
R ‾ 2 , t = g 2 ( t ) + ∫ 0 t K 2 ( s , t ) σ s 2 d s . \overline{R}_{2,t}=g_2(t)+\int_{0}^{t}K_2(s,t)\sigma_s^2\,\mathrm{d}s. R 2 , t = g 2 ( t ) + ∫ 0 t K 2 ( s , t ) σ s 2 d s .
The Cauchy–Schwarz inequality, Tonelli’s theorem, and the definition (11) of C 2 , T C_{2,T} C 2 , T give
( ∫ [ 0 , t ] R ‾ 2 , t − s R ( d s ) ) 2 ≤ R ( [ 0 , T ] ) ∫ [ 0 , t ] R ‾ 2 , t − s R ( d s ) ≤ R ( [ 0 , T ] ) 2 sup u ∈ [ 0 , T ] g 2 ( u ) \left(\int_{[0,t]}\sqrt{\overline{R}_{2,t-s}}\,\mathcal{R}(\mathrm{d}s)\right)^2\le\mathcal{R}([0,T])\int_{[0,t]}\overline{R}_{2,t-s}\,\mathcal{R}(\mathrm{d}s)\le\mathcal{R}([0,T])^2\sup_{u\in[0,T]}g_2(u) ( ∫ [ 0 , t ] R 2 , t − s R ( d s ) ) 2 ≤ R ([ 0 , T ]) ∫ [ 0 , t ] R 2 , t − s R ( d s ) ≤ R ([ 0 , T ] ) 2 u ∈ [ 0 , T ] sup g 2 ( u )
+ R ( [ 0 , T ] ) ∫ u ∈ [ 0 , t ] ∫ s ∈ [ 0 , t − u ] K 2 ( u , t − s ) R ( d s ) σ u 2 d u ≤ R ( [ 0 , T ] ) 2 sup u ∈ [ 0 , T ] g 2 ( u ) + R ( [ 0 , T ] ) C 2 , T ∫ 0 t σ u 2 d u . (17) \begin{aligned}
\qquad+\mathcal{R}([0,T])\int_{u\in[0,t]}\int_{s\in[0,t-u]}K_2(u,t-s)\mathcal{R}(\mathrm{d}s)\sigma_u^2\,\mathrm{d}u \\
&\le\mathcal{R}([0,T])^2\sup_{u\in[0,T]}g_2(u)+\mathcal{R}([0,T])C_{2,T}\int_{0}^{t}\sigma_u^2\,\mathrm{d}u. \tag*{(17)}
\end{aligned} + R ([ 0 , T ]) ∫ u ∈ [ 0 , t ] ∫ s ∈ [ 0 , t − u ] K 2 ( u , t − s ) R ( d s ) σ u 2 d u ≤ R ([ 0 , T ] ) 2 u ∈ [ 0 , T ] sup g 2 ( u ) + R ([ 0 , T ]) C 2 , T ∫ 0 t σ u 2 d u . (17)
Combining (17) with (16) and the inequalities x + y ≤ x + y \sqrt{x+y}\le\sqrt{x}+\sqrt{y} x + y ≤ x + y and a x − x 2 2 ≤ a 2 2 ax-\frac{x^2}{2}\le\frac{a^2}{2} a x − 2 x 2 ≤ 2 a 2 gives
∀ t ∈ [ 0 , T ] , log S t S 0 ≤ R ( [ 0 , T ] ) ( β 0 − β 1 + sup u ∈ [ 0 , T ] ( − g 1 ( u ) ) + + β 2 − β 1 sup u ∈ [ 0 , T ] g 2 ( u ) ) + β 2 − β 1 R ( [ 0 , T ] ) C 2 , T ∫ 0 t σ u 2 d u − 1 2 ∫ 0 t σ u 2 d u ≤ R ( [ 0 , T ] ) ( β 0 − β 1 + sup u ∈ [ 0 , T ] ( − g 1 ( u ) ) + + β 2 − β 1 sup u ∈ [ 0 , T ] g 2 ( u ) ) + β 2 2 2 β 1 2 R ( [ 0 , T ] ) C 2 , T . \begin{aligned}
\forall t\in[0,T],\qquad\log\frac{S_t}{S_0}\le\mathcal{R}([0,T])\left(\frac{\beta_0}{-\beta_1}+\sup_{u\in[0,T]}(-g_1(u))_++\frac{\beta_2}{-\beta_1}\sqrt{\sup_{u\in[0,T]}g_2(u)}\right) \\
&+\frac{\beta_2}{-\beta_1}\sqrt{\mathcal{R}([0,T])C_{2,T}}\sqrt{\int_0^t \sigma_u^2\,\mathrm{d}u}-\frac{1}{2}\int_0^t \sigma_u^2\,\mathrm{d}u \\
&\le\mathcal{R}([0,T])\left(\frac{\beta_0}{-\beta_1}+\sup_{u\in[0,T]}(-g_1(u))_++\frac{\beta_2}{-\beta_1}\sqrt{\sup_{u\in[0,T]}g_2(u)}\right) \\
&+\frac{\beta_2^2}{2\beta_1^2}\mathcal{R}([0,T])C_{2,T}.
\end{aligned} ∀ t ∈ [ 0 , T ] , log S 0 S t ≤ R ([ 0 , T ]) ( − β 1 β 0 + u ∈ [ 0 , T ] sup ( − g 1 ( u ) ) + + − β 1 β 2 u ∈ [ 0 , T ] sup g 2 ( u ) ) + − β 1 β 2 R ([ 0 , T ]) C 2 , T ∫ 0 t σ u 2 d u − 2 1 ∫ 0 t σ u 2 d u ≤ R ([ 0 , T ]) ( − β 1 β 0 + u ∈ [ 0 , T ] sup ( − g 1 ( u ) ) + + − β 1 β 2 u ∈ [ 0 , T ] sup g 2 ( u ) ) + 2 β 1 2 β 2 2 R ([ 0 , T ]) C 2 , T .
Exponentiation gives (12) . The upper bound is deterministic. Thus S S S is a bounded local martingale, hence a martingale.
Extensions to stochastic volatility models
The following proposition is a direct adaptation of the proof of Theorem 1. It applies to several stochastic volatility models at perfect negative correlation between their Brownian drivers, ρ = − 1 \rho=-1 ρ = − 1 , as illustrated in Table 1 .
Table 1. Applications of Proposition 2 with deterministic initial values and κ , θ , ν > 0 \kappa,\theta,\nu>0 κ , θ , ν > 0 .
Proposition 2. Let T > 0 T>0 T > 0 , S 0 , c > 0 S_0,c>0 S 0 , c > 0 , and let σ \sigma σ and G G G be continuous adapted processes on [ 0 , T ] [0,T] [ 0 , T ] satisfying
d S t S t = σ t d W t , \frac{\mathrm{d}S_t}{S_t}=\sigma_t\,\mathrm{d}W_t, S t d S t = σ t d W t , ∀ t ∈ [ 0 , T ] , Φ ( σ t ) = G t − c ∫ 0 t K ( t − s ) σ s d W s , (18) \forall t\in[0,T],\qquad\Phi(\sigma_t)=G_t-c\int_0^t K(t-s)\sigma_s\,\mathrm{d}W_s, \tag*{(18)} ∀ t ∈ [ 0 , T ] , Φ ( σ t ) = G t − c ∫ 0 t K ( t − s ) σ s d W s , (18) where Φ : R → R \Phi:\mathbb{R}\to\mathbb{R} Φ : R → R is continuous and Φ ( σ t ) ≥ 0 \Phi(\sigma_t)\ge0 Φ ( σ t ) ≥ 0 almost surely for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Suppose that K ∈ L l o c 1 ( R + ) K\in L^1_{\mathrm{loc}}(\mathbb{R}_+) K ∈ L loc 1 ( R + ) admits a nonnegative resolvent of the first kind R \mathcal{R} R . If deterministic constants a T ∈ R a_T\in\mathbb{R} a T ∈ R , b T ≥ 0 b_T\ge0 b T ≥ 0 and d T < c 2 d_T<\frac{c}{2} d T < 2 c satisfy, almost surely,
∀ t ∈ [ 0 , T ] , ( R ∗ G ) ( t ) ≤ a T + b T ∫ 0 t σ s 2 d s + d T ∫ 0 t σ s 2 d s , (19) \forall t\in[0,T],\qquad(\mathcal{R}*G)(t)\le a_T+b_T\sqrt{\int_0^t \sigma_s^2\,\mathrm{d}s}+d_T\int_0^t \sigma_s^2\,\mathrm{d}s, \tag*{(19)} ∀ t ∈ [ 0 , T ] , ( R ∗ G ) ( t ) ≤ a T + b T ∫ 0 t σ s 2 d s + d T ∫ 0 t σ s 2 d s , (19) then, almost surely,
sup t ∈ [ 0 , T ] S t ≤ S 0 exp ( a T c + b T 2 2 c ( c − 2 d T ) ) . (20) \sup_{t\in[0,T]}S_t\le S_0\exp\left(\frac{a_T}{c}+\frac{b_T^2}{2c(c-2d_T)}\right). \tag*{(20)} t ∈ [ 0 , T ] sup S t ≤ S 0 exp ( c a T + 2 c ( c − 2 d T ) b T 2 ) . (20) In particular, S S S is a bounded martingale on [ 0 , T ] [0,T] [ 0 , T ] .
Proof. Convolution with R \mathcal{R} R , justified as in Lemma 3.1, gives
log S t S 0 = ( R ∗ G ) ( t ) − ( R ∗ Φ ( σ ) ) ( t ) c − 1 2 ∫ 0 t σ s 2 d s \log\frac{S_t}{S_0}=\frac{(\mathcal{R}*G)(t)-(\mathcal{R}*\Phi(\sigma))(t)}{c}-\frac{1}{2}\int_0^t\sigma_s^2\,\mathrm{d}s log S 0 S t = c ( R ∗ G ) ( t ) − ( R ∗ Φ ( σ )) ( t ) − 2 1 ∫ 0 t σ s 2 d s ≤ a T c + b T c ∫ 0 t σ s 2 d s − c − 2 d T 2 c ∫ 0 t σ s 2 d s \le\frac{a_T}{c}+\frac{b_T}{c}\sqrt{\int_0^t\sigma_s^2\,\mathrm{d}s}-\frac{c-2d_T}{2c}\int_0^t\sigma_s^2\,\mathrm{d}s ≤ c a T + c b T ∫ 0 t σ s 2 d s − 2 c c − 2 d T ∫ 0 t σ s 2 d s ≤ a T c + b T 2 2 c ( c − 2 d T ) . \le\frac{a_T}{c}+\frac{b_T^2}{2c(c-2d_T)}. ≤ c a T + 2 c ( c − 2 d T ) b T 2 . The deterministic bound makes the local martingale S S S a martingale.
Example 3.6. Under the hypotheses of Theorem 1 , (PDV) has the form (18) with Φ ( x ) = x \Phi(x)=x Φ ( x ) = x , c = − β 1 c=-\beta_1 c = − β 1 , K = K 1 K=K_1 K = K 1 and G = β 0 + β 1 g 1 + β 2 R ‾ 2 G=\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{R}_2} G = β 0 + β 1 g 1 + β 2 R 2 . Estimate (17) gives (19) with d T = 0 d_T=0 d T = 0 , recovering (12) .
For the Heston and rough Heston models, Φ ( x ) = x 2 \Phi(x)=x^2 Φ ( x ) = x 2 , c = ν c=\nu c = ν and
G t = σ 0 2 + ∫ 0 t K ( t − s ) κ ( θ − σ s 2 ) d s , ( a T , b T , d T ) = ( σ 0 2 R ( [ 0 , T ] ) + κ θ T , 0 , − κ ) . G_t=\sigma_0^2+\int_0^t K(t-s)\kappa(\theta-\sigma_s^2)\,\mathrm{d}s,\qquad(a_T,b_T,d_T)=\left(\sigma_0^2\mathcal{R}([0,T])+\kappa\theta T,0,-\kappa\right). G t = σ 0 2 + ∫ 0 t K ( t − s ) κ ( θ − σ s 2 ) d s , ( a T , b T , d T ) = ( σ 0 2 R ([ 0 , T ]) + κ θ T , 0 , − κ ) .
For the Heston model, we have K = 1 K = 1 K = 1 and R = δ 0 \mathcal{R} = \delta_0 R = δ 0 , recovering the bound in [22 Proposition 4.2.8] . For the rough Heston model, we have
K ( t ) = t α − 1 Γ ( α ) , α ∈ ( 1 2 , 1 ) , R ( [ 0 , T ] ) = T 1 − α Γ ( 2 − α ) . K(t) = \frac{t^{\alpha-1}}{\Gamma(\alpha)}, \qquad\alpha\in\left(\frac{1}{2},1\right), \qquad\mathcal{R}([0,T]) = \frac{T^{1-\alpha}}{\Gamma(2-\alpha)}. K ( t ) = Γ ( α ) t α − 1 , α ∈ ( 2 1 , 1 ) , R ([ 0 , T ]) = Γ ( 2 − α ) T 1 − α .
Table 1 collects the resulting price bounds at ρ = − 1 \rho=-1 ρ = − 1 .
Thus, the endogenous specifications in Table 1 at ρ = − 1 \rho=-1 ρ = − 1 have an implied volatility smile with a flat, zero right wing beyond a finite strike. Reproducing a rising right wing motivates an additional volatility driver independent of W W W or, in the PDV model, a volatility function that increases again for sufficiently large positive values of the past-return factor, as discussed in Remark 3.3.
Sufficient conditions for positivity
We now return to the volatility σ \sigma σ in (PDV).
Notation 3. Under Assumption 1, we set
λ 1 : = inf supp ν . \lambda_1 := \inf\operatorname{supp}\nu. λ 1 := inf supp ν . We define the nonnegative measure
ν ^ ( d λ ) : = λ ( λ − λ 1 ) Λ 1 ν ( d λ ) . \hat{\nu}(d\lambda) := \frac{\lambda(\lambda-\lambda_1)}{\Lambda_1}\nu(d\lambda). ν ^ ( d λ ) := Λ 1 λ ( λ − λ 1 ) ν ( d λ ) . By condition (2) , it is finite and has total mass λ 1 , e f f − λ 1 \lambda_{1,\mathrm{eff}}-\lambda_1 λ 1 , eff − λ 1 . For λ > 0 \lambda>0 λ > 0 , we define
η 0 ( λ ) : = β 1 Λ 1 λ R 1 , 0 ( λ ) − σ 0 , D : = sup r ∈ [ λ 1 , ∞ ) ∫ λ ∈ ( 0 , r ] η 0 ( λ ) ν ^ ( d λ ) . \eta_0(\lambda) := \beta_1\frac{\Lambda_1}{\lambda}R_{1,0}(\lambda)-\sigma_0,\qquad D := \sup_{r\in[\lambda_1,\infty)}\int_{\lambda\in(0,r]} \eta_0(\lambda)\,\hat{\nu}(d\lambda). η 0 ( λ ) := β 1 λ Λ 1 R 1 , 0 ( λ ) − σ 0 , D := r ∈ [ λ 1 , ∞ ) sup ∫ λ ∈ ( 0 , r ] η 0 ( λ ) ν ^ ( d λ ) . Conditions (2) and (4) give
∫ λ ∈ ( 0 , ∞ ) ∣ η 0 ( λ ) ∣ ν ^ ( d λ ) ≤ ∣ β 1 ∣ ∫ λ ∈ ( 0 , ∞ ) λ ∣ R 1 , 0 ( λ ) ∣ ν ( d λ ) + ∣ σ 0 ∣ ( λ 1 , e f f − λ 1 ) < ∞ . (21) \int_{\lambda\in(0,\infty)}|\eta_0(\lambda)|\,\hat{\nu}(d\lambda)
\leq|\beta_1|\int_{\lambda\in(0,\infty)}\lambda|R_{1,0}(\lambda)|\,\nu(d\lambda)+|\sigma_0|(\lambda_{1,\mathrm{eff}}-\lambda_1)<\infty. \tag*{(21)} ∫ λ ∈ ( 0 , ∞ ) ∣ η 0 ( λ ) ∣ ν ^ ( d λ ) ≤ ∣ β 1 ∣ ∫ λ ∈ ( 0 , ∞ ) λ ∣ R 1 , 0 ( λ ) ∣ ν ( d λ ) + ∣ σ 0 ∣ ( λ 1 , eff − λ 1 ) < ∞. (21) Hence η 0 ∈ L 1 ( ν ^ ) \eta_0\in L^1(\hat{\nu}) η 0 ∈ L 1 ( ν ^ ) and D ≤ ∫ λ ∈ ( 0 , ∞ ) ∣ η 0 ( λ ) ∣ ν ^ ( d λ ) < ∞ D\leq\int_{\lambda\in(0,\infty)}|\eta_0(\lambda)|\,\hat{\nu}(d\lambda)<\infty D ≤ ∫ λ ∈ ( 0 , ∞ ) ∣ η 0 ( λ ) ∣ ν ^ ( d λ ) < ∞ . Moreover, ν ^ ( { λ 1 } ) = 0 \hat{\nu}(\{\lambda_1\})=0 ν ^ ({ λ 1 }) = 0 gives D ≥ 0 D\geq0 D ≥ 0 .
Under Assumption 1, R − Λ 1 − 1 δ 0 \mathcal{R}-\Lambda_1^{-1}\delta_0 R − Λ 1 − 1 δ 0 has a locally integrable density. Differentiating K 1 ∗ R = 1 K_1 * \mathcal{R}=1 K 1 ∗ R = 1 with [16 Chapter 3, Corollary 7.4(ii)] gives
Λ 1 R ( d t ) = δ 0 ( d t ) − ( K 1 ′ ∗ R ) ( t ) d t . (22) \Lambda_1\mathcal{R}(dt)=\delta_0(dt)-(K_1'*\mathcal{R})(t)\,dt. \tag*{(22)} Λ 1 R ( d t ) = δ 0 ( d t ) − ( K 1 ′ ∗ R ) ( t ) d t . (22)
Since R ( { 0 } ) = Λ 1 − 1 \mathcal{R}(\{0\})=\Lambda_1^{-1} R ({ 0 }) = Λ 1 − 1 , we have
( K 1 ′ ∗ R ) ( t ) = K 1 ′ ( t ) Λ 1 + ∫ ( 0 , t ] K 1 ′ ( t − s ) R ( d s ) . (K_1'*\mathcal{R})(t)=\frac{K_1'(t)}{\Lambda_1}+\int_{(0,t]}K_1'(t-s)\mathcal{R}(ds). ( K 1 ′ ∗ R ) ( t ) = Λ 1 K 1 ′ ( t ) + ∫ ( 0 , t ] K 1 ′ ( t − s ) R ( d s ) .
Continuity of K 1 ′ K_1' K 1 ′ and local finiteness of R \mathcal{R} R imply that the integral tends to zero as t ↓ 0 t\downarrow0 t ↓ 0 . Consequently,
( K 1 ′ ∗ R ) ( 0 ) = K 1 ′ ( 0 ) Λ 1 = − λ 1 , e f f . (23) (K_1'*\mathcal{R})(0)=\frac{K_1'(0)}{\Lambda_1}=-\lambda_{1,\mathrm{eff}}. \tag*{(23)} ( K 1 ′ ∗ R ) ( 0 ) = Λ 1 K 1 ′ ( 0 ) = − λ 1 , eff . (23)
By [16 Chapter 5, Theorem 5.4] , the density of R − Λ 1 − 1 δ 0 \mathcal{R}-\Lambda_1^{-1}\delta_0 R − Λ 1 − 1 δ 0 is completely monotone. Identity (22) therefore shows that − ( K 1 ′ ∗ R ) -(K_1'*\mathcal{R}) − ( K 1 ′ ∗ R ) is completely monotone on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Complete monotonicity and (23) imply local absolute continuity on R + \mathbb{R}_+ R + . Thus K 1 ′ ∗ R K_1'*\mathcal{R} K 1 ′ ∗ R is nonpositive and nondecreasing, which gives
for almost every t > 0 , ( K 1 ′ ∗ R ) ′ ( t ) ≥ 0 , \text{for almost every }t>0,\qquad(K_1'*\mathcal{R})'(t)\geq0, for almost every t > 0 , ( K 1 ′ ∗ R ) ′ ( t ) ≥ 0 ,
∀ t ≥ 0 , ∫ 0 t ( K 1 ′ ∗ R ) ′ ( s ) d s = ( K 1 ′ ∗ R ) ( t ) + λ 1 , e f f ≤ λ 1 , e f f . (24) \forall t\geq0,\qquad\int_0^t(K_1'*\mathcal{R})'(s)\,ds=(K_1'*\mathcal{R})(t)+\lambda_{1,\mathrm{eff}}\leq\lambda_{1,\mathrm{eff}}. \tag*{(24)} ∀ t ≥ 0 , ∫ 0 t ( K 1 ′ ∗ R ) ′ ( s ) d s = ( K 1 ′ ∗ R ) ( t ) + λ 1 , eff ≤ λ 1 , eff . (24)
Lemma 3.7. Under Assumptions 1 and 2, for every T > 0 T>0 T > 0 , the process σ \sigma σ is a semimartingale with dynamics
d σ t = ( − λ 1 , e f f σ t + ( ( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) ) d t + Λ 1 d d t ( R ∗ ( β 0 + β 1 g 1 + β 2 R ‾ 2 ) ) ( t ) d t + β 1 Λ 1 σ t d W t , (25) d\sigma_t=\left(-\lambda_{1,\mathrm{eff}}\sigma_t+((K_1'*\mathcal{R})'*\sigma)(t)\right)dt+\Lambda_1\frac{d}{dt}\left(\mathcal{R}*(\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{\mathcal{R}}_2})\right)(t)\,dt+\beta_1\Lambda_1\sigma_t\,dW_t, \tag*{(25)} d σ t = ( − λ 1 , eff σ t + (( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) ) d t + Λ 1 d t d ( R ∗ ( β 0 + β 1 g 1 + β 2 R 2 ) ) ( t ) d t + β 1 Λ 1 σ t d W t , (25) where the drift is defined for almost every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , almost surely.
Proof. By Proposition 1, almost surely, g 1 g_1 g 1 and R ‾ 2 \overline{\mathcal{R}}_2 R 2 are absolutely continuous on [ 0 , T ] [0,T] [ 0 , T ] and inf t ∈ [ 0 , T ] R ‾ 2 , t ≥ min t ∈ [ 0 , T ] g 2 ( t ) > 0 \inf_{t\in[0,T]}\overline{\mathcal{R}}_{2,t}\geq\min_{t\in[0,T]}g_2(t)>0 inf t ∈ [ 0 , T ] R 2 , t ≥ min t ∈ [ 0 , T ] g 2 ( t ) > 0 by Assumption 2. The chain rule therefore gives absolute continuity of R ‾ 2 \sqrt{\overline{\mathcal{R}}_2} R 2 . Convolving (22) with σ \sigma σ and using (10) gives
∀ t ∈ [ 0 , T ] , σ t = Λ 1 ( R ∗ ( β 0 + β 1 g 1 + β 2 R ‾ 2 ) ) ( t ) + ( ( K 1 ′ ∗ R ) ∗ σ ) ( t ) + β 1 Λ 1 ∫ 0 t σ s d W s . (26) \forall t\in[0,T],\qquad\sigma_t=\Lambda_1\left(\mathcal{R}*(\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{\mathcal{R}}_2})\right)(t)+((K_1'*\mathcal{R})*\sigma)(t)+\beta_1\Lambda_1\int_0^t\sigma_s\,dW_s. \tag*{(26)} ∀ t ∈ [ 0 , T ] , σ t = Λ 1 ( R ∗ ( β 0 + β 1 g 1 + β 2 R 2 ) ) ( t ) + (( K 1 ′ ∗ R ) ∗ σ ) ( t ) + β 1 Λ 1 ∫ 0 t σ s d W s . (26)
We show that the two convolution terms in (26) are absolutely continuous, so that their time derivatives can be taken to obtain (25) . By (23) and (24) , K 1 ′ ∗ R K_1'*\mathcal{R} K 1 ′ ∗ R is absolutely continuous on [ 0 , T ] [0,T] [ 0 , T ] and ( K 1 ′ ∗ R ) ( 0 ) = − λ 1 , e f f (K_1'*\mathcal{R})(0)=-\lambda_{1,\mathrm{eff}} ( K 1 ′ ∗ R ) ( 0 ) = − λ 1 , eff . Hence
d d t ( ( ( K 1 ′ ∗ R ) ∗ σ ) ( t ) ) = ( K 1 ′ ∗ R ) ( 0 ) σ t + ( ( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) = − λ 1 , e f f σ t + ( ( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) . (27) \frac{d}{dt}\left(((K_1'*\mathcal{R})*\sigma)(t)\right)=(K_1'*\mathcal{R})(0)\sigma_t+((K_1'*\mathcal{R})'*\sigma)(t)=-\lambda_{1,\mathrm{eff}}\sigma_t+((K_1'*\mathcal{R})'*\sigma)(t). \tag*{(27)} d t d ( (( K 1 ′ ∗ R ) ∗ σ ) ( t ) ) = ( K 1 ′ ∗ R ) ( 0 ) σ t + (( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) = − λ 1 , eff σ t + (( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) . (27)
The right-hand side is integrable on [ 0 , T ] [0,T] [ 0 , T ] , so ( K 1 ′ ∗ R ) ∗ σ (K_1'*\mathcal{R})*\sigma ( K 1 ′ ∗ R ) ∗ σ is absolutely continuous. For the other convolution in (26) , the measure identity (22) gives
Λ 1 ( R ∗ ( β 0 + β 1 g 1 + β 2 R ‾ 2 ) ) = β 0 + β 1 g 1 + β 2 R ‾ 2 − ( K 1 ′ ∗ R ) ∗ ( β 0 + β 1 g 1 + β 2 R ‾ 2 ) . (28) \Lambda_1\left(\mathcal{R}*(\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{\mathcal{R}}_2})\right)=\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{\mathcal{R}}_2}-(K_1'*\mathcal{R})*(\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{\mathcal{R}}_2}). \tag*{(28)} Λ 1 ( R ∗ ( β 0 + β 1 g 1 + β 2 R 2 ) ) = β 0 + β 1 g 1 + β 2 R 2 − ( K 1 ′ ∗ R ) ∗ ( β 0 + β 1 g 1 + β 2 R 2 ) . (28)
The function β 0 + β 1 g 1 + β 2 R ‾ 2 \beta_0+\beta_1g_1+\beta_2\sqrt{\overline{\mathcal{R}}_2} β 0 + β 1 g 1 + β 2 R 2 is absolutely continuous. Formula (27) , applied with σ \sigma σ replaced by this continuous function, shows that its convolution with K 1 ′ ∗ R K_1'*\mathcal{R} K 1 ′ ∗ R is absolutely continuous as well. Identity (28) therefore shows that the first convolution in (26) is absolutely continuous, completing the proof. □
Assumption 3. The following conditions hold:
(i) σ 0 ≥ 0 \sigma_0 \ge0 σ 0 ≥ 0 .
(ii) For a ∈ R a \in\mathbb{R} a ∈ R , we consider, for each s ≥ 0 s \ge0 s ≥ 0 and almost every t ≥ 0 t \ge0 t ≥ 0 ,
for t ≥ s t \ge s t ≥ s ,
∂ t K 2 ( s , t ) + ( 2 λ 1 − a ) K 2 ( s , t ) ≥ 0 (29) \partial_t K_2(s,t) + (2\lambda_1-a)K_2(s,t) \ge0 \tag*{(29)} ∂ t K 2 ( s , t ) + ( 2 λ 1 − a ) K 2 ( s , t ) ≥ 0 (29) g ˙ 2 ( t ) + ( 2 λ 1 − a ) g 2 ( t ) ≥ 0. (30) \dot{g}_2(t) + (2\lambda_1-a)g_2(t) \ge0. \tag*{(30)} g ˙ 2 ( t ) + ( 2 λ 1 − a ) g 2 ( t ) ≥ 0. (30) We define the set
{ a ∈ R : x r e f ‘ ( 3.20 ) ‘ and x r e f ‘ ( 3.21 ) ‘ hold } . \{a \in\mathbb{R} : {xref}`(3.20)` \text{ and } {xref}`(3.21)` \text{ hold}\}. { a ∈ R : x r e f ‘ ( 3.20 ) ‘ and x r e f ‘ ( 3.21 ) ‘ hold } . We assume that this set is nonempty. It is closed and, by integration of (30) over [ 0 , 1 ] [0,1] [ 0 , 1 ] , bounded above by 2 λ 1 + log g 2 ( 1 ) g 2 ( 0 ) 2\lambda_1 + \log\frac{g_2(1)}{g_2(0)} 2 λ 1 + log g 2 ( 0 ) g 2 ( 1 ) , where g 2 ( 0 ) , g 2 ( 1 ) > 0 g_2(0),g_2(1)>0 g 2 ( 0 ) , g 2 ( 1 ) > 0 by Assumption 2. We define m m m as its maximum. We also assume that either m ≥ 0 m \ge0 m ≥ 0 , or m < 0 m < 0 m < 0 and
∀ t ≥ 0 , g 2 ( t ) ≤ g 2 ( 0 ) , \forall t \ge0,\qquad g_2(t) \le g_2(0), ∀ t ≥ 0 , g 2 ( t ) ≤ g 2 ( 0 ) , for almost every s ≥ t s \ge t s ≥ t ,
∂ s K 2 ( t , s ) + 2 λ 1 K 2 ( t , s ) ≥ 0. \partial_s K_2(t,s) + 2\lambda_1 K_2(t,s) \ge0. ∂ s K 2 ( t , s ) + 2 λ 1 K 2 ( t , s ) ≥ 0. (iii) The following bound holds:
D ≤ β 0 λ 1 + β 2 m 2 R ‾ 2 , 0 . (31) D \le\beta_0\lambda_1 + \frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}. \tag*{(31)} D ≤ β 0 λ 1 + 2 β 2 m R 2 , 0 . (31)
Remark 3.8. By (25) , the diffusion coefficient β 1 Λ 1 σ t \beta_1\Lambda_1\sigma_t β 1 Λ 1 σ t vanishes at σ t = 0 \sigma_t=0 σ t = 0 , as does the drift term − λ 1 , e f f σ t -\lambda_{1,\mathrm{eff}}\sigma_t − λ 1 , eff σ t . By the nonnegativity of ( K 1 ′ ∗ R ) ′ (K'_1 * \mathcal{R})' ( K 1 ′ ∗ R ) ′ in (24) , ( ( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) ≥ 0 ((K'_1 * \mathcal{R})' * \sigma)(t) \ge0 (( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) ≥ 0 if σ s ≥ 0 \sigma_s \ge0 σ s ≥ 0 for all s ∈ [ 0 , t ] s \in[0,t] s ∈ [ 0 , t ] . To prevent the drift from pushing volatility below zero at σ t = 0 \sigma_t=0 σ t = 0 , it therefore remains to show that
for almost every t ≥ 0 , Λ 1 d d t ( R ∗ ( β 0 + β 1 g 1 + β 2 R ‾ 2 ) ) ( t ) ≥ 0. \text{for almost every }t \ge0,\qquad\Lambda_1 \frac{d}{dt}\left(\mathcal{R} * \left(\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{R}_2}\right)\right)(t) \ge0. for almost every t ≥ 0 , Λ 1 d t d ( R ∗ ( β 0 + β 1 g 1 + β 2 R 2 ) ) ( t ) ≥ 0.
The assumptions control the contributions of R ‾ 2 \sqrt{\overline{R}_2} R 2 and g 1 g_1 g 1 to this term as follows.
For a single exponential in each kernel, with initial inputs as in (5) , at σ = 0 \sigma=0 σ = 0 , β 1 R ‾ 1 \beta_1\overline{R}_1 β 1 R 1 relaxes toward zero at rate λ 1 , 1 \lambda_{1,1} λ 1 , 1 , raising σ \sigma σ , whereas β 2 R ‾ 2 \beta_2\sqrt{\overline{R}_2} β 2 R 2 decays at rate λ 2 , 1 2 \frac{\lambda_{2,1}}{2} 2 λ 2 , 1 , lowering σ \sigma σ . Similarly, Condition (ii) controls the contribution of R ‾ 2 \sqrt{\overline{R}_2} R 2 by limiting its decay relative to the return kernel K 1 K_1 K 1 . More precisely, (30) controls the decay of g 2 ( t ) g_2(t) g 2 ( t ) , while (29) controls that of the integral term ∫ 0 t K 2 ( s , t ) σ s 2 d s \int_0^t K_2(s,t)\sigma_s^2\,ds ∫ 0 t K 2 ( s , t ) σ s 2 d s .
Condition (iii) bounds, through D D D , the potentially negative drift contribution of β 1 g 1 \beta_1g_1 β 1 g 1 .
Remark 3.9. Since g 2 > 0 g_2>0 g 2 > 0 by Assumption 2, the definition of m m m in Assumption 3 (ii) equivalently reads
m = 2 λ 1 − max { ess sup t ≥ 0 − g ˙ 2 ( t ) g 2 ( t ) , sup s ≥ 0 ess sup t ≥ s , K 2 ( s , t ) > 0 − ∂ t K 2 ( s , t ) K 2 ( s , t ) } . m = 2\lambda_1-\max\left\{\operatorname*{ess\,sup}_{t\ge0}\frac{-\dot{g}_2(t)}{g_2(t)},\ \sup_{s\ge0}\operatorname*{ess\,sup}_{t\ge s,\ K_2(s,t)>0}\frac{-\partial_tK_2(s,t)}{K_2(s,t)}\right\}. m = 2 λ 1 − max { t ≥ 0 ess sup g 2 ( t ) − g ˙ 2 ( t ) , s ≥ 0 sup t ≥ s , K 2 ( s , t ) > 0 ess sup K 2 ( s , t ) − ∂ t K 2 ( s , t ) } .
We define Z Z Z by
d Z t = − λ 1 , e f f Z t d t + β 1 Λ 1 Z t d W t , Z 0 = 1 , (32) dZ_t=-\lambda_{1,\mathrm{eff}}Z_t\,dt+\beta_1\Lambda_1Z_t\,dW_t,\qquad Z_0=1, \tag*{(32)} d Z t = − λ 1 , eff Z t d t + β 1 Λ 1 Z t d W t , Z 0 = 1 , (32)
i.e., integrating this SDE,
∀ t ≥ 0 , Z t = exp ( β 1 Λ 1 W t − ( λ 1 , e f f + β 1 2 Λ 1 2 2 ) t ) > 0. \forall t \ge0,\qquad Z_t=\exp\left(\beta_1\Lambda_1W_t-\left(\lambda_{1,\mathrm{eff}}+\frac{\beta_1^2\Lambda_1^2}{2}\right)t\right)>0. ∀ t ≥ 0 , Z t = exp ( β 1 Λ 1 W t − ( λ 1 , eff + 2 β 1 2 Λ 1 2 ) t ) > 0.
Theorem 2. Under Assumptions 1, 2, and 3, P \mathbb{P} P -almost surely, the following bound holds:
∀ t ≥ 0 , σ t ≥ Z t ( σ 0 + ( β 0 λ 1 − ( D − β 2 m 2 R ‾ 2 , 0 ) ) + ∫ 0 t Z s − 1 d s ) ≥ 0. (33) \forall t \ge0,\qquad\sigma_t \ge Z_t\left(\sigma_0+\left(\beta_0\lambda_1-\left(D-\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}\right)\right)_+\int_0^t Z_s^{-1}\,ds\right)\ge0. \tag*{(33)} ∀ t ≥ 0 , σ t ≥ Z t ( σ 0 + ( β 0 λ 1 − ( D − 2 β 2 m R 2 , 0 ) ) + ∫ 0 t Z s − 1 d s ) ≥ 0. (33)
If σ 0 > 0 \sigma_0 > 0 σ 0 > 0 , the lower bound is strictly positive for every t ≥ 0 t \ge0 t ≥ 0 .
If σ 0 = 0 \sigma_0 = 0 σ 0 = 0 , it is strictly positive for every t > 0 t > 0 t > 0 provided that
β 0 λ 1 > ( D − β 2 m 2 R ‾ 2 , 0 ) + . \beta_0\lambda_1 > \left(D-\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}\right)_+. β 0 λ 1 > ( D − 2 β 2 m R 2 , 0 ) + .
Moreover, for every T > 0 T > 0 T > 0 , ( S t ) t ∈ [ 0 , T ] (S_t)_{t\in[0,T]} ( S t ) t ∈ [ 0 , T ] is a bounded martingale under P \mathbb{P} P .
Remark 3.10. The positivity proof requires K 1 ( 0 ) < ∞ K_1(0) < \infty K 1 ( 0 ) < ∞ and therefore does not cover kernels with K 1 ( t ) → ∞ K_1(t) \to\infty K 1 ( t ) → ∞ as t ↓ 0 t \downarrow0 t ↓ 0 . Whether volatility reaches zero can depend on whether the kernel is regular or rough. For Volterra CIR processes with regular kernels, Bondi and Pulido [7 Theorem 9 (ii)] give a Feller-type condition ensuring that zero is inaccessible. By contrast, the variance in the fractional rough Heston model has a positive atom at zero at every positive time, see [9 Theorem 3] and [14 Theorem 1.6, equation (1.9)] .
Remark 3.11. For a diffusion
d Y t = b ( Y t ) d t + Σ ( Y t ) d W t , C : = { y ∈ R d : y 1 ≥ 0 } , dY_t = b(Y_t)\,dt + \Sigma(Y_t)\,dW_t,\qquad C := \{y \in\mathbb{R}^d : y_1 \ge0\}, d Y t = b ( Y t ) d t + Σ ( Y t ) d W t , C := { y ∈ R d : y 1 ≥ 0 } ,
[13 Theorem 4.1] characterizes local stochastic invariance of C C C , meaning that solutions starting in C C C remain there locally. Under the hypotheses of [13 Theorem 4.1] , the criterion is given by
∀ y ∈ R d with y 1 = 0 , Σ 1 , ⋅ ( y ) = 0 , b 1 ( y ) ≥ 0. \forall y \in\mathbb{R}^d\text{ with }y_1=0,\qquad\Sigma_{1,\cdot}(y)=0,\qquad b_1(y)\ge0. ∀ y ∈ R d with y 1 = 0 , Σ 1 , ⋅ ( y ) = 0 , b 1 ( y ) ≥ 0.
In (25) , the coefficient of d W t dW_t d W t vanishes at σ t = 0 \sigma_t = 0 σ t = 0 . The drift depends on the past through ( ( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) ((K_1' * R)' * \sigma)(t) (( K 1 ′ ∗ R ) ′ ∗ σ ) ( t ) and on R ‾ 2 , t \overline{R}_{2,t} R 2 , t , so σ \sigma σ alone is not a diffusion with coefficients determined by its current value. The proof applies Grönwall’s lemma to σ − \sigma^- σ − without invoking this diffusion criterion.
Proof of Theorem 2. Lemma 3.7 gives the decomposition (25) . The dynamics (9) and conditions (29) and (30) give R ‾ ˙ 2 , t + ( 2 λ 1 − m ) R ‾ 2 , t ≥ 0 \dot{\overline{R}}_{2,t} + (2\lambda_1-m)\overline{R}_{2,t} \ge0 R ˙ 2 , t + ( 2 λ 1 − m ) R 2 , t ≥ 0 for almost every t ≥ 0 t \ge0 t ≥ 0 . Since R ‾ 2 , 0 = g 2 ( 0 ) > 0 \overline{R}_{2,0}=g_2(0)>0 R 2 , 0 = g 2 ( 0 ) > 0 by (1) ,
∀ t ≥ 0 , R ‾ 2 , t ≥ e − ( 2 λ 1 − m ) t R ‾ 2 , 0 > 0. (34) \forall t \ge0,\qquad\overline{R}_{2,t} \ge e^{-(2\lambda_1-m)t}\overline{R}_{2,0}>0. \tag*{(34)} ∀ t ≥ 0 , R 2 , t ≥ e − ( 2 λ 1 − m ) t R 2 , 0 > 0. (34)
In (25) , when σ t = 0 \sigma_t=0 σ t = 0 , both β 1 Λ 1 σ t \beta_1\Lambda_1\sigma_t β 1 Λ 1 σ t and − λ 1 , e f f σ t -\lambda_{1,\mathrm{eff}}\sigma_t − λ 1 , eff σ t vanish.
Step 1: Nonnegativity of the last drift term in (25) . Since R ∗ K 1 = 1 R * K_1=1 R ∗ K 1 = 1 , we have
Λ 1 d d t ( R ∗ ( β 0 + β 1 g 1 + β 2 R ‾ 2 ) ) ( t ) = Λ 1 d d t [ R ∗ ( β 0 + β 1 g 1 + β 2 R ‾ 2 − σ 0 Λ 1 K 1 ) ] ( t ) . \Lambda_1\frac{d}{dt}\left(R * \left(\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{R}_2}\right)\right)(t)
=\Lambda_1\frac{d}{dt}\left[R * \left(\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{R}_2}-\frac{\sigma_0}{\Lambda_1}K_1\right)\right](t). Λ 1 d t d ( R ∗ ( β 0 + β 1 g 1 + β 2 R 2 ) ) ( t ) = Λ 1 d t d [ R ∗ ( β 0 + β 1 g 1 + β 2 R 2 − Λ 1 σ 0 K 1 ) ] ( t ) .
We define, for almost every t ≥ 0 t \ge0 t ≥ 0 ,
H t : = ( d d t + λ 1 ) ( β 0 + β 1 g 1 ( t ) + β 2 R ‾ 2 , t − σ 0 Λ 1 K 1 ( t ) ) . (35) H_t:=\left(\frac{d}{dt}+\lambda_1\right)\left(\beta_0+\beta_1g_1(t)+\beta_2\sqrt{\overline{R}_{2,t}}-\frac{\sigma_0}{\Lambda_1}K_1(t)\right). \tag*{(35)} H t := ( d t d + λ 1 ) ( β 0 + β 1 g 1 ( t ) + β 2 R 2 , t − Λ 1 σ 0 K 1 ( t ) ) . (35)
Since K 1 ( 0 ) = Λ 1 K_1(0)=\Lambda_1 K 1 ( 0 ) = Λ 1 and σ 0 = β 0 + β 1 g 1 ( 0 ) + β 2 R ‾ 2 , 0 \sigma_0=\beta_0+\beta_1g_1(0)+\beta_2\sqrt{\overline{R}_{2,0}} σ 0 = β 0 + β 1 g 1 ( 0 ) + β 2 R 2 , 0 by (1) , the expression in parentheses in (35) is zero at t = 0 t=0 t = 0 . Hence we have
β 0 + β 1 g 1 ( t ) + β 2 R ‾ 2 , t − σ 0 Λ 1 K 1 ( t ) = ∫ 0 t e − λ 1 ( t − s ) H s d s . \beta_0+\beta_1g_1(t)+\beta_2\sqrt{\overline{R}_{2,t}}-\frac{\sigma_0}{\Lambda_1}K_1(t)=\int_0^t e^{-\lambda_1(t-s)}H_s\,ds. β 0 + β 1 g 1 ( t ) + β 2 R 2 , t − Λ 1 σ 0 K 1 ( t ) = ∫ 0 t e − λ 1 ( t − s ) H s d s .
We define, for almost every t ≥ 0 t \ge0 t ≥ 0 ,
J ( t ) : = Λ 1 d d t ( R ∗ e − λ 1 ⋅ ) ( t ) . J(t):=\Lambda_1\frac{d}{dt}\left(R*e^{-\lambda_1\cdot}\right)(t). J ( t ) := Λ 1 d t d ( R ∗ e − λ 1 ⋅ ) ( t ) .
Identity (22) gives Λ 1 R ( { 0 } ) = 1 \Lambda_1R(\{0\})=1 Λ 1 R ({ 0 }) = 1 . Convolving with R R R and differentiating gives, for almost every t ≥ 0 t\ge0 t ≥ 0 ,
Λ 1 d d t ( R ∗ ( β 0 + β 1 g 1 + β 2 R ‾ 2 ) ) ( t ) = Λ 1 d d t [ ( R ∗ e − λ 1 ⋅ ) ∗ H ] ( t ) = Λ 1 ( R ∗ e − λ 1 ⋅ ) ( 0 ) H t + Λ 1 ∫ 0 t ( R ∗ e − λ 1 ⋅ ) ′ ( t − s ) H s d s = Λ 1 R ( { 0 } ) H t + ∫ 0 t J ( t − s ) H s d s = H t + ( J ∗ H ) ( t ) . (36) \begin{aligned}
\Lambda_1\frac{d}{dt}\left(R*\left(\beta_0+\beta_1g_1+\beta_2\sqrt{\overline{R}_2}\right)\right)(t)
=\Lambda_1\frac{d}{dt}\left[(R*e^{-\lambda_1\cdot})*H\right](t) \\
&= \Lambda_1(\mathcal{R} \ast e^{-\lambda_1\cdot})(0)H_t+\Lambda_1\int_0^t(\mathcal{R}\ast e^{-\lambda_1\cdot})'(t-s)H_s\,ds \\
&= \Lambda_1\mathcal{R}(\{0\})H_t+\int_0^t J(t-s)H_s\,ds=H_t+(J\ast H)(t). \tag*{(36)}
\end{aligned} Λ 1 d t d ( R ∗ ( β 0 + β 1 g 1 + β 2 R 2 ) ) ( t ) = Λ 1 d t d [ ( R ∗ e − λ 1 ⋅ ) ∗ H ] ( t ) = Λ 1 ( R ∗ e − λ 1 ⋅ ) ( 0 ) H t + Λ 1 ∫ 0 t ( R ∗ e − λ 1 ⋅ ) ′ ( t − s ) H s d s = Λ 1 R ({ 0 }) H t + ∫ 0 t J ( t − s ) H s d s = H t + ( J ∗ H ) ( t ) . (36)
We prove that J ≥ 0 J\geq0 J ≥ 0 and H ≥ 0 H\geq0 H ≥ 0 , which makes the right-hand side of (36) nonnegative. Since K 1 ∗ R = 1 K_1\ast\mathcal{R}=1 K 1 ∗ R = 1 , associativity gives
( K 1 ∗ ( R ∗ e − λ 1 ⋅ ) ) ( t ) = ( ( K 1 ∗ R ) ∗ e − λ 1 ⋅ ) ( t ) = ∫ 0 t e − λ 1 s d s . \left(K_1\ast(\mathcal{R}\ast e^{-\lambda_1\cdot})\right)(t)=\left((K_1\ast\mathcal{R})\ast e^{-\lambda_1\cdot}\right)(t)=\int_0^t e^{-\lambda_1s}\,ds. ( K 1 ∗ ( R ∗ e − λ 1 ⋅ ) ) ( t ) = ( ( K 1 ∗ R ) ∗ e − λ 1 ⋅ ) ( t ) = ∫ 0 t e − λ 1 s d s .
Since ( R ∗ e − λ 1 ⋅ ) ( 0 ) = 1 Λ 1 (\mathcal{R}\ast e^{-\lambda_1\cdot})(0)=\frac{1}{\Lambda_1} ( R ∗ e − λ 1 ⋅ ) ( 0 ) = Λ 1 1 and ( R ∗ e − λ 1 ⋅ ) ′ = J Λ 1 (\mathcal{R}\ast e^{-\lambda_1\cdot})'=\frac{J}{\Lambda_1} ( R ∗ e − λ 1 ⋅ ) ′ = Λ 1 J , differentiation yields
K 1 + J ∗ K 1 = Λ 1 e − λ 1 ⋅ . (37) K_1+J\ast K_1=\Lambda_1e^{-\lambda_1\cdot}. \tag*{(37)} K 1 + J ∗ K 1 = Λ 1 e − λ 1 ⋅ . (37)
The signed measure δ 0 + e λ 1 t J ( t ) d t \delta_0+e^{\lambda_1t}J(t)\,dt δ 0 + e λ 1 t J ( t ) d t satisfies the resolvent identity for
e λ 1 t K 1 ( t ) Λ 1 = 1 Λ 1 ∫ λ ∈ ( 0 , ∞ ) λ e − ( λ − λ 1 ) t ν ( d λ ) . \frac{e^{\lambda_1t}K_1(t)}{\Lambda_1}=\frac{1}{\Lambda_1}\int_{\lambda\in(0,\infty)}\lambda e^{-(\lambda-\lambda_1)t}\,\nu(d\lambda). Λ 1 e λ 1 t K 1 ( t ) = Λ 1 1 ∫ λ ∈ ( 0 , ∞ ) λ e − ( λ − λ 1 ) t ν ( d λ ) .
This kernel is completely monotone since λ − λ 1 ≥ 0 \lambda-\lambda_1\geq0 λ − λ 1 ≥ 0 on supp ν \operatorname{supp}\nu supp ν . The map t ↦ e λ 1 t K 1 ( t ) Λ 1 t\mapsto\frac{e^{\lambda_1t}K_1(t)}{\Lambda_1} t ↦ Λ 1 e λ 1 t K 1 ( t ) belongs to C 1 ( R + ) C^1(\mathbb{R}_+) C 1 ( R + ) . By [16 Chapter 5, Theorems 5.2 and 5.4] , the unique locally finite signed measure satisfying the resolvent identity for this kernel is nonnegative. Hence J ≥ 0 J\geq0 J ≥ 0 . We now prove that H ≥ 0 H\geq0 H ≥ 0 . The integrability in (4) and the representation in (3) give
∀ t ≥ 0 , g ˙ 1 ( t ) = − ∫ λ ∈ ( 0 , ∞ ) λ e − λ t R 1 , 0 ( λ ) ν ( d λ ) . \forall t\geq0,\qquad\dot{g}_1(t)=-\int_{\lambda\in(0,\infty)}\lambda e^{-\lambda t}R_{1,0}(\lambda)\,\nu(d\lambda). ∀ t ≥ 0 , g ˙ 1 ( t ) = − ∫ λ ∈ ( 0 , ∞ ) λ e − λ t R 1 , 0 ( λ ) ν ( d λ ) .
Together with the definitions in Notations 2 and 3, this yields, for almost every t ≥ 0 t\geq0 t ≥ 0 ,
H t = β 0 λ 1 + β 2 ( R ‾ ˙ 2 , t 2 R ‾ 2 , t + λ 1 R ‾ 2 , t ) + β 1 ( g ˙ 1 ( t ) + λ 1 g 1 ( t ) ) − σ 0 Λ 1 ( K 1 ′ ( t ) + λ 1 K 1 ( t ) ) = β 0 λ 1 + β 2 ( R ‾ ˙ 2 , t 2 R ‾ 2 , t + λ 1 R ‾ 2 , t ) − ∫ λ ∈ ( 0 , ∞ ) e − λ t ( λ − λ 1 ) ( β 1 R 1 , 0 ( λ ) − σ 0 λ Λ 1 ) ν ( d λ ) = β 0 λ 1 + β 2 ( R ‾ ˙ 2 , t 2 R ‾ 2 , t + λ 1 R ‾ 2 , t ) − ∫ λ ∈ ( 0 , ∞ ) e − λ t η 0 ( λ ) ν ^ ( d λ ) . \begin{aligned}
H_t={}&\beta_0\lambda_1+\beta_2\left(\frac{\dot{\overline{R}}_{2,t}}{2\sqrt{\overline{R}_{2,t}}}+\lambda_1\sqrt{\overline{R}_{2,t}}\right)+\beta_1\left(\dot{g}_1(t)+\lambda_1g_1(t)\right)-\frac{\sigma_0}{\Lambda_1}\left(K_1'(t)+\lambda_1K_1(t)\right)\\
={}&\beta_0\lambda_1+\beta_2\left(\frac{\dot{\overline{R}}_{2,t}}{2\sqrt{\overline{R}_{2,t}}}+\lambda_1\sqrt{\overline{R}_{2,t}}\right)-\int_{\lambda\in(0,\infty)}e^{-\lambda t}(\lambda-\lambda_1)\left(\beta_1R_{1,0}(\lambda)-\frac{\sigma_0\lambda}{\Lambda_1}\right)\nu(d\lambda)\\
={}&\beta_0\lambda_1+\beta_2\left(\frac{\dot{\overline{R}}_{2,t}}{2\sqrt{\overline{R}_{2,t}}}+\lambda_1\sqrt{\overline{R}_{2,t}}\right)-\int_{\lambda\in(0,\infty)}e^{-\lambda t}\eta_0(\lambda)\,\widehat{\nu}(d\lambda).
\end{aligned} H t = = = β 0 λ 1 + β 2 2 R 2 , t R ˙ 2 , t + λ 1 R 2 , t + β 1 ( g ˙ 1 ( t ) + λ 1 g 1 ( t ) ) − Λ 1 σ 0 ( K 1 ′ ( t ) + λ 1 K 1 ( t ) ) β 0 λ 1 + β 2 2 R 2 , t R ˙ 2 , t + λ 1 R 2 , t − ∫ λ ∈ ( 0 , ∞ ) e − λ t ( λ − λ 1 ) ( β 1 R 1 , 0 ( λ ) − Λ 1 σ 0 λ ) ν ( d λ ) β 0 λ 1 + β 2 2 R 2 , t R ˙ 2 , t + λ 1 R 2 , t − ∫ λ ∈ ( 0 , ∞ ) e − λ t η 0 ( λ ) ν ( d λ ) .
Since (34) gives R ‾ 2 , t > 0 \overline{R}_{2,t}>0 R 2 , t > 0 , the chain rule, the dynamics (9) , the definition of R ‾ 2 \overline{R}_2 R 2 in (PDV), and conditions (29) and (30) in Assumption 3 give, for almost every t ≥ 0 t\geq0 t ≥ 0 ,
R ‾ ˙ 2 , t 2 R ‾ 2 , t + λ 1 R ‾ 2 , t = 1 2 R ‾ 2 , t ( g ˙ 2 ( t ) + 2 λ 1 g 2 ( t ) + K 2 ( t , t ) σ t 2 + ∫ 0 t ( ∂ t K 2 ( s , t ) + 2 λ 1 K 2 ( s , t ) ) σ s 2 d s ) ≥ m 2 R ‾ 2 , t ( g 2 ( t ) + ∫ 0 t K 2 ( s , t ) σ s 2 d s ) = m 2 R ‾ 2 , t . (38) \begin{aligned}
\frac{\dot{\overline{R}}_{2,t}}{2\sqrt{\overline{R}_{2,t}}}+\lambda_1\sqrt{\overline{R}_{2,t}}
={}&\frac{1}{2\sqrt{\overline{R}_{2,t}}}\left(\dot{g}_2(t)+2\lambda_1g_2(t)+K_2(t,t)\sigma_t^2+\int_0^t\left(\partial_tK_2(s,t)+2\lambda_1K_2(s,t)\right)\sigma_s^2\,ds\right)\\
\geq{}&\frac{m}{2\sqrt{\overline{R}_{2,t}}}\left(g_2(t)+\int_0^tK_2(s,t)\sigma_s^2\,ds\right)\\
={}&\frac{m}{2}\sqrt{\overline{R}_{2,t}}. \tag*{(38)}
\end{aligned} 2 R 2 , t R ˙ 2 , t + λ 1 R 2 , t = ≥ = 2 R 2 , t 1 ( g ˙ 2 ( t ) + 2 λ 1 g 2 ( t ) + K 2 ( t , t ) σ t 2 + ∫ 0 t ( ∂ t K 2 ( s , t ) + 2 λ 1 K 2 ( s , t ) ) σ s 2 d s ) 2 R 2 , t m ( g 2 ( t ) + ∫ 0 t K 2 ( s , t ) σ s 2 d s ) 2 m R 2 , t . (38)
Taking square roots in (34) gives
R ‾ 2 , t ≥ R ‾ 2 , 0 e − ( λ 1 − m 2 ) t . \sqrt{\overline{R}_{2,t}}\geq\sqrt{\overline{R}_{2,0}}e^{-(\lambda_1-\frac{m}{2})t}. R 2 , t ≥ R 2 , 0 e − ( λ 1 − 2 m ) t .
For t > 0 t>0 t > 0 and λ > 0 \lambda>0 λ > 0 , we have
e − λ t = t ∫ x ∈ [ λ , ∞ ) e − x t d x . e^{-\lambda t}=t\int_{x\in[\lambda,\infty)}e^{-xt}\,dx. e − λ t = t ∫ x ∈ [ λ , ∞ ) e − x t d x .
The integrability in (21) , Fubini’s theorem, and the definition of D D D in Notation 3 yield
∀ t > 0 , ∫ λ ∈ ( 0 , ∞ ) e − λ t η 0 ( λ ) ν ^ ( d λ ) = t ∫ x ∈ [ λ 1 , ∞ ) e − x t ( ∫ λ ∈ ( 0 , x ] η 0 ( λ ) ν ^ ( d λ ) ) d x ≤ D t ∫ x ∈ [ λ 1 , ∞ ) e − x t d x = D e − λ 1 t . (39) \begin{aligned}
\forall t>0,\qquad
\int_{\lambda\in(0,\infty)}e^{-\lambda t}\eta_0(\lambda)\,\hat{\nu}(d\lambda)
&=t\int_{x\in[\lambda_1,\infty)}e^{-xt}\left(\int_{\lambda\in(0,x]}\eta_0(\lambda)\,\hat{\nu}(d\lambda)\right)\,dx\\
&\le Dt\int_{x\in[\lambda_1,\infty)}e^{-xt}\,dx
=De^{-\lambda_1t}. \tag*{(39)}
\end{aligned} ∀ t > 0 , ∫ λ ∈ ( 0 , ∞ ) e − λ t η 0 ( λ ) ν ^ ( d λ ) = t ∫ x ∈ [ λ 1 , ∞ ) e − x t ( ∫ λ ∈ ( 0 , x ] η 0 ( λ ) ν ^ ( d λ ) ) d x ≤ D t ∫ x ∈ [ λ 1 , ∞ ) e − x t d x = D e − λ 1 t . (39)
H t ≥ β 0 λ 1 + β 2 m 2 R ‾ 2 , t − D e − λ 1 t ≥ β 0 λ 1 + e − λ 1 t ( β 2 m 2 R ‾ 2 , 0 e m t 2 − D ) ≥ β 0 λ 1 + e − λ 1 t ( β 2 m 2 R ‾ 2 , 0 − D ) ≥ β 0 λ 1 − e − λ 1 t ( D − β 2 m 2 R ‾ 2 , 0 ) + ≥ β 0 λ 1 − ( D − β 2 m 2 R ‾ 2 , 0 ) + . \begin{aligned}
H_t&\ge\beta_0\lambda_1+\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,t}}-De^{-\lambda_1t}\\
&\ge\beta_0\lambda_1+e^{-\lambda_1t}\left(\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}e^{\frac{mt}{2}}-D\right)\\
&\ge\beta_0\lambda_1+e^{-\lambda_1t}\left(\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}-D\right)\\
&\ge\beta_0\lambda_1-e^{-\lambda_1t}\left(D-\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}\right)_{+}\\
&\ge\beta_0\lambda_1-\left(D-\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}\right)_{+}.
\end{aligned} H t ≥ β 0 λ 1 + 2 β 2 m R 2 , t − D e − λ 1 t ≥ β 0 λ 1 + e − λ 1 t ( 2 β 2 m R 2 , 0 e 2 m t − D ) ≥ β 0 λ 1 + e − λ 1 t ( 2 β 2 m R 2 , 0 − D ) ≥ β 0 λ 1 − e − λ 1 t ( D − 2 β 2 m R 2 , 0 ) + ≥ β 0 λ 1 − ( D − 2 β 2 m R 2 , 0 ) + .
The first inequality follows from (38) and (39) ; the second follows from (34) , since β 2 m ≥ 0 \beta_2m\ge0 β 2 m ≥ 0 . The last three use e m t 2 ≥ 1 e^{\frac{mt}{2}}\ge1 e 2 m t ≥ 1 and 0 < e − λ 1 t ≤ 1 0<e^{-\lambda_1t}\le1 0 < e − λ 1 t ≤ 1 .
If m < 0 m<0 m < 0 , the dynamics (9) and the definition of R ‾ 2 \overline{R}_2 R 2 in (PDV) give, for almost every t ≥ 0 t\ge0 t ≥ 0 ,
R ‾ ˙ 2 , t + 2 λ 1 R ‾ 2 , t = g ˙ 2 ( t ) + 2 λ 1 g 2 ( t ) + K 2 ( t , t ) σ t 2 + ∫ 0 t ( ∂ t K 2 ( s , t ) + 2 λ 1 K 2 ( s , t ) ) σ s 2 d s ≥ g ˙ 2 ( t ) + 2 λ 1 g 2 ( t ) ≥ m g 2 ( t ) . \begin{aligned}
\dot{\overline{R}}_{2,t}+2\lambda_1\overline{R}_{2,t}
&=\dot{g}_2(t)+2\lambda_1g_2(t)+K_2(t,t)\sigma_t^2+\int_0^t\left(\partial_tK_2(s,t)+2\lambda_1K_2(s,t)\right)\sigma_s^2\,ds\\
&\ge\dot{g}_2(t)+2\lambda_1g_2(t)\\
&\ge mg_2(t).
\end{aligned} R ˙ 2 , t + 2 λ 1 R 2 , t = g ˙ 2 ( t ) + 2 λ 1 g 2 ( t ) + K 2 ( t , t ) σ t 2 + ∫ 0 t ( ∂ t K 2 ( s , t ) + 2 λ 1 K 2 ( s , t ) ) σ s 2 d s ≥ g ˙ 2 ( t ) + 2 λ 1 g 2 ( t ) ≥ m g 2 ( t ) .
The first inequality uses K 2 ( t , t ) ≥ 0 K_2(t,t)\ge0 K 2 ( t , t ) ≥ 0 from Assumption 2 and the additional kernel bound in Assumption 3 (ii); the second follows from (30) . Proposition 1, (1) , and Assumption 3 (ii) give R ‾ 2 , t ≥ g 2 ( t ) > 0 \overline{R}_{2,t}\ge g_2(t)>0 R 2 , t ≥ g 2 ( t ) > 0 and g 2 ( t ) ≤ g 2 ( 0 ) = R ‾ 2 , 0 g_2(t)\le g_2(0)=\overline{R}_{2,0} g 2 ( t ) ≤ g 2 ( 0 ) = R 2 , 0 . Since m < 0 m<0 m < 0 , we obtain
R ‾ ˙ 2 , t + 2 λ 1 R ‾ 2 , t 2 R ‾ 2 , t ≥ m g 2 ( t ) 2 R ‾ 2 , t ≥ m 2 g 2 ( t ) ≥ m 2 g 2 ( 0 ) = m 2 R ‾ 2 , 0 . \frac{\dot{\overline{R}}_{2,t}+2\lambda_1\overline{R}_{2,t}}{2\sqrt{\overline{R}_{2,t}}}
\ge\frac{mg_2(t)}{2\sqrt{\overline{R}_{2,t}}}
\ge\frac{m}{2}\sqrt{g_2(t)}
\ge\frac{m}{2}\sqrt{g_2(0)}
=\frac{m}{2}\sqrt{\overline{R}_{2,0}}. 2 R 2 , t R ˙ 2 , t + 2 λ 1 R 2 , t ≥ 2 R 2 , t m g 2 ( t ) ≥ 2 m g 2 ( t ) ≥ 2 m g 2 ( 0 ) = 2 m R 2 , 0 .
By (39) and D ≥ 0 D\ge0 D ≥ 0 from Notation 3, ∫ λ ∈ ( 0 , ∞ ) e − λ t η 0 ( λ ) ν ^ ( d λ ) ≤ D e − λ 1 t ≤ D \int_{\lambda\in(0,\infty)}e^{-\lambda t}\eta_0(\lambda)\,\hat{\nu}(d\lambda)\le De^{-\lambda_1t}\le D ∫ λ ∈ ( 0 , ∞ ) e − λ t η 0 ( λ ) ν ^ ( d λ ) ≤ D e − λ 1 t ≤ D . Hence
H t ≥ β 0 λ 1 + β 2 m 2 R ‾ 2 , 0 − D . H_t\ge\beta_0\lambda_1+\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}-D. H t ≥ β 0 λ 1 + 2 β 2 m R 2 , 0 − D .
Thus both cases give, for almost every t ≥ 0 t\ge0 t ≥ 0 ,
H t ≥ β 0 λ 1 − ( D − β 2 m 2 R ‾ 2 , 0 ) + ≥ 0 , (40) H_t\ge\beta_0\lambda_1-\left(D-\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}\right)_{+}\ge0, \tag*{(40)} H t ≥ β 0 λ 1 − ( D − 2 β 2 m R 2 , 0 ) + ≥ 0 , (40)
by (31) and β 0 λ 1 ≥ 0 \beta_0\lambda_1\ge0 β 0 λ 1 ≥ 0 .
Step 2: Grönwall’s lemma. Itô’s formula, the semimartingale dynamics (25) , and the dynamics (32) of Z Z Z give
d ( Z t − 1 ) = ( λ 1 , e f f + β 1 2 Λ 1 2 ) Z t − 1 d t − β 1 Λ 1 Z t − 1 d W t , d\left(Z_t^{-1}\right)=\left(\lambda_{1,\mathrm{eff}}+\beta_1^2\Lambda_1^2\right)Z_t^{-1}\,dt-\beta_1\Lambda_1Z_t^{-1}\,dW_t, d ( Z t − 1 ) = ( λ 1 , eff + β 1 2 Λ 1 2 ) Z t − 1 d t − β 1 Λ 1 Z t − 1 d W t ,
d [ σ , Z − 1 ] t = − β 1 2 Λ 1 2 σ t Z t − 1 d t . d[\sigma,Z^{-1}]_t=-\beta_1^2\Lambda_1^2\sigma_t Z_t^{-1}\,dt. d [ σ , Z − 1 ] t = − β 1 2 Λ 1 2 σ t Z t − 1 d t .
Hence, using again (25) together with (36) , we obtain
Z t − 1 σ t = σ 0 + ∫ 0 t Z s − 1 ( ( ( K 1 ′ ∗ R ) ′ ∗ σ ) ( s ) + H s + ( J ∗ H ) ( s ) ) d s . (41) Z_t^{-1}\sigma_t=\sigma_0+\int_0^t Z_s^{-1}\left(((K'_1*R)'\ast\sigma)(s)+H_s+(J*H)(s)\right)\,ds. \tag*{(41)} Z t − 1 σ t = σ 0 + ∫ 0 t Z s − 1 ( (( K 1 ′ ∗ R ) ′ ∗ σ ) ( s ) + H s + ( J ∗ H ) ( s ) ) d s . (41)
Since H , J ≥ 0 H,J\geq0 H , J ≥ 0 and σ 0 ≥ 0 \sigma_0\geq0 σ 0 ≥ 0 , Fubini’s theorem and the bounds (24) imply, for T > 0 T>0 T > 0 ,
∀ t ∈ [ 0 , T ] , σ t − ≤ ∫ s ∈ [ 0 , t ] Z t Z s ∫ u ∈ [ 0 , s ] ( K 1 ′ ∗ R ) ′ ( s − u ) σ u − d u d s ≤ ( sup ( u , v ) ∈ [ 0 , T ] 2 Z u Z v ) ∫ u ∈ [ 0 , t ] σ u − ( ∫ s ∈ [ u , t ] ( K 1 ′ ∗ R ) ′ ( s − u ) d s ) d u = ( sup ( u , v ) ∈ [ 0 , T ] 2 Z u Z v ) ∫ u ∈ [ 0 , t ] σ u − ( ∫ r ∈ [ 0 , t − u ] ( K 1 ′ ∗ R ) ′ ( r ) d r ) d u ≤ λ 1 , e f f ( sup ( u , v ) ∈ [ 0 , T ] 2 Z u Z v ) ∫ 0 t σ u − d u . \begin{aligned}
\forall t\in[0,T],\qquad\sigma_t^{-}\leq\int_{s\in[0,t]}\frac{Z_t}{Z_s}\int_{u\in[0,s]}(K'_1*R)'(s-u)\sigma_u^{-}\,du\,ds \\
&\leq\left(\sup_{(u,v)\in[0,T]^2}\frac{Z_u}{Z_v}\right)\int_{u\in[0,t]}\sigma_u^{-}\left(\int_{s\in[u,t]}(K'_1*R)'(s-u)\,ds\right)\,du \\
&=\left(\sup_{(u,v)\in[0,T]^2}\frac{Z_u}{Z_v}\right)\int_{u\in[0,t]}\sigma_u^{-}\left(\int_{r\in[0,t-u]}(K'_1*R)'(r)\,dr\right)\,du \\
&\leq\lambda_{1,\mathrm{eff}}\left(\sup_{(u,v)\in[0,T]^2}\frac{Z_u}{Z_v}\right)\int_0^t\sigma_u^{-}\,du.
\end{aligned} ∀ t ∈ [ 0 , T ] , σ t − ≤ ∫ s ∈ [ 0 , t ] Z s Z t ∫ u ∈ [ 0 , s ] ( K 1 ′ ∗ R ) ′ ( s − u ) σ u − d u d s ≤ ( ( u , v ) ∈ [ 0 , T ] 2 sup Z v Z u ) ∫ u ∈ [ 0 , t ] σ u − ( ∫ s ∈ [ u , t ] ( K 1 ′ ∗ R ) ′ ( s − u ) d s ) d u = ( ( u , v ) ∈ [ 0 , T ] 2 sup Z v Z u ) ∫ u ∈ [ 0 , t ] σ u − ( ∫ r ∈ [ 0 , t − u ] ( K 1 ′ ∗ R ) ′ ( r ) d r ) d u ≤ λ 1 , eff ( ( u , v ) ∈ [ 0 , T ] 2 sup Z v Z u ) ∫ 0 t σ u − d u .
The supremum is finite because Z Z Z is continuous and strictly positive by (32) . Grönwall’s lemma gives σ ≥ 0 \sigma\geq0 σ ≥ 0 . The identity (41) and the lower bound (40) yield
∀ t ≥ 0 , σ t ≥ Z t ( σ 0 + ∫ 0 t Z s − 1 H s d s ) ≥ Z t ( σ 0 + ( β 0 λ 1 − ( D − β 2 m 2 R ‾ 2 , 0 ) ) ∫ 0 t Z s − 1 d s ) ≥ 0 , \begin{aligned}
\forall t\geq0,\qquad\sigma_t\geq Z_t\left(\sigma_0+\int_0^t Z_s^{-1}H_s\,ds\right) \\
&\geq Z_t\left(\sigma_0+\left(\beta_0\lambda_1-\left(D-\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}}\right)\right)\int_0^t Z_s^{-1}\,ds\right)\geq0,
\end{aligned} ∀ t ≥ 0 , σ t ≥ Z t ( σ 0 + ∫ 0 t Z s − 1 H s d s ) ≥ Z t ( σ 0 + ( β 0 λ 1 − ( D − 2 β 2 m R 2 , 0 ) ) ∫ 0 t Z s − 1 d s ) ≥ 0 ,
which proves (33) . For every T > 0 T>0 T > 0 , the definition (11) and condition (8) give
C 2 , T ≤ R ( [ 0 , T ] ) sup r ∈ [ 0 , T ] sup u ∈ [ 0 , r ] K 2 ( u , r ) < ∞ . C_{2,T}\leq\mathcal{R}([0,T])\sup_{r\in[0,T]}\sup_{u\in[0,r]}K_2(u,r)<\infty. C 2 , T ≤ R ([ 0 , T ]) r ∈ [ 0 , T ] sup u ∈ [ 0 , r ] sup K 2 ( u , r ) < ∞.
Theorem 1 therefore implies that S S S is a bounded martingale on [ 0 , T ] [0,T] [ 0 , T ] under P \mathbb{P} P . □
Positivity and negativity with exponential kernels
Notation 4. For the inputs (5) , the corresponding factors are defined, for t ≥ 0 t\geq0 t ≥ 0 , 1 ≤ i ≤ M 1\leq i\leq M 1 ≤ i ≤ M and 1 ≤ j ≤ N 1\leq j\leq N 1 ≤ j ≤ N , by
R 1 , t ( λ 1 , i ) = e − λ 1 , i t R 1 , 0 ( λ 1 , i ) + ∫ 0 t λ 1 , i e − λ 1 , i ( t − s ) σ s d W s , R_{1,t}(\lambda_{1,i})=e^{-\lambda_{1,i}t}R_{1,0}(\lambda_{1,i})+\int_0^t\lambda_{1,i}e^{-\lambda_{1,i}(t-s)}\sigma_s\,dW_s, R 1 , t ( λ 1 , i ) = e − λ 1 , i t R 1 , 0 ( λ 1 , i ) + ∫ 0 t λ 1 , i e − λ 1 , i ( t − s ) σ s d W s ,
R 2 , t ( λ 2 , j ) : = e − λ 2 , j t R 2 , 0 ( λ 2 , j ) + ∫ 0 t λ 2 , j e − λ 2 , j ( t − s ) σ s 2 d s . (42) R_{2,t}(\lambda_{2,j}):=e^{-\lambda_{2,j}t}R_{2,0}(\lambda_{2,j})+\int_0^t\lambda_{2,j}e^{-\lambda_{2,j}(t-s)}\sigma_s^2\,ds. \tag*{(42)} R 2 , t ( λ 2 , j ) := e − λ 2 , j t R 2 , 0 ( λ 2 , j ) + ∫ 0 t λ 2 , j e − λ 2 , j ( t − s ) σ s 2 d s . (42)
so that R ‾ 1 , t = ∑ i = 1 M θ 1 , i R 1 , t ( λ 1 , i ) \overline{R}_{1,t}=\sum_{i=1}^{M}\theta_{1,i}R_{1,t}(\lambda_{1,i}) R 1 , t = ∑ i = 1 M θ 1 , i R 1 , t ( λ 1 , i ) and R ‾ 2 , t = ∑ j = 1 N θ 2 , j R 2 , t ( λ 2 , j ) \overline{R}_{2,t}=\sum_{j=1}^{N}\theta_{2,j}R_{2,t}(\lambda_{2,j}) R 2 , t = ∑ j = 1 N θ 2 , j R 2 , t ( λ 2 , j ) . For the return kernel K 1 K_1 K 1 , we have
ν = ∑ i = 1 M θ 1 , i δ λ 1 , i , Λ 1 = ∑ i = 1 M θ 1 , i λ 1 , i , λ 1 = λ 1 , M . \nu=\sum_{i=1}^{M}\theta_{1,i}\delta_{\lambda_{1,i}},\qquad\Lambda_1=\sum_{i=1}^{M}\theta_{1,i}\lambda_{1,i},\qquad\lambda_1=\lambda_{1,M}. ν = i = 1 ∑ M θ 1 , i δ λ 1 , i , Λ 1 = i = 1 ∑ M θ 1 , i λ 1 , i , λ 1 = λ 1 , M .
By Notation 3, ν ^ ( { λ 1 , M } ) = 0 \widehat{\nu}(\{\lambda_{1,M}\})=0 ν ({ λ 1 , M }) = 0 . We set
∀ i ∈ { 1 , … , M − 1 } , λ ^ i : = ν ^ ( { λ 1 , i } ) = θ 1 , i λ 1 , i ( λ 1 , i − λ 1 , M ) Λ 1 . \forall i\in\{1,\ldots,M-1\},\qquad\widehat{\lambda}_i:=\widehat{\nu}(\{\lambda_{1,i}\})=\frac{\theta_{1,i}\lambda_{1,i}(\lambda_{1,i}-\lambda_{1,M})}{\Lambda_1}. ∀ i ∈ { 1 , … , M − 1 } , λ i := ν ({ λ 1 , i }) = Λ 1 θ 1 , i λ 1 , i ( λ 1 , i − λ 1 , M ) .
The definitions give
ν ^ = ∑ i = 1 M − 1 λ ^ i δ λ 1 , i , λ 1 , e f f = λ 1 , M + ∑ i = 1 M − 1 λ ^ i , D = max 1 ≤ p ≤ M ∑ i = p M − 1 λ ^ i η 0 ( λ 1 , i ) . \widehat{\nu}=\sum_{i=1}^{M-1}\widehat{\lambda}_i\delta_{\lambda_{1,i}},\qquad\lambda_{1,\mathrm{eff}}=\lambda_{1,M}+\sum_{i=1}^{M-1}\widehat{\lambda}_i,\qquad D=\max_{1\leq p\leq M}\sum_{i=p}^{M-1}\widehat{\lambda}_i\eta_0(\lambda_{1,i}). ν = i = 1 ∑ M − 1 λ i δ λ 1 , i , λ 1 , eff = λ 1 , M + i = 1 ∑ M − 1 λ i , D = 1 ≤ p ≤ M max i = p ∑ M − 1 λ i η 0 ( λ 1 , i ) .
All sums over 1 ≤ i ≤ M − 1 1 \le i \le M-1 1 ≤ i ≤ M − 1 equal zero when M = 1 M=1 M = 1 ; in this case, ν ^ = 0 \hat{\nu}=0 ν ^ = 0 and D = 0 D=0 D = 0 . We write R 2 , t : = ( R 2 , t ( λ 2 , j ) ) 1 ≤ j ≤ N R_{2,t}:=(R_{2,t}(\lambda_{2,j}))_{1\le j\le N} R 2 , t := ( R 2 , t ( λ 2 , j ) ) 1 ≤ j ≤ N and
Λ 2 : = ∑ j = 1 N θ 2 , j λ 2 , j , λ 2 , e f f : = ∑ j = 1 N θ 2 , j λ 2 , j 2 Λ 2 , λ 2 ( r ) : = ∑ j = 1 N θ 2 , j λ 2 , j r j ∑ j = 1 N θ 2 , j r j . \Lambda_2:=\sum_{j=1}^{N}\theta_{2,j}\lambda_{2,j},\qquad\lambda_{2,\mathrm{eff}}:=\frac{\sum_{j=1}^{N}\theta_{2,j}\lambda_{2,j}^{2}}{\Lambda_2},\qquad\lambda_2(\mathbf{r}):=\frac{\sum_{j=1}^{N}\theta_{2,j}\lambda_{2,j}r_j}{\sum_{j=1}^{N}\theta_{2,j}r_j}. Λ 2 := j = 1 ∑ N θ 2 , j λ 2 , j , λ 2 , eff := Λ 2 ∑ j = 1 N θ 2 , j λ 2 , j 2 , λ 2 ( r ) := ∑ j = 1 N θ 2 , j r j ∑ j = 1 N θ 2 , j λ 2 , j r j .
for r = ( r 1 , … , r N ) ∈ R N \mathbf{r}=(r_1,\ldots,r_N)\in\mathbb{R}^{N} r = ( r 1 , … , r N ) ∈ R N such that ∑ j = 1 N θ 2 , j r j > 0 \sum_{j=1}^{N}\theta_{2,j}r_j>0 ∑ j = 1 N θ 2 , j r j > 0 .
For exponential kernels, Theorem 2 yields explicit conditions on the rates and initial factors.
Proposition 3. Suppose that K 1 K_1 K 1 and K 2 K_2 K 2 are respectively M M M - and N N N -exponential kernels, with deterministic initial factors and inputs as in Notation 4, with σ 0 ≥ 0 \sigma_0\ge0 σ 0 ≥ 0 . If
2 λ 1 , M ≥ λ 2 , e f f , (43) 2\lambda_{1,M}\ge\lambda_{2,\mathrm{eff}}, \tag*{(43)} 2 λ 1 , M ≥ λ 2 , eff , (43) max 1 ≤ p ≤ M ∑ i = p M − 1 λ ^ i η 0 ( λ 1 , i ) ≤ β 0 λ 1 , M + β 2 ( λ 1 , M − max { λ 2 ( R 2 , 0 ) , λ 2 , e f f } 2 ) R ‾ 2 , 0 , (44) \max_{1\le p\le M}\sum_{i=p}^{M-1}\hat{\lambda}_i\eta_0(\lambda_{1,i})\le\beta_0\lambda_{1,M}+\beta_2\left(\lambda_{1,M}-\frac{\max\{\lambda_2(R_{2,0}),\lambda_{2,\mathrm{eff}}\}}{2}\right)\sqrt{\overline{R}_{2,0}}, \tag*{(44)} 1 ≤ p ≤ M max i = p ∑ M − 1 λ ^ i η 0 ( λ 1 , i ) ≤ β 0 λ 1 , M + β 2 ( λ 1 , M − 2 max { λ 2 ( R 2 , 0 ) , λ 2 , eff } ) R 2 , 0 , (44) then P ( ∀ t ≥ 0 , σ t ≥ 0 ) = 1 \mathbb{P}(\forall t\ge0,\ \sigma_t\ge0)=1 P ( ∀ t ≥ 0 , σ t ≥ 0 ) = 1 , and S S S is a bounded martingale on every finite interval under P \mathbb{P} P .
Proof. The exponential kernels and initial inputs satisfy Assumptions 1 and 2, with g 2 > 0 g_2>0 g 2 > 0 by (6) . Proposition 1 therefore gives a pathwise unique continuous solution of (PDV). Definition 2.4 gives θ 2 , j , λ 2 , j > 0 \theta_{2,j},\lambda_{2,j}>0 θ 2 , j , λ 2 , j > 0 and R 2 , 0 ( λ 2 , j ) ≥ 0 R_{2,0}(\lambda_{2,j})\ge0 R 2 , 0 ( λ 2 , j ) ≥ 0 for every j ∈ { 1 , … , N } j\in\{1,\ldots,N\} j ∈ { 1 , … , N } . Together with (5) , this gives, for every t ≥ 0 t\ge0 t ≥ 0 ,
K 2 ( 0 , t ) ∂ t 2 K 2 ( 0 , t ) − ( ∂ t K 2 ( 0 , t ) ) 2 = 1 2 ∑ j , k = 1 N θ 2 , j θ 2 , k λ 2 , j λ 2 , k ( λ 2 , j − λ 2 , k ) 2 e − ( λ 2 , j + λ 2 , k ) t ≥ 0 , K_2(0,t)\partial_t^2K_2(0,t)-(\partial_tK_2(0,t))^2=\frac{1}{2}\sum_{j,k=1}^{N}\theta_{2,j}\theta_{2,k}\lambda_{2,j}\lambda_{2,k}(\lambda_{2,j}-\lambda_{2,k})^2e^{-(\lambda_{2,j}+\lambda_{2,k})t}\ge0, K 2 ( 0 , t ) ∂ t 2 K 2 ( 0 , t ) − ( ∂ t K 2 ( 0 , t ) ) 2 = 2 1 j , k = 1 ∑ N θ 2 , j θ 2 , k λ 2 , j λ 2 , k ( λ 2 , j − λ 2 , k ) 2 e − ( λ 2 , j + λ 2 , k ) t ≥ 0 ,
g 2 ( t ) g ¨ 2 ( t ) − g ˙ 2 ( t ) 2 = 1 2 ∑ j , k = 1 N θ 2 , j θ 2 , k R 2 , 0 ( λ 2 , j ) R 2 , 0 ( λ 2 , k ) ( λ 2 , j − λ 2 , k ) 2 e − ( λ 2 , j + λ 2 , k ) t ≥ 0. g_2(t)\ddot{g}_2(t)-\dot{g}_2(t)^2=\frac{1}{2}\sum_{j,k=1}^{N}\theta_{2,j}\theta_{2,k}R_{2,0}(\lambda_{2,j})R_{2,0}(\lambda_{2,k})(\lambda_{2,j}-\lambda_{2,k})^2e^{-(\lambda_{2,j}+\lambda_{2,k})t}\ge0. g 2 ( t ) g ¨ 2 ( t ) − g ˙ 2 ( t ) 2 = 2 1 j , k = 1 ∑ N θ 2 , j θ 2 , k R 2 , 0 ( λ 2 , j ) R 2 , 0 ( λ 2 , k ) ( λ 2 , j − λ 2 , k ) 2 e − ( λ 2 , j + λ 2 , k ) t ≥ 0.
Since K 2 ( 0 , t ) > 0 K_2(0,t)>0 K 2 ( 0 , t ) > 0 by Definition 2.4 and g 2 ( t ) > 0 g_2(t)>0 g 2 ( t ) > 0 by (6) , the following quantities are well defined and satisfy
∀ t ≥ 0 , d 2 d t 2 log K 2 ( 0 , t ) ≥ 0 , d 2 d t 2 log g 2 ( t ) ≥ 0. \forall t\ge0,\qquad\frac{d^2}{dt^2}\log K_2(0,t)\ge0,\qquad\frac{d^2}{dt^2}\log g_2(t)\ge0. ∀ t ≥ 0 , d t 2 d 2 log K 2 ( 0 , t ) ≥ 0 , d t 2 d 2 log g 2 ( t ) ≥ 0.
Thus K 2 ( 0 , ⋅ ) K_2(0,\cdot) K 2 ( 0 , ⋅ ) and g 2 g_2 g 2 are log-convex. Their logarithmic derivatives are nondecreasing, so evaluation at zero gives
∀ t ≥ 0 , − ∂ t K 2 ( 0 , t ) K 2 ( 0 , t ) ≤ − ∂ t K 2 ( 0 , 0 ) K 2 ( 0 , 0 ) = λ 2 , e f f , \forall t\ge0,\qquad-\frac{\partial_tK_2(0,t)}{K_2(0,t)}\le-\frac{\partial_tK_2(0,0)}{K_2(0,0)}=\lambda_{2,\mathrm{eff}}, ∀ t ≥ 0 , − K 2 ( 0 , t ) ∂ t K 2 ( 0 , t ) ≤ − K 2 ( 0 , 0 ) ∂ t K 2 ( 0 , 0 ) = λ 2 , eff ,
∀ t ≥ 0 , − g ˙ 2 ( t ) g 2 ( t ) ≤ − g ˙ 2 ( 0 ) g 2 ( 0 ) = λ 2 ( R 2 , 0 ) . (45) \forall t\ge0,\qquad-\frac{\dot{g}_2(t)}{g_2(t)}\le-\frac{\dot{g}_2(0)}{g_2(0)}=\lambda_2(R_{2,0}). \tag*{(45)} ∀ t ≥ 0 , − g 2 ( t ) g ˙ 2 ( t ) ≤ − g 2 ( 0 ) g ˙ 2 ( 0 ) = λ 2 ( R 2 , 0 ) . (45)
The logarithmic-derivative bounds (45) are attained at zero. Hence the constant in Assumption 3 is m = 2 λ 1 , M − max { λ 2 ( R 2 , 0 ) , λ 2 , e f f } m=2\lambda_{1,M}-\max\{\lambda_2(R_{2,0}),\lambda_{2,\mathrm{eff}}\} m = 2 λ 1 , M − max { λ 2 ( R 2 , 0 ) , λ 2 , eff } .
If m ≥ 0 m\ge0 m ≥ 0 , the first alternative in Assumption 3 holds.
If m < 0 m<0 m < 0 , (43) and the first bound in (45) , using K 2 ( s , t ) = K 2 ( 0 , t − s ) K_2(s,t)=K_2(0,t-s) K 2 ( s , t ) = K 2 ( 0 , t − s ) , give, for every 0 ≤ s ≤ t 0\le s\le t 0 ≤ s ≤ t , ∂ t K 2 ( s , t ) + 2 λ 1 , M K 2 ( s , t ) ≥ 0 \partial_tK_2(s,t)+2\lambda_{1,M}K_2(s,t)\ge0 ∂ t K 2 ( s , t ) + 2 λ 1 , M K 2 ( s , t ) ≥ 0 . Definition 2.4 and (5) give g 2 ( t ) ≤ g 2 ( 0 ) g_2(t)\le g_2(0) g 2 ( t ) ≤ g 2 ( 0 ) .
Finally, (44) is equivalent to D ≤ β 0 λ 1 , M + β 2 m 2 R ‾ 2 , 0 D\le\beta_0\lambda_{1,M}+\frac{\beta_2m}{2}\sqrt{\overline{R}_{2,0}} D ≤ β 0 λ 1 , M + 2 β 2 m R 2 , 0 . Thus Assumption 3 holds, and Theorem 2 applies.
Remark 3.12. For an N N N -exponential kernel K 2 K_2 K 2 , suppose that, for some Δ ∈ ( 0 , ∞ ] \Delta\in(0,\infty] Δ ∈ ( 0 , ∞ ] and a deterministic measurable history ( σ s ) − Δ < s < 0 (\sigma_s)_{-\Delta<s<0} ( σ s ) − Δ < s < 0 ,
∀ j ∈ { 1 , … , N } , R 2 , 0 ( λ 2 , j ) = λ 2 , j ∫ − Δ 0 e λ 2 , j s σ s 2 d s , ∫ − Δ 0 e λ 2 , N s σ s 2 d s ∈ ( 0 , ∞ ) . \forall j \in\{1,\ldots,N\},\qquad R_{2,0}(\lambda_{2,j})=\lambda_{2,j}\int_{-\Delta}^{0}e^{\lambda_{2,j}s}\sigma_s^2\,\mathrm{d}s,\qquad\int_{-\Delta}^{0}e^{\lambda_{2,N}s}\sigma_s^2\,\mathrm{d}s\in(0,\infty). ∀ j ∈ { 1 , … , N } , R 2 , 0 ( λ 2 , j ) = λ 2 , j ∫ − Δ 0 e λ 2 , j s σ s 2 d s , ∫ − Δ 0 e λ 2 , N s σ s 2 d s ∈ ( 0 , ∞ ) .
These assumptions give R 2 , 0 ( λ 2 , j ) ∈ ( 0 , ∞ ) R_{2,0}(\lambda_{2,j})\in(0,\infty) R 2 , 0 ( λ 2 , j ) ∈ ( 0 , ∞ ) for every j ∈ { 1 , … , N } j\in\{1,\ldots,N\} j ∈ { 1 , … , N } , hence g 2 ( 0 ) > 0 g_2(0)>0 g 2 ( 0 ) > 0 by (5) . Thus,
λ 2 ( R 2 , 0 ) ≤ λ 2 , e f f . (46) \lambda_2(\mathbf{R}_{2,0})\leq\lambda_{2,\mathrm{eff}}. \tag*{(46)} λ 2 ( R 2 , 0 ) ≤ λ 2 , eff . (46)
Consequently, condition (44) simplifies to
max 1 ≤ p ≤ M ∑ i = p M − 1 λ ^ i η 0 ( λ 1 , i ) ≤ β 0 λ 1 , M + β 2 ( λ 1 , M − λ 2 , e f f 2 ) R ‾ 2 , 0 , \max_{1\leq p\leq M}\sum_{i=p}^{M-1}\hat{\lambda}_i\eta_0(\lambda_{1,i})\leq\beta_0\lambda_{1,M}+\beta_2\left(\lambda_{1,M}-\frac{\lambda_{2,\mathrm{eff}}}{2}\right)\sqrt{\overline{R}_{2,0}}, 1 ≤ p ≤ M max i = p ∑ M − 1 λ ^ i η 0 ( λ 1 , i ) ≤ β 0 λ 1 , M + β 2 ( λ 1 , M − 2 λ 2 , eff ) R 2 , 0 ,
since max { λ 2 ( R 2 , 0 ) , λ 2 , e f f } = λ 2 , e f f \max\{\lambda_2(\mathbf{R}_{2,0}),\lambda_{2,\mathrm{eff}}\}=\lambda_{2,\mathrm{eff}} max { λ 2 ( R 2 , 0 ) , λ 2 , eff } = λ 2 , eff . By (5) , g 2 ( t ) = ∫ − Δ 0 K 2 ( 0 , t − s ) σ s 2 d s g_2(t)=\int_{-\Delta}^{0}K_2(0,t-s)\sigma_s^2\,\mathrm{d}s g 2 ( t ) = ∫ − Δ 0 K 2 ( 0 , t − s ) σ s 2 d s . Differentiating at zero and applying the first bound in (45) at t = − s ≥ 0 t=-s\geq0 t = − s ≥ 0 give
λ 2 ( R 2 , 0 ) = − g ˙ 2 ( 0 ) g 2 ( 0 ) = ∫ − Δ 0 − ∂ t K 2 ( 0 , − s ) σ s 2 d s ∫ − Δ 0 K 2 ( 0 , − s ) σ s 2 d s ≤ λ 2 , e f f ∫ − Δ 0 K 2 ( 0 , − s ) σ s 2 d s ∫ − Δ 0 K 2 ( 0 , − s ) σ s 2 d s = λ 2 , e f f . \lambda_2(\mathbf{R}_{2,0})=-\frac{\dot{g}_2(0)}{g_2(0)}=\frac{\int_{-\Delta}^{0}-\partial_tK_2(0,-s)\sigma_s^2\,\mathrm{d}s}{\int_{-\Delta}^{0}K_2(0,-s)\sigma_s^2\,\mathrm{d}s}\leq\lambda_{2,\mathrm{eff}}\frac{\int_{-\Delta}^{0}K_2(0,-s)\sigma_s^2\,\mathrm{d}s}{\int_{-\Delta}^{0}K_2(0,-s)\sigma_s^2\,\mathrm{d}s}=\lambda_{2,\mathrm{eff}}. λ 2 ( R 2 , 0 ) = − g 2 ( 0 ) g ˙ 2 ( 0 ) = ∫ − Δ 0 K 2 ( 0 , − s ) σ s 2 d s ∫ − Δ 0 − ∂ t K 2 ( 0 , − s ) σ s 2 d s ≤ λ 2 , eff ∫ − Δ 0 K 2 ( 0 , − s ) σ s 2 d s ∫ − Δ 0 K 2 ( 0 , − s ) σ s 2 d s = λ 2 , eff .
The next result gives conditions under which volatility becomes negative with positive probability, and hence necessary conditions for nonnegative volatility.
Theorem 3. Suppose that K 1 K_1 K 1 and K 2 K_2 K 2 are respectively M M M - and N N N -exponential kernels. We use Notations 3 and 4. Assume that the deterministic initial factors and inputs are as in (5) , with σ 0 > 0 \sigma_0>0 σ 0 > 0 . If at least one of the following conditions holds:
λ 1 , e f f < λ 2 , e f f 2 , (47) \lambda_{1,\mathrm{eff}}<\frac{\lambda_{2,\mathrm{eff}}}{2}, \tag*{(47)} λ 1 , eff < 2 λ 2 , eff , (47) ∑ i = 1 M − 1 λ ^ i η 0 ( λ 1 , i ) > β 0 λ 1 , M + β 2 ( λ 1 , M − λ 2 ( R 2 , 0 ) 2 ) R ‾ 2 , 0 . (48) \sum_{i=1}^{M-1}\hat{\lambda}_i\eta_0(\lambda_{1,i})>\beta_0\lambda_{1,M}+\beta_2\left(\lambda_{1,M}-\frac{\lambda_2(\mathbf{R}_{2,0})}{2}\right)\sqrt{\overline{R}_{2,0}}. \tag*{(48)} i = 1 ∑ M − 1 λ ^ i η 0 ( λ 1 , i ) > β 0 λ 1 , M + β 2 ( λ 1 , M − 2 λ 2 ( R 2 , 0 ) ) R 2 , 0 . (48) then we have
∀ t > 0 , P ( σ t < 0 ) > 0. \forall t>0,\qquad\mathbb{P}(\sigma_t<0)>0. ∀ t > 0 , P ( σ t < 0 ) > 0. Consequently, for every t 0 > 0 t_0>0 t 0 > 0 , ( S t ) t ∈ [ 0 , t 0 ] (S_t)_{t\in[0,t_0]} ( S t ) t ∈ [ 0 , t 0 ] is a strict local martingale under P \mathbb{P} P .
For M = 1 M=1 M = 1 , ν ^ = 0 \hat{\nu}=0 ν ^ = 0 , D = 0 D=0 D = 0 , and λ 1 , e f f = λ 1 , 1 \lambda_{1,\mathrm{eff}}=\lambda_{1,1} λ 1 , eff = λ 1 , 1 . Combining Proposition 3, Theorem 3, and Remark 3.12 yields the following corollary.
Corollary 1. Suppose M = 1 M=1 M = 1 and K 2 K_2 K 2 is an N N N -exponential kernel, with deterministic initial factors and inputs as in Notation 4, with σ 0 > 0 \sigma_0>0 σ 0 > 0 .
(i) The following equivalences hold:
P ( ∀ t ≥ 0 , σ t ≥ 0 ) = 1 ⟺ { 2 λ 1 , 1 ≥ λ 2 , e f f , 0 ≤ β 0 λ 1 , 1 + β 2 ( λ 1 , 1 − λ 2 ( R 2 , 0 ) 2 ) R ‾ 2 , 0 \mathbb{P}(\forall t\geq0,\ \sigma_t\geq0)=1\quad\Longleftrightarrow\quad
\begin{cases}
2\lambda_{1,1}\geq\lambda_{2,\mathrm{eff}},\\
0\leq\beta_0\lambda_{1,1}+\beta_2\left(\lambda_{1,1}-\dfrac{\lambda_2(\mathbf{R}_{2,0})}{2}\right)\sqrt{\overline{R}_{2,0}}
\end{cases} P ( ∀ t ≥ 0 , σ t ≥ 0 ) = 1 ⟺ ⎩ ⎨ ⎧ 2 λ 1 , 1 ≥ λ 2 , eff , 0 ≤ β 0 λ 1 , 1 + β 2 ( λ 1 , 1 − 2 λ 2 ( R 2 , 0 ) ) R 2 , 0 ⟺ S is a martingale on every finite interval under P . \Longleftrightarrow\quad S\text{ is a martingale on every finite interval under }\mathbb{P}. ⟺ S is a martingale on every finite interval under P . (ii) If g 2 g_2 g 2 is initialized from a deterministic past as in Remark 3.12, then
P ( ∀ t ≥ 0 , σ t ≥ 0 ) = 1 ⟺ 2 λ 1 , 1 ≥ λ 2 , e f f \mathbb{P}(\forall t\geq0,\ \sigma_t\geq0)=1\quad\Longleftrightarrow\quad2\lambda_{1,1}\geq\lambda_{2,\mathrm{eff}} P ( ∀ t ≥ 0 , σ t ≥ 0 ) = 1 ⟺ 2 λ 1 , 1 ≥ λ 2 , eff ⟺ S is a bounded martingale on every finite interval under P . \Longleftrightarrow\quad S\text{ is a bounded martingale on every finite interval under }\mathbb{P}. ⟺ S is a bounded martingale on every finite interval under P .
If ∑ k = 1 n r k 2 > 0 \sum_{k=1}^{n} r_k^2 > 0 ∑ k = 1 n r k 2 > 0 , this history satisfies the integrability assumption in Remark 3.12 and yields R 2 , 0 ( λ 2 , j ) = λ 2 , j ∫ − Δ 0 e λ 2 , j s σ s 2 d s R_{2,0}(\lambda_{2,j})=\lambda_{2,j}\int_{-\Delta}^{0}e^{\lambda_{2,j}s}\sigma_s^2\,\mathrm{d}s R 2 , 0 ( λ 2 , j ) = λ 2 , j ∫ − Δ 0 e λ 2 , j s σ s 2 d s , hence (46) .
Example 3.15. We use the following parameters and initial factors, with the parameters based on the market-smile calibration in [17 Table 8] :
( β 0 , β 1 , β 2 ) = ( 0.026 , − 0.138 , 0.69 ) , (\beta_0,\beta_1,\beta_2)=(0.026,-0.138,0.69), ( β 0 , β 1 , β 2 ) = ( 0.026 , − 0.138 , 0.69 ) , ( λ 1 , 1 , λ 1 , 2 , θ 1 , 1 , θ 1 , 2 ) = ( 62.11 , 32.25 , 0.77 , 0.23 ) , (\lambda_{1,1},\lambda_{1,2},\theta_{1,1},\theta_{1,2})=(62.11,32.25,0.77,0.23), ( λ 1 , 1 , λ 1 , 2 , θ 1 , 1 , θ 1 , 2 ) = ( 62.11 , 32.25 , 0.77 , 0.23 ) , ( λ 2 , 1 , λ 2 , 2 , θ 2 , 1 , θ 2 , 2 ) = ( 9.57 , 3.51 , 0.01 , 0.99 ) , (\lambda_{2,1},\lambda_{2,2},\theta_{2,1},\theta_{2,2})=(9.57,3.51,0.01,0.99), ( λ 2 , 1 , λ 2 , 2 , θ 2 , 1 , θ 2 , 2 ) = ( 9.57 , 3.51 , 0.01 , 0.99 ) , ( R 1 , 0 ( λ 1 , 1 ) , R 1 , 0 ( λ 1 , 2 ) ) = ( 0.2988 , 0.2397 ) , (R_{1,0}(\lambda_{1,1}),R_{1,0}(\lambda_{1,2}))=(0.2988,0.2397), ( R 1 , 0 ( λ 1 , 1 ) , R 1 , 0 ( λ 1 , 2 )) = ( 0.2988 , 0.2397 ) , ( R 2 , 0 ( λ 2 , 1 ) , R 2 , 0 ( λ 2 , 2 ) ) = ( 0.016 , 0.02 ) . (R_{2,0}(\lambda_{2,1}),R_{2,0}(\lambda_{2,2}))=(0.016,0.02). ( R 2 , 0 ( λ 2 , 1 ) , R 2 , 0 ( λ 2 , 2 )) = ( 0.016 , 0.02 ) . These parameters and initial factors, together with (1) , give
R ˉ 2 , 0 = 0.01996 > 0 , σ 0 ≃ 0.08412 > 0 , λ 2 ( R 2 , 0 ) ≃ 3.55858 < λ 2 , e f f ≃ 3.67242. \bar{R}_{2,0}=0.01996>0,\qquad\sigma_0\simeq0.08412>0,\qquad\lambda_2(\mathbf{R}_{2,0})\simeq3.55858<\lambda_{2,\mathrm{eff}}\simeq3.67242. R ˉ 2 , 0 = 0.01996 > 0 , σ 0 ≃ 0.08412 > 0 , λ 2 ( R 2 , 0 ) ≃ 3.55858 < λ 2 , eff ≃ 3.67242. Moreover, η 0 ( λ 1 , 1 ) ≃ − 0.12080 < 0 \eta_0(\lambda_{1,1})\simeq-0.12080<0 η 0 ( λ 1 , 1 ) ≃ − 0.12080 < 0 implies D = 0 D=0 D = 0 by Notation 4. The sufficient conditions in the second row of Table 2 are satisfied, since
λ 1 , 2 = 32.25 > λ 2 , e f f 2 ≃ 1.83621 , \lambda_{1,2}=32.25>\frac{\lambda_{2,\mathrm{eff}}}{2}\simeq1.83621, λ 1 , 2 = 32.25 > 2 λ 2 , eff ≃ 1.83621 , max { 0 , λ ^ 1 η 0 ( λ 1 , 1 ) } = 0 < β 0 λ 1 , 2 + β 2 ( λ 1 , 2 − λ 2 , e f f 2 ) R ˉ 2 , 0 ≃ 3.80333. \max\{0,\widehat{\lambda}_1\eta_0(\lambda_{1,1})\}=0<\beta_0\lambda_{1,2}+\beta_2\left(\lambda_{1,2}-\frac{\lambda_{2,\mathrm{eff}}}{2}\right)\sqrt{\bar{R}_{2,0}}\simeq3.80333. max { 0 , λ 1 η 0 ( λ 1 , 1 )} = 0 < β 0 λ 1 , 2 + β 2 ( λ 1 , 2 − 2 λ 2 , eff ) R ˉ 2 , 0 ≃ 3.80333. Thus Theorem 2 applies with m = 2 λ 1 , 2 − λ 2 , e f f ≃ 60.82758 > 0 m=2\lambda_{1,2}-\lambda_{2,\mathrm{eff}}\simeq60.82758>0 m = 2 λ 1 , 2 − λ 2 , eff ≃ 60.82758 > 0 and gives σ t > 0 \sigma_t>0 σ t > 0 for every t ≥ 0 t\geq0 t ≥ 0 , almost surely. With S 0 = 1 S_0=1 S 0 = 1 and T = 16 252 T=\frac{16}{252} T = 252 16 , bound (12) gives B T ≃ 1.40630 B_T\simeq1.40630 B T ≃ 1.40630 . With 2 × 10 6 2\times10^6 2 × 1 0 6 Monte Carlo paths and a time step of 1 25200 \frac{1}{25200} 25200 1 years, the call-price estimate first vanished at the tested strike K = 1.05 K=1.05 K = 1.05 : no simulated terminal value exceeded this strike.
For the proof of Theorem 3, K 1 K_1 K 1 and K 2 K_2 K 2 are respectively M M M - and N N N -exponential kernels as in Definition 2.4 , with M , N ≥ 1 M,N\geq1 M , N ≥ 1 and σ 0 > 0 \sigma_0>0 σ 0 > 0 . We define the processes
∀ t ≥ 0 , ∀ i ∈ { 1 , … , M − 1 } , z i , t : = β 1 ( Λ 1 λ 1 , i R 1 , t ( λ 1 , i ) − R ˉ 1 , t ) . (49) \forall t\geq0,\quad\forall i\in\{1,\ldots,M-1\},\qquad z_{i,t}:=\beta_1\left(\frac{\Lambda_1}{\lambda_{1,i}}R_{1,t}(\lambda_{1,i})-\bar{R}_{1,t}\right). \tag*{(49)} ∀ t ≥ 0 , ∀ i ∈ { 1 , … , M − 1 } , z i , t := β 1 ( λ 1 , i Λ 1 R 1 , t ( λ 1 , i ) − R ˉ 1 , t ) . (49)
By (42) and (49) , they have finite variation, since their Brownian coefficients are
β 1 ( Λ 1 λ 1 , i λ 1 , i − Λ 1 ) σ t = 0. \beta_1\left(\frac{\Lambda_1}{\lambda_{1,i}}\lambda_{1,i}-\Lambda_1\right)\sigma_t=0. β 1 ( λ 1 , i Λ 1 λ 1 , i − Λ 1 ) σ t = 0.
We write z t : = ( z i , t ) 1 ≤ i ≤ M − 1 z_t:=(z_{i,t})_{1\leq i\leq M-1} z t := ( z i , t ) 1 ≤ i ≤ M − 1 . By (42) and Notation 4, the drift of β 1 R ˉ 1 , t \beta_1\bar{R}_{1,t} β 1 R ˉ 1 , t is
− β 1 ∑ i = 1 M θ 1 , i λ 1 , i R 1 , t ( λ 1 , i ) = − λ 1 , M β 1 R ˉ 1 , t − ∑ i = 1 M − 1 λ ^ i ( z i , t + β 1 R ˉ 1 , t ) = − λ 1 , e f f ( σ t − β 0 − β 2 R ˉ 2 , t ) − ∑ i = 1 M − 1 λ ^ i z i , t . (50) \begin{aligned}
-\beta_1\sum_{i=1}^{M}\theta_{1,i}\lambda_{1,i}R_{1,t}(\lambda_{1,i})=-\lambda_{1,M}\beta_1\bar{R}_{1,t}-\sum_{i=1}^{M-1}\widehat{\lambda}_i\left(z_{i,t}+\beta_1\bar{R}_{1,t}\right) \\
&=-\lambda_{1,\mathrm{eff}}\left(\sigma_t-\beta_0-\beta_2\sqrt{\bar{R}_{2,t}}\right)-\sum_{i=1}^{M-1}\widehat{\lambda}_iz_{i,t}. \tag*{(50)}
\end{aligned} − β 1 i = 1 ∑ M θ 1 , i λ 1 , i R 1 , t ( λ 1 , i ) = − λ 1 , M β 1 R ˉ 1 , t − i = 1 ∑ M − 1 λ i ( z i , t + β 1 R ˉ 1 , t ) = − λ 1 , eff ( σ t − β 0 − β 2 R ˉ 2 , t ) − i = 1 ∑ M − 1 λ i z i , t . (50)
Together with the factor dynamics (42) and R ˉ 2 , t ≥ g 2 ( t ) > 0 \bar{R}_{2,t}\geq g_2(t)>0 R ˉ 2 , t ≥ g 2 ( t ) > 0 from (PDV) and (6) , the chain rule gives
d σ t = μ σ ( σ t , z t , R 2 , t ) d t + β 1 Λ 1 σ t d W t , (51) \mathrm{d}\sigma_t=\mu_\sigma(\sigma_t,z_t,\mathbf{R}_{2,t})\,\mathrm{d}t+\beta_1\Lambda_1\sigma_t\,\mathrm{d}W_t, \tag*{(51)} d σ t = μ σ ( σ t , z t , R 2 , t ) d t + β 1 Λ 1 σ t d W t , (51)
∀ i ∈ { 1 , … , M − 1 } , z ˙ i , t = − λ 1 , i z i , t + ∑ k = 1 M − 1 λ ^ k z k , t + ( λ 1 , e f f − λ 1 , i ) ( σ t − β 0 − β 2 R ˉ 2 , t ) . (52) \forall i\in\{1,\ldots,M-1\},\qquad\dot{z}_{i,t}=-\lambda_{1,i}z_{i,t}+\sum_{k=1}^{M-1}\widehat{\lambda}_kz_{k,t}+(\lambda_{1,\mathrm{eff}}-\lambda_{1,i})(\sigma_t-\beta_0-\beta_2\sqrt{\bar{R}_{2,t}}). \tag*{(52)} ∀ i ∈ { 1 , … , M − 1 } , z ˙ i , t = − λ 1 , i z i , t + k = 1 ∑ M − 1 λ k z k , t + ( λ 1 , eff − λ 1 , i ) ( σ t − β 0 − β 2 R ˉ 2 , t ) . (52)
∀ j ∈ { 1 , … , N } , R ˙ 2 , t ( λ 2 , j ) = λ 2 , j ( σ t 2 − R 2 , t ( λ 2 , j ) ) , (53) \forall j \in\{1,\ldots,N\},\quad\dot{R}_{2,t}(\lambda_{2,j})=\lambda_{2,j}\left(\sigma_t^2-R_{2,t}(\lambda_{2,j})\right), \tag*{(53)} ∀ j ∈ { 1 , … , N } , R ˙ 2 , t ( λ 2 , j ) = λ 2 , j ( σ t 2 − R 2 , t ( λ 2 , j ) ) , (53)
where we define the drift coefficient by
μ σ ( σ , z , r ) : = − λ 1 , e f f σ + λ 1 , e f f β 0 − ∑ i = 1 M − 1 λ ^ i z i + β 2 ( λ 1 , e f f − λ 2 ( r ) 2 ) ∑ j = 1 N θ 2 , j r j + β 2 Λ 2 σ 2 2 ∑ j = 1 N θ 2 , j r j . (54) \mu_\sigma(\sigma,z,\mathbf{r}) := -\lambda_{1,\mathrm{eff}}\sigma+\lambda_{1,\mathrm{eff}}\beta_0-\sum_{i=1}^{M-1}\hat{\lambda}_i z_i+\beta_2\left(\lambda_{1,\mathrm{eff}}-\frac{\lambda_2(r)}{2}\right)\sqrt{\sum_{j=1}^{N}\theta_{2,j}r_j+\frac{\beta_2\Lambda_2\sigma^2}{2\sqrt{\sum_{j=1}^{N}\theta_{2,j}r_j}}}. \tag*{(54)} μ σ ( σ , z , r ) := − λ 1 , eff σ + λ 1 , eff β 0 − i = 1 ∑ M − 1 λ ^ i z i + β 2 ( λ 1 , eff − 2 λ 2 ( r ) ) j = 1 ∑ N θ 2 , j r j + 2 ∑ j = 1 N θ 2 , j r j β 2 Λ 2 σ 2 . (54)
This coefficient is well defined and smooth for σ ∈ R \sigma\in\mathbb{R} σ ∈ R , z ∈ R M − 1 z\in\mathbb{R}^{M-1} z ∈ R M − 1 , and r ∈ R N \mathbf{r}\in\mathbb{R}^{N} r ∈ R N with ∑ j = 1 N θ 2 , j r j > 0 \sum_{j=1}^{N}\theta_{2,j}r_j>0 ∑ j = 1 N θ 2 , j r j > 0 . For i ∈ { 1 , … , M − 1 } i\in\{1,\ldots,M-1\} i ∈ { 1 , … , M − 1 } , the drift in (52) follows from (49) and (50) :
z ˙ i , t = − β 1 Λ 1 R 1 , t ( λ 1 , i ) + β 1 ∑ k = 1 M θ 1 , k λ 1 , k R 1 , t ( λ 1 , k ) = − λ 1 , i z i , t + ( λ 1 , e f f − λ 1 , i ) β 1 R ‾ 1 , t + ∑ k = 1 M − 1 λ ^ k z k , t . \begin{aligned}
\dot{z}_{i,t}=-\beta_1\Lambda_1R_{1,t}(\lambda_{1,i})+\beta_1\sum_{k=1}^{M}\theta_{1,k}\lambda_{1,k}R_{1,t}(\lambda_{1,k}) \\
&=-\lambda_{1,i}z_{i,t}+(\lambda_{1,\mathrm{eff}}-\lambda_{1,i})\beta_1\overline{R}_{1,t}+\sum_{k=1}^{M-1}\hat{\lambda}_k z_{k,t}.
\end{aligned} z ˙ i , t = − β 1 Λ 1 R 1 , t ( λ 1 , i ) + β 1 k = 1 ∑ M θ 1 , k λ 1 , k R 1 , t ( λ 1 , k ) = − λ 1 , i z i , t + ( λ 1 , eff − λ 1 , i ) β 1 R 1 , t + k = 1 ∑ M − 1 λ ^ k z k , t .
using β 1 R ‾ 1 , t = σ t − β 0 − β 2 R ‾ 2 , t \beta_1\overline{R}_{1,t}=\sigma_t-\beta_0-\beta_2\sqrt{\overline{R}_{2,t}} β 1 R 1 , t = σ t − β 0 − β 2 R 2 , t . When M = 1 M=1 M = 1 , z z z and all sums over 1 ≤ i ≤ M − 1 1\leq i\leq M-1 1 ≤ i ≤ M − 1 are omitted. For T > 0 T>0 T > 0 and a deterministic control u ∈ L 2 ( [ 0 , T ] ) u\in L^2([0,T]) u ∈ L 2 ([ 0 , T ]) , we denote by ( σ u , z u , R 2 u ) (\sigma^u,z^u,\mathbf{R}_2^u) ( σ u , z u , R 2 u ) the solution of (3.42) on its interval of existence, with d W t dW_t d W t replaced by u t d t u_t\,dt u t d t and the same initial state. We write
R 2 , t u : = ( R 2 , t u ( λ 2 , j ) ) 1 ≤ j ≤ N , R ‾ 2 , t u : = ∑ j = 1 N θ 2 , j R 2 , t u ( λ 2 , j ) . \mathbf{R}_{2,t}^{u}:=(R_{2,t}^{u}(\lambda_{2,j}))_{1\leq j\leq N},\qquad\overline{R}_{2,t}^{u}:=\sum_{j=1}^{N}\theta_{2,j}R_{2,t}^{u}(\lambda_{2,j}). R 2 , t u := ( R 2 , t u ( λ 2 , j ) ) 1 ≤ j ≤ N , R 2 , t u := j = 1 ∑ N θ 2 , j R 2 , t u ( λ 2 , j ) .
For h > 0 h>0 h > 0 and a continuous positive path s h s^h s h on [ 0 , h ] [0,h] [ 0 , h ] with s 0 h = σ 0 s_0^h=\sigma_0 s 0 h = σ 0 , we define the functions z ~ \widetilde{z} z and R ~ 2 : = ( R ~ 2 ( λ 2 , j ) ) 1 ≤ j ≤ N \widetilde{\mathbf{R}}_2:=(\widetilde{R}_2(\lambda_{2,j}))_{1\leq j\leq N} R 2 := ( R 2 ( λ 2 , j ) ) 1 ≤ j ≤ N as the solutions of (52) and (53) with σ = s h \sigma=s^h σ = s h and initial values z 0 z_0 z 0 and R 2 , 0 \mathbf{R}_{2,0} R 2 , 0 .
Remark 3.16. In (52) , the Brownian terms cancel and the drift contains σ \sigma σ ; in (53) , the variance-factor drift contains σ 2 \sigma^2 σ 2 . If a volatility path has size h − 1 2 h^{-\frac{1}{2}} h − 2 1 on [ 0 , h ] [0,h] [ 0 , h ] , these terms give, up to constants independent of h h h ,
∣ z ~ i , h − z i , 0 ∣ ≈ h × h − 1 2 = h , R ~ 2 , h ( λ 2 , j ) − R 2 , 0 ( λ 2 , j ) ≈ h × ( h − 1 2 ) 2 = 1. |\widetilde{z}_{i,h}-z_{i,0}|\approx h\times h^{-\frac{1}{2}}=\sqrt{h},\qquad\widetilde{R}_{2,h}(\lambda_{2,j})-R_{2,0}(\lambda_{2,j})\approx h\times\left(h^{-\frac{1}{2}}\right)^2=1. ∣ z i , h − z i , 0 ∣ ≈ h × h − 2 1 = h , R 2 , h ( λ 2 , j ) − R 2 , 0 ( λ 2 , j ) ≈ h × ( h − 2 1 ) 2 = 1.
When z ~ \widetilde{z} z and R ~ 2 \widetilde{\mathbf{R}}_2 R 2 stay bounded, the terms in their drifts that do not contain σ \sigma σ or σ 2 \sigma^2 σ 2 contribute O ( h ) O(h) O ( h ) . Thus the variance factors can change by an amount of order one while z ~ \widetilde{z} z remains close to z 0 z_0 z 0 . The proof uses a positive path s h s^h s h on [ 0 , h ] [0,h] [ 0 , h ] with s 0 h = σ 0 s_0^h=\sigma_0 s 0 h = σ 0 and
s h h = h 2 , ∫ 0 h s t h d t → h ↓ 0 0 , ∫ 0 h ( s t h ) 2 d t → h ↓ 0 A s_h^h=h^2,\qquad\int_0^h s_t^h\,dt\xrightarrow[h\downarrow0]{}0,\qquad\int_0^h(s_t^h)^2\,dt\xrightarrow[h\downarrow0]{}A s h h = h 2 , ∫ 0 h s t h d t h ↓ 0 0 , ∫ 0 h ( s t h ) 2 d t h ↓ 0 A
for a chosen A ≥ 0 A\geq0 A ≥ 0 . The controlled factors then satisfy
z ~ h → h ↓ 0 z 0 , R ~ 2 , h ( λ 2 , j ) → h ↓ 0 R 2 , 0 ( λ 2 , j ) + λ 2 , j A . \widetilde{z}_h\xrightarrow[h\downarrow0]{}z_0,\qquad\widetilde{R}_{2,h}(\lambda_{2,j})\xrightarrow[h\downarrow0]{}R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A. z h h ↓ 0 z 0 , R 2 , h ( λ 2 , j ) h ↓ 0 R 2 , 0 ( λ 2 , j ) + λ 2 , j A .
We choose A A A so that
μ σ ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) < 0. \mu_\sigma\left(0,z_0,\left(R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A\right)_{1\leq j\leq N}\right)<0. μ σ ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) < 0.
Since s h h = h 2 = o ( h ) s_h^h=h^2=o(h) s h h = h 2 = o ( h ) , setting u = 0 u=0 u = 0 on [ h , 2 h ] [h,2h] [ h , 2 h ] makes σ u \sigma^u σ u negative for all sufficiently small h h h . A control on [ 2 h , T ] [2h,T] [ 2 h , T ] keeps σ u \sigma^u σ u at its negative value at time 2 h 2h 2 h .
Proof of Theorem 3. We fix T > 0 T>0 T > 0 .
Step 1: Construction of a control. Since z i , 0 = η 0 ( λ 1 , i ) + β 0 + β 2 R ‾ 2 , 0 z_{i,0}=\eta_0(\lambda_{1,i})+\beta_0+\beta_2\sqrt{\overline{R}_{2,0}} z i , 0 = η 0 ( λ 1 , i ) + β 0 + β 2 R 2 , 0 (by (49) and Notation 3) and λ 1 , e f f − ∑ i = 1 M − 1 λ ^ i = λ 1 , M \lambda_{1,\mathrm{eff}}-\sum_{i=1}^{M-1}\widehat{\lambda}_i=\lambda_{1,M} λ 1 , eff − ∑ i = 1 M − 1 λ i = λ 1 , M , we have
μ σ ( 0 , z 0 , R 2 , 0 ) = β 0 ( λ 1 , e f f − ∑ i = 1 M − 1 λ ^ i ) − ∑ i = 1 M − 1 λ ^ i η 0 ( λ 1 , i ) + β 2 ( λ 1 , e f f − ∑ i = 1 M − 1 λ ^ i − λ 2 ( R 2 , 0 ) 2 ) R ‾ 2 , 0 = β 0 λ 1 , M − ∑ i = 1 M − 1 λ ^ i η 0 ( λ 1 , i ) + β 2 ( λ 1 , M − λ 2 ( R 2 , 0 ) 2 ) R ‾ 2 , 0 . (55) \begin{aligned}
\mu_\sigma(0,z_0,R_{2,0})={}&\beta_0\left(\lambda_{1,\mathrm{eff}}-\sum_{i=1}^{M-1}\widehat{\lambda}_i\right)-\sum_{i=1}^{M-1}\widehat{\lambda}_i\eta_0(\lambda_{1,i})\\
&+\beta_2\left(\lambda_{1,\mathrm{eff}}-\sum_{i=1}^{M-1}\widehat{\lambda}_i-\frac{\lambda_2(R_{2,0})}{2}\right)\sqrt{\overline{R}_{2,0}}\\
={}&\beta_0\lambda_{1,M}-\sum_{i=1}^{M-1}\widehat{\lambda}_i\eta_0(\lambda_{1,i})+\beta_2\left(\lambda_{1,M}-\frac{\lambda_2(R_{2,0})}{2}\right)\sqrt{\overline{R}_{2,0}}. \tag*{(55)}
\end{aligned} μ σ ( 0 , z 0 , R 2 , 0 ) = = β 0 ( λ 1 , eff − i = 1 ∑ M − 1 λ i ) − i = 1 ∑ M − 1 λ i η 0 ( λ 1 , i ) + β 2 ( λ 1 , eff − i = 1 ∑ M − 1 λ i − 2 λ 2 ( R 2 , 0 ) ) R 2 , 0 β 0 λ 1 , M − i = 1 ∑ M − 1 λ i η 0 ( λ 1 , i ) + β 2 ( λ 1 , M − 2 λ 2 ( R 2 , 0 ) ) R 2 , 0 . (55)
Moreover, we have
∑ j = 1 N θ 2 , j ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) = R ‾ 2 , 0 + Λ 2 A , \sum_{j=1}^{N}\theta_{2,j}\left(R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A\right)=\overline{R}_{2,0}+\Lambda_2A, j = 1 ∑ N θ 2 , j ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) = R 2 , 0 + Λ 2 A ,
λ 2 ( ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) = λ 2 ( R 2 , 0 ) R ‾ 2 , 0 + Λ 2 λ 2 , e f f A R ‾ 2 , 0 + Λ 2 A → A → ∞ λ 2 , e f f . \lambda_2\left(\left(R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A\right)_{1\leq j\leq N}\right)=\frac{\lambda_2(R_{2,0})\overline{R}_{2,0}+\Lambda_2\lambda_{2,\mathrm{eff}}A}{\overline{R}_{2,0}+\Lambda_2A}\xrightarrow[A\to\infty]{}\lambda_{2,\mathrm{eff}}. λ 2 ( ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) = R 2 , 0 + Λ 2 A λ 2 ( R 2 , 0 ) R 2 , 0 + Λ 2 λ 2 , eff A A → ∞ λ 2 , eff .
Substitution in (54) gives, for A > 0 A>0 A > 0 ,
μ σ ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) A = λ 1 , e f f β 0 − ∑ i = 1 M − 1 λ ^ i z i , 0 A + β 2 ( λ 1 , e f f − λ 2 ( R 2 , 0 ) R ‾ 2 , 0 + Λ 2 λ 2 , e f f A 2 ( R ‾ 2 , 0 + Λ 2 A ) ) Λ 2 + R ‾ 2 , 0 A → A → ∞ β 2 Λ 2 ( λ 1 , e f f − λ 2 , e f f 2 ) . (56) \begin{aligned}
\frac{\mu_\sigma\left(0,z_0,\left(R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A\right)_{1\leq j\leq N}\right)}{\sqrt{A}}
={}&\frac{\lambda_{1,\mathrm{eff}}\beta_0-\sum_{i=1}^{M-1}\widehat{\lambda}_iz_{i,0}}{\sqrt{A}}\\
&+\beta_2\left(\lambda_{1,\mathrm{eff}}-\frac{\lambda_2(R_{2,0})\overline{R}_{2,0}+\Lambda_2\lambda_{2,\mathrm{eff}}A}{2(\overline{R}_{2,0}+\Lambda_2A)}\right)\sqrt{\Lambda_2+\frac{\overline{R}_{2,0}}{A}}\\
&\xrightarrow[A\to\infty]{}\beta_2\sqrt{\Lambda_2}\left(\lambda_{1,\mathrm{eff}}-\frac{\lambda_{2,\mathrm{eff}}}{2}\right). \tag*{(56)}
\end{aligned} A μ σ ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) = A λ 1 , eff β 0 − ∑ i = 1 M − 1 λ i z i , 0 + β 2 ( λ 1 , eff − 2 ( R 2 , 0 + Λ 2 A ) λ 2 ( R 2 , 0 ) R 2 , 0 + Λ 2 λ 2 , eff A ) Λ 2 + A R 2 , 0 A → ∞ β 2 Λ 2 ( λ 1 , eff − 2 λ 2 , eff ) . (56)
If condition (48) of Theorem 3 holds, A = 0 A=0 A = 0 makes the drift negative by the identity (55) . If condition (47) holds, every sufficiently large A A A makes the drift negative by the limit (56) . Thus there exists A ≥ 0 A\geq0 A ≥ 0 such that
μ σ ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) < 0. (57) \mu_\sigma\left(0,z_0,\left(R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A\right)_{1\leq j\leq N}\right)<0. \tag*{(57)} μ σ ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) < 0. (57)
For h ∈ ( 0 , 1 ] h\in(0,1] h ∈ ( 0 , 1 ] , we define
∀ t ∈ [ 0 , h ] , s t h : = σ 0 ( 1 − t h ) + h t + 2 A h sin ( π t h ) . \forall t\in[0,h],\qquad s_t^h:=\sigma_0\left(1-\frac{t}{h}\right)+ht+\sqrt{\frac{2A}{h}}\sin\left(\frac{\pi t}{h}\right). ∀ t ∈ [ 0 , h ] , s t h := σ 0 ( 1 − h t ) + h t + h 2 A sin ( h π t ) .
For fixed h h h and every t ∈ [ 0 , h ] t\in[0,h] t ∈ [ 0 , h ] ,
s h h = h 2 , s t h ≥ σ 0 ( 1 − t h ) + h t ≥ min { σ 0 , h 2 } > 0. s_h^h=h^2,\qquad s_t^h\geq\sigma_0\left(1-\frac{t}{h}\right)+ht\geq\min\{\sigma_0,h^2\}>0. s h h = h 2 , s t h ≥ σ 0 ( 1 − h t ) + h t ≥ min { σ 0 , h 2 } > 0.
By (53) and (1) ,
∀ t ∈ [ 0 , h ] , ∑ j = 1 N θ 2 , j R ~ 2 , t ( λ 2 , j ) ≥ e − λ 2 , 1 h R ‾ 2 , 0 > 0. \forall t\in[0,h],\qquad\sum_{j=1}^{N}\theta_{2,j}\widetilde{R}_{2,t}(\lambda_{2,j})\geq e^{-\lambda_{2,1}h}\overline{R}_{2,0}>0. ∀ t ∈ [ 0 , h ] , j = 1 ∑ N θ 2 , j R 2 , t ( λ 2 , j ) ≥ e − λ 2 , 1 h R 2 , 0 > 0.
Thus μ σ ( s t h , z ~ t , R ~ 2 , t ) \mu_{\sigma}(s_t^h,\widetilde{z}_t,\widetilde{\mathbf{R}}_{2,t}) μ σ ( s t h , z t , R 2 , t ) is well defined and β 1 Λ 1 s t h ≠ 0 \beta_1\Lambda_1s_t^h \ne0 β 1 Λ 1 s t h = 0 . The control on [ 0 , h ] [0,h] [ 0 , h ] is therefore well defined by
∀ t ∈ [ 0 , h ] , u t : = s ˙ t h − μ σ ( s t h , z ~ t , R ~ 2 , t ) β 1 Λ 1 s t h . \forall t \in[0,h],\qquad u_t:=\frac{\dot{s}_t^h-\mu_{\sigma}(s_t^h,\widetilde{z}_t,\widetilde{\mathbf{R}}_{2,t})}{\beta_1\Lambda_1s_t^h}. ∀ t ∈ [ 0 , h ] , u t := β 1 Λ 1 s t h s ˙ t h − μ σ ( s t h , z t , R 2 , t ) .
It is continuous, hence bounded on [ 0 , h ] [0,h] [ 0 , h ] . This gives σ t u = s t h \sigma_t^u=s_t^h σ t u = s t h on [ 0 , h ] [0,h] [ 0 , h ] . Uniqueness of the factor equations then gives z t u = z ~ t z_t^u=\widetilde{z}_t z t u = z t and R 2 , t u = R ~ 2 , t \mathbf{R}_{2,t}^u=\widetilde{\mathbf{R}}_{2,t} R 2 , t u = R 2 , t for 0 ≤ t ≤ h 0\le t\le h 0 ≤ t ≤ h . Moreover,
∫ 0 h s t h d t = σ 0 h 2 + h 3 2 + 2 2 A π h = O ( h ) , \int_0^h s_t^h\,\mathrm{d}t=\frac{\sigma_0h}{2}+\frac{h^3}{2}+\frac{2\sqrt{2A}}{\pi}\sqrt{h}=O(\sqrt{h}), ∫ 0 h s t h d t = 2 σ 0 h + 2 h 3 + π 2 2 A h = O ( h ) ,
∫ 0 h ( s t h ) 2 d t = A + 2 2 A π ( σ 0 + h 2 ) h + h 3 ( σ 0 2 + σ 0 h 2 + h 4 ) = A + O ( h ) . \int_0^h (s_t^h)^2\,\mathrm{d}t=A+\frac{2\sqrt{2A}}{\pi}(\sigma_0+h^2)\sqrt{h}+\frac{h}{3}(\sigma_0^2+\sigma_0h^2+h^4)=A+O(\sqrt{h}). ∫ 0 h ( s t h ) 2 d t = A + π 2 2 A ( σ 0 + h 2 ) h + 3 h ( σ 0 2 + σ 0 h 2 + h 4 ) = A + O ( h ) .
Hence, the variance factors satisfy
R ~ 2 , h ( λ 2 , j ) = e − λ 2 , j h R 2 , 0 ( λ 2 , j ) + λ 2 , j ∫ 0 h e − λ 2 , j ( h − t ) ( s t h ) 2 d t = R 2 , 0 ( λ 2 , j ) + λ 2 , j ∫ 0 h ( s t h ) 2 d t + O ( h ) = R 2 , 0 ( λ 2 , j ) + λ 2 , j A + O ( h ) . (58) \begin{aligned}
\widetilde{R}_{2,h}(\lambda_{2,j})&=e^{-\lambda_{2,j}h}R_{2,0}(\lambda_{2,j})+\lambda_{2,j}\int_0^h e^{-\lambda_{2,j}(h-t)}(s_t^h)^2\,\mathrm{d}t\\
&=R_{2,0}(\lambda_{2,j})+\lambda_{2,j}\int_0^h(s_t^h)^2\,\mathrm{d}t+O(h)\\
&=R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A+O(\sqrt{h}). \tag*{(58)}
\end{aligned} R 2 , h ( λ 2 , j ) = e − λ 2 , j h R 2 , 0 ( λ 2 , j ) + λ 2 , j ∫ 0 h e − λ 2 , j ( h − t ) ( s t h ) 2 d t = R 2 , 0 ( λ 2 , j ) + λ 2 , j ∫ 0 h ( s t h ) 2 d t + O ( h ) = R 2 , 0 ( λ 2 , j ) + λ 2 , j A + O ( h ) . (58)
The O ( h ) O(h) O ( h ) remainder in (58) follows from 0 ≤ 1 − e − x ≤ x 0\le1-e^{-x}\le x 0 ≤ 1 − e − x ≤ x for x ≥ 0 x\ge0 x ≥ 0 and
∣ R ~ 2 , h ( λ 2 , j ) − R 2 , 0 ( λ 2 , j ) − λ 2 , j ∫ 0 h ( s t h ) 2 d t ∣ ≤ λ 2 , j h R 2 , 0 ( λ 2 , j ) + λ 2 , j 2 h ∫ 0 h ( s t h ) 2 d t = O ( h ) . \left|\widetilde{R}_{2,h}(\lambda_{2,j})-R_{2,0}(\lambda_{2,j})-\lambda_{2,j}\int_0^h(s_t^h)^2\,\mathrm{d}t\right|\le\lambda_{2,j}hR_{2,0}(\lambda_{2,j})+\lambda_{2,j}^2h\int_0^h(s_t^h)^2\,\mathrm{d}t=O(h). R 2 , h ( λ 2 , j ) − R 2 , 0 ( λ 2 , j ) − λ 2 , j ∫ 0 h ( s t h ) 2 d t ≤ λ 2 , j h R 2 , 0 ( λ 2 , j ) + λ 2 , j 2 h ∫ 0 h ( s t h ) 2 d t = O ( h ) .
Moreover, we have
∀ t ∈ [ 0 , h ] , ∀ j ∈ { 1 , … , N } , 0 ≤ R ~ 2 , t ( λ 2 , j ) ≤ R 2 , 0 ( λ 2 , j ) + λ 2 , j ∫ 0 h ( s r h ) 2 d r ≤ C . \forall t\in[0,h],\ \forall j\in\{1,\ldots,N\},\qquad0\le\widetilde{R}_{2,t}(\lambda_{2,j})\le R_{2,0}(\lambda_{2,j})+\lambda_{2,j}\int_0^h(s_r^h)^2\,\mathrm{d}r\le C. ∀ t ∈ [ 0 , h ] , ∀ j ∈ { 1 , … , N } , 0 ≤ R 2 , t ( λ 2 , j ) ≤ R 2 , 0 ( λ 2 , j ) + λ 2 , j ∫ 0 h ( s r h ) 2 d r ≤ C .
In the estimates on [ 0 , h ] [0,h] [ 0 , h ] , C C C denotes a constant depending only on A A A , the parameters and the initial state. The equation (52) for z z z therefore gives
∣ z ~ ˙ t ∣ ≤ C ( 1 + ∣ z ~ t ∣ + s t h ) . |\dot{\widetilde{z}}_t|\le C(1+|\widetilde{z}_t|+s_t^h). ∣ z ˙ t ∣ ≤ C ( 1 + ∣ z t ∣ + s t h ) .
Since ∫ 0 h s t h d t = O ( h ) \int_0^h s_t^h\,\mathrm{d}t=O(\sqrt{h}) ∫ 0 h s t h d t = O ( h ) , Grönwall's lemma yields
sup t ∈ [ 0 , h ] ∣ z ~ t ∣ ≤ C , ∣ z ~ h − z 0 ∣ ≤ C ∫ 0 h ( 1 + ∣ z ~ t ∣ + s t h ) d t = O ( h ) . \sup_{t\in[0,h]}|\widetilde{z}_t|\le C,\qquad|\widetilde{z}_h-z_0|\le C\int_0^h(1+|\widetilde{z}_t|+s_t^h)\,\mathrm{d}t=O(\sqrt{h}). t ∈ [ 0 , h ] sup ∣ z t ∣ ≤ C , ∣ z h − z 0 ∣ ≤ C ∫ 0 h ( 1 + ∣ z t ∣ + s t h ) d t = O ( h ) .
Since R ‾ 2 , 0 + Λ 2 A > 0 \overline{R}_{2,0}+\Lambda_2A>0 R 2 , 0 + Λ 2 A > 0 by (1) , (54) shows that the coefficients in (3.42) are bounded and Lipschitz on a neighborhood of ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) (0,z_0,(R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A)_{1\le j\le N}) ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) . For sufficiently small h h h , we extend the control by setting u t : = 0 u_t:=0 u t := 0 for t ∈ [ h , 2 h ] t\in[h,2h] t ∈ [ h , 2 h ] . The bounded and Lipschitz coefficients then give
sup t ∈ [ h , 2 h ] ( ∣ σ t u − h 2 ∣ + ∣ z t u − z ~ h ∣ + ∣ R 2 , t u − R ~ 2 , h ∣ ) = O ( h ) . \sup_{t\in[h,2h]}\left(|\sigma_t^u-h^2|+|z_t^u-\widetilde{z}_h|+|\mathbf{R}_{2,t}^u-\widetilde{\mathbf{R}}_{2,h}|\right)=O(h). t ∈ [ h , 2 h ] sup ( ∣ σ t u − h 2 ∣ + ∣ z t u − z h ∣ + ∣ R 2 , t u − R 2 , h ∣ ) = O ( h ) .
Together with z ~ h = z 0 + O ( h ) \widetilde{z}_h=z_0+O(\sqrt{h}) z h = z 0 + O ( h ) and the expansion (58) , this gives
σ 2 h u = h 2 + ∫ h 2 h μ σ ( σ t u , z t u , R 2 , t u ) d t = h 2 + h μ σ ( h 2 , z ~ h , R ~ 2 , h ) + O ( h 2 ) = h μ σ ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) + O ( h 3 / 2 ) . (59) \begin{aligned}
\sigma_{2h}^u&=h^2+\int_h^{2h}\mu_{\sigma}(\sigma_t^u,z_t^u,\mathbf{R}_{2,t}^u)\,\mathrm{d}t\\
&=h^2+h\mu_{\sigma}(h^2,\widetilde{z}_h,\widetilde{\mathbf{R}}_{2,h})+O(h^2)\\
&=h\mu_{\sigma}(0,z_0,(R_{2,0}(\lambda_{2,j})+\lambda_{2,j}A)_{1\le j\le N})+O(h^{3/2}). \tag*{(59)}
\end{aligned} σ 2 h u = h 2 + ∫ h 2 h μ σ ( σ t u , z t u , R 2 , t u ) d t = h 2 + h μ σ ( h 2 , z h , R 2 , h ) + O ( h 2 ) = h μ σ ( 0 , z 0 , ( R 2 , 0 ( λ 2 , j ) + λ 2 , j A ) 1 ≤ j ≤ N ) + O ( h 3/2 ) . (59)
For a fixed sufficiently small h > 0 h>0 h > 0 with 2 h < T 2h<T 2 h < T , the inequality (57) and the expansion (59) give c : = σ 2 h u < 0 c:=\sigma^u_{2h}<0 c := σ 2 h u < 0 . Since β 1 < 0 \beta_1<0 β 1 < 0 , Λ 1 > 0 \Lambda_1>0 Λ 1 > 0 , and c < 0 c<0 c < 0 , the denominator β 1 Λ 1 c \beta_1\Lambda_1c β 1 Λ 1 c is nonzero. We extend the control to [ 2 h , T ] [2h,T] [ 2 h , T ] by setting
∀ t ∈ [ 2 h , T ] , u t : = − μ σ ( c , z t u , R 2 , t u ) β 1 Λ 1 c . \forall t\in[2h,T],\qquad u_t:=-\frac{\mu_\sigma(c,z^u_t,R^u_{2,t})}{\beta_1\Lambda_1c}. ∀ t ∈ [ 2 h , T ] , u t := − β 1 Λ 1 c μ σ ( c , z t u , R 2 , t u ) .
This choice gives σ ˙ t u = 0 \dot{\sigma}^u_t=0 σ ˙ t u = 0 , so σ t u = c \sigma^u_t=c σ t u = c , and
∀ t ∈ [ 2 h , T ] , ∀ j ∈ { 1 , … , N } , R 2 , t u ( λ 2 , j ) = e − λ 2 , j ( t − 2 h ) R 2 , 2 h u ( λ 2 , j ) + ( 1 − e − λ 2 , j ( t − 2 h ) ) c 2 . \forall t\in[2h,T],\ \forall j\in\{1,\ldots,N\},\qquad R^u_{2,t}(\lambda_{2,j})=e^{-\lambda_{2,j}(t-2h)}R^u_{2,2h}(\lambda_{2,j})+\left(1-e^{-\lambda_{2,j}(t-2h)}\right)c^2. ∀ t ∈ [ 2 h , T ] , ∀ j ∈ { 1 , … , N } , R 2 , t u ( λ 2 , j ) = e − λ 2 , j ( t − 2 h ) R 2 , 2 h u ( λ 2 , j ) + ( 1 − e − λ 2 , j ( t − 2 h ) ) c 2 .
This formula gives, for every j ∈ { 1 , … , N } j\in\{1,\ldots,N\} j ∈ { 1 , … , N } ,
sup t ∈ [ 0 , T ] R 2 , t u ( λ 2 , j ) ≤ max { sup t ∈ [ 0 , 2 h ] R 2 , t u ( λ 2 , j ) , c 2 } < ∞ . \sup_{t\in[0,T]}R^u_{2,t}(\lambda_{2,j})\leq\max\left\{\sup_{t\in[0,2h]}R^u_{2,t}(\lambda_{2,j}),c^2\right\}<\infty. t ∈ [ 0 , T ] sup R 2 , t u ( λ 2 , j ) ≤ max { t ∈ [ 0 , 2 h ] sup R 2 , t u ( λ 2 , j ) , c 2 } < ∞.
We write ∥ ⋅ ∥ ∞ \|\cdot\|_\infty ∥ ⋅ ∥ ∞ for the supremum norm on [ 0 , T ] [0,T] [ 0 , T ] . Since R 2 u R^u_2 R 2 u is bounded and σ u = c \sigma^u=c σ u = c on [ 2 h , T ] [2h,T] [ 2 h , T ] , the equation (52) and Grönwall’s lemma give
∥ z u ∥ ∞ + ∥ R 2 u ∥ ∞ < ∞ . \|z^u\|_\infty+\|R^u_2\|_\infty<\infty. ∥ z u ∥ ∞ + ∥ R 2 u ∥ ∞ < ∞.
The variance equations (53) and the nonnegativity of ( σ u ) 2 (\sigma^u)^2 ( σ u ) 2 also give
∀ t ∈ [ 0 , T ] , ∀ j ∈ { 1 , … , N } , R 2 , t u ( λ 2 , j ) ≥ e − λ 2 , j t R 2 , 0 ( λ 2 , j ) . \forall t\in[0,T],\ \forall j\in\{1,\ldots,N\},\qquad R^u_{2,t}(\lambda_{2,j})\geq e^{-\lambda_{2,j}t}R_{2,0}(\lambda_{2,j}). ∀ t ∈ [ 0 , T ] , ∀ j ∈ { 1 , … , N } , R 2 , t u ( λ 2 , j ) ≥ e − λ 2 , j t R 2 , 0 ( λ 2 , j ) .
Summing with weights θ 2 , j \theta_{2,j} θ 2 , j and using R ‾ 2 , 0 > 0 \overline{R}_{2,0}>0 R 2 , 0 > 0 from (1) yields
inf t ∈ [ 0 , T ] R ‾ 2 , t u ≥ e − λ 2 , 1 T R ‾ 2 , 0 > 0. (60) \inf_{t\in[0,T]}\overline{R}^u_{2,t}\geq e^{-\lambda_{2,1}T}\overline{R}_{2,0}>0. \tag*{(60)} t ∈ [ 0 , T ] inf R 2 , t u ≥ e − λ 2 , 1 T R 2 , 0 > 0. (60)
The bound (60) keeps the controlled path in the domain of μ σ \mu_\sigma μ σ . Together with β 1 Λ 1 c ≠ 0 \beta_1\Lambda_1c\neq0 β 1 Λ 1 c = 0 and the boundedness of z u z^u z u and R 2 u R^u_2 R 2 u , it ensures that the control on [ 2 h , T ] [2h,T] [ 2 h , T ] is well defined and bounded. Thus u ∈ L ∞ ( [ 0 , T ] ) ⊂ L 2 ( [ 0 , T ] ) u\in L^\infty([0,T])\subset L^2([0,T]) u ∈ L ∞ ([ 0 , T ]) ⊂ L 2 ([ 0 , T ]) and σ T u = c < 0 \sigma^u_T=c<0 σ T u = c < 0 .
Step 2: Stability and the Cameron–Martin theorem. We set
∀ t ∈ [ 0 , T ] , w t : = ∫ 0 t ( u s + β 1 Λ 1 2 ) d s , G t : = exp ( β 1 Λ 1 ( W t − w t ) ) . \forall t\in[0,T],\qquad w_t:=\int_0^t\left(u_s+\frac{\beta_1\Lambda_1}{2}\right)\,\mathrm{d}s,\qquad G_t:=\exp\left(\beta_1\Lambda_1(W_t-w_t)\right). ∀ t ∈ [ 0 , T ] , w t := ∫ 0 t ( u s + 2 β 1 Λ 1 ) d s , G t := exp ( β 1 Λ 1 ( W t − w t ) ) .
The definition of w w w gives
G t = exp ( β 1 Λ 1 W t − β 1 Λ 1 ∫ 0 t u s d s − β 1 2 Λ 1 2 2 t ) . G_t=\exp\left(\beta_1\Lambda_1W_t-\beta_1\Lambda_1\int_0^t u_s\,\mathrm{d}s-\frac{\beta_1^2\Lambda_1^2}{2}t\right). G t = exp ( β 1 Λ 1 W t − β 1 Λ 1 ∫ 0 t u s d s − 2 β 1 2 Λ 1 2 t ) .
Since w ˙ t = u t + β 1 Λ 1 2 \dot{w}_t=u_t+\frac{\beta_1\Lambda_1}{2} w ˙ t = u t + 2 β 1 Λ 1 , Itô’s formula and (51) give
d ( G t − 1 ) = ( β 1 Λ 1 u t + β 1 2 Λ 1 2 ) G t − 1 d t − β 1 Λ 1 G t − 1 d W t , \mathrm{d}(G_t^{-1})=\left(\beta_1\Lambda_1u_t+\beta_1^2\Lambda_1^2\right)G_t^{-1}\,\mathrm{d}t-\beta_1\Lambda_1G_t^{-1}\,\mathrm{d}W_t, d ( G t − 1 ) = ( β 1 Λ 1 u t + β 1 2 Λ 1 2 ) G t − 1 d t − β 1 Λ 1 G t − 1 d W t ,
d [ σ , G − 1 ] t = − β 1 2 Λ 1 2 σ t G t − 1 d t . \mathrm{d}[\sigma,G^{-1}]_t=-\beta_1^2\Lambda_1^2\sigma_tG_t^{-1}\,\mathrm{d}t. d [ σ , G − 1 ] t = − β 1 2 Λ 1 2 σ t G t − 1 d t .
Since G > 0 G>0 G > 0 , the ratio σ G \frac{\sigma}{G} G σ is well defined and satisfies, for almost every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
d d t ( σ t G t ) = G t − 1 μ σ ( σ t , z t , R 2 , t ) + β 1 Λ 1 u t σ t G t . (61) \frac{\mathrm{d}}{\mathrm{d}t}\left(\frac{\sigma_t}{G_t}\right)=G_t^{-1}\mu_\sigma(\sigma_t,z_t,R_{2,t})+\beta_1\Lambda_1u_t\frac{\sigma_t}{G_t}. \tag*{(61)} d t d ( G t σ t ) = G t − 1 μ σ ( σ t , z t , R 2 , t ) + β 1 Λ 1 u t G t σ t . (61)
For G = 1 G=1 G = 1 , (61) reduces to (51) with d W t \mathrm{d}W_t d W t replaced by u t d t u_t\,\mathrm{d}t u t d t . By continuity of the controlled path and the lower bound (60) , for sufficiently small ρ > 0 \rho>0 ρ > 0 , the open set
U ρ : = { ( σ , z , r ) : inf s ∈ [ 0 , T ] ( ∣ σ − σ s u ∣ + ∣ z − z s u ∣ + ∣ r − R 2 , s u ∣ ) < ρ } . \mathcal{U}_\rho:=\left\{(\sigma,z,r):\inf_{s\in[0,T]}\left(|\sigma-\sigma^u_s|+|z-z^u_s|+|r-R^u_{2,s}|\right)<\rho\right\}. U ρ := { ( σ , z , r ) : s ∈ [ 0 , T ] inf ( ∣ σ − σ s u ∣ + ∣ z − z s u ∣ + ∣ r − R 2 , s u ∣ ) < ρ } .
has compact closure in { ( σ , z , r ) : ∑ j = 1 N θ 2 , j r j > 0 } \{(\sigma,z,r): \sum_{j=1}^{N}\theta_{2,j}r_{j}>0\} {( σ , z , r ) : ∑ j = 1 N θ 2 , j r j > 0 } . The coefficients in (3.42) are bounded and Lipschitz on U ‾ ρ \overline{\mathcal{U}}_{\rho} U ρ by (54) . For 0 < ε ≤ 1 0<\varepsilon\leq1 0 < ε ≤ 1 , on { ∥ W − w ∥ ∞ < ε } \{\lVert W-w\rVert_{\infty}<\varepsilon\} {∥ W − w ∥ ∞ < ε } , we have
∥ G − 1 ∥ ∞ + ∥ G − 1 − 1 ∥ ∞ ≤ C ε . (62) \lVert G-1\rVert_{\infty}+\lVert G^{-1}-1\rVert_{\infty}\leq C\varepsilon. \tag*{(62)} ∥ G − 1 ∥ ∞ + ∥ G − 1 − 1 ∥ ∞ ≤ C ε . (62)
Up to the first exit of ( σ , z , R 2 ) (\sigma,z,\mathbf{R}_{2}) ( σ , z , R 2 ) from U ρ \mathcal{U}_{\rho} U ρ , boundedness and Lipschitz continuity of μ σ \mu_{\sigma} μ σ , together with the bound (62) , give on U ‾ ρ \overline{\mathcal{U}}_{\rho} U ρ
∣ G t − 1 μ σ ( σ t , z t , R 2 , t ) − μ σ ( σ t u , z t u , R 2 , t u ) ∣ ≤ ∣ G t − 1 − 1 ∣ ∣ μ σ ( σ t , z t , R 2 , t ) ∣ + C ( ∣ σ t − σ t u ∣ + ∣ z t − z t u ∣ + ∣ R 2 , t − R 2 , t u ∣ ) ≤ C ε + C ( ∣ σ t − σ t u ∣ + ∣ z t − z t u ∣ + ∣ R 2 , t − R 2 , t u ∣ ) . (63) \begin{aligned}
\left\lvert G_{t}^{-1}\mu_{\sigma}(\sigma_{t},z_{t},\mathbf{R}_{2,t})-\mu_{\sigma}(\sigma_{t}^{u},z_{t}^{u},\mathbf{R}_{2,t}^{u})\right\rvert
&\leq\left\lvert G_{t}^{-1}-1\right\rvert\left\lvert\mu_{\sigma}(\sigma_{t},z_{t},\mathbf{R}_{2,t})\right\rvert\\
&\quad+C\left(\left\lvert\sigma_{t}-\sigma_{t}^{u}\right\rvert+\left\lvert z_{t}-z_{t}^{u}\right\rvert+\left\lvert\mathbf{R}_{2,t}-\mathbf{R}_{2,t}^{u}\right\rvert\right) \\
&\leq C\varepsilon+C\left(\left\lvert\sigma_{t}-\sigma_{t}^{u}\right\rvert+\left\lvert z_{t}-z_{t}^{u}\right\rvert+\left\lvert\mathbf{R}_{2,t}-\mathbf{R}_{2,t}^{u}\right\rvert\right). \tag*{(63)}
\end{aligned} G t − 1 μ σ ( σ t , z t , R 2 , t ) − μ σ ( σ t u , z t u , R 2 , t u ) ≤ G t − 1 − 1 ∣ μ σ ( σ t , z t , R 2 , t ) ∣ + C ( ∣ σ t − σ t u ∣ + ∣ z t − z t u ∣ + R 2 , t − R 2 , t u ) ≤ C ε + C ( ∣ σ t − σ t u ∣ + ∣ z t − z t u ∣ + R 2 , t − R 2 , t u ) . (63)
Moreover, the bound (62) and boundedness of σ u \sigma^{u} σ u imply
∣ σ t − σ t u ∣ = ∣ G t ( σ t G t − σ t u ) + ( G t − 1 ) σ t u ∣ ≤ C ∣ σ t G t − σ t u ∣ + C ε . (64) \left\lvert\sigma_{t}-\sigma_{t}^{u}\right\rvert=\left\lvert G_{t}\left(\frac{\sigma_{t}}{G_{t}}-\sigma_{t}^{u}\right)+(G_{t}-1)\sigma_{t}^{u}\right\rvert\leq C\left\lvert\frac{\sigma_{t}}{G_{t}}-\sigma_{t}^{u}\right\rvert+C\varepsilon. \tag*{(64)} ∣ σ t − σ t u ∣ = G t ( G t σ t − σ t u ) + ( G t − 1 ) σ t u ≤ C G t σ t − σ t u + C ε . (64)
We compare (61) , (52) , and (53) with the controlled equations. Estimates (63) and (64) give
∣ σ t G t − σ t u ∣ + ∣ z t − z t u ∣ + ∣ R 2 , t − R 2 , t u ∣ ≤ C ε \left\lvert\frac{\sigma_{t}}{G_{t}}-\sigma_{t}^{u}\right\rvert+\left\lvert z_{t}-z_{t}^{u}\right\rvert+\left\lvert\mathbf{R}_{2,t}-\mathbf{R}_{2,t}^{u}\right\rvert\leq C\varepsilon G t σ t − σ t u + ∣ z t − z t u ∣ + R 2 , t − R 2 , t u ≤ C ε
+ C ∫ 0 t ( 1 + ∣ u s ∣ ) ( ∣ σ s G s − σ s u ∣ + ∣ z s − z s u ∣ + ∣ R 2 , s − R 2 , s u ∣ ) d s . \quad+C\int_{0}^{t}(1+\lvert u_{s}\rvert)\left(\left\lvert\frac{\sigma_{s}}{G_{s}}-\sigma_{s}^{u}\right\rvert+\left\lvert z_{s}-z_{s}^{u}\right\rvert+\left\lvert\mathbf{R}_{2,s}-\mathbf{R}_{2,s}^{u}\right\rvert\right)\,ds. + C ∫ 0 t ( 1 + ∣ u s ∣) ( G s σ s − σ s u + ∣ z s − z s u ∣ + R 2 , s − R 2 , s u ) d s .
Since u ∈ L 2 ( [ 0 , T ] ) ⊂ L 1 ( [ 0 , T ] ) u\in L^{2}([0,T])\subset L^{1}([0,T]) u ∈ L 2 ([ 0 , T ]) ⊂ L 1 ([ 0 , T ]) , Grönwall’s lemma and the bound (64) give
sup s ∈ [ 0 , t ] ( ∣ σ s − σ s u ∣ + ∣ z s − z s u ∣ + ∣ R 2 , s − R 2 , s u ∣ ) ≤ C ε , \sup_{s\in[0,t]}\left(\left\lvert\sigma_{s}-\sigma_{s}^{u}\right\rvert+\left\lvert z_{s}-z_{s}^{u}\right\rvert+\left\lvert\mathbf{R}_{2,s}-\mathbf{R}_{2,s}^{u}\right\rvert\right)\leq C\varepsilon, s ∈ [ 0 , t ] sup ( ∣ σ s − σ s u ∣ + ∣ z s − z s u ∣ + R 2 , s − R 2 , s u ) ≤ C ε ,
where C C C does not depend on ε \varepsilon ε . For C ε < min { ρ , − c 2 } C\varepsilon<\min\{\rho,-\frac{c}{2}\} C ε < min { ρ , − 2 c } , the uniform bound on σ − σ u \sigma-\sigma^{u} σ − σ u , z − z u z-z^{u} z − z u , and R 2 − R 2 u \mathbf{R}_{2}-\mathbf{R}_{2}^{u} R 2 − R 2 u , together with continuity of σ , z , R 2 \sigma,z,\mathbf{R}_{2} σ , z , R 2 , excludes an exit from U ρ \mathcal{U}_{\rho} U ρ on [ 0 , T ] [0,T] [ 0 , T ] . Hence the terminal volatility satisfies
σ T ≤ σ T u + C ε = c + C ε < 0. \sigma_{T}\leq\sigma_{T}^{u}+C\varepsilon=c+C\varepsilon<0. σ T ≤ σ T u + C ε = c + C ε < 0.
Therefore the events satisfy { ∥ W − w ∥ ∞ < ε } ⊂ { σ T < 0 } \{\lVert W-w\rVert_{\infty}<\varepsilon\}\subset\{\sigma_{T}<0\} {∥ W − w ∥ ∞ < ε } ⊂ { σ T < 0 } . Since
w ˙ = u + β 1 Λ 1 2 ε ∈ L 2 ( [ 0 , T ] ) , \dot{w}=u+\frac{\beta_{1}\Lambda_{1}}{2}\varepsilon\in L^{2}([0,T]), w ˙ = u + 2 β 1 Λ 1 ε ∈ L 2 ([ 0 , T ]) ,
the Cameron–Martin theorem and P ( ∥ W ∥ ∞ < ε ) > 0 \mathbb{P}(\lVert W\rVert_{\infty}<\varepsilon)>0 P (∥ W ∥ ∞ < ε ) > 0 give
P ( σ T < 0 ) ≥ P ( ∥ W − w ∥ ∞ < ε ) > 0. \mathbb{P}(\sigma_{T}<0)\geq\mathbb{P}(\lVert W-w\rVert_{\infty}<\varepsilon)>0. P ( σ T < 0 ) ≥ P (∥ W − w ∥ ∞ < ε ) > 0.
The result in [19 Corollary 2.6 (ii)] applies with t 0 = T 2 t_{0}=\frac{T}{2} t 0 = 2 T , since P ( σ T 2 < 0 ) > 0 \mathbb{P}(\sigma_{\frac{T}{2}}<0)>0 P ( σ 2 T < 0 ) > 0 . Hence S S S is a strict local martingale on [ 0 , T ] [0,T] [ 0 , T ] .
References